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.ppat.1002746 | Propagation of RML Prions in Mice Expressing PrP Devoid of GPI Anchor Leads to Formation of a Novel, Stable Prion Strain | PrPC, a host protein which in prion-infected animals is converted to PrPSc, is linked to the cell membrane by a GPI anchor. Mice expressing PrPC without GPI anchor (tgGPI- mice), are susceptible to prion infection but accumulate anchorless PrPSc extra-, rather than intracellularly. We investigated whether tgGPI− mice could faithfully propagate prion strains despite the deviant structure and location of anchorless PrPSc. We found that RML and ME7, but not 22L prions propagated in tgGPI− brain developed novel cell tropisms, as determined by the Cell Panel Assay (CPA). Surprisingly, the levels of proteinase K-resistant PrPSc (PrPres) in RML- or ME7-infected tgGPI− brain were 25–50 times higher than in wild-type brain. When returned to wild-type brain, ME7 prions recovered their original properties, however RML prions had given rise to a novel prion strain, designated SFL, which remained unchanged even after three passages in wild-type mice. Because both RML PrPSc and SFL PrPSc are stably propagated in wild-type mice we propose that the two conformations are separated by a high activation energy barrier which is abrogated in tgGPI− mice.
| The agent causing transmissible spongiform encephalopathies, the prion, consists mainly if not entirely of PrPSc, an aggregated conformational variant of PrPC. PrPC is a host-specified glycoprotein which is attached to the outer surface of the plasma membrane by a glycosylphosphatidylinositol (GPI) anchor. PrPSc is thought to propagate by seeded conversion, in which PrPC accretes to an aggregate of PrPSc and thereby assumes its conformation. Prions occur as distinct strains, which have the same PrP sequence but different PrPSc conformations. Many strains are very stable but some can “mutate”, especially after transient propagation in another animal species. Here “mutation” is attributed to a conformational change of the underlying PrPSc protein rather than a change in the nucleotide sequence of a gene. In transgenic mice that express anchorless PrP, conversion to PrPSc occurs extracellularly rather than in a cell-associated compartment. Surprisingly, RML, a very stable prion strain, when propagated in tgGPI− mice and then returned to wild-type mice, was recovered as a novel strain, designated SFL, which was stable over many transmissions. Thus, faithful strain propagation was abrogated either as a consequence of the modified PrP or because of a different replication mode in the extracellular compartment of tgGPI− brain.
| Prions, the causative agents of transmissible spongiform encephalopathies, contain as their main component PrPSc, a multimeric conformer of the ubiquitous host protein PrPC. PrPC can carry one or two N-linked glycans, whose structure is highly variable [1]–[3], or remain unglycosylated; moreover the glycosylation state varies in different tissues and cell lines, and even in different brain regions [4]–[6]. Most PrPC is attached to the outer surface of the plasma membrane by a glycosylphosphatidylinositol (GPI) anchor [7], but transmembrane forms have been identified [8], [9]. In prion-infected cells PrPC is converted to PrPSc at the membrane [10] and within the endosomal-lysosomal compartment [11], and accumulates mainly in intracellular compartments [12], [13]. While PrPC is readily cleaved off the cell surface by PIPLC [7], PrPres is not, even though it retains its GPI anchor [14]. PrPSc presents as a partially proteinase K (PK)-resistant form, designated rPrPSc or PrPres, or as a PK-sensitive entity, sPrPSc or PrPsen [15]–[17].
Mouse prions occur in the form of a wide variety of strains, which have the same PrP sequence but differ in their incubation time in various mouse lines and in the pattern of lesions they cause in brain [18]. It is believed that strain-ness is encoded by the conformation of PrPSc and in many instances there are differences in the physicochemical properties of the cognate PrPres [19]–[24]. Interestingly, different prion strains populate distinct regions of the brain [25]–[27] and are selective in regard to the cultured cell lines they can chronically infect [28]–[33]. This raises the unanswered question as to which cellular property underlies prion strain tropism.
The seeding or nucleation model for prion propagation [34]–[36] predicates that PrPC monomers add to PrPSc and in doing so assume the conformation of its subunits, thus allowing faithful propagation of strains. A prerequisite for chronic infection is that PrPC monomers with a conformation allowing efficient accretion be present at a concentration sufficient to allow synthesis of PrPSc at a rate higher than that of its depletion [37]–[39]. PrPC is conformationally very flexible [40]–[42] but certain conformations may be more frequent in some cell types but less so or absent in others, which could account for the tropism of prion strains. This proposal begs the question as to why certain conformations would be favored in some cells but not in others. One possibility is that PrPC conformation is modulated by the association of PrPC with a cell-specific component, RNA and phospholipids being two possibilities among several [43], [44]. PrP has a strong binding affinity for polyanions, in particular RNA and homoribopolynucleotides [45], [46]. Binding of RNA causes conformational changes of PrP [47], [48] and in vitro conversion of purified PrPC or recombinant PrP to infectious PrPSc by Protein Misfolding Cyclic Amplification (PMCA) requires the presence of a polyanion and/or phospholipid [44], [49]–[51]. Another possibility is that the highly variable glycosylation of PrPC observed in different cells and tissues [52] might affect its susceptibility to conversion to PrPSc, because of an effect on its conformation [53] or its interactions with putative auxiliary host proteins such as the conjectured protein X [54]. Distinct PrPC glycoforms promote PMCA-mediated prion formation in a species-specific manner [55]. There have been many attempts to explore the role of glycans in the fidelity of strain propagation, in particular by generating amino acid replacements at the glycosylation sites [53], [56]–[58]. However, in these experiments failure to propagate prions or to propagate them faithfully could be due to sequence changes, which are known to affect efficiency of propagation and cause strain shifts [59], rather than the lack of glycans. We were therefore interested in investigating the effect of post-translational modifications of PrPC on the fidelity of strain propagation in the absence of an amino acid sequence change.
Some years ago Chesebro et al. generated transgenic mice lacking the PrP GPI signal sequence on a PrP null background (tgGPI− mice) [60]. In mice hemizygous for the transgenes, PrP devoid of the GPI anchor (GPI−-PrP) was expressed at about one quarter the level of PrP in wild-type controls, was largely unglycosylated and absent from the plasma membrane. Nonetheless, these mice, inoculated with RML or 22L prions accumulated high levels of infectivity and PK-resistant PrP in the form of extracellular amyloid plaques in their brains, but exhibited no striking clinical signs up to 600 and 400 days post infection (dpi), respectively. GPI−-PrPC to GPI−-PrPSc conversion is believed to occur extracellularly [11], [60]–[62]. Mice homozygous for the transgenes expressing GPI−-PrP at a twofold higher level than their hemizygous counterparts developed neurological signs from 300 to 480 dpi [63]. Cultured cells expressing GPI−-PrP secreted the bulk of the anchorless PrP into the medium, none was found at the cell surface but some was associated with membrane fractions [64]. Cells expressing epitope-tagged GPI−-PrP exposed to 22L prions did not become chronically infected, however during the acute infection phase, tagged PrPres was transiently formed [61], [65].
To determine whether prions formed from PrPC that had the same sequence but different post-translational modifications would retain their strain characteristics, we inoculated tgGPI− mice with various prion strains or isolates and determined the cell tropism of the resulting isolates by the Cell Panel Assay (CPA) [33]. We found that the CPA properties of 22L prions remained unchanged, whereas those of RML, 139A, 79A and ME7 prions became distinctly different when propagated in tgGPI− mice. RML, 139A and 79A prions are derived from drowsy goat scrapie and are related [66], [67] while ME7 and 22L are distinct and unrelated strains derived from sheep scrapie [68]. Because the changes in properties could have been due to the absence of the GPI anchor and/or to diminished glycosylation, RML and ME7 prions propagated in GPI− mice were transferred to wild-type mice and their properties were again assessed. As judged by the CPA, ME7 prions recovered their original properties after one passage, showing that their adaptation to the GPI− environment was readily reversible. RML-derived prions, however, did not recover their original cell tropism, even after three cycles of propagation in wild-type mice, and therefore constituted a novel, stable strain, which we designated “SFL” (“Scripps Florida Laboratory”). Because both RML and SFL prions were stably propagated in wild-type mice, we propose that the two conformations are separated by a high activation energy barrier which is abrogated in tgGPI− mice, either because of the different structure or reduced expression level of the hypoglycosylated GPI-less PrPC, or because of a different replication mode in the extracellular space.
TgGPI− mice express a PrP construct in which a nonsense codon replaces that for serine233, the residue to which the GPI moiety is added in wild-type PrP; thus, the polypeptide chains are identical in the transgenic and wild-type mice [60].
TgGPI− mice [60] (GPI− for short) and C57BL/6 wild-type mice (C57 for short) were inoculated with the mouse-adapted scrapie prion strains 22L, RML, 79A, 139A and ME7 to give GPI−[22L], GPI−[RML], GPI−[79A], GPI−[139A] and GPI−[ME7] prions (Table S1). GPI−[ME7] and GPI−[RML] brain homogenate was injected into GPI− mice for a second round of transmission, to give GPI−/GPI−[ME7] and GPI−/GPI−[RML] mice. Disease signs in GPI− mice occurred later than in C57 mice; for example, RML caused terminal disease at >300 days after inoculation (dpi) in GPI−, and at about 142 dpi in C57 mice, similar to the values reported previously [63]. GPI−[RML] elicited disease in GPI− mice after 263±47 dpi, suggesting adaptation of the RML prions to the novel environment.
We first analyzed the prion-infected brain samples for their content of PrPres. Western blot analysis of C57[RML] and C57[ME7] brain homogenates showed three PrP-specific bands, attributed to di-, mono- and unglycosylated species, whose mobility increased after truncation by PK digestion (Figure 1A, lanes 1,2 and 9,10 respectively). The same amount or threefold higher of GPI−[ME7] sample gave no detectable PrP signal (lanes 11–14), while GPI−/GPI−[ME7] (lane 17) and GPI−[RML] (lane 5) samples gave rise to a ladder of bands with mobilities corresponding to about 28 to >250 kDa. Treatment with PK converted the ladders to 2 bands (lanes 6, 8, 18), the major one corresponding to unglycosylated and the minor one to monoglycosylated GPI−PrP, as shown by the fact that PNGase digestion abrogates the slower-moving band (Figure 1B, lanes 2 and 12). The aggregated forms of GPI−-PrP are likely derived from the abundant amyloid plaques described earlier [60]. Although GPI−[ME7] brain homogenate from mice culled at 300 dpi showed barely a trace, if any at all, of PrPres on western blots (Figure 1A, lanes 12, 14) it nonetheless caused clinical disease by about 170 and 450 dpi when inoculated into wild-type and GPI− mice, respectively (Table S1, #20, 21), raising the suspicion that quantitation by western blot was erroneous. Nishina and Supattapone have reported that PrPC deprived of a GPI anchor by PIPLC-treatment was retained by the PVDF membranes used for western blotting at less than about 5% the efficiency of its GPI-linked counterpart [69]. We therefore compared the levels of PrPres in PK-treated samples from GPI− and C57 mice by western blotting and by sandwich ELISA, in which PK-treated, denatured samples were bound to wells coated with anti-PrP antibody D18 and visualized with biotinylated antibody D13. Astonishingly, sandwich ELISA revealed levels of PrPres in GPI−[RML] and GPI−/GPI−[ME7] brain 50- and 25-fold higher, respectively, than those in C57[RML] brain (Figure 1C) rather than 30% and >50% lower, as indicated by western blotting (Figure 1A). Figure 1D shows that quantitation of PrPres by western blotting and sandwich ELISA are consistent for wild-type brain but highly discordant for GPI− brain. Moreover, sandwich ELISA showed that the GPI−[ME7] samples, which were negative by western blotting, in fact contained PrPres at about 12% the level in C57[ME7] brains. Brains from GPI− mice infected with 22L, 79A and 139A were not examined by western blotting or sandwich ELISA.
In summary, using the sandwich assay rather than western blotting for the quantitation of PrPres, it became evident that RML and ME7 prions propagating in GPI− mice gave rise to unusually high levels of PrPres, likely due to greater stability of extracellular as compared to intracellular PrPres.
Homogenates of prion-infected brains from GPI− and C57 mice were analyzed by the CPA. In this assay, the cell lines CAD, PK1, LD9 and R332H11 are infected with serial dilutions of a prion preparation and the proportion of infected cells is plotted against the logarithm of the dilution. The Response Index (RI) is defined as the dilution of the prion preparation at which an arbitrary proportion of cells (in this paper usually 3%) becomes PrPres positive. The ratio of RI's for a prion preparation on a set of cell lines is a characteristic strain property [2]. Additional valuable strain discrimination is provided by the glycosylation inhibitors swainsonine (swa), kifunensine (kifu) and castanospermine (csp), which inhibit chronic infection of PK1 cells by various strains to different extents [70]. Swa [70]–[72] and kifu [71] are potent and selective inhibitors of class II and class I α-mannosidases, respectively, and lead to replacement of complex N-glycans by high-mannose glycans. Csp [72], by inhibiting glucosidases, causes replacement of complex glycans mainly by glucose-containing, high-mannose oligosaccharides.
As shown by the CPA in Figure 2A, RML prions from C57 brain were swa sensitive on PK1 cells and R332H11 incompetent, i.e. unable to infect R332H11 cells efficiently, while RML-derived prions from GPI− brain were swa resistant and R332H11 competent; moreover, the RICAD/RIPK1 ratio was lower for the C57-derived than for the GPI−-derived samples. The bar diagram (Figure 2B) shows log[RICAD/RIPK1] (blue) and log[RIPK1/RIPK1+swa] (red) values plotted for 22L, RML, 79A, and 139A, propagated in C57 or GPI− brain. The quantified data clearly show that C57[RML] and GPI−[RML] give vastly different patterns, as do their 79A and 139A counterparts. The statistical significance of the “log[ratio]” differences between two strains is given in the matrix of Figure 2C. For example, the difference between C57[RML] and GPI−[RML] is highly significant for both ratios, whereas there is no significant difference between C57[22L] and GPI−[22L] by the criteria used here. Interestingly, the differences between C57[RML] and C57[139A] are significant for both ratios, but not those between C57[RML] and C57[79A], contradicting the common assumption that RML and 139A are the same strains [67] and confirming the cognate conclusions of Browning et al. [73] and Oelschlegel et al. (personal communication).
From the RI values shown in the Figures and the PrPres content relative to RML (Figure 1D) the relative specific infectivities (RI/PrPres) of RML and ME7 prions (second passage) derived from GPI− mice were calculated and found to be 6- and 25-fold lower, respectively, than those from C57 mice (Table 1), suggesting that PrPres accumulating extracellularly was either inherently less infectious or lost infectivity over time.
Because PrPC and PrPres in GPI− mice lack the GPI anchor and are also largely unglycosylated (Figure 1B) [1], the question arose whether the differences in the CPA characteristics of GPI−-derived and wild-type prions were due to these structural features (which we considered unlikely because GPI-linked PrPres arises immediately after infecting the cells for the CPA) or to a conformational change of the PrPres. We therefore inoculated brain homogenate from GPI−[RML] mice into C57BL/6 mice to generate C57/GPI−[RML] prions, whose PrPres then carried a GPI anchor and was normally glycosylated (Figure 1B, lanes 5). The resulting brain homogenate was serially transmitted twice more to C57BL/6 mice, to yield C57/C57/GPI−[RML] (Figure 1B, lanes 7, 8) and C57/C57/C57/GPI−[RML] prions (Figure 1B, lanes 9, 10; Figure 1A, lanes 3, 4).
C57BL/6 mice inoculated with prions from GPI−[RML] brain exhibited pronounced clinical signs of RML scrapie disease after 150±9 days, similar to those of mice inoculated with the original C57[RML] (Table S1, #4a,b; #1). However, as shown in Figure 3A, C57/GPI−[RML] prions neither retained the drug susceptibility pattern of GPI−[RML], nor did they regain the pattern characteristic for the original C57[RML] prions, even after two additional transfers through C57 mice. From the bar diagram (Figure 3B) it can be seen that the log[RIPK1/RIPK1+kifu] (blue) and log[RIPK1/RIPK1+swa] (red) values for C57/GPI−[RML] and C57/C57/C57/GPI−[RML] prions, 0.6, 0.7, and 0.7, 0.8, respectively, were indistinguishable, but far lower than those for C57[RML] prions, >2.8 and 1.4, respectively. The statistical evaluation shown in the matrix of Figure 3C confirms the significance of these conclusions.
In a further experiment (Figure 3D), the susceptibility of prion replication in PK1 cells to inhibition by csp and, again, kifu was tested. The bar diagram (Figure 3E, blue bars) shows that kifu, as before, inhibited infection of PK1 cells by RML by more than 3 logs but had little effect on infection by C57/GPI−[RML] or C57/C57/GPI−[RML] prions, and Figure 3F documents the statistical significance of the conclusions. In addition, we cloned RML prions in PK1 cells, propagated them in mouse brain and determined that none of twelve RML subclones exhibited kifu resistance, excluding the possibility that SFL was a significant component of RML (Supporting Information, Figure S1). Csp inhibited RML infection of PK1 cells by about 2 logs, but that of GPI−-derived prions passaged once or twice in C57 mice by only about 1 log (Figure 3E, green bars). Thus, as judged by the CPA and by inhibitor susceptibility, RML prions passaged through GPI− brain acquired novel characteristics and these were retained even after three serial passages through C57BL/6 brain.
The conformational stability assay [26] showed no difference between the PrPres associated with wild-type RML, C57/GPI−[RML] and C57/C57/GPI−[RML], however GPI−[RML] PrPres seemed slightly more stable; whether this reflects a conformational difference or a difference due to the absence of GPI anchor and paucity of N-glycans remains unknown (Figure S2).
Typical scrapie signs were not observed in ME7-inoculated GPI− mice by 300–305 days, at which time mice showing mild clinical signs were euthanized and homogenates of their brains were subjected to the CPA. The level of infectivity of GPI−[ME7] brain homogenates as measured on CAD and PK1 cells was very low (RI<103), however LD9 cells showed a higher response (RI = 1.1×104), albeit at a level about 25 times lower than that found for C57[ME7] brains 2.8×105 (Figure 4). The GPI−[ME7] brain homogenates were infectious to wild-type mice, leading to disease at 170±0 dpi (Table S1, #20), as compared to 138±4 days for “normal” C57[ME7] (Table S1, #18b), and yielded prions indistinguishable from the original ME7, as judged by the CPA. A second transmission of GPI−[ME7] prions to GPI− mice resulted in disease at 447±64 dpi (Table S1, #21) and, interestingly, to high levels of PrPres, as determined by sandwich ELISA (Figure 1C, 1D), and infectivity, as measured on CAD cells (RI = 1.5×105; Figure 4A). GPI−/GPI−[ME7] prions differed from C57[ME7] prions by their log[RILD9/RICAD] value, -0.03 versus 0.7; whether or not they revert to the original ME7 after being returned to C57 is still under investigation (Table S1, 22.) (Figure 4B, C). In summary, repeated propagation of ME7 prions through GPI− brain resulted in a distinct alteration of their properties, reflecting progressive adaptation to the modified environment. However, prions returned from GPI−[ME7] mice to C57 mice resulted in reversion to prions indistinguishable from the original C57[ME7] prions.
Prion populations are considered to be “quasi-species”, i.e. to consist of a major component and a multiplicity of variants present at low levels, of which one may be selected as the major component if the population is exposed to a different environment [39], [74], [75].
We have reported previously that when 22L prions were transferred from brain to PK1 or R33 cells, their properties, as measured by the CPA, changed gradually in the course of many doublings, suggesting that a 22L variant present at low levels in the brain-derived population was being selected in the cellular environment. Conversely, when the cell-adapted 22L variants were again propagated in brain, the resulting population gradually re-acquired the properties of the original brain-derived 22L. When 22L prions were cloned by endpoint dilution into PK1 cells they were initially swa sensitive and incapable of developing swa resistance (“swa incompetent”), but as the populations were further propagated they became swa competent while remaining swa sensitive, suggesting that during propagation swa-resistant prions arose at a low level by “mutation” and could be selected when the population was challenged with swa [74]–[76].
In all these cases the changes were reversed when the prions were propagated in the original environment; assuming that properties of prions are encoded by the precise conformation of the cognate PrPSc, we concluded that the conformational states underlying adaptation to the cellular environment are separated by low activation energy barriers which allowed reversible conformational switches leading to “strain variants” or “sub-strains” [75]. In contrast to these earlier results with 22L prions, we now report that when RML prions from wild-type brain were propagated in brain producing anchorless PrP, a GPI−-PrPSc conformer adapted to that environment evolved, and when this was transferred to wild-type mice, a novel conformation of wild-type PrPSc, which we designated SFL, distinct from that of the original RML, was stably maintained. Because SFL does not revert to RML, it must be either better adapted to propagation in wild-type brain, prevented from reverting by a high activation energy barrier or both (Figure 5). Of note, methods commonly used to distinguish strains, such as incubation period, western blotting or conformational stability assays could not differentiate between RML and SFL, whereas this was readily achieved with the Extended Cell Panel Assay (ECPA) [77]. ME7 also acquired distinct properties when propagated in tgGPI− brain, however it reverted to what appears to be its original form when passaged through wild-type brain. Strain switching has been observed previously in transfer between animal species, when 139A prions were transferred from mouse to hamster and back to mouse [78]; in that case, a different amino acid sequence in the intermediate host may have led to the adoption of a more favorable conformation, which was preserved when the prions were returned to the original host. Mutated PrPSc could be formed if accretion of GPI−-PrPC to wild-type seed entailed adoption of a conformation slightly different from that of the seed, or if wild-type PrPSc contained a variety of mutant seeds –possibly at a low level– to one or some of which the resident GPI−PrPC preferentially accreted, adopting its conformation [76].
Considering the large variety of “classical” murine prion strains [79], the plethora of strains generated by Prusiner and his colleagues [80], [81] and the many variants we have observed, all encoded by a single murine PrP sequence, the number of stable conformational states must be vast. The structural elucidation of PrPSc and its variants continues to be the Holy Grail of the prion field.
When working with mice all efforts were made to minimize suffering. 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 Institute of Health. The protocol was approved by the Institutional Animal Care and Use Committee (IACUC). Scripps Florida has an Animal Welfare Assurance on file with the Office of Laboratory Animal Welfare (OLAW), National Institute of Health (assurance number #A4460-01). Scripps Florida's registration under USDA regulations is certificate 93-R0015. The Association for Assessment and Accreditation of Laboratory Animal Care International (AAALAC) awarded Scripps Florida full accreditation.
C57BL/6 mice were from Charles River Laboratories (Wilmington, MA). Breeding pairs of tgGPI− mice [60] were obtained from the Oldstone laboratory. Mice were anesthetized by isoflurane and 30-µl samples of 1% brain homogenates were inoculated in the prefrontal cortex. When pronounced clinical signs became evident, or earlier, the mice were asphyxiated with CO2 and subjected to cervical dislocation.
The 22L strain, cloned by two successive end-point dilutions, as well as strains 79A and 139A were obtained from the TSE Resource Centre, Compton, Newbury, UK (I. McConnell, R.M. Barron). The RML strain, from the MRC Prion Unit, University College, London, was propagated initially in CD1 mice and subsequently in C57BL/6 mice.
Brains were harvested from mice exhibiting neurological signs of prion diseases in C57BL/6. In some cases where inoculated, anchorless mice showed only mild clinical signs, brains were taken at 300 days post inoculation. For stock homogenates, frozen brains were pooled and homogenized for 10 sec in PBS (9 ml per g) using a hand-held Ultramax T18 basic homogenizer (IKA Works Inc., Bloomington, NC) at 20000–25000 rpm. Single frozen brains were homogenized in PBS using a ribolyser (FastPrep FP120, Bio 101, Thermo Electron Corp.,Thermo Fischer Scientific, MA) with ZrO 0.8–1 mm beads (cat. No. 7305-000010, Glen Mills Inc. Clifton, NJ), at maximum speed 6.5 for 15 seconds in Fast Prep tubes (MP Biomedicals).
The cell lines PK1 and R332H11 are derived from N2a neuroblastoma cells, CAD5 (CAD) from Cath.a-differentiated cells, and LD9 from L929 fibroblasts [33], [82]. The lines were maintained in OBGS (Opti-MEM [Invitrogen], 4.5% Bovine Growth Serum [Hyclone, Logan, UT], 90 units penicillin/ml, 90 µg streptomycin/ml [Invitrogen]). Cells were split 1∶10, or 1∶8 if sparse, twice a week. After 9 serial passages cells were replaced with freshly thawed samples.
The SSCA was performed as detailed earlier [82]. In summary, serial 1∶3 dilutions of the sample (300 µl) were placed in quadruplicate wells in 96-well plates with 5000 susceptible cells per well. Four days after infection the confluent monolayer was suspended and split 1∶7 in OBGS for total 3 splits, allowing cells 3–4 days to form a confluent layer between each split. After the third split, cells were grown to confluence again, 20000 cells were added to the wells of pre-activated Multiscreen IP96-well 0.45-µm Immobilon P membrane plates (Millipore, Danvers, MA) and vacuumed onto the membrane. After baking at 50°C for 1 h, samples were treated with 1 µg proteinase K (PK, Roche)/ml in lysis buffer, followed by PMSF and denaturation with guanidinium thiocyanate. In more recent assays the PMSF step was omitted. After washing with PBS, wells were blocked with 0.5% milk in 1× TBS and incubated with humanized anti-PrP antibody D18 [83] followed by mouse anti-human IgGγ AP-conjugated antibody in 0.5% milk-TBST. Signals were visualized with AP Conjugate Substrate Kit (BioRad) and PrPres-positive cells (“spots”) were counted using the Bioreader 5000-Eb (BioSys). To control for residual inoculum, the prion replication inhibitor pentosan polysulfate (PPS) was added during the 4-day infection at 10 µg/ml to the cells infected with the highest concentration of prion sample.
The number of PrPres-positive cells is plotted as a function of the logarithm of the dilution. The “Response Index” (RI) of a sample is the reciprocal of the concentration that gives rise to a designated proportion of PrPSc-positive cells under standard assay conditions (in our experiments, RI600, 600 positive cells per 20000 cells, or 3%).
The Cell Panel Assay (CPA) allows the characterization of strains by virtue of their cell tropism and their susceptibility to inhibitors, as determined by the SSCA performed on CAD5, PK1, LD9 and R332H11 cells. A strain or a substrain is characterized by the ratio of RIs on different cell lines and by the susceptibility to inhibition by drugs. When required, glycosylation inhibitors swainsonine (swa, Logan Natural Products; 2 µg/ml), kifunensine (kifu, Toronto Research Chemicals Inc.; 5 µg/ml) and castanospermine (csp, Toronto Research Chemicals Inc.; 50 µg/ml) were present during infection and up to the second split, and were diluted thereafter with each split.
PK1 cells were seeded at 100 cells/well in 96-well plates and inoculated with highly diluted RML-infected brain homogenates (final concentration: 10−9; 5×10−10). The cells were repeatedly grown to confluence and subjected to three 1∶3 splits, followed by eight 1∶10 splits; after a total of about 50 doublings 20000 cells from each well were subjected to the PK-Elispot Assay and samples containing PrPres-positive cells (spot numbers>[background+5 SDs]) were scored as positive. The proportion of positive wells was 27/168 (16%) for an inoculum dilution of 10−9 and 27/252 (11%) for an inoculum dilution of 5×10−10; the probability PN>1 = 1 - e−m (1+m) that under the conditions chosen a well was infected by more than one prion was 10−2–10−3 or less, where m is the average number of prions/well. Nine prion clones were expanded, conditioned media was harvested, concentrated and inoculated into C57BL/6 mice. The brain homogenates from terminally ill mice were characterized by the SSCA on PK1 cells in the presence or absence of 5 µg kifu/ml for all 9 clones from the 1st round of cloning. In addition, prions from several RML-infected PK1 (PK1[RML]) clones were subjected to another round of end-point dilution cloning in cells. Eight clones of 2nd- round RML-infected PK1 (PK1{PK1[RML]}) cells were expanded, conditioned media was harvested, concentrated and inoculated into C57BL/6 mice. The brain homogenates from three of the eight terminally-ill mice were characterized by the SSCA on PK1 cells in the presence or absence of 5 µg kifu/ml.
Samples were denatured by boiling in XT-MES sample buffer (BioRad), fractionated by SDS-PAGE on 4–12% Criterion gel (BioRad) for 1.5 h at 120 V and transferred to PVDF Immobilon membranes (Millipore) by wet transfer (Criterion, BioRad). Membranes were blocked in 5% non-fat dry milk/PBST and exposed to 0.5 µg D18 anti-PrP antibody/ml in 5% non-fat dry milk/PBST followed by mouse anti-human IgG HRP-conjugated secondary antibody (48 ng/ml, Southern Biotech) in 5% non-fat dry milk/PBST. Chemiluminescence was induced by ECL-Plus (Pierce) and recorded by CCD imaging (BioSpectrum AC Imaging System; UVP). PageRuler Plus Prestained Protein Ladder (Fermentas) was run as molecular weight marker.
The method is essentially that of Peretz et al. [23]. Prion-infected brain homogenates (30 µg total protein) were adjusted to between 0.5 M and 5 M guanidinium chloride (GndCl) in 10 mM Tris-HCl (pH 8.0) (final volume 60 µl) and incubated for 15 min at 25°C, shaking at 700 rpm in an Eppendorf thermomixer. The GndCl concentration in each sample was then adjusted to 0.2 M with 10 mM Tris-HCl (pH 8.0) and the volumes were equalized to 985 µl with 0.2 M GndCl in 10 mM Tris-HCl (pH 8.0). Samples were adjusted to 0.5% Triton X-100 and digested with 0.6 µg/ml PK (PK : protein = 1∶50 by weight) with shaking at 1000 rpm for 1 h at 37°C in an Eppendorf thermomixer; digestion was terminated by addition of PMSF to 2 mM. After addition of 5 µg BSA, proteins were precipitated with TCA (10% final concentration), chilled on ice for 30 min, and centrifuged 15 min at 16000× g and 4°C. Pellets were resuspended in 0.5 ml cold acetone, re-centrifuged and resuspended in 50 µl PBS-0.5% Triton X-100. Samples were heated 10 min at 100°C in MES sample loading buffer (Bio-RAD) and analyzed by western blotting. Chemiluminescence was induced by ECL-Plus (Pierce), recorded and quantified by CCD imaging (BioSpectrum AC Imaging System; UVP). The highest value of each curve was set to 100% and the Gnd1/2 value, i.e. the GndCl concentration at which 50% of the PrPres was digested by PK under standard conditions, was determined.
The procedure is essentially as described earlier [17]. Ninety-six-well plates (F16 Maxisorp Nunc Immune Module, Nunc) were rocked overnight with 15 µg/ml D18 antibody/well at 4°C, washed with PBST 3 times, blocked with 5% milk in PBST at 37°C for 1 h, washed again in PBST 3 times and stored with 200 µl PBS per well at 4°C. Brain homogenates were diluted to 3 mg total protein/ml in 0.5% Triton X-100. Samples with high PrPres content were diluted appropriately with PrPo/o brain homogenate to give a final protein concentration of 3 mg/ml. Samples were shaken with 12.5 µg proteinase K (PK)/ml at 37°C for 1 h, and adjusted to 6 mM PMSF, giving a protein concentration of 2.8 mg/ml. A 7.2-µl aliquot of each sample was denatured with an equal volume 8 M GndCl (Research Products International Corp.) for 5 min at 80°C and diluted in Sandwich Assay Buffer (50 mM Tris-HCl, pH 8.0, 2% Triton X-100, 2% sodium lauroylsarcosine, 2% BSA) to a final volume of 1 ml. Quadruplicate 300-µl sample aliquots were added to the wells of blank 96-well tissue culture plate (BD Falcon) and serially diluted 1∶2 in Sandwich Assay Buffer. When used as a standard, recombinant PrP (200 ng/ml in 2.8 mg/ml PrPo/o brain homogenate) was serially diluted into 2.8 mg/ml PrPo/o brain homogenate. A 150-µl aliquot of each sample was transferred to the D18-coated wells and rocked 1 h at 37°C. Plates were washed 5 times with PBST (0.1% Tween/PBS) and to each well was added 100 µl biotinylated D13 antibody [83] at 1.3 µg/ml 1% milk-PBST. After rocking for 1 h at 37°C in 1% milk-PBST, plates were washed 5 times with PBST and to each well was added 100 µl 1∶7500 HRP-streptavidin (Amersham GE Healthcare) in 0.5% BSA-PBST for 30 min at 37°C. After 5 washes with PBST, 100 µl TMB Super-sensitive HRP microwell substrate (SUB2) was added, and after 2 min at room temperature the reaction was stopped with 100 µl TMB Microwell Stop solution (STOP1; Immunochemistry Technologies, MN). Plates were read with a BioTek 450 plate reader (BioTek Instruments, VT) and analyzed with Gen5 software.
|
10.1371/journal.pntd.0007126 | Triple oral beta-lactam containing therapy for Buruli ulcer treatment shortening | The potential use of clinically approved beta-lactams for Buruli ulcer (BU) treatment was investigated with representative classes analyzed in vitro for activity against Mycobacterium ulcerans. Beta-lactams tested were effective alone and displayed a strong synergistic profile in combination with antibiotics currently used to treat BU, i.e. rifampicin and clarithromycin; this activity was further potentiated in the presence of the beta-lactamase inhibitor clavulanate. In addition, quadruple combinations of rifampicin, clarithromycin, clavulanate and beta-lactams resulted in multiplicative reductions in their minimal inhibitory concentration (MIC) values. The MIC of amoxicillin against a panel of clinical isolates decreased more than 200-fold within this quadruple combination. Amoxicillin/clavulanate formulations are readily available with clinical pedigree, low toxicity, and orally and pediatric available; thus, supporting its potential inclusion as a new anti-BU drug in current combination therapies.
| Buruli ulcer (BU) is a chronic debilitating disease of the skin and soft tissue, mainly affecting children and young adults in tropical regions. Before 2004, the only treatment option was surgery; a major breakthrough was the discovery that BU could be cured in most cases with a standard treatment that involved 8 weeks of combination therapy with rifampicin and streptomycin. However, the use of streptomycin is often associated with severe side effects such as ototoxicity, or nephrotoxicity. More recently, a clinical trial demonstrated equipotency of replacing the injectable streptomycin by the clarithromycin, which is orally available and associated with fewer side effects. BU treatment is now moving toward a full orally available treatment of clarithromycin-rifampicin. Although effective and mostly well tolerated, this new treatment is still associated with side effects and only moxifloxacin is additionally recommended by WHO for BU therapy. New drugs are thus needed to increase the number of available treatments, reduce side effects, and improve efficacy with treatments shorter than 8 weeks. In this work, we describe for the first time the potential inclusion of beta-lactams in BU therapy. More specifically, we propose the use of amoxicillin/clavulanate since it is oral, suitable for the treatment of children, and readily available with a long track record of clinical pedigree. Its inclusion in a triple oral therapy complementing current combinatorial rifampicin-clarithromycin treatment has the potential to counteract resistance development and to reduce length of treatment and time to cure.
| Buruli ulcer (BU) is a chronic debilitating mycobacterial disease of the skin and soft tissue. Although mortality is low, permanent disfigurement and disability are high, mainly affecting children and young adults. BU is found primarily in tropical regions of Africa, South America and the Western Pacific; however, it is also becoming a public health concern in some regions of Australia [1].
Before 2004, when the World Health Organization (WHO) published provisional guidance for the management of BU disease [2], antibiotics were viewed as relatively ineffective and surgery remained the mainstay of treatment for BU [3–5]. In the late 1990s and early 2000s, however, in vitro studies demonstrated anti-BU activity of some antibiotics used for the treatment of tuberculosis (TB) and other non-tuberculous mycobacteria [6–8], and in vivo studies the potential for combining two drugs to provide improved treatment outcomes [9–13]. Soon after, clinical evidence showed the effectiveness of a combination of rifampicin plus streptomycin when it was administered for at least 4 weeks [14], and that routine implementation of such a therapy was possible in the field [15, 16]; however, the use of the injectable streptomycin is often associated with adverse events and it is restricted in the treatment of pregnant women and young infants. In addition, the lack of an efficacious oral treatment remained one of the main obstacles to decentralizing care at local level in rural areas. These limitations motivated the scientific community to evaluate alternative oral treatments and clinical studies demonstrated that fluoroquinolones [17] or clarithromycin [18–20] could also be used in combination with rifampicin and were associated with fewer side effects compared to the injectable streptomycin. Thus, on March 24th, 2017, WHO recommended full oral treatment of 8 weeks daily combination therapy of rifampicin-clarithromycin [21].
While recommended regimens (rifampicin plus streptomycin or clarithromycin) allow cure of small lesions (<5 cm in diameter) without surgery [15, 18], controversy remains regarding the best surgery approach for large lesions (>10 cm) [22, 23]. Intermittent drug administration using rifapentine, a rifampicin analog with longer half-life, instead of daily rifampicin, has been also proposed as a strategy to facilitate treatment supervision in the field [24]. However, M. ulcerans strains resistant to rifampicin have been isolated after experimental chemotherapy in mice [25] and a recent report described the emergence of M. ulcerans strains resistant to rifampicin and streptomycin in the clinic [26]; further experiments would be, however, needed to identify the genetic basis of such resistance patterns and confirm the emergence of resistance in M. ulcerans clinical isolates. Nevertheless, these reports should be a warning sign since no alternatives for rifampicin are currently available. WHO currently recommends only four drugs for the treatment of BU: rifampicin, streptomycin, clarithromycin and moxifloxacin [2]. It would be thus desirable to increase the number of drugs available to treat BU and to develop a new therapy that would reduce both duration of treatment and time to healing after therapy completion for all type of lesions and suitable for children and pregnant women.
Drug discovery and development for neglected diseases is especially delayed due to lack of interest from the main scientific and industrial communities. To speed up the process in the BU field, we applied knowledge gathered in TB R&D drug repurposing programs [27–30] where we (and others [31]) showed that beta-lactams strongly increased the bactericidal and sterilizing properties of rifampicin [28]. Rifampin is the cornerstone drug for TB (and BU) therapy with a direct relation between dose increase and therapy efficacy [32] due to its bactericidal and sterilizing activity in a dose-dependent manner [33]. However, the current WHO recommended 10 mg/kg (600 mg daily) is not its optimal clinical dosage [34] and some recent studies suggest that it could be safely increased to 35 mg/kg daily for TB therapy with a bacteriological effect on time to culture conversion [32, 35, 36]. More recently, dose-ranging high-dose rifampicin studies using a murine model of M. ulcerans disease showed that shorter BU treatments might be also feasible [37], suggesting that synergistic partners could serve to improve rifampicin efficacy without compromising tolerability and toxicity.
Beta-lactams are one of the largest groups of antibiotics available today with an exceptional record of clinical safety in humans [38]. Used for decades, they had been traditionally considered ineffective for the treatment of mycobacterial infections (mainly TB) due to the presence of a beta-lactamase (BlaC) and the hydrophobic nature of the mycobacterial cell envelope [39]. However, after a seminal publication describing the in vitro activity of meropenem plus clavulanate against multi-drug (MDR) and extensively drug resistant (XDR) strains of M. tuberculosis [40] and its anecdotal use in salvage therapies for XDR patients [41], the first study convincingly demonstrating the clinical efficacy of beta-lactams was recently published [29]. These studies provided evidence of their anti-mycobacterial clinical potential, opening a new avenue to identify new drugs and optimize current BU therapy.
In this study, we are translating knowledge and concepts of drug repurposing and synergy generated in TB R&D programs to assess the potential inclusion of beta-lactams for BU therapy. We propose the combination of amoxicillin/clavulanate as a new anti-BU treatment in combination with current oral BU therapy, rifampicin and clarithromycin, with the potential of treatment shortening and readily implementation in the field.
M. ulcerans strain NCTC 10417 (ATCC Number: 19423; Lot Number: 63210551) was used for initial screening assays. Further validation studies were performed with clinical isolates from different geographical origins: ITM 063846, Benin; ITM 070290, China; ITM 083720 and ITM C05143, Mexico; ITM 941327, ITM C05142 and ITM M000932, Australia; ITM C05150, DR Congo; ITM C08756, Japan, purchased from the Belgian Co-ordinated Collection of Micro-organisms (BCCM).
M. ulcerans cells were initially grown at 30°C to an optical density at 600 nm (OD600) of 0.5–1.0 in tissue culture flasks containing 7H9 broth supplemented with 0.2% glycerol, 10% OADC and 0.05% (vol/vol) Tyloxapol. Aliquots of 500 μL were then stored at -80°C and the number of colony forming units (CFU) in the freeze stock enumerated. Every experiment was performed starting from a new frozen stock to avoid excessive passage of the original strain. Cells were also routinely passaged on Middlebrook 7H10 agar plates (Difco) supplemented with 10% (vol/vol) OADC to ensure purity of the isolate.
Rifampicin (R3501-1G; Lot Number: SLBH7862V) and meropenem (M2574; Lot Number: 055M4705V) were purchased from Sigma. GlaxoSmithKline provided clarithromycin, streptomycin, clavulanate and all other beta-lactams used in this study.
Minimal Inhibitory Concentrations (MIC) were determined in 7H9 broth supplemented with 0.2% glycerol, 10% OADC and without Tyloxapol using triplicate two-fold serial dilutions of compounds in polystyrene 384- or 96-well plates. MTT [3- (4,5-dimethylthiazol-2-yl)-2,5-diphenyl tetrazolium bromide] was used as the bacterial growth indicator [30, 42]. For cell density calculations, a culture having an OD600 of 0.125 was found to contain approximately 107 cfu/mL. Cultures were sampled (50 μL in 384-well plates or 200 μL in 96-well plates) at a final cell density of 106 cfu/mL and incubated at 30°C in the presence of the drug (or drug combinations) for 6 days before addition of 12.5 μL (384-well plates) or 30 μL (96-well plates) of a MTT / Tween 80 (5 mg/mL / 20%) solution mix. After further overnight incubation at 37°C, OD580 was measured. The lowest concentration of drug that inhibited 90% of the MTT color conversion (IC90) was used to define MIC values.
Checkerboard assays and calculations of the Fractional Inhibitory Concentration Index (FICI) were used to define the degree of pairwise drug interactions, as previously described [28] (for a visual representation and deeper understanding of the checkerboard assay refer to the supplementary information of Ramón-García et al.[30]). Up to quadruple combinations of rifampicin, clarithromycin, beta-lactams and clavulanate were also tested. For this, checkerboard plates were prepared with rifampicin in the y-axes, the beta-lactam in the x-axes, and clarithromycin added in the z-axes as fixed sub-MIC (1/2, 1/4, and 1/8xMIC values) concentrations for every checkerboard plate; typically the 1/8xMIC plate was used for synergy calculations of the quad combos. When assayed, clavulanate was added at a fixed sub-MIC concentration of 5 μg/mL. Increase efficacy of compounds (synergistic MIC, MICsyn) when in combination (fold-MIC reduction) was always reported versus the activity of drugs alone. The Most Optimal Combinatorial Concentration (MOCC) was defined as the lowest possible concentration of every compound that, when assayed together, prevented bacterial growth, i.e. in an isobologram representation this would be the closest point to the axes intersection.
Clinically approved beta-lactams representing different sub-families, i.e. meropenem (carbapenems), cephradine and cefdinir (cephems), faropenem (penems) and amoxicillin (penicillins) were assayed in vitro in a checkerboard format to assess their synergistic interactions with rifampicin in the absence and presence of clavulanate, a beta-lactamase inhibitor, against the M. ulcerans ATCC strain. A pattern of strong synergistic interaction was observed between rifampicin and all beta-lactams tested (Fig 1); however, no interaction was observed when the same assay was conducted using combinations of rifampicin and the currently WHO recommended anti-BU drugs, i.e. streptomycin, clarithromycin and moxifloxacin (Fig 1 and S1 Fig). Dose-response curves indicated that the activity of rifampicin (reflected in MIC reduction) was increased on average 16-32-fold (up to 128-fold in some cases) and, vice versa, the activity of the beta-lactams was strongly enhanced by rifampicin. In the case of amoxicillin, its activity was further increased 512-fold in combination with clavulanate (S2 Fig and S1 Table).
Beta-lactams not only had synergistic interactions with rifampicin but also with clarithromycin, the second drug recommended as first-line anti-BU therapy. These results prompted us to test the inhibitory effect of double clarithromycin-beta lactam, and triple rifampicin-clarithromycin-beta-lactam combinations (Fig 2). Our results indicated that, when in double or triple combinations, much lower sub-inhibitory concentrations were equally potent at inhibiting M. ulcerans growth than the additive effects of the compounds alone. MIC values were also lower than in other pairwise combinations. For example, the MIC of amoxicillin was greater than 32 μg/mL; however, its synergistic MIC (MICsyn) was reduced to 1 μg/mL in the presence of rifampicin, to 0.25 μg/mL when clavulanate was also added, or to 0.062 μg/mL when both clavulanate and clarithromycin were included together with rifampicin, i.e., an MIC reduction of ca. 500-fold for amoxicillin when in the quadruple combination. Similar results were obtained for combinations of meropenem or faropenem and rifampicin, with MIC reductions as high as 80-fold when tested within triple combinations. In these assays, clarithromycin was added at a fixed concentration of 1/8 its MIC value alone, being its presence critical to achieve the multiplicative effect observe in the quadruple combinations.
Our initial discoveries described above were performed using M. ulcerans ATCC 19423, a reference strain that does not produce mycolactone, a polyketide-derived macrolide secreted by M. ulcerans and responsible for the tissue damage pathology in Buruli ulcer [43]. In order to validate the pattern of synergistic interactions between rifampicin, clarithromycin and beta-lactams and to gauge the potential of introducing a beta-lactam in the treatment of BU, we expanded our analysis to a collection of M. ulcerans clinical isolates from different geographical locations and focused our synergy interaction studies on the amoxicillin /clavulanate combination. When in the quadruple combinations, the activities of all three drugs (clavulanate was added at a fixed 5 μg/mL concentration, more than 20-fold less its MIC) were strongly enhanced; depending on the strain tested, these interactions could range from ca. 5 to 600-fold (rifampicin), ca. 4 to 2,000-fold (amoxicillin) and ca. 20 to 80-fold (clarithromycin) (Table 1). Every possible pair-wise and triple combination was also evaluated showing strong synergism between amoxicillin and both rifampicin and clarithromycin, but not between rifampicin and clarithromycin, similar to previously described for the ATCC strain. Clavulanate enhanced the activity of amoxicillin (as expected) but had no effect over clarithromycin and only minor enhancements (in some cases) over rifampicin (S3 Table).
Amoxicillin is inactivated by beta-lactamase enzymes, limiting its clinical use. In fact, dose-response studies demonstrated that amoxicillin alone was not active (MIC > 16 μg/mL); however, a strong shift in the dose-response curve was observed by adding clavulanate and this shift was further enhanced when clarithromycin and rifampicin were also present in the combination at sub-MIC concentrations (Fig 3), thus confirming our MIC/synergy data and the potential of amoxicillin/clavulanate as a new anti-BU therapy, both alone and in combo with rifampicin, clarithromycin or both.
We have in vitro characterized the antimicrobial interactions of rifampicin with anti-BU drugs and beta-lactams and found that, while there was no synergy between rifampicin and current anti-BU drugs, it had strong synergistic interactions with beta-lactams. What is more, beta-lactams also displayed synergism with clarithromycin, the other first line drug in BU therapy. Our studies confirmed previous data showing lack of in vitro interaction between rifampicin and clarithromycin [44], thus reinforcing synergy data observed with beta-lactams. Similar observations of no interaction between rifampicin and clarithromycin were also recently described in murine models of M. ulcerans infection [37]
Working with M. ulcerans is challenging due to its slow generation time (ca. 48 hours), even slower than M. tuberculosis. Current methodologies to perform susceptibility testing against clinical isolates are often time consuming and cumbersome (use of agar proportion methods), thus, requiring several months to generate results [6, 8, 26]. Improvements have been introduced using luminescent reporter strains [45]; however, this technology is limited to specific engineered strains and cannot be widespread applied to clinical isolates. Here, we were able to perform synergy studies, obtaining results in just seven days, in a medium-throughput manner against a panel of clinical isolates by adapting a methodology previously described for TB research and already validated to determine MIC values in slow growing mycobacteria [28]. Similar redox-based assays (using alamar blue) have been previously employed in antimicrobial discovery programs targeting M. ulcerans [46, 47]. This methodology could provide a cost and time effective assay to implement in the clinical practice in Buruli ulcer drug resistance surveillance campaigns.
Over the last decade, BU therapy has undergone a tremendous improvement with the introduction of chemotherapy; however, there is only a limited number of drugs recommended by WHO for BU therapy, namely rifampicin, streptomycin, clarithromycin and moxifloxacin [2]. Current WHO recommended therapy is a fully oral eight weeks daily regimen with a combination of rifampicin and clarithromycin [20]. Although effective and mostly well tolerated, combination treatment of rifampicin plus streptomycin or clarithromycin is associated with undesirable side effects that might include mild (anorexia, nausea, abdominal pains and altered taste) or severe (deafness, skin rashes, jaundice, shock, purpura or acute renal failure) symptoms [2]. In a scenario where administration of any of these drugs needs to be interrupted, therapeutic options would remain limited to the use of moxifloxacin, an antibiotic contraindicated during pregnancy and within the pediatric population. A similar situation would occur in the eventual development of resistance to any of these drugs, especially rifampicin. Since they are administered in pairwise combinations, this would effectively imply monotherapy, further promoting the emergence of resistance. This is similar to another mycobacterial disease, leprosy, for which this threat was largely ignored and just recently WHO issued guidelines including procedures for the detection of drug resistance [48]. Thus, an alternative drug regimen would be required to treat resistant M. ulcerans strains [26], even though the emergence of resistance in M. ulcerans might follow different dynamics compared to M. tuberculosis and M. leprae (environment vs. host reservoirs, respectively) and clinical resistance has not been conclusively demonstrated to date.
BU is a neglected disease mainly affecting rural areas in under-resourced countries where medicine access and logistics might prove difficult, and hospitalization and loss of income for patients and families might compromise patient’s adherence to the 8-weeks antibiotic course. A shortened, highly effective, all-oral regimen is urgently needed to improve care for this neglected tropical disease; this would reduce indirect costs and barriers to therapy. The history of TB chemotherapy teaches us that combination therapy is critical for optimal cure outcome and treatment shortening [49]. Translation of this knowledge into BU therapy suggests that more drugs need to be added to the current rifampicin-clarithromycin combination in order to improve and shorten the duration of treatment.
Most beta-lactams tested in this study were active against M. ulcerans and enhanced the anti-BU activity of rifampicin to different degrees. Cephradine is a first-generation cephalosporins developed in the 1960s, also recently described to be active in vitro against M. tuberculosis [28]; however, cephradine was long ago discontinued and access to other first-generation cephalosporins, such as cefadroxil or cephalexin, is limited in many countries. Cefdinir is a third-generation cephalosporin active against pneumonia, skin and soft tissue infections, although with low oral absorption [50]. It is currently used in the clinic, widely distributed and access to it could be readily available; however, its synergistic profile with rifampicin was weaker compared to other beta-lactams (S1 Table). Although meropenem was active against pulmonary TB in a recent clinical study [29], it needs to be administered intravenously, not a practical approach in under-resourced countries where oral drugs are required. Faropenem, an orally administered beta-lactam, did not show activity in the same clinical trial due to the low drug exposure in plasma after oral administration [29]. Finally, amoxicillin/clavulanate showed good activity and very strong synergistic interaction with rifampicin.
The combination of amoxicillin plus clavulanate is a broad-spectrum antibacterial available for clinical use in a wide range of indications and is now used primarily in the treatment of community-acquired respiratory tract infections [51]. It was first launched in the UK in 1981; by the end of 2002, it was clinically available in various formulations in over 150 countries around the world. In addition to high efficacy, it has a well-known safety and tolerance profile, including for pregnancy and pediatric used, based on over 819 million patient courses worldwide, with the main contraindication being allergy to penicillin derivatives. Disruption of the gut microbiota is the main side effect of long-term use of amoxicillin/clavulanate, mainly caused by the presence of clavulanic acid in the formulation [52]. For TB treatment, amoxicillin/clavulanate is included in Group 5 (anti-TB drugs with limited data on efficacy and long-term safety in the treatment of drug-resistant TB) of the WHO 2011 TB drugs classification and in Group D3 (add-on agents, not core MDR-TB regimen components) of the WHO 2016 MDR-TB drugs classification [53]. In 1983, Cynamon et al. reported the in vitro bactericidal activity of amoxicillin/clavulanate against 15 isolates of M. tuberculosis, at concentrations of amoxicillin lower than 4 μg/mL [54], and some years later, Nadler et al. case reported the effective treatment of MDR-TB patients with the addition of amoxicillin/clavulanate to the second-line therapy [55]. Two contradictory follow up clinical studies, 2-days Early Bactericidal Activity (EBA), suggested that its activity was comparable to that reported for anti-TB agents, other than isoniazid [56], but also questioned its role in the treatment of tuberculosis [57]. The dosing interval of the amoxicillin/clavulanate therapy might explain these differences; while in the first EBA it was divided into three daily doses, it was given as a single high dose in the second one. More recently, a 14-days EBA study demonstrated activity of a combination of meropenem plus amoxicillin/clavulanate [29]; it remains to be determined whether this activity was due to any of the components alone or the combination therapy as a whole [58]. In fact, in vitro studies have demonstrated synergistic interactions among amoxicillin and meropenem (and other beta-lactams), rifampicin and ethambutol against M. tuberculosis [28, 58–60].
In our in vitro assays with M. ulcerans clinical isolates, we found that amoxicillin had no activity (typically MIC values > 16 μg/mL) but that its MIC could be reduced to 1 μg/mL in the presence of clavulanate (S3 Table), similar to previously reported to the closely related M. marinum species and other non-tuberculosis mycobacteria [61]. It also displayed strong synergistic interactions with rifampicin and clarithromycin and, in quadruple combinations, its activity was enhanced up to 2,000-fold in some cases, with average MIC ranges between 0.031 to 0.25 μg/mL (Table 1). For infections caused by other bacterial pathogens, susceptibility breakpoints of amoxicillin/clavulanate are established at ≤ 2 μg/mL (or ≤ 4 μg/mL for high-dose formulations) and mean peak plasma concentrations of amoxicillin range from 7.2 to 17 μg/mL, depending on the formulation [52], well above the synergistic MIC values reported in this work. The bacteriological efficacy of penicillins is dependent on the time its free plasma concentration remains above the MIC (time over the MIC value, fT>MIC). For other bacterial infections, it has been estimated that a fT>MIC of ca. 30–40% of the dosing interval is required for bactericidal activity [62]. In the case of M. ulcerans, this target therapy could be achieved using standard amoxicillin/clavulanate formulations of 500/125 mg (4:1) or 875/125 mg (7:1) administered three times a day, or the high-dose extended release formulation of 2000/125 (16:1) that would allow administration twice a day [52], an important consideration for treatment compliance in under-resourced settings. Thus, according to our in vitro data, amoxicillin/clavulanate could have an important role in the treatment of BU alone and, more importantly, in combination with current first-line anti-BU therapy since no pharmacological drug-drug interactions are described among amoxicillin/clavulanate and rifampicin or clarithromycin [51].
But, what could be the benefit of adding amoxicillin/clavulanate to the current anti-BU therapy? Besides being able to treat secondary infections associated with BU lesions, it has been proposed that the median time to healing is related to the bacterial load in the lesions at the beginning of therapy and the presence of persister bacteria [63]; in fact, healing of up to two thirds of patients occurs within 25 weeks from the start of treatment but for some patients this can take up to a year. One of the reasons for this slow healing could be due to a high initial bacterial load. In fact, active infection late into the recommended 8-week course of antibiotic therapy could be found in slowly healing lesions [22, 64]. Extensive histopathologic studies also demonstrated that M. ulcerans is essentially confined in extracellular areas of necrosis in skin [6]. Under this circumstances, amoxicillin/clavulanate would be extremely effective at targeting extracellular bacteria with rapid bactericidal activity, thus reducing initial bacterial burden, local levels of the immune-suppressive mycolactone toxin, and allowing local recovery of the host immune response to clear remaining bacteria. What is more, in vitro studies have demonstrated the sterilizing activity of synergistic combinations of beta-lactams and rifampicin [28], which could target those remaining persistent populations, thus shortening treatment and healing times. Rapid bacterial killing would also imply a reduction in the risk of development of resistance; even in the scenario of infections caused by bacteria resistant to rifampicin, this could still be re-introduced for BU therapy if it was administered with amoxicillin/clavulanate, as previously demonstrated in M. tuberculosis [28]. Finally, because of their synergistic interactions with clarithromycin, it could replace rifampicin in the treatment of HIV patients under anti-retroviral therapy.
Our study comes as well with some limitations. First, although representative of different geographical origins, we only tested one ATCC strain and 9 clinical isolates. Further in vitro studies with a larger set of M. ulcerans clinical isolates would be needed to assess the full clinical potential and coverage of a triple combination including rifampicin, clarithromycin and amoxicillin/clavulanate. Second, our results were generated using synergy assays based on MIC determinations. This implies two limitations: (i) although an established approach in antimicrobial synergy assays, the activities of drugs alone and in combination were determined at a single time point after seven days of drug exposure and, (ii) synergy calculations inherently rely on sub-MIC concentrations, instead of actual serum levels achieved by drugs in clinical therapy. In order to thoroughly assess the effect of an eventual rifampicin, clarithromycin and amoxicillin/clavulanate combination at therapeutic concentrations, time kill assays would be required. However, even these assays would be challenging; since serum levels would be much higher than MIC values, the individual activity of drugs would mask any synergy signal using bactericidal activity as endpoint readout. Under these circumstances, and the much longer generation time of M. ulcerans compared to M. tuberculosis, assessing the sterilization capacity of such combinations would require extensive (months) incubation times of M. ulcerans cultures. Nevertheless, time kill assays are static pharmacokinetic (PK) / pharmacodynamics (PD) models where drugs are only added at the beginning of the assays and do not reproduce clinical therapy. The hollow fiber system, a dynamic PK/PD model by which posology and length of treatment can be mimicked in vitro, might provide data with higher prediction potential of treatment outcomes. This technology has proven a useful tool in the TB field, recently endorsed by the European Medicines Agency [65]; however, to date no laboratory has reported work on M. ulcerans using the hollow fiber system. Reasons for this might include the difficulties of working with a BSL3 pathogen that forms colonies in 2–4 months. Finally, in the field of BU (and TB), it is common practice to perform preclinical evaluation of drugs (or drug combinations) identified by in vitro studies using murine models of M. ulcerans infection. Although promising, experience from TB research has revealed inconsistencies between murine model data and clinical predictability [66, 67]. In addition, mice are a sub-optimal in vivo model to evaluate the activity of beta-lactams since their pharmacokinetics and efficacy in mice do not predict those found in humans [68]; this is in part due to the fact that mice express an enzyme that degrades beta lactams (renal dehydropeptidase I, DPH-I) at levels that are several orders of magnitude higher than in humans [69, 70], thus effectively reducing the time beta-lactams are over the MIC value. As such, murine models might not be the most appropriate development strategy for the use of beta-lactams in BU therapy. Because amoxicillin/clavulanate is a well-known antimicrobial with a clear track record of safety over decades of use, we believe that direct evaluation in clinical trials would be the fastest route to improve treatment of BU patients.
In summary, using a repurposing approach and in vitro technology already developed in TB R&D programs, we have identified amoxicillin/clavulanate as a new potential anti-BU drug to be used alone or in combination therapy with rifampicin and clarithromycin, current first-line anti-BU drugs, with the potential to reduce length of therapy and time to healing. Based on the strong synergistic interactions among amoxicillin with rifampicin and clarithromycin, amoxicillin alone might be added to the full course of a shorter therapy. However, because the main role of amoxicillin/clavulanate in the anti-BU therapy would be to reduce the initial bacterial load found in the lesions, we propose the use of high-dose extended release formulations during the first two weeks of therapy.
|
10.1371/journal.ppat.1003753 | A Gammaherpesvirus Uses Alternative Splicing to Regulate Its Tropism and Its Sensitivity to Neutralization | Human gammaherpesviruses are associated with the development of lymphomas and epithelial malignancies. The heterogeneity of these tumors reflects the ability of these viruses to route infection to different cell types at various stages of their lifecycle. While the Epstein Barr virus uses gp42 – human leukocyte antigen class II interaction as a switch of cell tropism, the molecular mechanism that orientates tropism of rhadinoviruses is still poorly defined. Here, we used bovine herpesvirus 4 (BoHV-4) to further elucidate how rhadinoviruses regulate their infectivity. In the absence of any gp42 homolog, BoHV-4 exploits the alternative splicing of its Bo10 gene to produce distinct viral populations that behave differently based on the originating cell. While epithelial cells produce virions with high levels of the accessory envelope protein gp180, encoded by a Bo10 spliced product, myeloid cells express reduced levels of gp180. As a consequence, virions grown in epithelial cells are hardly infectious for CD14+ circulating cells, but are relatively resistant to antibody neutralization due to the shielding property of gp180 for vulnerable entry epitopes. In contrast, myeloid virions readily infect CD14+ circulating cells but are easily neutralized. This molecular switch could therefore allow BoHV-4 to promote either, on the one hand, its dissemination into the organism, or, on the other hand, its transmission between hosts.
| Gammaherpesviruses are highly prevalent human and animal pathogens. These viruses display sophisticated entry mechanisms, allowing them to infect different cell types inside a host but also to transmit between hosts in the presence of neutralizing antibodies. Here, we used bovine herpesvirus 4 (BoHV-4) to decipher how some gammaherpesviruses manage to do this. We found that, as function of the originating cell types, BoHV-4 is able to modify its tropism as well as its sensitivity to antibody neutralization just by controlling the alternative splicing of one of its genes. This virus therefore exploits post-transcriptional events to generate viral populations with distinct phenotypes.
| Gammaherpesviruses are ubiquitous pathogens in human and animal populations all over the world. The best studied gammaherpesviruses, the Epstein-Barr virus (EBV) and the Kaposi's sarcoma-associated herpesvirus (KSHV), infect respectively some 90% [1] and 30% [2] of human populations. Primary infections by these viruses are usually subclinical, however, long-term carriage of these viruses can be associated with the development of various malignancies [3], [4] such as Burkitt lymphoma, nasopharyngal carcinoma, primary effusion lymphoma or Kaposi's sarcoma. The variety of these pathologies reflects the different tropisms of these viruses for distinct cell types. Understanding how these viruses orient their tropism is therefore essential for the development of efficient antiviral strategies and approaches to control the consequences of these infections.
Attachment to and penetration into the host cells are two distinct events in herpesvirus entry [5], [6]. As enveloped viruses, gammaherpesviruses enter cells by fusion with a cell membrane. While the precise mechanism of action is still unclear, the core fusion machinery is closely conserved and made of gB, gH and gL [6], although gL can be non-essential [7]. In contrast, the glycoproteins that mediate attachment and trigger fusion differ between viral species and also differ for the same virus depending on its target cell. This is well described for EBV, which for the most part infects epithelial cells and B lymphocytes [8]. Gp350 is the most abundant protein of the EBV envelope and is responsible for the attachment of the virus with high affinity to B cells [9]. On the opposite, gp350 deleted viruses are more infectious for epithelial cells [10], and gp350-specific antibodies enhance epithelial cell infection [11]. In addition to gp350, gp42 can function as a switch for EBV tropism. Gp42 binds to the human leukocyte antigen (HLA) class II [12], and determination of gp42 structures in its bound [13] and unbound [14] forms has made possible the development of a model for understanding how gp42 functions as a fusion activator. However, EBV makes both three-part gH/gL/gp42 complexes and two-part gH/gL complexes [15]. While fusion with a B cell is triggered by an interaction between gp42 and HLA class II [16], [17], entry into epithelial cells requires complexes without gp42 [15]. Interestingly, HLA class II expression in the virus-producing cells alters the ratio of three-part to two-part complexes. Therefore, virus produced in epithelial cells is more infectious for B cells whereas B-cell-derived virus is more infectious for epithelial cells [18]. Presumably, this switch in tropism favors the movement between epithelial cells and B cells during the cycle of persistence [19].
The mechanisms that regulate cell tropism are less clear for KSHV and other rhadinoviruses such as murid herpesvirus 4 (MuHV-4) or bovine herpesvirus 4 (BoHV-4). However, while these viruses do not have any gp42 homolog, they all encode a virion glycoprotein positionally homologous to EBV gp350/220 (BLLF1 gene). These proteins are K8.1A encoded by K8.1 in KSHV [20], gp150 encoded by MuHV-4 M7 [21] and gp180 encoded by BoHV-4 Bo10 [22]. Similarly to gp350, these proteins are involved in binding to some receptors on target cells. Indeed, K8.1A, gp150 and gp180 interact with glycosaminoglycans (GAGs) [22]–[24]. Moreover, these proteins seem to block the infection of cells that do not express this receptor [22], [23]. It has therefore been proposed that these proteins might regulate viral tropism both positively and negatively depending on the presence or the absence of their receptor [22], [23], [25]. While this putative model explains GAG+ cells infection, it does not encompass explanation of the infection of GAG-, cells such as B cells or monocytes, which are the cells that participate in the dissemination of these viruses in vivo [26], [27]. Therefore the question of how rhadinoviruses infect these cells remains controversial.
In this study, we showed that the BoHV-4 Bo10 gene encodes two different mRNAs through alternative splicing and that the cell-type regulated expression of one or the other transcript leads to production of virions that are phenotypically distinct. Thus, similarly to EBV, BoHV-4 progenies derived from different cell types differ in their cell tropism. Moreover, they also differ in their susceptibility to neutralizing antibodies. While epithelial virions are hardly infectious for circulating cells, they are relatively resistant to neutralization. In contrast, myeloid virions readily infect CD14+ circulating cells but are easily neutralized. We therefore propose that by regulating Bo10 mRNA splicing, BoHV-4 could promote either, on the one hand, its dissemination into the organism, or, on the other hand, its transmission between hosts.
Initial DNA sequence analysis of the BoHV-4 genome identified an ORF between ORF50 and ORF52 that was made of two exons [28]. With respect to its relative position in the genome, Bo10 is therefore similar to K8.1 of KSHV [20] and BLLF1 of EBV [29]. These two genes encode transmembrane glycoproteins that have been shown to be translated from a spliced message. Recently, we demonstrated that a Bo10 mRNA splicing removes 77 nucleotides and generates an mRNA encoding a 273 aa protein with a signal sequence and a membrane anchor [22]. Interestingly, sequence analysis revealed the presence of an in-frame STOP codon inside the Bo10 intron (Figure 1A) that could generate a protein (Figure 1B) similar to the potential expression product of the K8.1γ message [20]. To verify the transcription of this unspliced Bo10 message during the BoHV-4 cycle, we used RT-PCR with different pairs of primers (Figure 1C) on cDNA from BoHV-4 infected Madin-Darby bovine kidney (MDBK) cells. Primers spanning the intron revealed the existence of two different Bo10 messages. As previously described, the major 739 bp PCR product (Figure 1C) corresponded to the expected size of the product generated from the spliced mRNA. However, although weaker, an unspliced PCR product (817 bp) was also detectable (Figure 1C). RT-PCR with primer pairs specific to either the spliced or the unspliced sequence confirmed the existence of both Bo10 messenger RNAs (Figure 1C). None of the fragments could be amplified without prior reverse transcription (Neg. control lines) and DNA sequencing confirmed the identity of the different PCR products (data not shown). Altogether, these results demonstrate that the BoHV-4 Bo10 gene undergoes alternative splicing.
We previously described two BoHV-4 Bo10 knockout strains [22], [30]. To unravel the function of both the spliced and the unspliced Bo10 expression products, we generated two additional Bo10 mutant viruses. In the Bo10 MuDir mutant, we punctually mutated the Bo10 splicing donor site (T to G) to only express the unspliced form. In contrast, the Bo10 Spliced strain only expresses the Bo10 spliced product (Figure 2A). Southern blots of viral DNA (Figure 2B) confirmed the expected genomic structures of the two mutants and their associated revertant strains. The expected mutations were further confirmed by DNA sequencing (data not shown). As expected, RT-PCR analysis showed that the Bo10 MuDir and the Bo10 Spliced strains express the unspliced or the spliced messenger RNA respectively (Figure 2C). Finally, immunoblotting with an anti-Bo10-c15 rabbit polyserum confirmed that only the Bo10 MuDir mutant virions lacked gp180, encoded by the spliced Bo10 product (Figure 2D) while content of other proteins such as gB appeared to be normal (Figure S1).
We previously showed that the Bo10 knockout strains display a growth deficit associated with reduced binding to epithelial cells. We interpreted this phenotype as being a consequence of the absence of the gp180 glycoprotein that is encoded by the Bo10 spliced message. However, as the Bo10 gene encodes two different transcripts, this growth deficit could also be associated with the absence of the unspliced Bo10 message. To address this question, multi-step growth assays were performed on MDBK cells with the different Bo10 mutants. Interestingly, only the Bo10 MuDir virus grew to lower titers than the WT BAC parental strain (Figure 3A).
With the Bo10 knockout strains, we interpreted the growth deficit on epithelial cells as a consequence of reduced binding to GAG [22]. Moreover, we showed that Bo10 deletion enhanced infection of GAG-deficient cells [22]. We therefore compared the relative dependence on GAGs of the different BoHV-4 strains constructed in this study by infecting CHO-K1 fibroblasts competent or not for GAG expression (Figure S2). As observed with the Bo10 deficient-viruses, Bo10 MuDir virions infected CHO GAG+ cells similarly to the WT virus but infected CHO GAG− cells much better (Figure S2). The Bo10 spliced virions (expressing gp180) behaved similarly to WT virions (Figure S2). In vivo, BoHV-4 infects monocytes, which are, however, relatively GAG deficient [31]. We therefore compared the capacity of the different viruses to infect rabbit (Figure S3) and cow (Figure 3B) peripheral blood mononuclear cells (PBMCs) ex vivo. While all the other strains infected rabbit and cow CD14+ PBMCs very poorly, Bo10 MuDir viruses infected them much better (Figure S3 and 3B).
We have previously shown that gp180 provides part of a glycan shield for otherwise vulnerable viral epitopes [30]. Indeed, antibodies have greater access to gB, gH and gL on Bo10 knockout virions and this correlates with a greater susceptibility to neutralization by immune sera [30]. As we have shown that Bo10 encodes two transcripts, we used Bo10 spliced and Bo10 MuDir mutants to investigate which of these transcripts is responsible for antibody evasion. Infected cell surfaces provide a means of probing antigenic differences between BoHV-4 glycoprotein mutants. We therefore compared epitope accessibility on cells infected by the different BoHV-4 Bo10 mutant viruses. Only the cells infected by the Bo10 MuDir and Stop mutants were more heavily stained with mAb 29 recognizing gB (Figure 3C). This result was not due to differences in protein expression, as permeabilized cells gave similar staining with each virus (Figure 3C). It therefore appears that only the spliced Bo10 transcript, encoding gp180, is involved in shielding other epitopes at the cell surface.
We therefore further compared the sensitivity of BoHV-4 WT BAC, Bo10 STOP, Bo10 MuDir and Bo10 Spliced strains to neutralization by sera of rabbits infected with the BoHV-4 V.test strain (Figure 3D). WT and Bo10 Spliced virions were poorly neutralized. In contrast, Bo10 STOP and Bo10 MuDir virions were much more efficiently neutralized. Especially, complete neutralization was now possible. Thus, Bo10 splicing and the subsequent gp180 expression are critical to limiting virion neutralization.
Different bovine cell lines were then investigated for expression of Bo10 spliced and unspliced transcripts by quantitative PCR. In order to be able to compare the numbers of both transcripts, a plasmid encoding both Bo10 spliced and unspliced specific sequences was constructed and used as a unique standard curve for both reactions. Interestingly, while the infected epithelial MDBK cells displayed around 1000 times more Bo10 spliced transcript than unspliced transcript, this ratio was more than 10 times lower in the BoMac myeloid cells (Figure 4A). We then compared the number of Bo10 spliced transcripts, encoding gp180, to the number of ORF8 transcripts, encoding gB, as we have shown that one of the glycoproteins shielded by gp180 is gB (Figure 3C) [30]. While the infected epithelial MDBK cells displayed around 10 times more Bo10 spliced transcripts than ORF8 transcripts, this ratio was around 10 times lower in the BoMac myeloid cells (Figure 4B), suggesting that in these cells, there could be less expression of gp180 glycoprotein in comparison to gB. Similar results were obtained when the amount of Bo10 spliced transcripts was compared to the number of ORF47 mRNA encoding gL (Figure S4), as gL is another viral glycoprotein shielded by gp180. Finally, these results were confirmed by western blotting on MDBK or BoMac cells (Figure 4C). Interestingly, while gB glycoprotein was readily detectable in both cell types, almost no gp180 glycoprotein was observed in BoHV-4 WT infected BoMac cells.
Finally, incorporation of gp180 in MDBK and BoMac virions was assessed by immunoblotting on purified virions. As observed on infected cells, BoMac virions contain relatively less gp180 than MDBK virions while amounts of gB appear to be similar (Figure 5). Blotting with a rabbit polyserum raised against BoHV-4 virions was used as loading control. These results therefore suggest that the BoHV-4 phenotype could change with the type of cell in which it is grown, namely epithelial or myeloid cells.
We first compared phenotypes of MDBK and BoMac derived virions by co-culture experiments (Figure 6A). While naïve MDBK cells co-cultured overnight with BoHV-4 WT BAC infected MDBK cells became >80% eGFP+, CD14+ PBMCs cultured under the same conditions remained nearly entirely uninfected (<1% eGFP+) (Figure 6A). In contrast, BoHV-4 WT BAC infected BoMac cells produced lower amount of virions, as demonstrated by the fact that only ∼20% of co-cultured MDBK cells became eGFP+. However, these few virions were much more infectious for CD14+ PBMCs (>10% eGFP+) in comparison with MDBK derived virions (Figure 6A). Therefore direct contact with infected myeloid BoMac cells allowed efficient BoHV-4 infection of circulating CD14+ PBMCs.
We next tested whether cell-free virions derived from myeloid BoMac cells could also better infect CD14+ PBMCs compared to cell-free virions derived from epithelial MDBK cells. While MDBK-derived virions barely infected 0.5% of PBMCs at the multiplicity of infection (MOI) of 3, according to titers measured on MDBK cells, BoMac-derived virions infected these cells substantially more efficiently as around 5% of CD14+ PBMCs became positive (Figure 6B).
The second phenotypic difference between Bo10 MuDir and Spliced virions was related to resistance to neutralization by immune sera. Therefore, we compared sensitivity to serum neutralization of BoHV-4 WT BAC virions derived from MDBK or BoMac cells. As previously observed [30], MDBK-derived WT virions were poorly neutralized by immune sera. In contrast, BoMac-derived WT virions were neutralized much more efficiently (Figure 6C).
Finally, in order to demonstrate that the phenotypic difference of BoHV-4 myeloid virions was mainly due to reduced Bo10 mRNA splicing, Bo10 Spliced virions were propagated on BoMac cells and compared to WT BAC virions derived from MDBK or BoMac cells (Figure 7). While WT BAC BoMac-derived virions readily infected CD14+ PBMCs, Bo10 Spliced virions derived from BoMac cells infected very few CD14+ PBMCs similarly to WT BAC virions propagated on MDBK cells (Figure 7A). Moreover, in contrast to BoMac-derived WT virions, Bo10 Spliced virions propagated on BoMac cells resisted neutralization similarly to BoHV-4 WT BAC virions derived from MDBK cells.
Altogether, these results show that regulation of Bo10 mRNA splicing offers the possibility of regulating both the tropism and the antigenicity of BoHV-4 virions.
Herpesvirus lifecycles are probably among the most complex lifecycles of all viruses. Indeed, these viruses are able not only to engender either latent or lytic infections but are also able to do so in different cell types at different stages of infection. This is particularly true of gammaherpesviruses, for which viral replication in mucosal epithelium appears to be mainly important for host entry [32] and exit [33], [34], whereas latency establishment in circulating leukocytes ensures host colonization [35]. Infection of these two key targets has proven to follow substantially different pathways [8] and regulation of these processes may allow these viruses to route infection in vivo. For example, EBV appears to use gp42 as a switch of cell tropism [18]: epithelial cells produce virions high in gH/gL/gp42 complexes, which promote B-cell infection, while B cells produce viruses low in gp42, which efficiently infect epithelial cells but not B cells.
While rhadinoviruses, such as KSHV or BoHV-4, share a similar alternate tropism in vivo, the mechanisms underlying this property were still unknown. Thus, BoHV-4 persists in circulating CD14+ cells in vivo, yet infects them poorly in vitro. In this study, we showed that alternative splicing of the Bo10 gene defines the cell tropism of BoHV-4 virions. In addition to the Bo10 spliced message, that encodes gp180, an unspliced Bo10 mRNA is transcribed during the BoHV-4 cycle (Figure 1). Viral strains expressing one or the other transcript showed different infectivity patterns associated with GAG expression (Figure S2). Exclusive expression of the unspliced transcript generated virions that were both less infectious for GAG+ cells than the wild-type and more infectious for GAG− cells. On the opposite, exclusive expression of the spliced message blocked GAG− cells infection. Moreover, relative proportions of these two transcripts determined the release of distinct viral populations. Epithelial cells, expressing high amounts of gp180, produced virions that were unable to infect CD14+ in contrast to virions derived from myeloid cells, expressing low amounts of gp180 (Figures 4–6).
These results suggest an epithelial/myeloid/circulating leukocyte infection pathway for BoHV-4 (Figure 8), similar to the one recently described for MuHV-4 [36]. Indeed, we had previously showed that the BoHV-4 Bo10 gene positively regulates the infection of GAG-bearing cells such as epithelial cells [22]. In this study we showed that this is through the expression of the Bo10 spliced product, encoding gp180. In vivo, such interaction probably occurs at specific locations displaying accessible heparin sulfates structures such as the olfactory neuroepithelium as shown for MuHV-4 [32]. Replication in epithelial cells likely allows infection of some myeloid cells which are susceptible to infection by epithelial virions. Finally, this myeloid infection allows infection of circulating leukocytes, such as CD14+ cells in the case of BoHV-4. Interestingly, a similar role of myeloid infection has been proposed in the case of MuHV-4 [36]. However, differences exist between the two models. Firstly, MuHV-4 infects mainly circulating B cells. Secondly, RT-PCR has never demonstrated MuHV-4 gp150 truncation by splicing nor were myeloid cell-derived MuHV-4 virions gp150-deficient [36]. Moreover, the results obtained with MuHV-4 suggested that the myeloid-derived virions phenotype could be associated with conformational changes in gB and gH. While this cannot be excluded in our model, the fact that the forced expression of gp180 in myeloid cells, as seen with Bo10 Spliced virions (Figure 7), or the absence of expression of gp180 in epithelial cells, as seen with Bo10 MuDir virions (Figure 3B), is sufficient to switch cell tropism suggest that Bo10 alternative splicing is the main determinant of BoHV-4 virion tropism. While our main hypothesis is that gp180 is directly involved in switching the tropism of BoHV-4 virions (direct effect), we cannot rule out that the amount of gp180 that is incorporated in virions affects either the recruitment of another protein or its conformation (indirect effect).
When seen only from the angle of cell tropism, the role of gp180 appears unclear. Indeed, it is unnecessary for growth on epithelial cells and it negatively regulates infection of GAG− cells (Figures 3, S2 and S3). Moreover, a BoHV-4 strain deleted for the Bo10 gene displays no latency establishment deficit in vivo [30]. Positive selection pressure for gp180 expression is therefore difficult to understand in these contexts. As was recently demonstrated for influenza, the evasion of neutralizing antibodies is one of the main forces that drive virus evolution [37]. This is particularly true for gammaherpesviruses that establish lifelong latency and are therefore continuously shed in the presence of neutralizing antibodies [34]. In this context, we recently showed that the deletion of BoHV-4 Bo10 gene markedly sensitized virions to neutralization by immune sera. The results obtained here demonstrate that this phenomenon is related, either directly or indirectly, to gp180 expression, as only Bo10 MuDir virions and not Bo10 spliced virions displayed increased sensitivity to neutralization (Figure 3D). Similarly, WT virions derived from myeloid cells were more easily neutralized than epithelial cell-derived WT virions (Figure 6C). Again, this was associated with reduced expression of the spliced Bo10 transcript, as the forced expression of gp180 in myeloid cells substantially reduced sensitivity to neutralization by immune sera (Figure 7B). These observations make sense in the context of an in vivo cycle as epithelial cell-derived virions have to face neutralizing antibodies that are likely to be present at mucosal sites. In contrast, because spread of a gammaherpesvirus within the host is more likely to involve direct cell–cell contact [38] rather than cell-free virion release, myeloid cell-derived virions may not have to face a similar constraint. Epithelial virions would therefore be better fit for virus transmission between hosts, while myeloid virions would be more efficient for host colonization.
A central question about this mechanism is why BoHV-4 uses alternative splicing of the Bo10 gene instead of expressing gp180 from an unspliced message that would be turned on or off. One of the reasons could be that the unspliced Bo10 message has its own function. Thus, as gp180 interacts with GAGs, the potential soluble form of gp180 could coat the infected cell surface and therefore promote the release of progeny virions by preventing their interaction with cell surface GAGs. Interestingly, a truncated isoform of the major glycoprotein is also secreted from infected cells during EBOLA virus infection [39]. This protein is involved in antibody evasion as it not only serves as a decoy for adsorbing preexisting neutralizing antibodies [40], [41] but also contributes to antigenic subversion of the host immune repertoire [41]. As antibody evasion is particularly important for gammaherpesviruses, the potential soluble form of BoHV-4 gp180 could share similar roles. This hypothesis would imply that gp180 itself is a neutralization target. At the moment, this is not known. This will therefore deserve future studies.
Similar mechanisms could occur in KSHV, where the gp180 homolog is encoded by K8.1. As observed for BoHV-4 Bo10, K8.1 displays several alternative spliced forms [20]. Two proteins, K8.1A and B, are generated from spliced message and encode transmembrane glycoproteins that bind to cell surface heparan sulfate [24], [42]. These proteins are thought to be involved in the initial steps of virion attachment. K8.1 also encodes a third unspliced message, K8.1γ [20], which might likewise code for a soluble protein. Interestingly, a novel alternative message has also recently been described for the EBV BLLF1 gene, encoding gp350/220 [43]. The potential protein encoded by this product also lacks a transmembrane domain. The usage of alternative splicing could therefore allow EBV and KSHV to regulate incorporation of their respective g180 homologs in a similar way to what we showed for BoHV-4 although other regulation mechanisms may exist. This phenomenon and the potential functional consequences on EBV or KSHV tropism will have to be tested in the future.
Deciphering the protein coding complexity of herpesviruses is far from complete. Numerous reports point to a major role for regulated use of alternative splicing in enabling tight temporal control of protein expression and allowing multiple distinct polypeptides to be generated from a single genomic locus [43]–[47]. However, the roles of these splicing events remain largely unknown. Our results show that alternative splicing of the BoHV-4 Bo10 gene orientates viral tropism and determines virion sensitivity to neutralizing antibodies.
The BoHV-4 V.test strain (WT), was initially isolated from a case of orchitis [48]. Bo10 Del, Bo10 Stop and the corresponding revertant strains have been described previously [22], [30]. The other viruses were derived from a cloned BoHV-4 BAC (WT BAC) [49]. Unless stated otherwise, the viruses were grown on MDBK cells.
MDBK (ATCC CCL-22) and BoMac [50] cells were cultured in Dulbecco's modified Eagle Medium (Invitrogen) containing 10% fetal calf serum (FCS), Penicillin (200 U/mL)/Streptomycin (200 µg/mL) (Invitrogen) and non-essential amino acids (Invitrogen) diluted following the manufacturer's recommendations. Bovine and rabbit PBMCs were prepared as described elsewhere [22]. Briefly, PBMCs were isolated from 10 ml of blood. Mononuclear cell suspensions were prepared with Ficoll-Paque Premium density gradient media (GE Healthcare) as follows. Cell suspension in sterile PBS was overlaid onto a 7 ml Ficoll-Paque density cushion and centrifuged (1825×g) for 20 min at room-temperature. Mononuclear cells at the interface were collected and washed twice in ice-cold PBS before further analysis. PBMCs were cultured in RPMI Glutamax Medium containing 10% FCS, Penicillin (200 U/mL)/Streptomycin (200 µg/mL) (Invitrogen), non-essential amino acids (Invitrogen) diluted following the manufacturer's recommendations, 1 mM Sodium pyruvate, 25 mM HEPES and 50 µM 2-mercaptoethanol.
For detection of gp180, encoded by the Bo10 spliced transcript, on western blotting, we used a rabbit monospecific polyserum raised against the C-term end of the gp180 protein (anti-Bo10-c15) [22]. Therefore, this polyserum does not recognize the potential protein encoded by the Bo10 unspliced transcript. The mouse monoclonal antibodies (mAbs) 29 and 35 were raised against gB [30]. The mouse anti-human CD14 Pacific Blue was purchased from Serotec. Rabbit anti-BoHV-4 WT polysera were obtained previously [30]. The PKH26 Red Fluorescent Cell Linker (Sigma) was used for general cell membrane labeling.
Sub-confluent monolayers of MDBK cells were infected at a MOI of 1 PFU/cell. 24 hours after infection (p.i.), cytoplasmic RNA was isolated by using RNeasy mini kit (Qiagen). Contaminating DNA was removed by DNase treatment. cDNA was produced by using the First Strand cDNA Synthesis Kit (Roche Applied Science) with poly d(T) primer. The cDNA products were amplified by PCR with Taq polymerase (New England Biolabs), and specific primer pairs. Bo10 23-43 (5′-TCATACATTCAAATTGCATGC-3′) and Bo10 839-818 (5′-CATTGAATGAGAACAAACACG-3′) were used to amplify both transcripts. Bo10 PBDF (5′-ATGAGGTTAAGAGTCAGATC-3′) and Bo10 IntronRev (5′-ACCATTTAGTCAAATTCCACAC-3′) were used to amplify unspliced Bo10 product. Bo10 PBDF and Bo10 SplicedOnlyRev (5′-GGATGTCTGTGTGCCTGAG- 3′) were used to amplify the spliced Bo10 product.
We modified the BoHV-4 V.test Bo10 coding sequence (genomic coordinates 65844 to 66743) by BAC mutagenesis [49] to generate two supplemental recombinant viral strains. In the first one (Bo10 MuDir strain), we introduced a single point mutation in the splicing donor site. In the second one, we replaced the entire Bo10 ORF with a sequence devoid of the intron (Bo10 Spliced strain). The Bo10 MuDir and the Bo10 Spliced strains were generated by a two-step mutagenesis procedure in bacteria using the shuttle plasmid pST76KSR [49]. Plasmids to induce homologous recombination were constructed as follows. For the Bo10 MuDir strain, we first PCR-amplified (Platinum Taq DNA Polymerase High Fidelity, Invitrogen) coordinates 65331–67427 of the BoHV-4 V.test genome, including SacI and XmaI restriction sites in the respective forward and reverse primers, and T/A cloned the PCR product into the pGEM-T Easy vector (Promega Corporation) resulting in pGEM-T Easy Bo10 zone Rec plasmid. Next, the Bo10 splicing donor site was mutated by site directed mutagenesis (Quikchange site directed mutagenesis, Stratagene) resulting in pGEM-T Easy Bo10 zone Rec MuDir plasmid. For the Bo10 Spliced strain, cDNA was prepared from WT BoHV-4 infected MDBK cells and Bo10 specific sequences were amplified with Bo10 23-43 and Bo10 839-818 specific primers. The lower 740 bp band was then BstBI/HpaI-restricted and cloned into the corresponding sites of the pGEM-T Easy Bo10 zone Rec plasmid resulting in pGEM-T Easy Bo10 zone Rec Spliced plasmid. Each construct was then subcloned as a SacI/XmaI fragment into the same sites of the pST76K-SR shuttle vector, and recombined into the BoHV-4 BAC [49]. In the same way, we also isolated revertants in which the Bo10 locus was restored to its wild-type form. Reconstitution of the infectious virus from BAC plasmids was obtained by transfection in MDBK cells.
Southern blot analysis of viral DNA digested with BamHI was performed with a probe corresponding to nucleotides 65900–66370 of the BoHV-4 V.test strain genome (coding for Bo10 Exon 1) [51].
BoHV-4 strains grown on MDBK cells were purified as follows. Briefly, after removal of the cell debris by low-speed centrifugation (1,000× g, 10 min at 4°C), virions present in the infected cell supernatant were harvested by ultracentrifugation (100,000× g, 2 h at 4°C) through a 30% (wt/vol) sucrose cushion. Virions were then banded by isopycnic gradient ultracentrifugation in a 20 to 50% (wt/vol) potassium tartrate gradient in PBS (100,000× g, 2 h at 4°C). The band containing virions was collected (∼3 mL), diluted ten fold in PBS and pelleted by ultracentrifugation (100,000× g, 2 h at 4°C). The pellet was finally resuspended in PBS, and virus enriched preparations were stored at −80°C.
Purified virions were lysed and denatured by heating (95°C, 5 min) in SDS-PAGE sample buffer (31.25 mM Tris-HCl pH 6.8, 1% (w/v) SDS, 12.5% (w/v) glycerol, 0.005% (w/v) Bromophenol Blue, 2.5% (v/v) 2-mercaptoethanol). Proteins were resolved by electrophoresis on Mini-PROTEAN TGX (Tris-Glycine eXtended) precast 7.5% resolving gels (Bio-Rad) in an SDS-PAGE running buffer (25 mM Tris-base, 192 mM glycine, 0.1% (w/v) SDS) and transferred to polyvinylidene difluoride membranes (Immobilon-P transfer membrane, 0.45 µM pore size, Millipore). The membranes were blocked with 3% non-fat milk in PBS/0.1% Tween-20, and then incubated with the anti-Bo10-c15 rabbit antibodies, with the anti-BoHV-4 polyserum or with the mAb 35 in the same buffer. Bound antibodies were detected with horseradish peroxidase-conjugated goat anti-rabbit IgG pAb or anti-mouse IgG pAb (Dako Corporation), followed by washing in PBS/0.1% Tween-20, development with ECL substrate (GEHealthcare) and exposure to X-ray film.
The growth kinetics of mutant and revertant viruses were compared to that of WT. Cell cultures were infected at a MOI of 0.01 (multi-step assay). After 1 h of adsorption, the cells were washed and then overlaid with MEM containing 5% FCS. Supernatants of infected cultures and infected cells were harvested together at successive intervals and the amount of infectious virus was determined by plaque assay on MDBK cells [52].
For PBMCs, staining, washes and incubation steps were performed in FACS buffer (PBS pH 7.4, 0.1% BSA, 0.05% NaN3). Cells were incubated with Pacific Blue-conjugated anti-human CD14 (1/50) on ice for 45 min. After washing, cells were analyzed for eGFP and Pacific-Blue fluorescences. Dead cells were excluded with 7-AAD staining. For cell surface staining of viral glycoproteins, cells were infected by the different virus strains at a MOI of 2 PFU/cell for 36 h. After trypsinization and one wash in PBS, the cells were stained with monoclonal antibodies in PBS containing 10% fetal calf serum for 1 h at 4°C. After one wash in PBS, the cells were incubated with goat anti-mouse IgG-Alexa Fluor 633 (Invitrogen) diluted 1∶1,000 in PBS containing 10% FCS for 1 h at 4°C. After one wash in PBS, the cells were analyzed. For intracellular staining, cells were fixed in 1% paraformaldehyde (PFA) for 30 min at room temperature and then permeabilized with 0.1% saponin. Cells were incubated in PBS 10% FCS 0.1% saponin (1 h, 4°C) with the different mAbs specific for BoHV-4 glycoproteins followed by Alexa 633-conjugated goat anti-mouse pAb (Invitrogen). Cells were then washed and analyzed. All the analyses were performed on a FACSAria flow cytometer (Becton Dickinson).
Neutralization was tested by incubating virus with antibody for 1 h at 37°C before adding the virus/antibody mixtures to MDBK cells for a further 3 h. The cells were then overlaid with 0.6% carboxymethylcellulose and the cell monolayers were fixed for plaque counting (based on eGFP expression) after a further 4–5 days. The neutralization assays were carried out with similar amounts of starting virions, in order to ensure that the starting ratio of viral particles to antibodies is similar.
Total cDNA was produced as described above and analyzed by quantitative PCR with iQ SYBR green supermix (Bio-rad) containing 625 nM of each primer. Quantitative PCR reactions were carried out under the following conditions: initial activation of the Taq polymerase (Bio-Rad) at 95°C for 3 min followed by 45 cycles comprising one step of 95°C for 30 sec, one step of 56°C for 45 sec and one step of 72°C for 45 sec. Dissociation curves were performed to check for the presence of a single peak corresponding to the required amplicon.
In parallel, fragments corresponding to both products were quantified. On the one hand, a fragment corresponding to the BoHV-4 Bo10 spliced product, was amplified with the forward primer Bo10 23-43 and the reverse primer Bo10 SplicedOnlyRev. On the other hand, the forward primer Bo10 intron sens (5′GTCCATGTGTGTTAAATCGGG3′) and the reverse primer Bo10 839-818 were used to amplify the specific Bo10 unspliced product. In order to be able to compare amounts of both products, a pGEM-T easy containing specific fragments of both Bo10 transcripts was constructed to establish a unique standard curve. Briefly, we first introduced by T/A-cloning in pGEM-T-easy vector a specific Bo10 spliced sequence generated by PCR on cDNA with Bo10 PBDF and Bo10 Spliced Rev BamHI 5′-GGATCCTGGGAGGTTGTGTTGAAGAGT-3′ as primers resulting in pGEM-T spliced vector. Then a sequence specific of Bo10 unspliced product was generated by using the Bo10 MuDir cDNA as a template and Bo10 intron sens BamHI 5′-GGATCCGTCCATGTGTGTTAAATCGGG-3′ and Bo10 PBDR BamHI 5′- GGATCCTCATAATAAATTATATCCCTGACTATAATT-3′ as primers. This PCR product was restricted with BamHI and ligated into the BamHI site of the pGEM-T spliced vector generating the pGEM-T spliced/unspliced plasmid.
To estimate the proportion of the spliced Bo10 transcript relative to the ORF8 (encoding gB) and ORF47 (encoding gL) transcripts, we amplified the Bo10 spliced as described above and fragments corresponding to BoHV-4 ORF8 and ORF47 with the forward primer 8startfw (5′- CAAATAGTTCATTAGCTGCCTCTCC -3′) and the reverse primer 8middlerev (5′- TCATCAGTAACAGTTGGAATAGTGG -3′), and with the forward primer 47startfw (5′- AAGGATCCGCCGCCACCATGAGAGATATCTATGTTTTTTGT -3′) and the reverse primer 47rev (5′- AACTCGAGCTATAATCTGCCCAGGCCAC -3′) respectively. For these comparisons, we used DNA from the BoHV-4 BAC G Bo10 Spliced plasmid as the unique standard curve.
For each comparison, serial dilutions of the standard curve were made and the same amounts were used in the different reactions. The numbers of copies have been determined based on the measure of the DNA concentration of the standard curve.
MDBK or BoMac cells in suspension were stained with PKH26 as described by the manufacturer. 1.105 labelled cells were transferred into 6-well plates and then infected with the WT BAC strain at a MOI of 1. Three hours post-infection, cells were washed with acidic solution (PBS pH 3) in order to inactivate and remove the inoculum. 12 hours post infection, 2.105 freshly isolated PBMCs or MDBK controls were added. After 24 hours, the co-cultivated cells were collected with a cell dissociation buffer (Invitrogen) and stained by the Pacific Blue-conjugated anti-CD14 for FACS analysis.
|
10.1371/journal.pcbi.1006377 | Characterising seasonal influenza epidemiology using primary care surveillance data | Understanding the epidemiology of seasonal influenza is critical for healthcare resource allocation and early detection of anomalous seasons. It can be challenging to obtain high-quality data of influenza cases specifically, as clinical presentations with influenza-like symptoms may instead be cases of one of a number of alternate respiratory viruses. We use a new dataset of confirmed influenza virological data from 2011-2016, along with high-quality denominators informing a hierarchical observation process, to model seasonal influenza dynamics in New South Wales, Australia. We use approximate Bayesian computation to estimate parameters in a climate-driven stochastic epidemic model, including the basic reproduction number R0, the proportion of the population susceptible to the circulating strain at the beginning of the season, and the probability an infected individual seeks treatment. We conclude that R0 and initial population susceptibility were strongly related, emphasising the challenges of identifying these parameters. Relatively high R0 values alongside low initial population susceptibility were among the results most consistent with these data. Our results reinforce the importance of distinguishing between R0 and the effective reproduction number (Re) in modelling studies.
| When patients present to their doctor with influenza-like symptoms, they may have influenza, or some other respiratory virus. The only way to discriminate between these viruses is with an expensive test, which is not performed in many cases. Additionally, results other than influenza may not be reported. This means that it can be difficult to determine how much influenza is circulating in the population each season. We used a unique dataset of confirmed influenza with denominators to fit models for seasonal influenza in New South Wales, Australia. Knowing the denominators allowed us to estimate population level trends. We found that the relationship between influenza transmission rates and immunity due to previous infections was critical, with relatively high transmission corresponding to substantial preexisting immunity likely. This existing immunity is critical to understanding and effectively modeling influenza dynamics.
| Influenza is a highly contagious, rapidly evolving respiratory virus that circulates globally on a seasonal basis [1, 2]. It can cause death in at-risk groups, and places a high burden on health systems (e.g., emergency room (ER) and general practitioner (GP) services). In some cases, antigenic shift creates circumstances where influenza can be particularly infectious or dangerous, resulting in influenza pandemics such as those that occurred in 1918 and 2009. However, understanding normal seasonal circulation is critical for effective healthcare resource allocation, forecasting future seasonal dynamics, and early detection of anomalous seasons.
Influenza has been widely studied, through both experimental studies on animal models such as ferrets (e.g. [3, 4]), and modelling studies based on data from human populations [5, 6]. There remains substantial uncertainty surrounding the way influenza spreads within populations. Model-based analyses depend on the underlying assumptions and available data, and knowledge gaps exist that may be difficult or impossible to test using animal models (e.g., for seasonal influenza, long-term patterns of immunity through and across many seasons). Parameter estimation can also be challenging. For example, a key epidemic parameter is the basic reproduction number, R0, which is defined using a completely susceptible population. However, in practice few populations are likely to be completely susceptible, and so estimates are instead of Re, the effective reproduction number [7, 8]. To determine R0 itself, it is necessary to also have an estimate of the proportion of individuals that are susceptible at the beginning of an outbreak. Unfortunately, estimating the initial susceptible proportion is close to impossible in practice, given the complexity of immunity and interactions between strains. Immunity to a strain may vary substantially between individuals with similar infection or vaccination histories [9], and the life course of antibodies and interactions between antibodies to different strains is complex [10]. The rapid evolution and multi-strain nature of influenza means that individuals may have variable immune responses to different strains, particularly in the presence of complex phenomena such as “antigenic seniority” [11, 12, 13], whereby individuals have increased long-term immunity to the first strain they were exposed to during their lifetime. Recent evidence suggests that there may be some level of immunity existing in populations even to novel influenza strains with pandemic potential [14, 15]. The presence of antibodies in an individual, particularly at low levels, may not prevent reinfection but may reduce the chance of infection when challenged or lead to a milder or asymptomatic illness [16, 17]. Further, many historical studies have relied on serological data as a means of determining if influenza infection occurred (e.g. [18, 19, 20]). However, the results of serological studies can vary substantially based on the criteria used [21], and produce estimates that are vastly different from active surveillance of symptomatic cases (e.g., [22]), and the relationships between antibodies and immunity is not completely understood [23].
Estimates for Re are often between 1.0–2.0 (e.g., see [24]), however in rare cases estimates of R0 have been as high as 22.7 [25, 26]. While these most extreme values may be highly unlikely, estimates in the intermediate range are plausible when associated with reduced population susceptibility. Assumptions around parameter values can propagate through a modelling analysis, and impact on how we understand and characterise epidemic dynamics, e.g., when estimating the final size of a seasonal epidemic, or when calculating optimal vaccination policy. As such, careful analysis of high-quality data, estimating both population susceptibility and R0, is critical to inform present and future modeling efforts.
Effective modelling of influenza dynamics is further complicated by the paucity and quality of available data. Influenza is difficult to identify clinically; patients presenting with influenza-like symptoms may have influenza (Fig 1), but they may also have some other respiratory virus, such as human rhinovirus, or respiratory syncytial virus [27]. To determine the actual cause of these symptoms requires a (relatively expensive) test, which may be appropriate for an at-risk individual presenting serious symptoms at a hospital but is unlikely to be necessary for an otherwise healthy individual in the 15–45 age group (for example). In addition, patients may be asymptomatic (while potentially still able to infect others), or those with mild symptoms may choose not to seek medical treatment at all. So, confirmed cases are likely to be only a subset of actual cases, with poor “denominator data”. The hierarchical observation process that links population level epidemic dynamics to observed confirmed cases (Fig 2) may make identifiability challenging and can in some cases introduce bias to estimates [28, 29, 30].
In this study, we used a new, high-quality dataset along with modern Bayesian parameter estimation methods to investigate seasonal influenza dynamics in New South Wales, Australia, from 2011–2016. We used Polymerase Chain Reaction (PCR) confirmed influenza infections from the Australian Sentinel Practices Research Network (ASPREN; [31]) to fit models. Within these data, approximately 0.5% of general practitioners (GPs; known as family physicians in North America) within Australia reported each Influenza-like illness (ILI) case that they observed, along with the number of consultations they performed, and submitted swab samples for c. 20% of these ILI cases for virological testing via PCR. These data were available weekly. Knowing the numbers of doctors, consultations, ILI observed, and samples tested provides critical and unparalleled denominator data against which to assess the hierarchical observation process between population-level influenza dynamics and the observed confirmed infections. We fit model parameters using approximate Bayesian computation (ABC), assuming the same strain-specific parameters for seasons in which the same predominant strain circulated. By performing this analysis using high-quality confirmed influenza data (including denominators), state of the art parameter estimation techniques, and minimal assumptions, we are able to characterize the epidemiology of influenza within populations. In doing so, we aim to both progress the science of disease modeling, and produce new epidemiological knowledge around population-level immunity to seasonal influenza.
We chose to use an SEIRS-type model, with an Erlang-2 infectious period and exponential exposed and immune times, to model seasonal influenza (Fig 3), which is consistent with known influenza biology. In this form of model, we track the number of individuals in the population who are susceptible (S), exposed but not yet infectious (E), infectious (I), or recovered and immune from reinfection (R) at any time. We include an observed class (O) to allow individuals to choose to seek treatment, or not. Note we allow immunity to wane, so that individuals return from the recovered class to the susceptible class, and include external importation of cases at a low rate. Transition events are stochastic, with rates that depend on the number of individuals in each class, and fitted parameters that relate to the epidemiology of the disease. We fit the initial proportion of individuals that are susceptible to the strain circulating in the given season as a parameter, and model seasons that had the same predominant circulating strain together, i.e., 2011 and 2013 both had predominantly H1N1pdm09 circulating, so we fit these seasons to have the same epidemic parameters, but each with a separate initial population susceptibility. The presence of an exposed class in the chosen model indicates that there is some lag between when an individual is exposed to the disease, and when they become infectious. We assume that they are not symptomatic during this exposed period. This is consistent with many existing models and known influenza epidemiology [32].
By using models aggregated at a state level, without any age- or spatial- structuring, we make a number of simplifying assumptions. Specifically, we lose the ability to model heterogeneity in mixing and thus infection rates, vaccination coverage, immunity, or case ascertainment. We know that these factors should, in reality, be heterogeneous, as there is strong evidence for variations based on (in particular) age structure [33, 34]. However, the data we have access to are unfortunately too sparse to effectively parametrize models at this fine level of detail. In some cases, this may introduce some degree of bias. For example, newborn children entering the population are necessarily susceptible, and by not tracking demographics we may overestimate the rate at which immunity wanes (i.e., underestimate the duration of immunity), which could potentially have some impact on estimates of other parameters, including R0.
Having chosen the SEIRS-type model, we used ABC to estimate posterior distributions for model parameters, and to evaluate the impact of these fitted model parameters on underlying epidemic dynamics at the population level. We parameterized transmission based on physical quantities, including the latent, infectious, and immune durations, the probability of an infectious person seeking treatment over their infectious period, the basic reproduction number, R0, and the proportion of individuals susceptible to the circulating strain at the beginning of each season. R0 was assumed to have seasonal forcing, for which we used variations in temperature (T) from the annual mean (T ¯), i.e., R 0 = R ¯ 0 + a ( T - T ¯ ), with R ¯ 0 and a being estimated parameters. Alternate seasonal forcing terms were tested (e.g., based on specific humidity with the same form as Yang et al. [35]), however these produced similar quality results while requiring more parameters, and were harder to interpret. We therefore chose to use temperature-based forcing for simplicity. Results were broadly similar between strains, so in the text we report primarily the H1N1pdm09 results (circulating in 2011 & 2013); a summary of H3N2 (2014 & 2016) appears in Table 1, and figures appear in S1 Text.
Initial susceptibility values in seasons of the same strain were highly correlated (e.g. correlation coefficient 0.90 between initial susceptibility in 2011 and 2013). The posterior densities had similar forms between strains, however the H3N2 seasons (2014 & 2016) had slightly lower median R ¯ 0, associated with slightly higher initial susceptibility (Tables 1 & 2). For H3N2 seasons, the maximum bivariate posterior density occurred at R ¯ 0 = 4 . 71 and initial susceptibility 0.22 (in 2014) (S1 Text).
An understanding of seasonal influenza transmission is of critical public health concern, however effective inference can be constrained by the availability of high-quality data. In this study, we have used confirmed, multi-year virological data from a primary care surveillance network to perform model parameterisation for seasonal influenza in New South Wales, Australia. Making minimal assumptions around parameter values, we observe strong posterior relationships between R0 and population susceptibility, with higher R0 values corresponding to lower levels of susceptibility to the circulating strain. These results emphasise the challenges of identifying parameters around susceptibility, transmission, and observation, and highlight the substantial uncertainty that remains in these areas. We encourage caution when making modeling assumptions, choosing priors, and interpreting results, particularly when parameters are related.
The point that R0 is not clearly identified in influenza data has been made previously [7, 8], observing that many studies report R0 while actually estimating Re = R0S0, where S0 is the proportion of the population that is initially susceptible. Our results provide an additional line of evidence to reinforce this idea, with the majority of accepted parameter sets having parameters and model states such that R0 times the proportion of individuals that are susceptible at the start of a season lies between 1 and 2, whereas actual R0 values are generally much higher. The maximum posterior estimates when both R0 and susceptibility were considered had R ¯ 0 = 5 . 81 (for H1N1pdm2009 seasons), however there was substantial uncertainty in this estimate, with high posterior density across the range 2–10. This is consistent with previously calculated R0 values for historic pandemic influenza seasons or within populations that were likely to have high levels of initial susceptibility, such as on the island of Tristan da Cunha in 1971, with R0 in the range 3.73–10.69 [37]. The highest extreme of our posterior estimates for R0 exceed these estimates (in association with very low population susceptibility), and may be so high as to be implausible outside of unique outbreak scenarios. It is likely that with more data, posterior estimates would become more precise, however given the data available these extreme values cannot be discarded. We also emphasise that in epidemic studies marginal parameter estimates should be considered with caution, as maximum posterior density estimates from marginal distributions are very different to those from bivariate posterior distributions, given the strong nonlinear relationships between key parameters. This is illustrated clearly by comparing Figs 4 and 5.
While there may be cases where only the effective reproduction number is important, it is, in our view, generally inappropriate to model a seasonal virus that has underlying immunity within a population without specifically considering that immunity. Doing so could create bias in predictions or model outputs, and, in the worst case, detrimentally affect public health outcomes if or when incorrect parameter estimates are used to inform resource allocation or vaccination strategies. For example, it may be more effective to target vaccination resources at individuals or households (or geopolitical regions) with higher likelihoods of susceptibility based on infection history in previous years. This likely differs from the optimal strategies if underlying immunity is ignored. Forecasting in the presence of mutation or antigenic drift is also of concern, with model assumptions likely to lead to vastly differing predictions particularly when made based on limited data at the start of a potential pandemic. Moreover, confusion over estimates may lead to problems when making comparisons between diseases. In every case, researchers should critically consider the modeling assumptions being made and the impact that they may have on parameter estimates and subsequent analyses.
Observing potentially high transmissibility alongside lower susceptibility for seasonal influenza supports the idea that population susceptibility, rather than increased transmission, is the primary factor differentiating influenza strains with pandemic potential at a particular time from strains exhibiting routine seasonal circulation. The H1N1pdm09 strain that emerged and caused a pandemic in 2009 provides a specific example: when we fit parameters to this same strain in 2011 and 2013, it had relatively low levels of population susceptibility, and therefore routine, seasonal epidemics. It is likely that the difference between the circulation of the strain in 2009 and later years is due to differences in population susceptibility: because there were no similar strains circulating prior to 2009, a large proportion of the population (and specifically, anyone born after the most recent pandemic of a similar strain, in the late 1970s) was effectively naive, or only protected by cross-immunity. A confounding observation is that many estimates of Re during pandemic seasons have fallen in the range 1.0-2.0 [24]; whereas we may expect these estimates to by higher given increased population susceptibility. There are myriad reasons which could contribute to lower than expected estimates of transmissibility in a pandemic. The most critical factor is that, while population susceptibility is increased relative to seasonal circulation, there remains some level of immunity within populations, particularly in older age cohorts, either due to previous exposure to similar strains or cross-immunity with other strains [8, 15, 38, 39]. Additionally, studies which estimated Re for the 2009 pandemic highlighted possible impacts of factors including: public-health containment measures or changes in public behaviour due to pandemic awareness [38, 40, 41, 42]; outbreak timing coincident with school vacations [40, 43]; climatic conditions (with the initial outbreak occurring during summer in the Northern Hemisphere [42, 43]); or challenges in data collection (including sampling associated specifically with public-health programs, and potential variation in case ascertainment across age groups [39, 40]). It is also important to note that novel strains (e.g., transmitted from animal hosts) require the capacity to spread in humans, in addition to novelty and sufficient population susceptibility, before they can induce a pandemic [44].
We acknowledge that our results, particularly surrounding high R0 estimates in the bivariate posterior distribution, are provocative, and may challenge conventional wisdom. While we have endeavoured to take a principled approach that makes as few assumptions as possible, there are aspects of the study with the potential to introduce bias, and there remains substantial challenges in identifiability between parameters, particularly R0, the proportion of individuals that are susceptible, and the probability an individual seeks treatment. Specific modelling choices that could contribute to higher-than-expected R0 values are discussed in subsequent paragraphs. These include:
It is difficult to be certain of the impact of particular assumptions on parameter estimates, given the complex interactions between parameters. It is possible that, combined, these modelling assumptions could have contributed to higher-than-expected posterior estimates of R0. Ultimately, there remains substantial uncertainty around parameter estimates, with wide credible intervals and correlated posterior distributions consistent with the challenges of identifiability for R0, susceptibility, and treatment seeking behaviour. The present study does not claim to provide a definitive answer, rather it should inspire further research with high-quality datasets and careful analysis that at least considers a full range of plausible priors, and that considers interactions between parameters rather than marginal estimates.
It is important to emphasise that throughout this study we chose to use general, uninformative priors, in an effort to avoid imposing any strong assumptions around parameter values. A reasonable alternative approach to this would be to use information from the literature to impose informative priors, such as by reducing the prior density for larger values of R0, or making assumptions around population susceptibility. It is possible that informative priors could shift the resulting posterior distributions, and even impact the way that the study is interpreted. We emphasise the importance of carefully considering the priors used for this type of modelling study, and in interrogating how these choices might impact modelling outcomes.
A key limitation of this study is that we considered each of the study years to have only a single circulating influenza strain, which disregards the low levels of circulation of other strains that did occur. The impact of the circulation of these other strains, even at low levels, could potentially bias parameter estimates. Constructing distinct models for the different strains means that we could not use these models to understand cross-immunity between strains. Models that consider multiple strains and their interactions, or asymptomatic infections, would be ideal, but parametrizing these more realistic models is likely to be very challenging without both a better epidemiological understanding of how immunity and cross-immunity between strains works, and high resolution data including these strains. A further critical limitation of this work is that we did not consider population turnover or age structure. Age structure plays a key role in immunity, treatment-seeking behaviour, and transmission between individuals. By not including age structure, we necessarily assume that individuals mix homogenenously with age, that immunity is uniform across ages and wanes consistently, and that vaccination coverage is uniform across age groups: all of which are known to be incorrect assumptions. Future work should seek to include age structure in models such as this, while seeking to ensure that the available data and computation resources are sufficient to effectively fit additional parameters. It is likely that to produce detailed models with cross-immunity and age structure, a detailed longitudinal cohort study would be required, including regular sampling of multiple types (e.g., serology and virology) and potentially movement or contact tracing. Existing studies such as the Fluscape cohort study [45] will provide a starting point, however more regular sampling regimes may be necessary to elucidate the most complex dynamics.
While high-quality data and modern parametrization methods are helpful, there are limits to what may be reasonably determined using highly filtered weekly data. We are not able to effectively discriminate infectious durations, nor latent periods, and credible intervals on many parameters (e.g., the probability of seeking treatment) were quite broad. It is conceivable that this may be possible with daily data, additional testing, or by incorporating other data sources.
It is important to emphasise that the modelling in this study directly considers confirmed influenza. Such data are not typically directly available, particularly with denominator data. Were it more common to report negative test results (or, test results positive for other respiratory viruses), then alternate data sources of confirmed influenza, such as national notifications databases, would be more valuable and provide higher resolution with which to investigate seasonal influenza.
Seasonal influenza can be a challenging process to characterize given the complex nature of immunity and the inability to discriminate influenza from other respiratory viruses without testing. In this study, we used confirmed virological influenza data, with known denominators, to ensure that the model is specifically based on true influenza cases and the hierarchical observation process is accurate, and modern Bayesian inference techniques, with uninformative, unbounded priors, to ensure that assumptions around these priors did not impact on the resulting parameters estimates. We identified strong posterior relationships between R0, population susceptibility, and the probability an infected individual seeks treatment. The bivariate posterior distribution had maximum density where relatively high R0 values correspond with low levels of population susceptibility (while the effective reproduction number Re remains within the expected range). This was in sharp contrast to the values that would be obtained by considering only the marginal posterior distribution of each parameter. This highlights the importance of carefully considering the challenges of identifiability in parameter estimates, and in particular the importance of considering immunity alongside transmissibility in modelling studies. While there remains uncertainty around parameter estimates, we encourage researchers to consider carefully the challenges of identifying related parameters, and the impact that strong prior and modeling assumptions can have on parameter estimates.
Data were obtained as part of the ASPREN project ([31]; https://aspren.dmac.adelaide.edu.au/), a database of general practitioners (GPs) distributed throughout Australia. The target coverage is approximately one GP per 200,000 people in metropolitan areas, and one GP per 50,000 people in regional areas. This database is continually updated and requires the voluntary participation of GPs, so there is variation between years and states. All GPs in the database reported to ASPREN the number of consultations they performed each week, along with every case of influenza-like-illness (ILI) that they observed. GPs were asked to take swab samples from a proportion of those patients presenting ILI symptoms—25% from 2012–2014, 20% in 2015–2016. In practice, not all GPs took samples, and the proportion of patients for which samples were taken varied substantially, however we use exact denominators accounting for the testing doctors and the number of ILI patients they observed. These swab samples were sent for PCR testing, resulting in respiratory virus virology data for each sample. In this study, we used those data from New South Wales, as this is the most populous state in Australia and has the most confirmed influenza cases.
Because influenza has multiple strains, interactions between those strains are complex, and immunity from one strain may not confer protection to a different strain, we chose to model seasons with different strains separately. Specifically, the predominant strain in 2011 and 2013 was H1N1pdm2009, and the predominant strain in 2014 and 2016 was H3N2, so we performed model fitting separately on these two datasets. In doing so, we assume that epidemic parameters (e.g., R0) were consistent across seasons for a single strain, and we explicitly fit the proportion of the population susceptible and number infected at the start of each season. In 2012 and 2015 there were multiple strains with substantial circulation, and so we chose not to fit these seasons rather than making strong assumptions around interactions between strains.
Historical climate data were obtained from the Australian Government Bureau of Meteorology (www.bom.gov.au/climate/data), at station 066062 (Sydney—Observatory Hill). We used climate data from Sydney, the state capital, as both the majority of the population and the majority of GPs in the study were located here. Climate data consisted of observations of temperature, precipitation, and absolute humidity, taken every three hours (i.e., 8 times per day), continuously from 1955–present. Note that three-hourly observations creates a cyclic temperature pattern with both daily and annual periods. So as to avoid having unrealistic diurnal dynamics, we used the average daily temperature rather than the three-hourly observations themselves. We used a fixed, total population size of 6.9 million, based on 2011 Australian Bureau of Statistics estimates. While the population of New South Wales is increasing over time, including demography would add considerable complexity, which we wished to avoid.
The critical concern when using these data is an understanding of how underlying population-wide influenza dynamics translate to observed, confirmed influenza cases in the available dataset. We propose that filtering occurs at three levels:
There may well be variation from these fixed parameters in practice. However, if all parameters were allowed to vary freely then it would not be possible to identify them, as only the result of all three levels of filtering is observed.
We considered a continuous-time stochastic epidemic model, with an SEIRS structure, with an observation class and Erlang-2 sojurn times in the infectious class (modelled by splitting these classes into two consecutive compartments) (Fig 3). This process was approximated in discrete time with 8 timesteps per day. So as to avoid diurnal dynamics with transmission rates varying within days and being maximised in the early hours of the morning, we used the daily mean temperature at each timestep. We also allow external importation of cases, at a per-susceptible rate ξ (constant through time) which we fit as part of the model. Note that because of the structure of the model, the mean infectious time of an individual is 2/(2γ + λ) (see S1 Text). The model was parameterized based on physical epidemiological quantities, which were then transformed into model-based rates, which appear in Table 3.
Given these transformed parameters, transitions for the SEIR model are as appears in Table 4, using the increments that appear in Table 5.
We chose this parameterisation so as to put priors on the physical quantities of interest. That is, priors were defined on R0, mean latent, infectious and immune durations, and the probability of observation over the infectious period, and then transformed to obtain epidemic model parameters such as β and γ. Initial conditions were fit for each season, so that susceptibility to the circulating strain in that season could be determined. Prior distributions for epidemic parameters appear in Table 1. Per-susceptible external importation of cases at a low rate was used to allow for the possibility of epidemic fade-out and reintroduction from elsewhere, and was assigned an exponential prior distribution with rate 109 (approximately 3 individual case importations per year per million susceptible individuals). We did not reject simulation runs in which fade-out occurred, except to require circulation of influenza in any week in which there were influenza cases observed.
We chose to use exponential prior distributions so as to avoid putting maximum limits on the values these parameters may take, i.e., in the same sense that you would choose uniform priors to be uninformative, but with infinite positive support. Rates for the infectious and latent period exponential priors were chosen so that the means were within a reasonable range of influenza epidemiological parameters. The prior for R ¯ 0 was initially chosen as Exponential with large mean so as to have a relatively flat density; we truncated this prior at the highest published estimate of R0 for influenza, 22.7 [25]. While arguments could be made for a more informative prior with further reduced density at higher R0 values, we chose to retain this relatively uninformative prior so as to minimise modelling assumptions. The probability that an individual chooses to seek treatment was assumed to be constant, both within and between seasons. In practice this may not be the case, particularly when a circulating strain is unusually transmissible, or produces stronger symptoms (in either case, potentially leading to greater media attention). However, the hierarchical observation structure means that we are able to capture some of this effect in the variation in testing probabilities, as we explicitly use the actual testing proportions given in our dataset, which varied each week.
Note that we incorporate vaccination into this model by removing, deterministically, a proportion of susceptibles to a separate vaccinated class (i.e., they are still counted as part of N but otherwise do not interact with the model dynamics). Specifically we remove 21% of the population, based on published vaccination rates and efficacy statistics in Australia [46, 47]. We chose to do this rather than take vaccination rate as a fitted parameter so as to minimize the number of parameters to be fit. Consequently, we set the prior for initial season susceptibility to be uniformly distributed from 0 to 0.75. We used temperature as the covariate with which to enable seasonal forcing, following the relationship R 0 = R ¯ 0 + a ( T - T ¯ ) (with T the daily mean temperature and T ¯ the mean temperature over the study period). A number of studies have considered a variety of climate covariates and their interaction with influenza transmission both in populations and experimental models (e.g., [48, 49]); with indications that relationships with and between climate factors are complex and may vary globally [50]. We noted similar quality model fits when using the specific humidity formulation presented by Shaman et al. [51], however we chose to use the simplest seasonal link, temperature, as this form produced similar quality fits while requiring fewer parameters and enabling greater ease of interpretation.
We used approximate Bayesian computation [52, 53] to calculate posterior distributions for the model parameters. Specifically, we generated candidate parameter sets from the prior distributions listed above. For a given candidate parameter set, we simulated a realisation of the model, and observed the sample path (i.e., the number of infected individuals attending ASPREN doctors each week in the simulated realisation) for 2011 and 2013 (H1 seasons), or 2014 and 2016 (H3 seasons). We compared this simulated realisation to the true sample path, using the square root of mean squared error as score function
D : = 1 # weeks ∑ i = 1 # weeks ( true i - candidate i ) 2 ,
with truei the observed number of cases in the ith week from the ASPREN dataset and candidatei the number in the simulated realisation from the candidate parameter set. There were 104 weeks in each of the study periods (# weeks). The parameter set is accepted if D is less than some tolerance level, in this case set to 4.5. We observe that this choice of score function and threshold produces reasonable model fits, while still being high enough to accept parameter sets regularly. Choice of score function in ABC is problem-specific [54], and it is likely that a variety of other metrics could reasonably be used to produce reasonable fits. Note that a candidate parameter set can produce a range of scores given that realisations are stochastic, and the accepted realisations are generally among the best possible realisations from the given parameter set (i.e., those that most closely fit the data), however, the average realisation from these parameter sets still fit the data relatively well (S1 Text).
Kernel density estimates were constructed from posterior samples using the algorithm of Botev et al. [55].
|
10.1371/journal.pntd.0006999 | Robustness of the reproductive number estimates in vector-borne disease systems | The required efforts, feasibility and predicted success of an intervention strategy against an infectious disease are partially determined by its basic reproduction number, R0. In its simplest form R0 can be understood as the product of the infectious period, the number of infectious contacts and the per-contact transmission probability, which in the case of vector-transmitted diseases necessarily extend to the vector stages. As vectors do not usually recover from infection, they remain infectious for life, which places high significance on the vector’s life expectancy. Current methods for estimating the R0 for a vector-borne disease are mostly derived from compartmental modelling frameworks assuming constant vector mortality rates. We hypothesised that some of the assumptions underlying these models can lead to unrealistic high vector life expectancies with important repercussions for R0 estimates.
Here we used a stochastic, individual-based model which allowed us to directly measure the number of secondary infections arising from one index case under different assumptions about vector mortality. Our results confirm that formulas based on age-independent mortality rates can overestimate R0 by nearly 100% compared to our own estimate derived from first principles. We further provide a correction factor that can be used with a standard R0 formula and adjusts for the discrepancies due to erroneous vector age distributions.
Vector mortality rates play a crucial role for the success and general epidemiology of vector-transmitted diseases. Many modelling efforts intrinsically assume these to be age-independent, which, as clearly demonstrated here, can lead to severe over-estimation of the disease’s reproduction number. Our results thus re-emphasise the importance of obtaining field-relevant and species-dependent vector mortality rates, which in turn would facilitate more realistic intervention impact predictions.
| Many infectious diseases of public health concern, such as dengue, Zika and malaria, are transmitted by insect vectors. Control effort required to curb their continued spread or even prevent their establishment in the first place are partially determined by the disease’s basic reproduction number, R0. Of particular importance for estimating R0 is the duration at which the vector can transmit the disease, which is limited by its life expectancy. Many R0 estimation methods are based on mathematical frameworks that assume constant vector mortality rates. Here we demonstrate how the resulting exponential distribution in life-expectancy can lead to significant over-estimations. By means of an individual-based model we elucidate the effect of vector mortality on R0 estimates and derive a correction factor that alleviates some of the discrepancies due to erroneous vector age distributions. Our results clearly demonstrate the need to obtain more realistic, i.e. field-relevant vector mortality rates in order to generate robust estimates of a disease’s reproduction number and provide guidance for setting-specific disease control.
| Over the last few decades there has been a global rise in the emergence and re-emergence of vector-borne infectious diseases [1]. The continuing threat of Plasmodium falciparum malaria [2, 3] and dengue [4], the rapid, near pandemic spread of Zika virus [5] or the recent epizootic outbreak of Yersinia pestis (plague) in Madagascar [6] are just some examples of pathogens transmitted by insect vectors that pose a major threat to global public health. Their dependence on insects for transmission between vertebrate hosts has a number of important implications. First, they are frequently subject to strong spatial and temporal fluctuations due to environmental and climatic variations, such as seasonality in rainfall or temperature. Second, these pathogens should be amenable to vector control. That is, disease transmission can, at least in theory, be interrupted simply by removing the insect vector (e.g. use of insecticides) or by preventing contact between the vector and the host (e.g. use of bednets). Furthermore, it has been suggested that only a fraction of insects need to be removed or vector-host contacts to be prevented for the disease to die out. This concept is largely based on mathematical theory that can be traced back to the first formal description and mathematical treatment of the malaria life-cycle by Ross [7]. Unfortunately, translating theoretical predictions to practical applications, especially with regards to disease elimination through vector control, has only resulted in partial success.
The epidemiological reasoning behind the theory relies on a particular threshold condition involving the so-called basic reproduction number, R0, which denotes the expected number of secondary cases arising from a single infection in a totally susceptible population [8]. To date, R0 is frequently used either to predict the extent of an epidemic outbreak or to derive the necessary conditions to prevent this outbreak from happening, e.g. by means of vaccination. The crux of the problem is how to robustly derive or estimate this number in the first place. Compartmentalised systems of ordinary differential equations (ODEs) have been in use for decades to understand infectious diseases at the population level and provide the backbone for most formulas for R0 [9]. These allow the reproduction number to be computed either exclusively using empirically informed parameter estimates or from the initial growth rate of an outbreak [10]. Although the latter is the more common approximation method for directly-transmitted disease [11–14], it has equally been applied to vector-borne pathogens [15–17].
An important consideration for R0 estimates of vector-borne pathogens is that these can vary substantially across space and time. For example, reported R0 estimates for the complete transmission cycle of Plasmodium falciparum in Africa range from 1 to more than 3,000 [18, 19]. Based on nine epidemics in Brazil between 1996 and 2003, the reproduction number for dengue has been estimated to be somewhere between 2 to 103 [20], and median estimates for Zika range between 2.6–4.8 in French Polynesia [21] and 4–9 in Rio de Janiero [22]. The reasons for such wide variations are manifold. As mentioned earlier, the dependence on insect vectors for transmission can naturally introduce large spatio-temporal heterogeneities. That is, a disease introduced during the dry season will behave very differently to the same disease being introduced during the rainy seasons. Equally, an outbreak in a densely populated urban area will likely take a different course than an outbreak in a sparsely populated rural area. Here we argue that in addition to these natural variations and potential differences in data collection and analyses, the actual methodologies used to derive R0 estimates can also introduce substantial discrepancies.
A crucial component of the reproduction number for a vector-borne disease is the mean time that an infected vector is able to transmit to a host, or the infectious vector-to-host transmission period (VHTP) [23]. As infectious vectors are assumed to continue to transmit the disease until death, the VHTP is determined both by the life expectancy of the vector and the extrinsic incubation period of the pathogen. For mathematical simplicity, most epidemiological models of vector-borne diseases assume that vectors have a constant (daily) mortality rate. However, this assumption is in stark contrast to findings from lab-based and field mark-and-recapture studies. For example, survival probabilities of the dengue mosquito vector Aedes aegypti and the principal malaria vectors Anopheles stephensi and An. gambia have been shown to be strongly age-dependent [24–27]. Although it should be clear that current lab and field-based studies of vector survivorship come with their own set of limitations and uncertainties, constant, i.e. age-independent mortality rates are biologically less likely than assuming a general decrease in the survival probability with age.
Previous work has looked into the effects of logistic mortality rates on the vectorial capacity, the mosquito-related components of R0 [28]. However, the effects of assuming constant vector mortality on R0 in a system where death rates are strongly age-dependent have not yet been explored. Here we compared a commonly used R0 formula based on continuous-time differential equation model using constant mortality rates to an R0 estimate derived from first principles under relaxed assumptions about vector mortality. Using a stochastic, individual-based simulation model (IBM), which permits the direct measurement of the average number of secondary cases, we demonstrate how the underlying assumptions of vector survivorship can significantly inflate R0 estimates. We further show how estimates based on endemic equilibria are generally more robust and derive a correction factor to ameliorate R0-inflations in estimation methods based on epidemic growth curves.
We derived R0 estimates from two different epidemiological frameworks: (i) a simple, single-strain vector-borne disease model based on ordinary differential equations (ODE), where vector mortality is assumed to be constant, leading to an exponentially distributed vector survivorship, and (ii) a stochastic individual-based model (IBM), which permits more explicit control over the demographic processes regulating birth and death rates.
Table 1 provides an overview of the parameters and parameter values used throughout this work (unless stated otherwise). The values were chosen to reflect the epidemiological dynamics of an arboviral disease, such as dengue or Zika. However, the results presented here are qualitatively independent of the particular choice of parameters; S2 Fig show the results of model sensitivity analyses with respect to the dependency of R0 estimates on particular parameter values.
A multitude of the mathematical models put forward to study the dynamics of vector-borne diseases are based on compartmental models described by systems of ordinary differential equations (ODE). Crucial to these types of models is the assumption of constant death rates. As vectors are assumed to remain infectious for life, such assumptions influence not only the resulting dynamics but also the estimates of the disease’s basic reproduction number R0 (and relatedly the (time-varying) effective reproduction number Re(t)). Here we aimed to quantify the effects of relaxing the assumption of constant vector death rate on R0 estimates within the same theoretical setting. This was done by comparing the R0 values derived from an SEIR-SEI system of ODEs with a formula derived from first principles using the transmission cycle of a generic vector-borne disease using different assumptions about vector mortality rates (see Methods). We then verified these estimates by means of a stochastic individual-based model, which allowed us to directly measure R0 from running repeat simulations of introducing an infected individuals into a fully susceptible population. The same model was also used to derive R0 estimates from simulated timeseries data.
Assuming constant vector mortality rates leads to exponentially distributed age profiles (Fig 1B), which permit some vectors to live in excess of four times their mean life expectancy and potentially to transmit the pathogen for an unusually long period of time. Even more concerning is that the vector life expectancy in each compartment of the infection process is essentially the same (Fig 1A). That is, independent of when in its life a vector becomes infected and infectious, its remaining life expectancy remains exponentially distributed around the mean life expectancy. As a consequence, all infectious vectors have a vector-to-human transmission period (VHTP) equal to the mean life expectancy of all vectors, 1/μV, with obvious consequences for R0 estimates.
In contrast to ODE models, individual-based models (IBM) permit much greater control over vector mortality rates. Here we used (Weibull distributed) age-dependent vector death rates (see Methods), which yield a range of sigmoid age profiles (Fig 1D) but which all prevent vectors from living severely extended lives. More importantly, an individual vector’s remaining life expectancy remains unchanged when transitioning between susceptible and infected state or between infected and infectious state, resulting in shorter and more realistic infectious periods.
We demonstrate the effect of assuming different vector mortality rates by comparing the R0 estimates derived from the ODE model to the individual-based model (see Methods). As expected, using parameters as listed in Table 1 we find that the reproduction numbers from the ODE and IBM systems are similar under the assumption of constant vector mortality rates (Table 2). The small discrepancy between the two models is due to the ODE model’s assumption of an exponentially distributed extrinsic incubation period, whereas the IBM assumes this to be a fixed length of time. Using the IBM approach to track individual mosquitoes and infection events we also find that under this assumption the mean age at which vectors become infected is 20 days and infectious at an age of 25 days, i.e. days beyond their average life-expectancy. Furthermore, those vectors that have become infectious live for an average of 46 days, which means that their infectious period is 21 days (equal to the life expectancy of all vectors). This clearly highlights the discrepancy between model outputs based on constant mortality rates and biological reality. In contrast, assuming age-dependent mortality rates (Weibull shape parameter, cV = 4) results in biologically more reasonable infectious periods of 11 days (Table 2) and R0 estimates that are less than half of those based on a model with constant mortality.
To further demonstrate the dependency of R0 on different distributions of mosquito survivorship, we changed the Weibull distribution of vector mortality to transition smoothly between an exponential (cV = 1) and a sigmoid (cV > 1) age profiles and by keeping the average life expectancy constant. As illustrated in Fig 2, relaxing the assumption of constant mortality and resultant exponential age profile shortens the average infectious period and lowers the reproduction number as derived from the transmission cycle of the pathogen, i.e. R 0 IBM. This clearly demonstrate that as well as the vector life expectancy, the actual shape of the survival curve strongly determines the estimated values of a pathogen’s reproduction number.
The scenario defined by the reproduction number, whereby a single infectious case enters an entirely susceptible population, is arguably unrealistic for most diseases. Furthermore, disease transmission is an inherently stochastic process, such that each realisation of a disease introduction event is likely to take a different course. We should therefore expect that R0 estimates derived from such introductory events should come with a certain degree of variation. In order to better understand the variability of the expected number of secondary cases and then to directly compare the above formula-based R0 estimates, we simulated disease introduction events into a completely susceptible population using our IBM framework and kept track of all secondary host infections resulting from the index case.
As before we compared the two different assumptions regarding vector life expectancy: constant vs. age-dependent mortality rates. As shown in Fig 3, there is a wide distribution in the number of secondary infections, particularly when we assumed constant vector death rates (Fig 3A and 3B). In that case it was not unusual to observe 40-60 secondary infections, due to the aforementioned unrealistically high life-expectancies for some of the vectors, permitting the accumulation of secondary cases well after the primary human case has recovered (Fig 3A). The mean number of secondary infection (i.e. R0) across 500 model simulations was around 7, more than twice that of the model which assumed age-dependent mortalities. In the latter case we observed secondary infections in the range of 0 to 18 (due to the model’s stochastic nature where some vectors may be infected for their entire life) and with a mean of around 3.2 (Fig 3C and 3D), in line with theoretical expectations. Please refer to S3 Fig for sensitivity on model parameters on the direct measurement of mean secondary infections from the IBM.
An interesting observation is that under both assumptions of vector mortality, over 30% of our simulations resulted in zero secondary infections, as either the single primary case did not infect any vectors, the infected vectors failed to survive the extrinsic incubation period, or the infectious vectors failed to transmit the pathogen. Shorter infectious periods for both the host and the vector, a longer extrinsic incubation period, and lower transmissibility naturally decrease the overall likelihood of transmission from primary to secondary cases. Therefore, the proportion of failed outbreaks crucially depends on all these factors (S4 Fig).
In most cases, only successful disease introductions that lead to epidemic outbreaks are observed. These outbreaks can then be used to estimate the reproduction number based on the initial epidemic growth rate λ (see Methods). Formulas to calculate R0 from λ are usually based on ODE modelling frameworks assuming constant vector death rates. To investigate the effect of this assumption on estimating a disease’s basic reproduction number from epidemic growth rates, R 0 λ, we used our IBM framework to generate 100 epidemic outbreaks (discounting failed introductory events) under identical initial conditions for both constant and age-dependent mortality rates.
As illustrated in Fig 4A, estimating the reproduction number from initial outbreak data is fairly reliable as long as the empirical age profiles of the mosquitoes match the one assumed in the model. That is, if mosquito mortality was indeed independent of age, leading to exponentially distributed age profiles, then R 0 λ can provide good estimates of the real reproduction number. However, if the risk of dying does increase with age, then R 0 λ, as derived form the ODE framework, is once again significantly over-inflated. Likewise, assuming age-dependent death rates when mortality is in fact constant, this could lead to an underestimation of the true reproduction number; note, however, that the latter scenario is arguably less relevant in biological terms.
Noticeable in all situations is the considerable variance in R 0 λ. This is due to the stochastic nature of our spatial IBM framework, which to a certain extent should also reflect the natural stochasticities underlying real vector-borne disease systems. Changing the model’s spatial and demographic set-up will obviously affect the variance reported here; however, the results, related to the mean values, are to be understood as independent of the model’s underlying structure.
As shown in Fig 4, using the initial epidemic growth rate is only appropriate when empirical vector mortality is indeed age-independent, whereas it can lead to significant over-estimations otherwise. In order to compensate for this and include age-dependent vector mortality rates into the ODE-derived formula for R 0 λ, we replaced this critical term by the vector to host transmission period (VHTP), denoted by ν I V, calculated directly from an assumed vector age profile (see Methods), which yields the corrected estimate
R ^ 0 λ= ( 1 + λ ν I V ) ( μ V μ V + λ ) R 0 λ (14)
where μV is the constant vector mortality rate in the classical system of ordinary differential equations. Crucially, a vector age profile has to be assumed explicitly to calculate the VHTP. And as before, if the assumed profile in R ^ 0 λ matches the simulations’ profile, we find that the derived reproduction numbers are good estimates of the actual ones, with the same variance as before (Fig 4).
Finally, we sought to estimate R0 from the dynamic equilibrium distribution of susceptibles in the human population (see Methods). Crucially, this approach does not require any a priori knowledge of mosquito survivorship and should therefore provide more robust estimates regardless of the underlying assumptions regarding vector mortality rates. Indeed, and as demonstrated in Fig 5, using the endemic state can provide reasonable estimates of a disease’s true (i.e. theoretical) R0 value, even though the formula itself was derived from a directly transmittable disease, which might explain why R 0 * slightly underestimates R0.
As before we find a significant degree of variation around the mean estimates, due to the stochastic nature of disease transmission. This can somewhat be reduced by taken longer term averages (compare single time point estimates with 10 year average in Fig 5), which in reality will be limited due to data availability. Equally, the model and population structure itself, including as population size, importation rates, spatial structuring and mixing, all affect the stability of the dynamic equilibrium and with it the variance and hence robustness of R 0 * (see S2 Fig). Although this method is only applicable for diseases that have reached at least a semi-endemic state, its parameter and assumption-free approach means that it should be considered as one of the most robust ways to estimate a disease’s reproduction number.
Mathematical models describing the population dynamics of an infectious disease provide the necessary frameworks by which we can calculate an infectious disease’s reproduction number, R0, based on specific parameters related to infection and transmission probabilities. One of the most important factors influencing R0 is the length at which an individual remains infectious. For vector-transmitted diseases this places huge significance on vector mortality rates as vectors usually do not clear an infection and instead remain infectious for life. Many formulas to estimate R0 are based on systems of ordinary differential equations (ODEs), which commonly assume that vector mortality is constant, i.e. independent of age. As we have demonstrated here, the resulting exponential distribution and the effective resetting of life expectancies as individuals transition through the infection stages permit some vectors to live for an extraordinary length of time. As a result, vectors are potentially able to transmit the disease multiple times that of what should biologically be possible, leading to significantly inflated R0 estimates.
In comparison to ODE models, individual-based models (IBMs) provide much greater control over the dynamics that govern both demography and disease transmission. Here we used an individual-based modelling approach to elucidate the influence of vector mortality on R0 estimates and to highlight the discrepancy between model predictions based on constant vs. age-dependent mortality. Because individual infection events can easily be tracked within an IBM, the basic reproduction number can essentially be measured simply by counting the number of secondary infections arising from a single index case. This in turn not only allowed us to compare different formulas for estimating R0 but also provided us with a better understanding of the degree of uncertainty surrounding these estimates.
As demonstrated here, the assumption of constant vector death rates can lead to significant over-estimation of R0. Importantly, it is not so much that the formulas commonly used to estimate R0 are inherently wrong but rather that the underlying assumption of the models from which they are derived are not necessarily aligned with biological reality. We found that one of the most robust methods to estimate a pathogen’s R0 is based on the proportion of susceptible individuals at endemic equilibrium, as this is entirely parameter free and does not require any assumption about vector death rates. Unfortunately, this only works for diseases that are well established in a population, and its reliability is strongly dependent on the stochasticity of the underlying endemic equilibrium, i.e. the (multi-annual) variations around the mean. For emerging diseases this is obviously not practical and estimation methods in those cases usually make use of epidemic growth curves instead. However, these also implicitly assume exponential vector age profiles and are therefore subject to inflation. In order to account for this we have here derived a correction factor that can be applied to classical R0 estimation formulas and which adjusts for most of the discrepancy between the vector-to-human transmission period (VHTP) of the biological system and the assumed system with constant vector mortality.
In this work we made use of an individual-based modelling framework to test the effect of non-exponential vector age distributions on R0. Alternative methods that allow for the (partial) relaxation of the assumption regarding constant mortality or vector senescence have also been proposed, including lumped-age class models [32] or systems of partial differential equations [33]. However, these methods can still suffer from the same issues as simpler ODE models, where transition rates between life and infection stages are usually exponentially distributed and where information about individual ages is lost at every transition stage. The ease at which different distributions that govern host and vector mortality, infection recovery and other epidemiological factors can be incorporated, make IBM frameworks the natural choice to examine the influence of vector mortality or other such factors on R0 estimations. Here we only concentrated on the effect of vector mortality, whereas similar arguments are equally valid for the distribution underlying the extrinsic incubation period [34], for example. Nevertheless, our work strongly suggests that vector mortality rates, or rather our assumptions about the age-dependency of survivorship, are the predominant factors, as our correction term for R0 estimates based on epidemic growth essentially recovers the true value.
Another important observation from this study was that when simulating the spread of a disease from a single infected individual and then calculating R0 based on the number of secondary infections, the stochastic nature of such events resulted in very wide distributions in R0. Although the assumption of age-dependent mortality rates generally prevented extremely high values of secondary cases, and therefore R0, the variance was still in the region of twice the mean and included a significant proportion of zero cases. That is, in around a third of the simulations we observed no secondary case at all despite starting off with the same initial conditions. This then begs the question whether these events should be counted towards the estimated R0 or not, as in reality we never observe such failed introductions. Comparing the expected with the observed R0 value would suggest that zero cases should be counted, which on the other hand implies that even high values of the reproduction number are by no means a guarantee that an outbreak should ensue if a disease gets introduced in a fully susceptible population (sufficient conditions to prevent stochastic fade-out at the start of an epidemic have been previously discussed [35]). The high variation also suggests that control strategies based on R0 estimates generated from initial growth rates should be treated with caution and that estimations based on one particular setting might not be adequate to generalize and predict pathogen behaviour across all other spatial contexts [36].
We here concentrated solely on the basic reproduction number, which describes an arguably unusual and often artificial situation. However, it should be clear that the same arguments also hold for the effective reproduction number, Re, which is essentially R0 multiplied by the fraction of the population that is susceptible to a disease, as well as their time-dependent counterparts R(t) and Re(t). Furthermore, the serial and generation intervals, which can be understood as temporal analogues of the reproduction number, also rely on the vector to host transmission period and are usually assumed to be exponentially distributed [37, 38]. This implies that these intervals, and alternative R0 estimation methods that depend upon them [39–41], may equally be over-estimated.
Our work thus reiterates the importance of obtaining empirical vector mortality rates in the field. The original Ross-MacDonald model for the spread of Plasmodium falciparum and P. vivax malaria assumed constant vector mortality as laboratory and field studies seemed to suggest that death rates were age independent [42]. However, re-analysis of laboratory data showed that mosquito mortality is in fact age-dependent for several Anopholes species [43]. More recent studies also confirmed that mosquito mortality is dependent on age for Anopheles mosquitoes [27] and Aedes aegypti [24, 25]. What is clear is that more work needs to be done to fully elucidate realistic, i.e. field-relevant vector mortality rates, perhaps with more accurate spectroscopic methods [44], as well as their environmental drivers. That is, seasonal variations in temperature and rainfall have been shown to affect the birth and death rate of vectors [45–48], the vectorial competence [49] as well as the extrinsic incubation period [50–53]. It has also been emphasized that other spatio-temporal heterogeneities, such as community structures and host and vector movement, should be considered when assessing R0 [54, 55]. All this needs to be factored in if we are to develop better models to understand the epidemiological and ecological determinants of vector-borne diseases, guide outbreak prevention strategies or monitor ongoing intervention measures.
|
10.1371/journal.pntd.0007014 | Development of sandwich ELISA and lateral flow strip assays for diagnosing clinically significant snakebite in Taiwan | Taiwan is an island located in the south Pacific, a subtropical region that is home to 61 species of snakes. Of these snakes, four species—Trimeresurus stejnegeri, Protobothrops mucrosquamatus, Bungarus multicinctus and Naja atra—account for more than 90% of clinical envenomation cases. Currently, there are two types of bivalent antivenom: hemorrhagic antivenom against the venom of T. stejnegeri and P. mucrosquamatus, and neurotoxic antivenom for treatment of envenomation by B. multicinctus and N. atra. However, no suitable detection kits are available to precisely guide physicians in the use of antivenoms. Here, we sought to develop diagnostic assays for improving the clinical management of snakebite in Taiwan. A two-step affinity purification procedure was used to generate neurotoxic species-specific antibodies (NSS-Abs) and hemorrhagic species-specific antibodies (HSS-Abs) from antivenoms. These two SSAbs were then used to develop a sandwich ELISA (enzyme-linked immunosorbent assay) and a lateral flow assay comprising two test lines. The resulting ELISAs and lateral flow strip assays could successfully discriminate between neurotoxic and hemorrhagic venoms. The limits of quantification (LOQ) of the ELISA for neurotoxic venoms and hemorrhagic venoms were determined to be 0.39 and 0.78 ng/ml, respectively, and the lateral flow strips were capable of detecting neurotoxic and hemorrhagic venoms at concentrations lower than 5 and 50 ng/ml, respectively, in 10–15 min. Tests of lateral flow strips in 21 clinical snakebite cases showed 100% specificity and 100% sensitivity for neurotoxic envenomation, whereas the sensitivity for detecting hemorrhagic envenomation samples was 36.4%. We herein presented a feasible strategy for developing a sensitive sandwich ELISA and lateral flow strip assay for detecting and differentiating venom proteins from hemorrhagic and neurotoxic snakes. A useful snakebite diagnostic guideline according to the lateral flow strip results and clinical symptoms was proposed to help physicians to use antivenoms appropriately. The two-test-line lateral flow strip assay could potentially be applied in an emergency room setting to help physicians diagnose and manage snakebite victims.
| Snakebite is a public health issue that causes life-threatening medical emergencies. Rapid diagnosis of snakebite in the clinic is a critical necessity in many tropical and subtropical countries, where various venomous snakes are common. Venoms from different snake species contain distinct protein components that require treatment with different antivenoms. However, given the similarity in clinical symptoms among some snake envenomations, it is often challenging for physicians to precisely define the snake species responsible for envenomation. Thus, a reliable method or assay for rapidly diagnosing envenoming species is urgently needed. Here, we present a two-step affinity purification procedure for generating species-specific antibodies (SSAbs) from antivenom, followed by the development of a sandwich ELISA (enzyme-linked immunosorbent assay) and lateral flow strip assay using these SSAbs. This feasible and cost-effective strategy allowed us to develop workable assays for distinguishing between venom proteins from hemorrhagic and neurotoxic snakes in Taiwan. The usefulness of this strategy was demonstrated in the clinic, where both diagnostic assays were shown capable of detecting venoms in blood samples from snakebite patients. Together with the observation of clinical symptoms, the two-test-line lateral flow strip assay is potentially applicable in an emergency room setting to improve snakebite diagnosis and management.
| Envenoming resulting from snakebites is a significant public health issue in many regions of the world, particularly in tropical and subtropical countries and some poor rural communities [1]. An estimated 1,800,000–2,700,000 envenoming cases and 81,410–137,880 associated deaths occur each year globally owing to snakebite [2]. The regions with the highest burden are South Asia, Southeast Asia, and Africa [2, 3].
Administration of antivenom is the standard treatment for snake envenomation. In most countries, multiple types of antivenom are clinically available, but uncertainty regarding the appropriate antivenom to use in any given situation remains an important issue. To date, the species responsible for envenomation of snakebite victims referred for medical treatment is initially identified primarily based on the shape of the wound or identification of dead snakes brought to the hospital. Thereafter, the physician monitors local symptoms to confirm which antivenom should be used. However, some clinical symptoms caused by envenomation are similar among species, and non-venomous snakes are often responsible for the patient’s snakebite [4]. Additionally, physicians are often misled by incorrect descriptions of the snake by victims or their family members [5]. Identification of venomous snake species is important for optimal clinical management, because it allows physicians to use the correct antivenom for effective treatment, thereby improving patients’ prognosis and preventing the waste of expensive antivenoms and exposing victims to antivenom-induced adverse reactions [6]. Although identification of snake species is important for the management of snakebite-related injuries worldwide, there are currently no developed standard platforms or guidelines for snakebite diagnosis globally.
Detection of venom proteins using antibodies is a simple and effective approach for identifying the species responsible for snakebite. To date, various immunoassays for detecting venom proteins in body fluids have been described [7–13], including radioimmunoassay [14], agglutination assays [9, 15], enzyme-linked immunosorbent assays (ELISAs) [10–12, 16, 17], and fluorescence immunoassays [18, 19]. In addition to immunoassays, immunology-based biosensors have been explored for detection of snakebite [20, 21]. ELISAs and lateral flow assays [22, 23] are arguably the best choice of immunoassays for snakebite identification. ELISAs, the most common and general immunoassays in clinical use, are sensitive to their target at pictogram per milliliter levels [18]. Although the antibodies are relatively costly, ELISA devices and reagents are affordable for routine diagnosis. Compared with ELISAs, lateral flow assays offer advantages in terms of detection time and required equipment: it takes only ~5–20 min to obtain assay results and no supporting instrumentation is needed [24, 25]. Although lateral flow assays mainly provide qualitative results, their simple design and operation compared with quantitative ELISAs make them the most user-friendly for the public, allowing rapid adoption in rural countries.
Snake venoms contain many proteins, and closely related snake species have some of the same or similar venom components, causing cross-reactions in immunoassays applied to detect venom proteins [11, 12, 26]. The venom antigens responsible for the observed cross-reactivity would further cause ambiguities and false-positive results in snake species detection [11, 27]. Hence, the direct use of polyclonal antibodies against whole venoms for snake species detection is inappropriate, and elimination of cross-reactive antibodies is critical for generating an immunoassay with high specificity for discriminating snake species [11, 12, 28]. Solving the problem of cross-reaction and improving the specificity of immunoassays might most efficiently be achieved through purification of species-specific antibodies (SSAbs) on affinity columns immobilized with venom proteins cross-reactive to the polyclonal antibodies or antisera [11, 12].
Six venomous snakes—Deinagkistrodon acutus, Trimeresurus stejnegeri, Protobothrops mucrosquamatus, Daboia russelii formosensis, Bungarus multicinctus and Naja atra—are indigenous to Taiwan, a subtropical island in East Asia [29]. Four kinds of antivenom had been produced by the Vaccine Center, Center for Disease Control, Taiwan, to treat envenomation by these six venomous snakes and effectively limit snakebite mortality [30]. Freeze-dried hemorrhagic antivenom (FHAV) is used to treat envenomation by T. stejnegeri and P. mucrosquamatus, and freeze-dried neurotoxic antivenom (FNAV) neutralizes venom of B. multicinctus and N. atra. Envenomation by the other two snake species is treated by monovalent antivenoms. A population-based study of venomous snakebites in Taiwan from 2005 to 2009 reported a total of 4647 snakebite cases, of which 380 (8.1%) received at least two types of antivenoms, mainly because of similarities in the clinical presentations of different snakebites and the inability of some patients to identify the culprit snake [31]. In some studies, such unidentified cases accounted for 12–45% of total cases [32–35]. In addition, according to a clinical survey of antivenom usage in Taiwan, more than 99% of snakebite patients that had received FHAV or FNAV treatment were rescued [36], indicating that most snakebite cases in Taiwan represent envenomation by T. stejnegeri, P. mucrosquamatus, B. multicinctus or N. atra. Unfortunately, there have been very few efforts to develop sensitive assays for detecting snake venom in Taiwan. Currently, only one ELISA-based blood assay has been developed to detect the N. atra venom, but it is not commercially available [7], and no laboratory test can be used to identify other types of venoms.
In the present study, we designed a workflow to develop immunoassays for snakebite detection based on clinical antivenom usage in Taiwan. We used FHAV and FNAV as resources for purification of hemorrhagic species-specific antibodies (HSS-Ab) and neurotoxic species-specific antibodies (NSS-Ab), and applied these two critical reagents to develop ELISAs and lateral flow strip assays. These assays hold the potential for use in identification of snake species responsible for snakebites in Taiwan.
Lyophilized venoms of T. stejnegeri, P. mucrosquamatus, B. multicinctus and N. atra were obtained from the Center for Disease Control, R.O.C (Taiwan). The venoms were collected from several adult specimens, then freeze-dried and stored at -20°C before use. Hemorrhagic venom (T. stejnegeri and P. mucrosquamatus)-immunized and neurotoxic venom (B. multicinctus and N. atra)-immunized horse plasma were also donated by the Center for Disease Control, R.O.C (Taiwan). The plasma was stored at -80°C before use.
For coupling of venom proteins onto Sepharose beads, CNBr-activated Sepharose 4B was swollen in 1.0 mM HCl (pH 3.0), then incubated with 10 mg hemorrhagic or neurotoxic snake venoms dissolved in coupling buffer (0.1 M NaHCO3 pH 8.3) overnight at 4°C on a round rotator. After washing with coupling buffer, any remaining active sites on beads were blocked by incubating overnight at 4°C with blocking buffer (1.0 M diethanolamine pH 8.0) on a rotator. The beads were then alternately washed three times with an acidic buffer (0.1 M C2H3NaO2 pH 4.0, 0.5 M NaCl) and basic buffer (0.1 M Tris pH 8.0, 0.5 M NaCl) and packed into a column, The resulting venom affinity columns were equilibrated with binding buffer (10 mM Tris-HCl pH 7.5) and stored at 4°C before use.
To purify HSS-Ab, 2 ml FHAV was diluted in 30 ml of binding buffer and the diluted sample was pumped into the neurotoxic venom affinity column at 4°C for 3 h. The flow-through fraction was then pumped into the hemorrhagic venom affinity column at 4°C for another 3 h. The hemorrhagic venom affinity column was then washed with 60 ml binding buffer and 60 ml wash buffer (10 mM Tris-HCl pH 7.5,0.5 M NaCl). After washing, each affinity column was eluted with 20 ml of acidic (100 mM glycine pH 2.5) or basic (100 mM triethylamine pH 11.5) elution buffer, and eluted fractions (1 ml/fraction) were collected into microcentrifuge tubes containing 100 μl of neutralized buffer (1.5 M Tris-HCl pH 8.0). Finally, all eluted fractions were pooled, concentrated, and exchanged into phosphate-buffered saline (PBS) by dialysis overnight. The concentrated antibodies in PBS were diluted with an equal volume of glycerol and stored at -20°C. Similar protocol was used to purify NSS-Ab from 2 ml FNAV, in which the diluted FNAV was passed through the hemorrhagic venom affinity column first, and the flow-through fraction containing NSS-Ab was further purified using the neurotoxic venom affinity column.
Snake venom proteins (100 ng) were diluted in 100 μl PBS and coated onto 96-well polystyrene microplates (Corning, USA) by incubating at 4°C overnight. The plates were washed six times with 200 μl of PBST (PBS contain 0.1% Tween-20) and blocked by incubating with 200 μl of 1% ovalbumin in PBS at room temperature for 2 h. After washing wells six times with PBST, antivenom or purified Ab (1 mg/ml) was serial diluted (from 1:2000 to 1:16000) and added to individual wells, then the plate was incubated at room temperature for 2 h. Wells were again washed six times with PBST, and then alkaline phosphatase-conjugated anti-horse IgG antibody (Santa Cruz Biotechnology, USA) was added to each well and the plate was incubated at room temperature for 1 h. After washing six times with PBST, the substrate 4-methyl umbelliferyl phosphate (100 μM, 100 μl/well; Molecular Probes) was added to each well, and fluorescence was measured with a SpectraMax M2 microplate reader (Molecular Devices, USA) at excitation and emission wavelengths of 355 and 460 nm, respectively.
Snake venom proteins (5 μg) were separated by sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE), transferred onto PVDF (polyvinylidene difluoride) membranes (Millipore, USA), and probed with antivenom or purified Ab. Immunoreactive proteins in PVDF membranes were detected by incubating for 1 h with the appropriate alkaline phosphatase-conjugated anti-horse IgG antibodies (Santa Cruz Biotechnology, USA) and visualized using the CDP-Star Western Blot Chemiluminescence Reagent (PerkinElmer, USA).
Antibodies were biotinylated using a Lightning-Link biotinylation kit (Innova Biosciences, USA) according to the protocol provided by the manufacturer. Briefly, 100 μl of SSAb (2 mg/ml) was mixed with 10 μl of modifier reagent, then added to the tube containing biotinylation powder and incubated for 15 min in the dark. After the biotinylation reaction, 10 μl of quencher reagent was added and the reaction mixture was stored at -20°C until use.
SSAb (100 μl at 2 mg/ml), diluted 1:1000 in PBS, was coated onto 96-well polystyrene microplates. Thereafter, wells were blocked by incubating with 1% bovine serum albumin (BSA) in PBS for 1 h, then washed six times with 200 μl PBST. Test samples (100 μl) were added into individual wells and incubated at room temperature for 2 h. After washing six times with PBST, 100 μl of biotin-labeled SSAb, diluted 1:16000 in PBS, was added and plates were incubated for 2 h. Plates were again washed six times with PBST, then alkaline phosphatase-conjugated streptavidin was added and allowed to interact with biotin. The alkaline phosphatase substrate, 4-methyl umbelliferyl phosphate (100 μM, 100 μl/well), was then added to each well, and fluorescence was measured with a SpectraMax M5 microplate reader at excitation and emission wavelengths of 355 and 460 nm, respectively.
Experiments were performed on male 7-wk-old littermate mice (C57BL/6Narl strain). Mice were maintained under specific pathogen-free conditions with a 12:12 h light-dark cycle at a temperature of 22°C and a humidity level of 60–70%. Animals had ad libitum access to food and water. Mice (n = 3/group) within a defined weight range (20–25 g) were subcutaneously (B. multicinctus and N. atra venom) or intraperitoneally (T. stejnegeri and P. mucrosquamatus venom) injected with a precise 0.1 ml volume of sterile saline solution containing a minimal lethal dose (MLD) of venom. Blood samples from each mouse were collected using a heparinized capillary blood collection system (Kent Scientific, USA) 0.5, 1, 1.5 and 2 h after venom injection. Collected blood was centrifuged at 3000 × g for 20 min. The resulting supernatant (plasma) was collected into a microcentrifuge tube and stored at -80°C before use.
A colloidal gold (40 nm) solution (REGA Biotechnology Inc., Taipei, Taiwan) was adjusted to pH 8.0 with 0.1 M potassium carbonate. The optimal concentration of SSAb (10 mg) was added to 2 ml of colloidal gold solution and incubated at room temperature for 10 min with gentle mixing. The mixture was blocked by incubating with 0.5 ml of 5% BSA in PBS at room temperature for 15 min with gentle mixing, and then centrifuged at 10,000 × g at 4°C for 30 min. The gold pellets were suspended in PBST containing 1% BSA, and washed by repeated centrifugation and suspension in the same solution. The final precipitates were suspended in 1 ml PBST containing 1% BSA and stored at 4°C until use.
The strips were manufactured by REGA Biotechnology Inc. (Taipei, Taiwan). Nitrocellulose membranes, sample pads, conjugate pads and absorbent pads were all from REGA Biotechnology Inc. Conjugate pads were saturated with HSS-Ab–or NSS-Ab–conjugated colloidal gold, then dried at 37°C for 1 h before assembling. The nitrocellulose membrane was pasted to the cardboard, after which conjugated and absorbent pads were also pasted to the cardboard such that they overlapped with each side of the nitrocellulose membrane by about 2 mm. The sample pad was also laid over the absorbent pad (2 mm overlap) and pasted onto the cardboard. The AGISMART RP-1000 rapid test immuno-strip printer (REGA Biotechnology Inc.) was used to dispense HSS-Abs and NSS-Abs (2 mg/ml) onto hemorrhagic and neurotoxic test lines, respectively, and goat anti-horse IgG antibody (2 mg/mL) (REGA Biotechnology Inc.) onto the control line on the nitrocellulose membrane. The distance between each line was 5 mm. The strips were prepared and assembled in a low-humidity environment, packaged into an aluminum pouch, and stored at room temperature before use.
Patients with suspected snakebite were admitted directly to the Emergency Departments of Taipei Veteran General Hospital, Linkou Chang Gung Memorial Hospital, Chiayi Chang Gung Memorial Hospital or Hualien Tzu Chi Hospital, and did not receive antivenom treatment before being enrolled in this study. After obtaining signed, informed consent forms from patients, 5 ml of blood was collected in SST blood collection tubes (BD, Franklin Lakes, New Jersey, USA) and centrifuged at 4°C for 10 min to obtain serum samples. A 100–200 μl aliquot of serum sample was immediately applied to lateral flow strip test in the emergency room, and results were determined by clinical physicians. The remainder of each sample was sent to the laboratory in Chang Gung University and stored at -80°C. All samples were re-analyzed using the lateral flow strip test in the laboratory to confirm emergency room result; samples were also analyzed by sandwich ELISA to measure the concentrations of venom proteins.
Each serum sample (100–200 μl) was diluted with 1 volume of reaction buffer (100 mM borax, 250 nM polyvinylpyrrolidone (PVP)-40 and 1% Triton X-100) in a microcentrifuge tube. The strips were directly soaked in the samples, and results were recorded after a 10-min reaction.
The Cohen's kappa coefficient (κ) statistic [37, 38] was used to assess the strength of inter-method agreement for diagnosis results. The value of kappa coefficient statistic over 0.75, between 0.75 to 0.40, or below 0.40 indicates excellent agreement, good to fair agreement, and poor agreement, respectively [39, 40].
All clinical serum samples were collected and obtained at Taipei Veteran General Hospital, Linkou Chang Gung Memorial Hospital, Chiayi Chang Gung Memorial Hospital or Hualien Tzu Chi Hospital from February 2017 to February 2018. All study subjects are adult participants and signed an informed consent form approved by the Institutional Review Board (IRB) of Taipei Veteran General Hospital (Approval No: 2017-06-013BCF) and Linkou Chang Gung Memorial Hospital (Approval No: 201800098B0) permitting the use of plasma samples for this study. Experiments involving the care, bleeding, and injection of mice with various venoms were reviewed and approved by the Institutional Animal Care and Use Committee of Chang Gung University (Permit Number: CGU14-024). The protocol for mouse studies was based on guidelines provided by the Council for International Organizations of Medical Sciences (CIOMS) [41].
To assess the cross-reactivity among four venoms and two antivenoms, we performed indirect ELISAs and immunoblotting. The results of indirect ELISAs showed that cross-reactivity of FHAV towards B. multicinctus and N. atra venom was very low (Fig 1A); however, FNAV strongly cross-reacted with T. stejnegeri and P. mucrosquamatus venom (Fig 1B). Cross-reaction signals increased gradually with increases in antivenom concentration, and both antivenoms showed stronger reactivity toward homologous venoms than heterologous venoms. As shown in Western blot profiling data, FHAV primarily cross-reacted with protein bands in the high molecular weight region (55–70 kDa) of N. atra venom (Fig 1C), whereas FNAV cross-reacted with multiple bands in T. stejnegeri and P. mucrosquamatus venoms, predominantly towards protein bands in the 15–25 kDa range in P. mucrosquamatus venom (Fig 1D). A comparison of the protein profiles of the four venoms (S1 Fig) showed that, generally, most venom components of these venoms were recognized by the corresponding homologous antivenom.
In this study, we used an affinity purification procedure to eliminate cross-reactive antibodies from antivenoms. Heterologous venom-immobilized affinity columns were prepared and used to remove cross-reactive antibodies from antivenoms, after which the remaining antibodies were purified using a homologous venom-immobilized affinity column, yielding SSAbs. SDS-PAGE analysis of the affinity-purified HSS-Abs and NSS-Abs showed a typical pattern of IgG heavy and light chains (S2 Fig). Indirect ELISAs and Western blotting assays were performed to evaluate the specificity of affinity-purified SSAbs, HSS-Abs and NSS-Abs. The results of indirect ELISAs showed that both SSAbs possessed high specificity toward the homologous venoms, and showed significantly decreased cross-reactivity with heterologous venoms compared with the original antivenoms (Fig 2A & 2B). The immunoreactivity of HSS-Ab towards P. mucrosquamatus venom was stronger than that towards T. stejnegeri venom, whereas NSS-Ab preferentially reacted with venom proteins from N. atra compared with those from B. multicinctus. Consistent with the ELISA data, Western blot analyses also showed the high specificity of HSS-Ab and NSS-Ab towards their homologous venoms (Fig 2C & 2D), although NSS-Ab did weakly react with high-molecular-weight proteins (55–70 kDa) in the two hemorrhagic venoms (Fig 2D). Proteins in the high molecular weight region (25–70 kDa) of T. stejnegeri and P. mucrosquamatus venom represented the dominant targets of HSS-Ab (Fig 2C); in contrast, NSS-Ab mainly recognized lower molecular weight proteins (<15 kDa) in the two neurotoxic venoms (Fig 2D).
To form sandwich complexes for ELISA measurements, we used HSS-Ab (or NSS-Ab) as the capture antibody and biotinylated HSS-Ab (or NSS-Ab) as the detection antibody. Antibody concentrations, buffers, and incubation times used for these sandwich ELISAs were optimized based on the ELISA development guide provided by the manufacturer (R&D Systems, Inc.). To determine the sensitivity of sandwich ELISA assays for snake venom detection, we serially diluted the four snake venoms in plasma and measured their reactivity by sandwich ELISA, generating standard curves for each venom (Fig 3). The limits of detection (LODs) of sandwich ELISAs for detecting T. stejnegeri, P. mucrosquamatus, B. multicinctus and N. atra venom were 0.39, 0.14, 0.56 and 0.23 ng/ml, respectively. In all cases, R2 values of standard curves were greater than 0.99. Taken together, these results suggest that our sandwich ELISA has the potential to identify snake species and quantify venom proteins in body fluids. For further application of this snakebite sandwich ELISA, the four venoms were used as the gold standards for venom quantification, and the LOD value determined as described above was set as the cutoff for detecting each venom.
To determine whether snake venoms are still detectable after neutralization by antivenoms, we individually neutralized a fixed amount of venom with serially diluted antivenoms and then performed sandwich ELISAs. ELISA signals produced by 10 ng of T. stejnegeri (Fig 4A) and P. mucrosquamatus (Fig 4B) venom were completely eliminated by 8–40 nl of FHAV. Similarly, 40–200 and 8–40 nl of FNAV totally blocked ELISA signals derived from 10 ng of B. multicinctus (Fig 4C) and N. atra (Fig 4D) venom, respectively. These observations show that our sandwich ELISA assays only detects “free” venom proteins, and not antivenom-neutralized venoms. Importantly, they also suggest that our assays are suitable for evaluating the amount of free venom proteins remaining in a snakebite victim, making it possible to determine whether the dosage of antivenom delivered is sufficient to treat the patient.
The MLD of each venom was determined using an experimental envenomation animal model. The MLD of T. stejnegeri, P. mucrosquamatus, B. multicinctus and N. atra were 1.5, 3, 0.3 and 0.65μg/g, respectively. All mice developed local symptoms within 10–20 min after injection of a lethal dose of venom. As soon as 30 min post injection, all four venoms could be detected by sandwich ELISA in plasma samples from mice injected with venom; as expected, none of the saline-injected control mice showed a positive reaction in these assays (Fig 5). The plasma concentrations of T. stejnegeri (Fig 5A), P. mucrosquamatus (Fig 5B) and N. atra (Fig 5D) venom proteins in these mice gradually increased during a 2-h period post injection. In contrast, the plasma concentrations of venom proteins in mice injected with B. multicinctus venom decreased dramatically during this period (Fig 5C). Collectively, these results demonstrate that the newly developed sandwich ELISA can successfully identify and quantify these four Taiwanese snake venoms in vivo.
Although the newly developed sandwich ELISA assay exhibited high specificity and sensitivity, the assay time in its current format is too long for use in clinical practice. To reduce the operation time and simplify the platform for snakebite diagnosis, we sought to develop another assay using a lateral flow strip format with two test lines (Fig 6A). To assess the specificity and sensitivity of this lateral flow strip, we tested it on the four venoms serially diluted (from 500 ng/ml to 5 ng/ml) in human plasma. The assay was evaluated based on the appearance of a control line, a hemorrhagic test line (H line), or a neurotoxic test line (N line) (Fig 6B). All strips showed a visible control line, confirming that all test samples were successfully flowed onto the strips (Fig 7). An H line was only observed in those strips used to test T. stejnegeri and P. mucrosquamatus venom (Fig 7A & 7B), and the N line appeared only in assays of N. atra and B. multicinctus venom proteins (Fig 7C & 7D). These results indicate that this newly developed strip assay does not exhibit sufficient cross-reactivity to cause ambiguous results. In assays of hemorrhagic venom, the H line was still detectable after reducing the concentration of T. stejnegeri and P. mucrosquamatus venom proteins to 50 ng/ml (Fig 7A & 7B). For neurotoxic venom detection, the N line was still visible when both venom protein levels were reduced to 5 ng/ml (Fig 7C & 7D).
Thirty-two victims of snakebite sent to Emergency Departments of the four participating hospitals from May 2017 to February 2018 were enrolled in this study. Among them, eleven patients were excluded because they had been treated with the appropriate antivenom before arrival in the Emergency Department (n = 9) or displayed no symptoms (n = 2). The serum samples obtained from the remaining 21 cases were analyzed by sandwich ELISA and lateral flow strip assay (Table 1). The lateral flow strip assay showed 100% (5/5) specificity and 100% specificity (5/5) for the detection of neurotoxic envenomation samples. However, the sensitivity for detecting hemorrhagic envenomation samples was only 36.4% (4/11). We used the kappa statistic to assess the strength of agreement between the two assays, and this analysis indicated good to fair agreement (κ = 0.53) between snakebite sandwich-ELISA and lateral flow strip assay (Table 1).
The clinical information of these 21 patients were summarized in Table 2. Most of the culprit snakes were initially identified by patients’ description or recognition of snake photograph (17/21), and 2 of them were definitely confirmed according to the killed snakes brought to the hospital. The aggressor snakes of case 18–21 cannot be identified at scenes of ED. In the laboratory identification, both ELISA and lateral flow strip assay were shown hemorrhagic venom positive results for case 18 and 19, and venom negative result for case 20 and 21. All patients were presented with local swelling except case 11 who was initially identified as B. multicinctus envenomation, and no neurologic symptoms appeared in all. Case 16, 18 and 19, who were performed surgery, have higher level of venom concentration than other victims. Seven cases with hemorrhagic venom-positive ELISA results appeared with negative lateral flow strip results. The venom concentration of them was ranged from 2.2 to 10.6 ng/ml, which are lower than other cases detected by lateral flow strip assay. Among them, five cases were shown mild clinical severity, and 2 cases shown moderate severity. Case 11, 15, 17, 20 and 21 have ELISA undetectable venom level. All of them have mild clinical severity that the local swelling restricted in fang mark area, or even did not have local swelling. The sample time after bite for the majority of the victims (15/21) was ≦3.5 h. Overall, there was no significant correlation between the blood venom concentration and sampling time after snakebite or the bitten area according to this small-scale clinical study.
The presence of common antigens in heterologous venoms has been demonstrated to be a major source of bias for the development of snakebite detection assays [26, 42]. The appearance of widespread cross-reactivity between heterologous snake venoms and polyvalent or monovalent antivenoms considerably hampers the specificity of such assays [11, 12, 28, 43]. Consistent with these previous observations, the current study also found that FHAV and FNAV cross-reacted towards heterologous venoms, as evidenced by the detection of 3–5 protein bands in Western blot analyses (Fig 2A & 2B). However, snake venoms are known to comprise multiple (10–100) proteins, many of which have the same or similar epitope(s), but with different molecular weights. At present, it is difficult to predict the venom components that contribute to this cross-reactivity. Immunoaffinity purification appears capable of removing antibodies in antiserum that recognize common epitopes of venom components. Even though the identity of the species-specific antigens and common epitopes that contribute to the cross-reactivity remain largely unknown, we were still able to successfully obtain venom protein antibodies with high specificity (i.e., low cross-reactivity among different snake species). In addition, detection of snake envenomation by monoclonal antibodies generated using a single species-specific venom protein can considerably improve assay specificity [44–47]. However, the sensitivity of these antibodies may not be high enough, because venoms contain numerous protein components and a mAb can only react with a single epitope on its target protein. Moreover, the targeted venom component may become degraded through metabolic processes in biological systems. Thus, the application of monoclonal antibodies to the development of snakebite kits remains a considerable challenge. The promising data shown in the present study suggest that purification of SSAbs from antivenoms could be a feasible and cost-effective strategy for generating effective probes for snake venom detection and species discrimination.
Sandwich ELISAs, which have been widely used in snake venom detection and snakebite diagnosis [10, 11, 44, 48], are capable of measuring venom proteins at the level of a few nanograms per milliliter. In conjunction with the biotin-streptavidin amplification system, the detection limit can be further improved, reducing the lower limit to less than 1 ng/ml [10]. Generally, two different antibodies are used for sandwich ELISA assay development. Because we used the same SSAb as both capture and detection antibody in our sandwich ELISA, the capture SSAbs in the solid phase only occupied one binding site on their cognate antigen molecules. Thus, the detection SSAb was still capable of recognizing the remaining epitopes on the captured antigens. With this approach, how to pair two suitable antibodies to form the sandwich complex for detection is not a concern, making it easy to adapt for snake venom detection. Although the sandwich ELISA assay is time consuming, and thus is likely not the most appropriate assay for use in emergency rooms, it is still a good tool for snakebite epidemiology and prognosis studies.
The usefulness of our sandwich ELISA assay was demonstrated by detecting venoms in blood samples from an experimentally envenomed mouse model (Fig 5). These experiments showed that this assay is capable of identifying the envenoming species and quantifying venom concentrations in blood. Application of this ELISA to the snakebite animal model revealed that concentrations of T. stejnegeri, P. mucrosquamatus and N. atravenom proteins gradually increased in mouse plasma during a 2-h period post-injection; in contrast, the concentration of B.multicinctus venom proteins dramatically decreased over this same time period (Fig 5). A previous study reported that more than half (nearly 60–80%) of B.multicinctus venom components are neurotoxins, including β-bungarotoxin,α-bungarotoxin and γ-bungarotoxin [49]. These bungarotoxins bind to specific receptor(s) on presynaptic and postsynaptic membranes, leading to paralysis and neurotoxicity [50–52]. Our findings suggest that, when injected into the victim, these bungarotoxins rapidly interact with specific receptors, and thus are immobilized in the neuromuscular junctions; this, in turn, causes a significant decrease in their bioavailability, accounting for the rapid decrease in their concentration in blood plasma.
The lateral flow strip assay is a sandwich-based immunostrip used to rapidly (5–20 min) examine whether target molecules are present in a sample [53]. This type of assay is appropriate for use in snakebite detection and diagnosis, and can offer guidance to physicians in administering antivenom [22, 23]. Furthermore, the visual diagnosis format of this assay is simple, making it desirable for use in developing countries, where snakebites are most prevalent. However, some factors and sampling conditions may profoundly affect strip assay results. For example, a high concentration of serum proteins and high viscosity of the test sample could interfere with the formation of the red line in test and control zones, and samples containing high concentrations of salt, such as urine, often cause false-positive results. Thus, in some situations, sample pretreatment is required. The lateral flow strip assay developed here has two test lines for discriminating hemorrhagic and neurotoxic snake envenomation in Taiwan. This strip assay successfully detected and identified snake venom in serum samples from snakebite patients.
Our small-scale clinical study demonstrated that the lateral flow strip assay is useful for assessing neurotoxic envenomation, exhibiting a sensitivity/specificity of 100%. It is suggested that the newly developed strip assay holds promise for the diagnosis of neurotoxic snakebite. However, the sensitivity of this assay for hemorrhagic envenomation was nearly 40%. ELISA results of these 11 hemorrhagic envenomation samples showed that the T. stejnegeri or P. mucrosquamatus venom protein concentrations in 7 lateral flow strip-negative samples were less than 10 ng/ml (Table 2), suggesting that this assay is not sensitive enough to detect snakebite cases with low blood concentration of hemorrhagic venom in clinical practice. Although, at this point, we cannot definitively establish the appropriateness of our lateral flow strip assay for precise diagnosis of all clinical snakebites, the combination of clinical symptoms and the results of lateral flow strip could improve the clinical utility of our lateral flow strip, especially in the weak aspect of diagnosis of hemorrhagic snake envenomation. A diagnosis flowchart which composed of clinical symptoms and the result of lateral flow strip was therefore proposed (Fig 8). Because of the relative high sensitivity and specificity of our lateral flow strip in diagnosis of neurotoxic snake envenoming, cases with negative lateral flow strip results have a great possibility of hemorrhagic snake envenoming when they have developed local tissue swelling. This diagnosis flowchart may further enhance the ability of our lateral flow strip to guide the usage of antivenom. Because only 21 snakebite cases were included, further study using a larger sample set is needed to verify the sensitivity, specificity, stability, and feasibility of this strip assay.
Seven of the 21 clinical samples examined in this study showed positive ELISA result but negative on the lateral flow strip test. All of them were identified as hemorrhagic snake envenomation with low venom concentration level accompanying with mild or moderate clinical severity. Even though these patients have been transferred to hospital and sampled nearly within 1–2 hrs, their blood venom concentrations were still lower than the others. It is highly possible that the amount of venom injected into these victims was originally low, which is hard to detect by lateral flow strip assay after dilution in the systemic circulation, and only induced mild clinical symptoms. Despite initial identification of envenoming species is almost the same as the test results in our small-scale study (Table 2), sometimes, envenoming species identified by patients or their family may mislead the physicians. Take case 8 as an example, this patient was initially identified as P. mucrosquamatus envenomation according to family members’ recognition of the snake pictures, however, both ELISA and lateral flow strip assay showed positive result of neurotoxic snake envenomation, indicating the culprit snake is N. atra. Furthermore, few cases with negative result of both assays may be bitten by non-venomous snakes. There are more than 50 snake species in Taiwan. It is hard for citizens to correctly recognize and distinguish all of them. Bringing the envenoming snake to the hospital, like cases 9 and 13, is the most reliable way for species identification.
In the present study, all five cases (case 11, 15, 17, 20 and 21) with negative quantification of venom displayed the mild clinical severity. These patients may be bitten by non-venomous snakes, or the dry bite. As mentioned above, snakebite victims have the chance to misidentify the envenoming species, and slight swelling usually occurred around the fang mark even if they were bitten by non-venomous snakes. It is one of the reasons leading to the negative results in both assays. In addition, although we did not observe a close relationship between the transcurrent time from the bites to the ER consult and the results of the diagnostic test, 3 of the 5 cases with negative ELISA result had longer transcurrent time. Case 11, 15 and 21 had their transcurrent time for 14, 10.5 and 34 hrs, respectively. The metabolism time more than 10 hours may allow the venom to be eliminated from patients’ body and resulted in negative test result. The delay in seeking medical help may be another reason leading to the negative test results. On the other hand, case 18 had 12.5 hours of transcurrent time, but displayed severe clinical symptoms and positive test results. It is reasonable to assume that the type and amount of venom injected into patients is the main factor to determine the outcome of the test results, and the effect of transcurrent time could be minor.
The current study used serum samples from snakebite patients to evaluate the performance of the snakebite lateral flow strip assay. Other types of specimen, such as urine, wound exudate and blister fluid, have been reported as alternatives for venom detection [7, 12]. The highest amounts of venom proteins (>100 ng/ml) are found in wound exudates and blister fluid; thus, venom proteins are more easily detected and measured in these types of specimens [12]. However, cases with blister fluid are very rare; in the current study, only one patient formed blister fluids after envenomation. Wound exudates are easier to obtain than blister fluids, but obtaining untreated wound exudates for pre-clinical trials is another challenge. Because people have been taught to perform first aid when bitten by snakes, snake venom remaining in the wound will typically have been washed out or swabbed out. Furthermore, fang marks have usually clotted by the time victims arrive at the Emergency Department. Thus, although wound exudate maybe the best sample type for venom detection, how to collect good quality samples for survey remains a daunting challenge.
Countries in tropical and subtropical regions have various indigenous venomous snake species. Two or more antivenoms are currently available for clinical treatment of snake envenomation. Directly using these antivenoms as a resource for the development of snakebite diagnostic assays could be a cost-effective approach for snakebite management. The use of an affinity purification strategy makes it possible to obtain SSAbs from antivenoms, thereby eliminating cross-reactive antibodies and preventing false-positive results in assays of snake venoms. This approach obviates the need to produce additional polyclonal or monoclonal antibodies, and alleviates concerns regarding whether the antigens targeted by the polyclonal or monoclonal antibodies produced are species specific. SSAbs purified from antivenoms are suitable for use in developing sandwich ELISAs and lateral flow assays for rapid detection of snake venoms. The ability of these purified SSAbs to detect venom in the blood of animal models as well as in blood samples taken from snakebite patients validates the usefulness of this strategy.
In conclusion, our data indicate the feasibility of a cost-effective approach (i.e. preparation of SSAbs from specific antivenoms available in Taiwan) to develop the snakebite diagnostic assay for discriminating hemorrhagic and neurotoxic snake envenomation in Taiwan. When combining the clinical observation of patient’s symptom, this assay would aid in the clinical decision of the appropriate antivenom to be used where the signs and symptoms of the envenoming did not allow a precise diagnosis by the clinician responsible to treat the envenomed patient. Although our present results are promising, further studies including improvement of detection sensitivity/specificity of the assays and application of the optimized assays to a larger sample set are needed to validate the clinical utility of the assays for snakebite management.
|
10.1371/journal.pcbi.1002963 | Computational Biomarker Pipeline from Discovery to Clinical Implementation: Plasma Proteomic Biomarkers for Cardiac Transplantation | Recent technical advances in the field of quantitative proteomics have stimulated a large number of biomarker discovery studies of various diseases, providing avenues for new treatments and diagnostics. However, inherent challenges have limited the successful translation of candidate biomarkers into clinical use, thus highlighting the need for a robust analytical methodology to transition from biomarker discovery to clinical implementation. We have developed an end-to-end computational proteomic pipeline for biomarkers studies. At the discovery stage, the pipeline emphasizes different aspects of experimental design, appropriate statistical methodologies, and quality assessment of results. At the validation stage, the pipeline focuses on the migration of the results to a platform appropriate for external validation, and the development of a classifier score based on corroborated protein biomarkers. At the last stage towards clinical implementation, the main aims are to develop and validate an assay suitable for clinical deployment, and to calibrate the biomarker classifier using the developed assay. The proposed pipeline was applied to a biomarker study in cardiac transplantation aimed at developing a minimally invasive clinical test to monitor acute rejection. Starting with an untargeted screening of the human plasma proteome, five candidate biomarker proteins were identified. Rejection-regulated proteins reflect cellular and humoral immune responses, acute phase inflammatory pathways, and lipid metabolism biological processes. A multiplex multiple reaction monitoring mass-spectrometry (MRM-MS) assay was developed for the five candidate biomarkers and validated by enzyme-linked immune-sorbent (ELISA) and immunonephelometric assays (INA). A classifier score based on corroborated proteins demonstrated that the developed MRM-MS assay provides an appropriate methodology for an external validation, which is still in progress. Plasma proteomic biomarkers of acute cardiac rejection may offer a relevant post-transplant monitoring tool to effectively guide clinical care. The proposed computational pipeline is highly applicable to a wide range of biomarker proteomic studies.
| Novel proteomic technology has led to the generation of vast amounts of biological data and the identification of numerous potential biomarkers. However, computational approaches to translate this information into knowledge capable of impacting clinical care have been lagging. We propose a computational proteomic pipeline for biomarker studies that is founded on the combination of advanced statistical methodologies. We demonstrate our approach through the analysis of data obtained from heart transplant patients. Heart transplantation is the gold standard treatment for patients with end-stage heart failure, but is complicated by episodes of immune rejection that can adversely impact patient outcomes. Current rejection monitoring approaches are highly invasive, requiring a biopsy of the heart. This work aims to reduce the need for biopsies, and demonstrate the power and utility of computational approaches in proteomic biomarker discovery. Our work utilizes novel high-throughput proteomic technology combined with advanced statistical techniques to identify blood markers that guide the decision as to whether a biopsy is warranted, reduce the number of unnecessary biopsies, and ultimately diagnose the presence of rejection in heart transplant patients. Additionally, the proposed computational methodologies can be applied to a range of proteomic biomarker studies of various diseases and conditions.
| After the first successful human-to-human heart transplant in 1967, cardiac transplantation became the primary therapy for patients with end-stage heart failure due to dilated cardiomyopathy or ischemic heart disease. Improvements in immunosuppressive drug therapies have significantly increased the number of successful transplants, yet episodes of acute rejection and progression of chronic rejection remain major factors that negatively impact long term graft survival. Acute rejection is predominantly considered to be an immunological reaction in response to the major and minor histocompatibility antigens recognized as ‘foreign’ by the graft recipient. This process triggers the subsequent activation, migration and infiltration of immune cells such as T- and B-lymphocytes, which can ultimately lead to cellular- and antibody-mediated destruction of the heart allograft tissue [1]. Endomyocardial biopsy (EMB), through which histological features such as cellular infiltration and myocyte damage can be observed, is currently considered to be the only reliable gold standard for diagnosis and monitoring of acute cardiac allograft rejection [2]. However, the invasive and qualitative nature, risk of complications, associated cost and lack of timeliness of the results render the EMB a suboptimal procedure for routine monitoring [3]. A more reliable, minimally invasive, inexpensive, and early diagnostic tool to monitor graft survival remains a significant clinical unmet need.
Since proteins may serve as molecular indicators (i.e., biomarkers) of cardiac allograft rejection, plasma proteomics offers an attractive and promising avenue for the development of diagnosis tools for cardiac transplantation [4]. Technical advances in the field of quantitative proteomics in the last decade have enabled the identification and quantitation of thousands of proteins and have stimulated a large body of research focused on the discovery of new biomarkers. However, the translation of candidate biomarkers from discovery research into proteomic tests for clinical use has faced significant challenges, due mostly to a lack of an adequate analytical pipeline [5], [6], [7], [8]. In a significant step forward, technological proteomic pipelines have recently been proposed, optimizing the design of the discovery, validation, and clinical implementation stages of biomarker studies [8], [9], [10], [11]. Nevertheless, the development of new clinical proteomic tests hinges on a tailored computational pipeline to distill the information contained in thousands of proteins into an accurate classifier score with demonstrable clinical utility.
Computational proteomics is a new and expanding field of research which primarily focuses on data management and mass-spectra analysis for the discovery phase of biomarker studies [12], [13], [14]. Although previous work has acknowledged the need of a tailored computational pipeline in proteomics (e.g.,[15]), a systematic and complete process that specifically addresses the challenges emerging from proteomic studies has not been proposed or demonstrated to date. Using unsuitable methodological tools to explore and analyze the data may result in the selection of biomarkers that ultimately fail in the final stages of validation, or may fail to select relevant biomarkers. For example, identifying a panel of candidate markers based only on the comparison of relative abundance between case and control samples, or the use of classical statistical tests when the sample size of the study is too small, may result in the identification of many false candidate markers.
We complement previous work by proposing a computation pipeline powered by extensive statistical analysis for all stages of quantitative proteomics biomarker studies (Figure 1). At the discovery stage, the pipeline focuses on selecting an appropriate experimental design and statistical methodologies to identify and assess a panel of candidate biomarkers. At the validation stage, the pipeline emphasizes on the migration of discovery results to the validation platform, and the development and validation of a biomarker classifier. At the clinical implementation stage, the main aims are to develop an assay suitable for clinical deployment, and to calibrate the biomarker classifier using the developed assay.
We demonstrate the power of our methodology in a proteomic biomarker study in the context of cardiac transplantation, with a goal towards the development of a more accurate and less invasive blood test for monitoring graft survival. Our work identified a panel of five candidate plasma proteins that clearly discriminates acute cardiac allograft rejection from non-rejection. These biomarker proteins distribute broadly among three relevant biological processes: cellular and humoral immune responses, acute phase inflammatory pathways and lipid metabolism. Of the five candidate biomarkers, we corroborated four using two independent platforms. A classifier score based on these four corroborated proteins measured by multiple reaction monitoring mass-spectrometry (MRM-MS) demonstrated that plasma protein biomarkers have significant potential in serving as a reliable, minimally-invasive, inexpensive, and timely diagnostic tool for acute cardiac allograft rejection. Our results advance the approaches to diagnosis with respect to cardiac transplantation biomarker, as well as the computational methodologies tailored for a wide range of proteomic biomarker studies.
A synopsis of the computational pipeline proposed in this study is illustrated in Figure 1. We first describe in detail each of its steps for the discovery, validation, and clinical implementation stages. We then present a brief description of the materials and methods related to the biomarker study in cardiac transplantation used to illustrate the proposed methodology.
Recent technological advances in quantitative proteomics have enabled the untargeted quantitation and identification of hundreds to thousands of proteins simultaneously from complex samples such as human plasma. The aim of the discovery pipeline is to create a list of candidate markers from an extensive set of proteins identified and measured within each sample.
As widely discussed in the literature, any list of candidate markers identified in a discovery stage must be validated in a large and independent cohort of patients before its clinical utility assessment. To bridge the gap between discovery and clinical technologies, the validation stage is usually performed in an independent platform which provides a timely and cost-effective approach to measure all samples. To overcome the dependence on antibody availability, we developed an MRM-MS assay to complete the validation stage. However, similar analytical steps would have been taken if another independent platform was used.
The final translation of proteomic results from the validation to the clinical implementation stage requires careful examination of many factors, including the development of assays suitable for clinical laboratories, considerations from health economics, as well as approval of regulatory agencies (e.g., Food and Drug Administration, Conformité Européenne mark) [8]. From a methodological point of view, the following steps are crucial to complete this last stage.
A brief summary of the materials and methods used in the proteomic biomarker study of cardiac transplantation are outlined here and further details are given as supporting material in Text S1.
The first two stages of the computational pipeline, discovery and validation, were applied to a biomarker study in cardiac transplantation. An overall schematic of the number of samples, design, and proteomics data used at each stage is summarized in Table 1.
In the discovery stage, multiplexed iTRAQ-LC-MALDI-TOF/TOF mass spectrometry was used to identify and quantitate proteins from 108 depleted plasma samples representing a time course of 20 weeks from the first 26 patients enrolled (Figure S2). These samples were processed in 50 independent iTRAQ runs, including other samples from the heart cohort. In addition, each iTRAQ run included a normal pooled control plasma sample to provide a common reference across multiple runs. A total of 924 protein groups (PGCs) was cumulatively identified from the 50 runs with an average of 273 PGCs within each run.
Following the selection criteria and the power calculation described in the supporting material (Text S1), the first AR samples from 6 (out of 8) AR patients were selected as cases, and samples from 14 (out of 18) NR patients at matching time points were selected as controls (Figure S2A and Table 1). The remaining 88 longitudinally collected iTRAQ samples were used as test samples to initially validate the results at the Discovery stage. Although samples in this test set are part of BiT cohort, none of them were previously used in the training set. As described in Step 3 of the Discovery stage, only those PGCs identified in at least 2/3 of the AR and the NR groups were considered for further analysis. The resulting data consisted of 127 PGCs measured in at least 4 (out of 6) AR patients and 10 (out of 14) NR patients. Of these 127 PGCs, 51 PGCs contained 133 missing values out of a total of 1020 values (i.e., 51 PGCs×20 patients).
A panel of 5 PGCs was identified with significant differential relative concentrations (robust eBayes p value<0.01) between AR and NR samples (Tables 2 and 3). This panel consisted of 3 PGCs that were more abundant in AR versus NR samples: B2M, F10, and CP, and 2 PGCs that were less abundant: PLTP, and ADIPOQ (Wilcoxon tests are shown in the Table S3).
The quality assessment of the proteomics data demonstrated a strong confidence regarding identified protein identities, wherein 98% of the 127 analyzed PGCs and all 5 PGCs candidate biomarkers were identified based on two or more peptides (Figure S5). Similarly, results showed an overall good coverage and quantitative levels for the analyzed proteins (Table S4). The potential confounding of the results was examined using all available clinical data close to the event time, including daily dose of immunosuppressants, weight, and blood pressure. The GlobalAncova p values (Table 4 and Table S5) demonstrate that the simultaneous relative concentrations of the 5 candidate PGCs remained significantly different in the AR group versus the NR group (p value<0.05) after adjusting for potential confounders. The correlation values in Table 4 show that none of the clinical variables were highly correlated with the LDA classifier score (r<0.5). Overall, the results demonstrated that the identification of the biomarker panel was not confounded by other clinical variables available for this study cohort.
To illustrate the joint performance of all candidate markers to discriminate AR from NR samples, the average LDA score was calculated for all the AR samples (n = 10) and the NR samples from NR patients (n = 40) available at each time point (Figure 2). Based on these initial results, the identified candidate markers together discriminated the two groups regardless of which week the rejection occurred after transplantation. Despite this differentiation, the two AR samples in week 2 were still classified as NR (negative score) by the classifier. Although the LDA classifier score was trained to discriminate AR from NR samples, Figure S7 also includes the score of 47 1R mild, non-treatable rejection samples. Average scores of 1R samples from NR patients were in general similar to those of NR samples, while those from AR patients were closer to the average scores of AR samples.
Figure 3 illustrates the temporal correlation of the score with the diagnosis of rejection. The classifier score for AR patients was at baseline before the rejection episode (pre-rejection point) with a similar average value to that of NR patients at matched time point(s) (no statistical evidence of differentiation). The score for AR patients was differentially elevated at the time point(s) of rejection (as determined by biopsy) compared to that of NR patients (alpha level = 0.05, two-sided t test, p value<0.001) at matched time point(s). The score for the AR patients returned to baseline following treatment and resolution of the rejection episode (post-rejection point, non-rejection determined by biopsy) with a similar average value to that of NR patients. In addition, the evaluation of the score across time shows that the biomarker signature is specific to the rejection episodes, rather than reflecting confounded differences or potential bias between the groups (e.g., different medication regimens).
Further results from an initial validation performed in this stage based on 88 test iTRAQ samples not included in the discovery are shown in Figure S6. A total of 3 out the 4 AR and 29 out of the 37 NR samples tested were correctly classified (non-highlighted cells). Similar results were obtained only if a single test sample per patient was randomly selected.
The results from the iTRAQ discovery analysis were corroborated and initially validated by two independent assays: ELISA/INA (available for ADIPOQ, F10, B2M, and CP), and MRM-MS (developed for ADIPOQ, F10, B2M, CP, and PLTP). Following the results of the power calculation illustrated in Figure S2B, a total of 43 patients were selected and plasma and serum samples were processed by ELISA/INA for an initial validation cohort that extends the discovery cohort. A subset of 25 of these 43 samples, 7 AR, 6 1R and 12 NR, were also part of the iTRAQ discovery cohort. Of these 25 samples, 23 were also processed by MRM-MS (Table 1 and Figures S2A and S2B). Samples measured by the three assays were used to perform the correlation analysis.
Results showed good levels of correlations for B2M, ADIPOQ, and CP (r>0.6, Figure 4-A). F10 measurements from both ELISA/INA and MRM-MS and PLTP measurements from MRM-MS did not show a similar degree of correspondence with iTRAQ as seen for other proteins. However, a good correlation was observed between ELISA/INA and MRM-MS for F10 measurements (r = 0.69, Figure 4A).
The differential protein levels between AR and NR samples observed in the discovery stage were successfully translated for 3 of 4 proteins measured by ELISA/INA (B2M, ADIPOQ, and CP, p value<0.05), and 4 of 5 proteins measured by MRM-MS (B2M, ADIPOQ, CP, and PLTP) (Figure 4A). Results from the ELISA/INA data provided additional validation in 12 new patients using a platform other than iTRAQ (Table 1, and 0R(E) and 2R(E) samples in Figure S2A). Taken together, with the exception of F10, the results showed that measurements from the three platforms were strongly correlated and corroborated most of the results from the iTRAQ discovery stage.
Figure 4B demonstrates the gain in classification performance by a panel of markers combined together into a multivariate classifier score. Although estimated on a small cohort, the sensitivity estimates improved from 17% for the classifier based only on B2M to 100% for the classifier based on the 4 corroborated protein panel (B2M&ADIPOQ&CP&PLTP), the specificity improved from 91% to 100%, and the AUC improved from 0.25 to the maximum of 1. Based on the classification performance of the evaluated MRM-MS classifier scores, a panel of 4 proteins (B2M, ADIPOQ, CP, and PLTP) was selected to complete the validation stage.
Figure 4C illustrates the resulting classifier score based on the 4-protein panel for the test samples resulting from a 6-fold cross-validation. Samples with a positive proteomic classifier score were classified as “rejection”, and those with a negative score were classified as “non-rejection”. In this initial validation, all test samples were correctly classified by the proteomic classifier score. However, because the test samples in the cross-validation were still part of the discovery analysis, these performance measures cannot be used to characterize the identified classifier. Although similar results were obtained using the ELISA/INA measurements on an extended cohort of patients (Figure S8B), a larger validation in an external cohort of patients is still required to complete this phase. A prospective clinical assessment of the value of these proteomic markers of acute rejection is currently underway using MRM-MS measurements of over 200 samples from six Canadian sites.
Over the last two decades, the accelerating pace of technological progress has initiated a new era in the field of clinical proteomics. In particular, plasma proteomics offers a powerful tool to examine the underlying mechanisms of various diseases and opens novel avenues for biomarkers discoveries. To date, the number and quality of technical resources available for proteomic biomarker studies are well recognized. However, the development of statistical methods to address the challenges that have arisen in the field has lagged behind, dramatically reducing the pace, quality and precision of biomarker studies. An important piece of the puzzle in clinical proteomics is to distill the information contained in the very rich data generated by new proteomic technologies via a tailored computational pipeline [15].
In this study we propose and apply a computational pipeline that provides a systematic process to analyze proteomics data related to biomarker studies. The computational steps described in this study are consistent with, and complement those described in previous technological and analytical pipelines [6], [8], [9], [10]. Our pipeline successfully generated an accurate classifier score based on four plasma proteins to diagnose acute rejection in patients who have received cardiac transplants.
There are likely additional plasma protein biomarkers that were not identified by this approach. For example, additional candidate markers may be identified using a different reference sample or an alternative proteomics platform. However, initial validation results indicate that a classifier score based on 4 corroborated biomarkers can achieve a satisfactory classification of acute rejection and non-rejection samples. If validated in a larger and external cohort of patients, the identified proteomic biomarker panel can be used to develop a more accurate and minimally invasive clinical blood test to monitor allograft rejection.
The analysis may also reveal some biomarkers previously associated with unrelated disease phenotypes, or that are not linked to cardiac transplantation. In general, looking at injury controls is a good idea and ideally one would want to include such to show that the identified panel is specific for the disease of interest. However, such comparisons would require additional carefully phenotyped cohorts, analyzed with the same analytical and technological methods on the same sample source, which for many relevant injuries are difficult or impossible to obtain. The data shown in our study do not address this point and much more work needs to be done on the comparison of acute rejection with other injuries.
The complex pathobiology of acute cardiac allograft rejection is reflected in the heterogeneity of markers identified in this study. The majority of proteins identified distribute broadly among three biological processes, consistent with the current understanding and pathogenesis of acute rejection: cellular and humoral immune responses, acute phase inflammatory pathways and lipid metabolism. Our results also highlight the anticipated distinction between the plasma proteome and that observed in tissue-based discovery studies [44]. As well, while circulating protein markers of acute allograft rejection found by others were mainly indicative of tissue damage and stress [44], we also identified markers that implicate immune and vascular processes, in addition to other aspects of the rejection process. In general, knowing the biological process of the identified markers may lead to a better understanding of disease pathogenesis, and to novel therapeutic targets.
Transplantation elicits a host immune response that encompasses both cellular and humoral immunity, which together lead to graft tissue damage, and episodes of acute and chronic rejection. B2M is a protein associated with MHC Class I histocompatibility antigens, with increased levels reflecting allograft rejection, autoimmune or lymphoproliferative diseases as a result of increased immune activation [45]. Several studies have reported higher circulating levels of B2M in cardiac or renal allograft rejection [46], [47], [48], consistent with our observations. Importantly, our data demonstrates improved classification performance when additional markers are used with B2M.
Acute rejection resulting from cellular infiltration of the graft leads to severe local inflammation, which has systemic consequences with a concomitant increase in circulating inflammatory markers. The acute phase response to inflammatory stimuli involves the production and release of numerous plasma proteins by the liver. CP, significantly up-regulated in AR relative to NR samples in this study, is a positive acute phase reactant. It is elevated in acute and chronic inflammatory states and elevated plasma CP is also associated with increased cardiovascular disease risk [49]. CP is a player in inflammation, coagulation, angiogenesis, and vasculopathy, but its role in the pathogenesis of acute rejection is unknown. Current evidence supports a relationship between inflammation and coagulation [50]. FX, a key mediator in the conversion of prothrombin to thrombin, is up-regulated in our acute rejection cohort, and this finding may reflect an intersection between inflammatory and coagulation responses in acute rejection. However, this protein was not validated in our study. C reactive protein (CRP), an acute phase reactant protein previously studied in the context of acute cardiac allograft rejection, was not identified in our study. Consistent with this finding, previous work has demonstrated conflicting evidence regarding the informative value of CRP in monitoring acute cardiac allograft rejection [51].
Dyslipidemia as a consequence of immunosuppressive therapy has been reported in cardiac allograft recipients, and is a risk factor for chronic rejection [52]. Lipid metabolism is represented by two proteins in our panel: ADIPOQ and PLTP. ADIPOQ is a circulating plasma protein involved in metabolic processes shown to play a role in atherosclerotic cardiovascular diseases [53]. Work by Nakano and others described elevated ADIPOQ as reflective of tolerance following a rat model of orthotopic liver transplantation, suggesting a mechanistic role for this protein and corresponding with the observed decrease in ADIPOQ levels during acute rejection episodes [54]. Further, recent work by Okamoto and colleagues [55] has demonstrated that ADIPOQ inhibits allograft rejection in a murine model of cardiac transplantation. PLTP plays a role in HDL remodeling and cholesterol metabolism but its involvement in acute rejection is unknown.
A comparison between the current panel identified for the diagnosis of cardiac allograft rejection, and that of renal allograft rejection [37], reveals that the biological roles of identified proteins are shared in the setting of both transplantation situations. Moreover, the relative regulation of proteins involved in these biological processes is likewise shared. Our current data reveals a differentiation of particular molecules involved in the pathogenesis of cardiac versus renal allograft rejection.
The plasma protein markers identified in this study have the potential to be further assessed in combinatorial analyses with Biomarkers in Transplantation (BiT) genomic and metabolomic data. Notably, numerous research groups, including the BiT group, have identified potential gene expression markers of cardiac allograft rejection using microarray and qPCR analyses of peripheral and whole blood [56], [57], [58], [59]. These studies provide an opportunity for a systems biology approach to understanding allograft rejection.
Taken together, the panel of protein markers identified and initially validated in this study offers a fresh approach to the diagnosis of acute cardiac rejection, providing novel avenues of investigation and potential new targets for therapeutic intervention. The computational pipeline proposed and applied in this biomarker is highly applicable to a wide range of biomarker proteomic studies.
|
10.1371/journal.pgen.1003560 | Cohesin and Polycomb Proteins Functionally Interact to Control Transcription at Silenced and Active Genes | Cohesin is crucial for proper chromosome segregation but also regulates gene transcription and organism development by poorly understood mechanisms. Using genome-wide assays in Drosophila developing wings and cultured cells, we find that cohesin functionally interacts with Polycomb group (PcG) silencing proteins at both silenced and active genes. Cohesin unexpectedly facilitates binding of Polycomb Repressive Complex 1 (PRC1) to many active genes, but their binding is mutually antagonistic at silenced genes. PRC1 depletion decreases phosphorylated RNA polymerase II and mRNA at many active genes but increases them at silenced genes. Depletion of cohesin reduces long-range interactions between Polycomb Response Elements in the invected-engrailed gene complex where it represses transcription. These studies reveal a previously unrecognized role for PRC1 in facilitating productive gene transcription and provide new insights into how cohesin and PRC1 control development.
| An important task for the cohesin protein complex that binds chromosomes is to ensure equal distribution of chromosomes into the daughter cells when a cell divides. Small changes in cohesin activity, however, can alter gene activity without affecting chromosome distribution, and disrupt physical and mental development. How cohesin controls gene activity and development is not well understood. In this study we show that cohesin controls the binding of the Polycomb Repressive Complex 1 (PRC1) to many genes. PRC1 silences many genes that control development. Surprisingly, we find that cohesin aids binding of PRC1 to active genes, where PRC1 ensures that RNA polymerase, the enzyme that transcribes genes, is properly modified before entering the gene body. We also find that cohesin antagonizes the binding and activity of PRC1 at genes silenced by PRC1, and can influence interactions between the DNA sequences that recruit PRC1 and other Polycomb complexes to silenced genes. These findings provide new and unexpected insights into how both cohesin and PRC1 control gene activity during development.
| The cohesin complex that encircles DNA is named for its role in mediating sister chromatid cohesion [1]. Cohesin has four subunits: the Smc1 and Smc3 structural maintenance of chromosomes proteins, Rad21, a kleisin protein, and Stromalin (SA). Cohesin is loaded onto chromosomes by the kollerin complex containing the Nipped-B (NIPBL, Scc2) and Mau-2 (Scc4) proteins, and removed by the releasin complex containing Pds5 and Wapl.
Minor alterations in cohesin function disrupt development without affecting sister chromatid cohesion or chromosome segregation [2]. In humans, dominant loss-of-function mutations in the NIPBL gene encoding a kollerin subunit, and dominant missense mutations in the Smc1 or Smc3 cohesin subunits cause Cornelia de Lange syndrome (CdLS) [3]. CdLS is characterized by poor growth, diverse structural abnormalities and intellectual impairment. Current data argue that small changes in cohesin function disrupt development because cohesin binds and regulates many genes important for growth and development [2], .
Genetic and molecular evidence suggests that cohesin also functionally interacts with Polycomb group (PcG) epigenetic silencing proteins during development. There are two major PcG protein complexes: Polycomb Repressive Complex 2 (PRC2), whose Enhancer of zeste [E(z)] subunit performs histone 3 lysine 27 trimethylation (H3K27me3), and PRC1, whose Polycomb (Pc) subunit binds the H3K27me3 histone modification [5]–[8]. The Drosophila Rad21 cohesin subunit is encoded by verthandi (vtd). Nipped-B kollerin and vtd cohesin subunit mutations suppress the body segment transformations of Pc mutant flies, and a wapl releasin mutation that stabilizes cohesin binding causes phenotypes similar to Pc mutations [9]–[11].
The in vivo findings argue that cohesin antagonizes PcG silencing at certain genes, and consistent with this idea, genome-wide mapping in Drosophila cells revealed that cohesin and Nipped-B preferentially bind promoters of active genes, and are usually excluded from PcG-silenced genes [12]. The few exceptional genes that bind cohesin and also have the PRC2-generated H3K27me3 mark are not fully silenced, and show unusually large increases in expression upon cohesin depletion [13]. Cohesin biochemically interacts with PRC1, suggesting that these two complexes might directly control each other's activities [14].
To clarify the functional relationships between cohesin and PcG complexes, we mapped their binding genome-wide in developing Drosophila wing and cultured nervous system cells, and measured the effects of cohesin on PRC1 binding, and vice versa. We also measured the genome-wide effects of PRC1 depletion on Pol II occupancy and mRNA levels. These studies revealed functional interactions between cohesin and PRC1 that give rise to their co-regulation of active genes, and antagonism to each other on expression of PcG-silenced genes. They uncovered an unexpected role for PRC1 in controlling RNA polymerase II (Pol II) function at active genes that provides new insights into how cohesin and PcG proteins regulate transcription.
The genomic data from cultured cells and genetic interactions between cohesin and PcG mutations suggest that cohesin and PcG proteins functionally interact to control gene expression and development. To test this idea in a developing tissue, we used genomic chromatin immunoprecipitation (ChIP-chip) to map Rad21, Nipped-B, RNA polymerase II (Pol II), and PRC2-generated histone 3 lysine 27 trimethylation (H3K27me3) in late 3rd instar Drosophila wing imaginal discs. Genetic data shows that both cohesin and PcG proteins regulate gene expression and development in wing discs (e.g. [15]–[17]).
Two biological replicates were used for each ChIP-chip experiment, and enrichment was calculated genome-wide using the model-based MAT algorithm [18], [19]. We found that the cohesin binding patterns seen in Drosophila cell lines [12] hold true in developing wing: (a) cohesin (Rad21) co-localizes with Nipped-B (genome-wide correlation r = 0.91), (b) cohesin and Nipped-B preferentially bind active genes occupied by Pol II (r = 0.63), and (c) cohesin and Nipped-B are largely absent from silenced genes with PRC2-generated H3K27me3 (r = 0.24). The genome-wide correlation coefficients are summarized in Table S1. These and the other genome-wide ChIP and expression assays performed for this study (see below) are listed in Table S2. Table S3 lists all annotated genes, and indicates whether or not they bind each of the factors that were mapped by ChIP-chip in wing discs and cultured cells (see below). Table S3 also gives the mRNA expression values determined by microarray.
The correlation between cohesin and H3K27me3 is higher in wing discs than in cultured cells. It is likely that sites of false overlap arise from the mixture of different cell types in the wing disc, in which a gene can be silent in one compartment and active in another. For example, the invected-engrailed (inv-en) gene complex, encoding two homeobox transcription factors that confer posterior fate, is expressed in the posterior wing compartment and PcG-silenced in the anterior (e.g. [16], [20]).
The inv-en complex is of particular interest because it is one of the rare examples of cohesin-H3K27me3 overlap in BG3 cells, where its expression is highly sensitive to cohesin dosage [13]. We thus mapped cohesin (Rad21), H3K27me3, and the PRC1 subunit Polycomb (Pc) in the posterior and anterior wing disc separately to determine if inv-en might be regulated by both cohesin and PcG proteins in either the active or silenced state. We used a transgene that expresses red fluorescent protein (RFP) only in the posterior compartment [21] as a guide to slice the discs into anterior and posterior portions (Figure 1A).
We found that inv-en binds cohesin primarily in the posterior disc, where it is expressed, and is marked by H3K27me3 primarily in the anterior disc, where it is silenced (Figure 1A). The low levels of cohesin detected in anterior chromatin, and low levels of H3K27me3 seen in posterior chromatin, likely reflect imperfect dissection (10–20% cross-contamination by visual inspection and by measuring inv and en RNA levels, Figure S1). This low level of contamination is unavoidable, given the difficulty of the dissection by hand of discs that are some 250 by 350 microns in size. Tissue dissociation and FACS sorting did not provide sufficient viable cells for chromatin preparation. Nonetheless, the data clearly demonstrate an inverse relationship between H3K27me3 and cohesin at the inv-en complex between the anterior and posterior compartments. The genome-wide correlation between Rad21 and H3K27me3 in the posterior wing disc (r = 0.11) is significantly lower than for whole discs (r = 0.24), further confirming that some sites of Rad21-H3K27me3 overlap seen with whole discs reflect tissue heterogeneity, and that H3K27me3 and cohesin have an inverse relationship genome-wide in wing discs as in cultured cells (Table S1).
We unexpectedly found that the Pc PRC1 subunit occupies inv-en and flanking active genes at virtually equal levels in anterior and posterior wing discs (Figure 1A). Even more surprising, by measuring the mRNA levels for some 13,000 genes in whole wing discs (Table S3) and comparing to the genomic ChIP data, we found that Pc associates with a large portion of active genes, including ubiquitously expressed genes such as the Act5C actin gene, diminutive (dm, the Drosophila myc gene) and some ribosomal protein genes (Figure 1B). Among active genes, Pc preferentially occupies those that bind cohesin (Figure 1D). In the posterior disc, Pc and Rad21 binding exhibit a genome-wide correlation of 0.75 (Table S1), and Pc is present at 90% of cohesin-bound genes (Figure 1D). Conversely, 76% of Pc sites exhibit cohesin binding.
The extensive overlap of cohesin and Pc at active genes was unexpected, given that PRC1 is generally thought to associate primarily with PcG-silenced genes. For instance, Pc was previously mapped in Drosophila cell lines with a different antibody (denoted Pc-VP; [22]) than the one used here (denoted Pc-RJ; [23]), and was generally found at silent genes marked by H3K27me3 [24], [25]. Both Pc-VP and Pc-RJ were made using the same Pc fragment as antigen (residues 191–354) and antigen-affinity purified. The Pc-VP antibody was validated for ChIP-chip in the initial studies, and re-validated by the modENCODE project, including RNAi depletion experiments. Two laboratories independently verified the Pc-RJ antibody by RNAi, western blots and ChIP [13], [23]. Figure S2 confirms that the Pc-RJ antibody detects a single protein of the expected size that is reduced in cultured cells subjected to Pc RNAi treatment.
The core PRC1 complex consists of Pc, Polyhomeotic (Ph), Posterior sex combs (Psc) and Sex combs extra (Sce/dRing). We thus further mapped PRC1 in whole wing discs using the Pc-VP Pc antibody and validated antibodies against the Polyhomeotic (Ph) and Posterior sex combs (Psc) PRC1 core subunits [22]. These antibodies were previously validated by westerns, in vivo expression of fusion proteins, and ChIP-chip [22], [24]–[26]. Figure S2 shows that the Ph antibody recognizes one major band of the correct size that is reduced upon treatment of cultured cells with Ph RNAi.
These experiments demonstrate that the PRC1 complex is present at active cohesin-binding genes in wing discs. Ph ChIP gave a nearly identical pattern to Pc-RJ, with a genome-wide correlation with Rad21 of 0.82 in whole discs (Figure 1B,C,E; Table S1; Figure S3). Thus two validated antibodies against different PRC1 subunits show essentially the same pattern. It is very unlikely that both cross-react with a non-PRC1 protein that co-localizes with cohesin by chance. Moreover, Psc was also detected at many cohesin-binding active genes, although the signals are noticeably lower than at silenced genes (Figure 1B,C,E; Figure S3). We cannot distinguish if the lower Psc signals reflect reduced epitope accessibility, or if Psc is present in a smaller fraction of the PRC1 complexes at active genes. For instance, Su(z)2, a Psc homolog encoded by a neighboring gene, may substitute for Psc in a much higher fraction of PRC1 complexes at active genes.
Figure S3 shows plots of the ChIP enrichment at all microarray features for each of the PRC1 subunits against each other over a 400 kb region on chromosome 2L that includes several active genes and the PcG-silenced dpp gene, along with the genome-wide correlation coefficient for each comparison. Ph shows a high genome-wide correlation with Pc-RJ (0.85), while Pc-VP shows a lower correlation, with the plots showing a clear separation of active and silenced genes. Psc correlates with Ph at both silenced and active genes, with a distinct separation into silenced genes with high Psc and active genes with low Psc. Figure S4 shows a high resolution view of the diminutive (dm, myc) gene to show the close similarities in the ChIP patterns of three PRC1 subunits (Pc, Ph, Psc) at a constitutively-active cohesin-binding gene. Figure S5 shows that additional independent anti-Pc and anti-Ph antibodies [27]–[29] also detect PRC1 at active cohesin-binding genes by ChIP-qPCR. We thus conclude that PRC1 is present at most active cohesin-binding genes.
As in cultured cells, the Pc-VP antibody detected Pc binding primarily at sites of H3K27me3 in wing discs, such as inv-en (Figure 1C,E). We do not know why Pc-VP, in contrast to several other PRC1 antibodies, does not detect PRC1 at active genes, but one clear possibility is that the Pc epitope recognized by Pc-VP is masked at active cohesin-bound loci. This idea is consistent with the direct interaction between cohesin and PRC1 detected by purification of biotin-tagged PRC1 [14]. The cohesin-PRC1 interaction was characterized using different cohesin and PRC1 antibodies than those used in our studies. With the antibodies we used for genomic ChIP, immunoprecipitation of Rad21 from soluble nuclear extracts treated with DNase I co-precipitated Pc and Ph, confirming that cohesin and PRC1 interact (Figure S2). These results also provide further validation of the specificity of the Pc-RJ and Ph antibodies.
The extensive overlap of cohesin and PRC1 at active genes in wing discs and the direct interaction between cohesin and PRC1 suggested that they might influence each other's binding. The presence of multiple cell types in wing discs creates ambiguities in interpreting ChIP signals at genes that are not active or silent in all cells. It is also difficult to reduce cohesin and PRC1 levels in vivo by more than 50% without causing lethality or substantially altering development. Homozygous cohesin and PcG mutants are early lethals, and many putatively tissue-specific RNAi drivers cause lethality even without large reductions in cohesin [30]. We thus chose to examine the effects of cohesin on PRC1 binding and vice versa in cultured ML-DmBG3 (BG3) cells derived from 3rd instar central nervous system as a more homogenous cell population, and in which cohesin or PRC1 subunits can be easily reduced by at least 80% using RNAi without causing lethality [13].
We mapped Pc and Rad21 binding genome-wide in BG3 cells using an antibody (Pc-RJ) that detects PRC1 at active genes. Cohesin binds primarily at active genes with promoter-proximal paused RNA polymerase II (Pol II) in BG3 cells, and controls the transition of paused Pol II to elongation [12], [13], [31], [32]. There is extensive overlap between cohesin and Pc in BG3 cells, although lower than seen in wing discs, with 72% of cohesin-bound genes binding Pc (Figure 2A). We conclude, therefore, that PRC1 also binds most active cohesin-binding genes in BG3 cells, including several highly-expressed constitutively-active genes such as some ribosomal protein genes, Act5C, and dm/myc (Figure 3).
We tested if cohesin and PRC1 influence each other's binding by performing genomic ChIP of Rad21 after Ph RNAi depletion, and Pc ChIP after Rad21 RNAi depletion, with two biological replicates for each RNAi treatment. Depletion of cohesin by 80% in BG3 cells does not alter chromosome segregation, cell morphology or viability, and modestly decreases proliferation [13], [31]. This depletion reduces cohesin-binding by 70% or more at genes with high levels of cohesin, and large changes in expression of several genes that bind cohesin [13],[31],[32]. Figure S6 shows that cohesin depletion increases the fraction of cells in G2 without a substantial effect on the proportion of cells in S phase. Prior analysis, including genome-wide mRNA measurements, suggests that the G2 delay reflects decreased diminutive (dm, myc) gene expression and cell differentiation, and not mitotic defects [13]. There are minimal effects on sister chromatid cohesion, and no detectable effects on chromosome segregation. There is a modest 1.5-fold increase in cyclin B mRNA and modest decrease in expression of genes involved in spindle formation and elongation, indicating a delay prior to entry into mitosis. There are no changes in expression of cell cycle checkpoint or apoptosis genes. Consistent with decreased dm/myc expression, the most significant down-regulated gene ontology is protein synthesis, and the top up-regulated gene ontology is development, suggesting a change in cell differentiation. Both the up- and down-regulated genes are highly enriched for cohesin-binding genes, indicating that a large proportion of the affected genes are directly regulated by cohesin, and that most expression changes are not attributable to changes in cell physiology. Although the levels of mRNA produced by most genes that do not bind cohesin do not change upon cohesin depletion, direct transcription measurements show that transcription initiation subtly decreases at most of them, which likely reflects decreased levels of Dm/Myc or other general transcription factors [32].
Similar to cohesin depletion, PRC1 depletion also increases the fraction of cells in G2 without decreasing the proportion of S phase cells (Figure S6). Depletion of Ph, similar to Pc depletion [13] does not reduce viability, but causes BG3 cells to form more distinct colonies, with longer cellular processes, and cessation of proliferation after six to seven days of treatment. We thus examined the effects of Ph on Rad21 binding after five days of RNAi treatment, when Ph is reduced by at least 80% (Figure S2) and cells are still dividing, to minimize binding changes that might be caused by cell differentiation.
Changes in Pc binding upon Rad21 depletion were measured by two methods (Figure S7), both of which showed that Rad21 depletion reduces Pc binding to active genes. In the first method, we integrated the Pc ChIP MAT scores for all microarray features within a gene body in the control and Rad21-depleted cells, and then calculated the difference in the total ChIP signal for each gene between the depleted and control cells (Figure S7, method 1). By comparing these differences to the control mRNA level for each gene (Table S3), we found that Pc binding decreases at many active genes upon Rad21 depletion (Figure 2B). We also measured the change in Pc ChIP enrichment after Rad21 depletion at each individual microarray feature (ΔPc) across the genome, and then mapped all locations where ΔPc was greater or smaller than the median genome-wide ΔPc by at least two standard deviations for at least three microarray features in a row against all annotated genes (Figure S7, method 2). These intervals are shown in the Δ tracks in Figure 3 and Figure S8. We found that 50% of active cohesin-binding genes show a Pc decrease, compared to only 10% of the genes that don't bind cohesin, and that Pc increases are 10-fold less frequent than increases at cohesin-binding genes (Figure 2C).
The finding that cohesin depletion frequently decreases but rarely increases PRC1 binding to active cohesin-binding genes argues that most of the changes in PRC1 binding are a direct consequence of cohesin reduction at the cohesin-binding gene, and not caused indirectly by cellular differentiation. Differentiation would be expected to have more selective, gene-specific effects, and would also be more likely to increase PRC1 at some active genes. Together with the direct interaction between cohesin and PRC1 (Strübbe et al., 2011; Figure S2), the decrease in PRC1 levels at many cohesin-binding genes argues that cohesin recruits or stabilizes PRC1 binding at active genes. Because cohesin depletion alters transcription of many genes, we cannot rule out the possibility that the PRC1 decreases upon cohesin depletion are caused in part by transcriptional changes. We note, however, that the effects of cohesin reduction on PRC1 binding do not fully mirror the effects on Pol II occupancy. Although cohesin depletion frequently reduces both total Pol II (Rpb3) and Pc at cohesin-binding genes, the genome-wide correlation between the change in Pc (ΔPc) and change in Rpb3 (ΔRpb3; [32]) is modest (r = 0.29).
In addition to the effects of cohesin depletion on PRC1 binding, we found that PRC1 depletion affected cohesin binding at active genes. Ph depletion increased Rad21 signals at nearly a third of active cohesin-binding genes (Figure 2B, Figure 2C, Figure 3, Figure S8). Cohesin increases were twice as frequent as decreases (Figure 2C), and increases were generally greater in magnitude than decreases (Figure 2B). These findings suggest that although cohesin facilitates PRC1 binding, PRC1 negatively influences cohesin binding at many active genes. We cannot rule out, however, the possibility that some increases in cohesin signals are caused by greater epitope accessibility in the absence of PRC1, or increased Pol II levels in the gene body (see below).
In contrast to active genes, cohesin and PRC1 binding are mutually antagonistic at PcG-silenced genes that are marked by H3K27me3 and bind Pc. Rad21 binding increases at 28% of these genes upon Ph depletion, and Pc binding increases at 60% of them upon Rad21 depletion (Figure 2B,C). The Rad21 increases upon Ph depletion, and the Pc increases upon Rad21 depletion occurred at all the genes in the Antennapedia and bithorax HOX gene complexes (ANT-C, BX-C). The changes at the Sex combs reduced (Scr) and Antennapedia (Antp) genes in the ANT-C are shown in Figure 3.
The increased cohesin binding is not caused by higher cohesin expression, because Ph depletion slightly reduces cohesin subunit mRNA levels, as determined from a genome-wide mRNA analysis described below. The increase in Rad21 at silenced genes, however, correlates with an increase in Pol II and mRNA for the silenced genes (see below), and thus increased cohesin binding to silenced genes could be caused by increased transcription. The increased Pc levels at silenced genes upon Rad21 depletion are unlikely to reflect higher PRC1 expression. Rad21 depletion increases the Psc and Su(z)2 mRNAs less than 2-fold, and has no significant effect on the expression of other PRC1 subunits [13]. Also, as described above, Rad21 depletion decreases PRC1 levels at active genes. The increase in Pc at silenced genes upon Rad21 depletion does not reflect a change in the expression of the silenced genes, because they remain silenced.
Increases in Pc and Rad21 binding at silenced genes upon cohesin and PRC1 depletion were substantially more frequent than decreases. Pc increases were detected at 60% of the H3K27me3-Pc genes upon Rad21 depletion, but decreases were detected at only 2% (Figure 2C). At the H3K27me3-Pc genes that bind cohesin, some of which are not fully silenced, Pc increases were 2.5-fold more frequent than decreases upon cohesin depletion, which is opposite to what occurs at cohesin-binding active genes, where Pc decreases are 10-fold more frequent than increases (Figure 2C). Thus Rad21 depletion causes coincident PRC1 decreases at active cohesin-binding genes and PRC1 increases at silenced genes, where cohesin binding is generally restricted to Polycomb Response Elements (PREs; [12]). PRC1 binding may increase at silenced genes because it is released from active genes, increasing the amount available for binding. We cannot rule out the possibility, however, that cohesin directly competes with PRC1 for binding at genes marked by H3K27me3, but that under normal conditions, the cohesin levels fall below the threshold for detection by ChIP. It is unlikely that cohesin is epitope-masked at silenced genes, because it was not detected outside of PREs at silenced genes with antibodies against Smc1, SA, Rad21 and Nipped-B [12].
The antagonism between cohesin and PRC1 binding at PcG-silenced homeotic genes in the ANT-C and BX-C such as Scr and Antp (Figure 3) provides a molecular explanation for dominant suppression of Pc mutant homeotic phenotypes by cohesin mutations [10], [11]. The ectopic sex comb phenotype of heterozygous PRC1 mutants results from reduced silencing of Scr, and the abdominal segment transformations are caused by derepression of genes in the BX-C. We further confirmed the in vivo genetic antagonism by testing the effects of cohesin and releasin mutations on the phenotypes displayed by several PRC1 mutants. Smc1 and Rad21 cohesin mutations suppressed these phenotypes and pds5 releasin mutations enhanced them, supporting the idea that cohesin antagonizes PRC1 function at silenced genes during development (Table S4).
The antagonism between cohesin and PRC1 binding at silenced genes predicts that heterozygous cohesin mutants would show phenotypic transformations opposite to those exhibited by PRC1 mutants because they would increase PcG silencing of homeotic genes. Indeed, we find that adult male flies heterozygous for Rad21 (vtd) loss-of-function mutations exhibit mild and weakly penetrant posterior to anterior abdominal transformations resulting in lighter pigmentation of abdominal segment A5, which is an A5 to A4 transformation opposite to the A4 to A5 anterior to posterior transformation caused by PRC1 subunit mutations (Figure S9). The penetrance and extent of this transformation is increased by heterozygous Nipped-B kollerin mutations (Figure S9). This transformation indicates increased silencing of genes in BX-C, consistent with the increased PRC1 levels at these genes upon cohesin depletion in BG3 cells.
Cohesin selectively binds active genes in which transcriptionally-engaged RNA polymerase II (Pol II) pauses several nucleotides downstream of the transcription start site [31], [32]. One mechanism by which cohesin controls gene transcription is influencing the transition of paused Pol II to elongation. Cohesin depletion decreases the levels of total and elongating phosphorylated Pol II in the bodies of most cohesin-binding genes in BG3 cells, indicating that it often facilitates transition of paused Pol II to elongation [32]. Cohesin also hinders transition of paused Pol II to elongation at genes that are strongly repressed by cohesin, which include those rare genes such as inv and en in BG3 cells that have an extensive cohesin - H3K27me3 overlap [31], [32].
Because cohesin facilitates PRC1 binding to active genes, we tested if PRC1 participates in the control of the transition of paused Pol II to elongation at active genes by mapping Pol II genome-wide before and after depletion of Ph in BG3 cells. We performed ChIP for Pol II subunit Rpb3 to measure total Pol II. Paused Pol II transitions to elongation after phosphorylation by P-TEFb [33], and thus we also conducted ChIP for the Pol II Rpb1 subunit phosphorylated on the serine 2 residues of the heptapeptide repeats in the C terminal domain (Ser2P Pol II) to detect elongating Pol II. We correlated the effects of Ph depletion on Pol II occupancy with changes in mRNA levels measured by microarrays (Table S3). As described below, the results show that PRC1 influences transition of paused Pol II to elongation, but the effects of PRC1 depletion differ from those of cohesin depletion, indicating that cohesin has roles in controlling Pol II activity at active genes beyond facilitating PRC1 binding. They also confirm that PRC1 inhibits transcription of PcG-silenced genes.
Plotting the change in Rpb3 levels versus the change in Ser2P Pol II levels shows that Ph depletion increased both Rpb3 and Ser2P Pol II at many PcG-silenced genes marked by H3K27me3, consistent with the idea that PRC1 is essential for PcG-silencing (Figure 3; Figure 4A; Figure S8). The increase in Ser2P Pol II is often accompanied by an increase in mRNA (Figure 4B,D).
By contrast, at a large fraction of active genes, which lack H3K27me3, Ph depletion increased total Pol II (Rpb3) in the gene body, but decreased Ser2P Pol II (Figure 3; Figure 4A; Figure S8). The genes with increased total Pol II and decreased Ser2P Pol II often show decreased mRNA (Figure 4B,C). Total Pol II increases and Ser2P Pol II decreases were much more frequent at cohesin-binding genes (40–50%) than at genes that lack cohesin (12–15%), and at Pc-binding genes (30–40%) than at genes that lack Pc (8–15%) (Figure S10).
Together, the findings that the total Pol II increases and Ser2P decreases upon Ph depletion (a) occur at a large fraction of active cohesin and Pc-binding genes, (b) happen more rarely at genes that don't bind cohesin, and (c) opposite changes (Ser2P Pol II increases or Rpb3 decreases) are rare, are compelling evidence that most of these Pol II changes are direct, and not caused indirectly by altered cell identity. Changes in cellular identity would be expected to affect a smaller fraction of active genes, and to alter Pol II occupancy more frequently in the opposite direction at cohesin and PRC1-binding genes and at non-binding genes.
In summary, although total Rpb3 increased at both silent and active genes upon PRC1 depletion and rarely decreased, Ser2P Pol II and mRNA often increased at PcG-silenced genes and frequently decreased at active genes (Figure 4A–D). The increase in total Pol II and decrease in Ser2P Pol II at many cohesin-binding active genes upon Ph PRC1 depletion argues that PRC1 blocks the release of non- or under-phosphorylated Pol II into elongation. Increases in total Pol II occurred 2.5-fold less frequently at promoters than in gene bodies upon Ph depletion, and the fold-increases in Pol II at promoters are also smaller (Figure S10). This argues that in most cases, transition to elongation, not Pol II recruitment or transcription initiation, is the key step regulated by PRC1 at active genes. Because promoter Pol II increases are more frequent than decreases, there may be modest effects on recruitment or initiation at some genes. The idea that PRC1 controls primarily the transition to elongation predicts that the pausing index, which is the ratio of total Pol II at the promoter to the total Pol II in the gene body will decrease upon Ph depletion. It also predicts that the ratio of phosphorylated Pol II to total Pol II in the gene body will decrease. Indeed, measuring the pause index and Ser2P to total Pol II ratio for all genes confirms global decreases in both cases, with a particularly dramatic decrease in the phosphorylated to total Pol II ratio in gene bodies (Figure 4E,F). The strong decrease in the fraction of phosphorylated Pol II is consistent with the observed decreases in mRNA, because Pol II phosphorylation is required for association of elongation and RNA processing factors with elongating Pol II [34]. The under-phosphorylated Pol II that enters the gene bodies upon PRC1 depletion may have reduced processivity, and the nascent RNA is also unlikely to be efficiently spliced or polyadenylated.
The effects of PRC1 depletion on Pol II activity at cohesin-binding genes differ substantially from those of cohesin depletion. Upon cohesion depletion, total Pol II and Ser2P Pol II levels usually change in the same direction, not in the opposite direction, with decreases being substantially more frequent than increases [32]. Thus the increase in total Pol II in the bodies of cohesin-binding genes upon PRC1 depletion likely requires cohesin, consistent with the finding that cohesin frequently facilitates the transition of paused Pol II to elongation [32]. Because the effects of cohesin depletion cannot be uncoupled from changes in transcription, we cannot rule out the possibility that the decreases in PRC1 upon cohesin depletion are caused in part by transcriptional changes instead of reduced recruitment or stabilization of PRC1 binding. We note, however, that the effects of cohesin reduction on PRC1 binding do not fully mirror the effects on Pol II occupancy, with the effects on PRC1 being more strongly skewed to the promoter. For instance, Rad21 depletion decreases Rpb3 at 12% of all active promoters [32] but decreases Pc on 18% of promoters, even though the starting Rpb3 signals are generally higher than Pc signals at promoters. In contrast, Rad21 depletion decreases Rpb3 in 50% of all active gene bodies, but Pc only in 35%. Thus reductions in Pol II levels at the promoter cannot explain all cases of reduced PRC1.
In anterior wing disc (Figure 1) and cultured Sg4 cells [13], the inv-en complex is PcG-silenced and has high H3K27me3 and PRC1, but low cohesin. In the posterior wing disc, inv-en is expressed, with low H3K27me3, but high cohesin and PRC1 (Figure 1). In BG3 cells, the inv-en complex has a different cohesin-PcG structure than either the silenced or active states [13]. H3K27me3 marks the entire complex, including the large flanking regulatory region, but inv and en and the regulatory region between them also bind cohesin (Figure 5A). In these cells, inv and en mRNAs are present at modest levels and increase dramatically upon cohesin or PRC1 depletion [13], [31], [32]. Thus the inv-en complex in BG3 cells has a cohesin-PcG “restrained” state distinct from both the silenced and active states.
Cohesin depletion can reduce long-range looping between enhancers and promoters, and between CTCF-binding sites in mammalian cells [4], [35]. We thus considered the idea that part of the mechanism by which cohesin represses inv-en in BG3 cells is by facilitating looping between the Polycomb Response Elements (PREs) that recruit PcG complexes. There is a bipartite PRE just upstream of the en promoter, and another upstream of inv (Figure 5; [36]). Other regulatory sequences include enhancers upstream of each gene, between the genes, and throughout the long region extending from en to the tou gene [37].
We used chromosome conformation capture (3C) [38] to determine if the two PRE-containing regions in inv-en interact in BG3 cells, and if cohesin depletion reduces these interactions. We used four anchor restriction sites in a 180-kilobase region encompassing the inv and en genes, the regulatory regions, and flanking DNA (Figure 5). Two anchors (a, d) are outside the complex, and one (c) includes the en promoter and PREs. The fourth (b) is upstream of inv in the region containing a PRE.
Figure 5A shows that the en PRE-promoter anchor (c) interacts with the inv PRE region. Similarly, the anchor near the inv PRE (b) interacts with the en PRE and with the region between inv and en, which harbors tissue-specific enhancers [37]. Both PRE anchors (b, c) also interact with the ends of the inv-en complex demarcated by the ends of the H3K27me3 domain. In contrast, the external anchors (a, d) interact only with their immediately surrounding regions, and not with sites in the complex.
Upon depletion of the Rad21 cohesin subunit, which dramatically increased inv and en mRNA levels 20 to 40-fold [13] interaction between the inv and en PREs decreased substantially (Figure 5A). Interaction of the inv PRE-containing region with the enhancer-containing region between inv and en also decreased, although the enhancer-containing region has lower, but significant cohesin levels. It may be that this interaction is stimulated secondarily by PRE-PRE looping. Facilitation of PRE-PRE looping by cohesin may explain why cohesin represses inv-en in BG3 cells. We cannot formally exclude, however, the possibility that cohesin represses by a different mechanism, and that the increased transcription caused by cohesin depletion reduces the PRE-PRE interaction.
In Drosophila Sg4 cells in which inv-en is fully PcG-silenced, cohesin binds only at the two PREs, and there is no detectable Pol II or transcripts (Figure 5B; [12], [13]). We see interactions between the PREs as in BG3 cells, indicating that the PRE-PRE interactions seen in BG3 cells are consistent with PcG silencing (Figure 5B). Interaction of the inv PRE with the end of the regulatory domain is lower than in BG3 cells. We were unable to test if cohesin depletion reduced the PRE-PRE interactions in the fully silenced state because Sg4 cells are relatively refractory to RNAi treatment [13]. However, cohesin also occupies all known PREs in the silenced BX-C in BG3 cells [12], consistent with the possibility that cohesin can facilitate PRE-PRE looping at silenced genes.
Our genome-wide studies in developing wing discs and cultured cells reveal three distinct cohesin-PcG states at their target genes in Drosophila, all of which occur at the invected-engrailed (inv-en) complex in different cell types. In Sg4 cells [12], [13] and anterior wing disc, inv-en is PcG-silenced and has high H3K27me3 showing PRC2 activity, high PRC1, and low or undetectable cohesin, except at the PREs (Figure 6, PcG silenced). This pattern also occurs at virtually all PcG-silenced genes in wing discs and BG3 cells. In this state, cohesin and PRC1 binding are mutually antagonistic in BG3 cells, consistent with the in vivo genetic antagonism between cohesin and PRC1 in determining segmental identity. Reducing cohesin increases silencing, while reducing PRC1 decreases silencing.
In BG3 cells, inv-en has the rare restrained state in which H3K27me3 (PRC2), PRC1 and cohesin overlap over long extended regions of several kilobases (Figure 6, cohesin – PcG restrained). The genes are not silenced, but depletion of either cohesin or PRC1 causes a large increase in transcription and mRNA [13], [31], [32]. In this case, cohesin inhibits transition of paused Pol II to elongation at a step distinct from those controlled by the DSIF and NELF pausing factors [31], [32]. As shown here, cohesin depletion reduces interactions between the two inv-en PREs in this state, suggesting that it represses by facilitating PRE-PRE looping. Although there are other genes with the cohesin-PcG restrained state in BG3 and Sg4 cells [13], there is currently no unambiguous evidence that this state occurs in vivo. It is possible that this rare state occurs at genes that are in transition from a silenced to active state or vice versa, but it may also be a stable state that serves to moderate gene expression.
Unexpectedly, we also found that PRC1 is present at the active inv-en complex in posterior wing disc, and at most other active cohesin-binding genes in both developing wing disc and BG3 cells (Figure 6, cohesin-PRC1 active). These genes lack H3K27me3, indicating that PRC2 is not present or active. Cohesin facilitates PRC1 binding to active genes in BG3 cells, and PRC1 depletion decreases the ratio of phosphorylated to total Pol II in the gene body and reduces mRNA levels. This finding greatly expands the known gene regulatory functions of both cohesin and PRC1 and provides new insights into the mechanisms by which they control transcription.
As outlined in Figure 6, we propose that the three cohesin-PcG states are functionally linked. We posit that cohesin facilitates binding of PRC1 to active genes via direct interactions, and that PRC1 then inhibits the transition of non-phosphorylated paused Pol II to elongation. We favor the idea that binding of PRC1 to active genes sequesters enough PRC1 at active genes to limit the amount of PRC1 available for binding to PcG-silenced genes targeted by PRC2. Although there are other possibilities, this explains why cohesin depletion simultaneously reduces PRC1 binding to active genes and increases binding to silenced genes. The cohesin-mediated sequestration of PRC1 at active genes can also explain the genetic antagonism between cohesin and PRC1 in silencing of the Scr gene and the BX-C in vivo. These and alternative ideas are discussed in more detail below.
PcG-silenced genes are targeted by both PRC1 and PRC2, and generally do not bind measurable levels of cohesin except at PREs, but cohesin reduction increases Pc levels at PcG-silenced genes and PRC1 depletion increases cohesin binding. This antagonistic relationship contrasts with the functional interactions between PcG proteins and the Trithorax (Trx) protein, which binds PREs of silenced genes and is generally antagonistic to PcG silencing, but whose knockdown does not alter PRC1 binding [25].
Multiple mechanisms, which are not mutually exclusive, could contribute to the antagonism between PRC1 and cohesin association with PcG-silenced genes. Upon Ph PRC1 subunit depletion, increased transcription could facilitate cohesin binding. It might be expected, based on their direct interaction, that PRC1 would recruit cohesin to silenced genes, but PRC1 also has negative effects on cohesin binding at many active genes, and the presence of PRC2 at silenced genes may enhance these effects. For example, binding of Pc to H3K27me3 at silenced genes could alter PRC1 conformation such that the PRC1-cohesin contacts inhibit cohesin loading by kollerin, and/or facilitate cohesin removal by releasin.
Cohesin depletion increases PRC1 levels at PcG-silenced genes, even though cohesin binding is very low at these genes outside of the PREs. One idea is that a reduction in cohesin dosage could release PRC1 from active genes, thereby making more PRC1 available to bind to silenced genes (Figure 6). Alternatively, cohesin and PRC1 binding could be dynamically competitive at silenced genes, with the competition favoring PRC1, making it difficult to detect cohesin. Both models are consistent with the finding that a dominant wapl releasin mutation, which globally increases and stabilizes cohesin binding, causes similar mutant phenotypes as PRC1 mutations, indicating reduced PRC1 at Scr and other silenced genes [9]. Higher cohesin levels at active genes in the releasin mutant could sequester more PRC1, making less available for binding to silenced genes. It more difficult to explain how more cohesin binding at silenced genes could directly reduce PRC1 binding in this mutant. One possibility is that binding of the Pc PRC1 subunit to H3K27me3 at silenced genes alters PRC1 conformation in a manner that allows cohesin to stimulate PRC1 removal instead of facilitate its binding.
Because cohesin directly interacts with PRC1 and cohesin depletion reduces Pc binding at active genes, we posit that cohesin directly recruits and/or stabilizes PRC1 binding to active genes. As discussed below, our studies reveal that PRC1 controls Pol II phosphorylation and transition of paused Pol II to elongation at active genes. This may be similar to the function of PRC1 at some silenced and bivalent genes. Although the mechanisms by which PcG complexes repress transcription are not well understood, in some contexts, they inhibit transition of paused transcriptionally-engaged Pol II to active elongation [39]–[41].
Our findings suggest that at active genes, PRC1 facilitates phosphorylation of Pol II to the Ser2P Pol II elongating form and/or blocks entry of non-phosphorylated paused Pol II into elongation. Cohesin preferentially binds genes with promoter-proximal paused Pol II, and often facilitates, or less frequently, inhibits transition of paused Pol II to elongation [31], [32]. Transition of paused Pol II to elongation requires phosphorylation of the NELF and DSIF pausing factors and Pol II by P-TEFb [33]. At most cohesin-binding genes, cohesin depletion reduces the levels of both total and phosphorylated Pol II in the gene body [32]. Here we find that PRC1 depletion, similar to cohesin depletion, reduces Ser2P Pol II at many active genes, but opposite to cohesin depletion, increases total Pol II in the same gene bodies. Thus both the global pausing index and ratio of phosphorylated Pol II to total Pol II in gene bodies decrease upon PRC1 depletion. This argues that Pol II enters into elongation with inadequate phosphorylation. One possible mechanism could be that PRC1 helps NELF and DSIF pausing factors restrain Pol II at the promoter until it is fully phosphorylated.
The decrease in the ratio of phosphorylated to total Pol II upon PRC1 depletion is accompanied by a decrease in mRNA at many genes. This argues that the under-phosphorylated Pol II is less processive, and/or that the nascent RNA is not efficiently processed to mRNA. This is expected, because Pol II phosphorylation, in addition to facilitating transition to elongation, is required for binding of RNA processing factors to the transcription complex [34].
The fact that PRC1 depletion affects Pol II activity and mRNA levels at a large fraction of active cohesin-binding genes confirms that PRC1 is present and plays an important role at these genes. The finding that PRC1 directly controls transcription of many active genes is surprising because it is generally thought that PRC1 functions primarily at PcG-silenced genes. There is, however, published evidence consistent with our findings. Müller and colleagues used chromatin from mixed imaginal discs to perform Ph ChIP-chip with a different antibody than the one used here [28], [42]. Although the mixture of many different cell types creates false overlaps, close examination of their data reveals that Ph associates with many constitutively active genes, including Act5C and some ribosomal protein genes, as we found in both wing discs and BG3 cells. Moreover, using different Pc, Psc and Ph antibodies, Paro and colleagues found that PRC1 preferentially binds promoters with paused RNA polymerase in cultured S2 cells [43]. Although the role of PRC1 in paused Pol II function was not investigated, this is consistent with the selective association of cohesin with genes that have paused Pol II [31], [32] and our finding here that cohesin aids binding of PRC1 to active genes.
There is also evidence that PcG complexes regulate some active genes in mammalian cells. In differentiated T helper cells, both PRC1 and PRC2 components are required for transcriptional activation of cytokine genes, although they repress HOX genes in the same cells [44]. Some expressed genes exhibit H3K27me3, and the Ezh1 alternative methyltransferase is in PRC2 complexes that promote transcription during myogenic differentiation [45], [46]. Ezh1 knockdown decreases Ser2P Pol II at the genes that it binds, similar to our findings with Ph depletion. We did not find any evidence, however, that PRC2 promotes transcription of active genes in Drosophila, which unlike mammals, has only one E(z) methyltransferase.
The work presented here clearly establishes functional connections between cohesin and PRC1, and suggests mechanisms for how cohesin and PRC1 control each other's activities and gene transcription. Our finding that three different cohesin-PcG states can exist at the invected-engrailed gene complex in different cells raises the question of how these various states arise. The presence or absence of PRC2 is likely to be one important factor that determines whether cohesin and PRC1 compete or collaborate, but the factors responsible for establishing these distinct states and controlling transitions between them remain to be determined.
Drosophila was cultured and genetic crosses were conducted at 25° as previously described [17]. The w1118; P{en2.4-GAL4}e16E, P{UAS-myr-mRFP}1, P{NRE-EGFP.S}5A line [21] was used for preparing chromatin from anterior and posterior wing imaginal discs. Stocks were obtained from the Bloomington stock center.
Nipped-B, pds5, and cohesin mutant alleles have been described previously [10], [17], [47]. PcG mutant stocks were obtained from the Bloomington stock center and Rick Jones (SMU). At least 30 male progeny from crosses between cohesin and PcG mutants were scored. Sex comb bristles on first, second, and third legs were counted, and abdominal transformations were scored using a defined arbitrary scale.
Chromatin was prepared from late third instar Oregon-R wing imaginal discs according to Papp and Müller [48], without dialysis. Sixty to seventy-five discs were used per immunoprecipitation, or 100 to 120 anterior or posterior segments. Chromatin preparation from BG3 cells, immunoprecipitation, and Affymetrix Drosophila 2.0R genome tiling microarray hybridization were as previously described [12]. Rick Jones (SMU) generously provided Pc antibodies (Pc-RJ), and Vincenzo Pirrotta (Rutgers) kindly provided Pc (Pc-VP), Ph, and Psc antibodies. Renato Paro (ETH Zürich), Jürg Müller (Max Planck Institute of Biochemistry), and Giacomo Cavalli (Institute of Human Genetics) provided additional Pc and Ph antibodies. Karen Adelman (NIEHS) generously provided Rpb3 antibodies. Ser2P Pol II and H3K27me3 antibodies were purchased from Abcam (ab5095, ab6002) and 8WG16 Pol II antibody was purchased from Covance (MMS-126R).
MAT software [19] was used to calculate ChIP enrichment. MAT scores measure enrichment relative to an input control and scale linearly with the log2 IP/control ratio. MAT has been experimentally demonstrated to be more sensitive and quantitative than other algorithms for measuring ChIP enrichment with Affymetrix tiling arrays, providing sensitivity equivalent to ChIP-seq at a density of 1 aligned read per genome base pair [18], [49]. MAT uses probe sequence to perform within-array normalization, avoiding assumptions associated with quantile normalization.
Bed files showing significant enrichment of proteins mapped by ChIP-chip (Rad21, Nipped-B, H3K27me3, Pc-RJ, Pc-VP, Ph, Psc, Rpb3, Ser2P Pol II) at statistical thresholds of p≤10−3 or p≤10−2 were generated using MAT. Genes bound by a given protein were determined using the bed files and FlyBase gene annotations (www.flybase.org, v5.28) with the Intersect tool of Galaxy/Cistrome [50] or BEDTools [51]. Venn diagrams were generated using eulerAPE (www.eulerdiagrams.org/eulerAPE/). Gene-based ChIP-chip signal data was extracted from MAT score files and aligned to gene expression data using custom programs. Data was analyzed using Microsoft Excel and R ([52]; http://www.R-project.org). For some analyses, differences in integrated ChIP signals (MAT scores) over each annotated gene were calculated (Figure S6, method 1). For others, increases and decreases in Rad21, Pc, Rpb3 and Ser2P binding were measured by the difference in the ChIP MAT scores at each measured point in the genome between the RNAi treated samples and the controls (Figure S6, method 2). Increases or decreases ≥2 standard deviations from the genome-wide median difference that extend for at least three contiguous microarray features (105 bp) were used to generate bed files that were matched with gene annotations using BEDTools [51]. Examples of the increase (+) and decrease (−) bed files are shown in Figure 3. Based on changes in Pol II occupancy upon cohesin depletion, these methods agree closely with results obtained by ChIP-qPCR at selected genes and genome-wide PRO-seq analysis [31], [32].
Total RNA from wing discs or BG3 cells was isolated using Zymo ZR RNAi MicroPrep columns (Zymo Research). Genome-wide analysis of wild-type wing disc mRNA with four biological replicates was performed using Affymetrix Drosophila GeneChip 2.0 microarrays as previously described [13]. Genome-wide measurements of three biological replicates of control BG3 cells and BG3 cells treated with Ph RNAi (see below) for five days was conducted in the same manner.
BG3 cells were cultured and proteins were RNAi-depleted for Rad21, Nipped-B or Ph as described [13]. The double-stranded RNAs used for Rad21 and Nipped-B RNAi were as described [13]. Two double-stranded RNAs targeting Polyhomeotic were used in concert; the constructs target regions of homology between polyhomeotic proximal and polyhomeotic distal. Primer sequences are as follows: Ph RNAi A, forward: TAATACGACTCACTATAGGGAGAAGCCATCAGCACCATGTCGC, reverse: TAATACGACTCACTATAGGGAGACGTAATTTCCGCCAGCGAATC. Ph RNAi B, forward: TAATACGACTCACTATAGGGAGATGCCCATTGATTCGCCCAAG, reverse: TAATACGACTCACTATAGGGAGATGCAACTTGTGGTAAAGGTGCC. These targets were designed using tools in the Drosophila RNAi Screening Center (DRSC) website (www.flyrnai.org/) to avoid off-target effects. All RNAi-treated chromatin and 3C samples were collected 5 days after RNAi treatment.
3C was conducted using a modification of the strategy outlined by Miele et al. [53]. 108 BG3 or Sg4 cells were collected, washed with phosphate buffered saline, pH 7.6 (PBS) and then with hypotonic buffer [10 mM HEPES pH 7.9, 50 mM NaCl, 1 mM dithiothreitol (DTT), 10 mM MgCl2, protease inhibitor cocktail (cOmplete Mini, EDTA-Free Protease inhibitor tablets, Roche)]. Cells were lysed in hypotonic buffer containing 0.35 M sucrose and 0.2% NP-40, vortexed for one minute, and immediately cross-linked with 1% formaldehyde for 10 minutes at room temperature. Cross-linking was stopped by adjustment to 135 mM glycine, and nuclei were isolated by centrifugation onto a 0.8 M sucrose cushion. Nuclei were washed in EcoRI restriction enzyme buffer, and 0.1% SDS was added to extract free protein. 1% Triton X-100 was used to sequester SDS, and EcoRI digestions (500–700 units) were performed overnight at 37°. SDS was added to sequester EcoRI, and the digested DNA was ligated at 18° for 2 hours at a concentration of approximately 4 ng per microliter. Cross-linking was reversed by Proteinase K incubation at 65° for 6 hr or overnight, and DNA was isolated by sequential phenol, phenol-chloroform, and chloroform extractions, followed by precipitation with 0.5 volumes of 7.5 M ammonium acetate and 2.5 volumes of ethanol. DNA was dissolved in TE, digested with 500 ng per microliter RNAse A for 30 min at 37° and used for RT-PCR quantification.
Digestion and religation of BAC DNA containing the entire inv-en locus was used as a normalization control (BACR13F13; BACPAC Resources, Oakland, CA). This template was used to determine the amplification efficiency of each primer pair. EcoRI digestion efficiency was confirmed to be over 85% in several experiments using PCR primer pairs spanning the EcoRI sites. When the experiment was conducted without formaldehyde cross-linking, we were unable to detect 3C ligation products except between two immediate adjacent EcoRI sites (likely resulting from incomplete digestion).
Genome-wide ChIP and expression data generated for this study is deposited in the GEO database (accession no. GSE42106).
|
10.1371/journal.pntd.0006599 | Variation in competence for ZIKV transmission by Aedes aegypti and Aedes albopictus in Mexico | ZIKV is a new addition to the arboviruses circulating in the New World, with more than 1 million cases since its introduction in 2015. A growing number of studies have reported vector competence (VC) of Aedes mosquitoes from several areas of the world for ZIKV transmission. Some studies have used New World mosquitoes from disparate regions and concluded that these have a variable but relatively low competence for the Asian lineage of ZIKV.
Ten Aedes aegypti (L) and three Ae. albopictus (Skuse) collections made in 2016 from throughout Mexico were analyzed for ZIKV (PRVABC59—Asian lineage) VC. Mexican Ae. aegypti had high rates of midgut infection (MIR), dissemination (DIR) and salivary gland infection (SGIR) but low to moderate transmission rates (TR). It is unclear whether this low TR was due to heritable salivary gland escape barriers or to underestimating the amount of virus in saliva due to the loss of virus during filtering and random losses on surfaces when working with small volumes. VC varied among collections, geographic regions and whether the collection was made north or south of the Neovolcanic axis (NVA). The four rates were consistently lower in northeastern Mexico, highest in collections along the Pacific coast and intermediate in the Yucatan. All rates were lowest north of the NVA. It was difficult to assess VC in Ae. albopictus because rates varied depending upon the number of generations in the laboratory.
Mexican Ae. aegypti and Ae. albopictus are competent vectors of ZIKV. There is however large variance in vector competence among geographic sites and regions. At 14 days post infection, TR varied from 8–51% in Ae. aegypti and from 2–26% in Ae. albopictus.
| Aedes aegypti is an efficient vector for arboviruses which is partially determined by its vector competence (VC) status, a highly variable trait. Based on previous reports, the VC of New World mosquitoes is limited for ZIKV. However, a VC assessment from additional geographical sources was lacking. Therefore, we evaluated the VC for ZIKV using recently colonized mosquitoes from Mexico. Aedes aegypti and Ae. albopictus were highly susceptible to ZIKV infection but varied greatly in their transmission rates. As with previous studies of other flaviviruses; VC for ZIKV was highly variable in both Ae. aegypti and Ae. albopictus but was greatest in Ae. aegypti, supporting its role as the main ZIKV vector.
| Zika virus (ZIKV, Flavivirus, Flaviviridae) was first isolated from a febrile sentinel rhesus macaque in the Zika forest of Uganda in 1947 and later in 1948 from Ae. africanus mosquitoes from the same area [1]. ZIKV circulated in Africa and Asia without much attention until 2007 when a major outbreak occurred in the Pacific Island of Yap in the Federate States of Micronesia [2, 3]. Outbreaks were later reported in other Pacific islands: French Polynesia, Easter Island, the Cook Islands and New Caledonia during 2013–2014 [4–6]. Making its arrival to the Americas in early 2015, ZIKV circulation was confirmed in Brazil in May and, as expected, ZIKV spread quickly to areas where the vectors were present. Mosquito-borne transmission has been reported in 48 countries of the Americas since its introduction [7]. In addition, ZIKV was associated with congenital abnormalities such as microcephaly and an increased incidence of Guillain-Barré syndrome, and was thus declared a Public Health Emergency of International Concern by the World Health Organization on February 1, 2016 [8], which ended nine months later [9]. Since its introduction, the Pan American Health Organization has reported more than a thousand cumulative Zika cases in the Americas. Mexico alone had a total of 129 cases [10], with its first case of congenital ZIKV syndrome in November of 2016 [11].
The main mechanism of ZIKV transmission in epidemic and endemic areas is through the bite of an infectious mosquito, with Ae. aegypti apparently serving as the primary vector [12]. From the screening of wild-caught mosquitoes in Mexico, ZIKV RNA has been detected in Ae. aegypti pools collected in and around houses of suspected ZIKV cases [13]. Aedes albopictus [14] have also been confirmed to be infected with ZIKV.
Vectorial capacity is a quantitative measure of the potential of an arthropod vector to transmit a pathogen. It is defined as the average number of potentially infective bites that will ultimately be delivered by all the vectors feeding on a single host in a day [15]. Vectorial capacity is impacted by extrinsic factors like vector density, vector longevity, length of the extrinsic incubation period (EIP) and blood feeding behavior [16, 17] and also by intrinsic factors like vector competence (VC). VC is defined as the intrinsic ability of an arthropod vector to acquire, maintain and eventually transmit a pathogen [18]. Upon ingestion, the arbovirus has to replicate to be transmitted to a susceptible host in a subsequent feeding episode. However the virus has to first bypass a series of physiological and anatomical barriers [19]. Briefly, upon entry of the virus into the mosquito midgut through an infectious blood meal, the virus has to establish an infection; if this does not occur the mosquito has a midgut infection barrier (MIB). Next, the virus has to replicate in the midgut and then escape the midgut to disseminate to other tissues. When this does not occur the mosquito is said to have a midgut escape or dissemination barrier (MEB). The virus may infect several mosquito tissues including, most importantly, the salivary glands where it again has to establish an infection. If this is prevented the mosquito has a salivary gland infection barrier (SGIB). Finally, the virus has to replicate and disseminate into the saliva from where it will be expectorated with the saliva while probing and feeding in a susceptible vertebrate host. If this is limited, the mosquito has a salivary gland escape barrier (SGEB) [19, 20]. Consecutively the MIB, MEB, SGIB and SGEB contribute to the overall VC phenotype.
By harvesting mosquitoes at 7 and 14 days post infection (dpi) we can obtain potential indicators of infection and dissemination and/ or transmission respectively [21, 22]. Previous studies have reported low ZIKV transmission rates for the Asian lineage of ZIKV using mosquitoes from a wide geographical range in the New World [23, 24]. We hypothesized that VC is variable and is highly dependent upon the geographic origin of the mosquito populations. Hence, we analyzed the ZIKV transmission potential of 13 recently colonized Aedes collections, 10 of Ae. aegypti and 3 of Ae. albopictus, from different locations across Mexico. These collections were analyzed for ZIKV (strain PRVABC59—Asian genotype) VC at 7 and 14 days dpi.
Herein we report that both Aedes species are competent for ZIKV transmission and that MIB, MEB, SGIB and SGEB vary by species, as well as by collection, region, and whether they were collected north or south of the NVA. TR ranged from 2–51% at 7 dpi and from 8–51% at 14 dpi in Ae. aegypti. Aedes albopictus had from 0–8% transmission at 7 dpi and 2–26% at 14 dpi. We describe the contribution of each of the barriers for ZIKV transmission showing that a SGEB may be an important barrier to ZIKV transmission in Ae. aegypti populations.
Collection protocols were approved by the ethics board at the Universidad Autonoma de Nuevo Leon. Written informed consents were obtained from the household owners for mosquito collections indoor and outside the houses. No special permit was needed for sampling in non-private properties.
Aedes eggs were collected from ovitraps set at different locations in Mexico (Fig 1) during 2016 with exception of the collections from the state of Chiapas (Huehuetan and Mazatan) where immature stages were obtained from at least 20 different containers (S1 Table). All 10 Ae. aegypti collections were analysed from north to south and were further grouped into regions (Northeastern, Yucatan, or Pacific). These regions are defined based upon a past survey of variation in mitochondrial DNA in 38 collections across Mexico [25]. That study indicated that northeastern Mexico collections were genetically differentiated from and had lower genetic diversity than Yucatan and Pacific coastal collections. Regions were further grouped as to whether they were located north or south of the NVA based upon earlier findings [26].
At each location where ovitraps were used, 4–5 were set and checked once a week. The eggs were dried and shipped to the laboratory at Colorado State University (PHS permit no. 2016-06-185), where they were hatched, reared to adults and then identified to species. Larvae were fed ad libitum with a 10% (w/v) liver powder solution. Adult mosquitoes were maintained on sucrose ad libitum and for egg production citrated sheep blood was given once a week through water-jacketed glass feeders using hog gut as a membrane through which to feed. Adults were maintained at insectary conditions (28°C, 70% relative humidity and 12:12 light:dark diurnal cycle). Mosquitoes were identified as Ae. aegypti or Ae. albopictus based on scale patterns on the thorax after adult eclosion [27].
The flow chart in Fig 2 indicates how each of the 13 collections were evaluated for VC using Ae. aegypti collected in Apodaca as an example (raw data in first four rows of S1 Table). The ZIKV strain used for these studies was PRVABC59 (Accession # KU501215) [28] obtained from the CDC. PRVABC59 strain had been passed four times on African green monkey kidney cells (Vero, ATCC CCL-81). For mosquito infections PRVABC59 was used to infect Vero cells at a MOI of 0.01. After 4 days infection, the supernatant was harvested and centrifuged at 3,000xg for 10 min at 4°C. The supernatant was then transferred to a clean tube and a sample was taken to perform ZIKV quantification by quantitative-reverse transcriptase PCR (RT-qPCR) with oligonucleotides for the ZIKV 3’ untranslated region (S2 Table) prior to the infection of mosquitoes. During this time, the supernatant was maintained at 4°C until it was mixed with blood. RNA was extracted from 50 μL of the clarified supernatant using the Direct-zol™ RNA MiniPrep Kit (Zymo Research Corp.) following manufacturer recommendations. Based upon the result, the supernatant was supplemented with Dulbecco’s modified Eagle’s medium (DMEM) and 20% FBS and further mixed 1:1 with defibrinated calf blood to a final concentration of 1 x109 genome equivalents (GE) / mL. Viral titers in the ZIKV infectious calf blood were confirmed by plaque assays on Vero cells, averaging 106 PFU/mL (Fig 2A).
Prior to feeding, 5–6 day old mosquitoes were deprived of sucrose and water for 24 hours. Mosquito infections were performed under BSL-3 containment where they were offered a ZIKV infectious blood meal through water-jacketed glass feeders covered with hog gut. After up to one-hour of feeding, mosquitoes were cold-anesthetized and engorged females were placed into new cartons and water and a sugar source were provided (Fig 2A.)
At 7 and 14 dpi, mosquitoes were cold anesthetized at 4°C. To measure the DIR, mosquito legs and wings were removed and placed into a tube with 250 μL mosquito diluent (1X phosphate buffer saline (PBS) supplemented with 20% heat-inactivated fetal bovine serum (FBS), 50 μg/mL penicillin/streptomycin, 50 μg/mL gentamycin, 2.5 μg/mL fungizone) (Fig 2B) and a stainless steel bead for homogenization. To obtain saliva to measure TR, the mosquito (without legs or wings) proboscis was placed into a capillary tube that contained immersion oil (~5 μL) and allowed to expectorate saliva for 30 minutes. Following salivation, the tip of the capillary tube was broken into a tube containing 100 μL of mosquito diluent. Subsequently, the midgut (to measure MIR) and salivary glands (to measure SGIR) were dissected, rinsed individually in PBS and placed in tubes with mosquito diluent and a stainless steel bead. Forceps were dipped in 70% ethanol and cleaned after each tissue was dissected and between individual mosquitoes. Mosquito tissues were stored at -80°C until further processing (Fig 2B).
Mosquito tissues (midguts, legs/wings and salivary glands) were thawed and homogenized at 25 cycles/second for one minute using a Retsch Mixer Mill MM400 (Germany) and centrifuged at 20,000xg for 5 minutes at 4°C while saliva samples were centrifuged at 20,000 x g for 3 minutes at 4°C, mixed by vortexing and centrifuged for 3 additional minutes. Clarified supernatant was titrated by plaque assay on Vero cells to determine whether individual mosquito tissues contained infectious ZIKV (Fig 2B).
Plaque assays were performed on Vero cells which were maintained in DMEM containing 8% FBS, 50 μg/mL penicillin and streptomycin and 50 μg/mL gentamycin at 37°C with 5% CO2. Twelve-well plates were seeded with Vero cells and allowed to reach 90 to 95% confluency. Media was then removed and replaced with 250 μL of DMEM containing 1% FBS, 50 μg/mL penicillin and streptomycin, and 50 μg/mL gentamycin. Subsequently, each sample (30 μL for mosquito saliva or 70 μL for midgut, salivary glands and legs/wings) was added to one well of the plate. The plates were rocked for 90 minutes to allow absorption after which 1 mL of overlay (tragacanth gum (6 g/L) in 1X DMEM supplemented with 10% FBS, 50 μg/mL penicillin/ streptomycin and 50 μg/mL gentamycin) was added to each well and plates were incubated at 37°C with 5%CO2. After 5 days, the plates were fixed with a staining solution (1 g/L crystal violet in 20% ethanol solution); plaques were visualized on a light box and recorded as plaque positive or negative. We used 10 fold dilutions for virus (blood meal) titration. But for testing the presence of the virus in the mosquito tissues or saliva samples we determined the presence of ZIKV infectious particles by observing plaques from a fixed volume of the sample (Fig 2B).
We determined the rates of midgut infection, dissemination, salivary gland infection and transmission in each of the mosquito populations tested [29, 30]. The dissemination rate (DIR) was defined as the number of mosquitoes with infectious ZIKV in the legs/wings divided by the number of blood-fed mosquitoes (Fig 2B.1). The transmission rate (TR) was defined as the number of mosquitoes whose saliva contained infectious ZIKV divided by the total number of salivary glands that were successfully dissected (Fig 2B.2). Salivary gland infection rate (SGIR) was defined as the number of mosquitoes with infectious ZIKV in the salivary glands divided by number of salivary glands that were successfully dissected (Fig 2B.3). Midgut infection rate (MIR) was defined as the number of mosquitoes with infectious ZIKV in the midgut divided by the total number of mosquitoes that had blood-fed. (Fig 2B.4).
Next we calculated the additive contribution of each of the four transmission barriers to blocking transmission by adjusting MIR, DIR, SGIR and TR so that they sum to 100%.
GraphPad PRISM version 7.03 (GraphPad Software, San Diego, CA, USA) and SigmaPlot 12 (Systat Software, Inc, San Jose, CA) were used for graph construction.
Blood titers were log10 (x+1) transformed (Fig 2C.1). All analyses were replicated at least twice with different (independent) virus preparations for each replicate (S1 Table). MIR, DIR, SGIR, and TR were compared using logistic regression in PROC GLIMMIX in SAS 9.4. The LSMEANS option was used to report mean proportions and their 95% confidence intervals. Main effects in the logistic regression were mosquito generation, dpi, collection, geographic region and whether they were from north or south of the NVA [26]. Blood meal titers (pfu/mL) were compared using Analysis of Variance (ANOVA) in PROC GLM in SAS 9.4 (Fig 2C.3). Pearson’s Correlation Coefficients were calculated using PROC CORR (Fig 2C.3). Proportions (MIR, DIR, SGIR, TR) was normalized with arcsine of the square root prior to ANOVA and correlation analyses.
Due to time and space constraints it was not feasible to measure the four components of vector competence (MIR, DIR, SGIR, TR) simultaneously in all 13 Aedes collections. Instead 6 generations (F1—F6) (S1 Table) of Ae. aegypti and the three generations of Ae. albopictus were required to complete all experiments. The PRVABC59 virus was grown fresh for each round of infections and titers varied significantly from 5.1 to 7.6 logs among the seven generations of Ae. aegypti (ANOVA P = 0.0252) and the three generations of Ae. albopictus (ANOVA P = 0.0203). In Ae. aegypti the blood titer was not correlated with 7 dpi MIR (r = 0.246, P = 0.2832), DIR (r = 0.046, P = 0.8420), SGIR (r = 0.157, P = 0.5210) or TR (r = 0.147, P = 0.525). The same was the case for 14 dpi MIR (r = 0.350, P = 0.1103), DIR (r = 0.314, P = 0.1541), SGIR (r = 0.304, P = 0.1800) or TR (r = 0.415, P = 0.0546). All correlation analyses had n = 22 observations. Even though infectious blood titers varied among the different mosquito generations they did not appear to affect vector competence parameters in Ae. aegypti.
Interestingly these same patterns did not hold for the three Ae. albopictus collections. S1–S4 Figs show that the four VC parameters vary greatly among the three generations. Furthermore, the four VC parameters were correlated with 7dpi MIR (r = 0.693, P = 0.127), DIR (r = 0.905, P = 0.0132), SGIR (r = 0.960, P = 0.0024) and Sal (r = 0.976, P = 0.0009) and at 14 dpi MIR (r = 0.765, P = 0.0452), DIR (r = 0.830, P = 0.0208), SGIR (r = 0.886, P = 0.0079) but not TR (r = 0.545, P = 0.2064). All correlation analyses had n = 7 observations. Both trends suggest that all four VC parameters are confounded by the large variation among the generations of Ae. albopictus measured. For these reasons, results can be seen in S1–S4 Figs but are not considered further.
In Ae. aegypti MIR varied significantly between mosquito generations F1 and F6 but otherwise broadly overlapped (Fig 3). MIR was similar between mosquitoes harvested at 7 or 14 dpi. MIR varied broadly between collections with Ciudad Madero and Monterrey having the lowest MIR and Poza Rica and Mazatan having the highest MIR. MIR was least in northeastern and Yucatan regions and MIR was highest in the Pacific collections. MIR was lowest in collections south of the NVA.
DIR varied significantly between mosquito generations F1, F2 and F6 but otherwise broadly overlapped among the six mosquito generations (Fig 4). Mosquitoes harvested at 7 dpi had a lower DIR than those harvested at 14 dpi. DIR varied broadly between collections with Ciudad Madero and Monterrey having the lowest DIR and Poza Rica and Mazatan having the greatest DIR. DIR was least in northeastern collections, intermediate in the Yucatan and highest in the Pacific collections. DIR was lowest in collections north of the NVA.
SGIR varied significantly between mosquito generations F2 and F6 but otherwise broadly overlapped among the six generations (Fig 5). Mosquitoes harvested at 7 dpi had a lower SGIR than those harvested at 14 dpi. SGIR varied broadly between collections with Ciudad Madero and Monterrey having the lowest SGIR and Poza Rica, Mazatan and Guerrero having the greatest SGIR. SGIR was least in the northeastern region, intermediate in the Yucatan and highest in Pacific collections. SGIR was lowest in collections north of the NVA.
TR varied significantly and broadly among mosquito generations F1—F6 (Fig 6). Mosquitoes harvested at 7 dpi had a much lower TR than those harvested at 14 dpi. SGIR varied broadly between collections with Ciudad Madero and Monterrey again having the lowest TR and Poza Rica, Coatzacoalcos and Guerrero having the greatest TR. TR, as with the three other measures, was least in the northeastern region, but the same in the Yucatan and Pacific collections. TR was lowest in collections north of the NVA. These patterns could be confounded by the large variation in TR among mosquito generations.
The DIR, MIR, SGIR and TR are shown together for each collection at 7 dpi (Fig 7A) and 14 dpi (Fig 7B). At 7 dpi the MIR was high except for Ciudad Madero. In most cases, the DIR and SGIR were lower than the MIR. This suggests that at 7 dpi most infections have not disseminated yet and only a few DI have progressed to infect the salivary glands. At 14 dpi the MIR, DIR, and SGIR are equivalent except for Ciudad Madero and Monterrey. The TR at 14 dpi are uniformly greater than they were at 7 dpi.
The DIR, MIR, SGIR and TR are shown together for each region at 7 dpi (Fig 8A) and 14 dpi (Fig 8B). At 7 dpi MIR in Pacific collections are significantly greater than in Yucatan collections which are significantly greater than in northeastern collections. That same pattern occurs in DIR, SGIR, and TR. By 14 dpi the MIR is the same in Pacific and Yucatan collections and both are greater than in northeastern collections. The DIR is the greatest in Pacific collections, significantly higher than in Yucatan collections and least in northeastern collections. The SGIR is the same in Yucatan and northeastern collections; both lower than in the Pacific. The TR in the Pacific and Yucatan collections are the same and both are greater than the TR in northeastern collections.
Fig 9 displays DIR, MIR, SGIR and TR for each of the three Ae. albopictus collections at 7 dpi (Fig 9A) and 14 dpi (Fig 9B). The four VC rates were not compared among the three Ae. albopictus collections because they are confounded by large variation among the generation measured and because the four rates were highly correlated with the log10(x+1) blood titer. However it was instructive to compare VC patterns between the two species. At 14 dpi in Ae. aegypti the MIR, DIR, and SGIR were equivalent (except for Ciudad Madero and Monterrey). However in Ae. albopictus the MIR, DIR, and SGIR are significantly different at 14d pi in all three collections. This may suggest that in Ae. albopictus there are distinct MIB, DIB and SGIB that block passage of ZIKV from one tissue to the next. In addition notice that, unlike Ae. aegypti, all four parameters only increase slightly from 7 dpi to 14dpi. These patterns are summarized for both species in Fig 10 across all collection sites combined to reiterate species differences.
Table 1 shows the relative contribution of each of the four barriers in all 10 collections using Eqs 1–10. Based on the contribution of each barrier, this analysis shows that the main barrier to ZIKV transmission in Ae. aegypti is the SGEB (Table 1) in nine of the 10 collections and SGIB was the main barrier in Ciudad Madero. TR ranged from 8–51% at 14 dpi. The main barrier to ZIKV transmission in Ae. albopictus was the SGEB in Coatzacoalcos (Table 1) and SGIBs in Ciudad Nicolas and Huehuetan. TR ranged from 2–21% at 14 dpi in Ae. albopictus.
This study demonstrates that Mexican Ae. aegypti and Ae. albopictus are competent vectors of ZIKV. However the patterns of MIR, DIR, SGIR and TR in general vary between the two species. The MIR, DIR, and SGIR were all statistically similar by 14 dpi in Ae. aegypti (Fig 10). In contrast in Ae. albopictus, the MIR, DIR, and SGIR are significantly different at 14dpi in all three collections. In Ae. albopictus there may be distinct MIB, DIB and SGIB that block passage of ZIKV from one tissue to the next. In addition notice that, unlike Ae. aegypti, all four rates only increase slightly from 7 dpi to 14dpi in Ae. albopictus.
Vector competence varied among collections and geographic regions in Mexico and depending on whether mosquitoes are collected north or south of the NVA. Lozano-Fuentes et al [26] previously reported an abrupt decline in MIR, DIR, SGIR and TR in collections just south of the NVA. This was consistent with a hypothesis that the intersection of the NVA with the Gulf of Mexico coast acts as a barrier to gene flow previously observed between Ae. aegypti collections north and south on coastal plain along the Gulf of Mexico. The Transverse Volcanic Belt of Mexico divides the state of Veracruz into northern and southern Coastal Plains. This belt began to develop during the Oligocene and then later, during the Pliocene–Pleistocene, intense orogenic activity raised the Neovolcanic axis. The NVA extends from near the Pacific Coast east to the Gulf of Mexico and intersects the Atlantic coast. In the present study an abrupt drop in MIR, DIR and SGIR was observed but far north of the NVA in Monterrey and Ciudad Madero. On the other hand, the lower VC in northeastern collections observed by Bennett et al [31] for DENV2 were also noted in the current paper for ZIKV.
In general both Ae. aegypti and Ae. albopictus were highly susceptible to midgut infection, had high dissemination and salivary gland infection rates but these led to only low to moderate transmission rates. The potential role of the SG barriers has been previously suggested for ZIKV [23, 32], and herein we also provide evidence of a strong SGEB limiting ZIKV transmission in Mexican Ae. aegypti populations. Salivary gland escape barriers may therefore be the most important factors limiting ZIKV transmission by Ae. aegypti. In contrast Ae. albopictus had lower rates of midgut infection and dissemination but transmission was limited mainly by a salivary gland infection barrier.
TR at 14 dpi varied from 8–51% in Ae. aegypti and from 2–26% in Ae. albopictus. SGEBs are indicated when mosquitoes have an infected salivary gland but are unable to transmit virus. The existence of SGEBs has been definitively demonstrated for Japanese Encephalitis Virus in Culex tritaeniorhynchus[33], Snow Shoe Hare virus in Ae. triseriatus[34], La Crosse virus (LACV) in Ae. hendersoni [35] and Sindbis virus in Cx.theileri [36]. More recently, SGEBs have been reported for Rift Valley Fever Virus [37, 38].
Alternatively, an apparent SGEB may simply reflect a low sensitivity of saliva collection for detecting live ZIKV followed by plaque assays. A major difference between studies of SGEB made in the mid 1970–1980’s and similar studies done now is that earlier studies were able to use 8-day-old (suckling) mice to test for arboviral transmission since suckling mice are susceptible to fatal infection. Although there are recent examples [39], use of suckling mice is now prohibited by many institutions including CSU for bioethical reasons. Use of suckling mice is also prohibitively expensive when analyzing many mosquitoes as in the current study.
The protocol used in the present study involves removing the legs and wings of the mosquito and sliding its proboscis into the narrow end of a glass capillary tube filled with mineral oil. This “oil-technique” is well established at CSU and can be viewed as an online video which shows CSU personnel performing the technique [40]. Plaque assays are then used to detect live virus and/or measure the amount of live virus in the collected saliva. However this quantitative step was not taken in the present study because of the large numbers of individuals analyzed.
Measuring TR in collected saliva remains problematic because mosquitoes vary in the amount of saliva that they produce during probing and ingestion of a blood meal. Furthermore the concentration of virus may not be uniform in the saliva such that small volumes may carry as much virus as a larger volume. As examples, transmission electron microscopy in the salivary glands of Ae. albopictus revealed mature CHIKV particles (an alphavirus) that are highly clumped (almost crystalized) in some acinar cells [41] but in other cells mature particles appear to be more uniformly distributed. The same pattern appears in the salivary glands of Culex pipiens quinquefasciatus [42] infected with West Nile Virus (a flavivirus).
However, using transmission electron microscopy can also be misleading because many of the particles that are visualized may be defective rather than live virus. Defective particles are also a problem when trying to quantify viral RNA using Quantitative RT-PCR [43]. With DENV-2, we found that lower detection and quantitation limits were 20 and 200 copies per reaction, respectively. Amounts of positive and negative strand viral RNA strands were correlated and the numbers of plaque-forming units (PFU) were correlated with DENV-2 RNA copy number in both C6/36 cell cultures and mosquitoes. PFU were consistently lower than RNA copy number by 2–3 log10. Some investigators use pilocarpine (a parasympathomimetic plant alkaloid) to promote salivation [44] when attempting to collect saliva. But in our experience, while pilocarpine increases the volume of saliva produced by a mosquito, it also largely depletes the salivary glands of saliva and makes them brittle and difficult to remove intact. Furthermore, pilocarpine treated females probably produce more saliva than they probably do in a normal bite. Thus it is possible that the salivary escape barriers reported here are due to lack of sensitivity of our in-vitro assay for virus transmission.
It would be interesting to repeat this study using ZIKV isolates from Mexico. At the time that we initiated this study no viral isolates from Mexico were available. Genotype x genotype interactions have been observed for other flaviviruses [45]. However, it has been also observed that virus from the Asian strain behaves similarly to a virus from the Americas (Mexico) [24]. A strain from Mexico and the Puerto Rican PRVABC59 was previously shown to be very similar [29].
|
10.1371/journal.pgen.0030072 | Regional Variation in the Density of Essential Genes in Mice | In most species, and particularly in vertebrates, the percentage of genes absolutely required for survival, the essential genes, has not been estimated. To obtain this estimation, we used the mouse as an experimental model to carry out high-efficiency N-ethyl-N-nitrosourea (ENU) mutagenesis screens in two balancer chromosome regions, and compared our results to a third previously published screen. The number of essential genes in each region was predicted based on allele frequencies. We determined that the density of essential genes differs by up to an order of magnitude among genomic regions. This indicates that extrapolating from regional estimates to genome-wide estimates of essential genes has a huge variance. A particularly high density of essential genes on mouse Chromosome 11 coincides with a high degree of regional linkage conservation, providing a possible causal explanation for the density variation. This is the first demonstration of regional variation in essential gene density in the mouse genome.
| The genome sequences of many organisms are now complete. However, speculation remains regarding the function of many newly discovered genes. There is also debate about the percentage of genes that are required to build an organism. These genes, which are necessary for the development of the organism, are essential genes. We have performed mutagenesis screens that allow the identification of mutations in essential genes from specific regions of the mouse genome. From these data we have predicted the number of essential genes in three regions of the mouse genome. When we compared these predictions, we found that the density of essential genes varies in different regions of the mouse genome. We then analyzed these regions of the genome to identify similar regions in other mammals. We found that regions of the mouse genome with a high density of essential genes are more similar to other species than those regions with fewer essential genes, suggesting that throughout evolution genomic regions with many essential genes remain intact.
| In the era of complete genomes, the total number of genes in a sequenced organism can now be predicted, but the function and selective importance of a substantial fraction of genes remains unknown. Some gene functions may be of central importance to the organism, whereas other gene functions may be useful, but not critical, or may have functions that are partially redundant. Genes are classified as essential if an organism cannot develop to maturity without them. Here, employing balancer chromosome mutagenesis studies on specific regions of the mouse genome, we evaluate the distribution of essential genes in these regions. Our data also show that in mammals, similar to worms [1], essential gene clusters are located in genomic regions with high linkage conservation.
Essential genes in two genomic regions were targeted using balancer chromosome screens: a 35-Mb region of mouse Chromosome 11 between the Trp53 and Wnt3 loci [2] and a 20-Mb region of mouse Chromosome 4 between markers D4Mit281 and D4Mit51 [3]. For comparison, we also analyzed results from an earlier mutagenesis study that identified nine essential loci in a 20-Mb deletion region on mouse Chromosome 7 [4]. In our study, we considered essential genes to be those that when mutated cause lethality at or before birth. To improve the accuracy of the analysis, we performed pair-wise complementation tests of fully penetrant mutant lines from each screen to identify alleles at each locus. From 785 pedigrees bred in the Chromosome 11 balancer screen, we isolated 45 mutant lines that die at or before birth (Table 1). These 45 lines formed 40 complementation groups, and thus only five loci were detected more than once (Table 1). From 551 pedigrees bred in the Chromosome 4 balancer screen, we isolated 16 mutant lines that die at or before birth (Table 1). These mutants formed 12 complementation groups (Table 1). In comparison, the deletion screen on Chromosome 7 bred 4,557 pedigrees to generate 24 fully penetrant lethal mutant lines that fell into nine complementation groups [4]. Notably, only a third of the number of pedigree groups were screened on Chromosome 11 as compared to Chromosome 7. However, we obtained about two and a half times as many mouse lines carrying essential genes, and almost six times as many complementation groups.
To predict the number of essential genes in each chromosomal region, we employed a Bayesian approach that incorporates variation in the degree of mutability among loci to provide a credible range of values rather than a point estimate [5]. This analysis requires knowledge of the number of complementation groups in each region, and cannot be applied to studies that fail to consider allelism. Evidential support for gamma and mixture models that incorporate variation in mutability among loci was minimal based on the datasets alone, although previous analyses show that variation in mutability is the norm [5]. When mutabilities vary, genes with low mutabilities tended to be under-counted if a model with a single mutability rate (Poisson) is assumed; the numbers of lethal mutations predicted from a Poisson distribution are therefore probably an underestimate [6,7]. To obtain an accurate measurement, we considered gamma-distributed mutabilities with the shape parameter constrained to reasonable values (a = 0.2–5.0) based on previous observations [5].
There were 222 essential genes (between 98 and 943 based on a very conservative 99% credible region) predicted in the Chromosome 11 balancer region (Figure S1A; Table S1). Similarly, 31 essential genes (16 to 124) were predicted in the Chromosome 4 balancer region (Figure S1B). The Chromosome 7 mutagenesis experiment was more highly saturated, with 12 essential genes estimated (10 to 25, Figure S1C).
These three regions clearly vary considerably in their density as well as their number of essential genes. The predicted mean density of essential genes per Mb in the Chromosome 11 balancer region is four times greater than the density on Chromosome 4, and 11 times greater than the density on Chromosome 7. All density differences between chromosomes are significant, and the chromosome 11/4 density ratio is at least 2.26 (p < 0.05), while the 11/7 ratio is at least 7.0 (p < 0.05). The number of essential genes predicted in each region is also significantly different (p < 0.05) as a proportion of the total number of predicted genes (739, 373, and 237, respectively).
The Chromosome 11 balancer region has unusually high synteny in addition to its high essential gene density: human Chromosome 17 is entirely conserved with this region of mouse Chromosome 11, making it the most conserved mouse–human autosomal linkage group (Figure S2). Chromosomes 4 and 7 have less synteny conservation with human chromosomes (data not shown). Although gene density (as well as essential gene density) is high on Chromosome 11, we found that on other mouse chromosomes the relationship between gene density and synteny conservation was weak (Figure S3).
The number of essential genes appears to be predictive of microsynteny and sequence conservation as well as large-scale synteny. We examined homologs among mouse, rat, human, dog, and cow to determine which genes had the same neighbors in all five species, and found that 26% of the genes on mouse Chromosome 11 had conserved microsynteny. In contrast, only 22% of the genes on Chromosome 4 and 13% of the genes on Chromosome 7 had conserved microsynteny in all five species (Table 1). These frequency differences are significant (Table 1). At the sequence level, a previous comparison between the C57BL/6J and 129S5 mouse strains demonstrated that Chromosome 11 has much higher sequence conservation than Chromosomes 4 or 7 [8]. Overall, Chromosome 11 is the third most-conserved chromosome between these two strains [8].
In this first comparative study of essential gene densities in a mammalian genome, we have identified surprising differences as large as an order of magnitude. Our region-specific mutagenesis screens combined with complementation testing were laborious but necessary for these calculations. Our statistical accommodation of variation in mutability, although more complex than most previous studies, allowed a more accurate assessment of the variability in essential gene density.
Sequence conservation of regions dense in essential genes is perhaps not surprising, but synteny conservation is more so. A weak correlation between essential gene density estimates and synteny was previously observed in roundworms based on RNAi [1], but our observations in mammals use a more precise assessment of essential function and a more definitive assessment of large-scale synteny among more species, as well as an assessment of microsynteny. Thus, it is reasonable to consider a general causal relationship between essential genes and reduced rates of chromosomal translocation and rearrangement. If adjacent essential genes generally reduce the probability of productive chromosomal translocations between them, essential gene-dense regions would be expected to expand over time as essential genes randomly join a cluster, but then have a reduced probability of departing. Thus, it appears that the large number of densely packed essential genes on the balancer region of mouse Chromosome 11 may have forced it to remain as a unit in spite of millions of years of divergence and speciation. This also predicts that syntenically conserved regions should be especially attractive targets for future essential gene detection.
It is traditional to use regional estimates of essential gene density to estimate the total number of essential genes in the genome. If we extrapolate the number of essential genes as a proportion of predicted genes in each region, there would be 5,749 essential genes overall (20% of the genome). If we extrapolate based on the density of essential genes per Mb, we predict about twice as many (10,849). The results of our own research, however, indicate that the variability on this extrapolation is huge. If the variability of the regional estimates, as well as the variability among the regional estimates (up to 11-fold), is taken into account, the estimate ranges from ∼1,100 essential genes up to more genes than the total predicted number of genes in the genome (28,594). It is a near certainty that such variability is not specific to our study, but applies to all previous estimates of essential genes that utilized one or a few genomic regions. If the relationship between essential genes and synteny, particularly microsynteny, is consistently upheld in a variety of organisms, more accurate and believable estimates could be obtained by using microsynteny and conservation in essential gene predictions.
The fraction of lethal mutations remaining to be isolated from each screen was calculated using Saturate [5]. We considered gamma-distributed mutabilities with the shape parameter constrained to reasonable values (a = 0.2–5.0) based on previous observations. For the gamma model, alpha was constrained to be less than 5.0.
Genomic sequences of mouse, human, chimp, rat, cow and dog were downloaded from Ensembl v.38 (http://www.ensembl.org/info/data/download.html). Each region of mouse sequences was divided into 150-kb fragments, which were then blasted using Megablast (http://www.ncbi.nlm.nih.gov/BLAST/download.shtml). The sequence comparison was carried out on a Sun cluster with SunFire 280R (http://www.sun.com). Mouse genomic annotation was downloaded from Ensembl BioMart v.38 (http://www.ensembl.org/Multi/martview). To visualize the blast results, we developed in-house software written in Microsoft Visual Basic (http://www.microsoft.com). All blast results were uploaded in a MS SQL server database, and the results displayed on a PC. Microsynteny comparisons were performed using gene annotation from Ensembl Biomart v.38. A list of genes with conserved microsynteny will be provided upon request.
An explanation of calculations is found in Table S2. All predictions are based on protein-coding known genes found in Ensembl Biomart v.39. The extremes of two distributions such that they were as similar as possible but the joint probability was no less than 5% was taken to obtain the minimal ratio of the two essential gene predictions. In no case did the density distributions overlap with greater than 5% probability. |
10.1371/journal.pgen.1008124 | Estimating variance components in population scale family trees | The rapid digitization of genealogical and medical records enables the assembly of extremely large pedigree records spanning millions of individuals and trillions of pairs of relatives. Such pedigrees provide the opportunity to investigate the sociological and epidemiological history of human populations in scales much larger than previously possible. Linear mixed models (LMMs) are routinely used to analyze extremely large animal and plant pedigrees for the purposes of selective breeding. However, LMMs have not been previously applied to analyze population-scale human family trees. Here, we present Sparse Cholesky factorIzation LMM (Sci-LMM), a modeling framework for studying population-scale family trees that combines techniques from the animal and plant breeding literature and from human genetics literature. The proposed framework can construct a matrix of relationships between trillions of pairs of individuals and fit the corresponding LMM in several hours. We demonstrate the capabilities of Sci-LMM via simulation studies and by estimating the heritability of longevity and of reproductive fitness (quantified via number of children) in a large pedigree spanning millions of individuals and over five centuries of human history. Sci-LMM provides a unified framework for investigating the epidemiological history of human populations via genealogical records.
| The advent of online genealogy services allows the assembly of population-scale family trees, spanning millions of individuals and centuries of human history. Such datasets enable answering genetic epidemiology questions on unprecedented scales. Here we present Sci-LMM, a pedigree analysis framework that combines techniques from animal and plant breeding research and from human genetics research for large-scale pedigree analysis. We apply Sci-LMM to analyze population-scale human genealogical records, spanning trillions of relationships. We have made both Sci-LMM and an anonymized dataset of millions of individuals freely available to download, making the analysis of population-scale human family trees widely accessible to the research community. Together, these resources allow researchers to investigate genetic and epidemiological questions on an unprecedented scale.
| Genealogical records can reflect social and cultural structures, and record the flow of genetic material throughout history. In recent years, very large pedigree records have come into existence, owing to collaborative digitization of large genealogical records [1,2] and to digitization of large cohorts collected by healthcare providers, spanning up to millions of individuals [3–7]. Such population-scale pedigrees allow investigating the sociological and epidemiological history of human populations on a scale that is orders of magnitude larger than existing studies.
Traditional human pedigree studies collect a large number of independent families which are analyzed separately and then meta-analyzed. However, this approach is not suitable for population-scale pedigrees, because such pedigrees cannot be decomposed into mutually exclusive families [1]. Hence, the analysis of such pedigrees requires modeling complex covariance structures between trillions of pairs of individuals.
Pedigree studies often employ LMMs to decompose the phenotypic variation among individuals into variance components such as genetic effects and shared environment [8]. LMMs have been the statistical backbone of animal and plant breeding programs for almost six decades [9], and have been continuously developed over the years [10–22]. LMMs are routinely used nowadays to analyze pedigrees of millions of animals and plants [13,23], hundreds of thousands of which are often genotyped (e.g. [22,24,25]).
LMMs and their extensions have recently gained considerable popularity in human genetics studies for the purposes of estimating heritability [26–32] and genetic correlation [33–37], predicting phenotypes [38–41] and modeling sample relatedness [42–46]. Unlike classical animal and plant studies, human studies typically do not include pedigree data, but instead measure genetic relatedness via dense genotyping of single nucleotide polymorphisms (SNPs).
In recent years, animal and human studies have been using different techniques to scale LMMs to datasets with millions of individuals. Animal studies typically fit large-scale LMMs via restricted maximum likelihood (REML) [12], by exploiting the sparsity of pedigree data. Specifically, a pair of individuals with no known common ancestor can be regarded as having no genetic similarity. Consequently, these pairs induce a zero entry in the genetic similarity matrix. Such sparse matrices can be stored and analyzed efficiently with suitable numerical techniques [21,47,48].
Human genotyping studies do not give rise to sparse data structures. Instead, human studies have managed to scale LMMs to large datasets via two approaches. The first approach applies REML, either via supercomputers with thousands of CPUs and terabytes of memory [46], or by approximating the restricted likelihood gradient via Monte-Carlo techniques [28]. However, the latter technique is only suitable for specific types of covariance matrices whose decomposition is known beforehand.
The second approach to scale LMMs uses the method of moments rather than REML, by solving a set of second moment matching equations [49–53]. Such approaches have become increasingly popular recently [30–35,54–60] owing to their computational tractability and their compatibility with privacy-preserving summary statistics [61]. Although moment estimators are less statistically efficient compared to REML estimators, they have several advantages: the loss of efficiency has been found to often be small [56]; they are more robust to modeling violation because they make fewer distributional assumptions; and they are more flexible, which enables applying techniques to limit confounding factors such as assortative mating (Methods). Moment estimators have also recently been explored in animal breeding studies [62–64] and were found to be faster than REML while providing similar accuracy, but they have not been widely adopted in animal studies to date.
Here we present Sci-LMM, a statistical framework for analyzing population-size pedigrees that combines techniques from animal and human genetic studies. Sci-LMM uses sparse data structures as is common in animal studies, and supports both moment and REML estimators. The moment estimator is based on a common technique called Haseman-Elston (HE) regression [65,66] (Methods). Sci-LMM scales HE regression to population-sized pedigrees via sparse matrix tools [67]. The REML estimator combines a direct sparse REML solver [47] with Monte-Carlo gradient approximation [68]. Importantly, existing packages for pedigree-based REML [69–73] cannot handle the analyses performed in this paper because they require the inverse of the epistatic interactions matrix [47,74,75], which is extremely difficult to compute in large pedigrees [76]. Hence, Sci-LMM provides a comprehensive solution for LMM-based pedigree analysis.
To demonstrate the capabilities of Sci-LMM, we carry out an extensive analysis of simulated data with millions of individuals, which we complete within a few hours. We additionally estimate the heritability of longevity and of reproductive fitness (quantified via number of children), using a large cohort spanning millions of genealogical records and several centuries of human history. We estimate that both traits have a substantial heritable component, with an estimated 22.1% heritability for longevity and 34.4% for reproductive fitness. Sci-LMM enables analysis of large pedigree records that was not previously possible.
Consider a sample of n individuals with observed phenotypes y1,…,yn, and covariates vectors C1,…,Cn, and consider a set of n×n covariance matrices M1,…,Md, where Mi,jk encodes the covariance between the phenotypes of individuals i and j according to the kth covariance structure, up to a scaling constant. We assume that the vector y = [y1,…,yn]T follows a multivariate normal distribution:
y∼N(Cβ;Σ)
(1)
Σ=G+σe2I
(2)
G=∑k=1dσk2Mk.
(3)
Here, C = [C1,…,Cn]T is an n×c matrix of covariates (including an intercept), β is a c×1 vector of fixed effects, Σ is the covariance matrix of the vector y, σk2 is the kth variance component, and I is the identity matrix. The parameters to estimate are the fixed effects β and the variance components σ12,…,σd2,σe2. The Sci-LMM software can currently compute an identity by descent (IBD) matrix, an epistatic covariance matrix and a dominance matrix, as described below.
The restricted log-likelihood lR(β,σ12,…,σd2,σe2) is given by [77]:
lR(β,σ12,…,σd2,σe2)=l(β,σ12,…,σd2,σe2)+c2log(2π)+12log|CTC|−12log|CTΣ−1C|
(4)
l(β,σ12,…,σd2,σe2)=−12(y−Cβ)Σ−1(y−Cβ)−12log|Σ|−n2log(2π),
(5)
where l(β,σ12,…,σd2,σe2) is the non-restricted likelihood and c is the number of covariates.
An alternative form of Eq 4 often used in animal breeding literature is:
lR(β,σ12,…,σd2,σe2)=−12(log|B|+log|Σ|+yTPy),
(6)
where B=[σe−2CTCσe−2CTσe−2Cσe−2I+G−1],P=σe−2I−σe−4WB−1WT,W=[CI] and we ignored additive constants. This form is particularly convenient when the inverse of each of the matrices M1,…,Md is known, as it can be solved efficiently using mixed model equations via Gaussian elimination, without having to directly invert or factorize the matrix Σ [47,72]. This makes it particularly convenient to use this form in the presence of only an additive IBD matrix, because the inverse of this matrix is sparse and can be computed analytically [78,79].
REML estimation consists of finding the parameters β,σ12,…,σd2,σe2 that maximize Eq 4. When the inverse of each of the matrices M1,…,Md is known, the REML can be found efficiently by using Eq 6, using the so-called mixed model equations method [47,72]. Here we describe a direct solution that can be applied when the inverse of M1,…,Md is unknown.
Our solution combines several ideas: (1) we maximize Eq 4 directly, rather than the equivalent form of Eq 6; (2) instead of directly inverting Σ, we compute its Cholesky factorization Σ = LLT via sparse matrix routines; (3) any product of the form Σ−1v for some vector v is computed using L and two triangular solvers (forward and backward substitution); and (4) the gradient of Eq 4 is approximated using Monte Carlo techniques. We now describe our REML approach in detail.
We first describe a solution to the unrestricted log-likelihood (Eq 5) and then extend the solution to the restricted log-likelihood (Eq 4). To compute Eq 5 we need to compute the terms Σ−1(y−Cβ) and log|Σ|. The first term can be computed exactly via either conjugate gradient iterations or by explicitly computing the Cholesky factorization of Σ and then applying forward and back substitution. The second term can be computed via the Cholesky factorization of Σ. The Cholesky factorization can be computed efficiently via the CHOLMOD routines [80]. It remains to find the maximum likelihood estimates of the model parameters.
To find the MLE of β^ we note that given Σ, β^ can be computed analytically by deriving Eq 5 with respect to β as follows:
∂l(β,σ12,…,σd2,σe2)∂β=−12(y−Cβ)TΣ−1C
(7)
By setting the transpose of the gradient to 0, we obtain the MLE:
β^=(CTΣ−1C)−1CTΣ−1y.
(8)
The MLEs of the variance components σ^12,…,σ^d2,σ^e2 are estimated via an optimization procedure, which requires computing the gradient of Eq 5. The partial derivative with respect to each variance component σk2 is given by:
∂l(β,σ12,…,σd2,σe2)∂σk2=−12yTΣ−1MkΣ−1y−12Tr[Σ−1Mk].
(9)
The first term on the right-hand side of Eq 9 can be computed efficiently given the Cholesky factorization of Σ. Unfortunately, the second term cannot be solved efficiently via the above technique because it requires solving n different linear equations, where n can be in the millions. Instead, we use the approximation technique used in [28,68,81]. We first rewrite this term as an expectation (ignoring the scaling factor) as follows:
Tr[Σ−1Mk]=Tr[Σ−1MkΣ−1Σ]=Tr[Σ−1MkΣ−1E[y′y′T]]=E[Tr[Σ−1MkΣ−1y′y′T]]=E[Tr[y′TΣ−1MkΣ−1y′]]=E[y′TΣ−1MkΣ−1y′],
(10)
where y′∼N(0,Σ) and we used the fact that the trace of a scalar is equal to the scalar. We therefore approximate Eq 10 by sampling a small number of y′ vectors to approximate the expectation. These vectors can be sampled efficiently given the Cholesky factorization Σ = LLT by sampling a vector yΣ∼N(0,I) and then using the fact that LyΣ∼N(0,Σ). The Cholesky factorization can be computed efficiently via the CHOLMOD routines. We found that 100 vectors often yields a very good approximation at a modest computational cost.
We note that [28] proposes an alternative estimation method by completely foregoing the likelihood computation, and instead only trying to minimize the squared gradient elements. However, we found that in sparse settings, this solution often converges into local maxima at the edge of the parameter space (where many variance components are equal to zero) rather than the true maximum likelihood estimate.
We now extend the solution to handle restricted maximum likelihood (Eq 4). Clearly, the restricted maximum likelihood estimate of β is the same as the MLE. The derivative of the restricted log likelihood with respect to each variance component σk2 is given by:
∂lR(β,σ12,…,σd2,σe2)σk2=∂l(β,σ12,…,σd2,σe2)σk2+12Tr[(CTΣ−1C)−1CTΣ−1MkΣ−1C].
(11)
The term Σ−1C can be computed by solving c different linear equations, which can be performed efficiently given the Cholesky factorization of Σ. All the other terms can be computed efficiently, assuming that c is small compared to n.
The standard errors of σ12,…,σd2 can be approximated via the average information REML (AI-REML) procedure [82], which consists of approximating each entry of the Hessian of the restricted log likelihood as follows:
∂l(β,σ12,…,σd2,σe2)σk2σl2≈−12yTPMkPMlPy,
(12)
where P = Σ−1− Σ−1C(CTΣ−1C)−1CTΣ−1. Afterwards we approximate the standard errors via the square roots of the diagonal entries of the inverse of the negative Hessian. Following [28], we multiply these entries by (1+1100) to account for sampling variance introduced by the 100 y′ vectors sampled in the Monte-Carlo approximation.
We implemented our REML solver in Python, using an L-BFGS-B algorithm [83] as implemented in the SciPy package [67]. To prevent the parameters from inducing a non positive-definite matrix, We enforced non-negative parameters by using a log-transformation, which transforms the problem into an unconstrained optimization problem.
HE regression estimates variance components via the method of moments, by finding the set of variance components σ12,…,σd2,σe2 that minimize the expression:
∑i,j(cov(yi−Ciβ,yj−Cjβ)−Σij)2.
(13)
Typically, the fixed effects β are first estimated without considering the covariance matrices, by solving the multivariate linear regression problem y=Cβ+ϵ,ϵ∼N(0,σe2I), where I is the identity matrix. This solution yields a consistent estimator under mild regularity conditions [84]. Afterwards we plug the fixed effect estimate β^ into Eq 13 and estimate the variance component estimates σ^12,…,σ^d2 as follows:
[σ^12,…,σ^d2]T=([V1,…,Vd]T[V1,…,Vd])−1[V1,…,Vd]TY,
(14)
where Vk is a vector representation enumerating the elements Mijk for all pairs of distinct individuals i,j, and Y is a vector representation of the corresponding elements (yi−Ciβ^)(yj−Cjβ^). Each element q,r of the d×d matrix ([V1,…,Vd]T[V1…,Vd]) can be computed via an element-wise multiplication of the upper-diagonal elements of the matrices Mq, Mr, which can be performed efficiently via sparse matrix routines. The vector [V1,…,Vd]TY can also be computed efficiently in a similar manner.
By following the notation of [56] and denoting q≜[V1,…,Vd]TY, S = [V1,…,Vd]T[V1,…,Vd], we have:
[σ^12,…,σ^d2]T=S−1q.
(15)
By applying a few matrix manipulations, we can compute q and S efficiently as follows:
qk=yTMky−∑iMiikyi2=yT(Mk−I)y
(16)
Skl=∑i,jMijkMijl−∑iMiikMiil,
(17)
where we used the assumption Miik=1. Both these quantities can be computed explicitly via sparse matrix routines.
The sampling variance of the estimators is given by S−1var(q)S−1, where var(q) is given by:
var(q)kl=2tr(Σ^(Mk−I)Σ^(Ml−I)).
(18)
This quantity can be computed in two ways:
Exactly, via: tr(Σ^(Mk−I)Σ^(Ml−I))=∑ij[Σ^(Mk−I)]ij[Σ^(Ml−I)]ij
Approximately, via: tr(Σ^(Mk−I)Σ^(Ml−I))=Ey′[y′T(Mk−I)Σ^(Ml−I)y′],
where y′ is sampled from N(0,I), using a derivation similar to the one in Eq 10.
The approximate approach uses Monte-Carlo approximations, by randomly sampling y′ vectors and approximating the right hand side. It can be substantially faster than the exact approach (because it circumvents expensive matrix-matrix multiplications) and obtain excellent accuracy. The Sci-LMM software uses the approximate approach by sampling ~100 random y′ vectors.
HE regression provides a simple technique for excluding specific pairs of individuals (e.g. spouses) from the analysis without excluding the individuals themselves. This can be useful when trying to limit confounding due to factors such as assortative mating. Excluded pairs can be omitted by zeroing the covariance matrix entries of corresponding pairs. Importantly, this technique cannot be used in REML, because the resulting covariance matrices may not be positive definite. We note that another potential approach to capture environmental risk factors is including shared effects with a suitable incidence matrix [10], but this approach requires additional assumptions and has not been used here.
HE regression is a convenient theoretical framework to analyze the factors affecting estimation accuracy. HE regression can be considered as a special form of linear regression, where off-diagonal entries of covariance matrices serve as explanatory variables. Hence, good accuracy is obtained when measured and unmeasured explanatory variables are uncorrelated with other explanatory variables (Eq 18).
Specifically, obtaining accurate estimates requires (1) that the off-diagonal entries of the LMM covariance matrices are uncorrelated with each other; and (2) that they are uncorrelated with covariance due to unmeasured environmental factors. While the first requirement can be easily tested, the second one requires making strong assumptions about the structure of environmental covariance. For example, if latent environmental factors are shared between spouses but not between parents and children, we may wish to exclude spouses from the analysis. Unfortunately, we not know the structure of environmental covariance for the traits studied in this work, and we leave its investigation for future work.
The IBD kinship coefficient of two individuals, denoted as aij, is the probability that a randomly selected allele in an autosomal locus was originated from the same chromosome of a shared ancestor between individuals i and j [85,86], and is given by:
ai,j={1+fi,i=jrijaiiajj,i≠j
Here, fi is the inbreeding coefficient, defined as half of the IBD coefficient of the parents of individual i [85], and rij is the coefficient of relationships, defined as:
rij=∑path1+fA2|path|+1(1+fi)(1+fj)
The quantities in the above equation are defined as follows: A is a least common ancestor of individual i and j in the pedigree graph (a graph where every node is an individual connected to her parents and children); the summation is performed over every path connecting individuals i,j in the pedigree graph, culminating at some ancestor A, such that the path does not contain the same individual twice; and |path| is the path length.
To efficiently compute the IBD matrix we first construct the matrices L and H of its decomposition A = LHLT, where L is a lower triangular matrix such that Lij contains the fraction of genome shared between individuals i and her ancestor j, H is diagonal, and the matrices are ordered such that ancestors precede their descendants (Fig 1A–1C). The matrices L and H can be computed efficiently via iterative techniques [78,79] using sparse matrix routines [80] (S1 Text).
Dominancy represents the genetic variance due to co-ancestry of two alleles, and can be approximated by 14(Afi,fj⋅Ami,mj+Afi,mj⋅Ami,fj), where Ak,l is the IBD coefficient of individuals k,l, and fk,mk are the parents of individual k [10,87]. A necessary condition for nonzero dominancy entry is a nonzero IBD relationship, which enables rapid computation of the dominance matrix.
Epistatic covariance encodes the assumption that variants interact multiplicatively to affect a given phenotype, and is proportional to the exponent of the corresponding IBD coefficient, i.e., (Ak,l)2 for two-loci epistasis, (Ak,l)3 for three-loci epistasis and so on [75]. Therefore, an epistatic covariance matrix is simple to compute given the IBD matrix.
Population scale pedigree data typically presents heterogeneity of the completeness of records. However, individuals with missing data may still be required for IBD computation. For example, consider a pedigree of two siblings with phenotypic data, and two parents and one uncle without phenotypic data. The parents are important for the IBD computation of the siblings, but the uncle is non-informative.
Sci-LMM applies pedigree-pruning techniques to remove non-informative individuals, similarly to other REML packages for pedigree analysis [70–73]. Briefly, we defined required individuals as individuals who have phenotypic and explanatory variables data, or individuals who appear in a lineage path connecting two individuals with such data with one of their least common ancestors (S1 Text; S1 Fig). This algorithm reduces the matrix construction time by several hours.
In addition to covariance matrices, Sci-LMM can include the top principal components (PCs) of the IBD matrix as fixed effects, using sparse matrix routines [88]. The inclusion of PCs can capture major linear sources of variation in a dataset, and is motivated by large scale human genetic studies, where such PCs often capture population structure [89]. However, we caution that PCs computed from an IBD matrix are not guaranteed to capture population structure [90]. An alternative approach often employed in animal studies is the assignment of unobserved parents to genetic groups [91], but this approach requires knowledge about the location of birth of all individuals without known parents.
We generated pedigrees mimicking real family patterns in the United States, partially based on publications by the United States Census Bureau [92,93]. We iteratively generated generations of individuals, where the first generation included two individuals, and the number of individuals in each successive generation increases by 40% (approximately the same ratio as in the GENI dataset), until obtaining the desired sample size. Each generation included 50% females and 50% males.
In each generation we generated households, where every household includes either one individual or two individuals with different genders, and every individual can belong to zero, one or multiple households. The number of households in each generation was 62.5% of the number of individuals in that generation. 68% of the households included pairs of individuals, and the rest included a single individual. Every individual in every generation (except for the top one) was born to parents from a randomly selected household from the previous generation (for 80% of individuals) or from two generations in the past (for the remaining 20% of individuals).
After generating all individuals, we omitted randomly selected edges until obtaining the desired sparsity factor, up to 10% error. We then created corresponding IBD, dominance and epistasis matrices.
Finally, we generated phenotypes using Eq 1 by (1) generating variance components σk2 for each covariance matrix Mk from U(0,1) and scaling them such that they sum to 1.0; (2) Generating 5 binary and 5 normally distributed covariates; and (3) generating fixed effects from N(0,1000/n), where n is the sample size.
The parameters differentiating the various experiments are: (1) cohort size (50K, 100K, 250K, 500K, 1M or 2M); (2) sparsity factor (0.0005, 0.001, or 0.005); and (3) the subset of matrices used. We generated 10 different datasets for every unique combination of settings, except for matrices with 2M individuals, for which we generated a single pedigree with ten different phenotype vectors due to runtime considerations.
All experiments were conducted using a Linux workstation with a 24-cores 2GHz Xeon E5 processor and 256Gb of RAM.
To evaluate the capabilities of Sci-LMM, we generated large synthetic pedigrees spanning 20 to 40 generations and various family structures, under a wide variety of settings. The pedigrees included 50,000–2,000,000 individuals, amounting to trillions of pairs of relatives. A subset of the individuals in each generation consists of children of individuals from either the previous generation, or from two generations in the past. To simulate patterns observed in real datasets, the simulations also included consanguinity, half-siblings, and individuals with less than two recorded parents (Methods).
In each simulation we generated a normally distributed phenotype, using a covariance matrix with additive, epistatic and dominance effects and ten binary covariates. Unless otherwise stated, the sparsity factor (the fraction of non-zero entries in each matrix) was 0.001. Ten different datasets were generated for each combination of sample size and sparsity factor.
In all settings, Sci-LMM yielded empirically unbiased estimates of the variance components, using both REML and HE regression. As expected, estimation accuracy increased with sample size, though the estimators became slightly less accurate when increasing the number of variance components, (Fig 2A–2C). Specifically, the root mean square error (RMSE) was < 0.03 for all methods under all settings with more than 250,000 individuals, indicating <3% average error (because the phenotype was standardized to have unit variance).
A comparison of the REML and the HE results shows that that HE was slightly more accurate in the presence of <100,000 individuals (Fig 2A–2C), and REML was slightly more accurate otherwise. These results possibly indicate that REML convergence is difficult in the presence of sparse covariance matrices with limited sample sizes. We also found that estimation accuracy was anti-correlated with relatedness sparsity, indicating that the estimators efficiently exploit the information found in non-zero covariance entries (Fig 2D and 2E).
We have described a statistical framework for analysis of large pedigree records spanning millions of individuals. Our framework includes methodologies for constructing large sparse matrices given raw pedigree data, and methodologies for LMM analysis with random effects described by these matrices. Taken together, the proposed solution enables an end-to-end analysis of population scale human family trees.
In this work we focused on partitioning phenotypic variance into additive genetics, epistasis and dominance. However, the LMM framework is flexible and can be extended in various directions. For example, sparse LMMs are often used to model transmissible phenomena [99–104], which enables combining pedigree-based and geography-based covariance structures. Both Sci-LMM and the data studied in this paper are freely available for download, which makes the analysis of population-scale human family trees widely accessible to the research community. Combined, these resources allow researchers to investigate genetic and epidemiological questions on unprecedented scales.
We evaluated two methods for variance components estimation: REML and HE regression. REML is more accurate than HE and provides a likelihood-based solution, which can be used for model comparison and hypothesis testing. HE estimates are less accurate but can be more robust to modeling violation. Importantly, HE regression can mitigate confounding due to environmental factors by zeroing selected entries in the covariance matrices, which may be especially suitable for studying human genealogical records (Methods). Hence, the two methods are complementary in terms of their strengths and weaknesses. In practice, we found that it is difficult to scale REML to datasets with more than 500,000 individuals with a sparsity factor of 0.001. Our recommendation is to use REML when it is feasible and all model assumptions hold, and HE regression otherwise. We note that REML estimation can be substantially faster when not fitting epistatic interactions by using a mixed model equations approach [47], which is implemented in several software packages [69–72].
Our work demonstrates the technical feasibility of studying population scale human family trees. However, the analysis of human genealogical records is challenging due to imperfect data and the difficulty of controlling for confounding factors. Potential issues include non-paternity, cryptic relatedness, missing or false genealogical records, genetic nurture [105,106], environmental bias [97,107], assortative mating [2] and correlation between additive and epistatic effects [17,18]. As such, our estimates should be considered as a first order approximation, and our heritability estimates are likely upper biased due to confounding. We expect that recently proposed techniques to address these issues (e.g. [2,106,107]) could be integrated into the Sci-LMM framework in the future.
In this work we focused on analyzing large pedigree records with no measured genotypes. In recent years, the advent of biobank-sized datasets allows analyzing population-scale genotyped cohorts. The two study types are complementary because biobanks cannot be used to investigate longevity, traits with a late age of onset, or epidemiological and sociological questions on historical scales. We anticipate that cohorts combining both types of data will become increasingly common. For example, we and other online genealogy platforms allow users to upload their genetic information and link it with their genealogical profile. Such combined datasets have been extensively explored in the animal breeding literature [19,21,108–112]. However, privacy and logistical concerns limit public access to human genetic data, necessitating methods based on summary statistics [61]. Thus, approaches for analysis of such combined datasets will combine state of the art techniques from the animal breeding and the human genetics literature, and remain a potential avenue for future work.
|
10.1371/journal.pcbi.1002744 | Evolutionary Analysis of Human Immunodeficiency Virus Type 1 Therapies Based on Conditionally Replicating Vectors | Efforts to reduce the viral load of human immunodeficiency virus type 1 (HIV-1) during long-term treatment are challenged by the evolution of anti-viral resistance mutants. Recent studies have shown that gene therapy approaches based on conditionally replicating vectors (CRVs) could have many advantages over anti-viral drugs and other approaches to therapy, potentially including the ability to circumvent the problem of evolved resistance. However, research to date has not explored the evolutionary consequences of long-term treatment of HIV-1 infections with conditionally replicating vectors. In this study, we analyze a computational model of the within-host co-evolutionary dynamics of HIV-1 and conditionally replicating vectors, using the recently proposed ‘therapeutic interfering particle’ as an example. The model keeps track of the stochastic process of viral mutation, and the deterministic population dynamics of T cells as well as different strains of CRV and HIV-1 particles. We show that early in the co-infection, mutant HIV-1 genotypes that escape suppression by CRV therapy appear; this is similar to the dynamics observed in drug treatments and other gene therapies. In contrast to other treatments, however, the CRV population is able to evolve and catch up with the dominant HIV-1 escape mutant and persist long-term in most cases. On evolutionary grounds, gene therapies based on CRVs appear to be a promising tool for long-term treatment of HIV-1. Our model allows us to propose design principles to optimize the efficacy of this class of gene therapies. In addition, because of the analogy between CRVs and naturally-occurring defective interfering particles, our results also shed light on the co-evolutionary dynamics of wild-type viruses and their defective interfering particles during natural infections.
| A long-standing challenge in efforts to control human immunodeficiency virus type 1 (HIV-1) is the rapid evolution of the virus. Any effective therapy quickly gives rise to so-called escape mutants of the virus, potentially resulting in treatment failure. A distinct class of gene therapy based on conditionally replicating vectors has been suggested to have potential to circumvent the problem of viral evolutionary escape. A conditionally replicating vector cannot replicate on its own, but when it coinfects the same cell with HIV-1, it is packaged into a virion-like particle and can be transmitted from cell to cell. Importantly, these vectors replicate using the same machinery that HIV-1 uses, and so they mutate at the same rate. This opens the possibility that conditionally replicating vectors could ‘keep up’ with HIV-1 evolution and prevent HIV-1 escape. In this study, we present mathematical analyses of the co-evolutionary dynamics of HIV-1 and conditionally replicating vectors within a patient. Our results show that with proper genetic design, conditionally replicating vectors can keep pace with HIV-1 evolution, leading to persistent reduction in HIV-1 viral loads. Therefore, this class of gene therapies shows potential for ‘evolution-proof’ control of HIV-1, and merits further investigation in laboratory trials.
| The HIV-1 pandemic has been a major challenge in global public health for decades, and continues to impose crippling burdens of morbidity and mortality worldwide. While recent years have brought major breakthroughs in identifying the protective effect of male circumcision [1], [2], [3] and the transmission-blocking potential of early antiretroviral (ARV) drug therapy [4], the scalability and sustainability of these strategies remains in doubt. ARV therapy can reduce viral loads to undetectable levels, but affected populations must be reached, continuous treatment (and hence investment) is required, long-term ARV use can result in side effects, and there is an ever-present risk that the virus will evolve drug resistance [5], [6], [7]. Meanwhile efforts to develop a protective vaccine, the conventional tool for broad-scale disease prevention, have been unsuccessful so far [8]. As a result, alternative strategies are being investigated, including gene therapy approaches. Gene therapies offer many advantages compared with pharmaceutical drugs, such as low economic cost, ease of administration, and potential to reduce HIV-1 viral loads by sustained interference with the viral life cycle (reviewed in [9], [10]).
All methods of HIV-1 treatment, and many methods of prevention, are challenged by the extremely rapid evolution of the HIV-1 genome. The error-prone reverse transcription of HIV-1 generates a ‘viral swarm’ of HIV-1 genotypes within an individual host. This population of HIV-1 virions is under continual selection, from immune system effectors and medical treatments, resulting in the rapid generation of immune escape variants and drug-resistant mutants. Such resistant mutants are frequently observed in patients under long-term drug treatment and gene therapy [11], [12], [13], [14], [15].
One family of gene therapy approaches offers the intriguing possibility of overcoming the challenges caused by rapid HIV-1 evolution. Conditionally replicating vectors (CRVs) have been discussed as a gene therapy strategy to combat HIV-1 for two decades [16], [17], [18]. Like other approaches to antiviral gene therapy, CRVs are capable of delivering genetic elements into the host cell's nuclear genome to block HIV-1 replication through inhibition and competition [16], [19], [20], [21], [22]. Many proposals for CRVs center on genetic modification of HIV-1 or other lentiviruses, including the addition of viral inhibitory machineries and the deletion of essential genes for viral replication or packaging. The defining characteristic of CRVs is that they can create new virions, and hence transmit from cell to cell, by complementation with wild-type HIV-1; because of this trait, CRVs are also known as mobilization-competent vectors [19], [23]. Because of the clear analogy to naturally-occurring defective interfering particles [24], [25], [26], [27], [28], [29], the term therapeutic interfering particle (TIP) has also been proposed [30].
In contrast to other therapeutic approaches, the CRV approach offers a unique opportunity to overcome the problem of HIV-1 evolution [19], [30]. Because CRVs replicate using the exact same machinery as HIV-1, they have potential to evolve as rapidly as HIV-1. Consequently, it is possible that CRVs can hold their ground in a co-evolutionary arms race, and continuously interfere with HIV-1 replication by generating mutants that match HIV-1 escape mutants. This phenomenon has been reported previously for defective interfering particles [31], [32], [33], [34]. Recently, an experimental study has shown that a CRV expressing small non-coding RNAs that targeted the long terminal repeat (LTR) of HIV-1 was able to suppress HIV-1 viral production without viral escape for one month [19]. A clinical trial has shown that a CRV can be maintained stably within patients, and the long-term presence of the vector does not induce adverse clinical effects [21]. These results highlight the unique promise of CRVs as a strategy for HIV-1 intervention. However, the co-evolutionary dynamics of HIV-1 and CRVs have not been investigated in detail, and the conditions that would allow CRVs to persistently suppress HIV-1 abundance are not known. Important questions have been raised about the conditions that would allow escape mutants of HIV-1 to arise, and the outcomes of co-evolution between the HIV-1 and CRV genomes [9], [11], [16], [19], [21], [22], [30], [35]. In addition, it is possible that CRVs could select for faster replicating HIV-1 strains, thereby potentially causing more severe disease in those susceptible individuals who are not protected [36].
Here we provide the first investigation of the co-evolutionary dynamics associated with the use of a CRV gene therapy against HIV-1. We use a mathematical modeling approach, building upon the best-developed model framework for the within-host dynamics of a CRV therapy against HIV-1 [30], [37]. Thus our analysis focuses on a particular proposed therapy, the therapeutic interfering particle or TIP, but this case study enables us to address general questions about the evolutionary robustness of CRV therapies against HIV-1. Mathematical modeling is well established as a tool for elucidating the mechanisms underlying viral dynamics and evolution [38], [39], [40], and has contributed greatly to our understanding of the dynamics of HIV-1 and the immune system [41], [42], [43], as well as the mechanisms by which antiviral drugs act on the HIV-1 population within hosts [43], [44], [45]. Models have been also used for assessing the potential properties and refining design principles of those proposed interventions, such as novel gene therapies, for which in vivo data are currently lacking. Recently, several modeling studies have investigated the properties of TIPs. Weinberger et al. modeled the in vivo dynamics of HIV-1 and TIPs, and showed that TIPs could reduce the HIV-1 viral load by orders of magnitude to a level comparable to that caused by highly active ARV treatment [37]. In another study, Metzger et al. extended the previous model to study the dynamics of HIV-1 and TIP at three levels (intracellular, within-host and population) to understand the population-level consequences of TIP intervention [30]. That study made qualitative arguments about the likely direction of selection acting on TIP and HIV-1 at different scales, and predicted that competing selection pressures across scales would lead TIP therapies to be evolutionarily robust. However the co-evolutionary dynamics are not considered explicitly in either study, so these hypotheses remain untested.
We study the co-evolutionary dynamics of TIPs and HIV-1 in the peripheral blood within a host with the aim of establishing design principles for the class of CRV gene therapies and addressing safety concerns associated with them. Our aims are to evaluate the long-term persistence and efficacy of TIPs, to clarify the likely selective pressures on HIV-1 arising from the presence of TIPs, and to test whether HIV-1 can evolve resistant mutants that escape TIP inhibition. By constructing and analyzing mathematical models, we show that the dynamics of HIV-1 and TIP follow a characteristic three-phase pattern. The first two phases reflect temporary efficacy of the therapy followed by evolutionary escape of HIV-1, as observed in studies of antiviral drugs as well as other gene therapy approaches. The TIP is strikingly different, though, because under most conditions, it is able to catch up to HIV-1 evolution and continue to exert its therapeutic effects even after the HIV-1 escape mutant has arisen. We hypothesize that this third phase of sustained suppression of HIV-1 may be a general feature of well-designed CRV therapies, as indicated by preliminary empirical findings [19]. By performing sensitivity analysis, we find that the qualitative behavior of the system is robust to differing assumptions about the detailed interactions between HIV-1 and TIP, supporting this notion. We conclude by using our model to propose possible design criteria to enhance the efficacy and robustness of this class of CRVs in the context of HIV-1 evolution.
The TIP is a proposal for a genetically engineered CRV to combat HIV-1 [30], [37]. It would have a genomic structure that is closely similar to HIV-1, but lacking all the structural and envelope genes required for replication and packaging. Upon infection, the TIP integrates its genome into the host cell, but it can only replicate when the cell is further infected by HIV-1 such that the gene products required for viral replication and packaging are available. The TIP would be designed so that its genome is synthesized at a much higher rate than that of HIV-1, resulting in higher production of TIP virions from a dually infected cell. In addition, the TIP can be genetically modified to encode inhibitory agents, such as RNAi [46], [47] and ribozyme nuclease [48], to inhibit HIV-1 production by blocking its life cycle. As explored in earlier modeling studies [30], [37], the efficacy of TIPs is driven by dynamics at several scales. At the cellular scale, resource competition between TIPs and HIV-1 and the direct inhibition of HIV-1 production result in reduced production of new HIV-1 virions [30]. At the scale of within-host dynamics, infection with TIP results in a lower overall HIV-1 viral load (see Refs. 30,37 for detailed analyses of the dynamics between HIV-1 and the TIP without considering evolution). Because TIPs are packaged in HIV-1 structural elements, it has hypothesized that they could transmit between hosts using the same routes as HIV-1, but we do not consider this phenomenon here.
We developed a cross-scale model of the co-evolutionary dynamics of HIV-1 and TIP, based on the biological framework proposed previously by Metzger et al. [30] (Fig. 1). The model keeps track of the competition of multiple strains of HIV-1 and TIP at both the cellular level and the host level. At the host level, infection of susceptible CD4+ T cells (U) by the ith HIV-1 viral strain (denoted as xi, i = 1…m) or by the jth TIP strain (denoted as yj, j = 1…n) leads to HIV-1 infected cells (Hi) or TIP infected cells (Tj), respectively. T cells infected by yj (Tj) can be further infected by xi, becoming dually infected cells (Mij). New HIV-1 viral particles are produced from both HIV-1 infected cells and dually infected cells, whereas new TIP particles are produced only from dually infected cells, since the materials needed for replication and packaging are lacking in TIP infected cells (Tj).
At the cellular level, both the HIV-1 genome and the TIP genome are transcribed to genomic RNAs (gRNAs) in dually infected cells. We follow the assumption of Metzger et al. [30] that dimerization and encapsidation of two copies of genomic RNAs result in three types of proviruses: HIV-1 homozygotes, TIP homozygotes, and heterozygotes with one copy of HIV-1 gRNA and one copy of TIP gRNA. Heterozygotes are not viable to infect other cells owing to the difference in length between the gRNAs of HIV-1 and TIP [30]. Therefore, TIP production is proposed to reduce HIV-1 yield in a cell through pairing with HIV-1 genomes, i.e. a form of resource competition. To further reduce HIV-1 production, the TIP can be engineered to encode an inhibitory factor, such as an RNAi, which blocks the process of HIV-1 genome production and virion formation. Broadly speaking, for the class of gene therapies based on CRVs, the number of HIV-1 virions produced by the dually infected cell can be reduced by two factors: competition with CRV genomes for resources, and direct inhibition by inhibitory factors encoded by CRV genomes. The intracellular model constructed for the TIP is a special case for the general interactions of competition and inhibition between CRVs and HIV-1 genomes.
To model the molecular evolution of HIV-1 and TIP, we assume that the cellular-scale phenotypes of HIV-1 and TIP are determined by the genomes of HIV-1 and TIP and the interactions between them. The genotype-phenotype mapping is determined using a parsimonious approach recently introduced for co-evolutionary models of viruses and the immune system [41], [49]. Each HIV-1 or TIP strain is represented by a digital sequence, reflecting the genetic elements involved in controlling the traits of interest, and trait values arising from interactions between HIV-1 and TIP are determined by a string-matching algorithm (see Methods).
Building on the model of Metzger et al. [30], we consider three critical phenotypic parameters (D, P, and A) in the cellular model. Parameter D is the fraction of inhibition/upregulation of HIV-1 gRNA production, and it models the degree of inhibition exerted by the TIP-encoded repressor on the production of HIV-1 gRNA in a dually infected cell. Mutations in the gene encoding the repressor in the TIP genome, or in the region where the repressor binds in the HIV-1 genome, can change the value of D [30], [37]. We allow the value of D to vary over a range from the situation where HIV-1 production is completely blocked by the inhibitory factor (D = 1) to the situation where the inhibitory factor has evolved to become an activator of HIV-1 production (D<0). Note that since TIP replication and packaging require materials produced by the HIV-1 genome, inhibition of HIV-1 production affects TIP production in a similar way (see Methods). Parameter P is the production ratio of TIP gRNA over HIV-1 gRNA in a dually infected cell. This is the relative rate of TIP genome replication in a dually infected cell, compared to the rate of wild-type HIV-1 genome replication. The value of parameter P will be influenced by the interaction between TIP gRNA and the HIV-1 proteins Rev and Tat, as well as TIP's genome length and manipulations to its splice sites [30], [37]. Accordingly we assume that the value of P is jointly controlled by the two genomes. Parameter A is the replication coefficient, which characterizes the genome replication rate of a HIV-1 strain in a cell infected only with HIV-1, relative to the genome replication rate of the wild-type HIV-1 strain. This factor is a simple phenomenological representation of the evolutionary constraints on the HIV-1 genome, since there may be a fitness cost to mutating away from the wild-type genotype that prevails in the absence of TIPs. Only mutations in the HIV-1 genome can change the value of A.
We implemented these intracellular and within-host mechanisms using a hybrid deterministic-stochastic framework, following techniques developed elsewhere [41], [50]. The dynamics of existing strains of HIV-1 and TIP are modeled by ordinary differential equations, while mutation events are modeled stochastically.
To gain insight into the competition between different strains of HIV-1 and TIP, we started with the two simplest evolutionary scenarios: the ‘HIV-1 mutant’ model and the ‘TIP mutant’ model (see Eqn.S7 and S8 in Text S1). Each of these models considers only three strains: wild-type or ‘resident’ strains of HIV-1 and TIP, and a mutant strain of either HIV-1 or TIP. We assumed that infection with a mutant strain results in different D, P and A values compared to the wild-type strain (see Methods section for details). To reveal the fitness landscape of the mutant against a fixed resident viral population in a host, we performed invasion analysis, in which the effective reproduction number, Reff, of the mutant is calculated when it is introduced into a system where both the resident HIV-1 and the resident TIP are at equilibrium [51]. Reff is a measure of the relative fitness of the mutant compared to the resident strain. When Reff>1, the mutant can invade the system and replace the resident strain, and when Reff<1, the mutant is unable to invade and dies out. We used Reff,H and Reff,T to denote the effective reproduction numbers of the mutant HIV-1 in the ‘HIV-1 mutant’ model and the mutant TIP in the ‘TIP mutant’ model, respectively. Analytic expressions for both quantities are shown in equations (5) and (6) in the Methods section.
We first analyzed the selection pressure on HIV-1 replication, i.e. how the fitness of mutant HIV-1 strains varies as a function of the replication coefficient A. Several lines of evidence show that increases in viral gRNA replication rate beyond the wild-type rate have negative impacts on overall fitness, since the death rate of the infected T cell increases dramatically as viral gRNAs and virus-encoded proteins accumulate inside the cell [52], [53]. In the model, we follow previous studies [54], [55] in assuming that the wild-type replication rate (A = 1) has evolved to be optimal in the absence of TIP, and modeling the dependence of the death rate of the HIV-1 infected T cell on the production rate of HIV-1 gRNA with a concave-up function (Ω(A)) (see Methods). In a dually infected host, we found that the optimal replication rate at which HIV-1 attains maximal fitness in the presence of TIP is very close to the wild-type value (Fig. 2A). This is because the majority HIV-1 virions are produced from singly infected cells, and thus the overall fitness of HIV-1 within hosts is more affected by changes of HIV-1 production in singly infected cells than in dually infected cells. Therefore, HIV-1 strains with notably higher replication rates, which potentially could be more virulent, are unlikely to be selected and transmitted to other hosts.
The parameters D and P are determined by the genomes of both HIV-1 and TIP, so we considered their impact on the fitness of both types of virus. For the fraction of inhibition (D), we allowed it to vary between −1 and 1 to consider both upregulation and inhibition of HIV-1 production by TIP-encoded elements. The optimal value of D for HIV-1 varies with the rate of HIV-1 gRNA replication (Fig. 2A). However since we have shown that the HIV-1 gRNA replication rate will stay near the wild-type level of A = 1 in the presence of TIP, the optimal value of D for HIV-1 is 0 (Fig. 2A). Considering the fitness of TIP mutants, we found that the optimal value of D for TIP is also 0 when A = 1 (Fig. 2C). This suggests that selection will lead the inhibitory factor encoded by TIP to become non-functional over time.
The parameter P characterizes the relative rate of TIP gRNA production in a dually infected cell. In the model, we allowed P to vary from 0 to 30 in line with arguments presented by Metzger et al. [30]. Higher P values cause dually infected cells to produce more TIP virions, while at the same time wasting HIV-1 resources. Consequently, selection on HIV-1 favors low P values (Fig. 2B), while selection on TIP favors high P values (Fig. 2C), as argued previously [30].
To ensure robustness of these results, we further tested two major assumptions made in this analysis. We first tested the sensitivity of the optimal A and D values for HIV-1 to the assumed relationship between the HIV-1 replication rate and the death rate of infected T cells (Fig. S1A). If the death rate of infected T cells is a concave-up function of HIV-1 replication rate, then fitness decreases as the replication rate increases beyond the wild-type level. When the curvature of this relationship is at least moderate as assumed in our main analyses (α = 1 in the function Ω), then the optimal values of A and D remain close to 1 and 0 (the wild-type values), respectively (Fig. S1B). As the curvature becomes weaker, i.e. the stiffness parameter α becomes smaller, the optimal values of A and D become larger. If the death rate of HIV-1 infected cells is a linear function of HIV-1 replication (Ω(x) = 0.7*x; the black line in Fig. S1A), the total number of virions produced in a cell would stay the same irrespective of variations in the HIV-1 genomic production rate, i.e. variations in HIV-1 replication rate do not change its fitness. It can then be concluded from Eqns. (1) and (5) in the Methods section that changes in the values of A and D do not affect the value of Reff,H, and therefore, the presence of CRVs does not select for HIV-1 variants with higher replication rate in this scenario. Second, we explored the possibility of mutation in the Dimerization Initiation Signal (DIS) region of the HIV-1 and TIP genome, which changes the rates of dimerization of different single-stranded genomic RNAs, and thereby changes the distribution of diploid genomes. Intuitively, a lower rate of heterodimer formation would result in higher production of both HIV-1 and TIP virions in dually infected cells, and thus would raise HIV-1 and TIP fitness. However, this benefit may be balanced by fitness costs arising from mutating the DIS region. Numerous experimental studies, as well as the conservation of the DIS sequence in wild-type HIV-1, indicate that mutations in the DIS region lead to reductions in viral fitness [56], [57], [58]. By incorporating conservative assumptions about this reduction in viral fitness into our invasibility model, we found that mutations in the DIS region are not likely to invade for either HIV-1 or TIP (see Text S1).
Taken together, our analyses of mutant invasibility models made the following predictions for HIV-1 and TIP co-evolution (summarized in Table 1): 1) the HIV-1 replication rate will stay at approximately the wild-type level; 2) the TIP-encoded inhibitory factor, if any, will evolve toward a non-functional state (D = 0); and 3) there is a conflict between selection pressures on HIV-1 and TIP regarding the production ratio, P, i.e. selection on HIV-1 favors low P values, but selection on TIP favors high P values. This final conflict sets the stage for a co-evolutionary arms race.
To explore the co-evolutionary dynamics arising from the conflict in selection pressure on P, and to test the predictions of the invasibility analysis, we constructed a multi-strain deterministic-stochastic hybrid model. In the model, each HIV-1 or TIP strain has a unique digital sequence, representing the relevant part of the genome of HIV-1 or TIP (Fig. 3A), and the genotype-phenotype mapping is determined using a recently developed string-matching approach to modeling co-evolution [41], [49]. The parameters P and D (phenotypes) are determined by the degree of matching between the digital sequences of the HIV-1 and TIP (genotypes) infecting a cell (GH,P and GH,D in the HIV-1 genome, and GT,P and GT,D in the TIP genome, shown in Fig. 3A and explained in the Model Overview section); the parameter A is determined by the match between the HIV-1 genotype and a genotype that allows HIV-1 to replicate at a maximal rate (GH,A in HIV-1 genome and GMax in TIP genome, Fig. 3A). Therefore, mutations in the genomes lead to altered intracellular parameters, which in turn lead to changes in the fitness of the mutant virions (see Methods section for details).
Simulation results confirmed the predictions from the invasion analysis with respect to parameters A and D, which stayed close to 1 and 0, respectively (Fig. 3B). The dynamics for parameter P are more complex as a result of the opposing selection pressures on HIV-1 and TIP. Considering the broader dynamics of co-evolution, we found that the system exhibited a robust pattern of three distinct phases, namely the ‘preliminary’ phase (PR), the ‘escape’ phase (ES) and the ‘set-point’ phase (SP). Below, we first describe the three phases in a typical realization of the model (shown in Fig. 3B), then we describe the sensitivity analysis that tested the robustness of the results to the assumptions made.
The ‘preliminary’ (PR) phase is characterized by the dominance of both the wild-type HIV-1 and the wild-type TIP in the viral population. During this phase, the average P value (mean across all extant infected T cells, denoted <P>) was high (close to the maximum value of 30) and the HIV-1 viral load was reduced to a low level, as a result of the high efficacy of the genetically engineered wild-type TIP strain. However, mutant HIV-1 strains were generated rapidly within the host. Those mutant HIV-1 particles that lead to a lower P value when co-infecting a T cell with the wild-type TIP (i.e. those with mutations in the GH,P region in Fig. 3A) possess selective advantages over the wild-type, and thus rose in frequency within the host. Eventually the ‘full escape’ mutant (highlighted as heavy black line in Fig. 3B), whose genomic sequence in the GH,P region is completely different from the GT,P region of the wild-type TIP, was generated. The level of the full-escape HIV-1 mutant increased exponentially after emergence, since it is completely released from suppression by TIP. Consequently <P> decreased rapidly at the end of the PR phase. In contrast to the rising genetic diversity of the HIV-1 population, the TIP mutants with one point mutation in the GT,P region (i.e. the one-point TIP mutants) were generated quickly but remained at a relatively low level during the PR phase. Two-point TIP mutants arose repeatedly throughout the PR phase, but were cleared rapidly from the system each time. TIP evolutionary dynamics differed from HIV-1 in the PR phase because TIP mutants were not selectively advantageous, due to their low fitness when the HIV-1 population was dominated by the wild-type strain.
The ‘escape’ (ES) phase is the time period during which the full-escape HIV-1 mutant dominates the population. The HIV-1 viral load increased to a high level at the beginning of this phase, as a result of HIV-1 escape from suppression by TIP. The average P value in the host (<P>) dropped almost to 0. However, because the majority of the HIV-1 infected T cells were infected with the full-escape mutant, the extant TIP mutants (the 1-point and 2-point TIP mutants) that partially match the full-escape HIV-1 genome possessed selective advantages over the wild-type TIP at the beginning of the ‘ES’ phase, and thus increased in frequency. These mutants gave rise to 3-point TIP mutants, which possessed further advantage in P. Once the 3-point TIP mutants became the dominant strains, the HIV-1 full-escape mutant was replaced by another strain that has the lowest match with the three-point TIP mutant in their representative genome sequences for parameter P. This marked the end of the ES phase; diverse HIV-1 and TIP strains were present in relatively high abundance, resulting in intermediate values of <P> fluctuating around 15.
The system then entered the ‘set-point’ (SP) phase where HIV-1 and TIP settled into a stable coexistence, in which the abundance of each HIV-1 or TIP strain oscillated around a fixed point, and the populations of HIV-1 and TIP were each dominated by two strains (the uppermost red lines in Fig. 3B, with mean long-term abundances at least 10-fold higher than any other strains). By examining the genome sequences of these strains in the regions determining P values, we found that they consist of two matched pairs that are opposite to each other, i.e. each dominant HIV-1 strain has a perfectly matched TIP strain, and the two such pairs are complete mismatches of each other. Therefore, the oscillatory dynamics between the two dominant strains of HIV-1 and TIP can be understood as a ‘match-escape’ cycle: a TIP strain increases in abundance by matching the dominant HIV-1 strain, the HIV-1 strain that escapes the control of this TIP is selected, leading to selection for the other TIP strain, and so on. The oscillatory dynamics resulted in average P values oscillating near an intermediate value of around 15. The set-point viral load of HIV-1 is at 15 virions/µL on average; in contrast, this model predicts a mean HIV-1 viral load of 100 virions/µL in the absence of TIP. As emphasized in earlier work [30], [37], this sustained suppression of HIV-1 is possible because of the high abundance of TIP particles, which leads to a majority of T cells being TIP-infected. As a result, a large proportion of T cells infected with HIV-1 at set-point are co-infected with TIP, and HIV-1 replication is robustly suppressed.
To test whether the results above are robust to changes in the genomic structure, we performed simulations for models assuming higher genome dimensions instead of binary sequences, and different lengths of genome regions corresponding to the phenotypes of interest. For all simulations performed, TIP evolution was able to catch up with HIV-1 after the emergence of the full-escape mutant, leading to establishment of the set-point phase. We quantified the average HIV-1 and TIP viral loads, the average values of P, D and A, and the duration of each phase of the dynamics, and found that the qualitative behavior of the system is robust to changes in genome structure (Fig. 4). There were some quantitative changes in the dynamics, which accord with our intuitive understanding of the system. Longer representative genome lengths led to lower HIV-1 viral loads (Fig. 4A), since it took longer for an HIV-1 escape mutant to arise in those simulations (Fig. 4B). In contrast, higher genome dimensions led to higher HIV-1 viral loads. This is because, during the set-point phase, more HIV-1 mutant strains were available to escape the dominant TIP, however, only one TIP strain was able to match the dominant HIV-1 strain. In essence, the HIV-1 population had a larger genotype space in which to escape suppression by TIP, leading to a lower set-point <P> (Fig. 4E).
The fraction of dually infected cells is an important parameter determining the viability of TIP, as well as other evolutionary outcomes for HIV-1 [59]. In peripheral blood of untreated HIV-1 patients, the frequencies of multiple infection among all infected CD4+ T cells are 2.6% and 7.0% for acute and chronic infection, respectively [60]. Substituting our model parameters into a model for HIV-1 superinfection following the approach developed in a recent study [59], we obtain a prediction that 2.6% of all infected cells will be multiply infected, indicating that our assumed superinfection rate maps onto the lower range of observed values (Text S1). In our simulations, the frequency of T cells dually infected by HIV-1 and TIP among all infected cells is around 0.4%, which is much lower than the predicted value. This is because the majority of dually infected cells would be infected by two TIPs (which is not considered in our model), as a result of effective TIP suppression of HIV-1 viral loads. To explore the potential impact of higher frequencies of dual infection, we extended the model to include the superinfection of HIV-1 infected cells by TIP, and further tested the sensitivity of our results to increases in the superinfection rate. Importantly, the three-phase co-evolutionary dynamics are robust to changes of superinfection rate and the frequency of dually infected cells (Text S1). Increasing the rate of the superinfection rate leads to higher TIP fitness and therefore a lower minimum value of P (Pthreshold) required for TIP invasion (compare Eqn.(4) in the Method section and Eqn.(S15) in Text S1), suggesting that TIP performs better when the superinfection rate increases. However, increasing the rate of superinfection shortens the time required to generate the HIV-1 full-escape mutant, because higher frequency of dually infected cells leads to higher selection pressure on HIV-1 to escape, but the TIP always catches up and the system approaches the same set-point as in our main analysis.
A fundamental challenge in modeling evolutionary processes is defining the genotype-phenotype relationship. For the current analysis, this is most important for determining the parameter P (phenotype) from the GHP and GTP regions of the HIV-1 and TIP genomes (genotype). In the results presented above, we assumed a linear relationship between the production ratio, P, and the percentage of match between the sequences of HIV-1 and TIP genomes (red line in Fig. 5A). Here we tested how the co-evolutionary dynamics are affected by two alternative assumptions about this genotype-phenotype mapping: the production ratio, P, is either a concave-down function or a concave-up function of the percentage of genomic match (blue and green curves in Fig. 5A).
In almost all scenarios, the co-evolutionary dynamics arising from these alternative genotype-phenotype maps were qualitatively similar to the model with a linear function (Fig. S2). In general, models assuming a concave-down function showed better TIP performance (longer time to HIV-1 escape, shorter duration of escape phase, and higher set-point <P>) than models assuming linear and concave-up functions (Fig. S2A). This occurred because the concave-down function gave higher P values for relatively poorly-matching genomes, favoring TIP in the co-evolutionary arms race. Accordingly, the average HIV-1 viral load was always lowest in models assuming a concave-down function. In contrast, for models assuming a concave-up function, it is possible for HIV-1 to escape control by TIPs completely, and for TIPs to be eliminated from the system, when the length of GHP is 4 or greater (Fig. 5B).
A representative simulation showing the elimination of TIP is shown in Fig. 5C. The PR phase showed dynamics similar to those seen before (Fig. 3B), but after the full-escape HIV-1 mutant dominated the HIV-1 population at the beginning of the ES phase, the extant TIP strains (i.e. the wild-type and the one-point mutant strains) declined in abundance and eventually were eliminated from the system. This happened because the production rate of TIP in the dually infected cells was lower than the minimum required for persistence. We calculated the minimum threshold value of the production ratio to be Pthreshold = 2.80 (shown as dotted line in Fig. 5A). When the length of GHP is 4 or 5 sites, both the wild-type TIP and the 1-point TIP mutant give P<Pthreshold when co-infecting a cell with the full-escape HIV-1 mutant (Fig. 5A), causing these TIP strains to decline to extinction. Note that low levels of 2-point TIP mutants are sometimes present at the beginning of the ES phase. In the simulation shown in Fig. 5C, the frequency of T cells infected by these 2-point TIP mutants was not high enough to be further infected by HIV-1 to become dually infected cells, as needed to complete the life-cycle of TIP. However, in some simulations the abundance of these 2-point TIP mutants was slightly higher due to the stochasticity of the system, which enables these mutants to complete their life cycle so that TIP persists in the host. As a generality, the ability of TIP to persist in the system depends on whether the TIP strains present at the onset of the ES phase are able to persist in a system dominated by the full-escape HIV-1 mutant. Under the assumptions of a concave-up function with representative genome length of more than 3, TIP extinction is possible because low P values arising from coinfection of cells with extant TIP strains and the full-escape HIV-1 strain.
In the simulations above, we assumed that HIV-1 and TIP mutate at the same rate and that mutations of HIV-1 and TIP change the phenotypic parameters in dually infected cells in the same way. The assumption of equal mutation rate is appropriate, since TIP genomes are replicated by the exact same mechanisms as HIV-1 genomes. However, since TIP interacts with HIV-1 in dually infected cells in a complicated way involving processes such as genome-protein binding, mutations in the TIP and HIV-1 genomes may impact differently on changes of the phenotypic parameters in dually infected cells. Focusing on the crucial interactions that determine the parameter P, one way that HIV-1 can escape TIP repression is by a mutation in the Tat gene which changes the Tat protein conformation. This change may lead to a lowered binding affinity for extant TIP genomes, thus giving the mutated HIV-1 strain a selective advantage over other strains. In order to regain the higher binding affinity, TIP needs to mutate the Tat-binding region on its genome. Hence, in this particular scenario, the parameter P is determined by mutations in HIV-1 and TIP that act at the amino acid and nucleotide levels, respectively. Therefore, the rate of change in parameter P could be differentially affected by mutations in the HIV-1 and TIP genomes. To account for this complexity, we examined the probability of TIP elimination as a function of the rate at which TIP mutation changes the crucial phenotypic parameter P relative to HIV-1 mutation. We found that the system is broadly robust to variation in the relative rate of phenotypic changes in HIV-1 and TIP (Fig. S3). When a concave-up function was used for the genotype-phenotype mapping, so that elimination of TIP is a possibility, the relative rate of evolution has a strong influence on the outcome in an intuitive manner: the probability of TIP elimination is lower when TIP mutates faster than HIV-1 in the phenotypic space, and vice versa (Fig. S3). When other genotype-phenotype mappings were used, the probability of TIP elimination was unaffected by relative mutation rate.
In the simulations above, we have assumed that HIV-1 mutants in the GH,P region are as competent as the wild-type strain in terms of replication and infection of new cells. However, mutations in the HIV-1 genome change the properties of HIV-1 viral particles, and in general, they are likely to reduce viral replication and infectivity [61], [62], [63], [64]. If the fitness cost of these reductions is greater than the gain from lowering the production ratio, then HIV-1 mutations lowering the value of P would not be selected, i.e. HIV-1 would not be able to escape from TIP. We analyzed the impact of a mutation-induced reduction in HIV-1 fitness on the selection of the full-escape mutant. To make the analysis clear, we considered the scenario when HIV-1 mutants with mutations in the GH,P region have decreased viral infectivity. Reduction in HIV-1 fitness arising from other mechanisms would lead to similar results. We performed invasion analysis for the mutant, and found a tradeoff between lowering P values and decreasing infectivity (Fig. 6). For a given decrease in P, if the corresponding reduction in infectivity is smaller than a threshold value (white line in Fig. 6), then Reff,H>1 and the mutant with lower P value is selected (the shaded area in Fig. 6); otherwise, the wild-type HIV-1 strain is maintained in the system. To confirm that mutant HIV-1 strains are not selected if the fitness reduction (via reduced infectivity) is too high, we simulated the multi-strain model under the simple assumption that all HIV-1 mutants have a 30% reduction in infectivity, with all other parameter values the same as in Fig. 5C. In stark contrast with the dynamics shown in Figs. 3C and 5C, the viral populations were dominated by the wild-type HIV-1 and the wild-type TIP throughout the simulation, and no escape events occurred (Fig. S4). The average P value remained fixed at 30, and the set-point HIV-1 viral load was reduced to 5 virions/µL.
The surface shown in Fig. 6 can be viewed as a fitness landscape for HIV-1 mutants during the PR phase, given a tradeoff between decreasing P and maintaining HIV-1 infectivity. Generation of a full-escape mutant requires several mutational steps. If each step results in a fitness gain, i.e. moving uphill on the surface, then the full-escape strain can be generated by Darwinian selection (route 1 in Fig. 6). However, if one or more mutants along the mutational trajectory have lower fitness than their parents, i.e. deleterious mutations, then the generation of the full-escape mutant requires a low-probability event such as double mutation or stochastic tunneling to cross the fitness valley (route 2 in Fig. 6) [65], [66]. This suggests a design principle for TIPs: it is desirable to design TIPs so that HIV-1 must mutate its conserved genome region (which may induce a high fitness cost to HIV-1) to reduce the production ratio (P). In this way, the selection of HIV-1 escape mutants can be prevented due to the high cost associated with HIV-1 mutation. Interestingly, with regard to their effect on HIV-1 infectivity, the mutations along route 1 act synergistically to reduce the fitness of HIV-1 (i.e. exhibiting negative epistasis), while the mutations along route 2 act antagonistically (i.e. exhibiting positive epistasis). Previous work has shown that the majority of deleterious mutations act antagonistically in HIV-1, i.e. with positive epistasis [67]. This property of HIV-1 genetics suggests that the generation of a full-escape mutant of HIV-1 may be constrained by the current design of TIPs.
In this study, we have used mathematical models to analyze the co-evolutionary dynamics of HIV-1 and a gene therapy delivered by a conditionally-replicating vector (CRV) in the peripheral blood within a host. We have considered the proposed therapeutic interfering particle (TIP) as a case study for this analysis, which has enabled us to build on recent modeling studies, and has provided specificity and context for our findings. We have investigated questions about HIV-1 escape mutants and virulence evolution, and the potential to achieve long-term viral suppression despite viral evolution, as raised in recent CRV studies [19], [21], [30], [35].
Linking models describing dynamics at both the cellular and the within-host level, we have shown that, under most conditions, the TIP strategy is able to circumvent the evolution of resistance by HIV-1. The TIP population is able to keep pace with the evolution of HIV-1, and thus maintains effective suppression of the HIV-1 viral load in the long-term. The long-term dynamics of HIV-1 and TIP have characteristics of a co-evolutionary arms race (also termed as ‘Red-Queen’ dynamics) [68]. HIV-1 mutants that escape TIP suppression have a fitness advantage and rise in frequency, leading to selection for TIP mutants that ‘catch up’ and can suppress the HIV-1 mutants. For the broad range of parameter values and model structures that we analyzed, this cycle of escape and catch-up continues indefinitely, and the TIP population results in long-term control of the HIV-1 infection within a host. This finding points to the potential for a new generation of CRV-delivered gene therapy agents, which co-opt the viral evolutionary process to design robust and ‘evolution-proof’ disease control.
This distinctive pattern of co-evolutionary dynamics is driven by selection on the production ratio, P, which describes the fold increase in genomic RNA production for TIP relative to HIV-1. The invasion analysis shows that selection on this parameter acts in opposing directions for the HIV-1 and TIP populations in a co-infected host (Fig. 2). TIP benefits from high values of P, while HIV-1 benefits from low values and hence is under selection to acquire substitutions that escape TIP by decreasing the degree of genome matching. Then the TIP population is under selection to catch up with HIV-1 mutation to restore a high P value. This arms race underlies the characteristic three-phase dynamics that arise as a robust pattern in our co-evolutionary simulations (Fig. 3). In the preliminary phase at the onset of treatment, HIV-1 mutants appear and rise in frequency due to the lower P they experience. Soon a ‘full-escape’ HIV-1 mutant appears, which experiences no suppression by the wild-type TIP, and rises to dominate the population throughout the escape phase. For most scenarios we considered, this is followed by the set-point phase where the TIP population is able to mutate to match the HIV-1 strains, and sustained suppression of the HIV-1 viral load is achieved.
We explored the circumstances under which TIP could not establish a set-point phase, and found that the TIP population could be eliminated when two conditions are met. First, any fitness cost associated with HIV-1 mutations must be sufficiently low that the HIV-1 mutational steps to escape the repression of TIP are always increasing in fitness (Fig. 6). Second, once the full-escape HIV-1 mutant takes over the population, the production ratios (P) in cells dually infected by extant TIP strains and the full-escape HIV-1 must be below the invasion threshold, i.e. the production of TIP particles is too slow to sustain the TIP population (Fig. 5C). In addition, we find that when elimination of TIP becomes possible, the probability of elimination is decreased if TIP mutations change the parameter P faster than HIV-1 mutations, and vice versa.
Our analysis reveals particular design principles that would enhance the efficacy and safety of TIPs, and we propose that similar principles would apply to other CRV gene therapies. Consideration of the factors that enable TIP to persist leads to the finding that the mapping between genotype and phenotype for the production ratio has paramount importance for the efficacy of TIPs. In simulations with a concave-down curve for the mapping (Fig. 5A), i.e. a lower reduction in P for intermediate mutants that lead to the full-escape mutant, a lower HIV-1 viral load is observed relative to simulations with either a linear or a concave-up curve (Fig. S2A). Therefore, a high priority in designing the TIP must be that it maintains a high P value when it coinfects T cells with HIV-1 strains that have acquired a few escape mutations (i.e. so the TIP and HIV-1 genome regions corresponding to P match partially). A related conclusion is that design principles that increase the minimum value of P, such that even full-escape mutants of HIV-1 do not reduce P below the critical value Pthreshold, will lead to much greater evolutionary robustness.
Another facet of the proposed TIP design is to encode an inhibitory factor that interferes with the HIV-1 life cycle to reduce HIV-1 viral loads [46], [47]. However, the model analysis shows that both HIV-1 and TIP attain maximal fitness when this factor is non-functional, i.e. with no inhibition (D = 0; Fig. 2). This leads to the prediction that both populations will evolve to diminish the activity of the inhibitory factor, which is borne out by simulation results in the multi-strain co-evolutionary model (Fig. 3B). Since HIV-1 viral load can be suppressed when the inhibitory factor is non-functional [30], we propose that the inhibitory factor is not needed for the TIP design. Note that earlier modeling work predicted that mutants that upregulate HIV-1 gRNA production in dually infected cells (i.e. those with D<0) are selectively more advantageous than mutants with no inhibition (D = 0) [30], which differs from our results here. This discrepancy arises because we added the assumption, based on experimental evidence [52], [53], that elevated HIV-1 gRNA production induces costs to HIV-1 fitness by reducing T cell lifetime.
The possible adverse consequences of co-evolution are considered by analyzing the selection exerted by the presence of TIP on the replication coefficient A of HIV-1. Our analysis suggests that the optimal genome replication rate of HIV-1 is dependent on how the infected cell lifetime changes as the HIV-1 genome replication increases, i.e. the function Ω. Under our model formulation, if the death rate of HIV-1 infected cells increases linearly with increases in HIV-1 replication, HIV-1 intra-host fitness remains the same and the presence of TIP does not select for HIV-1 strains with higher replication rate. If the death rate of HIV-1 infected cells increases non-linearly with increases in HIV-1 replication (as assumed in our main analyses), and the curvature α of this relationship is moderate or stronger, then the optimal HIV-1 replication rate stays near the wild-type level in the presence of TIP. However, if the curvature is weak, so that HIV-1 intra-host fitness decreases only slightly as HIV-1 replication rate increases beyond the wild-type level, then the optimal rate of HIV-1 replication can be higher than the wild-type rate in the presence of TIP. Empirical evidence for the relationship between HIV-1 replication rate and infected cell lifetime is not conclusive. Some indirect evidence suggests that higher viral replication rates incur a significant cost in cell lifetime [52], [53]; however, one study reported that HIV-1 replication induces little cytopathic effect on host cells [69], and two other studies showed that the death rate of HIV-1 infected cells appears to be unaffected by the presence of cytotoxic CD8+ T cells [70], [71]. Further experiments examining how HIV-1 fitness changes with variation in the HIV-1 genome replication rate would enable more precise predictions.
As for all models, we made simplifying assumptions during model construction. Most importantly, the representations of the HIV-1 and TIP genomes, and the relationship between these genotypes and the resulting cellular phenotypes, are highly simplified. In the multi-strain co-evolutionary model, each genome is represented by a digital sequence, and mutations in the genomes are mapped to changes in phenotype (i.e. the values of parameters D, P and A) via simple matching algorithms. We have ignored epistatic interactions and assumed that mutations affect the phenotypic parameters additively. In reality, changes in phenotype are affected by viral mutations in complicated ways, which often are not understood completely; this is necessarily the case for therapies like TIP that are still hypothetical. However we note that the digital sequences in our model are an abstract representation of any information encoded in the genome, and are not restricted to representing a particular set of nucleotide loci. Hence any genome properties that influence the phenotype of interest can be represented. To test the robustness of our conclusions, we performed extensive sensitivity analyses for different parameter values, genome structures, genotype-phenotype mappings and superinfection rates. The results show that the qualitative behavior of the system does not depend on these assumptions, beyond the broad findings discussed above. Other factors recently shown to influence the infection-limiting effects of defective interfering particles, such as dose-dependent responses [26], host cell limitation [72], and potential synergy with the host immune response [73], should be explored in future work.
The assumptions pertaining to the parameter P merit special attention, given that parameter's central role in the co-evolutionary dynamics. We have assumed that P is determined by the degree of matching between relevant regions on the HIV-1 and TIP genomes. This assumption is motivated by the necessary interactions between the TIP genome and HIV-1-encoded elements that regulate genome replication [37]. However, some factors proposed to contribute to TIP gRNA over-expression, such as TIP's shorter genome length or re-engineered splice sites [30], may not be influenced directly by the HIV-1 genotype present in the cell. If these factors do act to increase P in a manner independent of HIV-1, then they will have the effect of raising the minimum value of P corresponding to a full-escape mutant. As noted above, this will benefit TIP in the co-evolutionary process, making it more robust to elimination and HIV-1 escape. Our assumption that P is fully controlled by the interaction between genomes is thus conservative with respect to estimating the benefits of TIP.
In our model, TIP competes with HIV-1 for resources through dimerization with HIV-1 gRNA, and we have assumed that TIP and HIV-1 gRNAs dimerize randomly in dually infected cells. Under this assumption, the efficacy of TIP would be undermined if the DIS of HIV-1 or TIP mutates to reduce the rate of heterodimer formation. Our analysis showed that, if the cost of DIS mutation to viral fitness is greater than the gain from lowering the rate of heterodimer formation, then mutations in the DIS are unlikely to arise. Currently available experimental data support this conclusion (see Text S1). However, the margin of safety in these results was small, so if fitness costs for TIP are less than projected or if other mechanisms favor mutations in the DIS region, then this could be a vulnerability of the proposed TIP strategy. We suggest that robustness of heterodimer formation be a focus of on-going research on TIP design. Further experimental investigation is needed to test the stability of the DIS regions of the HIV-1 and TIP genomes in dually infected cells, and alternative approaches to increasing TIP production relative to HIV-1 should be explored.
In the model, we have only considered infection of productively infected T cells in the peripheral blood. There are two other sites of HIV-1 production that can play important roles in the dynamics of infection: long-lived cells and tissue cells. It has been shown that long-lived cells may have a significant influence on viral dynamics during the chronic phase of infection [74]. Modeling both short- and long-lived cells is beyond the scope of our study, but we have performed simulations for models considering only long-lived cells, and found that the three-phase dynamics are robust to this change though they occurred on a longer time scale (data not shown). Infections of tissue cells pose greater challenges. Data show higher multiplicity of infection for HIV-1 in tissue cells, probably resulting from formation of virological synapses [75]. Our model does not consider cells infected with multiple strains of HIV-1 or of TIP, since the intracellular interactions among different strains of HIV-1 and TIP are not currently understood. Intuitively, the majority of infecting viruses in these multiply-infected cells would be TIPs, because of the much higher viral load of TIP. We speculate that this would lead to broader and stronger selective pressure on HIV-1 due to the presence of different TIP variants within the cell. Therefore, the period of the ‘preliminary’ phase would be shorter, and the three-phase dynamics would remain. However, we emphasize that there are fundamental uncertainties about how these processes would play out, and the genetic exchange between viral populations in tissues and in peripheral blood will make the co-evolutionary dynamics more complex. Because a substantial amount of HIV-1 replication occurs in tissues, these dynamics could have significantly impact on the prediction of sustained control of infection in a patient. This prediction should be evaluated further using models that explicitly consider multiply-infected cells in tissue, and the interactions between viral populations in different body compartments, once the biological processes are understood in sufficient detail.
The evolutionary principles developed in this study, including the robust pattern of three-phase dynamics of evolutionary escape and catch-up, provide general lessons for other CRV-based gene therapies against HIV-1. All CRVs transmit from cell to cell via the same basic mechanism of complementation with viable viral genomes, and therefore the within-host model is a general representation of the population dynamics of CRV and HIV-1 strains in a patient. The cellular model was constructed based on the proposed properties of TIPs in particular, but is easily related to other systems by noting its key outputs. As shown in the Methods, the two quantities that link the cellular level to the within-host level are the HIV-1 production rate in both singly and dually infected cells, and the CRV production rate in dually infected cells. These are the fundamental properties of any CRV or mobilization-competent gene therapy system.
From a conceptual perspective, the various strategies of constructing CRVs differ only in how these key quantities change depending on the molecular mechanisms of CRV replication and interference with HIV-1. The evolutionary principles regarding the competition and inhibition between CRVs and the HIV-1 population at the within-host level are the same. Our findings, in conjunction with those of Weinberger et al. [37] and Metzger et al. [30], show that any CRV-based approach has potential to achieve sustained control of an HIV-1 infection if it has the following traits: 1) the CRV is replicated at a sufficiently high level in dually infected cells that it can be persistently transmitted within a host; 2) the CRV competes with HIV-1 for essential resources required for replication and packaging, creating an evolutionary conflict that leads to a co-evolutionary arms race between HIV-1 and CRV; 3) the CRV is designed such that a) it targets a conserved region of HIV-1, and consequently HIV-1 escape mutants cannot be generated due to the high fitness cost, or b) when the full-escape mutant takes over the HIV-1 population, the extant CRV strains are able to catch-up with the full-escape HIV-1, by replicating at a sufficiently high level in cells infected by the full-escape mutant. These general conclusions suggest design principles to ensure the evolutionary robustness of viral gene therapies based on CRVs.
In addition, due to the similarities between CRVs and naturally-occurring defective interfering particles, the results in this study also shed light on the dynamics of defective interfering particles observed both in experimental studies [31] and in natural populations [76]. Earlier theoretical studies have examined the effect of defective interfering particles on viable viral populations [72], [77], [78], [79]. In particular, Kirkwood and Bangham developed a mathematical model to understand the evolutionary dynamics of a wild-type virus with its associated defective particles in serial passage experiments, and concluded that the effects of defective particles were intrinsically unpredictable [77]. However, their model assumed that the defective interfering particles were generated constantly from the extant viable viral population and only interfered with their parent strains. Mutants generated within lineages of defective interfering particles, and the potential impacts of interference between multiple strains, were not considered in their model. By accounting for the possibility of continuous evolution in the interfering particle population, our study shows that the system can approach a sustained co-evolutionary arms race as observed experimentally [32].
The rapid evolution of HIV-1 poses fundamental challenges for all strategies of treatment and prevention. In the long term, HIV-1 evolution can compromise the efficacy of these treatments and even render them useless. This phenomenon is well-known for ARV drug therapies [6], [44] and vaccine candidates [8], but is increasingly recognized for gene therapies as well. For example, Aviron et al. recently analyzed the evolutionary dynamics of HIV-1 for another class of gene therapy approaches, in which T cells are genetically modified such that they confer resistance to HIV-1 infection and replication [50]. The dynamics of HIV-1 under this gene therapy show many parallels with the dynamics of HIV-1 under other non-evolving therapies, such as ARVs. During the initial period of the treatment, the wild-type HIV-1 remains the dominant strain, and the viral load is effectively suppressed. More and more HIV-1 mutants are generated, until the full-escape mutant appears and quickly takes over the virus population. Once the full-escape mutant arises, the genetically modified T cell does not exert its protective effects anymore, and HIV-1 returns to its pre-treatment abundance [50]. Our study suggests that gene therapies based on CRVs, such as the proposed TIP, could continue to inhibit HIV-1 viral production by keeping pace with HIV-1 evolution even after the full-escape mutant is generated. This phenomenon, which mirrors the co-evolutionary dynamics of an immune system, stands in stark contrast with drug treatments and ‘static’ gene therapy approaches [44], [50]. In addition, in the modeling study by Metzger et al., it has been shown that CRVs can potentially act synergistically with ARVs [30]. Since ARVs and CRVs target different components of the HIV-1 life cycle, it is likely that a combination of ARV and CRV therapies would reduce the risk of generation of HIV-1 mutants that escape ARVs and CRVs. Furthermore, our results suggest potential for rational design of gene therapies based on conditionally-replicating vectors to avoid undesirable evolutionary outcomes—in keeping with recent calls to construct ‘evolution-proof’ approaches to disease control [80]. By harnessing the remarkable evolutionary potential of CRVs in this way, this new class of gene therapy agents could contribute a valuable new dimension to the increasingly successful effort to combat the HIV-1 pandemic worldwide.
Following the framework proposed by Metzger et al. [30], the intracellular model keeps track of the abundance of single-stranded HIV-1 and TIP genomic RNAs, as well as dimerized diploid genomes including the HIV-1 homozygote, the TIP homozygote, and the HIV-1-TIP heterozygote. We assume that the level of single-stranded HIV-1 gRNA reaches equilibrium quickly, and that packaging materials are present in excess [30], so the dimerization of genomic RNAs is the limiting step in the process of viral particle formation. Thus, the rate of formation of new HIV-1 virions can be approximated as proportional to the rate of RNA dimerization. The detailed model equations are presented in the Text S1.
We define two parameters, ψ and ρ to link the intracellular model to the within-host model, as in previous work [30]:
The expressions for ψ and ρ are related to intracellular parameters by [30]:(1)where P is the production ratio for genomes and D is the fraction of inhibition, as defined in the main text. The replication coefficient A is the ratio of the HIV-1 replication rate for a mutant genotype to the replication rate of the wild-type HIV-1 genotype. According to Eqn. (1), the value of ψ decreases with increases in D and P, because either stronger inhibition on HIV-1 gRNA production or higher CRV production that competes with HIV-1 production will lead to lower HIV-1 viral production. The value of ρ depends on the value of P, but not D. This is because inhibition of HIV-1 gRNA production limits the gene products available to both HIV-1 and CRV, therefore changes in D affect HIV-1 and CRV production similarly. For an arbitrary HIV-1 genotype, the production rate of HIV-1 virions in dually infected cells, relative to the production rate of wild-type HIV-1 in singly-infected cells, is Aψ. The production rate of TIP virions in dually infected cells, relative to the production rate of the wild-type HIV-1 in singly-infected cells, is then Aψρ.
Note that the parameters P, A, D, ψ and ρ are subscripted in the multi-strain model below in order to specify the value for particular HIV-1 and/or TIP strains. For example, A0 denotes the A value for the wild-type HIV-1 (‘0’ is used throughout to denote the wild-type), and P01 denotes the P value in a cell dually infected by the wild-type HIV-1 and the 1st TIP mutant strain.
To model the dynamics of viral populations within individual hosts, we consider the infection dynamics of CD4+ T cells by m strains of HIV-1 and n strains of TIP. We assume that each T cell can only be infected by a single strain of HIV-1 and a single strain of TIP. The dynamics are described by the following system of ODEs:(2)Uninfected T cells (U) are generated at a constant rate λ, and cleared from the blood at rate d. The ith HIV-1 and jth TIP strains infect T cells at rate k, resulting in HIV-1 infected T cells (Hi) and TIP infected T cells (Tj), respectively. TIP infected T cells can be further infected by HIV-1, becoming dually infected cells (Mij). Because TIP genomes do not encode any protein that is toxic to the cell or that induces immune response, the death rate of TIP infected cells is assumed to be the same as the death rate of uninfected cells. The death rates of HIV-1 infected cells and dually infected cells are modeled as a concave-up function (Ω(), described below) depending on the HIV-1 genomic RNA (gRNA) production rates (A in singly infected cells and A(1-D) in dually infected cells as shown in Eqn S1 and S3 in Text S1).
Viral particles of the ith HIV-1 strain (xi) are produced from both singly and dually infected cells, whereas viral particles of the jth TIP strain (yj) are produced only from dually infected cells. The rate of HIV-1 viral production in cells infected only with wild-type HIV-1 is π. The rates of viral production in cells infected with mutant HIV-1 strains, and in dually infected cells, are scaled relative to π by the relationships given in the previous section on intracellular dynamics. Both HIV-1 and TIP particles are cleared from the system at a constant rate, c.
|
10.1371/journal.pcbi.1006077 | New methods for computational decomposition of whole-mount in situ images enable effective curation of a large, highly redundant collection of Xenopus images | The precise anatomical location of gene expression is an essential component of the study of gene function. For most model organisms this task is usually undertaken via visual inspection of gene expression images by interested researchers. Computational analysis of gene expression has been developed in several model organisms, notably in Drosophila which exhibits a uniform shape and outline in the early stages of development. Here we address the challenge of computational analysis of gene expression in Xenopus, where the range of developmental stages of interest encompasses a wide range of embryo size and shape. Embryos may have different orientation across images, and, in addition, embryos have a pigmented epidermis that can mask or confuse underlying gene expression. Here we report the development of a set of computational tools capable of processing large image sets with variable characteristics. These tools efficiently separate the Xenopus embryo from the background, separately identify both histochemically stained and naturally pigmented regions within the embryo, and can sort images from the same gene and developmental stage according to similarity of gene expression patterns without information about relative orientation. We tested these methods on a large, but highly redundant, collection of 33,289 in situ hybridization images, allowing us to select representative images of expression patterns at different embryo orientations. This has allowed us to put a much smaller subset of these images into the public domain in an effective manner. The ‘isimage’ module and the scripts developed are implemented in Python and freely available on https://pypi.python.org/pypi/isimage/.
| An important component of research into the function of genes in the developing organism is an understanding of both when and where the gene is expressed. Well established molecular techniques can be used to colour the embryo in regions where the gene of interest appears, and researchers will photograph such treated embryos at different stages of development to build up the story of the gene’s use. Small numbers of these expression pattern images may easily be examined by eye, but getting usable information from large collections of such images would take an enormous investment in time by trained scientists. Computational analysis is much to be preferred, but the task is complex and difficult to generalise. The frog Xenopus is an important model for studying vertebrate development, but up till now has had no purely computational methods available for analysing gene expression. Here we present a suite of computational tools based on a range of mathematical methods, capable of recognising the outline of the embryo against a variety of backgrounds, and within the embryo separately recognising areas of both gene expression and natural pigmentation. These tools work over a wide range of embryo shapes and imaging conditions, and, in our opinion, represent a major step towards full automation of anatomical gene expression annotation in vertebrate embryology.
| A significant challenge for current bioinformatics is the computational analysis of large data sets. Recent developments in sequencing technologies have allowed, for example, the investigation of the time course of gene expression in early development of Xenopus tropicalis at high time resolution [1,2]. For a robust understanding of gene expression, the precise anatomical or cellular location of expression is as important as the timing of expression, yet this presents significant challenges for computational analysis. The most advanced work has been done in Drosophila, with the analysis of the time evolution of the spatial pattern of gene expression revealing genes with co-localised expression[3,4,5,6].
The spatial distribution of RNA within an embryo or tissue is typically obtained by in situ hybridisation (WISH) of a probe sequence to the endogenous RNA under study or by protein immunofluoresence, followed by photographic imaging of the required stages and views or sections. Preparation of reagents and optimisation of conditions for a specific protein/gene target may take time, but once done it is straightforward to generate images covering (for example) many different developmental stages.
For studies on the localisation of small numbers of genes, analysis by inspection of the resultant images is likely to be feasible and may provide sufficient descriptive data to answer the biological question at hand. In larger scale screens the number of generated images can grow rapidly to tens of thousands [3] or more [7], and at this level will either require computational analysis or significant commitment by members of the respective model organism community to manually annotate the images; for example with zebrafish [8], Drosophila [6] or Xenopus [9,10]. However, although manual annotation is generally of high quality, it is slow and the required effort is not easily replicated.
Computational analysis is clearly preferred for large numbers of images, although this is not a straightforward task, and may require significant investment of time and expertise to develop a suitable system. The goals of computational analysis are easily stated: to recognize the relevant physical anatomy of the organism in the image, locate the regions which show gene expression, and either label these regions with suitable anatomical terms or transfer them to a model coordinate system within which the expression patterns may be analysed and/or compared. These goals are usually achieved by two distinct processes described as segmentation (recognising compartmentalisation in the image) and registration (fitting the embryo shape in the image to a model), as well as recognising which parts of the segmented image correspond to gene expression.
Image analysis in Xenopus has several specific challenges: the embryos are not normally fully transparent; embryos may display distinct pigmented regions; in embryos that are cleared to make them transparent the outline of the embryo may merge into the image background; and experimental data frequently cover a wide range of development stages and concomitant variety of embryo shapes and sizes. In addition, the earlier development stages are quasi-spherical, and, unlike fly embryos, may present some difficulty in determining the axial orientation within the image. To date there are no published methods for computational image analysis developed for Xenopus.
Here we report a first suite of tools developed for computational analysis of Xenopus in situ images. These tools are capable of cleanly separating the embryo from the image background over a wide range of developmental stages without requiring the background to be either uniform or any specific colour; in situ hybridisation stain and natural pigment are detected independently and can be marked up accordingly; and analysed images at the early quasi-spherical stages can be compared with each other to identify groups of images photographed at the same axial orientation. Application of this solution of the segmentation problem and partial solution of the registration problem has enabled us to analyse a large and highly redundant image collection, selecting a usefully condensed and representative set for public dissemination. Although it remains to provide the ability to register the images in a model coordinate system, we have laid some useful ground work for future progress. The reduced image set may also now be considered for manual image registration, expression pattern extraction and annotation in existing Xenopus community tools such as Xenbase (http://www.xenbase.org, RRID:SCR_003280) [11,12] and XenMARK (https://genomics.crick.ac.uk/apps/XenMARK, RRID:SCR_014924) [9].
Two of us (MG and IP) were motivated to undertake this research by the desire to complement our high resolution time series data in Xenopus tropicalis [1,2] with expression localisation data mined from public image collections, and to promote and enable further work on computational image analysis within the community. Earlier work [9] had suggested a way forward through crowd sourcing of manual annotation, but the generation, and donation to the community, by others of us (AC-U and RP) of a large collection of 33,289 informative in situ images, suggested that we consider computational approaches. This large set of images contained multiple images at given developmental stages for each gene, and we reasoned that a systematic reduction of this (around 10-fold) technical redundancy would yield a more useful and tractable set of images for use by other researchers via submission to Xenbase, the Xenopus the model organism database. Computational tools devised to achieve this would necessarily form a sound basis for further progress in image analysis in Xenopus. We do not at this stage provide a solution to the problem of registering embryo outlines with a model representation.
In this section we present the functional outline and application of each part of our method. More details of the basis of each algorithm are provided in the Methods section below.
In summary, we developed two primary algorithms: (i) embryo-masking through image segmentation to separate the part of the image containing the embryo from the image background without prior constraints on the background colour or texture; and (ii) colour-separation within the embryo outline, to identify the approximate hues of in situ stain, and pigmented and un-pigmented embryo in each image, and mark up the image accordingly. In addition, we developed algorithms to automatically classify large sets of images by background characteristics, and to perform image-clustering on spherical stage images under transformations of scale, rotation, and shear and hence identify groups of images for the same gene and stage but photographed at different orientations. This last was a key tool in applying our methods to the large, redundant image collection described above, and may also provide a way forward to a solution of the general registration problem for Xenopus embryos.
To develop the methods we have drawn on two sources of Xenopus in situ images: firstly, locally hosted images from the XenMARK project[9] covering a wide range of imaging conditions, and second, a collection of 33,289 images provided by two of us (AC-U and RP) which is described in more detail below. Randomly selected images from both these collections were used for validation.
A brief overview of the primary algorithms is given here, with more technical detail presented in Methods §2 and §3.
This algorithm locates the outline of the embryo within the image. We made two assumptions (i) that the distribution of colour and texture within the embryo is distinct from the distribution of colour and texture in the background, and (ii) that at least part of the embryo is more or less centrally located within the image. These are generally reasonable for the great majority of images we have seen during the development of this work. As a useful side-effect we can also detect images where we believe the embryo touches or is intersected by the edges of the image frame.
Images are first processed to remove potential illumination artefacts and then downscaled. The degree of downscaling depends on the image, but is usually between 4- and 32-fold. Colour content and context are analysed for each downscaled pixel, and modelled as a mixture of either two (un-cleared images) or three (cleared images) Gaussian distributions. Pixels are assigned to the most likely distribution, and the image is mapped accordingly. The spatial distribution of each set of assigned pixels over the image is then considered: if a component is spread more uniformly across the image than other more compactly and centrally distributed component(s), then that component is considered to represent background. Isolated foreground regions that are small or close in colour to the background are re-assigned to the surrounding value. Embryo outlines are thus defined as the border between the background and other regions. Then the embryo outline is smoothed and moved inwards by the width of the low resolution pixels used at this stage. A detailed technical description of these processes can be found in Methods §2, and also see Figs 1 and 2 for illustration and examples. The embryo outline and its bounding box, with sides parallel to the image edges, are recorded with the image data, and a flag is set if the embryo outline touches the edges of the image.
This algorithm determines the most likely hues within the previously detected outline of the embryo for stain and pigmentation. Analysis of the colour distribution inside the embryo outline is used to find statistically independent colour components of each image. These components are compared to the (detected) background colour (as bleed-through is possible), and likely stain and pigment colours: stain is assumed to be relatively blue-green and pigment relatively red-brown (see Methods §3). Heat maps of the determined stain and pigment colours are extracted from the data using adaptive thresholding, and overlaid on the image (Fig 1 and Fig 2). A score representing the degree of stain is associated with each image, and can be used to rank image for selection amongst sets of known duplicates (see Methods §6). This was useful in our analysis of the large, highly redundant, image collection (see below).
The overall image analysis workflow consists of 8 steps, these are summarised as follows (see also the visualisations in Figs 1 and 2):
To validate the performance of our algorithms we used visual inspection of significant numbers of images selected by random sampling from our available collections. We would have preferred a purely computational method, but to our knowledge there are no suitable data sets of manually marked up in situ images of Xenopus embryos available. The closest available images were the manually marked up images from the XenMARK project[9], where the stained regions had been ‘registered’ by eye with, and transferred to, the embryo model diagrams. Even had we solved the registration problem for these embryos to enable a computational comparison with these data, we note that the subjective judgement applied during the XenMARK annotation process, as to the presence and limits of stained regions, would be much the same as using an expert annotator to compare side-by-side images of stained embryo and extracted stained regions. We further note that at this stage we were validating the correct identification of in situ stained regions within the images, irrespective of our understanding of their anatomical location.
We therefore validated our algorithms in two ways: quick visual inspection of two thousand images after steps (iii) & (iv) (see Outline Workflow above) to check correct identification of the embryo, as opposed to the background, within the image; and more intensive inspection of two hundred images after final mark up at step (viii) by WISH experts for correct interpretation of the in situ stain within the embryo.
The quick tests assessed three characteristics of the segmentation process: whether the image background component was correctly identified, whether the selected connected region corresponded generally to the embryo, and whether intersections of the frame edge with the embryo were correctly identified. For these tests we randomly sampled 1000 images from our local hosting of the XenMARK database, as well as 1000 images from the Ciau-Uitz/Patient collection. The errors observed were sufficently distinctive as to be effectively non-subjective, and tests results were scored true or false, with false positive and false negative results distinguished for the embryo/image boundary collision test. Most of the tests were passed at well over 99%, with the exception being for embryo/image edge collisions where false positive results were around 5%, depending on which collection images were from. These data are presented in more detail, along with expected error rates and corresponding 95% confidence intervals, in Table 1.
The more intense inspection by two WISH experts looked at the precision of identification of the embryo outline and the extent of both the stained and pigmented regions within the embryo. The 200 tested images were randomly sampled from both the local XenMARK images and the Ciau-Uitz/Patient collection in proportion to the numbers of images in each collection. Each expert assessed the same set of images, but were instructed not to compare notes during this process. The experts were presented side-by-side with the original and marked up images and asked to give a subjective assessment of good, intermediate or bad for each of the three criteria: embryo outline, stained region and pigmented region. For subsequent analysis these assessments were converted to scores of 1.0, 0.5 and 0.0 respectively. These data are presented in Table 2.
Overall the results were encouraging, with both experts rating the algorithm for outline detection and expression domain extent (stained region) as close to or better than 90% in the good or intermediate categories. However it is quite notable that the correlation between the experts’ individual converted numerical scores was only a little over 0.5 for stained regions and 0.6 for pigmented regions. The underlying cause of this apparent discrepancy is likely in the different interpretation of the terms good and intermediate between the two experts, with Expert 1 being consistently more generous at the intermediate/good boundary then Expert 2. These results underscore the general problem in converting the variable intensity of the stained region into a computationally tractable expression pattern. We address this problem in part by providing a two-tone scale for mark up in situ stained regions. The pigmented region generally scored worse than the other criteria, although this is obviously of lower concern. We suspect, but have not shown in detail, that the dissimilarities in scores by the experts were attributed to their different assessment of the impact of artefacts caused by imaging conditions on the extracted pigment patterns. One of the more obvious artefacts affecting annotated stain pattern occurred in images where the embryo was illuminated from one side. In these cases the algorithm tended to interpret darker areas caused by shadow as more intensely stained.
To test the effectiveness of the algorithms, and to give us the opportunity to produce coherent sets of time dependent gene expression images, we applied them to a highly redundant image collection comprised 33,289 individual in situ images of Xenopus laevis embryos. These represented expression of 548 genes over the classical Nieuwkoop & Faber developmental stages [13], mostly between NF stage 6 (32-cell stage) and NF stage 50 (late tadpole stage), with an approximate 10-fold redundancy at each genes and development stage. We refer to this image set in the text as the Ciau-Uitz/Patient collection after its originators (AC-U and RP). This collection has been described previously [14], as have the methods by which they were produced[15]. The images from this collection used to illustrate our method have not been previously published.
The collection had been pre-screened by one of the originators (AC-U) to retain only images of stages with clearly detectable gene expression, and in total the collection contains 2781 gene/stage groups. Embryos had been imaged either directly after histological staining, or after additional treatment with a clearing agent. In general, both types of preparation were available for each gene and stage. Un-cleared embryos had been imaged against an orange/red background, and cleared embryos against a grey background. All images were whole-mount, and although the majority of the images included the whole embryo, almost a third of the images contained close-ups of specific regions of the embryo. Images generally had associated meta-data, notably the gene name or probe/sequence ID and the developmental stage, all embedded in the name of the image file. Nearly all the Stage 22 and later images were lateral views; early stages included mixture of views. See Fig 3 for a visual depiction of the problem and its resolution.
Our aim was to reduce redundancy in this collection by extracting single representative whole-embryo images for each gene in the collection at each developmental stage, and for the cleared and uncleared embryos. In addition, for the earlier quasi-spherical embryonic stages, we also wished to select images to represent the different anatomical views of the embryo and expression patterns.
To achieve these ends we needed two additional algorithms: the first of these simply classifies the images into cleared and un-cleared on the basis of their statistical distributions of pixel colours, whilst the second uses image similarity clustering to identify different views of the spherical stage embryos using the previously detected in situ stain patterns. These algorithms are described in outline here, and more detail is given in Methods §4 and §5.
This algorithm classified the Ciau-Uitz/Patient images into two groups, those with un-cleared and those with cleared embryos. These had been consistently photographed against an orange/red background or a grey background respectively. This knowledge was used to sort the images on the basis of the statistical properties of the distribution of pixel colour in LAB space within each image, using a Gaussian mixture approach. See Methods §4 for details.
We found 18,254 un-cleared images, 15,034 cleared images, and 1 image was rejected as un-classified. Classification was important (a) to allow selection of both cleared and un-cleared images for each gene/stage where both were present, and (b) to allow a mixture of either two (un-cleared) or three (cleared) Gaussian distributions for the embryo/background analysis.
The algorithm assesses similarity between expression patterns, and clusters images into groups, ideally representing different orientations of spherical embryos when photographed from different angles. Image comparison is performed after discovery of the embryo boundary and mark up of stain regions: within a group of same (spherical) stage images, each image is compared to all others by finding the combination of relative shift, rotation, scaling and shearing that maximises the overlap of stained regions. The minimum discrepancy achieved between two images is used as a dissimilarity metric, and pair-wise dissimilarities are used to perform clustering, of which the sub-groups represent the diversity of views of the expression pattern (Methods §5 and Fig 4). This algorithm represents a possible first step towards resolution of the image registration problem, although we take it no further in the current work.
We tested the image classification method using the quick visual inspection described above. We used the same set of 1000 images randomly selected from the Ciau-Uitz/Patient collection, and compared the predicted classification (cleared/un-cleared) with our observations. The image classification tool made no errors. These data are presented in Table 1.
In addition to this, we also assessed the expression pattern clustering performed with spherical stage embryos from the Ciau-Uitz/Patient collection. We had earlier noted that 79 image groups had embryo orientation information embedded in the image file names; this had not been used to support the computational clustering. We therefore compared the partitioning of the images by clustering to the partitioning provided by the image name annotation (see Methods §8), finding the sensitivity and the specificity of expression pattern clustering to annotated orientation to be 60.5% and 74% respectively. This lower sensitivity is primarily caused by non-informative expression patterns (i.e. uniform staining or absence of it) in some embryos, compounded by some imprecision between the described and likely actual viewing angles. On the other hand, the specificity value is explained by imperfect clustering of intrinsically variable expression domains and stain intensities.
The distinctive step for this analysis is to take any group of similar images and rank them according to the extent of in situ stain detected, from which a representative image can be easily selected (Methods §6 and Fig 3A).
The image selection pipeline runs as follows: (a) images are first classified as un-cleared or cleared to determine whether the initial image analysis needs to use two or three Gaussian components; (b) images are then analysed using the primary algorithms (described above) for embryo outline and in situ stain, recording whether the embryo touched the image frame or not, the position of the embryo outline and its bounding box, and the location and overall amount of in situ stain; (c) images are sorted into groups according to gene and developmental stage information; (d) images of spherical stages embryos are further grouped by anatomical viewpoint by clustering on the in situ stain patterns; (e) images within each final group are then ranked by in situ stain content to enable selection of one representative image, in our case the one with most stain, and selecting whole embryo images ahead of partial ones; and finally (f) images where the embryo was rather small are cropped to +15% of the embryo outline bounding box for display purposes. These functions are all provided within the command line Python program ‘select_images’, included in the ‘isimage’ module described below.
Application of the ‘select_images’ program (based on an earlier but fully functional version of the ‘isimage’ module) to the Ciau-Uitz/Patient collection resulted in the selection of 4,852 images, suitable for immediate web display, from the original 33,289 images. This smaller set was submitted to Xenbase, and is displayed on the appropriate gene pages. This effective consolidation of the original collection would have been extremely difficult to achieve by any other method.
To illustrate the power of this analysis to organise this large pool of data, showing the evolution of gene expression patterns during development, we include the selected images for the developmentally important genes prdm1, ank1 and hoxb3 (Fig 5). This illustrates well the importance of separating the spherical stage images by orientation, giving a clear picture of the intricate gene expression patterns developing through gastrulation and the setting up of neural patterning. To view the non-redundant version of the Ciau-Uitz/Patient image collection go to the Expression Search page in Xenbase, enter Patient Lab in the Experimenter field and click on the Search button.
The two main algorithms, for embryo outline detection and stain/pigment decomposition, are the backbone of our image analysis suite and form the primary image analysis workflow described above. The image clustering algorithm was developed as a useful tool for grouping expression patters for early stage embryos, but is also a potential step towards image registration. These algorithms are implemented as parts of a Python module 'isimage’, including the program ‘analyse_image’ which provides access to the algorithms from the command line and allows expression pattern extraction to be performed on a per image basis. Code for these algorithms is made freely available on https://pypi.python.org/pypi/isimage/.
We have presented a general framework for analysis of whole mount in situ hybridisation images in Xenopus which is based on two specific advances. The first advance is to base segmentation around a novel method for building a statistical model of the image based on analysing colour and colour gradient in separate scales. The second advance is in the separation of in situ stain and pigment colouration using a hint based method taking in the prior (per image) determination of likely background colour in the segmentation step.
For the first advance, we have introduced an approach for unsupervised building of the explicit statistical model of the image background. It is based on capturing both colour and colour gradient in two scales and then using Gaussian mixture analysis to find the best separation of segments having different properties. The spatial distribution of segments is then analysed, and the background model is selected. Finally, the resulting boundary is smoothed employing re-normalised probabilities under the background model as the external force in the curvature minimizing PDE.
The novelty of this segmentation algorithm is in the way it utilises colour, spatial and edge information. This is in contrast to existing general purpose image analysis algorithms [16,17], which augment colours with spatial information, using edges as external constraint. Here we have jointly modelled the colour and the colour gradient, thus incorporating edge information into the GMM, whilst spatial distribution of pixels is used downstream to the GMM to classify the Gaussian mixture components. This approach was motivated by the known difficulties encountered by edge detectors in whole mount in situ images: (i) finding the correct transition between background and unstained regions at the edges of cleared and half-transparent parts of un-cleared embryos, and thus missing the correct outline[18]; and (ii) when the embryo is imaged against a feature rich background (for instance, and commonly, crushed ice), and the given approach detects spurious segments in the background[19,20].
Joint modelling of the colour and texture cues brings our algorithm closer to the texture classification method published by Permuter and colleagues[21,22]. This is, however, not completely suitable for in situ images because wavelets (employed in their approach) capture the texture at all scales [23], whereas in segmentation of in situ images variation smaller than a certain scale is unlikely to be significant. Thus the colour and the gradient data modelled by the GMM in our approach were captured in two consecutive layers of the Gaussian pyramid, ensuring that only significantly large texture elements are captured. Such an approach allowed us to put an upper limit on the number of Gaussian mixture components: in un-cleared images the model consisted of 2 components representing the embryo(s) and the background, and in cleared images no more than 3 components were considered, representing staining, the unstained embryo body and the background.
For the second advance, we have suggested a hint based method for pigment/stain separation and subtraction of the background colour, corresponding to bleed-through, from the embryo region of the image prior to analysis. The algorithm estimates the number of independent colours in a masked embryo image based on information theoretic considerations. Then, by employing FastICA algorithm and colour hints provided (primarily that stain is relatively blue-green and pigment relatively red-brown), it estimates and classifies stain and pigment colours. These hints would be configurable for application to other systems.
Our image analysis framework allows the determination of an object’s outline (i.e. the embryo’s outline) in an image, with minimal assumptions about background and object properties. It also allows the extraction of stain patterns from the image whilst excluding natural pigmentation from consideration. The quality of the analysis was independently checked by two WISH experts, who found it performed well. In addition, application of these tools allowed us to reduce redundancy in a set of 33,289 Xenopus embryo WISH images, resulting in 4,852 high quality representative images, making this collection amenable to display in a public resource. Analysis of the image collection, including pattern clustering where needed, took around 12 hours on 40 compute cores. The framework is published as a Python module ‘isimage’, and includes the image selection pipeline and a command line utility to extract the expression pattern from the image.
The significance of our approach to the initial segmentation of the image is well illustrated by its ability to find the embryo boundary in a wide range of image background colours and textures, not to mention variation in the shape, position and orientation of the embryo (seen clearly in Fig 2). This makes it potentially an ideal tool for retro-analysis of existing image collections, and stands in contrast to some of the earlier successes in the field which relied on controlling aspects of the image appearance such as background colour or texture compared to the embryo[3,5], effectively tuning the performance of their algorithms towards the images sets for which they were developed. We believe that our approach has great promise for the development of a more widely applicable tool set.
Our choice of manual validation at different steps in the image analysis pipeline was driven by a number of considerations, not the least of which was the availability of WISH experts to assess performance. In addition, we had some concerns about the potential for unconscious bias in the construction of a gold-standard reference set of manually annotated images, especially in the delineation of in situ stain regions. It is clear that notional boundaries of stained regions are often poorly defined as strong staining shades gradually into weaker and unstained regions, and that subjective judgments of these are inevitably made even by experienced annotators. This might be self-fulfilling if these image sets were constructed by ourselves, or lock the algorithm onto a particular operator bias, producing results with which other experts might not agree. The potential for different interpretation we saw clearly within our own experts and their judgment of how well the stain and pigment recognition algorithm worked. In the absence of a suitable gold standard we felt it was more effective to understand the actual performance of our algorithms, improving them iteratively by studying their behavior, and ultimately allowing other experts to assess their effectiveness. Nevertheless, we suspect there may be room for improvement, and are keen to put these codes into the public domain where others may build on our ideas.
A computational approach to validation was used in a recent paper describing image analysis in Drosophila [24]. They randomly sampled 200 in-situ images and tested the performance of their segmentation and registration algorithms against manually segmented and registered embryos. This may have been important, given the inclusion of the more complex registration step, and given that both segmentation and registration are (presumably) less prone to subjective variation in manual operations than in situ staining.
The primary weakness of the method is in the colour identification of pigmented regions of the embryo, and a tendency to be affected by brightly illuminated or shadowed sections of the embryo, which may say as much about the limitations of digital imaging under extremes of contrast. In this sense, our project has some clear pointers for optimising image generation where it is likely to be associated with subsequent computational analysis: notably avoiding bright and non-uniform illumination, and shadows. The difficulty in correctly identifying the extent of pigmented regions is less of a problem, as we are primarily interested in mapping the in situ stain; but we do believe that mapping the pigmented areas independently ensures that stain identification is more robust. In future work we will turn to the problem of registration, where we hope to be able to use models of known regional expression in combination with a refinement of our image comparison methods based on transformations of position, scale, rotation and shear to identify the likely anatomical viewpoint.
The algorithms developed were implemented as a set of functions and classes in the programming language Python. The code is based on numpy, scipy, sklearn, OpenCV libraries and organised in Python module ‘isimage’ in a way that allows use of either the individual algorithms or the image selection pipeline as a whole. The code can be freely downloaded from https://pypi.python.org/pypi/isimage/.
For each image in the LAB colour space, a Gaussian pyramid [25] is constructed {Ik(x,y)}k=1n; where n is the number of layers in the pyramid and Ik:R2→R3 is the k-th layer of the pyramid. Search for an object is performed simultaneously in two adjacent layers of the Gaussian pyramid. The largest dimension of the biggest layer used is less than 200 pixels. From each of the two layers, colour and edge information are extracted. The edge information is represented as partial derivatives ∂∂xIk(x,y) and ∂∂yIk(x,y) computed with the Scharr operator [26]. The data extracted from the low-resolution layer are interpolated to match the high-resolution layer dimensions, with the same Gaussian kernel used for the pyramid creation. The information from both layers is combined resulting in 18 parameters for each pixel.
The data is then ‘whitened’ and only informative principal components are used in the subsequent analysis. Principal components whose singular values, divided by the sum of all singular values, exceeded 10−6 are considered informative.
To learn the borderline between the background and the foreground, the data is modelled as a mixture of Gaussian distributions. Since in un-cleared images the embryo is very distinct from the background, those images are modelled as the mixture of up to two Gaussians representing an embryo and the background. On the other hand, in cleared images the difference in colour and texture between unstained parts of embryo and the background can be subtle, compared to their difference from stained regions. In these cases, the two-component mixture will often draw the line between stained and unstained areas of the embryo thus counting unstained regions as the background. To handle this, cleared images are modelled as having up to 3 components, with an assumption that background is captured in one component and the embryo is captured in two other components. The actual number of components is determined with Bayesian information criterion[27]. The GM model was fitted using a random sampling of the image, but excluding data within three pixels of the image edges. GM fitting and component classification is repeated three times, then the most likely foreground/background decomposition is brought forward for further analysis.
Once the model is fitted and pixels were assigned to one of the components, there is a need to classify the components themselves as representing either the object or the background.
Here Xi is the set of pixels of the image I, which is classified as belonging to a component Ci. w and h are the width and the height of the image respectively.
To choose the most likely classification of the components, a Bayesian model selection approach is used:
P(M|X)∝P(X|M)P(M)
P(X|Mj)=P(Xi=j|Mj)P(Xi≠j|Mj)
Here M is a random variable representing a model and Mj is the model in which the jth component is believed to be the background. The prior distribution of the models was uniform.
Based on the assumption that the embryo resides in the middle of the picture, background pixels are modelled to be distributed uniformly across the image.
And object pixels are modelled to be distributed normally around the centre of the image.
Here N(0,Σ) is a two-dimensional normal distribution with zero mean and Σ covariance matrix. W−1(Ψ,ν) denotes the inverse Wishart distribution, the conjugate prior to multivariate normal with known mean. The parameters of the prior distributions Ψ and ν were chosen to be ν = 1 and Ψ=[(w2)200(l2)2]. Γ2 is the multivariate gamma function. nj is the number of pixels which belong to the foreground under the jth model.
If none of the models is substantially better, the component whose pixels are present the most often at the image edge is classified as the background.
The assumption is that the image contains only one embryo, but the foreground found above, along with the embryo outline, will have some noise in the form of a number of disconnected islands. In order to remove the noise in the foreground, a connectivity graph is created by connecting adjacent pixels belonging to the foreground. Disconnected sub-graphs of the connectivity graphs are extracted by a spectral graph theory approach[28]. The graph is recursively cut at points where the sorted elements of the eigenvector associated with zero valued eigenvalue of the Laplacian matrix of the graph exhibit the biggest jump exceeding the threshold 10−4.
The sub-graphs are compared based on their size and the difference of the average colour from the average colour of the background. The island having the maximum product of the square root of its size and the difference from the background is selected as the embryo outline.
Where Si is a disconnected sub-graph of the connectivity graph S; MSi is the average colour of pixels in sub-graph Si; Mbg is the average colour of the background.
The steps of splitting the foreground into disconnected components and selecting the most outstanding island are repeated twice, with the first round taking the entire foreground into account and the second round ignoring the parts of the image that are 3 pixels away from the image edges. The outline is considered touching the image edge if the selected foreground component touches the image edge in both cases. The embryos in the image are assumed to have no holes, thus all holes inside the closed contour around the selected foreground element are filled. In case the photographed embryo extends beyond the image, it is sometimes necessary to close the embryo contour along the image edge before filling the holes. To close the embryo contour along the image edge a randomly selected quarter of pixels in the 2-pixel band around the image edge is marked as foreground and then islands containing a single pixel are reverted to the background. The process is repeated until the size of the biggest foreground element increases by less than 5% during the last iteration, but no more than 100 times.
The resulting outline is smoothed by minimizing local curvature of the outline contour using the geodesic active contour framework proposed in [29]. The framework assigns to a curve an energy functional, which depends on the contents of an image. To minimize the functional [29] proposes a contour evolution partial differential equation (PDE), the stationary state of which minimizes the functional. The PDE contains three members on the right hand side; first corresponding to curvature force, which minimizes local curvature; second is balloon force, which tends to expand or contract the contour; third is the image attraction force, which makes the contour reflect the contents of an image. The stationary state of the equation is found using the level set approach as suggested in [29], with zero balloon force everywhere and 8 iterations for curvature force. Since the unsmoothed contour is the collection of points where the probability of belonging to the background under GMM equals the probability of belonging to the foreground, it is natural to use the probabilities as a base for the image attraction force; the log-likelihoods of pixels are first divided by the minimum log-likelihoods for the background and foreground respectively and then summed.
Images with the outline touching the image edge are considered to be presenting incomplete embryo, and a flag is set in the image data to record this. Next, the outline is scaled up to match the original image dimensions. The resulting contour is then contracted by the number of pixels corresponding to half the scaling factor between the original image and the smallest layer in the Gaussian pyramid that is used for the embryo search, to make the final embryo outline. A bounding box is recorded in the image data, which is the rectangle with sides parallel to the images edges that just contains the embryo outline.
In the ‘select_images’ program there is a pre-processing step before stain distribution extraction, since differences in embryo illumination can affect the estimation of the stain intensity. Where appropriate, the background colour is assumed to be the same for all compared images (either cleared or un-cleared images) in a specific gene/stage group where the images are from the same collection. Thus any differences in background colour amongst those images are assumed to be caused by imaging conditions. The differences in average luminosity in all images to be compared are compensated. Then the images are processed independently.
An image is converted into CMY colour space. The model behind the analysis assumes that an image is “painted” with a small number of paints, with each pixel colour being a linear mixture of different amount of each paint:
xi=Asi
Where xi is a CMY colour of ith pixel, A = [as,⋯] is a matrix containing CMY values of each of the paints normalised to unit length as its columns, si is a vector representing amounts of each paints in the pixel.
One approach to find a solution to the equation is independent component analysis; here we use FastICA algorithm to find the independent components [30]. The algorithm solves the equation:
K(X−X¯)=MS
Where M is a symmetric n×n “mixing” matrix; K is a n×m whitening matrix; m is the dimensionality of the data; n≤m is a number of independent components; S is latent “source” variables.
Despite the speed and underlying assumptions of the FastICA algorithm aligned well with the needs of this project, the algorithm has some drawbacks. As seen from the above equation, the FastICA algorithm finds a solution up to a multiplicative constant; it rather finds independent axes in data since the signs or magnitudes of the independent components cannot be determined. Furthermore, since FastICA finds the solutions for mean-centred rather than for zero-centred data an independent axis would correspond to a spectrum of colours rather than to the single colour of the respective paint.
To get around these issues, prior knowledge of the expected colours of the paints is used. From the equation above, the maximum possible number of components cannot exceed the dimensionality of the data, three colour channels in our case. Thus no more than three colour components are expected: stain, pigment, and background where C = [cs,cp,cb]. The background colour is estimated by averaging the colour of pixels outside the embryo outline, whilst stain and pigment colours, blue and brown respectively, were the same for all images in Ciau-Uitz/Patient collection; all expected colours are normalised to unit length.
Not all of the three components will always be present in an embryo image, thus there is a need to estimate the actual number of independent components. The FastICA algorithm finds the solution by choosing such entries in M that minimize the normality of the distribution of “source” variables. From that, it appears natural to estimate the number of components by maximizing the average information content per component as measured by Kullback-Leibler divergence of the empirical distribution of a “source” component from the fit Gaussian distribution.
If the number of components does not match the number of expected colours, some of the expected colours are assumed not present in the image and are removed from the set. At this stage the stain colour is assumed to always be present, thus if the estimated number of components is one the expected stain colour is the only one left in the set. If the estimated number of independent components is two whilst the number of expected colours is three, there is a need to identify which colour is missing. Since it is preferable to detect faint stain, whilst faint pigment can safely be ignored, the assumption at this step is that stain colour is always present and the missing colour is either background colour or pigment colour.
The ambiguity is resolved by maximising the linear combination of the absolute values of the determinant of the correlation matrix between the distribution of the colours in the image and the distribution of independent “sources”, and the cosine between the normals to the planes formed by the vectors left in the set and two first principal components (principal plane) of the pixel colours.
Where C is a matrix containing in its columns n expected colours including stain colour. Xe is colour values of pixels inside the embryo outline. T = M−1K is matrix of independent components. u×v^ means cross-product of u and v normalized to unit length.
If the resulting set of expected colours includes the pigment colour, the stain and pigment colours are adjusted. The adjustment is done by rotating vectors representing the colours in the CMY colour space around their mean by an angle ranging from -15 to +15 degrees in order to make the plane formed by the vectors as parallel as possible to the principal plane of the pixel colours.
Using both independent axes and expected colours, it is possible to estimate the components of A. The estimation is done in two steps: first, the proposed components of A are computed from the data and independent axes; second, the proposed components are compared to expected colours; the set of proposed components closest to the expected colours are accepted.
The computation of the proposed components is based on the assumption that the components of A should be as far as possible from the average colour. Thus, after picking an independent axis and choosing a direction from the mean, the proposed component then equals the normalised to unit length point on the independent axis where the projection of image pixels close to the axis in colour space is maximal. To increase robustness of the method to the colour imprecisions, only pixels sufficiently distant from white colour are used.
Where x is the colour of a jth pixel inside embryo outline, ‖x‖>lmin, it is set to 0.15; mk*^ is the kth column of M* = K+M normalised to unit length; zl∈{1,−1} is a direction with respect to mk*. Subscript notation x∥v means the parallel and x⊥v the orthogonal component of x with respect to v such that x = x∥v+x⊥v; dmax controls the effective distance of the pixel colours from the independent axis, and is set to 0.05.
The best set of proposed components is found by minimizing the weighted average distance between proposed components and the expected colours multiplied by the specificity of the match.
Where A′ is the proposed components matrix; n is the estimated number of independent components; ai′ is the i-th column of A′; ci is the i-th column of Ce; wi is a weight associated with each expected colour, the weights reflected a prior confidence in that colour: ws = 0.5,wp = 0.05,wb = 1. Since the background colour is computed from the image it has the highest confidence. There is less confidence in prior knowledge of the stain colour since it can vary due to imaging conditions, and choice of stain reagents. The argument behind the much smaller confidence level for pigment colour is two-fold; firstly, the colour can vary due to biological differences or imaging conditions. Secondly, the model used here assumes linear colour mixing whereas the imaging condition can produce saturation effects and hence non-linear colour mixing in lighter or darker parts of images, the number of independent components can be overestimated with the superfluous components being far from any of expected. The low weight for the pigment allows that false component to be associated with the pigment in case saturation occurs in the lighter part of the spectrum. In case the pigment component is associated with saturated stain, the estimation of A is done assuming no pigment component is present.
The estimated stain colour as^ is considered confidently estimated if the relative positive contribution of expected stain colour in as^ is over 5%.
The spatial stain distribution is found by solving the equation X=A^S for all image pixels and taking the first “source” components.
‘Adaptive thresholding’ is applied to the stain distribution to make sure that only significant staining is taken into account. This is done by modelling the stain distribution inside the embryo outline as a mixture of two Gaussian distributions, one of which would represent ‘noise’ and the other would be considered the ‘signal’. The noise is filtered out by selecting a threshold so that 95% of noise is under the threshold. The threshold is range limited by the interval [0.25, 0.67] because values of staining/pigmentation below 0.25 level would be too close to white to be significant; on the other hand staining/pigmentation above 0.67 would be significant anyway, even if doesn’t form a pattern.
Where S is the random variable representing staining; t is the threshold.
If the background colour has non-zero projection on the stain colour it results in the “background” noise. To remove the noise from the spatial stain pattern without creating sharp artefacts, a smooth mask is created from the embryo outline as follows. Mean and standard deviation of stain amount in the area outside the embryo outline are computed. The mean amount of stain inside morphological gradient of the embryo outline is computed and recorded. The embryo outline is eroded for one round. These two operations are repeated several times, or until the mean of the stain in the gradient band exceeds the mean plus two standard deviations of the stain in the background. Then a smooth mask is created with a sigmoid profile with the inflexion point located at the distance from the embryo outline where the mean amount of stain in the corresponding gradient band is at the minimum, or if a minimum is not reached, at half way to the maximum of the mean amount of stain.
In the Ciau-Uitz/Patient collection of 33,289 images, un-cleared images had orange/red backgrounds and cleared images had grey backgrounds. They were sorted into groups according to the colour distribution of pixels within each image. Initially, images were converted into LAB colour space, and pixel values accessed via standard library functions.
To capture the colour distribution of pixels in an image, means Mi and covariance matrices Ci were computed for each image. Components of the mean vector and lower triangular parts of the Cholesky decomposition of the covariance matrices were combined to produce a data point representing colour distribution in a particular image.
To learn the best separation between cleared and un-cleared images, the distribution of the data was modelled as a 2–component Gaussian mixture [31]. The model was fit using the expectation maximization algorithm. As a result images were assigned to one of the two components, hence separating cleared and un-cleared images.
Embryo images, in a particular gene/stage groups with recorded stage earlier than 22, are clustered on their expression patterns, with cleared and un-cleared images clustered separately.
To find the distance between images, the spatial stain distribution of each image is aligned with those of all other images in the group. Images of stain distribution are normalized by the standard deviation of the pixel intensities and down-sampled so none of their sides exceeds 100 pixels. Alignment is done by minimizing a function with the squared Euclidian distance between spatial stain distributions of the images as the external energy with respect to linear-affine transformation of one of the images[32]. The distance is penalized for scaling. Minimization is done with BFGS algorithm.
Where Si(⋅) is the spatial stain distribution in an image; α is the regularization constant.
For each pair of images in the group, one of the images is taken as a reference and minimization is done from eight initial positions of the template image: 4 rotations by 90 degrees of original image and the same of a flipped image. Then the process is repeated with the other image taken as a reference. The minimum of the 16 minimal distance values is taken as the distance between the expression patterns.
All the pairwise distances taken with the minus sign formed a similarity matrix. The clustering is done by adaptive affinity propagation algorithm [33] with the number of clusters not exceeding four.
In general, images are ranked on the 85th percentile of the stain distribution and the image of the highest rank in the gene/stage group is selected. In case of clustered images of early embryos, the total similarity of an image in the cluster is added to 85th percentile of the stain distribution to account for how well the image represents the cluster.
Expected error rates and corresponding 95% confidence intervals were calculated under the Bernoulli model with uninformative Beta(1, 1) prior.
Partition of the images by the clustering was compared to the partition by embryo orientation using Wallace pairwise agreement coefficient [34,35]. The Wallace coefficient from partition A to partition B is a ratio WA→B=a(a+b); where a and b are entries of a mismatch matrix [abcd], which in row 1 has the numbers of pairs in the same cluster and in row 2 the numbers in the different clusters of A, and in columns the same for the partition of B. In our case, the Wallace coefficient has the meaning of sensitivity of the clustering to embryo orientation, i.e. the proportion of pairs of images put in the same cluster that have the same embryo orientation. We augmented the clustering sensitivity by a clustering specificity coefficient d(c+d).
All algorithm parameters were tuned using manual procedure similar to cross validation, with a training image set that reflected image collection variability. The procedure consisted of recursively applying the following two steps until optimal parameter values were found. First, parameters were adjusted to allow the algorithm to perform best on a small subset of the training set containing images the algorithm performed worst at. Then the algorithm with parameter value found at the previous step was applied to the whole training set to assess the generality of the value.
|
10.1371/journal.pgen.1000132 | Myc Oncogene-Induced Genomic Instability: DNA Palindromes in Bursal Lymphomagenesis | Genetic instability plays a key role in the formation of naturally occurring cancer. The formation of long DNA palindromes is a rate-limiting step in gene amplification, a common form of tumor-associated genetic instability. Genome-wide analysis of palindrome formation (GAPF) has detected both extensive palindrome formation and gene amplification, beginning early in tumorigenesis, in an experimental Myc-induced model tumor system in the chicken bursa of Fabricius. We determined that GAPF-detected palindromes are abundant and distributed nonrandomly throughout the genome of bursal lymphoma cells, frequently at preexisting short inverted repeats. By combining GAPF with chromatin immunoprecipitation (ChIP), we found a significant association between occupancy of gene-proximal Myc binding sites and the formation of palindromes. Numbers of palindromic loci correlate with increases in both levels of Myc over-expression and ChIP-detected occupancy of Myc binding sites in bursal cells. However, clonal analysis of chick DF-1 fibroblasts suggests that palindrome formation is a stochastic process occurring in individual cells at a small number of loci relative to much larger numbers of susceptible loci in the cell population and that the induction of palindromes is not involved in Myc-induced acute fibroblast transformation. GAPF-detected palindromes at the highly oncogenic bic/miR-155 locus in all of our preneoplastic and neoplastic bursal samples, but not in DNA from normal and other transformed cell types. This finding indicates very strong selection during bursal lymphomagenesis. Therefore, in addition to providing a platform for gene copy number change, palindromes may alter microRNA genes in a fashion that can contribute to cancer development.
| Genetic instability is a key process in the development of naturally occurring cancer. Gene amplification is one important consequence of underlying oncogenic instability. Long DNA palindrome formation is a rate-limiting early step in gene amplification. The development of a new functional genomic tool called genome-wide analysis of palindrome formation (GAPF) led to the recent discovery of the widespread occurrence of such palindromes in both animal tumor models and human tumor cells. Using a Myc oncogene-induced lymphoma model system, this paper describes clustering of tumor-specific palindromes throughout the genome as well as an association of sites of palindrome formation with both preexisting short inverted DNA repeat sequences and occupied Myc DNA-binding sites. We discovered consistent palindrome formation at a cancer-associated microRNA gene called bic/miR-155, beginning at an early stage of tumor development, and without significant further amplification of the locus. Thus, DNA palindromes themselves may directly influence tumorigenesis.
| Oncogenic deregulation of the c-myc protooncogene was first reported in lymphomas of the chicken bursa of Fabricius resulting from Avian Leukosis Virus (ALV) insertional mutagenesis [1],[2]. Retroviral vector-mediated overexpression of c-myc in embryonic bursal precursors induces multi-staged tumorigenesis beginning with preneoplastic transformed follicles (TF) [3],[4] that progress to metastatic B-cell lymphomas. Extensive investigation of deregulated expression of c-Myc more broadly implicates this oncoprotein in a wide variety of animal and human neoplasms [5],[6] as a significant contributor to many of the key processes underlying cancer [7]. These tumor-related phenotypes include genomic instability [8]–[10], which is thought to be essential for the development of most naturally occurring cancers.
Using tools of functional genomic analysis we have recently examined DNA copy number instability during development of Myc-induced bursal lymphomas [11]. For this purpose we employed a 13K chicken cDNA microarray, specifically enriched for chicken immune system and bursal lymphoma ESTs [12], With this tool we carried out array-based comparative genomic hybridization (array-CGH) and detected genome-wide DNA copy number change at many loci in both early TF and end stage lymphomas. Formation of long inverted repeats or palindromes is thought to be a conserved rate-limiting step in eukaryotic gene amplification [13]–[15]. Such palindromes, detected by their ability to form “snap-back” structures in single stranded DNA, have been found in well-characterized tumor-specific amplicons, for example at the c-myc locus itself [16]. A functional genomics tool for the genome-wide analysis of palindrome formation (GAPF) recently has been developed and used to detect the widespread presence of palindromes in human cancer [17],[18]. We adapted this technique to interrogate this Myc-induced tumor system [11], and detected extensive palindrome formation in early TF and end-stage lymphomas. The population of loci showing amplification by array CGH was enriched for palindromes detected by GAPF providing strong evidence for genetic instability early in Myc-induced tumorigenesis and further support for the role of palindromes in gene amplification.
In this report we test, in the bursal lymphoma model, the relationship between the non-random formation of palindromes and pre existing short inverted repeats and describe a positive relationship between Myc expression levels and the abundance and genomic distribution of GAPF-detected palindromes. By combining GAPF with chromatin immune precipitation (ChIP) we discovered an association between palindrome formation and the occupancy of nearby Myc binding sites in chromatin, and, finally, report a strong selection for palindrome formation at an oncogenic micro RNA gene, bic/miR-155, known to cooperate with c-myc in bursal lymphomagenesis [19]–[21].
Conditions used for palindrome detection favor rapid formation of duplex fold back structure in denatured DNA followed by nuclease digestion of residual single strands, as described [15],[16] and detailed in Materials and Methods. In our previous GAPF studies we compared palindromes in normal chicken bursal DNA with DNA from preneoplastic and neoplastic bursal tissue, and employed constructs with direct repeats and inverted palindromic repeats as internal negative and positive specificity controls respectively [11],[22]. Using our chicken cDNA microarray GAPF detected very few palindromes in DNA from normal bursa. In view of this result, and In order to combine GAPF with other techniques for these studies, we modified our GAPF technique to compare bursal tumor cell DNA before and after the “snap-back” procedure, an approach employed successfully by other investigators [15],[17]. We employed this form of GAPF on DNA from a clone (clone 8) of DT40 cells, a bursal lymphoma-derived cell line [23] with a typical genomic complement of amplified and palindromic loci [24]. In repeated experiments we consistently detected hundreds of loci containing presumptive palindromes in DT40 clone 8 cells. We carried out a permutation analysis that compared the distribution of genomic sequence intervals separating 643 putative palindromes in DT-40 DNA with a 1000-fold reiterated random distribution of 643 loci in the genome. This study, described in Figure 1, indicated that the distribution of palindromes was distinctly non-random, and suggested the presence of sites or regions of preference for palindrome formation in this system.
In Tetrahymena, yeast and mammalian cells short inverted repeats (IRs), adjacent to sites of double strand DNA cleavage, mediate long palindrome formation and subsequent gene amplification [13]–[15]. Sequences encoding such IRs, known to be plentiful in mammalian DNA [25], have been postulated to provide substrates for palindrome formation [15]. In order to evaluate this possibility we scanned the chicken genomic sequence for IRs similar to those known to stimulate palindrome formation, and looked for an association with sites of palindrome formation detected in this study. Figure 2 summarizes the results of a chicken genomic scan for short IRs and the overlap of these sites with GAPF detected palindromes in DT40 cells. If we set an arbitrary size range of up to 20 nucleotides for the non complementary loop separating the palindromic arms of the IRs, the number of IRs detected varied from about 104,000 to only 123 as the perfectly paired arm length increased from a minimum 10 to a minimum of 20 nucleotides. Only one palindrome with an over-all length of greater than 200 nucleotides was detected. Figure S1 provides a more complete survey of the frequency of IRs in the chicken genome with a range of perfect complementary arm lengths and a loop size of up to 25 bp.
As summarized in Figure 2, 31.3 percent of the 284 GAPF positive sites that were consistently detected in replicate studies with DT40 DNA contained IRs within 1 Kb up or downstream of the center of the EST sequence for the GAPF positive feature on the array. The hyper geometric probability of a chance association of this magnitude is about 1.8×10−5. A majority of these sites contained multiple IRs with a median loop size of 9 nucleotides, arm lengths between 10 and 20 nucleotides and a range of overall length between 22 and 45 nucleotides. These findings represent a minimum correspondence between long palindromes and preexisting short IRs because IR loop sizes could easily be larger, and we don't have complete sequences for cDNAs on the array.
In order to test for a relationship between Myc binding and palindrome formation we employed a chromatin immune precipitation assay for genome-wide analysis of Myc binding sites (ChIP on chip analysis), and combined this approach with GAPF. We began by analyzing chromatin from DT40 bursal lymphoma cells, which express high levels of cMyc and then examined v-Rel transformed chicken B cell lines in which levels of Myc expression varied >50 fold. The results of the ChIP on chip
Analysis of DT40 bursal lymphoma cells is detailed in Figure 3 and indicates a strong association between some Myc binding sites and some loci of palindrome formation, but does not prove a causal relationship. One prediction of that possibility would be that increasing levels of Myc should increase both the numbers of occupied Myc binding sites and the numbers of palindromes detected by our functional genomic assays. As described in Figure 4 we tested this possibility in chicken B cells transformed by the NF Kappa B related viral oncogene vRel[26]. This oncogene can be used to acutely transform both normal bursal B-cells expressing low levels of cMyc and pre-neoplastic bursal TF cells expressing high levels of cMyc [27]. Figure 4A charts a protocol for generating sister vRel transformed bursal cell lines expressing either high or low levels of cMyc [27]. Figures 4B and 4C demonstrate a marked increase in numbers of occupied Myc binding sites detected by ChIP on chip assays in the high cMyc expressing cells compared to low Myc expressing vRel transformed bursal B-cells. Figure 4D demonstrates the increased numbers of GAPF-detected palindromes in the high Myc vRel transformants and compares these totals detected by GAPF in several DT40 clonal cell lines as well as normal bursa. As also shown in Figure 4D the numbers of GAPF-detected palindromes were not reduced in DT40 cells in which homologous recombination had been reduced by targeted deletion of RAD-54 [28].
These results suggest a quantitative relationship between the level of deregulated Myc overexpression and palindrome formation. Transformation by vRel in the presence of normally low levels of cMyc in bursal cells produced only minimal increases in palindromes suggesting that neither B-cell transformation itself, nor all types of transcription factor activity itself, could induce large numbers of palindromes. Finally, DT40 cells are unusual among vertebrate cell lines in their exceptional capacity for homologous recombination [29], which might contribute to the increased palindrome formation we have observed. This property has been useful for targeted gene deletion studies. Homozygous deletion of Rad 54 in DT40 clone 18 cells has been shown markedly to reduce homologous recombination toward normal levels [28], but did not appear to affect palindrome formation in our experiments (Figure 3 D).
Cultured primary chick embryo fibroblasts express low, barely detectable levels of c-myc RNA and protein. DF-1 is an immortalized morphologically normal chick embryo fibroblast cell line that can be transformed acutely by the oncogenes of a number of avian retroviruses [31]. We carried out GAPF studies with DNA from mass DF-1 cultures and new clones of uninfected DF-1. Low numbers of palindromes were detected in DNA from mass cultures of DF-1 cells. In contrast, when examined by GAPF, DNA from 10 individual DF-1 clones contained a variable, but on average about an 8-fold, increase in total palindromes compared to mass DF-1 cultures. These results in DF-1 cells suggested that formation of palindromes is a stochastic process, and the sequence target size is relatively large compared to the number formed in any one cell, so that patterns are more readily detected in single cell clones than in mass cultures where contributions from individual cells may be diluted below the sensitivity of the assay. DF-1 cells are immortal, and have been cultivated for many years. Therefore we could not distinguish between palindromes generated in-vivo before establishment and those developing, perhaps quite slowly, after establishment in culture.
We then rapidly transformed DF-1 cells by high multiplicity infection with a Myc-transducing retrovirus, HB1 [32]. GAPF studies of Myc-transformed DF-1mass cultures and a series of 10 transformed DF-1 clones initiated two weeks after infection and transformation also revealed palindromes principally, in transformed cell clones rather than mass cultures. As shown in Figure S2 the numbers of palindromes detected by GAPF were not altered by acute vMyc transformation of DF-1 in either mass culture or individual clones. Therefore, and In contrast to our observations in the multistage development of bursal lymphomas, palindrome formation is unlikely to play an essential role in acute Myc transformation of DF-1 fibroblasts
Bic is a non-coding RNA gene discovered as a site of ALV insertional mutagenesis (in addition to activating insertions at c-myc ) selected during tumor development in a large proportion of ALV-induced bursal lymphomas [19],[20]. Subsequently bic was found to encode a highly conserved micro RNA, miR-155, shown to be essential for normal B-cell development [33], to be lymphomagenic when over expressed in transgenic mice [34], and to be over expressed in several forms of human B-cell malignancies [35]–[37]. Although enforced bic overexpression was shown to accelerate the development of ALV induced bursal lymphomas [21], bic was not activated by retroviral insertion in all ALV-induced bursal lymphomas [19] or in derived cell lines like DT 40 [24]. Furthermore, in many experiments with bursal lymphomas induced by Myc-transducing defective retroviral constructs, where ALV insertional mutagenesis does not occur, we did not detect increases in spliced bic RNA [11]. We did, however, detect palindrome formation at bic. In GAPF studies from all of our analyses of preneoplastic TF (three studies), bursal lymphomas (three experiments ) and multiple studies of derivative high Myc expressing transformed B- cell lines such as DT 40 or TF 26 (Figure 3D). In contrast, none of the three GAPF studies of DNA from normal red blood cells or normal bursa, and four studies of low Myc TF 32 cells (vRel transformed normal bursal cells) detected palindromes at bic. Our twelve GAPF studies of untransformed DF-1 mass cultures and clones, and Myc-transformed DF-1 cultures and clones, also failed to detect palindromes at bic. Thus a high level of deregulated Myc oncogene expression was not sufficient by itself to result in a high frequency of palindrome formation at bic. The specific 1 to 1 correlation of palindrome formation at bic with all stages of bursal lymphomagenesis (preneoplastic TF, metastatic lymphoma and derivative established cell lines) indicates a strong selection in the development of these tumors for this otherwise uncommon change at the bic locus, and indicates a mechanistic alternative to retroviral insertional mutagenesis for alteration of bic/miR-155.
In order to validate the GAPF results for the bic locus, and gain an initial insight into the structure of palindromes formed near bic, we carried out blot-hybridization studies of DNA from normal bursa, DT40 cells and bursal lymphoma from an in-vivo tumor before and after the snap-back procedure and digestion with S1 nuclease. Figure 5 depicts a chart of the genomic region around bic including the positions of a Myc-binding E box (CACGTG) exons 1 and 2, the location of miR-155, key restriction endonuclease sites and the results of Southern blot hybridization analysis. These results indicate that the region around bic exon 1 and a segment of exon 2 in DT40 and bursal tumor DNA, but not in normal chicken DNA, retains an S1 resistant duplex character after the snapback procedure, consistent with the GAPF results. It is difficult to assess the significance of the relatively weaker blot-hybridization signal of the S1 resistant bic DNA recovered from tumor cells, in comparison to that from the pre-snap back normal control DNA. Influences on signal strength arising from the amplification procedure, S1 digestion of one normal bic allele, and/or structural heterogeneity at the bic locus in DNA of the tumor cell population may all play a role in diminishing the signal. In any case we conclude that the S1 resistant segments detected are included in a complementary paired region of a palindrome at the bic locus of DT40 and bursal tumor cells that is not present in normal chicken DNA.
We concluded that the deregulated overexpression of Myc oncoproteins in the bursa results in the widespread accumulation of long DNA palindromes not normally present in chicken bursal B-cells. This response did not extend to constitutive overexpression of another oncogenic transcription factor, vRel, nor was it a consistent general property of the transformed state of B-cells or fibroblasts. Analysis of the genomic distribution of these structures with respect to clustering, the presence of pre-existing IRs and occupied high affinity Myc binding site indicated the existence of preferred sites for palindrome formation. However, the sub genomic target size for induction of palindromes still appears to be quite large in comparison to the numbers of palindromic sites found in individual tumors. As with our previous study of gene amplification in this system [11] there was no unique signature of affected genes detected with the very important exception of the Bic/miR-155 locus. These observations, at the DNA level, are reminiscent of the historic detection of highly variable cytogenetic instability accompanying clonal evolution early in tumorigenesis leading to the selection of single essential oncogenic loci at sites of chromosomal rearrangement [38].
The GAPF studies of clones of untransformed DF-1 fibroblasts, in comparison with studies of parental mass cultures, suggests some important caveats in drawing conclusions about DNA palindromes in clonal neoplasms. The known genotoxic effects of long palindromes may dictate the instability of newly formed DNA palindromes in both germ line and normal somatic cells [39] and explain the relatively low numbers of palindromes detected In normal tissue DNA. Clones of morphologically normal DF-1 fibroblasts, however, contained significantly increased numbers of GAPF detected palindromes, which could have accumulated in single cells either during somatic life and/or as a result of immortalization and extended growth in tissue culture after establishment. Comparison of clonal tumor DNA with normal somatic tissue DNA for differences in palindromes, therefore, needs to be interpreted with some caution.
The results of the effects of Myc overexpression in this set of experiments, however, are not explained simply by the clonal nature of tumor cells. For example low Myc expressing TF 32 is a clonal cell line derived from normal bursal cells with many fewer palindromes than the clonal TF 26 line derived from Myc-transformed TF cells. Furthermore non-clonal Myc-induced preneoplastic transformed bursal follicle populations show marked increases in palindrome formation over normal bursal DNA[11]. Therefore genomic instability detected by GAPF in DF-1 clones, but not in either mass transformed cell cultures or transformed DF-1 clones, suggests that this process is not involved in acute single step Myc-induced fibroblast transformation. In contrast multistage development of bursal lymphomas may rely on this type of genomic instability established early in the preneoplastic phases of this disease. This conclusion is strongly supported by the consistent selection of palindromes at the highly oncogenic bic/microiRNA-155 locus during bursal lymphomagenesis. Why bursal target cells apparently differ from DF-1 fibroblasts with respect to palindrome formation and transformation in response to Myc is unknown. These bursal cells are exquisitely sensitive to radiation and other genotoxic agents [3]. Therefore double strand DNA breaks, the initiating lesions in palindrome formation, may be formed and/or repaired in a manner favoring establishment of palindromes in these cells.
It would be desirable to know the molecular consequences of this change. Palindrome formation at bic was not accompanied by gene amplification that was detectable by array CGH in our experiments [11]. Furthermore, and in contrast to ALV-induced lymphomagenesis, in this experimental model we have not detected any change in the low levels of spliced bic RNA during tumor development. We have carried out PCR-based assays (QuantiMir RT kit, SBI Systems Bioscience, Mountain View, CA) of miR-155 in whole normal embryonic and post hatching bursa as well as from preneoplastic TF and bursal lymphomas(not shown). As might be anticipated from the complex regulatory role of this micro RNA in vertebrate B-cell development [33], there was universal expression in all these tissues, but no simple pattern of accumulation correlating with neoplastic change. We do not know, however, the normal level of miR-155 expression in target normal bursal stem cells, which is the critical factor for comparison with expression in Myc-induced neoplastic development. We are left with the speculation that there may be a regulatory alteration in expression, in critical target cells, of this highly conserved pleiotropic micro RNA, which will require further study. In any case these results suggest that, in addition to providing a platform for gene amplification, palindrome formation can be a novel mechanism for oncogenic alteration of micro RNAs.
There are a number of known mechanisms for Myc-induced genomic instability that could contribute to the formation of palindromes that we have detected in these studies. In principle any process that produces DNA double strand breaks could initiate the formation of long palindromes, for example at short IRs. As mentioned, cytogenetic abnormalities associated with Myc overexpression, which are known to result from such breaks, have been described in mammalian cell lines [8],[9] and in DT40 cells [40]. Recently, Gp1bα, a subunit of the vonWillebrand factor receptor complex and a transcriptional target of Myc, has been implicated in a pathway from Myc-driven overexpression to tetraploidy and DNA strand breaks [41],[42]. Additionally, c-Myc has been shown to play a direct role in DNA replication by localizing at early sites of DNA synthesis. When over expressed, c-Myc drives increased replication origin activity and subsequent DNA damage [43]. Either or both of these of mechanisms in principle could contribute to the widespread formation of palindromes that we have observed. Neither mechanism, however, either requires or explains the striking association that we have detected on a cDNA micro array between sites of Myc DNA binding at or near exons and sites of palindrome formation. The association of course could reflect an unknown structural feature shared between sites of Myc binding and palindrome formation with no mechanistic connection to genomic instability. Alternatively it could also reflect an unknown cis-acting effect of Myc binding site occupancy. For example, with respect to DNA strand cleavage, activation of transcription by a number of DNA binding proteins produces transient DNA double stand breaks near the start of transcription as a normal physiological event [44]. As shown in this study pathologically deregulated Myc overexpression results in a marked increase in occupied Myc binding sites where abnormal activation events might, in a stochastic fashion, perturb the normally rapid repair of such breaks and thereby lead to palindrome formation. While at this point clearly speculative, such a mechanism fits well with our observations in this study.
The bursal transplantation model, employing defective c-myc transducing retroviral vectors to generate preneoplastic transformed bursal follicles (TF), and metastatic bursal lymphomas, has been described previously [11]. Generation and culture of TF 26 vRel- transformed TF, and TF 32 vRel-transformed normal bursal cell lines is outlined in Figure 4 and was described previously [27]. DT40 bursal lymphoma cells and DF-1 immortalized chick embryo fibroblasts were maintained as described respectively [23],[31]. DNA for functional genomic analyses was prepared from tissues and cell lines as described [45].
These experiments employed a 13K chicken cDNA microarray [12] used previously for expression profiling, comparative genome hybridization analysis of gene copy number and initial GAPF studies in this system [11], As mentioned in the text we modified our original GAPF protocol [11],[22] essentially as described previously [15]. In brief, one µg of DNA was denatured in 100 µl of 10 mM Tris, 100 mM NaCl at 100° for seven minutes, cooled rapidly on ice for 2 minutes to form fold back duplexes, and single stranded DNA was digested for 60 minutes with 1.6 units/µl of S1 nuclease in 1×New England Biolabs buffer 2 plus 20 mM NaCl at 37°. Duplex “snap back” DNA was then amplified using a whole genome amplification kit (Repli-g, Qiagen, Valencia, CA), reduced in size by digestion with Alu-1 plus Rsa-1, labeled with Cy-5, mixed with equal amounts of amplified, digested and Cy-3 labeled input DNA and hybridized to the microarray in the presence of 0.7 µg/µl of normal chicken Cot-1 DNA. Hybridization, scanning and spot intensity signal filtration (Genepix pro and GD filter software) parameters were as described, and spot-level Cy-5/Cy-3 ratios log2 transformed as before.
Naturally-occurring DNA palindromes in the Gallus gallus genome were detected by scanning the UCSC chicken genome sequence database, May 2006, version 2.1 draft assembly [46],[47]. DNA palindromes were defined as inverted repeat (i.e. cis-wise, self-complementary) sequences having arms of specified minimum length that are connected by loops of specified maximum length. Individual chromosomal maps were searched directly using an R-script [48] routine (see code in Protocol S1) that iteratively identified all sequences meeting user-specified arm- and loop-length criteria For examples see Figure 2. Culling shorter sequences having the same loop retained the longest unique palindromes. Each palindrome center was defined as the midpoint of its loop.
These studies followed, essentially, the procedure of Ren et. al. [49]. In brief, formaldehyde cross-linked and sonicated chromatin fragments were prepared from DT40, TF-26 and TF-32 cells and immunoprecipitated with either 9E11 Myc monoclonal antibody or control 9E10 monoclonal antibody and IgG-2A isotype control immunoglobulin (Abcam, Cambridge, MA). Formaldehyde cross-linking was then reversed and DNA recovered. Immunoprecipitated DNA was amplified by ligation-mediated PCR, and for genome-wide localization, labeled with CY-5 mixed with sonicated Cy-3 labeled input DNA and hybridized to the microarray. Signals were detected and processed in a fashion identical to that employed for GAPF. For gene locus-specific Myc ChiP assays, amplified DNA from immunoprecipitates with chicken Myc antibody and control antibody was subjected to gene-specific quantitative PCR using primers spanning apparent high affinity Myc binding sites, CACGTG, near transcription start sites. For Metaxin-BP the primer pairs were, forward: CTGAGAAGTGAAGCTGCAGTCCTG, reverse: GATTACTCACCCG AGGTATGGCTC; for bic they were, forward: GCTGAGGTGCTCCAGTGGCAG, reverse: CTAGTCTTCTCTTTGTTGCAGGTC. PCR limit products were evaluated by electrophoresis on 3% agarose gels.
Ten µg of input normal chicken DNA was digested with restriction endonucleases pst1, cla 1+apa 1 or cla 1+bgl 1. One µg of normal and DT40 DNA was subjected to the same procedure as for the GAPF studies including rapid denaturation - renaturation, digestion with S1 nuclease and amplification. Ten µg of amplified snap-back normal and DT40 DNA was then digested with the same sets of restriction endonucleases. Blot-hybridization was carried out on 1% agarose gels and evaluated with the same bic cDNA probe used on the microarray. |
10.1371/journal.pcbi.1000715 | Patient Referral Patterns and the Spread of Hospital-Acquired Infections through National Health Care Networks | Rates of hospital-acquired infections, such as methicillin-resistant Staphylococcus aureus (MRSA), are increasingly used as quality indicators for hospital hygiene. Alternatively, these rates may vary between hospitals, because hospitals differ in admission and referral of potentially colonized patients. We assessed if different referral patterns between hospitals in health care networks can influence rates of hospital-acquired infections like MRSA. We used the Dutch medical registration of 2004 to measure the connectedness between hospitals. This allowed us to reconstruct the network of hospitals in the Netherlands. We used mathematical models to assess the effect of different patient referral patterns on the potential spread of hospital-acquired infections between hospitals, and between categories of hospitals (University medical centers, top clinical hospitals and general hospitals). University hospitals have a higher number of shared patients than teaching or general hospitals, and are therefore more likely to be among the first to receive colonized patients. Moreover, as the network is directional towards university hospitals, they have a higher prevalence, even when infection control measures are equally effective in all hospitals. Patient referral patterns have a profound effect on the spread of health care-associated infections like hospital-acquired MRSA. The MRSA prevalence therefore differs between hospitals with the position of each hospital within the health care network. Any comparison of MRSA rates between hospitals, as a benchmark for hospital hygiene, should therefore take the position of a hospital within the network into account.
| The prevalence of hospital acquired infections is widely believed to reflect the quality of health care in individual hospitals, and is therefore often used as a benchmark. Intuitively, the idea is that infections spread more easily in hospitals with a poor quality of health care. This assumes that the rate at which admitted patients introduce new infections is the same for all hospitals. In this article, we show that this assumption is unlikely to be correct. Using national data on patient admissions, we are able to reconstruct the entire hospital network consisting of patients referred between hospitals. This network reveals that university hospitals admit more patients that recently stayed in other hospitals. Consequently, they are more likely to admit patients that still carry pathogens acquired during their previous hospital stay. Therefore, the prevalence of infections does not only reflect the quality of health care but also the connectedness to hospitals from which patients are referred. This phenomenon is missed at the single hospital level; our study is the first to address the connectedness between hospitals in explaining the prevalence of hospital acquired infections. Our findings imply that interventions should focus on hospitals that are central in the network of patient referrals.
| Pathogens that typically cause hospital-acquired infections have an opportunistic nature. These organisms are usually part of the normal bacterial flora of humans and only cause disease when reaching body sites that are normally free from bacterial colonization e.g. when anatomical barriers are breached due to trauma or medical/surgical interventions. For this reason, severe problems with nosocomial pathogens are mainly seen in the very young and elderly and most frequently in institutions such as hospitals and long-term care facilities where patients are treated for acute or chronic conditions.
Methicillin-resistant Staphylococcus aureus (MRSA) is an antimicrobial resistant variant of S. aureus, a common bacteria frequently colonizing healthy humans and animals. Emergence of MRSA is due to the acquisition of a large DNA fragment, which seems to be rare [1],[2]. The expansion of a limited number of MRSA clones that characterizes the current epidemic in hospitals worldwide is therefore believed to be the result of between patient transmission and only to a minor extent due to the ‘de novo’ emergence in patients exposed to antibiotics. MRSA has therefore become the marker with which the success or failure of hospital infection control [3].
The prevalence of the MRSA differs considerably within and between countries [4],[5]. Currently about 30% of the S. aureus causing bloodstream infections in the UK is resistant to methicillin, against only 1% in the Netherlands and Scandinavian countries [6]. Although in high endemic countries MRSA infections are frequent in all hospitals, the proportions are highest in large teaching (tertiary care) hospitals [4],[7], which also report the highest frequency of newly occurring MRSA clones [8]–[11]. The severity of underlying medical condition of the patients, as well as higher antibiotic use and frequency of invasive procedures have been proposed as the main reasons for this difference [3].
Patients can carry MRSA, asymptomatically, for a long time [12]. When readmitted, they may introduce the pathogen acquired during a previous admission into a new hospital [13]. Failure of one hospital's infection control measures can therefore affect the prevalence in hospitals with which it shares patients [14]. Patients are referred to hospitals at different rates depending on the function of hospitals within the health-care system, which likely affect the prevalence at different institutions. These referral patterns might therefore offer an explanation for high MRSA incidence in hospitals of the tertiary referral level [7]. But can referral patterns account for differences in spread between hospitals, and for differences in observed prevalence? To answer these questions, we have been mapping the health care network based on a large national medical registry, and evaluated the occurrence of hospital-acquired infections in different care categories under simulated epidemic conditions.
In 2004, hospital care in the Netherlands was provided through 71 general hospitals, 19 top clinical hospitals and 8 university medical centres (Figure 1A). During the observation period of one year (2004) 1,676,704 patients were admitted from the population of 16.7 million. These patients were admitted for a total of 2,611,452 times, the majority of patients were hospitalised once. The frequency with which patients were readmitted showed a right-skewed distribution (Figure 1B), with still 86 patients being readmitted for more than 52 times. Patients stayed on average 4.3 days per hospital admission, patients who had less hospital admissions stayed longer per admission (Figure 1C), and those who had four hospital admissions had on average the longest (5.6 days) episodes of hospital admission. Moreover, these patients had the highest rate of readmission in different hospitals (Figure 1E&F), whereas patients who were readmitted more frequently tended to return to the same hospital. These frequent attendees were also most likely to stay for only one day.
The individual-based model emulated the dynamics of patient referrals and allows us to assess the spread of hospital-acquired infections. Colonized patients from one hospital spread the pathogen to nearby hospitals within days, but it takes more time –5 to 10 years– before all hospitals encounter it (Figure 2A). The median time to first infection (TFI) for university medical centers (UMCs) was 755 days, the TFI for top clinical hospitals was 1,087 days and the TFI for general hospitals was 1,346 days. At any stage of the epidemic the expected prevalence in UMCs was higher than in general and top clinical hospitals (Figure 2C).
We reconstructed the Dutch national network of hospitals (Figure 3A) with respect to the potential spread of hospital-acquired infections, using patient referral patterns taken from national medical registration (LMR [15]). Within this network, the UMCs show a higher degree of connectedness than the general and top clinical hospitals (Figure 3B). General hospitals had a higher outdegree than indegree, whereas the reverse was true for UMCs, resulting in an 8-fold difference in the indegree between both types of institutions. Top clinical hospitals assumed an intermediate position and showed little difference between indegree and outdegree. Moreover, the indegree relative to the total number of admissions (including patients admitted directly from the community) was much higher in the UMCs compared to the general hospitals. The patient flow through the network was thus directed towards the UMCs.
In order to determine the effect of the directionality of the network, we repeated the analysis of the individual-based model using a dataset with alternative direction. We created a dataset in which all referral probabilities to hospitals were set equal. In the resulting network, both the indegree and outdegree of the UMCs were higher than the other hospital categories, but the outdegree is now higher than the indegree (Figure 4A, B & C). The relative indegree was higher for the general hospitals compared to the other two categories, although there was only a small difference between UMCs and top clinical hospitals. These simulations resulted in slightly higher prevalence in the general hospitals, compared to the top clinical hospitals and UMCs. The differences between the hospitals in connectedness and prevalence are caused by the different hospital sizes, the only parameter that varied between hospitals in this model. This suggests that the short time to first infection of UMCs is due to their absolute high degree of connectedness, while their high relative indegree causes the higher prevalence in UMCs relative to other hospital categories.
We also used two other networks with alternative directions, to test if our observation holds under different conditions. First, we reversed the direction of the network by reversing time in the original dataset, the patients who first visited a general hospital and then a UMC now do the opposite. In this dataset the UMCs still have a higher relative indegree, compared to the general hospitals, although their outdegree is now higher than their indegree (Figure 4D, E & F). These simulations reduced the difference in prevalence between hospitals, with still the highest prevalence in the UMCs. This exact reversion had almost no effect on the TFI of all hospital categories.
Second, we increased the reversed direction in order to decrease the relative indegree of the UMCs to a level below the relative indegree of the general hospital, while keeping both the absolute degree of the UMCs (both indegree and outdegree) above the degree of the general hospitals. These simulations resulted in a lower prevalence in the university medical centres compared to the hospitals of other care categories, whereby the top clinical hospitals had the highest prevalence, reflecting their highest relative indegree (Figure 4G, H & I). This reversion of direction in the network had, just like the previous ones, little effect on the order of TFI for the hospital categories. The results of all three simulation studies with alternative directions, when taken together, strongly suggest that the high prevalence in UMCs relative to other hospital categories is due to directionality of referral patterns, reflected by their high relative indegree.
This study sets a precedent by using data about all hospital admissions obtained from the National Medical Register (LMR [15]) to explore the potential spread of hospital-acquired infections through the Dutch national network of hospitals and describing the effect of nationwide referral patterns on the spread of nosocomial infections like MRSA. This method shows properties of hospitals, such as connectedness within the network, that on the level of a single hospital would not be visible.
In the Netherlands, 98 hospitals provide various forms of specialist care. Within the category of general hospitals, there are considerable differences from hospital to hospital, with some smaller hospitals providing only basic hospital care. Therefore, patients who need advanced medical treatment need to be referred to so-called top clinical hospitals or university medical centres. Top clinical hospitals are large institutions that provide a wide range of clinical specialities and are involved in specialists training and education of doctors and other health care workers. In contrast to university medial centres they are not affiliated with universities and do not include the same comprehensive spectrum of specialities. Within the health care system, the university medical centres occupy a special place as leading hospitals with advanced specialist and final referral functions.
In the Netherlands the hospital admission rate is rather low compared to international standards with 15.6 admissions per 100 inhabitants [16] and an average stay of only 4.3 days. This figure is low, as it also includes day care treatment when patients occupied a bed but do not stay overnight. The majority of patients (73%) are admitted only once to any hospital. Few return twice (17%), three times (5%), or more (5%). Importantly, patients who are admitted twice or three times in a one year period not only have the longest per admission treatment episodes, but are also more frequently readmitted to different hospitals. In this way, all hospitals in the Netherlands become connected and form a network consisting of referred patients who form a bridge between hospitals and provide a path that can facilitate the spread of hospital-acquired infections, such as MRSA, between hospitals.
The individual-based model which emulates the referral characteristics recorded in the LMR, describes the spread of nosocomial infections among hospitals on an individual patient level. It shows that patients who are admitted only two or three times contribute significantly to the inter-hospital spread of the infection and suggests that the prevalence is directly related to the referral level of different hospital categories. This model is, however, unable to provide a mechanistical explanation for the predicted differences in prevalence between hospital categories. For this reason, a simplified model of the hospital network was created. This model weights the contact pattern between hospitals on the basis of average patient referrals between any two hospitals without taking individual referrals and catchment populations into consideration. Despite being a coarse simplification, the hospital network model provides excellent heuristic value as it is able to demonstrate the directionality of the entire network, which is the driving force behind the difference in prevalence between different hospital categories.
Our methods rely on three key assumptions that should be addressed. First, all of our methods do not take account of transmission outside of the hospitals. If community transmission of hospital-acquired infections become a significant factor, the dynamics of the epidemic will ultimately change and the effect of patient referrals between hospitals will be diluted. Community transmission of MRSA is mainly seen in families [17], among military recruits [18], in relation with competitive sport activities [19] and among children in day-care centres [20]. Typical community-acquired (CA-) MRSA is a phenomenon widely described in the USA [21]–[24] but still rather uncommon in Europe. Although CA-MRSA has been identified in Europe in countries with high as well as low MRSA prevalence, it so far remains much less prevalent than health-care associated (HA-) MRSA. Indeed a recent comprehensive study among patients consulting general practitioners in the Netherlands could not find any CA-MRSA in this population [25]. For MRSA, our models will lose validity when CA-MRSA becomes widespread in the general population and the prevalence in the population reaches levels comparable with those in hospitals.
Second, we have assumed a specific measure of connectedness to create the network. However, the construction of hospital networks can be done based on other measures than the one we used, like weighting the contact between two hospitals by the number of patients these hospitals share, or by taking only subsequent admissions into account. These measures would slightly alter the difference in connectedness between the hospital types, but the differences between referral levels would remain (data not shown). However, we feel that exclusion of data about the length of stay and time between admissions would disguise the true utilization patterns that govern the spread of HA-MRSA.
Third, both the individual based model and the measure of connectedness assume homogeneous mixing within the hospital and leave out any ward structure. However, because the medical condition of a patient determines both the ward of admission and his/her health-care use, patients with a certain utilization pattern may mainly meet patients with comparable utilization patterns. This assortative behavior of patients [26] can potentially alter the dynamics of the epidemic, and especially the rate of growth of the epidemic. However, although the different wards may show different dynamics with the different patients they admit, the general direction of the referred patients will still be towards the university hospitals. We therefore expect the difference between hospital categories to still hold in the long run, despite some likely transient effects during the growth of the epidemic.
A higher prevalence of health care-associated infections has been repeatedly demonstrated for tertiary referral centres such as university and teaching hospitals, which also witness the majority of outbreaks of these types of infections. As a conventional explanation, the severity of underlying conditions, more invasive diagnostic and therapeutic procedures and higher rates of antibiotic prescription have been incriminated for this difference. Our model predictions based on the observed admission pattern in the Netherlands, however, suggest a more parsimonious explanation. In the Dutch health care network, the university medical centres admit a large number of referred patients from other hospitals, much more than the top clinical hospitals (Figure 3B). Each university medical centres is therefore connected to a large number of general hospitals as well as a number of top clinical hospitals. This central position within the hospital network puts these hospitals at higher risk of encountering colonized patients. Moreover, the flow of infectious patients through the hospital network is directed towards the university medical centres and we could show that as a direct result of this directionality, prevalence in these hospitals is predictably higher relative to the other categories.
These observations can have important implications concerning hospital infection control. When hospital infection control fails within a single hospital, hospital-acquired infections will start to spread between hospitals, with the most connected ones at the highest risk of both acquiring and spreading the disease. Differentiation of intervention measures over hospital categories, for instance by making the university medical centres the focal point, could then be considered. The exact implementation of such a differentiation is, however, beyond the scope of this paper and should be the focus of further research. Furthermore, our results suggest that differences in prevalence of nosocomial infections between hospitals do not necessarily reflect the success of the hospital infection control measures of individual hospitals. Direct comparisons of infection rates between hospitals may therefore give a distorted view of hospital standards, if national (or regional) health care utilization patterns are not considered. The use of such comparisons, for benchmarking, may therefore lead to a false conclusion about a hospitals effort to reduce nosocomial infections.
In summary we predict that (1) Hospital-acquired infections can spread rapidly from index hospitals to the next referral level. (2) Secondary and tertiary referral hospitals must be prepared for rapid response. (3) High connectedness and the directionality in the health care network towards the university medical centres cause a local build-up of nosocomial pathogens, such as MRSA, and thus a higher prevalence in these hospitals. This should be taken into consideration for benchmarking and the design of national control strategies.
We used the Dutch national medical register from 2004 (Landelijke Medische Registratie LMR [15]), which contains the data about all individual hospital admissions for the total of Dutch hospital organizations of that year. We stratified patients in the LMR based on the number of admissions, , in the one year of data. Per stratum we counted the number of patients, , and measured the distribution of the length of stay, , the time between admissions, , number of hospitals visited, , and the changes between hospitals, . We defined a change between hospitals as an admission to a hospital different from the hospital of the previous admission. For each hospital we counted the number of next admissions in other hospitals to determine the referral probability, , and counted the the number of admissions per hospital to determined the size, .
For reasons of privacy protection, we were not authorized to use the data at individual record level for detailed analysis. We therefore generated a simulated dataset based on the recorded frequencies which describes the individual patient referral patterns that is consistent with the observed patient characteristics in the LMR. This also enabled us to expand the simulated dataset beyond the recorded single year in the LMR to 20 years.
We assumed that each patient's health-care use comes in sequences of a given number of hospital admissions, , and that the time between these sequences, i.e. between the moment of discharge of the last admission in the sequence and first admission in the next sequence, is exponentially distributed. Patients were assigned a hospital of initial admission from the hospital size distribution, , and a number of admissions in this sequence from distribution . The number of changes between hospitals during these admissions was picked from the distribution . If the number of changes was larger than 0, the same was done for the number hospitals visited, picked from the distribution . We assumed that the moment of changing between hospitals was distributed uniformly over the admissions and the choice for the new hospital was based on the current hospital's referral distributions. The length of stay was picked from distribution and time between admissions from distribution for all sequential admissions.
We picked the rate of initial admission, , based on over 1.6 million admitted patients for an entire population of 16 million individuals, at 1/3650 day−1. After the last admission in the sequence, the time to next admission is therefore picked from an exponential distribution with mean . Because the average time between admission sequences is much longer than the average length of colonization, we thus assumed that the colonization status of an individual at the start of an admissions sequence does not depend on this individuals colonization status in the previous admission sequence. We created a dataset for 20 years to allow the epidemic to reach equilibrium level.
Using the individual entries of the simulated dataset we subsequently created a mathematical model that describes the effect of individual patient movements through the hospital network on the spread of hospital-acquired infections. These individuals can either be susceptible or infected. No distinction was made between colonization and clinical infection for the sake of simplicity. Infected individuals () infect susceptible individuals () within the same hospital during one day with rate , where is the total number of patients in the hospital. Therefore, each susceptible has a probability of of getting infected per day. We assume that infectious patients spread the infection to a random sample of the patients within the hospital, and take no ward structure into account. Individuals lose the infection with rate and the mean duration of colonization was set at 365 days [12].
In order to explore the dynamics, we infect 10% of the patients that are admitted to an index hospital on a randomly chosen starting date, and monitor how the infection spreads to other hospitals. The number of colonized individuals at each time step and the time to first encounter of a colonized patient in each hospital (time to first infection, TFI) was recorded. For each index hospital we perform 200 simulations, sequentially repeating these sets of simulations for each 98 hospitals as index hospital, thus performing a total of 19600 simulations. In further analysis, we only include simulation runs resulting in an outbreak larger than a threshold of 1000 colonized persons, to exclude runs that resulted only in small local outbreaks. The results are not sensitive to the exact value of this threshold.
In order to reduce the complexity inherent to the individual-based model, we created a hospital network model assuming transmission parameters between hospitals. All transmission parameters were based on the patient characteristics as observed in the LMR. Thereby, we calculated the infection rate, , from hospital to hospital , using the probability that any referred patient transmits the infection after referral. This probability depends on the patient's length of stay in both hospitals and the rate of losing colonisation between admissions. The infection rates between all hospitals form a 9898 matrix, , which describes the national network of hospitals in terms of potential transmission.
For each admission we calculate the probability that the patient transmits the infection from the referring hospital to the admitting one, . This probability can basically be divided into three separate probabilities: contracting the infection in a referring hospital, , still being colonized on readmission, , and spreading the infection in the admitting hospital, :(1)
The probability of being colonized depends on the length of stay in each referring hospital, , the number of colonized patients in each of these hospital, , and the transmissibility of the pathogen, ; . If we assume that both the infectivity and the number of colonized patients are at a fixed low level, we can simplify this to , where encompasses the transmissibility and low prevalence in the hospital. Because we assume the transmissibility and prevalence are equal in all hospitals, and because the matrix scales linearly with we can leave at unity:(2)
The probability of introduction in the admitting hospital, , in turn depends on the length of stay in the admitting hospital, , the number of susceptible patients, , and the transmissibility of the pathogen, ; . Here, we can assume that the number of newly infected patients is not dependent on the size of the hospital, because ward size is generally not related to hospital size. Therefore, the probability of transmission is directly related to the basic reproduction number per admission, , and becomes . Where denotes the average length of stay in the dataset. Just as before, we assume that the number of colonized patients is low, and the process is not limited by the number of available susceptible individuals:(3)
The probability that a patient is still colonized upon readmission, , depends on the time between discharge and admission, , and the recovery rate, ; . Although overlapping admissions do occur in the data –patients can for instance be moved to another hospital for a specific procedure without being discharged from the initial hospital– we simplify by only taking sequential admissions into account. Any overlapping admission is treated as having a time between admissions, , of 0, thus with :(4) gives the infectious referral rate, per day, between hospitals, where denotes the time span of the dataset. now denotes the probability that any patient will transmit the disease from hospital to within one day. All admissions of all patients combined result in the national hospital network .(5)(6)
The degree with which hospitals connect with the rest of the hospital network through referrals of patients can be divided into two parts. These consist of the indegree , reflecting the total of introductions a single hospital receives from the rest of the hospital network, and the outdegree which reflects the total amount of colonized patients a single hospital exports to the rest of the hospital network. Because the matrix is asymmetric, and may differ.
In order to determine the effect of the difference between inward and outward degree of connectedness, we created a number of datasets with alternative directions. One of these has no direction, the other two have reversed directions. In all three alternatives the university medical centers still have a high degree of connectedness, consistent with the LMR-based network, but a higher outdegree than indegree, contrary to the LMR based network.
We first created a dataset without direction, by setting all referral probabilities in the referral matrix equal, but leaving all other parameters the same as the original simulated dataset. We then created a reverse dataset by reversing the time of the original simulated dataset. The new date of admission of a patient, , is simply calculated as , where is the end date of the dataset, in our case day 7300, and is the discharge date of the patient. This then gives the exact reversion of the original simulated dataset.
In order to reverse the direction of the dataset even further, we created another dataset in the same way as the generated dataset with the characteristics of the LMR, in which we set all referral probabilities to university medical centers, in the referral matrix, to zero. This, however, also lowered the overall degree of connectedness of these hospitals. In order to raise the degree we increased the size of the university medical centers 7 fold. The university medical centers now have a higher outdegree than indegree, while their indegree is still higher than the outdegree of the top clinical hospitals.
Furthermore, we created a number of small datasets of only five hospitals, in which we varied network properties such as directionality and hospital size (See Text S1).
|
10.1371/journal.pcbi.1001073 | Pharmacodynamic Modeling of Anti-Cancer Activity of Tetraiodothyroacetic Acid in a Perfused Cell Culture System | Unmodified or as a poly[lactide-co-glycolide] nanoparticle, tetraiodothyroacetic acid (tetrac) acts at the integrin αvβ3 receptor on human cancer cells to inhibit tumor cell proliferation and xenograft growth. To study in vitro the pharmacodynamics of tetrac formulations in the absence of and in conjunction with other chemotherapeutic agents, we developed a perfusion bellows cell culture system. Cells were grown on polymer flakes and exposed to various concentrations of tetrac, nano-tetrac, resveratrol, cetuximab, or a combination for up to 18 days. Cells were harvested and counted every one or two days. Both NONMEM VI and the exact Monte Carlo parametric expectation maximization algorithm in S-ADAPT were utilized for mathematical modeling. Unmodified tetrac inhibited the proliferation of cancer cells and did so with differing potency in different cell lines. The developed mechanism-based model included two effects of tetrac on different parts of the cell cycle which could be distinguished. For human breast cancer cells, modeling suggested a higher sensitivity (lower IC50) to the effect on success rate of replication than the effect on rate of growth, whereas the capacity (Imax) was larger for the effect on growth rate. Nanoparticulate tetrac (nano-tetrac), which does not enter into cells, had a higher potency and a larger anti-proliferative effect than unmodified tetrac. Fluorescence-activated cell sorting analysis of harvested cells revealed tetrac and nano-tetrac induced concentration-dependent apoptosis that was correlated with expression of pro-apoptotic proteins, such as p53, p21, PIG3 and BAD for nano-tetrac, while unmodified tetrac showed a different profile. Approximately additive anti-proliferative effects were found for the combinations of tetrac and resveratrol, tetrac and cetuximab (Erbitux), and nano-tetrac and cetuximab. Our in vitro perfusion cancer cell system together with mathematical modeling successfully described the anti-proliferative effects over time of tetrac and nano-tetrac and may be useful for dose-finding and studying the pharmacodynamics of other chemotherapeutic agents or their combinations.
| Clinical treatment protocols for specific solid cancers have favorable response rates of 20%–25%. Cancer cells frequently become resistant to treatment. Therefore, novel anti-cancer drugs and combination regimens need to be developed. Conducting enough clinical trials to evaluate combinations of anti-cancer agents in several regimens to optimize treatment is not feasible. We showed that tetrac inhibits the growth of various cancer cell lines. Our newly developed in vitro system allowed studying the effects of tetrac over time in various human cancer cell lines. Our mathematical model could distinguish two effects of tetrac and may be used to predict effects of other than the studied dosage regimens. Human breast cancer cells were more sensitive to the effect on success of replication than the effect on growth rate, whereas the maximum possible effect was larger for the latter effect. Nanoparticulate tetrac, which does not enter into cells, had a larger effect than unmodified tetrac. The combinations of tetrac and resveratrol, tetrac and cetuximab (Erbitux), and nano-tetrac and cetuximab showed approximately additive effects. Our in vitro perfusion system together with mathematical modeling may be useful for dose-finding, translation from in vitro to animal and human studies, and studying effects of other chemotherapeutic agents or their combinations.
| Tetraiodothyroacetic acid (tetrac) is a deaminated thyroid hormone analogue that binds to the integrin αvβ3 receptor for thyroid hormone [1], [2]. Tetrac inhibits binding of agonist L-thyroxine, T4, and 3,5,3′-triiodo-L-thyronine, T3, to the integrin on cultured cells [1], blocking nongenomically-initiated effects of T4 and T3 on signal transduction pathways [2]–[4]. Tetrac also has actions at the receptor independent of T4 and T3, including inhibition of cancer cell proliferation [2]–[4] and angiogenesis [5], [6]. The integrin is largely expressed on tumor cells and dividing blood vessel cells [7]. Acting at the surface of cancer cells, tetrac alters expression of differentially-regulated cancer cell survival pathway-relevant genes. These include upregulation of expression of pro-apoptotic BcL-x short form [3] and other pro-apoptotic genes [8], upregulation of anti-angiogenic thrombospondin 1 and downregulation of several families of anti-apoptotic genes [8], [9]. Covalently bound to the exterior of a nanoparticle, tetrac does not gain access to the cell interior—where it may have thyromimetic activity [10]—and has biological activity at the integrin receptor similar to that of unmodified tetrac, but with desirable effects on cell survival pathway genes that differ from the parent thyroid hormone analogue [8], [9].
To further characterize in vitro the anti-proliferative pharmacodynamics (PD) of tetrac and nanoparticulate tetrac (nano-tetrac), with and without other chemotherapeutic agents, we developed a perfusion bellows cell culture system based on a perfusion (‘hollow fiber’) model. The hollow fiber model was modified by two co-authors (AL, GLD) from a previous system that explored antibiotic pharmacodynamics [11]. The hollow fiber model and perfusion bellows cell culture system allow simulation of concentration-time profiles (pharmacokinetics) expected in humans in an in vitro system and study of the effects over time (PD) of anti-infective and anti-cancer agents in vitro [12], [13]. Such in vitro systems in combination with mathematical modeling can support translation from in vitro to animal models and human clinical trials. The developed pharmacodynamic model describes the full time course of drug effects at various concentrations simultaneously and may be used to predict the effects of other than the studied dosage regimens.
We report here that tetrac and nano-tetrac inhibit cancer cell proliferation on a concentration-dependent basis that can be cell line-specific. Harvesting cancer cells from the perfusion bellows cell culture system permits fluorescence-activated cell sorting (FACS) analysis of cell cycle, and for apoptosis, quantitation of specific pro-apoptotic and anti-apoptotic gene expression by RT-PCR or microarray. Unmodified tetrac and nano-tetrac were tested in this model system for anti-proliferative efficacy alone or in combination with two other anticancer agents, the stilbene resveratrol [14], and commercially-available anti-epidermal growth factor receptor (EGFR) monoclonal antibody (cetuximab, Erbitux). Additive effects were obtained with combinations of tetrac or nano-tetrac and those other chemotherapeutic agents. We report studies in several human cancer cell lines to infer the applicability of the model and to confirm, not surprisingly, that there are dose-dependent differences in responses of specific cell lines to the chemotherapeutic agents tested.
Overall our aim to develop a mechanism-based pharmacodynamic model that characterizes the action of tetrac on human cancer cells in a newly developed perfusion bellows cell culture system was well achieved as described in the present report.
The pharmacodynamics of tetrac as an anti-proliferative agent against human cancer cell lines were examined in the perfusion bellows cell culture system depicted in Fig. 1. Stability of tetrac in the culture system was determined by LC/MS/MS. Without cells, 75% of the original tetrac concentration was detected after 24 h incubation in medium with 10% FBS at both room temperature and 37°C. Tetrac decayed by 12% when incubated with cells at 37°C, indicating that tetrac is relatively stable in the perfusion bellows cell culture system.
At first tetrac induced anti-proliferation of cancer cells was studied in the non-perfusion system. Human glioblastoma U87MG cells were treated with different tetrac concentrations (10−9–10−5 M) for 7 d, with daily replenishment of tetrac. A model incorporating effects of tetrac on both growth rate and success of replication (Fig. 2) adequately described the time course of cell counts as shown by comparison of the model fitted lines to the observed data (Fig. 3A). Tetrac caused a concentration-dependent reduction in U87MG cell proliferation (Fig. 3A), where 10−9 M was least effective, and 10−8 and 10−7 M caused 15% and 28% decreases in cell counts compared to untreated cells after treatment for 7 d (Fig. 3A). Both effects on growth rate and probability of successful replication were required to describe inhibition of cell proliferation of U87MG cells, as shown by a statistically significant (p<0.01) difference in NONMEM's objective function.
Parameter estimates suggested U87MG cells being more sensitive to tetrac's effect on growth rate than to the effect on success of replication (IC50k<IC50R, Table 1). However, the capacity (i.e. the largest possible effect at very high concentrations of tetrac) was higher for the effect on success of replication than the effect on rate of growth (ImaxR>Imaxk). For this model the cell count on day 0 was fixed based on the number of seeded cells. From simulation-estimation experiments (50 replicates, very rich sampling, additive error on log10-scale = 0.1) the median bias was −4% for Imaxk, +25% for IC50k, +0.4% for ImaxR, and −2% for IC50R, using the MC-PEM algorithm in S-ADAPT. When the same bootstrap datasets plus 50 additional ones were run in NONMEM, the median bias was −2% for Imaxk, +16% for IC50k, −2% for ImaxR and −7% for IC50R (nominal results from NONMEM shown in Table 1). Bootstrap results for the actual sampling times in the experiments were similar to those from the rich design (Table 1). The two effects were therefore estimable and distinguishable, both under ideal conditions and in the actual sampling schedule which was employed in our experiment. For additional model evaluation, S-ADAPT with the MC-PEM algorithm was also used to estimate the parameters from the observed data. The S-ADAPT results for ImaxR and IC50R were within 15% of the results from NONMEM, while Imaxk was 22% lower and IC50k was 70% higher than the results from NONMEM. All other parameters were within 40% of their NONMEM estimates. The satisfactory agreement of parameter estimates from two completely different algorithms suggests adequate estimability of the model parameters.
In addition, estrogen receptor-α (ERα)-negative human breast cancer MDA-MB-231 cells (MDA-MB) were treated with 7 different concentrations of tetrac (10−8 to 10−5 M) for 19 d and total cell counts determined every 1–2 d (Fig. 3B). A model with effects on both rate of growth and success of replication (Fig. 2) adequately described the data (Fig. 3B). Parameter estimates from NONMEM are shown in Table 1. The parameter estimates suggest a higher sensitivity for the effect on probability of successful replication (IC50R<IC50k, Table 1) and a larger capacity of the effect on growth rate (Imaxk>ImaxR). Simulation-estimation experiments (50 replicates, additive error on log10-scale = 0.1) showed a median bias of +3% for Imaxk, +9% for IC50k, −2% for ImaxR, and +6% for IC50R, using the MC-PEM algorithm in S-Adapt. For 100 datasets in NONMEM the median bias was +0.5% for Imaxk, −0.4% for IC50k, +0.5% for ImaxR and +4% for IC50R. The bootstrap results based on the actual sampling design which was also rich were similar (Table 1). As for the action of tetrac on U87MG cells, both effects were therefore estimable and distinguishable. In S-ADAPT (MC-PEM), the parameter estimates based on the observed data were within 20% of those from NONMEM for 5 parameters and were within 50% for the other 3 parameters.
Although tetrac had a growth-suppressive effect late in the treatment period, it may also have a proliferative effect on cancer cells (results not shown here). This is thought to reflect access of the agent to the cell interior where it is a modest thyroid hormone agonist (thyromimetic) [9], [10], [15] rather than an inhibitor, as it is exclusively at the cell surface receptor.
To prevent uptake of tetrac by cancer cells, it was reformulated as poly[lactide-co-glycolide] nanoparticle [8], [9], [16]. MDA-MB cells were treated with constant concentrations of 10−6 and 2.5×10−6 M tetrac or nano-tetrac for 9 d. Results indicate that the anti-proliferative effect of nano-tetrac in MDA-MB cells is greater than that of unmodified tetrac (Fig. 4A). MDA-MB cells were also treated with 4 different concentrations of nano-tetrac (10−9 to 10−6 M) for 9 d (Fig. 4B). Based on mathematical modeling, the sensitivity of the MDA-MB cells for the nano-tetrac effect on probability of successful replication was considerably higher than for the effect on growth rate (IC50R = 0.0086 µM, IC50k = 6.3 µM, Table 1), while the capacity was similar for both effects (Imaxk = 1.0, ImaxR = 1.0 at time = 0). Simulation-estimation experiments (50 replicates, additive error on log10-scale = 0.1) showed a median bias of +12% for IC50k, −0.8% for kiR, and +2.5% for IC50R, using the MC-PEM algorithm in S-ADAPT. For 100 datasets in NONMEM the median bias was +4.0% for IC50k, −2.5% for kiR, and −1.3% for IC50R. The bootstrap results based on the actual sampling design are shown in Table 1.
The anti-proliferative effect of nano-tetrac was also concentration-dependent in human glioblastoma U87MG cells. At a nano-tetrac concentration of 10−9 M, cell number was reduced by 36% (control vs. 10−9 M nano-tetrac = 1.45×108±3.3×107 vs. 2.28×108±1.9×107, average±S.D.) after 7 treatment days (Fig. 4C). Modeling suggested a higher sensitivity for the effect on rate of growth (IC50k<IC50R, Table 1) and a higher capacity for the effect on replication (Imaxk < ImaxR). Both IC50k and IC50R were lower for nano-tetrac than unmodified tetrac in U87MG cells indicating a higher sensitivity to nano-tetrac. For both MDA-MB and U87MG cells, the model includes a decrease in ImaxR of nano-tetrac over time in order to adequately describe the observed cell counts. Such a decrease in ImaxR might potentially be due to functional adaptation or presence of subpopulations with different sensitivities to tetrac. Simulation-estimation experiments (50 replicates, additive error on log10-scale = 0.1) showed a median bias of +2.1% for Imaxk, −2.8% for kiR, and +5.7% for IC50R, using the MC-PEM algorithm in S-Adapt. For 100 datasets in NONMEM the median bias was +1.5% for Imaxk, −1.5% for kiR, and +1.3% for IC50R. The bootstrap results based on the actual sampling design are shown in Table 1. The individual measurements presented as symbols in Fig. 4B and 4C are the results from 3 repeat experiments, i.e. one data point represents one experiment at each time point. The error bars in Fig. 4A are standard deviations from 3 experiments.
The plots of observed versus predicted cell counts are presented in Fig. 5 for unmodified and nano-tetrac in U87MG and MDA-MB cells and show that the time course of cell counts was adequately described.
Cells were harvested from the perfusion bellows cell culture system for flow cytometry analysis after 1–3 d of treatment with 10−8 to 10−5 M tetrac. There was a 1.8-fold increase of apoptotic cells with 10−5 M tetrac compared to untreated cells at 1 d (Fig. 6A). By days 2 and 3, all tetrac concentrations caused apoptosis, as determined by TUNEL assay. In cells continuously exposed to tetrac for more than 10 d, only 10−5 M tetrac produced apoptosis consistently (Fig. 6B), suggesting that tetrac may induce some cell proliferation, although the G1 phase was decreased by 50% after 12 d of tetrac treatment. The degree of apoptosis induced by 10−6 M nano-tetrac was 3-fold that of 10−6 M tetrac (Fig. 6C).
We have recently reported that tetrac and nano-tetrac induce gene expression profile changes in MDA-MB cells [8] and medullary thyroid carcinoma cells [9]. Experiments presented here examined pro-apoptotic gene expression in tetrac- and nano-tetrac-treated glioblastoma U87MG cells and MDA-MB cells in the perfusion bellows cell culture system. RNA was extracted from the harvested cells at the end of treatment for RT-PCR studies. Treatment of cells for 2 d with nano-tetrac (10−6 M) increased expression of PIG3, BAD, p21 and p53 in both U87MG and MDA-MB cells (Fig. 7). In contrast, tetrac (10−6 M) did not significantly increase expression of this panel of genes in U87MG cells and, except for c-jun, gene expression in the MDA-MB cells was enhanced to a lesser extent by tetrac than by nano-tetrac. We have previously observed several differences between gene expression profiles that result from treatment with unmodified tetrac and nano-tetrac [9].
Experiments of flow cytometry and gene expression demonstrate the practicality of harvesting tumor cells from polymer flakes in the perfusion bellows cell culture system for studies of post-treatment states of the cells.
We also determined whether tetrac and nano-tetrac had anti-proliferative actions on immortalized non-malignant cells. Monkey kidney epithelial CV-1 cells and human embryonic kidney 293T cells were treated daily with 10−6 M tetrac or 10−6 M nano-tetrac for 7 d. There was no significant change in cell numbers or morphology (results not shown here) when untreated cells were compared with those exposed to tetrac or nano-tetrac.
A naturally-occurring stilbene, resveratrol [14], induces apoptosis in human follicular thyroid cancer cells [4], [17]. Thyroid hormone analogue T4 inhibits the apoptotic action of resveratrol [3], [4] and tetrac has been shown to restore the pro-apoptotic effect of the stilbene in presence of T4 [3]. This effect of tetrac reflects displacement by tetrac of T4 from the iodothyronine receptor site on integrin αvβ3. Resveratrol is capable of binding to the integrin αvβ3 [3], [18], at a site distinct from that for tetrac and other thyroid hormone analogues [3], [4]. In the present studies, the anti-proliferative effect of combined resveratrol and tetrac exposure was tested. Cancer cells were treated with resveratrol (0.1 µM) in presence or absence of 10−7 M tetrac. Both tetrac and resveratrol individually caused anti-proliferative effects in MDA-MB cells (Fig. 8A), while their combination was additive, based on comparison of cell counts on day 8 and Loewe additivity. Human follicular thyroid cancer (FTC) cells were treated daily with resveratrol (0.1 µM) in presence or absence of 10−7 M tetrac. Compared with breast cancer cells, FTC236 cells were less sensitive to tetrac (Fig. 8B). The inhibitory effects of resveratrol and tetrac in combination were additive also in FTC cells, based on cell counts on day 10.
Cetuximab is a monoclonal antibody targeted to the extracellular domain of the EGFR intended for use in patients with metastatic colorectal carcinoma and certain other tumors [19], [20]. Effectiveness is variable [21], [22]. The drug has been combined clinically with various other chemotherapeutic agents in colorectal cancer patients [21], [22] and recently has been tested adjunctively in vitro against breast cancer cells [23]. Combining cetuximab with various chemotherapeutic agents has revealed additive or potentiated growth inhibition in various cancer cell lines [21], [22]. To determine whether tetrac potentiates cetuximab-induced anti-proliferation, human breast cancer MDA-MB cells were treated with cetuximab (0.1 µg/mL) in presence or absence of 10−7 M tetrac. Individually, both agents suppressed proliferation of MDA-MB cells (Fig. 9A). After 8 d treatment with cetuximab and tetrac the average total cell counts were decreased by 34% and 38%, compared to control. Combined tetrac and cetuximab was more effective, reducing total cell numbers on average by 63%. Application of an empirical mathematical model to all treatments and time points simultaneously also suggested an approximately additive effect of both compounds. The empirical model was a disease progression type model where the cell counts in the control treatment were described by a simple exponential function. The effect of tetrac was described as an offset, i.e. a change from baseline cell counts while tetrac is present. The effect of cetuximab was described in the same way, only including an additional lag-time of effect. When both drug effects were added the resulting profile adequately described the cell counts during combination treatment for the studied concentrations and observation period.
An approximately additive effect was also found for the combination of nano-tetrac and cetuximab in human colon cancer Colo-205 cells (Fig 9 B). Colo-205 cells grown in T-75 flask were treated with either nano-tetrac (10−8 and 10−7 M), cetuximab (4 and 40 µg/ml), or combination. Medium was refreshed with agents daily. Cells were harvested and counted as indicated up to 16 days. A model including effects of both drugs on the probability of successful replication and a noncompetitive interaction adequately described the observed cell counts (Figs. 9B and 9C). The effect of the combination treatments was slightly larger than predicted by a competitive interaction model, where both drugs work on the same pathway, and slightly smaller than predicted by a purely noncompetitive interaction model, where the drug works on completely different pathways. Therefore a factor ψ was included (see equation in the Materials and Methods section) which was estimated at 5.6. The ImaxR and IC50R for inhibition of the probability of successful replication were 0.12 and 7.0 nM for nano-tetrac and 0.13 and 3.3 µg/mL for cetuximab.
Using a novel perfusion bellows cell culture system developed in our laboratory (Fig. 1), we have compared the pharmacodynamics in vitro of unmodified and nanoparticulate formulations of tetrac as anti-proliferative agents. The system revealed that nano-tetrac had a higher potency than tetrac as an anti-proliferative agent (Fig. 4). Neither nano-tetrac nor tetrac affected proliferation of two non-cancer cell lines even at high concentrations (10−6 M).
The anti-proliferative effect of tetrac and nano-tetrac on cancer cells in the perfusion bellows cell culture system was seen starting 3 d after start of treatment (Fig. 3, 4). The anti-cancer effects of tetrac and nano-tetrac in human tumor cell xenografts were well-established within 3 d after onset of drug administration [9]. These results in the perfusion system thus reproduce findings obtained earlier in cells grown in culture dishes and xenografts. While the tetrac effects in xenografts have been shown to involve both primary effects on tumor cell proliferation and an anti-angiogenesis effect [6], the effect of tetrac and nano-tetrac in the perfusion bellows cell culture system of course is limited to suppression of cell proliferation.
In vitro models such as described here can save animals by decreasing the number of animal studies which need to be conducted, by employing well-defined conditions which allow for investigation of individual factors impacting the PD and permitting the simulation of human pharmacokinetics (PK) based on data from clinical trials. Limitations of the method described here which need to be considered are that the impact of tissue penetration and the effect of the immune system are usually not directly taken into account; PK/PD models based on animal or clinical studies that include measurement of drug concentrations in tumor need to be developed.
In the perfusion system cells are exposed alternately to fresh medium and air. This paradigm optimizes growth conditions for cancer cells by maximizing nutrient uptake and oxygen transfer and supported experiments of up to 3 weeks' duration (Fig. 3B). Information obtained in longer studies about both the slope of the growth/proliferation phase and the plateau of the cell count with regard to time permitted mathematical modeling to identify two different effects of tetrac on cancer cells: inhibition of growth rate and inhibition of success of replication (Fig. 2).
In addition to treating the cells with constant drug concentrations, reflecting in vivo continuous infusion treatment, the in vitro system described here allows to study other dosing regimens. Multiple short-term or intermittent infusions or brief injections can be studied in the perfusion system by adjusting the flow rate of the medium and the dosing schedule. Drug concentration/time profiles that are expected or have been obtained in human or animal studies can be simulated and effects on cancer cells of changing drug concentrations as anticipated in vivo may be observed in the system. Together with mathematical modeling, these in vitro paradigms can support optimization of design of subsequent animal and human studies thereby saving time and expense. Because a wider range of drug concentrations can be studied in vitro than in animal models, dose selection for in vivo studies may become more efficient.
Mathematical modeling was utilized to increase the amount of information gained from the reported experiments. By considering the entire time course of cell counts in response to multiple concentrations of tetrac and control treatment simultaneously, more insight can be gained into the dose-response relationship and mechanism of action of a drug. Purely empirical growth models, e.g., the Weibull model, often do not include meaningful parameters, but offer arbitrary coefficients. For simulating other scenarios, e.g., cells with faster growth rates, mechanism-based models may be more adequate. While only total cell counts were available in the perfusion bellows cell culture system experiments reported here the applied model is based on mechanisms of action. Inclusion of flow cytometry results in the model will be performed for future experiments in order to enhance the mechanism-based modeling approach.
For U87MG cells studied here, mathematical modeling suggested a higher maximum effect but lower sensitivity of the effect on probability of successful replication, compared to the effect on growth rate for both unmodified and nano-tetrac. For both effects the sensitivity favored nano-tetrac over unmodified tetrac. This may be explained by the ability of unmodified tetrac to penetrate into cells and thereby exert low-grade thyromimetic (proliferative) effects in addition to the anti-proliferative effects initiated at the cell surface integrin receptor. Therefore the net anti-proliferative effect of unmodified tetrac is decreased. The model describes the net effects of unmodified tetrac. Nano-tetrac does not gain access to the cell interior and shows a more robust anti-proliferative effect.
MDA-MB cells had growth rate sensitivity to nano-tetrac that was similar to unmodified tetrac, but a higher sensitivity to nano-tetrac for the effect on success of replication. For both unmodified tetrac and nano-tetrac MDA-MB cells were more sensitive to the effect on success of replication than the effect on growth rate. The uncertainty of the parameter estimates was explored by bootstrap runs. A very rich sampling design was used to ensure the general estimability of the model by two different algorithms. In addition, the estimability was tested under the sampling designs of the actual experiments. For the models of unmodified tetrac, the 10% percentile to 90% percentile intervals (P10–P90) were relatively narrow. A larger uncertainty was seen for the IC50 parameters in the nano-tetrac models, especially for the IC50k in MDA-MB cells. The latter suggests that the effect on rate of growth was not apparent in all of the randomly created bootstrap datasets. Optimal design was not applied to structuring those experiments but will be utilized in future studies. It is important to note that the studied concentrations were 10-fold different between the treatment arms and, based on that factor, the uncertainties in IC50 and the differences in the estimates between NONMEM and S-ADAPT are acceptable. Overall our mechanism-based models have adequately described the cell counts over time in our studies and the effects of a wide range of tetrac concentrations and will support the design of future experiments. In addition to the pharmacodynamic studies in vitro and in animals, also the pharmacokinetics of tetrac will be studied in vivo to more fully characterize the pharmacokinetic / pharmacodynamic relationship for tetrac in vivo.
We have previously shown that resveratrol induces apoptosis in human cancer cells, an effect which requires the nuclear translocation of COX-2 and activated ERK1/2 for support of p53-dependent apoptosis [3], [17]. Resveratrol and tetrac both bind to plasma membrane integrin αvβ3 [1], [18], but at discrete sites that apparently do not interfere with one another [3], [7]. In the present studies, the combination of resveratrol and tetrac was additive in the in vitro perfusion bellows cell culture system in terms of suppression of cell proliferation in two human cancer cell lines. The ability to detect such additivity—or potentiation, if present—is obviously a requirement of the perfusion system.
Therapeutic epidermal growth factor receptor (EGFR) targeting with cetuximab, either as single agent or in combination with chemotherapy, has demonstrated variable clinical activity [19] and may benefit only select patients [20]. In the perfusion bellows cell culture system, concurrent treatment with tetrac and cetuximab resulted in highly effective inhibition of proliferation of MDA-MB cells by day 8 (Fig. 9A). The model system thus offers the prospect of efficiently exploring a variety of drug combinations. An empirical disease progression model was employed for the combination treatment of MDA-MB cells with tetrac and cetuximab, and revealed an approximately additive effect for the combination. While such an empirical model has limitations it is not feasible to develop a receptor occupancy model for a drug combination without data at multiple drug concentrations. Two concentrations each of nano-tetrac and cetuximab and all four combinations were studied in Colo-205 cells in cell culture flasks. The effects of nano-tetrac and cetuximab were adequately described as inhibition of the probability of successful replication. Modeling of all treatment arms simultaneously revealed an approximately additive effect of the combination. The effect of the combination treatment was slightly smaller than predicted by a purely noncompetitive interaction and slightly larger than predicted by a purely competitive interaction model. This suggests that there is a partial overlap between the mechanisms and pathways of action of nano-tetrac and cetuximab. That interpretation of the modeling results is supported by previous studies in our laboratory where we showed that tetrac interferes with crosstalk between the cell surface receptor for thyroid hormone and EGFR [24] and it can be assumed with confidence that nano-tetrac also interferes with this crosstalk. In addition, nano-tetrac, but not unmodified tetrac, decreases the expression of the EGFR gene [8]. For this study in cell culture flasks it was observed that cell counts in all treatment arms decreased noticeably and approximately in parallel after Day 6 (Fig 9B) which cannot be attributed to drug effect. Such observations further support the use of the perfusion bellows cell culture system which provides optimal nutrient uptake and oxygen transfer for the cells and will be utilized for future combination studies in colon cancer cells.
The perfusion bellows cell culture studies we described provide useful pharmacodynamic information on the application of new drugs or combinations of various agents in vitro to human cancer cell lines. In combination with pharmacodynamic modeling and by including information about the expected pharmacokinetics of a drug, the perfusion bellows cell culture system permits study of the dose-response relationships of anti-neoplastic agents over a very wide concentration range in vitro, and can support translation from in vitro models to animal models and human clinical trials.
Human glioblastoma cells (U87MG), human breast cancer MDA-MB-231 cells (MDAMB), human colon cancer Colo-205 cells, African green monkey kidney epithelial CV-1 cells and human embryonic kidney 293T cells were purchased from ATCC. Human follicular thyroid cancer FTC236 cells were generously provided by Dr. Orlo Clark (University of California at San Francisco-Mt. Zion Medical Center, San Francisco, CA). U87MG cells were maintained for study in MEM (Gibco, Carlsbad, CA) supplemented with 10% fetal bovine serum (FBS, Sigma Aldrich, St. Louis, MO). Colo-205 cells were maintained in RPMI (Gibco) supplemented with 10% FBS. MDA-MB, CV-1 and 293T cells were maintained in DMEM (Gibco) supplemented with 10% FBS. Follicular thyroid cancer cells were supported in 50% DMEM/50% Ham's F-12 (Gibco) plus 10 mU/ml of TSH (Sigma Aldrich). Cells were cultured in a 5% CO2/95% air incubator at 37°C.
Shown in Fig. 1 is a newly developed perfusion bellows cell culture system that is a disposable bioreactor capable of high density cell culture for studies of anti-cancer drugs. Each cell culture system is a compressible (bellows) 500 mL bottle that contains cell culture medium and specially-treated polymer flakes to which cells spontaneously attach and then proliferate. Through moving bellows and porous membranes the level of the medium in the bottle changes periodically. Consequently, the cells are alternately submerged in the culture medium and exposed to 5% CO2/95% air which creates a dynamic interface between air and medium on the plated cell surface that maximizes nutrient uptake and oxygen transfer. The system provides a low shear, high aeration and foam-free culture environment. Proprietary treatment of the surfaces of the flakes enables seating and the harvesting of cells and secreted proteins are readily isolated from the perfusate.
In a non-perfusion bellows cell culture system that was also used, the medium in each bottle was replaced by fresh medium every 24 h. In the perfusion bellows cell culture system, medium was progressively refreshed over 24 h, so that one complete change of medium occurred over 24 h.
To establish the cultures, 5×107 cells were seeded in perfusion and non-perfusion bellows bottles and incubated overnight at 37°C. Flakes were then harvested, trypsinized, and the cells were collected and counted. The number of cells that attached to flakes was 10–15×106 per bottle. For experiments, the perfusion bellows cell culture system was run for 2 d prior to starting the experiments. The cell numbers at this point were about 30–50×106 cells per bottle. Cultured cells were then exposed to 1% FBS-containing medium. Tetrac or nano-tetrac was added to the medium in the reservoir bottle to achieve the final concentrations reported for each experiment.
Nano-tetrac utilized in the studies of proliferation of MDA-MB, U87MG, and Colo-205 cells was manufactured on contract by Azopharma (Miramar, FL). Nano-tetrac for all other experiments was prepared at the Pharmaceutical Research Institute, Rensselaer, NY [9]. Unmodified tetrac was synthesized on contract by Peptido GmbH (Bexbach, Germany).
In LC/MS/MS experiments, medium samples (20 µL) were injected onto an HP 1100 series HPLC system (Agilent Technologies, Palo Alto, CA, USA), equipped with a narrow-bore Zorbax Eclipse XDB-C18 column (5 µm, 150×2.1 mm; Agilent). Separation was performed using a mobile phase of 0.1% (v/v) acetic acid (A) and 100% acetonitrile (B), with a linear gradient of 20–60% B over 25 min. Flow rate was maintained at 0.2 mL min−1 and elution was monitored by a diode array detector (200–600 nm). The LC effluent was then introduced into a turbo ion-spray source on a Q/STAR-XL quadruple/time-of-flight (TOF) hybrid mass spectrometer (Applied Biosystems, Foster City, CA, USA). Negative ESI mass spectra were acquired over the range m/z 100 to 400. The electrospray voltage was set at −4.5 kV and the source temperature was maintained at 475°C. CID spectra were acquired using nitrogen as the collision gas under collision energies of 25–55 V. High purity nitrogen gas (99.995%) was used as the nebulizer, curtain, heater and collision gas source.
Total RNA was isolated as described previously [25]–[27]. First strand complementary DNAs were synthesized from 1 µg of total RNA, using oligo dT and AMV Reverse Transcriptase (Promega, Madison, WI). First-strand cDNA templates were amplified for GAPDH, c-fos, PIG3, c-Jun, and BAD mRNAs by polymerase chain reaction (PCR), using a hot start (Ampliwax, Perkin Elmer, Foster City, CA). Primer sequences were GAPDH (5′-AAGAAGATGCGGCTGACTGTCGAGCCACA-3′ [forward] and 5′- TCTCATGGTTCACACCCATGACGAACATG-3′ [reverse), c-fos (5′-GAATAAGATGGCTGCAGCCAAATGCCGCAA-3′[forward] and 5′-CAGTCA-GATCAAGGGAAGCACAGACATCT-3′ [reverse]), PIG3 (5′-TGGTCACAG-CTGGCTCCCAGAA-3′ [forward] and 5′-CCGTGGAGAAGTGAGGCAGAATTT-3′ [reverse]), c-jun (5′-GGAAACGACCTTCTATGACGATGCCCTCAA-3′ [forward] and 5′-GAACCCCTCCTGCTCATCTGTCACGTTCTT-3′ [reverse) and BAD (5′-GTT-TGAGCCGAGTGAGCAGG-3′ [forward] and 5′-ATAGCGCTGTGCTGCCCAGA-3′ [reverse]). The PCR cycle was an initial step of 95°C for 3 min, followed by 94°C for 1 min, 55°C for 1 min, 72°C for 1 min, then 25 cycles and a final cycle of 72°C for 8 min. PCR products were separated by electrophoresis through 2% agarose gels containing 0.2 µg of ethidium bromide/mL. Gels were visualized under UV light and photographed with Polaroid film (Polaroid Co., Cambridge, MA). Photographs were scanned under direct light for quantitation and illustration. Results from PCR products were normalized to the GAPDH signal.
Cells were harvested from flakes by trypsinization, washed with PBS, fixed in ice-cold 70% ethanol and stored in a freezer overnight. Cells were labeled to detect apoptosis with the In situ Cell Death Detection Kit, Fluorescein (Roche Diagnostics Corporation, Roche Applied Science, Indianapolis, IN). The recommended procedures were used with modifications in permeabilization time and temperature to improve results. Fixed cells were centrifuged and washed once in PBS containing 1% bovine serum albumin (BSA), then resuspended in 2 mL permeabilization buffer (0.1% Triton X-100 and 0.1% sodium citrate in PBS) for 25 min at room temperature, followed by a wash in 0.5 mL PBS/1% BSA. Cells were resuspended in 50 µL TUNEL reaction mixture (TdT enzyme and labeling solution) and placed in an incubator for 60 min at 37°C in a humidified dark atmosphere. Labeled cells were washed in PBS/1% BSA, then resuspended in 0.5 mL ice-cold PBS/0/1% BSA Triton X-100 that contained 1 µg/mL 4′, 6-diamidino-2-phenylindole (DAPI) for at 20 min. Cell samples were analyzed with a BD™ LSR II (BD Biosciences, San Jose, CA), using BD FACSDiva™ software. Fluorescence histograms were gated on forward scatter (FSC) and side scatter (SSC) to exclude debris and clumped cells. Gating on height vs. area fluorescence of DAPI signal was set to eliminate clumped cells and to obtain the singlet population for analyzing the cell cycle phase ratios in G1, S or G2/M.
Immunoblot and nucleotide densities were measured with a Storm 860 phosphorimager, followed by analysis with ImageQuant software (Molecular Dynamics, Sunnyvale, CA). Student's t test, with P<0.05 as the threshold for significance, was used to evaluate the significance of the hormone and inhibitor effects. Where cell counts were tested for statistical significance, the data were log-transformed prior to testing. For the cell count data, an α-adjustment to account for multiple comparisons was utilized according to the Holm t test. The concept of Loewe additivity [28] was applied to cell count data from combination treatments. For experiments involving cells counts at many time points, for multiple treatments, or both, multiple t tests were not an adequate method of analysis due to the large number of comparisons. In addition, multiple comparison tests treat the observations at each time point independently, whereas mathematical modeling, as described below, takes into account the full time course. Observed data are presented in the figures as individual data points or average ± standard deviation (SD).
The time course of cell counts of the several cancer cell lines treated with different concentrations of tetrac or nano-tetrac (or a combination of tetrac with cetuximab or resveratrol, or nano-tetrac with cetuximab) was modeled utilizing a naïve pooled approach in NONMEM VI (version 6.2). The pooled approach does not distinguish any potential unexplained variability between the bottles (treatment arms) from general assay error, e.g., uncertainty in cell counts, but expresses both in the residual error. The perfusion bellows cell culture system experiments included one bottle per treatment arm with the multiple observations per time point being different cell counts of one sample for the tetrac experiments and the nano-tetrac with cetuximab combination study, and average cell counts from three studies for the nano-tetrac experiments. The population approach in NONMEM (FOCE) did not succeed in distinguishing inter-subject variability (variability between bottles) and unexplained random variability (e.g. general assay error). The naïve pooled analysis in NONMEM was equivalent to a pooled analysis using the Maximum Likelihood approach in ADAPT, for example. S-ADAPT was also utilized as described below in order to make use of the MC-PEM algorithm and for additional model evaluation. All time points and treatment arms within each experiment were modeled simultaneously. A mechanism-based model [29] was adapted to describe the proliferation of cancer cells and the inhibition of proliferation by tetrac. This model assumes two populations of cells in different phases of the cell cycle: cells that are preparing for replication (phase 1) and cells that are immediately ‘pre-replication’ (phase 2). Cells transition from phase 1 to phase 2 by a first-order growth rate constant, while replication from phase 2 to phase 1 is assumed to be fast (Fig. 2).
The number of cells in phase 1 and 2 are described by:where C1 is the number of cells in phase 1, C2 the number of cells in phase 2, k21 the first order rate constant for replication (transition from phase 2 to phase 1), and k12 the first-order growth rate constant for transition from phase 1 to phase 2. The k21 was assumed to be fast and therefore was fixed to 100 day−1, which resulted in a ratio of k21/k12 of approximately 50 to 100, depending upon the cell line. The total number of cells Ct is the sum of C1 and C2. Rep is the replication efficiency factor which is described by:where Cmax is the maximum number of cells. Without tetrac (or nano-tetrac), the replication efficiency factor approaches 2, which reflects a 100% probability of successful replication. When Ct approaches Cmax, Rep approaches 1, representing a 0% probability of net replication, that is, cells in reality still transition between the phases, but the number of cells does not increase further. The InhR describes the inhibitory effect of tetrac on the probability of successful replication:where ImaxR is the maximum effect of tetrac (or nano-tetrac) on the probability of successful replication and IC50R is the tetrac concentration needed to achieve a half-maximal effect. In the case of InhR< 0.50, this effect results in cell killing, as it then follows that Rep • InhR< 1.0. The latter case also illustrates that cells which do not replicate successfully are eliminated in this process. For some studies, inclusion of a decrease in ImaxR over time was necessary in order to adequately describe the data:where ImaxR0 is the ImaxR at time = 0 and kiR is a constant describing the decrease of ImaxR over time.
Inhk describes the inhibitory effect of tetrac on the rate of growth:where Imaxk is the maximum effect of tetrac on rate of growth and IC50k is the tetrac concentration needed to achieve a half-maximal effect. Both IC50R and IC50k are measures for the sensitivity of the cancer cells to the effects of tetrac. A low IC50 corresponds to a high sensitivity of the cells to a particular drug effect, and vice versa. While the InhR describes an irreversible removal of cells from the cell cycle, Inhk only slows down the transitioning of cells through the cell cycle. The cells remain in state 1 for a longer period of time which represents growth and preparation for replication. This is reflected in a decreased slope of the growth curve.
Although cells in state 1 and state 2 were not measured separately in the perfusion bellows cell culture system experiments reported here, the two effects were distinguishable and the parameters estimable. The effect on rate of growth decreases the slope of the growth curves whereas the effect on successful replication results in lower plateaus at the end of the growth curves for the treatment arms compared to control. As described below simulation estimation runs were performed to confirm the estimability of the parameters.
The effects of nano-tetrac were modeled by the same equations as described above for unmodified tetrac. However the IC50 estimates for nano-tetrac are hypothetical concentrations that assume all of the tetrac bound to the nanoparticle is available for binding to the integrin receptor.
A lag time for growth was included in order to describe the data successfully. The parameter k12 was low at the start of the experiment and increased over time:Here, k12max is the maximum growth rate constant and b and c are empirical constants. The residual variability was described by an additive error on log-scale.
A model for non-competitive interaction was applied to the experiment on the effects of nano-tetrac, cetuximab, and their combination on Colo-205 cells. The effects of nano-tetrac (InhRNPT) and cetuximab (InhRCET) were described as:The effect of the combination was:which describes a non-competitive interaction [30], [31] when ψ = 1, that is both drugs act by completely separate pathways [32], [33]. When ψ>1, then the effect of the combination is less than would be expected from two drugs acting completely independent of each other. The decrease of cell counts in all treatment arms towards the end of the observation period in this study in cell culture flasks was modeled by a series of transit compartments.
Model discrimination was based on comparison of the objective function in NONMEM, visual comparisons of observed and fitted cell counts over time, and observed vs. fitted plots. Simulation estimation experiments (bootstraps) were performed for the models of tetrac and nano-tetrac effects on U87MG and MDA-MB cells in order to explore the estimability of the model and the bias and uncertainty in the parameter estimates. The simulations were done in Berkeley Madonna (v.8.3.14). The estimations were performed in both NONMEM (pooled approach) and the MC-PEM (Monte Carlo parametric expectation maximization) algorithm in parallelized S-ADAPT (v.1.56). One hundred bootstrap datasets in NONMEM and fifty bootstrap datasets in S-ADAPT, each with 10 profiles per treatment arm, were run for each of the four experiments (two cell lines and two formulations), assuming a very rich sampling schedule and an additive error on log-scale of 0.1 (Bootstraps based on additive errors on log-scale of 0.02, 0.05, and 0.1 had been previously conducted for the model of tetrac effects in U87MG cells). As the bootstraps were performed in order to obtain a point estimate for the parameters and not to characterize their distribution, and also due to long run times, 50 to 100 bootstrap runs each were adequate. Those bootstraps based on the rich sampling schedule were conducted to evaluate the mathematical estimability of the model parameters under ideal experimental conditions, i.e. many sampling time points. One hundred bootstrap datasets each with 10 profiles per treatment arm were run in NONMEM for each of the four models with the sampling schedules that were actually used in the experiments and assuming an additive error on log-scale of 0.1. The bootstraps based on the actual sampling schedules were performed to test whether the model parameters were well-estimable based on both the model and the experimental conditions. The median and 10% and 90% percentiles were calculated from each of those simulation estimation experiments.
|
10.1371/journal.pbio.1000577 | Analysis of LhcSR3, a Protein Essential for Feedback De-Excitation in the Green Alga Chlamydomonas reinhardtii | In photosynthetic organisms, feedback dissipation of excess absorbed light energy balances harvesting of light with metabolic energy consumption. This mechanism prevents photodamage caused by reactive oxygen species produced by the reaction of chlorophyll (Chl) triplet states with O2. Plants have been found to perform the heat dissipation in specific proteins, binding Chls and carotenoids (Cars), that belong to the Lhc family, while triggering of the process is performed by the PsbS subunit, needed for lumenal pH detection. PsbS is not found in algae, suggesting important differences in energy-dependent quenching (qE) machinery. Consistent with this suggestion, a different Lhc-like gene product, called LhcSR3 (formerly known as LI818) has been found to be essential for qE in Chlamydomonas reinhardtii. In this work, we report the production of two recombinant LhcSR isoforms from C. reinhardtii and their biochemical and spectroscopic characterization. We found the following: (i) LhcSR isoforms are Chl a/b– and xanthophyll-binding proteins, contrary to higher plant PsbS; (ii) the LhcSR3 isoform, accumulating in high light, is a strong quencher of Chl excited states, exhibiting a very fast fluorescence decay, with lifetimes below 100 ps, capable of dissipating excitation energy from neighbor antenna proteins; (iii) the LhcSR3 isoform is highly active in the transient formation of Car radical cation, a species proposed to act as a quencher in the heat dissipation process. Remarkably, the radical cation signal is detected at wavelengths corresponding to the Car lutein, rather than to zeaxanthin, implying that the latter, predominant in plants, is not essential; (iv) LhcSR3 is responsive to low pH, the trigger of non-photochemical quenching, since it binds the non-photochemical quenching inhibitor dicyclohexylcarbodiimide, and increases its energy dissipation properties upon acidification. This is the first report of an isolated Lhc protein constitutively active in energy dissipation in its purified form, opening the way to detailed molecular analysis. Owing to its protonatable residues and constitutive excitation energy dissipation, this protein appears to merge both pH-sensing and energy-quenching functions, accomplished respectively by PsbS and monomeric Lhcb proteins in plants.
| Reactive oxygen species are formed during photosynthesis, particularly when electron transport is saturated in high light. The process of non-photochemical quenching (NPQ) helps protect plants against excess light by dissipating the excited states of chlorophyll into heat. By doing so, it prevents the formation of triplet excites that otherwise would react with molecular oxygen to form singlet oxygen, a damaging reactive oxygen species. In plants, NPQ is triggered by the PsbS protein, which senses pH changes caused by excess light and consequently triggers energy-quenching functions in other proteins. The green microalga C. reinhardtii lacks the PsbS proteins, and NPQ depends on the LhcSR3 protein. In this study, we show that, unlike PsbS, LhcSR3 not only binds pigments but is also a strong quencher for chlorophyll excited states. LhcSR3 carries protonatable residues that enable it to sense pH change. Its quenching activity is further enhanced by low pH, suggesting that this algal protein merges the functions of pH sensor and of excited state quencher into a single gene product.
| In photosynthetic organisms, feedback dissipation of chlorophyll (Chl) singlet excited states balances light harvesting with metabolic energy consumption, in order to prevent photodamage due to reactive oxygen species (ROS) formation when excess energy is transferred to O2. Both plants and algae can dissipate Chl excited states into heat through mechanisms involving xanthophyll-binding Lhc proteins.
The light-harvesting complex (Lhc) gene family is present in all photosynthetic eukaryotes [1]. Lhc proteins act in light harvesting, owing to their capacity to bind Chl and carotenoid (Car) chromophores, a characteristic shared by most members of the family, with few exceptions [2]. Lhc proteins are also involved in photoprotection through their xanthophyll ligands, which are active in quenching Chl singlets and triplets as well as in scavenging ROS [3]–[9], with lutein and zeaxanthin (Zea) playing a predominant role [10],[11].
Among Lhc proteins, LhcSR, PsbS, and ELIPs are more specifically involved in photoprotective mechanisms and are over-expressed under stress [12]–[14]. ELIPs, transiently expressed in plants [15], repress Chl biosynthesis by sequestering precursors in order to prevent free Chl accumulation in high light (HL) [16]. PsbS acts in chloroplast lumenal pH sensing [17],[18] and in activation of the fast component energy-dependent quenching (qE) of non-photochemical quenching (NPQ) [17],[19]. Photosynthetic organisms thermally dissipate light energy absorbed in excess with respect to their needs for photosynthesis through NPQ. LhcSR orthologs are widely distributed among green and brown algae, and are also found in some mosses [20],[21]. Knock-out mutants disrupted in psbS and lhcSR3.1/lhcSR3.2 genes have similar qE-null phenotypes, respectively, in plants and algae [17],[22], suggesting similar functions and mechanisms of action for their gene products. Biochemical analysis of PsbS, both in vivo and in vitro, identified two lumenal-exposed, dicyclohexylcarbodiimide (DCCD)–binding glutamate residues essential for qE triggering [18],[23],[24]. PsbS has also been shown to be unable to bind pigments, owing to the non-conservation of Chl-binding residues [19],[25]. Thus, protonation of PsbS leads to activation of a lutein- and Zea-dependent quenching process in Lhcb proteins [19],[26]. Consistent with this model, deletion of PSII antenna subunits affects qE kinetics and amplitude [27]–[30], and these subunits have been shown in vitro to be active in energy-quenching processes involving the formation of Zea and/or lutein radical cations, by means of a Chl-Car charge-transfer quenching (CT quenching) mechanism [6],[31],[32].
Although the biogenesis of LhcSR proteins has been studied [14], their biochemical properties are still unknown. For information on the mechanism by which LhcSR activates energy dissipation in C. reinhardtii, we characterized LhcSR3 isoforms after in vitro refolding of the purified apoprotein, a procedure that has been shown to be effective in yielding pigment-protein complexes with the same biochemical and spectral properties as many Lhc proteins [6],[32]–[35]. We show that LhcSR3, unlike PsbS, forms complexes with pigments containing Chl a, Chl b, lutein, and violaxanthin/Zea. Spectroscopic analysis of LhcSR3 shows the presence of very short fluorescence lifetimes compared with other members of the Lhc family, implying that energy-dissipating mechanisms are very active in this protein. These findings, together with the capacity of LhcSR3 to bind DCCD, a marker for proton-sensitive residues in proteins, and its increased quenching activity upon acidification, suggest that it combines the functions of pH sensor and of the Chl excited state quenching needed for NPQ, which in plants is performed by two distinct protein components: PsbS and Lhcb subunits.
In C. reinhardtii, LhcSR3 has been reported to be essential for qE, consistent with its increased abundance in thylakoid membranes upon acclimation of cells at HL, a condition which up-regulates qE capacity [22]. In C. reinhardtii, three genes (lhcSR1, lhcSR3.1, and lhcSR3.2) encode LhcSR isoforms [36], but two of them, lhcSR3.1 and lhcSR3.2, encode the same 259-amino-acid polypeptide. The LhcSR1 isoform has 253 residues with 87% identity with respect to LhcSR3.1/LhcSR3.2. For information on the properties of the LhcSR proteins, we cloned and expressed LhcSR1 and LhcSR3 (corresponding to genes lhcSR1 and lhcSR3.1) in Escherichia coli. The LhcSR3 isoform, purified from inclusion bodies, was injected into rabbits to obtain an antiserum that was found to recognize both the LhcSR1 and LhcSR3 recombinant proteins in SDS-PAGE (data not shown). Figure 1 shows the immunodetection of LhcSR following SDS-PAGE separation of thylakoid membranes from HL (500 µE) and low light (LL) (50 µE) acclimated cells: three bands are detected at approximately 25 kDa. The fastest migrating band was always faint, whereas the bands with higher apparent molecular weight (MW) were strongly over-accumulated in HL matching results from a recent report [22]. To check whether the retardation of Lhc protein bands in SDS-PAGE was the result of protein phosphorylation, as previously shown for CP29 [37], we treated samples with alkaline phosphatase (Figure 1B). The intensity of the upper band in the HL lane was considerably decreased by the treatment, with concomitant intensification of the intermediate band. Similar effects of phosphatase treatment were observed in the LL sample, suggesting that phosphorylation of LhcSR3 did not depend on light intensity. In order to verify the possibility that LhcSR3 phosphorylation is involved in NPQ, we proceeded in two steps. We first verified that phosphorylation was almost absent in the stt7 mutant, which lacks the thylakoid kinase responsible for LHCII phosphorylation and State 1–State 2 transitions [38],[39]. We then compared the NPQ kinetics of wild type (WT) and stt7 upon acclimation to LL and HL conditions. Figure 2 shows that NPQ amplitude was below 0.5 in WT LL cells, but increased to 2.5 in HL acclimated cells. NPQ in stt7 HL cells was even higher than in WT. These results imply that Stt7 kinase is the major agent responsible for phosphorylation of LhcSR3, and that phosphorylation is not needed for NPQ activity.
Most Lhc proteins bind both Chl a and Chl b, but LhcSR proteins are also found in organisms, like diatoms, that lack Chl b [20]. We therefore analyzed the Chl b–less cbs3 mutant [40] in order to test whether this pigment species was required for LhcSR stability in vivo. Figure 1 demonstrates that this is not the case, since the same three bands were present in both mutant and WT. However, upon HL acclimation, the two upper LhcSR3 bands did not increase as in WT. Consistent with the similar level of LhcSR3 accumulation, NPQ was similar to WT in LL cells (Figure 2).
Titration of LhcSR3 protein abundance in thylakoid membranes can be performed by exploiting the availability of a specific antibody. For this purpose, various dilutions of thylakoid membranes from HL-grown cells were loaded on an SDS-PAGE gel, together with a dilution series of the recombinant pigment-protein obtained as described below. Following transfer to nitrocellulose and immunodetection with anti-LhcSR-specific antibody, the intensity of the immunological reaction was estimated by densitometry and related to the amount of Chl loaded. Based on a PSII/PSI ratio of 1.18 and antenna sizes of 240 and 222 Chls/reaction Center for PSI and PSII, respectively [41], and on the proposed number of Chls bound per LhcSR3 polypeptide of 6.7±1.9 (see below), we calculated a LhcSR/PSII ratio of 0.17±0.11 in HL acclimated thylakoids. Although this estimation should be used with caution, it clearly suggests that LhcSR is substoichiometric with respect to PSII reaction centers.
In order to clarify the function of LhcSR3, it is essential to establish its possible interactions with other proteins. We therefore separated pigment-protein complexes after solubilization with α-dodecyl-maltoside by native electrophoresis, as shown in Figure 3. As previously shown [42], the band with the highest mobility, at the electrophoretic front, contains protein-free pigments, and the slower migrating bands represent pigment-protein complexes or their oligomers (Figure 3). When thylakoids from HL- and LL-grown cells were separated and immunoblotted with the anti-LhcSR antibody, two reactive bands were detected, migrating, respectively, to the level of green band 2 and between green bands 2 and 3. Since LhcSR apoproteins have a MW similar to that of the monomeric Lhc proteins forming band 2, this indicates that LhcSR in thylakoid membranes form dimers in both HL and LL conditions. In principle, LhcSR may also form heterodimers with a similar MW protein, e.g., a monomeric Lhcb subunit, such as CP29, CP26, and/or Lhcbm1 [43]. In order to check this possibility, we proceeded to a second-dimension separation of native gel lanes in denaturing conditions, followed by immunoblotting with antibodies specific for Lhcb proteins [44]. No co-migration of Lhcb proteins or LhcSR corresponding to the upper LhcSR-reactive band was detected, thus excluding stable interactions between LhcSR and Lhcb proteins (data not shown). Alkaline phosphatase treatment did not affect this pattern (Figure 3B), indicating that phosphorylation plays no role in dimer formation.
In order to clarify the role of a protein in energy dissipation, it is essential to assess its capacity for binding pigments. For example, PsbS, essential for NPQ in plants [17], was suggested to be the actual quenching site, but its role was later revised after it was found not to bind pigments [19],[25]. The sequence alignment of LhcSR versus two Arabidopsis thaliana Lhcb sequences (Lhcb1 and CP29) in Figure 4 shows that six out of eight amino acid residues responsible for Chl binding in LHCII and CP29 [33],[35],[45],[46] are conserved, i.e., residues coordinating four Chl a–specific sites (A1, A2, A4, and A5) and two Chl a/Chl b promiscuous sites (A3 and B5). The two other residues (B3 and B6 sites) are not conserved. Sequence requirements for xanthophyll-binding residues are not well known, with the exception of tyrosine residue 111 (147) in Lhcb1 (CP29), which is involved in neoxanthin binding [47]: this tyrosine is not conserved in LhcSR1 or LhcSR3, indicating that no neoxanthin binding takes place.
In order to demonstrate pigment-binding capacity, the protein can be purified from C. reinhardtii cells acclimated to HL. Unfortunately, this approach was not successful, because of the presence of many other Lhc subunits with similar physico-chemical properties, hindering isolation of this low-abundance subunit. An alternative approach, successfully used in similar circumstances, consists of over-expressing the encoding gene in bacteria and refolding the apoprotein in vitro with pigments, giving rise to a holoprotein with biochemical and spectroscopic properties identical to those of the purified protein, as demonstrated in a large variety of Lhc members [2],[33],[34],[48]–[51].
The in vitro reconstitution of recombinant LhcSR1 and LhcSR3 yielded a pigment-binding complex having the same mobility as monomeric Lhcb proteins in a sucrose gradient. Refolding efficiency was higher in LhcSR3 than in LhcSR1. Since the yield of isoform LhcSR1 was limiting for full biochemical and spectroscopic characterization, we focused attention on the LhcSR3 isoform only, which is more physiologically important, as the subunit has been shown to be essential for NPQ.
Figure 5A shows the LhcSR3 absorption spectrum with a Qy transition peak at 678.8 nm, i.e., strongly shifted with respect to the 670-nm peak of the free Chl a in detergent solution, and even more red-shifted than any PSII antenna protein so far analyzed. The fluorescence emission spectra (Figure 5B) were characterized by a 681-nm peak, independently of exciting Chl a (440 nm), Chl b (475 nm), or xanthophylls (495 nm), implying efficient energy transfer between pigments, although a low level of direct emission from uncoupled Chl b was detected. These data, all together, show that LhcSR3 forms stable complexes with Chl a, Chl b, and xanthophylls, with chromophores having mutual interactions similar to those previously described for Lhcb proteins.
Quantitative analysis of chromophore binding to LhcSR3 was carried out by determination of Chl binding to the complex versus dye binding to apoprotein [52]. With the well-characterized Lhcb1 recombinant protein, binding 12.6 Chl per polypeptide [35],[53], as a reference, a Chl/apoprotein stoichiometry of 6.7±1.9 was obtained for LhcSR3. On the basis of this stoichiometry and on the conservation of six out of eight Chl-binding residues, we propose a stoichiometry of six Chls per apoprotein (Table 1), which is the lowest figure ever calculated for an Lhc complex, even below the eight Chl per CP29 holoprotein [54]. Nevertheless, a figure of seven Chl per polypeptide cannot be excluded. High-performance liquid chromatography (HPLC) analysis (Table 1) showed that LhcSR3 is characterized by a high Chl a/b ratio (Chl a/b = 6.3±0.3), demonstrating strong affinity for Chl a. We also performed in vitro reconstitution with Chl a and xanthophylls only, in the absence of Chl b, and again obtained a pigment-protein complex with characteristics similar to those of the control (reconstituted in the presence of Chl a and Chl b), including a red-shifted absorption peak at 678.4 nm and efficient excitation energy transfer from xanthophyll to Chl a. In that complex, Chl b chromophores were substituted by Chl a, as shown by the decreased absorption at 635–645 nm accompanied by an increased absorption at 660–675 nm in the difference absorption spectra (Figure S1). Xanthophyll composition was the same as in the sample containing Chl b (Table 1), thus confirming that LhcSR3 does fold well in the absence of Chl b. It is worth noting that there was a 20-fold decrease in refolding efficiency with respect to the Chl b–containing sample, indicating that Chl b, although not indispensable, does contribute to pigment-protein complex stability.
Besides Chl a and Chl b, Lhc proteins bind xanthophylls into specific sites, i.e., sites L1, L2, N1, and V1 [45]–[47],[55],[56]. However, in various members of the Lhc family, the affinity of each site for xanthophyll species is variable, and confers functional specialization on different Lhc proteins.
HPLC analysis of reconstituted LhcSR3 shows lutein and violaxanthin bound to the apoprotein, whereas neither neoxanthin nor loroxanthin, although present in the pigment mix during refolding, were bound. Based on six Chls per polypeptide and a Chl/Car ratio of 2.5, the number of Cars per apoprotein molecule in LhcSR3 is between two and three, indicating the presence of two binding sites with strong selectivity for lutein and violaxanthin, probably sites L1 and L2. A third binding site, partially unoccupied, is also present. The same conclusions hold, even under the assumption of seven Chls per polypeptide.
The absence of neoxanthin in the refolded complex indicates that site N1 is absent, consistent with the non-conservation of tyrosine 121, as in the case of CP24 [47]. The third Car binding site is thus of V1 type. When LhcSR3 was reconstituted in the presence of Zea (LhcSR3 LVZ), the Chl/Car and Chl a/b ratios were the same and slightly decreased, respectively, relative to the control protein. Since the transition energies of xanthophylls are tuned differentially by binding to different protein sites [55],[57], spectral deconvolution of the absorption spectra in the Soret region was performed as in [57]–[60] on the LV and LVZ complexes, in order to identify the Zea-binding sites. Results are shown in Figure S2. In all solutions yielding the best fits, the xanthophyll absorption forms showed three levels of red shifts with respect to absorption in organic solvent, i.e., 18–19, 15–16, and 9 nm, consistent with L2, L1, and V1 binding sites, respectively [55]. In the presence of Zea (0.5 mol per mole of protein), multiple new absorption forms were needed for optimal fitting, implying that Zea entered the three sites, although to different extents. This was unexpected, since in other recombinant Lhc proteins Zea binds selectively to site L2, and a single spectral form is needed for fitting [60].
Information on chromophore organization within pigment-proteins can be obtained by circular dichroism (CD) [61]. The LhcSR3 CD spectrum (Figure 5C) has features previously reported for native and recombinant Lhcb proteins, with signals in the Qy region at 683 nm (−) and 660 nm (+). In the Soret range, a strong broad negative signal is observed at 495 nm, associated with xanthophylls [61]. Interestingly, the amplitude of the CD signal in the Qy region is almost twice as strong as that of the homologous protein CP29 from Zea mays, upon normalization at the same protein molar concentration, which indicates an enhanced level of excitonic interactions between Chl a chromophores in LhcSR3 [61].
The absorption spectrum of LhcSR3 is shown in Figure 5A, compared with that of CP29 from Z. mays [33], used as a reference, since it has the most similar pigment-binding properties among all known Lhc proteins. The Qy transition of LhcSR3 peaks at 679 nm, 2 nm red-shifted compared with CP29, and has a tail that is redder than in CP29. Analysis of this Chl spectral contribution to the LhcSR3 LV absorption spectrum with Chl a and Chl b spectral forms in a protein environment (Figure 5E and 5F) yielded six major spectral forms, including four Chl a forms, peaking at 667, 674, 679, and 684 nm (accounting for 0.5, 1.0, 2.3, and 1.2 Chls, respectively). Chl b spectral forms were detected at 639 and 648 nm, together accounting for nearly one Chl b molecule per polypeptide. Two additional low-amplitude absorption forms (0.15× Chl a), peaking at 660 nm and 688 nm, were indispensable for optimal fitting of the spectra. Similar analysis in CP29 or any other Lhcb polypeptide did not require such strongly red-shifted spectral forms. The transition energy of absorption forms is quite well conserved within the Lhcb family, and the red-most component peaks at 682 nm [5],[33],[35]. In LhcSR3, the red-most forms were further red-shifted to 684 and 688 nm. Such strong shifts have been reported in the PSI-associated LHCI proteins to be caused by excitonic interactions and accompanied by increased bandwidths [49],[62], indicating that the 684- and 688-nm forms represent a single, wider absorption form deriving from excitonic interaction of two Chl a molecules. This possibility was probed by performing deconvolution, including a Gaussian spectral component with larger bandwidth in the red-most part of the spectrum (Figure 5G and 5H), and the goodness-of-fit was in fact significantly improved. The best description of the experimental spectrum was obtained by using a Gaussian peaking at 685.7 nm, with 16 nm full width at half maximum, compared with 12 nm of a monomeric Chl a spectral form [57]; the 684- and 688-nm forms disappeared. The spectrum of Zea-binding LhcSR3 was also best fitted with a 16-nm spectral form peaking at 685.7 nm, implying that no major conformational changes were induced by binding of Zea (data not shown).
The presence of low-energy excited states in LhcSR3 was confirmed by the 77K fluorescence emission peak at 685 nm, substantially red-shifted compared with CP29 (680 nm) (Figure 5D).
Red-shifted spectral forms have been involved in both energy dissipation [6],[29],[63] and light harvesting [49]. We therefore proceeded to single photon counting fluorescence lifetime analysis (Figure 6) in order to verify whether LhcSR3 acts as a quencher of the antenna proteins, which are well known to have a long lifetime, on the order of 3–4 ns [64],[65]. Decay curves upon excitation at 435 nm and detection at 685 nm are shown in Figure 6, and the results of their deconvolution are listed in Table 2. As a reference, we also analyzed a sample of CP29. Matching previous studies, CP29 showed two lifetimes of 4.6 ns (62%) and 1.8 ns (38%), thus yielding an average lifetime of 3.5 ns [66]. The case of LhcSR3 was clearly different, since three exponential components were needed for best fitting of the decay curves. Besides two components of 4 and 1.95 ns, accounting, respectively, for 10% and 25% of total decay, a dominant short component was obtained, with a lifetime less than 100 ps, accounting for nearly 65% of fluorescence. Similar results were obtained when fluorescence emission was collected at different wavelengths. LhcSR3 binding Zea (LhcSR3 LVZ) also showed decay with three components, like LhcSR3 LV. However, the intermediate lifetime component was slightly faster, 1.5 ns versus 1.95 ns, and the longest component had increased amplitude, probably because of partial uncoupling of a small fraction of pigment (Figure 6; Table 2). When measurements were repeated at pH 5.5, fluorescence decays were even faster because of increased amplitude of the fastest component, a shortening of the intermediate component (τ2), and decreased amplitude of the long-living component (τ3). In this case the Zea-containing sample has a shorter intermediate-lifetime component at 1.1 ns and a longer τ3 component. These effects, together, result in an average lifetime very similar to that of the LV sample. It is worth mentioning that no difference in fluorescence decay was observed in CP29 upon acidification.
A recently proposed mechanism for qE involves the transient formation of a Chl−Car+ radical cation, followed by charge recombination to the ground state (CT quenching). The formation of Car radical cation species can be detected by near-infrared (NIR) transient absorption (TA) spectroscopy, both in intact systems [67] and in purified Lhcb proteins [6],[31].
We performed ultrafast TA measurements on LhcSR3 samples by exciting Chl a at 670 nm and recording the absorption decay in the picoseconds timescale at 980 and 940 nm (NIR), respectively corresponding to Zea and lutein radical cations [31],[68]. In the NIR spectral region, Chl-excited state absorption was detected in addition to Car radical species. However, Chl-excited state absorption can easily be distinguished according to its kinetics, characterized by decay components only, whereas Car+ also displays a rise component. Figure 7 shows TA traces from LhcSR LV and LVZ samples. At both 980- and 940-nm detections, both LV and LVZ traces display a clear rise component, followed by a decay similar to what was previously described for higher plant monomeric Lhc proteins involved in NPQ [6],[67]. In the present study, both quenching and Car cation TA signal were a constitutive property of the LhcSR3 protein. Thus, the kinetic parameters were obtained directly from decay curves [31] rather than from quenched-minus-unquenched difference kinetic curves [6]. At 940 nm, the rise time was 14±3 ps, and decay time 428±103 ps, in both samples. The rise time at 980 nm was 7±4 ps, and decay time 501±224 ps. Interestingly, unlike previous results obtained in Lhc complexes from plants, the amplitude of the fast-rising component at 980 nm was not enhanced in the LVZ sample relative to the LV sample, indicating that the contribution of Zea radical cations to Zea-binding LhcSR3 is not dominant. Consistently, the LV sample displayed a clear rise component at both wavelengths, with similar kinetics. This indicates that Zea binding is not essential for Car radical cation formation in LhcSR3, at variance with previous findings in higher-plant monomeric Lhcb proteins [6],[31]. We then proceeded to verify the effect of pH on the formation of Car radical cation(s). To this aim we repeated the measurements upon lowering the pH to 5.5. Results (Figure 7C and 7D) show that the amplitude of the TA signal at 940 nm is increased by 40% and 90%, respectively, in LV and LVZ samples upon acidification. At 940 nm, the fastest component (<1 ps) is enhanced in both complexes at pH 5.5, while an increased amplitude of the slow rise component (∼23 ps) was observed only in LhcSR LVZ. It was previously reported that both the fastest and slowest rise components of TA are associated to Car radical cation formation, the former being related to energy transfer among Chls strongly coupled with the Chl-Car heterodimer responsible for charge separation, and the latter being associated with energy transfer among Chls in the complex [69]. This effect was not observed at 980 nm, consistent with the minor contribution of Zea in the process.
For information on which xanthophyll species are involved in generating the TA signal of LhcSR3, we reconstructed the NIR TA spectrum by recording kinetic traces in the 840–1,080 nm region and plotting TA signals at a delay time of 15 ps, which corresponds to peak amplitude. The resulting spectra are shown in Figure 7E and 7F. Major contributions are observed at various wavelengths, i.e., 920 and 980 nm in the LVZ sample, with the peaks of shorter wavelengths exhibiting the highest amplitude. The LV spectrum was shifted by 20 nm (900 and 960 nm). The signal amplitude at shorter wavelengths rose again towards 850 nm, the spectral range where violaxanthin radical cations are expected [68]. Although the signal-to-noise ratio of our NIR TA data decreases below 900 nm, violaxanthin involvement in CT quenching cannot be excluded in LhcSR3. Thus, LhcSR3 appears to have a high yield of radical cation(s) and also the capacity to produce this chemical species from various xanthophylls in the presence or absence of Zea. This is at variance with plant Lhcb proteins, in which Zea contributes to CT quenching both directly [6],[31] and as an allosteric activator of lutein CT quenching [32].
The above data (Figures 6 and 7) clearly suggest that acidification up-regulates quenching in LhcSR3, while Zea binding has a smaller effect. In order to further assess whether the effects observed in vitro on the isolated LhcSR3 protein are reflected in the level and/or kinetics of NPQ in vivo, we proceeded in two steps. First, we verified that NPQ in C. reinhardtii is sensitive to DCCD, a protein-modifying agent specific for reversibly protonatable residues. NPQ kinetics of HL acclimated cells with and without incubation with 20 µM DCCD is shown in Figure 8A and clearly demonstrates a strong inhibitory effect of DCCD on NPQ in vivo. In the second step, we measured NPQ kinetics in the npq1 mutant [70], unable to synthesize Zea, relative to in WT. The total NPQ amplitude was similar in the two strains. If anything, it was somehow rising faster and higher in the Zea-less mutant. Also, the dark recovery was faster in the npq1 mutant. We verified that the HL treatment was effective in inducing Zea synthesis in WT but not in the npq1 mutant by analyzing pigment content by HPLC before and after the actinic light treatment (1,600 µmol m−2 s−1). In doing that, we observed that HL-grown WT cells did contain a significant amount of Zea even after the 1-h dark incubation before the onset of the illumination, and the de-epoxidation index increased from 0.2 to 0.4 during measurement (Table S1). It should be noted that this behavior is significantly different from the case of higher plants, in which Zea is absent in dark-adapted plants, and the de-epoxidation index reaches 0.6 upon HL exposure.
Feedback energy dissipation is triggered by low lumenal pH. In plants, pH transduction is operated by PsbS through the protonation of two glutamate residues, which can be identified by labeling with DCCD, a chemical covalently binding to protonatable protein sites [18]. Although a psbS gene is present, the PsbS protein is not accumulated in algae [71], thus opening the question as to whether LhcSR is the molecule responsible for pH-dependent triggering of qE in algae. This hypothesis is supported by the presence of several acidic residues, potential candidates as pH-sensors in the LhcSR lumenal region sequence. We verified the capacity of LhcSR to bind DCCD by labeling the recombinant protein with 14C-DCCD, followed by autoradiography. Other Lhcb proteins, including CP29 from plants carrying a DCCD-binding site [72] and the algal CP29, CP26, and Lhcbm1, were analyzed for comparison. The results are shown in Figure 9: LhcSR3 revealed very efficient binding of DCCD, higher than plant CP29. It is interesting to note that algal Lhcb proteins show DCCD binding of approximately 50% with respect to LhcSR3, but clearly higher than plant LHCII, indicating that some level of pH responsiveness may be a general property of algal PSII antenna proteins (Figure 9B and 9C).
A recent report showed that although a psbS gene is present in algae, the corresponding protein is not accumulated [71]. This is of particular interest, since the npq4 mutant in C. reinhardtii, blocked in thermal dissipation, is disrupted in the lhcSR3.1 and lhcR3.2 genes encoding identical Lhc-like proteins [14],[20],[22]. This implies that the mechanism of feedback de-excitation differs in algae versus plants and that PsbS action can be carried out by a different component(s). The properties of LhcSR3 are thus of primary importance for understanding qE in algae.
Accumulation of LhcSR3 strongly depends on light intensity during growth: of the three LhcSR immuno-reactive bands that can be resolved by SDS-PAGE, the upper two are strongly up-regulated in HL-grown cells, whereas the fast-migrating isoform is of low intensity and decreases in HL. The correspondence of these bands with the three lhcSR genes found in the C. reinhardtii genome [36] is based on the fact that mutants deleted in the lhcSR3.1 and lhcSR3.2 genes are also missing the two higher bands [22], consistent with predictions based on polypeptide MW, with lhcSR1 encoding a smaller protein than lhcSR3.1 and lhcSR3.2. LhcSR1 and LhcSR3 are both over-expressed in HL [22] and iron starvation conditions [73]. The two bands with higher apparent MW correspond to the phosphorylated and unphosphorylated LhcSR3 isoforms, according to the results of phosphatase treatment (Figure 1). The increase in LhcSR3 accumulation strongly correlates with the amplitude of NPQ, consistent with the report that LhcSR3 protein is responsible for high NPQ levels in C. reinhardtii [22]. We show here that the stt7 mutant [39] is unable to phosphorylate LhcSR3 to any significant extent, yet it exhibits NPQ as in WT, or even higher. We thus conclude that phosphorylation is not indispensable for NPQ, but interpret the increased NPQ in stt7 as a consequence of the block in State 1–State 2 transitions, a mechanism active in energy pressure balancing in algae [74], thus increasing PSII over-excitation and the need for energy dissipation through NPQ. Phosphorylation does not appear to affect the aggregation state of LhcSR either, as detected by native Deriphat-PAGE (Figure 3). Migration of dimeric Lhc proteins between monomeric and trimeric Lhcbs has been reported for Lhca1–Lhca4 [2],[42],[75]. Since LhcSR1 is much less abundant than LhcSR3, and as we found no evidence for the presence of other Lhcb proteins in dimers, homodimeric organization is most likely, although alternative hypotheses cannot be entirely excluded.
Pigment binding is an important property for evaluating the role of LhcSR in excitation energy quenching: a pigment-binding protein may be directly involved in the quenching reaction [6], whereas a non-pigment-binding protein cannot, although it may play an ancillary role such as pH sensing, as previously found for PsbS [18],[19]. Both LhcSR isoforms are here shown to form stable and specific complexes with Chl and xanthophyll chromophores, as clearly demonstrated by (i) the spectral shift induced by pigment-protein interactions (Figure 5), (ii) the capacity for excitation energy transfer from Chl b and xanthophylls to Chl a, and (iii) for the LhcSR3 isoform, the strong optical activity: free pigments in detergent solution or unspecifically bound to proteins have very low amplitude CD spectra, with a single broad positive component in the Qy region [34]. Thus, recombinant LhcSR has the same properties as Lhc antenna complexes, unlike PsbS, which cannot form pigment proteins in vitro or in vivo [19],[25]. It is worth noting that, since LhcSR can coordinate pigments in vitro, it is highly unlikely that this does not occur in vivo. Although nonspecific binding of Chl to proteins cannot, in principle, be excluded, it is hard to imagine that coordination is carried out in such a specific way that it provides stoichiometric binding of Chls and xanthophylls, CD signals, and efficient energy transfer between chromophores, essentially very similar to other members of Lhc protein family, without reflecting an original capacity for pigment binding in vivo. The conservation of six pigment-binding residues with respect to other Lhc protein members further supports the pigment-binding nature of LhcSR3.
As regards the number and organization of chromophores in LhcSR3, the Chl/protein ratio indicates six or seven Chls per polypeptide, consistent with the non-conservation of Chl-binding residues at B3 and B6 binding sites compared with CP29, which binds eight Chl [33] (Figure 4). Interestingly, the Chl b complement is slightly below one per polypeptide: while Chl b may have a stabilizing effect because of the establishment of hydrogen bonds through its vinyl group [46], a small fraction of the LhcSR3 pigment-protein complexes may bind only Chl a and xanthophylls. This is consistent with the observation that LhcSR3 protein accumulates in the cbs3 mutant, lacking Chl b, and the finding of LhcSR orthologs in diatoms, which lack Chl b [76]. In fact, a Chl a/lutein/violaxanthin complex may be obtained with spectral properties and pigment composition similar to the Chl b–containing holoprotein.
Lutein and violaxanthin are the major xanthophylls in LhcSR3; neoxanthin and loroxanthin are absent. Neoxanthin is mainly bound to the major LHCII trimeric antenna [77],[78] and is involved in scavenging superoxide anions, not in qE [3]. The function of loroxanthin, still unknown, is probably related to enhancing light-harvesting efficiency, since its content is increased in LL cells and decreased in HL conditions [78]. Based on six Chls per polypeptide, more than two xanthophylls per polypeptide are calculated with lutein and violaxanthin bound, respectively, to sites L1 and L2, although site selectivity is less strict than in the case of most Lhc proteins [56],[58],[60],[79]–[81], not only for lutein and violaxanthin but also for Zea. According to its spectral shift (Figure S2), the additional xanthophyll ligand is bound to a third V1-like site. A model of the LhcSR3 holoprotein with bound pigments is shown in Figure 10.
Although the recombinant proteins obtained by in vitro reconstitution are monomers, as determined from their mobility in sucrose gradient and native PAGE (data not shown), the native state of the complex appears to be at least partly dimeric (Figure 3). So far, dimeric Lhcs have been reported for PSI only, i.e., Lhca1/Lhca4 and Lhca2/Lhca3 [2],[82]. We cannot exclude the possibility that LhcSR dimers have other properties in addition to those reported here, due to pigment-pigment interactions between subunits and/or additional pigments bound at protein-protein interfaces [52],[83]. Nevertheless, recombinant Lhca proteins have been shown to have all the major biochemical and functional properties typical of LHCI complexes isolated from thylakoids [49].
Accumulation of LhcSR3 is greatly enhanced in HL conditions (Figure 1) [22] or nutrient deficiency [73]. This is similar to the expression pattern of other Lhc-like gene products such as PsbS and ELIPs, whose involvement in photoprotection mechanisms has been well documented [15],[16],[18] and contrasts with the case of the major LHCII antenna complex, whose expression/accumulation is enhanced in light-limiting conditions [12],[84]. Also, like PsbS [85] and ELIPs, LhcSR3 is present in substoichiometric amounts with respect to the PSII reaction center. These characteristics indicate that it is involved in excitation energy dissipation in order to prevent photoinhibition, and that its activity/abundance can be modulated, depending on environmental requirements. The substoichiometric amount of a strong quenching molecule does not exclude the possibility of its being an efficient quencher for photosystems II with a high degree of connectivity [86]. Even though we have evidence that LhcSR3 might fulfill a PsbS-like role as a sensor of lumen acidification, as supported by its decrease in lifetime (Figure 6; Table 2) and by the presence of protonatable DCCD-binding residues (Figure 9), our data suggest that it displays intrinsic capacity for direct excitation energy quenching. This conclusion is supported by the pigment-binding capacity and the spectroscopic properties of LhcSR3: the dominant fluorescence lifetime components, shorter than 100 ps, imply that an energy dissipation channel is constitutively active in recombinant LhcSR3, and its activity is further enhanced upon acidification. Other Lhc proteins have much slower fluorescence decay rates, in the range of 3.4–4.5 ns, consistent with their function as sensitizers for photosystem II [65]. Faster components (1.4–2.8 ns) have been resolved in low abundance, deriving from alternative conformations [64]. In the case of LhcSR3, the dominant lifetime is below 100 ps, suggesting that a third state of the protein is stabilized, having strong energy dissipation activity. Nevertheless, a significant fraction of the fluorescence was detected with a lifetime of 1.95 ns. Since energy equilibration within a monomeric Lhc protein is completed within a few picoseconds [87]–[89], we suggest that the two lifetime components (<100 ps versus 1.95 ns) are due to different molecular forms whose abundance is regulated by pH. Heterogeneity may be related to the lower-than-one stoichiometry of Chl b versus LhcSR apoprotein, implying that a particular site may be occupied by either Chl a or Chl b. One additional source of heterogeneity is evidenced by the resolution of two Chl b spectral forms (Figure 5) deriving from binding to distinct protein sites, matching previous reports on Lhc proteins [33],[90]. According to the hypothesis that Chl a–Chl a excitonic interactions lead to red-shifting of the LhcSR3 spectrum (Figure 5), occupancy by Chl b of a site potentially involved in interactions would prevent excitonic coupling, because of the large difference in site energy of the two chromophores. In CP29, Chl a in binding site B5 can undergo excitonic interaction with Chl a in site A5, thus promoting CT quenching [6]. We propose that excitonic interaction between Chl a molecules in sites A5 and B5 is involved in the quenching process active in LhcSR3, and that the heterogeneity of site B5 occupancy is responsible for lifetime heterogeneity. This matches the fact that Chl A5 is a Chl a–specific site in all Lhc proteins, whereas site B5 is promiscuous in monomeric Lhcs [33],[90]. Spectral/lifetime heterogeneity may (also) be provided by mixed Chl a/b occupancy of site A3 [35].
Energy dissipation in CP29 has been reported to derive from the transient formation of a Zea radical cation and a Chl anion, followed by charge recombination to the ground state [67] in an Lhcb protein domain including Chl A5, Chl B5, and Car in site L2 [6]. Triggering of the energy dissipation reaction is obtained by displacement of violaxanthin in site L2 with Zea or lutein: this event induces a conformational change, leading to the establishment of an excitonic interaction between Chl a molecules in sites A5 and B5 [6],[32]. The coupled Chl dimer is more favorable to CT quenching, because the charge delocalization over the two Chls lowers the energy requirement for CT quenching [6]. We observed a 684–688 nm red-shifted absorption form in LhcSR3 with large bandwidth, which is not present in CP29 or in any other Lhcb protein so far described, to our knowledge. We suggest that this form derives from the strong excitonic interaction between Chl A5 and B5, constitutively present in this protein, without the need for binding of Zea. This view is supported by the high levels of Car radical cations measured in LhcSR3 (Figure 7), approximately ten times higher than in plant Lhcb4–Lhcb6 proteins in their active, Zea-binding form. With the concomitant presence of the short fluorescence lifetime component (<100 ps), we conclude that LhcSR3 is predominantly stabilized in the energy dissipation conformation that is transiently induced in plant monomeric Lhcb4–Lhcb6 proteins. Mutation analysis is in progress in order to confirm this hypothesis. To our knowledge, this is the first example of an Lhcb protein exhibiting a dominant dissipative conformation when isolated in detergent solution. Besides CT quenching, direct energy transfer from a Chl a Qy transition to a lutein S1 state [91] or a Chl-Chl charge transfer [92] have been proposed as alternative mechanisms for qE. Although we have no evidence that these processes are important for energy dissipation in algae, we cannot exclude the possibility that they may contribute to quenching. In fact, the fast lifetime component we resolved in LhcSR3 is below 100 ps, i.e., it is significantly faster than the relaxation of the lutein radical cation (see Results), indicating the involvement of multiple quenching mechanisms. More detailed spectroscopic analysis is needed to assess whether CT quenching is the only component of qE in algae or whether other mechanisms are also involved, as well as high-resolution structural studies on LhcSR3 in order to elucidate the molecular architecture of the quenching site in its active, energy-dissipating state.
Energy quenching is dependent on lumenal pH in both plants and algae, as clearly shown by its sensitivity to uncouplers [93] and the inhibition of NPQ by DCCD (Figure 8A). Yet, algae lack PsbS, the sensor of lumenal pH [18],[19]. The observation that the quenched conformation of the LhcSR3 protein is stable in detergent solution (i.e., in the absence of a transmembrane pH gradient) raises the question of how qE is modulated by the onset of light: in fact, active quenching in the dark or in LL conditions would impair photosynthesis and cell growth. In order to explain pH regulation of quenching in algae, we propose that LhcSR3, although present as an active quencher in thylakoids, is disconnected from other Lhc proteins, thus minimizing energy dissipation, while it establishes interactions with PSII antenna component(s) upon lumen acidification and protonation of lumen-exposed, negatively charged residues both in LhcSR and in PSII antenna components. This model is consistent with both earlier and new observations. First, lack of Lhcbm1, a major component of trimeric LHCII, has been shown to reduce qE strongly [43]. Second, the cbs3 mutant, although accumulating LhcSR3 in HL, cannot develop high NPQ, perhaps because of the lack of an Lhcbm1 partner for LhcSR3. Low lumenal pH also increases the formation of lutein radical cation (Figure 7C and 7D) and increases the amplitude of short-living fluorescence lifetime components (Figure 6B) in the isolated protein, suggesting that lumen acidification, besides promoting connection of the LhcSR quencher to the light-harvesting antenna system, also enhances the quenching activity of the pigment-protein complex.
Besides low lumenal pH, an additional factor in triggering NPQ in plants is synthesis of Zea in excess light. Zea is incorporated into Lhc proteins [60],[94],[95] and promotes dissociation of a pentameric Lhc complex, which is needed to trigger NPQ [27]. The fluorescence lifetime of isolated LhcSR3 in detergent solution is not strongly affected by Zea. The npq1 mutation, preventing Zea synthesis, has been shown to decrease NPQ in plants [10], while in C. reinhardtii the effect of Zea is much reduced [26],[95]. We observed essentially the same NPQ activity in WT and in the npq1 mutant, consistent with the small effect of Zea on the lifetime properties of LhcSR3 in vitro and with LhcSR3 being essential for NPQ in vivo [22]. It is thus possible that some level of NPQ dependence on Zea can be observed in some conditions as a consequence of its binding to antenna protein interacting with LhcSR3, possibly Lhcbm1 [43].
We have shown that LhcSR3, essential for energy quenching in C. reinhardtii, is a pigment-binding protein with the properties of a constitutive quencher, since it has a short lifetime component (<100 ps) when isolated in detergent solution. This is different from the case of plant monomeric Lhcb proteins, which have long lifetimes and whose quenching mechanisms are activated in vivo by the action of the PsbS protein and/or Zea synthesis. We propose that LhcSR3 regulates energy dissipation by establishing reversible interactions with other Lhcb antenna proteins, in particular Lhcbm1 [43], and that these interactions are induced by low lumenal pH through protonatable DCCD-binding sites present in both Lhcb proteins and LhcSR3. Thus, LhcSR3 has the properties of both an energy quencher, a function catalyzed by Lhcb proteins in vascular plants [6],[31],[91], and a sensor for lumenal pH, which is a function of PsbS in plants [18],[19],[27].
WT strains cw15 [96] and CC425 (Chlamydomonas Genetics Center, Duke University) and mutants stt7 [97], npq1 [70], and cbs3 [40] C. reinhardtii cells were grown in high-salt medium [96] at light regimes of 500 and 50 µmol m−2 s−1 for HL and LL samples, respectively. In Figure 8B, the npq1 mutant and WT were from the CC425 strain [70].
Cells acclimated to HL or LL conditions were harvested in the exponential growth phase (∼2×106 cells/ml), pelleted, and resuspended at a concentration of ∼108 cells/ml). Cells were pre-illuminated for 2 min with a weak (3 µmol m−2 s−1) far-red LED before NPQ analysis with a PAM-101 (Waltz); actinic light was 1,600 µmol m−2 s−1 and saturating light, 4,080 µmol m−2 s−1. The far-red LED was kept on during dark recovery. In the experiment of Figure 8A, cells were pre-incubated with 20 µM DCCD (Sigma) for 15 min in the dark before measurements.
C. reinhardtii thylakoids were purified as previously described [98], and membrane dephosphorylation was carried out by incubating one sample at 28°C for 1 h in the presence of calf intestinal alkaline phosphatase (1 Unit/3 µg Chl).
Denaturing SDS-PAGE as described previously [52] was performed in the presence of 6 M Urea with the Tris-sulfate acrylamide and Tris-glycine buffer systems [99]. The gel was transblotted to a nitrocellulose filter, decorated with an anti-LhcSR serum, and developed by means of the alkaline phosphatase detection system.
Thylakoid membranes were solubilized in the presence of 1.2% α-dodecyl-maltoside and loaded on native electrophoresis gels [42].
The DNA sequence coding mature LhcSR1 and LhcSR3.1 was cloned in the pET28 vector (Novagen) and transformed in E. coli BL21de3 cells. The recombinant proteins were purified as inclusion bodies from bacterial lysate as previously described [34].
The refolding procedure was performed as described in [34].
CD spectra were obtained with a Jasco J-600 spectropolarimeter with scan rate 200 nm/min. Absorption spectra were obtained with an AMINCO DW2000 spectrophotometer, with scan rate 2 nm/s, bandwidth 1 nm, and optical path length 1 cm. Fluorescence spectra were obtained at room temperature with a Fluoromax 3 fluorometer (Horiba Jobin Yvon). Time-resolved fluorescence spectroscopy was carried out at room temperature with the single-photon-timing method on a FluoTime 200 from PicoQuant. Kinetics were analyzed with FluoFit from PicoQuant. Excitation was at 435 nm, and detection was at 680, 690, 700, and 710 nm.
Pigments were extracted from pelleted cells, and samples were frozen in liquid nitrogen and resuspended in 80% acetone buffered with Na2CO3. The supernatant of each sample was then recovered after centrifugation (15 min at 15,000 g, 4°C). Separation and quantification of pigments was performed by HPLC [100]. Chl a/b and Chl/Car ratios were corrected through fitting analysis of the absorption spectrum [80].
The NIR TA laser system has previously been described [6],[31],[67]. Briefly, the repetition rate was 250 kHz, and the pump pulses were tuned to ∼670 nm. The maximum pump energy and full width at half maximum of the pulse auto-correlation trace were ∼24 nJ/pulse and ∼40 fs, respectively. White light continuum probe pulses were generated in a 1-mm quartz plate. Observation of the cross-correlation function of the pump and probe overlap was approximately 85 fs. The mutual polarizations of the pump and probe beams were set to the magic angle (54.7°). A monochromator (Spectra Pro 300i; Acton Research) and an InGaAs photodiode (DET410; Thorlabs) were used to monitor transmission. A sample cell for isolated LHCs with a path length of 1 mm was chilled by a circulating water bath (VWR Scientific 1160; PolyScientific) set at 7°C during data acquisition to prevent sample degradation. For TA measurements at lower pH, samples were placed in 40 mM citrate buffer (pH 5.5) with 0.2% α-dodecyl-maltoside.
Recombinant C. reinhardtii Lhcbm1, CP29, and CP26 were expressed in E. coli and refolded with pigments [34]. Z. mays LHCII and CP29 were purified in their native form as described in [72]. All samples were labeled with 14C-DCCD (Amersham) following the methods of [72] and loaded on SDS-PAGE electrophoresis gels [101]. After Coomassie staining, gels were dried, and radioactivity was revealed through autoradiography.
|
10.1371/journal.pntd.0001044 | Efficacy and Safety of Single and Double Doses of Ivermectin versus 7-Day High Dose Albendazole for Chronic Strongyloidiasis | Strongyloidiasis, caused by an intestinal helminth Strongyloides stercoralis, is common throughout the tropics. It remains an important health problem due to autoinfection, which may result in hyperinfection and disseminated infection in immunosuppressed patients, especially patients receiving chemotherapy or corticosteroid treatment. Ivermectin and albendazole are effective against strongyloidiasis. However, the efficacy and the most effective dosing regimen are to be determined.
A prospective, randomized, open study was conducted in which a 7-day course of oral albendazole 800 mg daily was compared with a single dose (200 microgram/kilogram body weight), or double doses, given 2 weeks apart, of ivermectin in Thai patients with chronic strongyloidiasis. Patients were followed-up with 2 weeks after initiation of treatment, then 1 month, 3 months, 6 months, 9 months, and 1 year after treatment. Combination of direct microscopic examination of fecal smear, formol-ether concentration method, and modified Koga agar plate culture were used to detect strongyloides larvae in two consecutive fecal samples in each follow-up visit. The primary endpoint was clearance of strongyloides larvae from feces after treatment and at one year follow-up.
Ninety patients were included in the analysis (30, 31 and 29 patients in albendazole, single dose, and double doses ivermectin group, respectively). All except one patient in this study had at least one concomitant disease. Diabetes mellitus, systemic lupus erythrematosus, nephrotic syndrome, hematologic malignancy, solid tumor and human immunodeficiency virus infection were common concomitant diseases in these patients. The median (range) duration of follow-up were 19 (2–76) weeks in albendazole group, 39 (2–74) weeks in single dose ivermectin group, and 26 (2–74) weeks in double doses ivermectin group. Parasitological cure rate were 63.3%, 96.8% and 93.1% in albendazole, single dose oral ivermectin, and double doses of oral ivermectin respectively (P = 0.006) in modified intention to treat analysis. No serious adverse event associated with treatment was found in any of the groups.
This study confirms that both a single, and a double dose of oral ivermectin taken two weeks apart, is more effective than a 7-day course of high dose albendazole for patients with chronic infection due to S. stercoralis. Double dose of ivermectin, taken two weeks apart, might be more effective than a single dose in patients with concomitant illness.
ClinicalTrials.gov NCT00765024
| Strongyloidiasis, caused by an intestinal helminth Strongyloides stercoralis, is common throughout the tropics. We conducted a prospective, clinical study to compare the efficacy and safety of a 7-day course of oral albendazole with a single dose of oral ivermectin, or double doses, given 2 weeks apart, of ivermectin in Thai patients who developed this infection. Patients were regularly followed-up after initiation of treatment, until one year after treatment. Ninety patients were studied (30, 31 and 29 patients in albendazole, single dose, and double doses ivermectin group, respectively). The average duration of follow-up were 19 (range 2–76) weeks in albendazole group, 39 ( range 2–74) weeks in single dose ivermectin group, and 26 ( range 2–74) weeks in double doses ivermectin group. Parasitological cure rate were 63.3%, 96.8% and 93.1% in albendazole, single dose oral ivermectin, and double doses of oral ivermectin respectively. No serious adverse event associated with treatment was found in any of the groups. Therefore this study confirms that both a single, and a double dose of oral ivermectin taken two weeks apart, is more effective than a 7-day course of high dose albendazole for patients with chronic infection due to S. stercoralis.
| Infection with the intestinal helminth Strongyloides stercoralis remains a common problem throughout the tropics, including Thailand [1], [2]. It is estimated that 30 to 100 million people are infected worldwide [1]. Most infected individuals are asymptomatic or developed minimally symptomatic chronic infection through autoinfection [3]. Potentially fatal disseminated infections, due to an acceleration of the autoinfection cycle, are seen in immunocompromised patients, such as those with concurrent human T-lymphotropic virus-1 (HTLV-1) infection, or those on corticosteroid therapy [3], [4]. Other recognized risk factors for disseminated strongyloidiasis include malignancies especially lymphoma, organ transplantation and diabetes mellitus [5], [6]. Gastrointestinal symptoms associated with strongyloidiasis include diarrhea, abdominal discomfort, nausea/vomiting and anorexia. The diagnosis of strongyloidiasis should be suspected if there are clinical signs and symptoms, or eosinophilia [7]. Definitive diagnosis of strongyloidiasis is usually made on the basis of detection of larvae in the stool [8]. The combination of diagnostic approaches such as repeated direct microscopic examination of fecal smear, fecal concentration methods such as formol-ether concentration (FEC), and modified Koga agar plate culture have been used to improve the likelihood of detecting this parasite [7]–[11].
In the past, the treatment of choice for strongyloidiasis has been thiabendazole, but this drug has unpleasant side effects and is no longer available. Albendazole, another broad-spectrum antihelmintic agent, was previously shown to be effective against S. stercoralis [12]–[15]. More recent reports suggest ivermectin, a macrolide-like agent developed primarily for the treatment of onchocerciasis, is as effective as thiabendazole [16] and superior to albendazole against intestinal strongyloidiasis [17]–[21].
Although a single dose of ivermectin 200 microgram/kilogram body weight (µg/kg) was shown to be effective in uncomplicated chronic strongyloidiasis, repeated treatment at two or three week intervals was thought to be necessary to eliminate larvae generated by autoinfection [22].
A preparation of oral ivermectin licensed for human use has recently become available in Thailand. However, albendazole remains the most widely used antiparasitic drugs for the treatment of this infection in this country. The purpose of the present study was to assess the safety and efficacy of a single dose of ivermectin (200 µg/kg), or two doses of ivermectin given 2 weeks apart, and a 7-day course of high dose albendazole for the treatment of chronic strongyloidiasis in adult patients who were at high risk of hyperinfection or disseminated infection.
This was a prospective open-label, randomised, controlled study conducted between July 2008 and April 2010 at Siriraj Hospital, Faculty of Medicine Siriraj Hospital, Mahidol University, Bangkok, Thailand. The study was approved by the Ethical Committee on Research Involving Human Subjects, Siriraj Hospital, Faculty of Medicine, Mahidol University, Thailand. All patients were informed about the purpose of the trial and gave written informed consent before enrollment. The study enrollment was stopped in December 2009 after 100 eligible patients had been recruited.
Adult patients (>18 years) were recruited from Siriraj Hospital if characteristic rhabditiform larvae of S. stercoralis were present on fecal microscopy. Exclusion criteria included a history of allergic reaction to either study medication, treatment within the month prior to the study with any drug known to have anti-strongyloides activity, pregnancy or lactation and any suggestion of disseminated strongyloidiasis.
Computer generated, simple, random allocation sequences were prepared for 3 study groups by the investigator team. These were sealed in an opaque envelope and numbered. The investigator (YS) assigned study participants to their respective treatment group after opening the sealed envelope. Once an eligible patient was identified and informed consent was obtained, the patient was randomly allocated to one of the following group (1∶1∶1 ratio):
Baseline evaluation included history, detailed physical examination, and laboratory investigations such as complete blood count (CBC), urinalysis, and biochemistry. Patients were requested to collect two consecutive fecal samples at every hospital visit. The coprodiagnosis for the detection of S. stercoralis larvae using direct smear, formol-ether concentration method [9], and modified Koga agar plate culture method [10] was performed for each patient at the Infectious Diseases and Tropical Medicine Laboratory, Division of Infectious Diseases and Tropical Medicine, Faculty of Medicine Siriraj Hospital, Mahidol University, Thailand.
Patients were required to make seven hospital visits to complete the study: at baseline evaluation and initiation of treatment, at 2 weeks after initiation of treatment, then at 1 month, 3 months, 6 months, 9 months, and 1 year after treatment. Patients who completed 1 year of follow-up were invited for further follow-up visits every 3 months or at their convenience.
One hundred and fifty one patients had detectable rhabditiform larvae of S. stercoralis on fecal microscopy during the study period. One hundred patients were enrolled (36, 32, and 32 patients in albendazole, ivermectin-I, and ivermectin-II groups respectively). Ten patients were excluded from analysis because they did not receive or complete the study treatment (3 in albendazole group, 2 in ivermectin-II group), or they were lost to follow-up immediately after treatment (3 in albendazole group, 1 each in ivermectin-I and ivermectin-II respectively). Overall, 90 patients were eligible for the modified intention to treat analysis. Detail of the total number of enrollment, randomization, follow-up and inclusion in the final analysis comparing among the three treatment groups is shown in Figure 1.
The demographic data, concomitant diseases, baseline clinical and laboratory investigations are shown in Table 1 and 2. All except one patient had an associated medical problem, including concurrent other parasitoses. These patients also had abnormal serum aspartate aminotranferase (AST) and alanine aminotransferase (ALT) levels prior to entering the study due to their underlying conditions.
The intensity of initial infection of the three study groups was similar, i.e. S. stercoralis larvae were found from the direct fecal examination in 24 (80%), 25 (80.6%), and 28 (96.6%) in albendazole, Ivermectin-I, and ivermectin–II groups, respectively (P = 0.123). Larvae were also detected from modified Koga agar plate culture in 22/26 (84.6%) patients in the albendazole group, in 22/26 (84.6%) patients in the ivermectin-I group, and in 24/29 (82.8%) patients in the ivermectin-II group, respectively (P = 0.976).
Diarrhea was detected in half of the patients and it was relieved after treatment in most patients. Abnormal bowel movement at second week of follow-up was reported in 4 patients in the albendazole group, 2 patients in the ivermectin-I group, and 3 patients in the ivermectin-II group, respectively (P = 0.641). S. stercoralis larvae were detected in one patient in the ivermectin-I group at third month of follow-up. In ivermectin-II group, S. stercoralis larvae were detected at second week prior to the second dose of ivermectin in two patients. No patients had reinfection/relapse after the second dose of ivermectin treatment. In albendazole treated patients, S. stercoralis larvae were detected at second week of follow-up in 2 patients, at first month of follow-up in 2 patients, between 3–6 months of follow-up in 3 patients, and between 6–12 months of follow-up in 4 patients. All of the relapses/ reinfections found during follow-up were clinically inapparent.
Parasitologically, parasite elimination was documented in 19 (63.3%) albendazole treated patients, in 30 (96.8%) single-dose ivermectin treated patients, and 27 (93.1%) two-dose ivermectin treated patients (P = 0.006) (Table 3). Cox regression analysis showed that albendazole treated group had 14.7 times (95%CI 1.8–111.9), and 5.7 times (95%CI 1.3–25.7) higher risk for reinfection/ relapse of strongyloidiasis than ivermectin-I and ivermectin-II group, respectively. Kaplan- Meier Plot compares the parasitological cure rate between these study groups is shown in figure 2. No hyperinfection syndrome or disseminated infection was found among these patients during the study period. S. stercoralis larvae were detected after treatment using FEC in 8 patients, and by modified Koga agar plate culture only in 6 of them. All patients with relapse/reinfection were retreated with two doses of ivermectin in two weeks apart.
Overall albendazole and ivermectin were well tolerated. Transient elevation of AST, and ALT levels was detected in one patient in ivermectin-II group. The AST and ALT levels returned to normal 2 weeks after the second dose of ivermectin treatment. Severe nausea and vomiting was reported in one patient in the albendazole group.
Fifteen patients died after enrollment (5 patients in each treatment group). Causes of death were not related to the study drugs, and were considered to be due to an underlying disease or its complications (solid tumor in 5, hematologic malignancies in 3, diabetes mellitus, or systemic lupus erythrematosus (SLE), or hypertension with complications such as myocardial infarction or sepsis in 7 patients). The median duration from enrollment to death was 2 weeks (range 2–14 weeks) in the albendazole group, 5 weeks (range 2–38 weeks) in the ivermectin-I group, and 2 weeks (range 1–27 weeks) in the ivermectin-II group, respectively.
Strongyloidiasis remains a significant health problem in many developing countries, mainly due to the potential for lethal disseminated disease [1], [5]. Gastrointestinal symptoms associated with strongyloidiasis found in this study included diarrhea, abdominal discomfort, nausea/vomiting and anorexia. Chronic infection with S. stercoralis was clinically inapparent in half of the patients at enrollment, and in all of relapses/ reinfections found during follow up. Peripheral eosinophilia (>500 eosinophils/µL.) was detected in half of the patients at enrollment. S. stercoralis larvae were detected after treatment using FEC in 8 patients, and by modified Koga agar plate culture in 6 patients. This information confirmed that fecal examination, including culture and/or serology, every 3–6 months of follow-up should be recommended for early detection and treatment of latent infection to prevent hyperinfection or disseminated disease in these patients [3], [5].
Results of this study corroborate the results from previous randomised controlled studies on the higher efficacy of ivermectin compared to various dosage regimens of albendazole for treating chronic strongyloidiasis [17]–[21]. A summary of results from these previous controlled trials of ivermectin treatment for chronic strongyloidiasis is shown in Table 4. Although these studies were conducted in different geographical areas and population groups, i.e. in children and adults, they were considered to be within a community-based setting, such as schools or primary care clinic. The duration of follow-up varied from 3 weeks to 12 months. The present study was conducted in a tertiary hospital. The majority of patients had known risk factors for disseminated strongyloidiasis, and approximately one-third of them received corticosteroid or chemotherapy. Results of this study confirmed that ivermectin was also effective in this population who were at high risk of severe infection.
Albendazole remains an option of treatment for chronic strongyloidiasis in many countries in South East Asia, where oral ivermectin is not widely available. Cure rates of a regimen consisting of albendazole 400 mg daily for three to five days varied from 38–87% in those without underlying diseases [14], [17]–[20]. In this study, the cure rate was found to be 63% when a 7-day course of high dose albendazole was used. The efficacy of albendazole varied widely even when the same dose and duration of treatment was used. Differences in duration of follow-up examinations could be one explanation, and re-infection from the environment may also be a factor when the efficacy is monitored for an extended period in endemic areas. The study which reported the highest cure rate (87%) was conducted in Okinawa, Japan [19], where the chance of re-infection from the environment was less likely to occur compared to other studies conducted in endemic areas [17], [18], [20], [21].
Two patients in the ivermectin-II group had detectable S. stercoralis larvae in the second week prior to the second dose of ivermectin treatment. One patient in ivermectin-I group also had detectable S. stercoralis larvae 3 months after treatment. This observation supports the recommendation that repeated doses of ivermectin should be the preferred treatment in patients with chronic strongyloidiasis who have an underlying or concomitant illness [22].
The limitation of this study was the high loss to follow-up rates over time. High mortality associated with the concomitant illnesses was an unavoidable cause of concern in this study. The median duration of follow up was 19 weeks in albendazole group, 39 weeks in ivermectin-I, and 26 weeks in ivermectin-II group. The non-significant shorter duration of follow up found in albendazole treatment group was due to the significant higher rate of treatment failure compared to ivermectin. However, the study still had sufficient power to detect a difference between albendazole and ivermectin treatments. This study, however, was too small to detect any but the most severe and common side- effects of both albendazole and ivermectin. Only one of albendazole treated patients and one treated with ivermectin had transient changes in transaminases, a well-recognized and reversible adverse event.
In conclusion, this clinical study confirms that both a single and a double dose of oral ivermectin taken at a two-week interval is more effective than a 7-day course of high dose of albendazole for patients with chronic infection due to S. stercoralis.
|
10.1371/journal.ppat.1002540 | Transient Reversal of Episome Silencing Precedes VP16-Dependent Transcription during Reactivation of Latent HSV-1 in Neurons | Herpes simplex virus type-1 (HSV-1) establishes latency in peripheral neurons, creating a permanent source of recurrent infections. The latent genome is assembled into chromatin and lytic cycle genes are silenced. Processes that orchestrate reentry into productive replication (reactivation) remain poorly understood. We have used latently infected cultures of primary superior cervical ganglion (SCG) sympathetic neurons to profile viral gene expression following a defined reactivation stimulus. Lytic genes are transcribed in two distinct phases, differing in their reliance on protein synthesis, viral DNA replication and the essential initiator protein VP16. The first phase does not require viral proteins and has the appearance of a transient, widespread de-repression of the previously silent lytic genes. This allows synthesis of viral regulatory proteins including VP16, which accumulate in the cytoplasm of the host neuron. During the second phase, VP16 and its cellular cofactor HCF-1, which is also predominantly cytoplasmic, concentrate in the nucleus where they assemble an activator complex on viral promoters. The transactivation function supplied by VP16 promotes increased viral lytic gene transcription leading to the onset of genome amplification and the production of infectious viral particles. Thus regulated localization of de novo synthesized VP16 is likely to be a critical determinant of HSV-1 reactivation in sympathetic neurons.
| Herpes simplex virus is a widespread human pathogen that establishes permanent infections in nerves innervating the lips, eyes and other surfaces. The viral DNA genome is transported to the neuronal nucleus located in the nerve ganglia, where it establishes a semi-dormant state known as latency. Periodically, latent viruses undergo reactivation, a process that leads to the production of infectious particles, allowing for person-to-person transmission and acting as the major source for painful lesions (cold sores) and other more severe pathological outcomes. How latency and reactivation are controlled is not well understood. Using cultured nerve cells, we show that reactivation involves a unique, two stage program of viral gene expression. We find that the essential control protein VP16 is synthesized during the first stage but accumulates in the cytoplasm rather than the nucleus where it functions. Nuclear entry is determined by host signaling and marks the onset of the second reactivation stage. This work provides important new insights into the virus-host interaction and reveals a natural control point that could be used in innovative therapies that for the first time target the latent virus.
| The remarkable success of the herpesviruses as infectious agents stems from their ability to alternate between productive (acute) replication and latency; distinct genetic programs that achieve very different outcomes for both the virus and the host cell. Acute replication results in release of infectious particles by cell lysis and produces a strong immunological stimulus, whereas in latency the lifespan of the host cell is often extended and the viruses use various strategies to minimize antigen presentation. By alternating between these two programs, herpesviruses can often remain in their host indefinitely but at the same time retain the ability to spread through reactivation, a process whereby latent virus reenters the productive replication cycle and infectious particle are shed at the surface.
The prototype example for this successful strategy is herpes simplex virus type-1 (HSV-1), a widespread human pathogen that infects epithelial cells in the oral cavity, eyes and other regions of mucosa. Latency is established in the ganglia of peripheral nerves that innervate these sites, creating a lifelong reservoir that is shielded from immune clearance (reviewed in [1], [2]). Intermittent reactivation events give rise to infectious particles that travel to the periphery by anterograde axonal transport. Continuous reemergence of virus from the permanent neuronal reservoir ensures lifelong transmission and is often associated with clinical disease.
How the HSV-1 regulates the transition from one program to the other is not well understood. Latent genomes reside in the nucleus of the host neuron as extra-chromosomal circles that are assembled into chromatin resembling that of the host [3], [4]. Transcription is limited to the latency-associated transcripts (LATs) that are spliced into a stable 1.5 to 2.0-kb intron and processed into several microRNAs [5], [6], [7]. The rest of the viral transcriptome corresponding to 80 or so genes is effectively silenced. Although the details are incomplete, it appears that lytic gene transcription is blocked by a combination of mechanisms involving histone modifications that create a repressive chromatin structure, recruitment of factors that prevent the assembly of pre-initiation complexes and the absence of activators need to stimulate RNA polymerase II recruitment and elongation [3], [4]. Chromatin immunoprecipitation (ChIP) studies have shown that the histones associated with the promoter regions of the key regulatory genes carry modifications such as trimethylated histone H3 lysine-9 (H3K9me3) and lysine-27 (H3K27me3) that are typical of facultative heterochromatin [8], [9]. Also present are polycomb (PcG) repressor subunits that may promote chromatin compaction and limit RNA polymerase elongation. Additional layers of repression are also imposed by the REST/coREST silencing complex that recruits the histone deacetylases HDAC1 or 2 and the demethylase LSD1 [10], [11], [12]. When latent episomes undergo reactivation, this heterochromatic signature is replaced with chromatin that is permissive for transcription, enabling the expression of the full repertoire of lytic genes needed to build new infectious virus and neutralize host defenses [13].
Despite intense study by many laboratories there is still much to be learned about the molecular processes that maintain the latent program and orchestrate the transition back into productive replication. Studies have been hampered by the very small numbers of infected neurons in ganglia obtained from human cadavers or animal models, our limited understanding of the signaling pathways that connect natural reactivation stimuli to the latent episome and difficulties in manipulating HSV-infected ganglia in order to test the roles of cellular factors implicated in the process. To circumvent these issues we have taken advantage of a robust cell culture model of HSV-1 latency that uses primary sympathetic neurons isolated from the superior cervical ganglia (SCG) of unborn rats [14]. Homogenous neuron cultures are maintained in the presence of nerve growth factor (NGF) and HSV-1 latency can be established with wild type virus by infection in the presence of acyclovir (ACV). Lytic gene products and infectious virus disappear within a few days but the viral genomic DNA (average of c.25 copies/neuron) remains and LAT RNA can be readily detected in the nuclei of >20% of the neurons by fluorescent in situ hybridization, signifying the presence of latent virus. At this stage, ACV can be removed and cultures maintained for up to five weeks without reactivation [15]. Latency is dependent on continuous NGF signaling and interruption of the TrkA/PI3-kinase/Akt branch of the pathway leads to reactivation [14].
This study focuses on the program of viral lytic gene transcription that initiates after NGF signaling is interrupted. Classic studies in non-neuronal cell types have shown that acute (lytic) HSV-1 replication follows an ordered gene expression cascade that can be divided into three temporal phases: immediate-early (IE, α), early (E, β) and late (L, γ) [16], [17]. During low multiplicity infections, the IE phase is initiated by the tegument protein VP16 (also termed α-TIF, UL48, vmw65), which collaborates with cellular factors to drive IE gene transcription [18], [19], [20]. The IE proteins then act as positive regulators of the E genes, leading to amplification of the HSV-1 DNA genome and transcription of the L genes, which encode structural components of the viral particle, including the tegument protein VP16. Whether reactivation follows the same transcription program is uncertain, not least because the context of the viral genome is radically different in the two situations. At the onset of reactivation, the viral DNA is incorporated into ordered chromatin consisting of regularly spaced nucleosomes, whereas at the beginning of acute infection cycle the DNA is essentially nucleosome-free. Additionally, regulatory proteins such as VP16 that incorporated into the virion tegument are almost certainly absent at the beginning of reactivation.
Using pharmacological inhibition of PI3-kinase to achieve synchronous reactivation, we find that viral lytic mRNAs accumulate in two discrete waves (dubbed Phase I and II) that differ in the need for viral protein synthesis, DNA replication and the initiator factor VP16. In Phase I, lytic transcripts of all three kinetic classes accumulate simultaneously. New viral protein synthesis is not required, consistent with a short-term reversal of episome silencing. Genome amplification begins in Phase II and is followed by infectious virus production. VP16 accumulates in the cytoplasm during Phase I but most interestingly, does not contribute to viral transcription until Phase II, when nuclear VP16 is first detected. During acute infections, VP16 assembles an activator complex on viral promoters in the nucleus through interactions with host factors Oct-1 and HCF-1 [21], [22], [23], and providing human Oct-1, which has a higher affinity for VP16 than its rodent counterparts, to the neurons increases viral transcript levels during Phase II, arguing for the assembly of an equivalent VP16-induced complex. We conclude that reactivation follows a unique program of gene expression initiated by a transient reversal of host-mediated gene silencing and suggest that regulated entry of VP16 into the neuronal nucleus represents a host-imposed barrier that must be overcome for successful viral reactivation to occur.
As a prelude to our analysis of reactivation in sympathetic neurons, we profiled viral mRNA accumulation during acute replication (Figure 1). Week-old neuron cultures were infected with recombinant HSV-1 GFP-Us11 at a multiplicity of infection of 3 PFU/neuron (MOI = 3). Under these conditions, the virus replicated productively and the majority of neurons expressed the GFP-Us11 fusion protein within the first day. RNA was collected at intervals between 0 and 12 h of infection and analyzed by quantitative reverse transcription PCR (qRT-PCR) using primers specific to lytic transcripts representing each kinetic class: IE (ICP27/UL54), E (DNA polymerase/UL30), leaky-late/γ1 (VP16/UL48) and true-late/γ2 (tegument phosphoprotein ICP1-2/UL36). The first transcripts were detected within 3 h and became increasingly abundant at later time points. Importantly, we found that the accumulation of these mRNAs followed the classic IE-E-L profile (Figure 1A).
To pinpoint the onset of viral genome amplification, total DNA was also collected at each time point and assayed by qPCR using primers complimentary to a region in the UL30 gene (Figure 1B). Aside from a small increase in DNA content between 0 and 3 h that likely reflects the binding and uptake of viral particles into the neurons, robust amplification of the viral genome began between 6 and 9 h post-infection. This correlates with the appearance of transcripts from UL36 (Figure 1A and 1D), which is defined as a replication-dependent (true-late or γ2) gene [24]. To confirm that the increase in genome content reflected virus-directed replication, the time course was repeated in the presence of phosphonoacetic acid (PAA), an inhibitor of the virus-encoded DNA polymerase (Figure 1B). As expected, there was no increase in DNA content in the presence of the inhibitor, even at 12 h. By definition, leaky-late (γ1) and true-late (γ2) genes are distinguished by the degree of sensitivity to PAA and other inhibitors of viral DNA synthesis. Accordingly, VP16 mRNA (γ1) showed only a limited decrease in the presence of PAA (Figure 1C), whereas the levels of UL36 mRNA (γ2) were significantly decreased (Figure 1D). Expression of GFP-Us11 also conforms to true-late kinetics and was sensitive to PAA (data not shown). Thus, productive HSV-1 infection in SCG-derived neurons follows a sequential program of gene expression that mirrors the classic lytic cascade defined in non-neuronal cells. In agreement with an earlier study using SCG neurons [25], we find that the duration of the lytic program is 2 to 3 times longer in primary neurons than in non-neuronal cells such as rat embryo fibroblasts (Figure S1).
Ex vivo studies of reactivation have relied on the analysis of latently infected ganglia that are surgically axotomized from the animal and then explanted into culture. Under these circumstances the neurons are severely stressed and it is likely that multiple reactivation triggers occur. Alternatively, reactivation can be induced by physiological stress prior to analysis (discussed in [26]). After explant, viral lytic mRNAs are synthesized very rapidly, making it difficult to establish the order of events and limiting the opportunities for manipulating the neuronal environment by non-pharmacological means. Mindful of these issues, we turned to the SCG neuron culture model as a tractable experimental system to study the first stages of reactivation. Treatment of latently-infected SCG cultures with the phosphatidyl inositol-3 kinase (PI3-K) inhibitor LY294002, is sufficient to activate latent HSV-1 episomes, resulting in the expression of a lytic reporter (GFP-Us11) in 15–20% of the neurons within 48 h [14]. This provides a single, defined reactivation stimulus that simplifies subsequent analyses.
Latently infected cultures were established over a one-week period in the presence of 100 µM acyclovir (ACV) to suppress lytic replication and then reactivated following the scheme outlined in Figure 2A. Once latency was established, fresh media containing LY294002 and lacking ACV was added to induce reactivation. After inducing reactivation, total RNA was collected at successive time points over a four-day period and analyzed by qRT-PCR using viral primer sets normalized to 18S rRNA (Figure 2B). A baseline for each lytic mRNA was established from samples collected when the inducer was first applied (abundance ranges from 94 copies/sample for UL30 to 347 copies/sample for VP16) and all other values are expressed relative to these low but measurable levels. The analysis revealed a protracted time course of lytic mRNA accumulation compared to the acute infection studies (see Figure 1A), with a strong but transient accumulation of IE, E and L viral mRNAs at 20 h, followed by a decline between 20 and 25 h and then a gradual accumulation from 48 to 96 h, the last time point. The analysis shown summarizes the data from three independent experiments prepared from separate neuron preparations made on different dates. Mock treatment with DMSO resulted in no reproducible change in baseline levels over a 72 h period (data not shown). In all of these experiments, and in others not shown here, we observed a similar protracted profile corresponding to a discernable spike in lytic mRNA accumulation at 15 to 20 h post-induction, a striking decrease at around 25 h and then a second, more gradual, rise at later times. This profile is clearly very different from the rapid and sequential accumulation of lytic mRNAs observed during acute infection of similar cultures. We will refer to these two discrete periods or waves of lytic mRNA accumulation as Phase I and Phase II.
A simple explanation for the biphasic profile is that a limited number of (primary) reactivation events produce infectious virus that subsequently spreads to neighboring cells and replicate further as (secondary) acute infections. We ruled out this out by including WAY-150138, a stable compound that prevents encapsidation of the DNA genome and blocks spread of infectious virus [27], [28]. Repeating the reactivation time course in presence of WAY-150138 produced a very similar, biphasic profile (Figure 2C), indicating that secondary infection was not responsible for the second peak of lytic mRNA accumulation.
It is known that brief exposure to the protein synthesis inhibitor cycloheximide (CHX) is sufficient to induce HSV-1 reactivation in both SCG and dorsal root ganglion neuron cultures [29]. To ask whether the synthesis of viral proteins is required for Phase I, we induced reactivation with LY249002 first and then maintained the cultures for 10 h to allow the earliest reactivation steps to proceed before CHX (10 µg/ml) was added. At this time, viral lytic transcripts remained at base line levels making it highly unlikely that any viral proteins had been synthesized. RNA was collected after a further 10 h in the presence of CHX (i.e. at 20 h post-induction with LY249002) and analyzed by qRT-PCR (Figure 2D). When compared to mock treated cultures, no significant changes were observed in the levels of lytic mRNAs, indicating that Phase I initiates normally in the absence of viral protein synthesis.
Next we sought to determine when viral genome amplification begins. DNA was collected at successive intervals and analyzed by qPCR (Figure 3A). Under these assay conditions we can reliably detect an increase in viral DNA content of 15% or greater (data not shown). No increase in viral DNA content was observed until the 48-h time point, which corresponds to Phase II and was not seen when the experiment was repeated in the presence of the PAA. The lack of demonstrable amplification during Phase I (15 and 20 h post induction) at least within the limits of the assay, was especially interesting given that we could readily detect representative late gene transcripts at 20 h. To examine this more closely, we compared the expression of UL36 mRNAs in the presence or absence of PAA at 20 h (Figure 3B). No significant difference was observed, showing that accumulation of UL36 mRNA in Phase I was replication-independent. By contrast, inclusion of PAA decreased the accumulation of UL36 mRNA during Phase II in a reproducible and significant manner (P-value<0.045).
We also performed plaque assays to detect infectious virus particles (Figure 3C). Acknowledging the sensitivity issues associated with this biological assay [27], we were unable to detect plaques until 48 h post LY294002 treatment, with a significant increase in plaque number from 72 h onwards. This suggests that infectious virus is not assembled until well into Phase II, concomitant with viral genome amplification and reinforces the notion that the transition from Phase I to Phase II does not reflect secondary infection.
VP16 is essential for HSV-1 replication, functioning in the assembly of the tegument layer and as a potent transactivator of the IE genes, which in turn drive the expression of the E genes [30], [31]. Recombinant viruses have been characterized that carry mutations in VP16 that preserve the essential tegument forming function but selectively impair the transactivation function, either by preventing the formation of the VP16-induced complex on DNA response elements found in each IE promoter [32] or by removing the C-terminal transactivation domain [33], [34]. Previous work has shown these mutants can still establish latency in animal models, and depending on the reactivation assay used, can display severe defects in reactivation [35], [36]. To examine the contribution of VP16 to LY294002-induced reactivation in the SCG system, we performed infection studies using in1814, a derivative of strain 17syn+ that contains a 12-bp insertion at codon 397 (Figure 4A) [32]. Neuron cultures were infected in the presence of ACV with equal titers (MOI = 1) of in1814 or a repaired version (in1814R) that behaves similar to wild type strain 17syn+ and allowed to establish latency. To obtain equivalent titers each virus stock was grown on complementing human U2OS cells and then plaque assayed on rat embryo fibroblasts in the presence of 5 mM hexamethylene bisacetamide (HMBA), a differentiation agent that overcomes the requirement for transactivation by VP16 during productive replication at low multiplicity [37]. The in1814 and in1814R viruses established latency at a similar frequency as judged by quantitation of viral genomic DNA after six days in the presence of ACV (Figure 4B). Slightly more genomes were detected with in1814, however, this difference was not statistically significant (P>0.29). The ability of in1814 to establish latency mirrors its behavior in several latency models [33], [35], [36], [38], [39]. Next, latently infected cultures were induced with LY294002 and the yield of infectious virus was determined by plaque assay on Vero cells in the presence of HMBA (Figure 4C). In spite of the similar numbers of genome at the time of induction, we observed a profound difference in the yield of infectious virus. Indeed, over multiple experiments we were unable to detect a single infectious particle for in1814. This indicates that loss of the transactivation function of VP16 has a highly deleterious effect on the ability of HSV-1 to reactivate successfully in cultured neurons following a specific stimulus. This agrees with previous studies using the murine hyperthermal stress model for reactivation [36].
To better understand the nature of the in1814 reactivation defect, we measured the relative levels of viral mRNAs produced by the mutant and rescued viruses in each reactivation phase (Figure 4D). For ease of comparison, values for in1814 are shown relative to the corresponding in1814R samples and the entire profile for mRNA accumulation from in1814 is shown in Figure S2A. When the two viruses are compared at 15 or 20 h post-induction (Phase I), no significant differences were observed, however at 72 h (Phase II) all five representative lytic mRNAs were expressed at approximately 5-fold lower levels by in1814 compared to in1814R (Figure 4D). Likewise, no amplification of the in1814 genome was detected (Figure S2B), consistent with the inability of the mutant virus to produce infectious progeny. Thus we conclude that the transactivation function of VP16 contributes to lytic gene transcription during the Phase II only, and that reduced expression of viral proteins results in a failure to replicate the genome. This would account for the absence of infectious progeny following reactivation.
An important caveat in using mutant viruses to study reactivation is the possibility that the mutations alter some aspect of the establishment of latency that subsequently impacts reactivation. To circumvent this concern and independently validate the striking phenotype of the in1814 virus, we developed a short-hairpin RNA (shRNA) capable of depleting newly synthesized VP16 from wild type virus. We have previously shown that latently infected SCG neurons can be successfully infected with lentiviruses without inducing reactivation and that shRNAs delivered in this manner are effective at depleting endogenous proteins [14]. The ability to deplete viral products in this system was previously untested. Neurons were latently infected with the HSV-1 GFP-Us11, allowed to establish latency and then secondarily infected with lentiviruses encoding either the VP16 shRNA or a non-silencing control. The vector includes a constitutive mCherry marker and visual inspection confirmed that the majority of neurons were infected (data not shown). After 5 days of dual infection, LY294002 was added to the media and reactivation allowed to proceed. VP16 depletion was confirmed by qRT-PCR to detect VP16 mRNA (Figure S2C) and by immunoblot of lysates prepared at 5 days post-induction (Figure 4E). VP16 is required for tegument assembly and accordingly, infectious virus was not detected in cultures expressing the VP16 shRNA, corroborating the knockdown (Figure 4F). More importantly, we compared the levels of lytic transcripts at 15 and 20 h (Phase I) and 48 h (Phase II). As before, we observed no significant difference in Phase I, but saw a 4 to 5-fold reduction in each of the lytic mRNAs in Phase II (Figure 4G). This result confirms the phenotype of the in1814 virus and underscores the importance of VP16-mediated transactivation during Phase II.
VP16 is recruited to the TAATGARAT sequences found in each HSV-1 IE promoter through physical association with HCF-1 and the POU domain transcription factor Oct-1 [30]. Selection of Oct-1 (also called POU2F1) is highly specific and the presence of other POU proteins may actually antagonize the VP16-induced complex by competing for the TAATGARAT sites. The Oct-1 POU DNA-binding domain is sufficient for VP16-induced complex formation, with VP16 recognizing a solvent exposed surface on the POU-homeo (POUH) subdomain [40]. Although the amino acid sequence of the POU domain has been conserved through vertebrate evolution (Figure 5A), there are four differences between the POUH sequences of human and either rat or mouse Oct-1 [41]. As a consequence, VP16 has a significantly lower affinity for rodent Oct-1 and reduced ability to drive transcription from reiterated TAATGARAT elements in murine 3T3 cells [42]. Whether this difference between rat and human Oct-1 has an equivalent effect on the activity of the IE promoters in the context of the reactivating virus is unknown. To address this, we used a lentiviral expression vector to introduce human Oct-1 into rat SCG neurons that were already latently infected with HSV-1 Us11-GFP. After the 5-day establishment period, separate cultures were transduced with either a lentivirus expressing wild type human Oct-1 or a control lentivirus expressing GFP. After a further 5 days, reactivation was induced with LY294002 and RNA samples prepared at regular intervals. Effective delivery of the Oct-1 lentivirus was confirmed by qRT-PCR analysis using human Oct-1 specific primers (Figure 5B). Amplification products were readily detected in the samples containing the Oct-1 expression lentivirus but not with the control virus and the levels remained essentially unchanged through the time course of the experiment. Analysis of viral lytic mRNAs performed in parallel revealed the expected biphasic profile in both sets (data not shown). During Phase I (15 to 20 h) no significant differences in transcript levels were observed between neurons expressing human Oct-1 or GFP, but in Phase II viral mRNA levels were increased of up to 5-fold in the presence of human Oct-1 (Figure 5C).
To determine whether this increase was simply a consequence of providing more Oct-1, we performed parallel infections with a human Oct-1 derivative in which the two non-conserved residues on helix-2 (corresponding to POUH sequence glutamic acid-30 and methonine-33, highlighted in Figure 5A) were swapped to the rodent equivalent (aspartic acid-30 and leucine-33). These two residues are reported to have the greatest impact on VP16 binding [42] and the deleterious effect of the combined mutation was confirmed by gel-shift assay using in vitro translated Oct-1, VP16 and HCF-1 proteins (Figure 5D, E). Both wild type (lane 4) and mutant human Oct-1 (lane 10) bind with similar efficiency to a 32P-labeled probe containing an (OCTA+)TAATGARAT sequence from the HSV-1 ICP0 promoter [23]. As expected, wild type Oct-1 supports VP16-induced complex (VIC) formation (lanes 5–9) but this is reduced by 100-fold or greater with the E30D/M33L mutation (lanes 11–15). Efficient lentiviral infection for the mutant was evident by a strong qRT-PCR signal using human Oct-1 primers (Figure 5B) although levels were slightly lower than with wild type. Importantly, we observed no significant change in viral lytic mRNA levels in the Oct-1 E30D/M33L transduced cultures, either in Phase I or in Phase II compared to the control lentivirus expressing GFP (Figure 5C), indicating that the stimulation observed by wild type human Oct-1 in Phase II is most likely due to improved interactions with VP16. We conclude from this experiment that the transactivation function provided by VP16 in Phase II requires Oct-1, presumably through the formation of a multi-protein complex on the TAATGARAT elements found in each IE promoter.
Although the transactivation function of VP16 is dispensable for Phase I, VP16 mRNA is readily detected (Figures 2–4). This prompted us to ask when the VP16 protein is synthesized. To do this we performed indirect immunofluorescence microscopy on latently infected neurons undergoing reactivation (Figure 6). Using a rabbit polyclonal antibody against VP16, we detected a strong signal in approximately 5% of the neurons at the onset of reactivation (0 h, see discussion), which rose to more than 20% of the culture at 15 and 20 h, which corresponds to Phase I (Figure 6A and 6B). The signal corresponds to VP16 because the antibody recognizes a single species of the correct size on immunoblots (Figure 4E), is absent from the remaining 80% or more neurons in the cultures (Figure 6A) and can be suppressed or eliminated by coinfection with the VP16 shRNA lentivirus (data not shown).
Strikingly, at 0 h and later time points the VP16 signal was preferentially localized to the soma (cell body) of the neurons rather than the nucleus (indicated by arrow heads and counterstained with DAPI), an unexpected distribution for a transcription factor. By contrast IE proteins ICP0 and ICP27 were detectable as speckles in the nucleus during Phase I showing that multiple viral antigens are expressed at this time and that nuclear uptake was not inhibited (Figure S3). The cytoplasmic VP16 signal persisted through the Phase I–II transition (25 h) and into Phase II (48 and 96 h), however most interestingly, a small but increasing number of neurons showed a predominantly nuclear VP16 signal at 25 and 48 h (Figures 6A and 6C). The nuclear signal appeared diffuse throughout the nucleoplasm and was excluded from the single large nucleolus characteristic of neurons. This distribution is consistent with a general association with chromatin, rather than formation of discrete viral replication compartments within the nucleus [43].
HCF-1 is nuclear in most cell types with the notable exception of sensory neurons, where it accumulates in the cytoplasm [44]. Ganglia explant and other stresses, trigger relocalization of HCF-1 into the nucleus, where it can associate with HSV-1 IE gene promoters [44], [45]. In light of the dynamic localization of VP16 in SCG neuron cultures, we also probed for HCF-1 using a polyclonal antibody directed against the HCF-1C subunit (Figure 7A). The immunofluorescence signal is widely dispersed throughout the cytoplasm including the axons and dendrites (upper panels). Nuclear sparing is clearly evident and the cytoplasmic HCF-1 did not appear to co-localize with major cytoplasmic organelles. When cultures are treated with LY294002 for 25 h, we saw significant accumulation of nuclear HCF-1 in some but not all of the neurons (Figure 7A, lower panels). Counting revealed a strong nuclear signal in more than 30% of the LY294002-treated neurons (n = 293) compared to less than 2% in mock-treated cultures (n = 1244) (Figure 7B). Control studies using antibodies to the HCF-associated proteins Ash2L [46] and LSD-1 [11] revealed predominantly nucleoplasmic staining with obvious nucleolar sparing, irrespective of treatment (Figure S4). Thus we can reproduce the differential and dynamic localization of HCF-1 first described in murine trigeminal ganglia, and show for the first time that HCF-1 localization can be manipulated by the inhibition of a cellular signal transduction pathway critical for the maintenance of HSV-1 latency.
Latent infection of primary SCG neuron cultures provides a powerful system to pick apart the mechanisms that control HSV-1 reactivation in response to defined physiological changes in the host neuron. Viral and cellular regulatory factors can be more readily manipulated in culture than in live-animal models and use of relatively pure populations of neurons allows us to set aside potentially confounding layers of control imposed by HSV-specific CD8+ T cells and other cell types resident in the ganglia. Thus analyses can be focused on the intimate and fundamental relationship between the latent virus and it's host neuron. Using this reductionist approach, we have profiled the transition from the latent to lytic transcription programs after treating latently infected cultures with a selective PI3-kinase inhibitor that interrupts NGF signaling [14], [29]. In contrast to acute infections, we find that lytic mRNAs accumulate in two discrete waves - termed Phase I and Phase II - that differ profoundly in their requirement for viral protein synthesis and DNA replication (see model in Figure 7C). Accumulation of viral transcripts during Phase I was unaffected by addition of the protein synthesis inhibitor CHX prior to the onset of viral mRNA accumulation, indicating that changes in pre-existing factors may be sufficient to bring about a generalized, but apparently temporary, de-repression of the lytic transcriptome. Studies published while this work was in revision also find that IE, E and L transcripts accumulate concurrently in the presence of CHX following explant-induced reactivation from the trigeminal ganglia of mice infected by corneal inoculation [47]. Our data does not provide evidence for the selective activation of key viral genes and suggests instead that there is activation of a broad spectrum of lytic genes. We acknowledge, however, that only a small fraction of the lytic transcriptome has been surveyed and that exceptions may exist.
While Phase I appears to be unique to the reactivation program, Phase II more closely resembles the ordered cascade seen during acute infection of both neuronal and non-neuronal cells. Lytic mRNA levels are enhanced by the transactivation function of the VP16-induced complex, true-late gene transcription is dependent on viral DNA replication and this phase culminates in the production of infectious virus. From the evidence obtained so far, we favor the notion that Phase I might serve as a ‘priming’ stage from which virus may or may not progress into a ‘production’ or ‘synthesis’ stage (Phase II). In this context, about a quarter of the neurons engage in Phase I as judged by VP16 and ICP27 fluorescence, but it seems likely that a much smaller number advance into Phase II as signified by either the presence of nuclear VP16 (this study) or the accumulation of true-late protein Us11 (Us11-GFP) in the presence of a compound that prevents secondary infection [14]. Further work is needed to establish the exact relationship between entry of VP16 into the nucleus and onset of Phase II-specific processes such as genome amplification and virion assembly.
Initiator protein VP16 is synthesized in Phase I but evidently does not contribute a necessary function until Phase II. Using two strategies we show that in the absence of VP16-mediated transactivation, the second phase of reactivation program proceeds at a reduced level and ultimately does not yield infectious virus. Thus we can conclude that VP16 is essential for reactivation in the SCG model following interruption of NGF signaling. Whether this is the result of insufficient transcription of the IE genes or loss of some less well understood function of VP16 remains to be addressed. A similar requirement for VP16 is observed in vivo during reactivation from murine trigeminal ganglia in response to thermal stress [36] but interestingly is not required for reactivation from the same ganglia after axotomy/explant, which is generally considered a more severe and multifaceted stimulus [35], [48]. This difference between assay approaches highlights the complex nature of the reactivation literature in which a number of different experimental systems have been used to explore the viral and cellular requirements for reactivation [49]. Although different conclusions have been drawn, the results are not necessarily incompatible.
While speculative, the idea of two-stage process for synthesizing and using VP16 (and possibly other viral regulatory factors) is appealing because it mirrors the delivery of tegument proteins including VP16 during acute infections [18], [20]. This strategy would perhaps allow HSV-1 to evaluate the capacity of the host neuron to support virion production, which is both metabolically demanding and involves extensive remodeling of the viral chromatin, before proceeding to point that might threaten the survival of the neuron or render the viral genome vulnerable to innate defenses. Reactivation events that do not meet the necessary criteria for completion of Phase II, for instance if there was insufficient nuclear accumulation HCF-1 and/or VP16, may even retain the option of re-establishing latency. This ability would buffer the virus from wide scale reactivation under moderate stress conditions and thereby help to maintain the latent reservoir. Although the overall amounts of viral mRNA were similar in samples collected during Phase I and Phase II this is unlikely to be a true representation of the transcriptional activity at each stage. Assuming that a smaller number of genomes progress into Phase II, the actual rates of transcription for any given reactivating episome will be much higher, consistent with the strong activator functions associated with VP16 and HCF-mediated remodeling of viral chromatin. In theory, this intriguing idea could be substantiated or refuted by following the progression of individual neurons through the reactivation program.
The transition from a repressed (silenced) chromatin state to one that allows high levels of transcription requires the concerted action of a variety of histone modifying enzymes and chromatin remodeling factors. An active area of study, there is already compelling evidence that HSV-1 uses a combination of viral and cellular factors to overcome silencing imposed by polycomb (PcG) repressor complexes [8], a REST/coREST/LSD1 repressor complex [1], [11] and possibly other mechanisms. ChIP studies of chromatin isolated from infected ganglia have established that core histones associated with the latent episome are enriched for H3K27me3, a repressive mark that is the signature of control by PcG proteins [8]. This and other repressive epigenetic marks are lost during reactivation and replaced with di/trimethylated histone H3 lysine-4 (H3K4me2/3), a mark of active transcription [50], [51]. Future work will determine whether silencing marks are removed at the onset of Phase I or are retained until Phase II, perhaps providing an epigenetic imprint that facilitates re-establishment of the latency program for genomes that do not progress into Phase II.
Exhaustive studies in mammalian cell lines and more heterologous systems such as yeast have shown that the individual components of the VP16-induced complex can recruit a variety of factors that promote active transcription [46], [50], [51], [52]. The well-studied acidic activation domain of VP16 is capable of recruiting or altering the activity of several general transcription factors, the Mediator complex, histone acetyltransferases and ATP-dependent chromatin remodeling factors as well as RNA polymerase II itself (reviewed in [53]). Whether the same processes and interactions apply during reactivation is not known. It is also worth noting that HCF-1 can bind to a variety of cellular transcription factors either directly or indirectly, and that some of these factors have binding sites in the HSV-1 IE and E gene promoters [54]. Moreover, HCF-1 is a component of several coactivators complexes, including the Set1 histone H3K4 methyltransferase complexes that is an active modifier of HSV-1 chromatin during acute infections [52], [55]. Thus it remains a possibility that under some circumstances, the necessary chromatin-modifying activities are recruited to viral chromatin through the association of HCF-1 and cellular DNA binding proteins rather than VP16. This might explain why the transactivation function of the VP16-induced complex is not required under all reactivation circumstances [35], [48].
As our understanding of reactivation advances it is helpful to draw a clear distinction between the overarching process of reactivation that ends with the release of infectious progeny and a more narrowly defined ability of the viral episome to transcribe lytic genes. Feldman and colleagues used ‘spontaneous molecular reactivation’ to describe events in which viral lytic genes are abundantly expressed but no infectious virus is detected [56]. Whether reactivation is curtailed by the action of adjacent immune cells within the ganglia or is intrinsic to the individual neurons to remains to be established. More recently, Penkert and Kalejta applied the succinct term ‘animation’ to describe the initial departure from the latent program prior to the onset of virion production [2]. This strikes us as a good descriptor for the Phase I stage detailed for the first time in this study because it allows for a flexible program in which the virus can either advance to full reactivation or re-establish latency if certain thresholds are not met.
Beyond the mechanistic details, partitioning of the reactivation program into two discrete phases has important implications for the control of HSV-1 latency by the host immune system. In the absence of an applied reactivation stimulus, we found that up to 5% of the neurons in our cultures were VP16 positive by immunofluorescence assay but otherwise the cultures lacked obvious signs of reactivation such as strong GFP-Us11 expression or detectable infectious particles. Although the levels of viral antigen expression seem high compared to previous in vivo studies, this is likely a reflection of the very high colonization rate (25–50% of neurons) in this in vitro system ([14], unpublished studies). In vivo, similar animation events will appear much less frequently because far fewer neurons are colonized however these rare events could still serve as the source of viral antigens detected by HSV-specific CD8+ T cells that infiltrate sensory ganglia in latently infected animals [57]. By manipulating the Phase I–II transition, it may be possible to propel a greater number of animation events into active replication phase rendering the virus vulnerable to both cell-mediated immunity and the action of current antiviral compounds. This would offer a new and worthy strategy to target the latent reservoir, a long-sought-after but elusive goal.
This study was carried out in strict accordance with the recommendations laid out by the NIH Guide for the Care and Use of Laboratory Animals. A detailed protocol for the isolation of rat primary neurons was approved by the Institutional Animal Care and Use Committee (IACUC # 101009) of the NYU Langone Medical Center (PHS Assurance of Compliance Number: A3435-01). Rats were euthanized by CO2 inhalation prior to ganglia harvest.
Primary neuron preparation, culture and infection was performed using our established procedures [14]. Briefly, superior cervical ganglia (SCG) were isolated from E21 Sprague Dawley® rat pups, placed in Leibovitz's L-15 medium supplemented with 2 mM L-glutamine, 0.4% D-glucose and incubated with 0.1% trypsin at 37°C for 30 min. Next, ganglia were washed in C-medium (minimum essential medium with Earle's salts, 0.4% D-glucose, 10% FBS, 2 mM L-glutamine) and dissociated by passing through 21G and 23G needles and a 70 µm nylon cell strainer. Dissociated cells were seeded in 24-well plates at a density of 4–5×104 cells per well with C-medium supplemented with NGF (50 ng/ml, Harlan Bioproducts). To improve adherence, wells were pre-coated with collagen (0.6 mg/ml, Millipore) and laminin (2 µg/ml, Sigma). The following day, media was changed to NBM (neural basal media, 0.4% D-glucose, 2 mM L-glutamine, 1x B-27) supplemented with NGF and with aphidicolin (5 µM, Calbiochem) and 5-fluorouracil (20 µM, Sigma) to remove proliferating cells.
For acute infections, cultures were maintained for 6 days as described above and then incubated with HSV-1 (Patton) GFP-Us11 (MOI = 3) for 2 h at 37°C. Latent infections were established by adding 100 µM acyclovir (ACV, Calbiochem) to the culture media on day 6, infecting the following day with HSV-1 (MOI = 1). After 2 h at 37°C, the media was replaced with NBM with 50 ng/ml NGF and 100 µM ACV and maintained for 7 days to allow latency to establish. Cultures were induced to reactivate with by replacing the media with fresh NBM containing the PI3-kinase inhibitor LY294002 (Calbiochem, 20 µM) and 50 ng/ml NGF but omitting ACV. When using HSV-1 GFP-Us11, individual wells were inspected prior to induction for GFP fluorescence that might be indicative of low level spontaneous reactivation, and discarded if positive. Typically no more than 10–20% of cultures undergo spontaneous reactivation after mock treatment, most likely due to the stress of handling.
Total RNA was prepared from neuron cultures using the RNeasy Mini kit (QIAgen) with some minor modifications. After washing with PBS, neurons were lysed in 350 µl RLT Buffer (QRNeasy Mini Kit, QIAgen) and homogenized using QIAshredder (QIAgen). The total volume was brought up to 900 µl with water and treated with 90 µg/ml Protease K for 10 minutes at 55°C. RNA was precipitated with 450 µl of 100% ethanol and applied to columns provided by the kit. Columns were washed as recommended and eluted with RNase-free water. To eliminate contaminating DNA, RNA was treated with DNase I (New England Biolabs) for 30 min at 37°C and the DNase I heat inactivated at 75°C for 10 min in the presence of 2 mM EDTA. RNA concentrations were determined using a NanoDrop spectrophotometer (ThermoScientific) and 300 ng/sample were used to generate cDNA with Superscript III (Invitrogen) and random hexamer primers (Fermentas). Levels of selected viral lytic mRNAs were quantified by qRT-PCR using the following primer sets:
ICP27 FW: 5′-TTTCTCCAGTGCTACCTGAAGG-3′
ICP27 RV: 5′-TCAACTCGCAGACACGACTCG-3′
UL5 FW: 5′-ACGTCGAGCTGTTGTTCGTCCA-3′
UL5 RV: 5′-GGCGAGCGTGCGTTTGATTT-3′
UL30 FW: 5′-CGCGCTTGGCGGGTATTAACAT-3′
UL30 RV: 5′-TGGGTGTCCGGCAGAATAAAGC-3′
VP16 FW: 5′-TCGGCGTGGAAGAAACGAGAGA-3′
VP16 RV: 5′-CGAACGCACCCAAATCGACA-3′
UL36 FW: 5′-CGCTGCACGAATAGCATGGAATC-3′
UL36 RV: 5′-CCAGCTCCCCGGAACACATTTA-3′
Primers were designed against the HSV-1 17syn+ strain reference sequence (GenBank NC_001806) using Primer 3 design software [58]. Input RNA was normalized using 18S rRNA primers (SA Biosciences) and the following equation: dCt (threshold cycle) = Target gene Ct- 18s rRNA Ct. Real-time qPCR analysis was performed using FastStart Universal SYBR Green Master-ROX (Roche) and a MyiQ™ single-color real-time thermal cycler (BioRad). Relative changes in transcript levels were calculated using the ΔΔCt method. Data plots and statistical calculations were made using Prism 5.0 software (GraphPad).
Neurons were washed in PBS, lysed with DNA extraction buffer (150 mM NaCl, 10 mM Tris pH 8, 10 mM EDTA, 10% SDS) and transferred to microcentrifuge tubes. Lysates were treated with 100 µg/ml Protease K overnight at 55°C, followed by two rounds of phenol extraction, one round of chloroform extraction and ethanol precipitation. DNA was re-suspended in water and analyzed by qPCR using primers to HSV-1 UL30 and normalized to cell number by amplification of rat RPL19 (RPL19 FW 5′-ATGTATCACAGCCTGTACCTG-3′ and RPL19 RV 5′-TTCTTGGTCTCTTCCTCCTTG-3′). To selectively inhibit lytic DNA replication, 0.3 mg/ml phosphonoacetic acid (PAA, Sigma) was added to the media 1 h before acute infection or with LY294002 for reactivation.
Viral stocks were amplified on Vero (GFP-Us11) or U2OS (in1814, in1814R) cells. Semi-confluent monolayers were infected at MOI = 0.01 and maintained at 34°C for 2–3 days before harvest by freeze-thaw lysis and sonication. Infectious titers were determined by plaque assay using either Vero cells or rat embryo fibroblasts. Uninfected cells were seeded at a density of 5×105/well in 6-well plates and incubated with lysate for 2 h at 37°C before being overlaid with 0.5% agarose in MEM supplemented with 1% calf serum. After incubation for 3 days at 37°C, cells were fixed with 10% TCA for 15 min and stained with 1% crystal violet.
Neurons were seeded at density of 4–5×104 onto coverslips pre-coated with poly-D lysine, collagen and laminin and later fixed using 4% paraformaldehyde/20% sucrose in PBS for 15 min, quenched with 100 mM ammonium chloride for 5 min and permeabilized with 0.1% Triton X-100 in PBS. After blocking in 1% BSA in PBS for 15 min, coverslips were incubated with polyclonal antibodies against VP16 (α-VP16 pep153, diluted 1∶1,000, a kind gift of Michael Gregory Peterson, Tularik Inc.) for 2 h at room temperature or against HCF-1 (α-rHCF H12 [59], diluted 1∶1,000) and detected using Alexa555-conjugated α-rabbit IgG antibody (diluted 1∶1,000, 1 h at 37°C). Nuclei were stained with 1 µg/ml 4′, 6-diamidino-2-phenylindole (DAPI) and the coverslips were mounted with DakoCytomation fluorescent mounting medium (Dako Corp.). Samples were visualized using a Zeiss LSM 510 META laser scanning confocal microscope. Images were captured with the Zeiss AIM software and exported into Adobe Photoshop 7.0 for cropping and minor adjustments.
Cells were washed in PBS and then lysed in medium salt extraction buffer (250 mM KCL, 20 mM Tris-HCl pH 7.9, 10% glycerol, 0.25% NP40) on ice for 1 h. The supernatant was collected after centrifugation for 10 min at 11,000 rpm at 4°C and fractionated by 10% SDS-PAGE. Separated proteins were transferred onto nitrocellulose membranes and blocked with 5% milk in TBS-T (200 mM Tris, 1.4 M NaCl, 1% Tween) for 30 min at room temperature prior to incubation overnight at 4°C with α-VP16 (diluted 1∶1,000) or α-Rho-GDI (Santa Cruz, diluted 1∶5,000), followed by HRP-conjugated α-rabbit IgG antibody (Roche, diluted 1∶2,500).
Depletion of HSV-encoded VP16 was achieved using shRNA delivered into latently infected neurons by lentivirus (pLVTHM-VP16shRNA). An oligonucleotide duplex 5′-cgcgtccccGAGTGTAAATTCCTATCAATtcaagaGATTGATAGGAATTTACACTCtttttggatccat-3′ containing the hairpin sequence (upper case) was inserted between the unique Mlu I and Cla I sites of a pLVTHM derivative that has been modified to constitutively express mCherry [14]. A human Oct-1 (GenBank NM_002697) cDNA was expressed from lentiviral vector pEZ-Lv105 vector (GeneCopoeia). Altered sense mutations (E30D/M33L) were introduced by successive rounds of QuikChange II XL Site-Directed Mutagenesis (Stratagene) and verified by DNA sequencing. Lentiviral stocks were generated by transfection of HEK 293LTV cells (Cell Biolabs) using calcium phosphate co-precipitation to introduce a mixture of the lentiviral vector plasmid and two packaging plasmids (pΔ8.9 and pMD2.G) in an equal ratio. Precipitates were left on the cells overnight and then washed. A control virus encoding GFP was prepared in the same manner. Supernatants containing lentivirus were collected on days 3 and 4. To infect neuron cultures, 500 µl of lentiviral-containing supernatant was added to each well, incubated overnight and replaced with fresh media the following day. For latently infected cultures, lentivirus was added 5 days after HSV-1 infection in the presence of ACV and reactivation with LY294002 was performed 5 days after lentivirus infection.
Recombinant human Oct-1, VP16ΔC and HCF-1N380 were synthesized by in vitro translation using the TNT quick-coupled transcription/translation system (Promega) in the presence of Easytag L-[35S]-methionine (Perkin Elmer) as previously described [60], [61]. The E30D/M33L mutation was introduced into wild type human Oct-1 by site-directed mutagenesis. A radiolabeled double-stranded DNA probe was prepared by PCR amplification using 32P-labeled primers to amplify a subcloned ICP0 (OCTA+)TAATGARAT sequence [23]. Binding reactions containing labeled probe, 1 µg poly-dI-dC, 10 mM Hepes pH 7.9, 75 mM KCl, 1 mM EDTA, 10 mM DTT and up to 4 µl reticulocyte lysate were incubated at room temperature for 20 min and loaded on a 4% native polyacrylamide gel with 1x Tris-glycine-EDTA buffer. Electrophoresis was carried out at room temperature and run at a constant 170 V. Gels were fixed in methanol/acetic acid, dried and the probe visualized by autoradiography using an X-ray film to block the 35S signal.
The Genbank (http://www.ncbi.nlm.nih.gov/Genbank/) accession number for the HSV-1 17syn+ strain reference sequence is NC_001806. Database accession numbers for proteins analyzed in this study are: rat HCF-1 (NM_001139507.1); human Oct-1 (NP_002688); rat Oct-1 (NP_001094109), and VP16 (Swiss-Prot: P04486.2).
|
10.1371/journal.ppat.1006083 | Role of Occult and Post-acute Phase Replication in Protective Immunity Induced with a Novel Live Attenuated SIV Vaccine | In order to evaluate the role of persisting virus replication during occult phase immunisation in the live attenuated SIV vaccine model, a novel SIVmac239Δnef variant (SIVrtTA) genetically engineered to replicate in the presence of doxycycline was evaluated for its ability to protect against wild-type SIVmac239. Indian rhesus macaques were vaccinated either with SIVrtTA or with SIVmac239Δnef. Doxycycline was withdrawn from 4 of 8 SIVrtTA vaccinates before challenge with wild-type virus. Unvaccinated challenge controls exhibited ~107 peak plasma viral RNA copies/ml persisting beyond the acute phase. Six vaccinates, four SIVmac239Δnef and two SIVrtTA vaccinates exhibited complete protection, defined by lack of wild-type viraemia post-challenge and virus-specific PCR analysis of tissues recovered post-mortem, whereas six SIVrtTA vaccinates were protected from high levels of viraemia. Critically, the complete protection in two SIVrtTA vaccinates was associated with enhanced SIVrtTA replication in the immediate post-acute vaccination period but was independent of doxycycline status at the time of challenge. Mutations were identified in the LTR promoter region and rtTA gene that do not affect doxycycline-control but were associated with enhanced post-acute phase replication in protected vaccinates. High frequencies of total circulating CD8+T effector memory cells and a higher total frequency of SIV-specific CD8+ mono and polyfunctional T cells on the day of wild-type challenge were associated with complete protection but these parameters were not predictive of outcome when assessed 130 days after challenge. Moreover, challenge virus-specific Nef CD8+ polyfunctional T cell responses and antigen were detected in tissues post mortem in completely-protected macaques indicating post-challenge control of infection. Within the parameters of the study design, on-going occult-phase replication may not be absolutely required for protective immunity.
| Development of an HIV vaccine remains a global health priority. In non-human primates live-attenuated SIV induces a potent vaccine effect. Following disappearance of vaccine virus from the peripheral circulation replication persists in lymphoid tissue. To address whether this occult replication is critical to the generation of protective immunity we used a novel construct (SIVrtTA) based on the prototypic live attenuated SIVmac239Δnef but which requires the presence of the antibiotic doxycycline to replicate. Protection appeared independent of doxycycline status at the time of virulent virus challenge suggesting that occult replication may not be absolutely necessary for persistence of immunity; however, stronger protection was observed in monkeys vaccinated with SIVrtTA where vaccine replication persisted for longer after peak viraemia. Moreover, some evidence of very low level breakthrough of vaccine virus replication was seen and protection was weaker than that obtained with SIVmac239Δnef. Both vaccination and challenge perturbed circulating T cell populations, but only the frequency of SIV-specific CD8+ polyfunctional T cells measured on the day of challenge was associated with protection. Replication-conditional mutants such as SIVrtTA have great potential in unlocking the complex interactions between the vaccine virus and host responses in the generation of potent anti-viral protection in vivo.
| Live attenuated SIV has proven to be a highly effective vaccination strategy in non-human primate (NHP) models of HIV/AIDS [1,2], in many cases protecting macaques from detectable superinfection following re-challenge with both homologous and heterologous wild-type SIV administered systemically and mucosally [3–23]. Although safety concerns such as reversion to virulence and recombination with wild-type strains preclude direct application of this vaccine approach in humans, a clearer understanding of mechanisms of pathogenesis and protection may inform the development of more clinically acceptable HIV vaccines. Studies have been performed using vaccine viruses attenuated by genetic disruption of key regulatory genes including nef, vpx, vpr and vif; although the moderately attenuated prototypic vaccine strain SIVmac239Δnef has been used for the majority of studies.
Attempts to establish clearly defined immune correlates of protection have not been conclusive, particularly where studies have measured responses in peripheral blood. Indeed, the only robust correlate identified so far is the observation between increasing attenuation of the vaccine virus and decreasing protection [11]. Recently, a detailed comparative study of different attenuated virus strains derived from SIVmac239 concluded that protection was associated with the induction of an effector memory T cell (TEM) response and protection of the T follicular helper (TFH) cell subset in lymphoid tissue [10]. This association, however, is not definitively established as the mechanism of protection.
A crucial property of minimally-attenuated SIV vaccines, which are the most effective, is the widespread distribution of the vaccine virus in multiple lymphoid tissues [22] but the role of occult replication (i.e. replication in lymphoid tissue when virus is no longer or only intermittently detected in the peripheral circulation) in the generation of protective immunity is not fully understood. Vaccine virus persistence may result in multiple alterations in the host innate immune system that contribute to protection, in addition to the induction of adaptive immune responses [22]. In the study reported here we have sought to influence occult phase persisting turnover of live attenuated SIV using a novel approach: the conditionally live attenuated SIVmac239Δnef vaccine (SIVrtTA) that in vitro is absolutely dependent on the presence of doxycycline (dox) to replicate [24, 25]. Previously, we have shown that SIVrtTA is infectious in Indian rhesus macaques and induced reversible up-regulation of the frequency of global circulating TEM [26].
Here, we report the outcome of an intravenous challenge of two groups of SIVrtTA-vaccinated macaques with wild-type SIVmac239 in comparison with macaques vaccinated with the prototypic SIVmac239Δnef live attenuated vaccine. One group of SIVrtTA vaccinates macaques remained on daily administration of dox, whereas another group received the final dose of dox 8 weeks prior to wild-type virus challenge during the occult phase of virus replication. Protection against detectable infection with wild-type, highly virulent SIVmac239 was observed at various levels; however, the pattern of protection did not associate directly with the experimental treatment protocol, but with the kinetics of vaccine-virus replication in the acute and immediate post-acute period of vaccine viraemia and with vaccine-driven T cell immune responses.
Two groups (A & B) of four Indian-derived rhesus macaques were injected intravenously with 5 x 103 TCID50 SIVrtTA vaccine (genetically engineered from the SIVmac239 backbone as indicated in Fig 1A) and treated with dox for 6 months followed by a period of 8 weeks without dox (Group A; E61, E63, E65, E66) or treated with dox for 6 months and then maintained on dox(Group B; E67, E68, E70, E71). A further 4 macaques (Group C; E73, E75, E76, E77) were vaccinated with SIVmac239Δnef for 6 months and four unvaccinated, naïve macaques (E79-E82) were included as challenge controls (Fig 1B). Total SIV gag vRNA profiles are shown for Groups A-C as a continuum of vaccination and wild-type challenge profiles (Fig 1C). As previously reported [26], the SIVrtTA vaccinates displayed a transient peak in plasma vRNA kinetics which is characteristic for attenuated SIVmac239Δnef with two exceptions: E65 (Group A) and E70 (Group B). These animals exhibited a persisting shoulder of ~ 102 vRNA copies/ml to ~100 days post-vaccination. From day 110 post-infection (p.i), prior to removal of dox, E65 plasma vRNA fell below the limit of detection. Plasma vRNA remained stably elevated in E70, which was maintained on daily dox treatment to the time of wild-type challenge. Another macaque, vaccinated with SIVmac239Δnef (E76, Group C) also failed to completely control viraemia below the limit of detection. Hence, at the time of wild-type SIVmac239 challenge detectable vRNA signals were present in the plasma of vaccinates E65, E70 (SIVrtTA) and E76 (SIVmac239Δnef) (Fig 1, S1 Fig).
Challenge outcome was initially assessed by the individual comparison of total SIV gag plasma vRNA profiles for each group (Fig 1C). As expected, all four naive, unvaccinated control macaques challenged with wild-type SIVmac239 (Group D) exhibited high plasma vRNA loads (1.86 x107 mean SIV RNA copies/ml) by day 14 p.i. which also exhibited a high vRNA steady state (106–108 wtSIVmac239 copies/ml) throughout the 20 week follow-up period. In contrast, a marked vaccine effect was observed in all animals vaccinated with either SIVrtTA or SIVmac239Δnef (Fig 1C, Groups A and B or C respectively).
Statistical analyses of suppression of vRNA levels post wild-type challenge were determined using a Kruskal-Wallis analysis with a Dunn’s post-hoc test to determine significance levels. Peak viraemia was statistically significantly suppressed when Groups A, B and C were compared individually to Group D, although the vaccination with SIVmac239Δnef resulted in the most significant outcomes. P values for Groups A-C were, p = 0.05, p = 0.05 and p = 0.001 respectively at day 14. When viraemia levels were analysed at day 84 (steady-state) significance was retained in Groups A and C (p = 0.012 and p = 0.002 respectively) although, interestingly, significance was lost at day 84 in Group B (p = 0.135). All SIVrtTA vaccinates analysed together (A and B combined) exhibited significant differences from Group D challenge controls at peak (day 14) and steady-state (day 84) time-points (p = 0.025 and 0.021 respectively).
E73, E76 and E77 (vaccinated with SIVmac239Δnef) and E65 (vaccinated with SIVrtTA) exhibited plasma vRNA levels that remained <100 SIV RNA copies/ml. Macaques E76 (vaccinated with SIVmac239Δnef) and E70 (vaccinated with SIVrtTA) exhibited plasma viral loads of the vaccine virus higher than 100 SIV RNA copies/ml prior to wild-type SIV challenge, with vRNA viral loads gradually increasing in the 20 week follow-up period. Of the remaining macaques vaccinated with SIVrtTA, with undetectable plasma vRNA on the day of challenge (E61, E63, E66, E67, E68, E71), a significant peak in plasma viremia was detected 14 days after wild-type challenge (mean 5.43 x 105 SIV RNA copies per ml), which was partially resolved, but remained between 1 x 103–1 x 106 SIV RNA copies/ml 20 weeks post-challenge.
From these initial analyses it was possible to classify macaques into two levels of protection: (1) complete-protection defined as having no secondary peak after wild-type virus challenge (E65, E70: SIVrtTA; E73, E75, E76, E77: SIVmac239Δnef). (2) partially-protected macaques that exhibited a clear secondary peak of viraemia 14 days post SIVmac239 wild-type challenge (E61, E63, E66, E67, E68, E71; SIVrtTA). When these data were re-plotted, also as a continuum, two patterns of plasma vRNA profiles were revealed immediately prior to and post wild-type SIVmac239 challenge, reflecting these two general classifications of protection as represented in Fig 2. Interestingly, in the partially protected group, the secondary spike in vRNA is immediately preceded by a virtual absence in detectable vaccine-virus replication prior to wild-type challenge. By comparison, in the completely protected group, total SIV gag vRNA signals are clearly evident in the same period (~100 days) up to challenge with little perturbations in these levels post-challenge. However, as total plasma SIV RNA levels reveal only part of the overall biomarker of infection picture, discriminatory PCR assays were required to fully evaluate the protection status of each macaque.
To discriminate further between superinfection with wild-type virus and recrudescence/persistence of vaccine virus, discriminatory PCR assays were established that selectively detect either vaccine-derived or wild-type vRNA in plasma, total vDNA signals in tissues or cell-associated viral RNA (CA-RNA). From these combined analyses a clear picture of superinfection status emerged with the ability to detect and quantify each viral nucleic acid species in blood and/or selected lymphoid tissues (Figs 2 and 3; S2 Fig). Wild-type SIVmac239-specific vRNA determinations partitioned macaques into completely protected (E65, E70, E73, E75, E76, E77) or partially protected (E61, E63, E66, E67, E68, E71) as indicated in Fig 3A. There was a highly statistically significant difference between completely protected macaques and naïve challenge controls (p<0.001 at days 14 and 84 post SIVmac239 wild-type challenge) using a Kruskal-Wallis analysis with a Dunn’s post-hoc test. Partially-protected macaques all demonstrated a spike in plasma vRNA that was unambiguously attributed to establishment of wild-type virus infection which at days 14 and 84 were non-significant (p = 0.095) compared to challenge controls, applying the same statistical test as for the completely protected group. Additionally, a broad range of lymphoid tissues was assessed for wild type SIV DNA (S2 Fig). High levels were detected in all tissues in naïve challenge controls, with lower but detectable levels in most tissues in E61, E63, E66, E67, E68 and E71, reflecting profiles of plasma viral RNA. No wild type SIV DNA was detected in any tissue from the completely-protected animal vaccinated with SIVrtTA (E65) and a single signal of wild-type SIVmac239 DNA detected in the spleen of E70. No wtSIVmac239-specific DNA was detected in animals of Group C vaccinated with SIVmac239Δnef.
Additional information relating to the ability to detect apparently replication-competent virus, rather than proviral signals, was gained for a number of tissues by measuring CA-RNA concentrations for vaccine and wild-type viruses 20 weeks after SIVmac239 challenge (Fig 3B). Wild-type SIV was never detected by any molecular biomarker of infection in those macaques vaccinated with SIVmac239Δnef (E73, E75, E76, E77), further confirming the complete protection status of this group. No wtSIVmac239 CA-RNA was detected in macaques E65 and E70 vaccinated with SIVrtTA, compared with high levels of wtSIVmac239 detected in all naïve challenge controls, particularly in the spleen and mesenteric lymph nodes (MLN). Lower levels of wtSIVmac239 CA-RNA were detected in spleen samples from E61, E63, E66, E67, E68 and E71 and more sporadically from MLN and peripheral lymph node (PLN) samples (Fig 3B). Hence, with information gained from virus-specific differential PCR techniques, taking only wild-type SIVmac239 levels as measures of outcome, there was a statistically significant difference in outcome between completely protected macaques and wild-type challenge controls and partially protected and wild-type challenge controls.
Although E65 and E70 displayed undetectable signals for wild-type specific plasma and CA-RNA, both vaccinates signalled positive by SIVrtTA-specific RT-PCR, particularly E70 in the plasma, spleen and PLN (Fig 3B and 3C). These data reflect the plasma vRNA signal in E70 post-wtSIV239 challenge which was unambiguously attributable to continuous SIVrtTA replication in the continued presence of dox. Remarkably, SIVrtTA replication did not fluctuate over time in this macaque, nor was it perturbed by administration of the wild-type challenge virus (Figs 1C, 2 and 3). In this respect, E70 was comparable to macaque E76 (Group C; SIVmac239Δnef) that displayed similar continuous viral kinetics post SIVmac239 challenge despite resistance to wild-type superinfection as confirmed by lack of wild-type SIVmac239 RNA signals in either plasma or tissues.
Perhaps the most interesting vaccinate of all groups was SIVrtTA-vaccinated macaque E65, which resisted wtSIVmac239 but displayed highly controlled vRNA kinetics in the later post-acute phase. However, in the absence of dox, four blips of plasma vRNA were noted as determined by total SIV-gag qPCR (Figs 1C and 2), two prior to wild-type challenge but after dox removal and two blips after wild-type challenge. Analysis of tissues for CA-RNA indicated low, but clearly detectable SIVrtTA in the PLN at termination. Taken together, these data suggest evidence of very low, but persistent replication of SIVrtTA in E65 when there was no, or little, dox present. Moreover, both SIVrtTA vaccinates E65 and E70 had detectable levels of SIVrtTA-specific CA-RNA at termination, many weeks after initial vaccine administration. Extending these observations to Group C vaccinates (SIVmac239Δnef) all had some level of residual detectable vaccine virus replication at termination (Fig 3). Indeed, all 6 completely protected vaccinates signalled positive for the vaccine virus post-mortem in PLN suggesting this to be an important site for virus sequestration, which as well as the spleen represents an important reservoir for the vaccine virus.
All vaccinates seroconverted to SIV Gag p27 prior to challenge with wild-type SIV (S3 Fig). Anti-p27 responses were broadly similar amongst all macaques vaccinated with SIVrtTA regardless of dox withdrawal and anti-p27 titres were lower than those in SIVmac239Δnef vaccinates. All fully protected animals, E65 and E70 vaccinated with SIVrtTA and E73, E75, E76 and E77 vaccinated with SIVmac239Δnef, showed only minor perturbations in anti-p27 titre after challenge with wild-type SIVmac239, whereas a marked increase in anti-p27 titres was detected in all other macaques (S3 Fig).
In order to address the possibility that mutations arising in SIVrtTA as a result of selection in vivo may have occurred, SIVrtTA RNA recovered from vaccinates was sequenced. For this, plasma vRNA was isolated at several times during the immediate post-acute phase period, when qRT-PCR revealed a vRNA load of >102 SIV RNA copies/ml including where there was the persisting shoulder of prolonged SIVrtTA replication in E65 and E70 SIVrtTA vaccinates.
In SIVrtTA, the Tat-TAR transcription activation mechanism has been functionally replaced by the dox-inducible Tet-On gene expression system [24, 25, 27]. To achieve this (1) TAR was inactivated through mutations in the binding sites for Tat and pTEFb, (2) the gene encoding the dox-inducible rtTA transcriptional activator was inserted at the site of the accessory nef gene and (3) tet operator (tetO) elements to which the dox-rtTA complex can bind were inserted between the NFκB and Sp1 binding sites in the U3 domain of the LTR promoter (Fig 1A). Sequencing of the LTR and leader RNA region of different SIVrtTA RNA samples demonstrated the stable presence of the TAR-inactivating mutations and no additional changes were observed in TAR. The virus also stably maintained the tetO elements but whereas the vaccine strain contained a triplicated NFκB-tetO repeat (resulting from in vitro evolution; [28]), deletion of one of these repeats was frequently observed (S1 Table; S4 Fig). Previous experiments demonstrated that such a deletion slightly reduces the transcriptional activity of the LTR promoter, but does not affect dox-control. In all macaques, a point mutation was observed in the primer binding site (PBS) sequence (T731C). This mutation was due to the fact that the SIVrtTA vaccine construct contained a PBS complementary to the infrequently used tRNAlys5 primer for reverse transcription [29, 30]. As expected, the in vivo replicating virus demonstrated a PBS sequence corresponding to the more frequently used tRNAlys3 primer. Sequencing of the tat gene did not reveal any sequence changes.
However, sequence analysis of the rtTA gene identified two non-silent codon changes (R80W and E191K) in the E65 samples isolated at 6 weeks after vaccination and later (S1 Table). The E70 sample isolated at 6 weeks after vaccination demonstrated an R80Q change, whereas later E70 samples (from 14 weeks) also demonstrated the R80W substitution. We did not identify such rtTA mutations in the other macaques vaccinated with SIVrtTA. The identified amino acid changes had never been observed previously in multiple long-term in vitro evolution experiments with SIVrtTA or with a similar dox-controlled HIVrtTA variant and hence represents a novel finding.
Testing the transcriptional activity of the new R80W and E191K rtTA variants demonstrated that the mutations did not increase the background activity in the absence of dox (no loss of dox control) nor significantly alter the dox-induced activity (S5 Fig). As both E65 and E70 showed prolonged SIVrtTA replication, the mutations may improve in vivo replication of the virus. Importantly, these results demonstrate that the in vivo replicating virus stably maintains the integrated dox-control mechanism and did not restore the Tat-TAR axis of transcription control.
Since TRIM5α status and MHC type may influence vaccine challenge outcome [31, 32], the TRIM5α/TRIMcyp status and MHC type of all study macaques was determined (S2 Table). While no direct associations were identified between either MHC or TRIM5/cyp status and outcome it is interesting to note that the two macaques which failed to control the vaccine virus (E70: SIVrtTA; E76: SIVmac239Δnef) and were protected from wild-type SIVmac239 did not express any of the major mamu A alleles analysed (S1 Fig). In this study, we could not identify any confounding factors associated with TRIM5α or TRIMcyp genotype.
We have previously reported that under replication permissive conditions, during the period when live attenuated virus RNA was essentially below the limit of detection in plasma, the global circulating T effector memory (TEM) cell frequency was upregulated [26]. Hence, we were interested to determine if this effect was associated with the degree of protection from superinfection. Comparison of partially and completely protected macaques on the day of challenge revealed that for both CD4+ and CD8+ CD95+ T cells, completely protected macaques had a lower median frequency of TCM and reciprocally a higher median frequency of CD28- CCR7- TEM than partially protected macaques; however, the difference in TEM frequencies between these groups only reached significance in CD8+ T cells (ρ = 0.026; Mann-Whitney rank sum test) (Fig 4). Comparison with results from naïve macaques showed that median frequencies of both CD4+ and CD8+ TCM were significantly reduced in completely protected macaques (ρ = 0.003 and ρ = 0.008 respectively; Mann-Whitney rank sum test). Conversely, the frequencies of CD4+ and CD8+ TEM (CD28- CCR7-) were significantly elevated in completely protected macaques compared with naïve macaques (ρ = 0.002 and ρ = 0.007; Mann-Whitney rank sum test). Despite these differences at the population level exceptions were seen: T cell frequencies for macaque E65, challenged under conditions of dox withdrawal were similar to those for naïve or partially protected macaques. Conversely, partially protected macaque E67 challenged under dox maintenance had a high frequency of CD28- CCR7- CD8+ TEM and partially protected macaque E71 also challenged under replication permissive conditions had relatively high frequencies of both CD28- and CD28+, CCR7- CD95+ CD8+ T cells. So, although there was an association between a high frequency of global CD8+ TEM in the circulation on the day of challenge and complete protection, a high TEM frequency alone was not predictive of complete protection status.
The phenotype of circulating T cells was examined again at day 130 following superinfection challenge. At this time-point no significant differences were found between partially and completely protected macaques; moreover, global circulating CD4+ memory T cell populations were significantly perturbed (Fig 4). CD4+ TCM were significantly elevated in superinfection-challenged animals, regardless of protection status compared with frequencies in naïve macaques (ρ = <0.001; Mann-Whitney rank sum test). Likewise, comparison of TCM frequencies on the day of challenge with day 130 post-challenge showed significantly elevated frequencies regardless of protection status (ρ = <0.031; Wilcoxon signed rank test). The median frequencies of CD4+ CD28+CCR7- (intermediate) T cells remained elevated post challenge compared with naïve animals (ρ = 0.003 and ρ = <0.001 for partially and completely protected groups respectively; Mann-Whitney rank sum test) and showed no statistical difference between day of challenge and day 130 post challenge for either group. In contrast, the median frequencies of TEM were significantly reduced 130 days after challenge compared with those in naïve animals; although this was most marked in completely protected animals (ρ = 0.049 and ρ = 0.006, partially and complete protection groups respectively; Mann-Whitney rank sum test). Similarly, pairwise comparison of completely protected animals revealed a significant reduction in TEM cell proportions between the day of challenge and 130 days post challenge (ρ = 0.031; Wilcoxon signed rank test). Five of 6 partially protected macaques also had lower frequencies at day 130 post challenge (ρ = 0.063; Wilcoxon signed rank test).
A somewhat different pattern of perturbation in circulating CD8+ T memory cell populations was seen following superinfection challenge (Fig 4). These changes were again, as for CD4+ T cells, independent of superinfection status. Frequencies of TCM and TEM (CD28-CCR7-) were not significantly different from frequencies in naïve animals; whereas CD28+CCR7- cell frequencies were significantly elevated compared with naïve macaques for both partially and completely protected macaques (ρ = 0.014 and ρ = 0.007 respectively; Mann-Whitney rank sum test) and were not significantly different from day of challenge frequencies. Five of 6 completely protected macaques, the exception being macaque E65, had elevated TCM and reduced TEM (CD28-CCR7-) at day 130 post-challenge compared with day of challenge but failed to reach statistical significance (p = 0.063 Wilcoxon signed rank sum test). Pairwise comparison of TCM and TEM frequencies at day 130 post-challenge and day of challenge in partially protected macaques showed no significant changes. Thus, the polarisation of circulating CD8+ T memory populations observed in completely protected macaques on the day of challenge was not evident 130 days after superinfection challenge.
In order to evaluate the possible influence of SIV-specific T cell quantity and quality on protection status, PBMC were stimulated in vitro with peptide pools from SIV Gag, Rev and Tat and intracellular cytokine staining for IL-2, IFN-γ, TNF-α and IL-17 was analysed by flow cytometry for CD4+ and CD8+ T cells. The total frequency (ie mono + bi + tri + quadruple) of SIV-specific CD8+ T cells was found to be significantly higher in completely protected compared with partially protected macaques on the day of challenge (ρ = 0.041; Mann-Whitney rank sum test); whereas no difference was seen with CD4+ cells (Fig 5). A similar analysis at day 130 after challenge failed to show a difference between groups for either CD8+ or CD4+ T cells (S6 Fig). It was noted, however, that the frequency of CD4+ T cells was markedly elevated in both groups regardless of protection status when compared with day of challenge and was statistically significantly different for completely protected animals (p = 0.063 for partially protected and p = 0.031 for fully protected animals; Wilcoxon signed rank test). In only one animal, E61, were frequencies similar on the two occasions tested (3.57% and 3.51%, day of challenge and day 130 post challenge respectively) and were largely confined to mono-functionality (see below). Although total frequencies of SIV-specific CD8+ T cells also showed an upwards trend at day 130 after challenge the differences were not statistically significant.
Deconvolution of cytokine combinations showed that on the day of superinfection challenge 6/6 completely protected animals had circulating SIV-specific quadruple cytokine expressing CD8+ cells at a frequency of >0.02% compared to only 1/6 partially protected macaques (ρ = 0.015; Fisher’s exact test). Differences in median frequencies of maximally polyfunctional CD8+ T cells between the groups did not reach statistical significance (ρ = 0.065; Mann-Whitney rank sum test) due to the outlier E71 (Fig 6). Similarly, IFN-γ + TNF-α dual positive CD8+ T cells were absent or below 0.02% in partially protected animals whereas in the completely protected group 5/6 macaques had frequencies markedly above 0.02% (ρ = 0.015; Fisher’s exact test) with a significantly elevated median frequency (ρ = 0.026; Mann-Whitney rank sum test). Significant differences were not seen for any cytokine combination 130 days after challenge (Fig 7). No significant differences were seen between partially and completely protected groups in the frequencies of circulating SIV-specific CD4+ T cells expressing individual cytokine combinations at either the day of superinfection challenge (S7 Fig) or 130 days after challenge (S8 Fig).
Although it was not possible to discern protection status-specific differences in circulating CD8+ T cell responses 130 days after wt-challenge, responses in lymphoid tissue may be more informative. Mononuclear cells extracted from mesenteric lymph nodes at necropsy were stimulated in vitro with a pool of Nef unique region-specific peptides. SIVrtTA and SIVmac239Δnef vaccine strains do not produce Nef protein, whereas the SIVmac239 challenge virus expresses full-length Nef. Surprisingly, poly and mono-functional CD8+ T cells were detected regardless of protection status (Fig 8). Although statistically different frequencies of functional cells were not detected between the groups, there was a trend towards higher reactivity in completely protected animals.
Sections of spleen from vaccinates, naïve challenge controls and unchallenged macaques were stained with KK77 monoclonal antibody specific for Nef (Fig 9). Positive cells were detected in partially-protected SIVrtTA vaccinates as well as fully-protected macaques E65 and E70. In contrast, macaques of Group C vaccinated with SIVmac239Δnef were indistinguishable from negative controls. Although clearly detectable staining for Nef was present in E65, the staining pattern was more diffuse with occasionally identifiable foci of positive cells, as distinguished from productively infected macaques which were partitioned into the partially protected group.
The reported breadth and duration of protection conferred in macaques following vaccination with live attenuated SIV has many of the features required of an effective vaccine against HIV/AIDS. Understanding the mechanisms of protection may allow the informed design of intrinsically safe vaccines. Earlier attempts to improve the safety profile of live attenuated SIV by introducing multiple attenuating mutations revealed that the degree of protection was inversely proportional to the degree of attenuation [11]. Hence, it was perhaps not unexpected that SIV clones molecularly engineered to be limited to a single round of replication conferred only limited protection compared with more vigorously replicating attenuated vaccine strains [23, 33]. The development of SIVrtTA with potential to be temporally modulated for replication in vivo provides a novel tool to further dissect the processes of protection elicited by live attenuated SIV. Previously, we reported this novel virus to replicate in vivo and being fully infectious in rhesus macaques, with the ability to disseminate to lymphoid tissues and elicit a range of immunological responses including reversible changes in the frequency of memory T cell subsets dependent upon the withdrawal of dox [26].
Here, we report that vaccination with SIVrtTA confers protection against homologous wild-type challenge, in some cases similar to the ‘gold-standard’ SIVmac239Δnef vaccine. However, levels of protection were variable. Full or complete protection (based on absence of a wild-type post-challenge viraemia) was associated with a prolonged shoulder of persisting SIVrtTA vRNA signal in plasma during the dox-on period rather than the later modulation of replication in lymphoid tissues (occult replication) mediated by the administration of dox. This aberrant viraemic profile may be dependent upon intrinsic host factors, for example the availability of alternative secondary receptors, or mutational events in the vaccine virus. Notably, mutations in rtTA which do not affect dox-dependence were detected only in the fully protected macaques and may have contributed to the fitness of SIVrtTA in vivo. Interestingly, a similar virological profile was seen also in one animal vaccinated with SIVmac239Δnef. Despite the replicative fitness cost introduced by the dox-dependent regulatory elements, the remaining animals vaccinated with SIVrtTA demonstrated significant protection from wild-type SIVmac239 challenge, as breakthrough of challenge virus was at lower levels than naive challenge controls with reduced lymphoid virus sequestration. These results support the observation that in the SIV/macaque model, and in common with other live attenuated vaccines, a defining feature of efficacy is related to the ability of the vaccine virus to replicate in the early phases of vaccination and in addition, suggest that limited acute phase replication may be compensated by subsequent persistence.
SIVrtTA shows absolute dependency upon dox for its replication in vitro [24, 25] and as we have shown previously, dox status influences the TEM circulating pool [26]. Nevertheless, we have not formally directly demonstrated that dox status completely controls replication in vivo in all anatomical compartments. Whilst we consider that loss of dox-dependency is unlikely, given the lack of mutations in the known critical sites, future experiments could include challenge of naïve macaques in the absence of administration of doxycycline.
Application of discriminatory PCR assays able to unravel the relative contributions of each virus to detectable PCR signals was a critical component of this study. These assays unequivocally established that plasma RNA following challenge of Group C animals, and of fully protected SIVrtTA vaccinates E70 and E65 was vaccine-virus specific. Moreover, this was corroborated by analysis of CA-RNA in lymphoid tissues at the termination of the study. A hallmark of complete vaccine protection appeared to be the persistent replication of vaccine virus in lymphoid tissue. Surprisingly however, only a very low level of vaccine CA-RNA was detected in a single tissue of macaque E75 suggesting there may have been persistence elsewhere such as the gut and/or vaccine generated immunity had cleared infection to limits below detection at least in the tissues examined. The mechanism for the persistent low-level replication of SIVrtTA in the absence of dox in macaque E65 is unknown. As we have reported previously, low levels of vRNA have been detected by in situ hybridisation in small intestine from SIVrtTA-infected rhesus macaques following dox withdrawal [26]. Therefore, we are unable to formally exclude the possibility that dox-dependency in vivo is conditional.
It was notable that where breakthrough virus was detected in lymphoid tissues of Group B animals, maintained on dox throughout the experiment, there was no evidence of residual vaccine virus. We reported previously that proviral DNA was detected pre-challenge in the spleen, PLN and MLN of animals maintained on dox although concentrations were lower than in macaques vaccinated with SIVmac239Δnef [26]. Presumably, given the fitness disadvantage, any extant replicating SIVrtTA was displaced by the challenge wild-type virus.
Although in this study we were unable to definitively address whether persisting vaccine virus replication in lymphoid tissue is an absolute requirement for complete protection because of the reduced replication of SIVrtTA, the opportunity was available nonetheless to compare T memory cell frequencies and cellular immune responses in partially and completely protected groups. The T memory cell results showed a strong association with protection status, which in most analyses reached statistical significance. The complete loss of these associations when analysis was done 130 days after challenge is striking, particularly in (1) the polarisation of CD4+ memory T cells toward the TCM phenotype regardless of protection status and (2) the changes in proportions of CD8+ memory T cells in completely protected animals. Although not reaching statistical significance due to outliers there was a clear trend for reduction in the number of CD8+ TEM with a concomitant increase in TCM. This latter effect probably reflects a reduction in on-going antigen re-stimulation in vivo at this time and/or a redistribution of TEM to tissue compartments. We did attempt analysis in gut tissues taken post mortem; however, cell recovery was poor making interpretation of flow cytometric data unreliable.
Despite the reported lack of association between responses detected in the blood and subsequent protection [10, 12, 6], we identified a statistically significant association between high frequencies of global TEM in peripheral blood at the time of challenge and outcome. Moreover, total frequencies of SIV-specific polyfunctional CD8+ T cells were significantly higher in macaques exhibiting complete protection, compared with partially protected macaques, on the day of challenge. Interestingly however, macaque E65, which demonstrated continuous very low-level replication of SIVrtTA in the absence of dox, failed to show this association, perhaps suggesting that other factors may be associated with complete protection in this animal. Further analysis of cytokine combinations revealed that CD8+ memory T cells with quadruple cytokine staining and cells staining for IFN-γ and TNF-α were present at higher frequency in complete protection compared with the frequencies in partially protected animals. Interestingly, the one macaque that did not have detectable dual-stained CD8+ T cells, E70, had an exceptionally high frequency of quadruple staining cells. Clearly, this analysis represents only a fraction of the total picture, since proteome-wide expansion of T cells was not performed and only 4 cytokines were analysed.
ICS staining for IL-17 was included in the present study since perturbations in CD4+ and CD8+ IL-17-staining cells in both the periphery and mucosal compartments reportedly reflect SIV-induced changes in disease status [34–36] and therefore could be a useful marker particularly in animals that may become dually-infected after challenge with virulent virus (i.e. may indicate sparing from disease progression). Several animals displayed unexpectedly high IL-17 positivity either before or following superinfection challenge. The reasons for this are not known; however, it is worth pointing out that these results were obtained in the context of infection with a novel SIV construct and it is possible that in certain genetic backgrounds this virus stimulates a strongly regulatory T cell phenotype.
Analysis of SIV-specific CD8+ T cell frequencies in mesenteric lymph nodes did not reveal a difference between completely and partially protected animals; however, it did reveal evidence of a challenge virus footprint. The Nef-specific T cell responses seen could only be stimulated by wild-type virus challenge. As Nef is not a structural component of the virus, this would require de novo synthesis of Nef in infected cells. The absence of Nef-staining in the spleen of SIVmac239Δnef vaccinated animals is consistent with the notion that the mechanism of complete protection from wild-type virus challenge operates through early clearance of challenge virus; whereas in partially-protected animals T cells may suppress wild-type virus replication rates relative to those in vaccine naïve animals. It was however surprising that a low level of Nef staining was detected in the apparently completely-protected SIVrtTA vaccinated animals. Thus, although by the criteria of RNA detection and Gag-specific antibody responses post-challenge these animals appeared to be completely protected, they should perhaps be considered falling into an intermediate category between completely and partially protected. Clearly, however, these macaques were protected from overt, productive superinfection.
In this regard the timing between exposure to wild-type virus and recovery of tissues at autopsy for analysis may be critical. In this study a relatively long period (20 weeks) was allowed to elapse from time of wild-type challenge to autopsy, which is likely to have allowed sufficient time for a response to wild-type virus to be generated but where the virus was no longer detected at termination. In such a scenario, the challenge virus is likely to have been present at some level but which had been subsequently cleared by host T cell responses to wild-type virus infection reflecting previous reports in the literature where much earlier sampling for virus post-challenge (eg 14 days after wt challenge) resulted in detection of virus in tissues at necropsy but the overall virological phenotype was that of protection [37]. The likely role of T cells in this protection has been further demonstrated by CD8 T cell depletion experiments where control of the replication of both the challenge and vaccine viruses have been linked to a CD8 T cell response [38, 39].
Recently reported detailed analysis of immune responses and deep sequence characterisation of SIVmac239Δnef post-vaccination indicated that there is a shift following early, rapid virus escape due to immune pressure to variable regions targeted during the acute phase to a re-focussed immunological response to more conserved epitopes [40]. However, the level of sub-clinical antigenic drive required to deliver such an anentropic state requires clarification, perhaps also in the face of host responses to the vaccine virus, since it was also noted that macaques with undetectable plasma viraemia experienced ongoing sequence evolution of the vaccine virus. It is perhaps noteworthy that in our study we observed distinct sequence changes in the rtTA gene rescued from viral RNA in plasma, taken as a measure of recently replicating virus, in the two SIVrtTA protected macaques (E65, E70) during the early, post-acute phase of virus replication which further marked these macaques out as being virologically distinct from the other SIVrtTA vaccinates. Hence, viral evolution as a driver for improved virological fitness in vivo during the post-acute phase appears to have had a marked effect in terms of the overall protection status conferred on these two macaques. SIVrtTA replication in macaques will also probably induce immune responses not only against viral proteins but also against rtTA itself [41]. Therefore, it is plausible that the observed amino acid changes mediate a mechanism of immune-escape of the rtTA protein, which would likely improve persistent virus replication, but this was not formally investigated.
Hence, SIVrtTA vaccination of Indian rhesus macaques appears on the cusp of delivering potent vaccine protection. If SIVmac239Δnef-induced protection correlates with an expanded T cell anentropy to highly conserved epitopes with an associated increased depth of response generated over time, this likely explains the relatively poor ability of a ‘one-hit’ vaccine response, such as single cycle SIV to ensure long-lived vaccine protection. Compensations in vaccine replication appear important in conferring protection mediated by SIVrtTA, although whether these are sufficient to explain features of early protection from heterologous challenge, for example, remains unclear. Highly attenuated viral vaccines such as SIVmacΔ4 [11] which have a reduced replication potential in vivo, but which fail to persist, exhibit an intermediate protection profile. Hence the ability of SIVrtTA to exhibit low, continuous replication provides a clear advantage compared to these approaches. Lack of an increased magnitude of SIV-specific CD8 T cell responses in lymph nodes correlating with proposed mechanisms of protection for cellular responses at key sites of virus replication in the body [40], suggest that the role of CD8 T cells in this mode of vaccine protection is far from resolved, whereby a higher viral replication in turn leads to higher CD8 T cells responses in lymphatic tissue [10].
On the face of it our data appears to strongly support the view that CD8+ polyfunctional TEM are critical in protective immunity induced by live attenuated SIV as suggested by Fukazawa et al [10] for lymph node responses. However, technical limitations precluded the ability to assign ICS responsiveness specifically to memory phenotype in our study, and as in other studies, our observations remain correlative. Indeed, antibody responses to Gag p27 before and after vaccine challenge are also predictive of outcome but are unlikely of mechanistic significance. If the current paradigm of live attenuated vaccine protection is correct, it must also explain why superior responses in the host that prevent viral infection are established in the same host where host control of the vaccine is poorest. This counterintuitive observation requires a cogent answer irrespective of localisation of the vaccine virus e.g. in T-follicular helper cells which may be subject to immune privilege, or magnitude and breadth of measurable immune responses such as CD8 T cell responses which are potentially capable of targeting and controlling both vaccine and challenge viruses, yet the vaccine virus is able to persist at these key sites.
Taken together, our data provide further insight into the highly dynamic process of live attenuated SIV vaccine outcomes where the replicative properties and persisting nature of the vaccine virus appear crucial to vaccine efficacy. SIVrtTA provides a novel tool in our armoury to understand more fully processes of occult and patent virus replication at niche anatomical sites where issues of viral latency and persistence are crucial in understanding retrovirus and immune interactions.
Non-human primates were used in strict accordance with UK Home Office guidelines, under a licence granted by the Secretary of State for the Home Office which approved the work described. Animal work at NIBSC is governed by the Animals (Scientific Procedures) Act 1986 that complies with the EC Directive 86/609 and performed under licence (PPL 80/1952) granted only after review of all procedures in the licence by the NIBSC local Animal Welfare and Ethical Review Body. All study macaques were purpose bred and group-housed for the entire duration of the study, with daily feeding and access to water ad libitum. Given the limited availability of suitable macaques, age, sex and weight matching was not possible, nor central to the study outcome. Regular modifications to the housing area were made by husbandry staff including introduction of novel structures (eg swings and perching stations) and foodstuffs in novel manners to encourage foraging for food, to further enrich the study environment. The environmental temperature (15–24°C), was appropriate for macaques and rooms were subject to a 12 hour day/night cycle of lighting. Animals were acclimatised to their environment and deemed to be healthy by the named veterinary surgeon prior to inclusion on the study.
All animals were sedated with ketamine prior to bleeding or virus inoculation by venepuncture. Frequent checks were made by staff and any unexpected change in behaviour by individuals on study followed up, including seeking of veterinary advice where necessary. Regular blood evidence of incipient disease and veterinary advice were sought when persisting abnormalities detected. The study was terminated and all animals killed humanely by administering an overdose of ketamine anaesthetic prior to development of overt symptomatic disease. All efforts were made to minimise animal suffering, including provision of a high standard of housing quarters and monitoring of animal health and well-being and the absence of procedures not essential to the study.
16 UK purpose-bred Indian rhesus macaques (Macaca mulatta) were used in the study, in accordance with UK Home Office guidelines (Code of Practice 1988) and local ethical approval. The basic construction and mode of action of the SIV-rtTAΔnef (SIVrtTA) vaccine, based on a SIVmac239 genetic backbone, is depicted in Fig 1A. In a challenge study experiment, eight macaques were inoculated intravenously with 5 x 103 TCID50 SIVrtTA vaccine receiving 100mg daily oral dosing with dox. In four macaques (Group A), dox was removed eight weeks prior to SIVmac239 wild-type challenge. In the remaining four SIVrtTA vaccinates (Group B) dox dosing was maintained at 100 mg daily oral dosing. Group C comprised four macaques inoculated with 104 TCID50 SIVmac239Δnef. All vaccinates were challenged with wild-type SIVmac239 in addition to four additional macaques which served as naïve challenge controls (Group D). The study outline is summarised in Fig 1B. Veterinary procedures deployed the use of ketamine hydrochloride prior to sedate macaques. Plasma concentrations of dox were monitored ex vivo using a previously described assay [41].
Macaques were genetically characterised for host MHC profiles, by Dr David Watkins (Univ. Wisconsin, S2 Table). Distribution of TRIM5α and TRIMcyp alleles was determined as previously described. Mamu7 represents macaques harbouring the TRIMcyp allele [42].
Initial quantitative measures were made in peripheral blood using quantitative gag-based real-time PCR assays as previously described [6]. Plasma vRNA levels were determined for EDTA-treated plasma samples with a limit of detection of 50 SIV RNA copies/ml and SIV DNA levels on PBMCs with limit of detection one SIV DNA copy/100,000 cell equivalents. SIVrtTA-specific levels were determined using primers designed to amplify a region of the rtTA gene using PCR conditions comparable to those described for the total gag estimations against an rtTA plasmid containing unique sequences to the rtTA gene. SIVrtTA-specific amplification sequences were CGCCGTGGGCCACTT (forward), and CTTTCCTCTTTTGCTACTTGATGCT (reverse); internal rtTA probe sequence was FAM-CACTGGGCTGCGTATTGGAGGAACAG-BHQ1; primers and probes were used at 100nM concentrations. Wild-type SIVmac239-specific amplifications were made with CTCAGGACCAGGAATTAGATACC (forward), AAGGGTCATCCCACTGGGAAGT (reverse) and internal probe sequence FAM-TCCCTGTAAATGTATCAGATGAGGCACAGGAGG-BHQ1 targeting the nef gene. Primers were used at 100nM and probe at 120nM concentrations. Detection limits of virus-specific amplification in plasma were determined to be 100 RNA copies/ml with an amplification efficiency of >98%.
Cell-associated RNA determinations were made for SIVrtTA, SIVmac239Δnef and wild-type SIVmac239 respectively by adapting a previously reported method [7]. Total RNA was isolated from spleen, mesenteric and peripheral lymph nodes using an RNeasy kit (Qiagen), subjected to on-column DNAase treatment in accordance with the manufacturers’ protocol. Virus-specific targets were amplified by one-step RT-PCR using 50ng total RNA input, adapting the SIVrtTA and SIVmac239 wild-type specific primers described above and employing those described previously in [43] for SIVmac239Δnef-specific amplification as follows: cttaggagaggtggaagatggatactc (forward), CTTTTCTTTTATAAAGTGAGACCTGTTCC (reverse) and internal probe sequence FAM- CAATCCCCAGGAGGATTAGACAAGGGCTTG -BHQ1. Primers were used at 300nM and probe at 75nM. All CA-RNA determinations were made using normalised values of GAPDH, in co-amplification reactions as described in [7]. All amplifications were performed with Invitrogen Ultrasense kits with a thermoprofile of RT step 52°C for 30 mins; 10 mins at 95°C then 40 cycles of 95°C for 30 seconds and 60°C for 60 seconds. Limits of detection for SIVrtTA, SIVmac239Δnef, wild-type SIVmac239 CA-RNA assays were determined as 50, 34 and 80 SIV RNA copies per 50ng total RNA. All CA-RNA quantitative PCR assays had an efficiency of >95%, typically 98–99% efficiency of amplification. The SIVrtTA assays were validated using a plasmid construct denoted rtTAV16 diluted to an extinction end-point in quantitative assays.
Plasma anti-SIV p27 IgG responses were quantified by ELISA. Briefly, medium binding 96-well plates (Greiner, UK) were coated with 1μg/ml recombinant SIV p27 (CFAR, UK, Cat no: EVA664). Test plasma and standard positive and negative control samples were added to washed plates and bound IgG detected with goat anti monkey IgG-HRP (Serotec) followed by addition of substrate to induce a colour reaction in reactive samples.
Memory phenotype and intracellular cytokine staining were performed separately in each sample per animal due to limitations of the flow cytometry capability available. Peripheral blood lymphocytes (PBL) were isolated using Percoll gradient centrifugation and mesenteric lymph node mononuclear cells (MNC) were isolated by mechanical disaggregation of tissue. To delineate memory T cell subsets, PBL were simultaneously surfaced stained with anti-CD3-V500 (clone SP32, BD Horizon), anti-CD4-V450 (clone L200, BD Horizon), anti-CD8-APCCy7 (clone SK1, BD Biosciences), anti-CD95-PECy7 (DX2, BioLegend), anti-CD28-PerCP-Cy5.5 (eBiosciences), and anti-CCR7-FITC (R&D systems). Gates on lymphocyte subpopulation were defined as central memory CD8+C95+CD28+CCR7- and CD8+CD95+CD28- CCR7- as effector memory.
SIV-specific T cell responses were determined by cytokine production after incubation with 5 μg/ml of either SIV Gag, Tat, Rev or (for MLN MNC additionally Nef peptides from the nef -unique coding region) (15mers overlapping 11 residues, CFAR/NIBSC, Potters Bar, UK) plus 10 μg/ml CD49d, 50μg/ml anti-CD28, Golgi Stop (10ng/ml, BD), and incubated at 37°C in a 5% CO2 environment with RPMI 1640/10% FCS for 14h. Stimulated cells were surfaced stained for CD3, CD4 and CD8, permeabilised (Fix and Perm kit, Caltag), and then stained for intracellular cytokine detection with anti-IFNγ-PErCPCy5.5 (clone B27), anti-IL-2-PE (MQ1-17H21, eBiosciences), anti-TNF-α-APC (MAB11, eBiosciences) and anti-IL-17-Pacific Blue (BioLegend). Polyfunctional T cells were determined by a gating strategy as shown in the representative plots (S9 Fig). In detail, within CD4 and CD8 subsets, distribution of TNF-α and/or IL-2 producing cells were specified using contour FACs profile quadrants. Each quadrant within these cell populations were sequentially analysed for IFN-γ and/or IL-17 production in combinatory plots. For group comparisons (partial versus complete), total frequencies of ICS-stained cells were derived by adding mono, bi, tri and quadruple functional frequencies for each animal. The relative distribution of the cytokine producing cells in each animal was summarised in pie charts using SPICE software.
All peripheral and tissue derived mononuclear cells were acquired and analysed using a BD Canto II flow cytometer (BD Immunocytometry) with FACS DIVA software as described previously [26].
Graphing and associated statistical analyses, as specified, were performed using Sigma Plot 11 (Systat Software, Inc.). Kruskal-Wallis analyses of variance with Dunn’s post-hoc test were determined using the Minitab version 17 software. In addition, analysis and graphical representation of cytokine production were conducted using the data analysis programme Simplified Presentation of Incredibly Complex Evaluations (SPICE, version 5.3) provided by M. Roederer, National Institutes of Health, Bethesda, MD.
Immunochemical staining for Nef was performed with the KK77 antibody (CFAR; ARP3093) which is an IgG2a isotype raised to recombinant SIVmac251 Nef and which detects wild-type Nef only, using protocols as previously described [7].
|
10.1371/journal.pgen.1006398 | Fine Mapping of a Dravet Syndrome Modifier Locus on Mouse Chromosome 5 and Candidate Gene Analysis by RNA-Seq | A substantial number of mutations have been identified in voltage-gated sodium channel genes that result in various forms of human epilepsy. SCN1A mutations result in a spectrum of severity ranging from mild febrile seizures to Dravet syndrome, an infant-onset epileptic encephalopathy. Dravet syndrome patients experience multiple seizures types that are often refractory to treatment, developmental delays, and elevated risk for SUDEP. The same sodium channel mutation can produce epilepsy phenotypes of varying clinical severity. This suggests that other factors, including genetic, modify the primary mutation and change disease severity. Mouse models provide a useful tool in studying the genetic basis of epilepsy. The mouse strain background can alter phenotype severity, supporting a contribution of genetic modifiers in epilepsy. The Scn1a+/- mouse model has a strain-dependent epilepsy phenotype. Scn1a+/- mice on the 129S6/SvEvTac (129) strain have a normal phenotype and lifespan, while [129xC57BL/6J]F1-Scn1a+/- mice experience spontaneous seizures, hyperthermia-induced seizures and high rates of premature death. We hypothesize the phenotypic differences are due to strain-specific genetic modifiers that influence expressivity of the Scn1a+/- phenotype. Low resolution mapping of Scn1a+/- identified several Dravet syndrome modifier (Dsm) loci responsible for the strain-dependent difference in survival. One locus of interest, Dsm1 located on chromosome 5, was fine mapped to a 9 Mb region using interval specific congenics. RNA-Seq was then utilized to identify candidate modifier genes within this narrowed region. Three genes with significant total gene expression differences between 129S6/SvEvTac and [129xC57BL/6J]F1 were identified, including the GABAA receptor subunit, Gabra2. Further analysis of Gabra2 demonstrated allele-specific expression. Pharmological manipulation by clobazam, a common anticonvulsant with preferential affinity for the GABRA2 receptor, revealed dose-dependent protection against hyperthermia-induced seizures in Scn1a+/- mice. These findings support Gabra2 as a genetic modifier of the Scn1a+/- mouse model of Dravet syndrome.
| Epilepsy is a neurological disorder affecting approximately 3 million Americans and 1% of the worldwide population. Approximately 70% of patients diagnosed with epilepsy have a genetic basis for their disease. The same genetic mutation can result in epilepsy with varying clinical severity in some individuals. This suggests that additional factors modify the effect of the primary mutation, resulting in the differences observed. Mouse models of epilepsy also exhibit variable severity depending on the inbred strain background. This is a useful model system that enables us to determine other genetic factors that influence phenotype presentation. This study investigates a mouse model of Dravet syndrome. Similar to Dravet syndrome patients, the mice experience a variety of seizure types and have increased risk for premature death. Interestingly, the Dravet syndrome phenotype is completely masked on one mouse strain background. Through a series of genetic and pharmacological techniques we were able to identify a gene that is likely responsible for the strain-dependent phenotype difference of Dravet mice. Identifying genes that modify diseases can help predict patient outcomes and suggest new therapies for the treatment of epilepsy.
| Epilepsy is one of the most common neurological disorders, affecting approximately 50 million people worldwide. Approximately two-thirds of human epilepsies are presumed to have a genetic basis. Voltage-gated sodium channel gene mutations are the most common cause of monogenic epilepsy. Mutations in SCN1A, encoding Nav1.1, result in a broad spectrum of disorders ranging from simple febrile seizures to Dravet syndrome, a severe, infant-onset epileptic encephalopathy with devastating outcomes [1]. Dravet syndrome patients have multiple seizures types that are refractory to treatment, as well as delays of psychomotor and cognitive development and a high risk of sudden unexpected death in epilepsy (SUDEP) [2].
Within genetic epilepsies, diverse phenotypes often result from identical sodium channel mutations [3]. This suggests that disease severity is influenced by other factors, including genetic modifiers. The genetic basis of epilepsy can be studied using mouse models that recapitulate hallmark features of human epilepsies. Frequently, mutations that result in seizures show strain-dependent phenotype variability, suggesting a contribution of genetic modifiers in epilepsy [4]. Heterozygous deletion of mouse Scn1a models features of Dravet syndrome, including spontaneous seizures, thermal seizure sensitivity, cognitive deficits, and increased mortality [5,6]. Interestingly, expressivity of the Scn1a+/- phenotype is highly strain-dependent. When the Scn1a+/- mutation is maintained on the 129S6/SvEvTac (129) strain, 129.Scn1a+/- mice have no overt phenotype and a normal lifespan. When 129.Scn1a+/- mice are crossed with C57BL/6J (B6), the resulting [129xB6]F1.Scn1a+/- (F1.Scn1a+/) offspring exhibit spontaneous seizures and 75% mortality by 8 weeks of age [7]. We previously mapped several Dravet syndrome modifier (Dsm) loci that influence the strain-dependent difference in survival [7].
In the current study, we used interval-specific congenic (ISC) strains to confirm and fine map the Dsm1 locus on chromosome 5. We then performed candidate gene analysis using an RNA-Seq approach, and identified Gabra2 as a high priority candidate gene. Further evaluation of Gabra2 by expression analysis and pharmacological manipulation support Gabra2 as a putative modifier gene that influences survival in the Scn1a+/- Dravet model.
Inheritance of B6 alleles in Dsm1 increased the risk of early death in Scn1a+/- mice [7]. Low resolution mapping identified the Dsm1 1.5 LOD support interval as 27.3–41.3 cM on chromosome 5, with a peak at 34 cM [7]. To refine the Dsm1 map interval, five interval specific congenic (ISC) lines were generated with varying 129-derived chromosome 5 segments within Dsm1 on a congenic B6 background (Fig 1A). Each ISC line was crossed with 129.Scn1a+/- to generate F1.Scn1a+/-offspring carrying either homozygous (129/129) or control heterozygous (129/B6) alleles in Dsm1. Survival of F1.Scn1a+/- mice was then monitored until 8 weeks of age (Fig 1B).
A significant improvement in survival was observed with lines ISC-A (4–118 Mb; p<0.01), ISC-B (4–74.9 Mb; p<0.0001), ISC-C (53–118 Mb; p<0.05) and ISC-D (64.6–74.9 Mb; p<0.0002) (Fig 1B). The strongest effect on survival was observed in mice carrying homozygous 129/129 alleles in ISC-B and ISC-D, with 8 week survival of 91.3% and 92.8%, respectively, compared to only 28.5% in heterozygous F1.Scn1a+/- control mice (Fig 1B). ISC-D carried the smallest 129-derived segment (64.6–73.9 Mb) and refined the map interval to ~9 Mb (Fig 1A).
We next addressed whether the survival advantage conferred by 129 alleles in the ISC-D interval correlated with a decreased seizure burden. F1.Scn1a+/-offspring carrying either homozygous (129/129) or control heterozygous (129/B6) alleles in ISC-D were monitored from postnatal day 21 (P21) through P24 to determine the frequency of generalized tonic-clonic (GTC) seizures. Scn1a+/- mice carrying homozygous (129/129) alleles in the ISC-D interval did not have a significantly reduced seizure frequency compared to heterozygous F1.Scn1a+/- control mice (ISC-D: 0.024 ± 0.015 seizures/hr; F1 control: 0.022 ± 0.02 seizures/hr). Dissociation of the survival and seizure frequency phenotypes was somewhat surprising, and raised the question of whether seizure frequency and survival are correlated in F1.Scn1a+/- mice. To address this question, we examined a separate phenotyping cohort and determined seizure counts for F1.Scn1a+/- mice between P19 to P24. In F1.Scn1a+/- mice experiencing seizures, total numbers of seizures were compared with survival throughout the monitoring period (Survivors: 7.95 ± 1.42 seizures; Lethals: 5.13 ± 1.55 seizures). This comparison revealed that high seizure frequency did not correlate with premature lethality in F1.Scn1a+/- (Pearson r2 = 0.01, p>0.8111), consistent with our observations with ISC-D.
The Dsm1 locus was narrowed to a 9 Mb region that contains 109 known and predicted genes, of which 40 are known to be expressed in the brain (Fig 2A). Within that interval, we identified four genes with non-synonymous coding sequence differences: Phox2b, Tmem33, Slc30a9, and Fryl. (Table 1). All resulting amino acid differences were predicted to be tolerated by SIFT [8].
To assess expression differences between 129 and [129xB6]F1 strains, we performed RNA-Seq on hippocampal RNA isolated from wildtype 129 and [129xB6]F1 mice at P24. Within the Dsm1 interval, RNA-Seq identified three genes, Nsun7, Bend4 and Gabra2, with significant expression differences between the 129 and [129xB6]F1 strains (Fig 2B, S1 Table). Nsun7 is a putative RNA methyltransferase that when mutated, leads to reduced sperm motility in humans and mice [9,10]. Bend4, or BEN domain containing 4, is of unknown function. Gabra2 encodes the GABAA receptor α2 subunit, which has previously been associated with addiction and dependence [11–13]. Considering the importance of GABA signaling in epilepsy and reported dysfunction in GABAergic signaling in Dravet syndrome [14–16], Gabra2 was identified as the top candidate modifier gene at Dsm1. Expression of Gabra2 was significantly elevated in wildtype 129 mice compared to [129XB6]F1 mice (Fig 2B).
We further evaluated forebrain Gabra2 transcript and protein expression in ISC-D mice carrying homozygous 129/129, heterozygous 129/B6 or homozygous B6/B6 alleles in Dsm1. We used digital droplet RT-PCR (ddRT-PCR) to evaluate transcript expression of Gabra2, as well as expression of other GABAA receptor genes located within Dsm1, including Gabra4, Gabrb1 and Gabrg1. Gabra2 expression was significantly different between all genotypes (F2,15 = 206.4, p<0.0001). We observed the highest levels of Gabra2 transcript with homozygosity for 129 alleles in Dsm1, intermediate levels with heterozygosity, and the lowest levels with B6 homozygosity (Fig 3A). Transcript expression of the other Gabr genes in the interval did not differ between genotypes (S1 Fig). Relative GABRA2 protein expression determined by immunoblotting correlated with the transcript levels, showing the highest expression with homozygous 129 alleles in Dsm1, intermediate with heterozygous alleles, and lowest with B6 homozygosity (Fig 3B and 3C). These results demonstrate that the level of Gabra2 expression is regulated by alleles within the Dsm1 interval, and that high expression of Gabra2 is correlated with the protective allele for Scn1a+/- survival.
We used clobazam, an anticonvulsant drug that has preferential affinity for the GABRA2 subunit [17–19], as a pharmacological tool to further evaluate Gabra2 as a candidate modifier gene. Both Dravet syndrome patients and F1.Scn1a+/- Dravet mice experience seizures triggered by hyperthermia. We evaluated the effect of clobazam on survival and hyperthermia seizure threshold in F1.Scn1a+/- mice. Beginning at P18, F1.Scn1a+/- mice were fed either chow containing clobazam (320 mg per kg of chow; estimated dosage 40 mg/kg/day) or control chow, and survival was monitored to 30 days of age. This dose of clobazam, selected to achieve plasma concentrations within the human therapeutic range, was not sufficient to provide protection against survival. We used the hyperthermia-induced seizure assay to perform an acute dose-response study. Between P14–16, F1.Scn1a+/- mice received an intraperitoneal injection of vehicle or clobazam (0.3, 1, or 30 mg/kg) prior to the induction of hyperthermia. Clobazam administration provided dose-dependent protection against hyperthermia-induced seizures (Fig 4). Clobazam dosed at 30 mg/kg provided complete protection against hyperthermia-induced seizures, while 83% of vehicle controls experienced GTC seizures with an average threshold temperature of 42.3°C (p < 0.0006). Clobazam doses of 0.3 and 1 mg/kg significantly increased GTC seizure thresholds to 43.0°C (p < 0.0022) and 42.9°C (p < 0.0079), respectively. To ensure plasma concentrations of clobazam reflected the human therapeutic range (0.1–0.4 μg/mL), plasma samples were assayed by HPLC. The commonly cited clobazam rodent dose of 30 mg/kg [20–22] resulted in an average plasma concentration that was 15 times higher (5.84 ug/mL) than the human therapeutic range, while the 1 mg/kg dose resulted in an average plasma concentration (0.16 ug/mL) within the range. Plasma concentrations were below the detection limit of the HPLC assay following the 0.3 mg/kg dose.
Candidate gene and expression analysis, along with pharmacological testing, identified Gabra2 as a putative modifier gene that influences survival in the Scn1a+/- Dravet mouse model. Considering the significance of GABAergic pathways in relation to epilepsy, strain-specific expression of Gabra2 may differentially influence excitatory/inhibitory balance. Several GWAS studies have demonstrated that human GABRA2 genetic variants are associated with alcohol dependence [23–27]. However, to date, no GABRA2 genetic variants have been associated with human epilepsy. Altered GABRA2 transcript and protein expression has observed in a number of rodent seizure models and patients with focal cortical dysplasia, tuberous sclerosis and temporal lobe epilepsy, suggesting a potential link between GABRA2 expression and epilepsy [28–32]. There is evidence for variation in human cortical GABRA2 expression that is associated with a trans eQTL on chromosome Xp11.4 [33]. Our mouse RNA-Seq and expression analysis demonstrated that low Gabra2 expression associated with the B6 allele and correlated with reduced survival in the Scn1a+/- model. Previous work has shown that in some mouse strains, Gabra2 expression is strongly modulated by a cis eQTL that overlaps with Dsm1 and a trans eQTL on chromosome 10 [34]. The difference in Gabra2 expression was maintained when we examined transcript and protein levels in brain tissues from the ISC-D line, which contains 129 alleles only in the minimal interval and is B6 elsewhere in the genome. From this, we infer that Gabra2 expression difference arises from a local eQTL effect within the Dsm1 interval. This is supported by a recent study that demonstrated differential allele expression of Gabra2 transcript in B6/129SF1 mice, with the 129 allele expressed at 2–3 times the level of the B6 allele [35]. Future work will focus on identifying the responsible regulatory variants within the Gabra2 genomic region.
The failure of line ISC-D to reduce seizure frequency in F1.Scn1a+/- mice was not surprising. As we report here, and have observed in our previous studies, high seizure counts do not correlate with premature death in F1.Scn1a+/- mice. Therefore, we could exclude the survival improvement of the ISC-D line to be a result of altered seizure frequency. GABAergic neurons are found in several brain regions involved in the control of respiration and cardiac activity, and dysfunction in either pathway is believed to be primarily responsible for SUDEP [36–39]. Mice with a conditional Scn1a+/- deletion in forebrain GABAergic interneurons experience SUDEP and cardiac dysfunctions similar to the Dravet Scn1a+/- mouse model [40]. When Scn1a+/- was conditionally deleted in the heart, the phenotype was not observed, demonstrating that SUDEP observed in Scn1a+/- mice originates from the central nervous system [40]. Increased expression of GABRA2 may exert a protective effect by altering autonomic control of cardiac and/or respiratory function and may explain the improved survival observed with 129/129 homozygosity in ISC-D. Additionally, GABAA receptor activity plays a pivotal role in neuronal development and altered levels of Gabra2 during development may alter network excitability [41].
Interestingly, clobazam, which has preferential affinity for the GABRA2 subunit, offers some therapeutic benefit in Dravet syndrome patients [17–19,42,43]. In our studies, threshold temperature for hyperthermia-induced seizures was elevated in Scn1a+/- mice treated with supratherapeutic levels of clobazam, while lower doses did not provide significant protection from hyperthermia-induced seizures or premature death. The 30mg/kg clobazam dose provided complete protection against hyperthermia-induced seizures, but resulted in plasma levels far above the human therapeutic range, highlighting the necessity of determining drug concentrations in similar studies. Frequently, Dravet syndrome patients are prescribed stiripentol as an add-on treatment to clobazam. In a randomized placebo-controlled study in children with Dravet syndrome, 71% of patients treated with a combination of clobazam and stiripentol experienced a reduction in seizure frequency [44]. Although the mechanism of action for the anticonvulsant benefit of stiripentol is yet to be determined, stiripentol has been identified as an inhibitor of CYP450s, specifically CYP3A4 and more potently CYP2C19, both important for the metabolism of clobazam. In vivo, a significant increase in the plasma concentrations of both clobazam and its active metabolite N-desmethylclobazam has been shown with concomitant stiripentol treatment [45], suggesting that elevated levels of clobazam may account for some of the therapeutic benefit.
Identification of Gabra2 as a putative modifier gene at the Dsm1 locus furthers our understanding of the genetic basis of Dravet syndrome. This provides information that may help predict the risk of SUDEP and the clinical course of epilepsy due to a sodium channel mutation. Furthermore, modifier genes may suggest new targets for the improved treatment of epilepsy.
All studies were approved by the Vanderbilt University [M/06/499 JAK] and Northwestern University Animal Care and Use Committees [IS00000539 JAK] in accordance with the National Institutes of Health Guide for the Care and Use of Laboratory Animals. Principles outlined in the ARRIVE (Animal Research: Reporting of in vivo Experiments) guideline and Basel declaration (including the 3R concept) were considered when planning experiments.
Scn1atm1Kea mice, with deletion of the first coding exon, were generated by homologous recombination in TL1 ES cells (129S6/SvEvTac) as previously described [7]. The 129.Scn1a+/- line has been maintained by continuous backcrossing to 129S6/SvEvTac (129). To generate mice for hyperthermia-induced seizure experiments, 129.Scn1a+/- mice were crossed with C57BL/6J (B6) resulting in [129XB6]F1.Scn1a+/- (F1.Scn1a+/-) mice. Mice were maintained in a Specific Pathogen Free (SPF) barrier facility with access to food and water ad libitum.
We generated five ISC lines carrying 129-derived chromosome 5 segments on a B6 background. [129 x B6]F1 progeny were continually backcrossed to B6 to generate congenic lines. Genotyping for chromosome 5 markers was performed at each generation and animals retaining 129 alleles in the Dsm1 interval were propagated. Whole genome and selective genotyping was performed at generations N2–N7 to select breeders with the lowest percentage of 129 alleles in the rest of the genome. All ISC lines were backcrossed to B6 for ≥N6 generations prior to experiments.
Dsm1 ISC females were crossed with heterozygous Scn1a+/- males to generate (B6.ISC x 129.Scn1a+/-)F1 offspring carrying heterozygous (129/B6) or homozygous (129/129) alleles in Dsm1. Phenotyping was performed similarly to our low-resolution mapping study [7]. Mice were ear-tagged at P12–14. At P21, mice were weaned into holding cages containing four to five mice of the same age and sex. Wild-type littermates were included in all holding cages. Survival was monitored to 8 weeks of age. During that time, all mice were monitored daily for general health and any mouse visibly unhealthy (e.g. underweight, dehydrated, poorly groomed, or immobile) was euthanized and excluded from the study. The focus of the study was sudden and unexpected death in the Scn1a+/− mice occurring in otherwise healthy appearing animals.
We used continuous video monitoring to determine GTC seizure frequency in Scn1a+/- mice with homozygous or heterozygous alleles in ISC-D mice. At P20, mice were placed in a recording chamber and video was captured using a Day/Night camera (Samsung SCB5003) equipped with an infrared lens (Tamron 13FG04IRSQ) and saved to a DVR (Samsung SRD-876D). During recording, mice had ad libitum access to food and water, and were maintained on a 14:10 light-dark cycle. Video records were analyzed offline by an observer blinded to genotype. We validated use of video capture for seizure evaluation in Scn1a+/- mice in a preliminary study. Behavioral GTC seizures were correlated with electroencephalographic seizures using video-electroencephalography (EEG) monitoring as previously described. On P16, F1.Scn1a+/- mice were implanted with prefabricated headmounts (Pinnacle Technology, Inc., Lawrence, KS, USA) and mice were allowed to recover for >72 hours. Continuous video-EEG data monitoring was performed from P20-P26. Data were acquired and analyzed with Sirenia software (Pinnacle Technology, Inc.). Electrographic GTC seizure activity was scored manually using video-EEG data. Separately, behavioral GTC seizures were counted using only the video record. There was perfect agreement between the video-EEG and video-only results (κ = 1.0; n = 25 mice; 39 seizures), validating video recording as a reliable method for measuring GTC seizure frequency in Scn1a+/- mice.
DNA was prepared from tail biopsies using the Gentra Puregene Mouse Tail Kit according to the manufacturer’s instructions (Qiagen, Valencia, CA, USA). Scn1a genotype was determined by multiplex PCR as previously described [7]. Microsatellite genotyping was performed by analysis of PCR products on 7% denaturing polyacrylamide gels stained with ethidium bromide.
Hippocampi were dissected from 129 and [129 x B6]F1 mice at P24. Primary pools were created with tissue from both male (n = 2) and female (n = 2). Total RNA was isolated using TRIzol reagent according to the manufacturer’s instructions (Life Technologies). Following RNA isolation, RNA integrity was assessed and all samples had a RIN of ≥ 7.7. For each strain, three superpooled biological replicates were generated by combining total RNAs from 3–4 primary pools (n = 12–16 mice/superpool). RNA integrity was assessed on the superpool samples and all samples had a RIN of ≥ 8.1.
Samples were processed for RNA-Seq using the TruSeq RNA Library Preparation Kit v2 (Illumina, San Diego, CA, USA). Samples were sequenced on an Illumina HiSeq 4000 at BGI (Hong Kong, China). Three multiplexed lanes of 50-bp single-end sequencing resulted in almost 167 million mapped reads. Base calling and filtering of sequence reads were performed with the Illumina pipeline [46]. Bioinformatic analysis was performed using Tuxedo Tools on the GALAXY platform [47–49]. Tophat2 aligned reads with Bowtie2 and identified splice junctions [50]. StringTie assembled alignments into transcripts [51]. Cuffmerge and Cuffcompare combined biological replicate transcript files and assigned reference annotations to transcripts [52]. Cuffdiff provided significance values for expressed genes and transcripts [52].
Forebrain RNA was extracted from P24 ISC-D mice carrying homozygous B6, homozygous 129, or heterozygous alleles. Total RNA was isolated using TRIzol reagent according to the manufacturer’s instructions. First-strand cDNA was synthesized from 2 micrograms of RNA using oligo(dT) primer and Superscript IV reverse transcriptase according to the manufacturer’s instructions (Life Technologies). First-strand cDNA samples were diluted 1:10 and 5 μl was used as template. Quantitative digital droplet PCR (ddPCR) was performed using ddPCR Supermix for Probes (No dUTP) (Bio-Rad, Hercules, CA, USA) and TaqMan Gene Expression Assays (Life Technologies) for mouse Gabra2 (FAM-MGB-Mm00433435_m1), Gabra4 (FAM-MGB-Mm00802631_m1), Gabrb1 (FAM-MGB-Mm00433461_m1), Gabrg1 (FAM-MGB-Mm00439047_m1) and Tbp (VIC-MGB-Mm00446971_m1). Reactions were partitioned into 20,000 droplets (1 nL each) in a QX200 droplet generator (Bio-Rad). Thermocycling conditions were 95°C for 10 minutes, then 40 cycles of 95°C for 15 seconds and 60°C for 1 minute (ramp rate of 2°C/sec) and a final inactivation step of 98°C for 10 minutes. Following amplification, droplets were analyzed with a QX200 droplet reader with QuantaSoft v1.6.6.0320 software (Bio-Rad). All assays lacked detectable signal in no-RT and no template controls. Relative transcript levels were expressed as a ratio of the gene of interest concentration to Tbp concentration. Statistical comparison between groups was made using ANOVA with Tukey’s post-hoc tests. Data are presented as mean ± SD of 5–6 biological replicates.
Forebrain membrane proteins were isolated from P24 ISC-D mice carrying homozygous B6, homozygous 129, or heterozygous alleles. Membrane fractions (50 μg/lane) were separated on a 7.5% SDS-PAGE gel and transferred to nitrocellulose membranes. Immunoblots were probed for GABRA2 using a rabbit polyclonal antibody (Phosphosolutions, 822-GA2C) and for Mortalin using a mouse monoclonal antibody (NeuroMab, 75–127). Alexa-conjugated fluorescent secondary antibodies (Jackson ImmunoResearch) were used to detect bound primary antibody using an Odyssey imaging system (Licor).
F1.Scn1a+/- mice between P14 to P16 received an intraperitoneal injection of 0, 0.3, 1, or 30 mg/kg doses of clobazam solubilized in vegetable oil (10 ml/kg volume). After 15 minutes, each mouse was placed in a cylindrical container and fitted with a RET-3 rectal probe (Physitemp) connected to a heat lamp via a temperature controller (TCAT-2DF (Physitemp) reconfigured with a Partlow 1160+ controller). At 20 minutes post-injection, mouse body temperature was elevated by 0.5°C every two minutes until a maximum of 42.5°C was reached. When a GTC seizure occurred (rearing and falling with forelimb clonus), mice were removed from the container and temperature was recorded. If after 3 minutes at 42.5°C no seizure occurred, the mouse was removed and recorded as a seizure-free.
Mice that underwent hyperthermia-induced seizure threshold testing were immediately removed from the assay and anesthetized with isoflurane. Cardiac puncture was performed and whole blood was collected in heparin microtainers (BD365965). Plasma was isolated by centrifugation (9000 x g for 10 minutes at 4°C). Plasma samples (50–100 μL) were spiked with 5 μg/mL prazepam (250 μg/mL in methanol) as an internal standard and vortexed well. Extraction of clobazam and prazepam was achieved by vortex-mixing with diethyl ether (4x volume). The organic layer was isolated following centrifugation at 2000 rpm for 5 minutes and was evaporated by heating at 75°C. The residue was reconstituted in acetonitrile:20 mM potassium phosphate buffer, pH 7 (42:58, v/v).
Clobazam was detected in plasma samples using a HPLC 9 Flexar Binary LC Pump Platform and Flexar UV/Vis Detector (Perkin Elmer, Waltham, MA) with a Polaris C-18A 5 μm column (4.6 x 100 mm; Agilent Technologies, Santa Clara, CA). The mobile phase consisted of acetonitrile and 20 mM potassium phosphate, pH7 (42:58, v/v) with a flow rate of 1.5 mL/min and detection at a wavelength of 228 nm. Quantitative analysis clobazam was performed using a calibration curve (0.05–1.0 μg/mL) by spiking clobazam into blank mouse plasma and preparing as described above.
|
10.1371/journal.pcbi.1002670 | The Regulation of Ant Colony Foraging Activity without Spatial Information | Many dynamical networks, such as the ones that produce the collective behavior of social insects, operate without any central control, instead arising from local interactions among individuals. A well-studied example is the formation of recruitment trails in ant colonies, but many ant species do not use pheromone trails. We present a model of the regulation of foraging by harvester ant (Pogonomyrmex barbatus) colonies. This species forages for scattered seeds that one ant can retrieve on its own, so there is no need for spatial information such as pheromone trails that lead ants to specific locations. Previous work shows that colony foraging activity, the rate at which ants go out to search individually for seeds, is regulated in response to current food availability throughout the colony's foraging area. Ants use the rate of brief antennal contacts inside the nest between foragers returning with food and outgoing foragers available to leave the nest on the next foraging trip. Here we present a feedback-based algorithm that captures the main features of data from field experiments in which the rate of returning foragers was manipulated. The algorithm draws on our finding that the distribution of intervals between successive ants returning to the nest is a Poisson process. We fitted the parameter that estimates the effect of each returning forager on the rate at which outgoing foragers leave the nest. We found that correlations between observed rates of returning foragers and simulated rates of outgoing foragers, using our model, were similar to those in the data. Our simple stochastic model shows how the regulation of ant colony foraging can operate without spatial information, describing a process at the level of individual ants that predicts the overall foraging activity of the colony.
| Social insect colonies operate without any central control. Their collective behavior arises from local interactions among individuals. Here we present a simple stochastic model of the regulation of foraging by harvester ant (Pogonomyrmex barbatus) colonies, which forage for scattered seeds that one ant can retrieve on its own, so there is no need for pheromone trails to specific locations. Previous work shows that colony foraging activity is regulated in response to current food availability, using the rate of brief antennal contacts inside the nest between foragers returning with food and outgoing foragers. Our feedback-based algorithm estimates the effect of each returning forager on the rate at which foragers leave the nest. The model shows how the regulation of ant colony foraging can operate without spatial information, describing a process at the level of individual ants that predicts the overall foraging activity of the colony.
| The fundamental question about the collective behavior of animals is how the actions of individuals add up to the dynamic behavior we observe. In many systems, including animal groups, distributed networks are regulated using feedback based on local interactions. It is not yet clear how the analogies among diverse complex systems reveal general underlying processes [1], [2]. Here we propose a simple stochastic model of collective behavior in ants. Our first goal is to account for the details of a particular system, as a step toward further insight on whether similar processes are at work in othersystems. A second goal of our work is to contribute to the study of collective behavior from the perspective of evolutionary biology. If the outcome of collective behavior is ecologically important, then natural selection can act on variation in that behavior. Modeling the parameters that produce collective behavior can provide the basis for detailed measures of variation among ant colonies.
The best-studied algorithms for collective behavior in animals are those that regulate spatial patterns [3], based on local interactions that influence whether one animal stays close to another [4]. Social insect colonies provide many fascinating examples of collective behavior. There is no central control; no insect directs the behavior of another. Like other social insects, ants use local interactions to regulate colony behavior [5].
The most familiar example of feedback based on local interaction in ants is recruitment to a food source using a pheromone trail. In some ant species, an ant that finds food lays a chemical trail on its way back to the nest. Studies of the algorithms used by ants in forming recruitment trails show that a slight tendency on the part of other ants to move toward the trail leads to the formation of trail systems [6], [7], [8] that can channel ants to the best food source [9] or trace the shortest path toward the food [10]. Ants also use feedback from other forms of interaction, such as brief antennal contact, in recruitment to food [11] and in other spatial decisions. The perception of the local density of digging ants generates branches in nest chambers [12]. The rate of brief antennal contact as a cue to local density [13] is used in spatial decisions, in combination with other information about location, such as the choice of new nest sites by acorn ants [14], [15].
The regulation of activity by a simple stochastic process is characteristic of many biological systems. Local interactions in social insects, like those in other dynamical networks, can regulate the flow or intensity of activity as well as its location or spatial pattern. For example, a social insect colony must adjust the allocation of individuals to various tasks, in response to changing conditions [16]. Various models have been proposed to explain the dynamics of the intensity of activity, or numbers of workers, devoted to colony tasks [e.g. 17], [18], [19].
Here we present a simple stochastic model that explains the process underlying the regulation of foraging activity in harvester ant (Pogonomyrmex barbatus) colonies. Foraging activity, the numbers of ants currently foraging, changes from moment to moment within a foraging period and from day to day. This species does not use pheromone trails to recruit to localized food sources. The ants forage for seeds that are scattered by wind and flooding [20], not distributed in patches, and a single ant can retrieve a seed on its own. The model uses an algorithm based on local interactions among individuals in the form of brief antennal contacts, without any spatial information such as pheromone trails.
Harvester ants searching for food in the desert undergo desiccation, and the ants obtain water from metabolizing the fats in the seeds they eat. Thus a colony must spend water to obtain water, as well as food. The intensity of foraging is regulated from moment to moment, and from day to day, to adjust foraging activity to current food availability, while maintaining sufficient numbers of ants foraging to compete with neighbors for foraging area [21].
A long-term study of the foraging ecology of this species has shown how the moment-to-moment regulation of foraging is accomplished. Regulation depends on feedback from returning foragers, who stimulate the outgoing foragers to leave on the next trip. Forager return rate corresponds to food availability, because foragers almost always continue to search until they find a seed, then immediately bring it back to the nest [22], [23]. The more food is available, the less time foragers spend searching and the more rapidly they return to the nest.
The crucial interactions between returning and outgong foragers take place in a narrow entrance tunnel, 5–10 cm long, that leads to a deeper entrance chamber. Observations with a videoscope show that returning foragers drop their seeds in the tunnel, and then other ants pick up the seeds and take them deeper into the nest. Once the returning forager has dropped its seed, it becomes an outgoing forager, available to go out on its next trip. Experiments using artificial ant mimics coated with extracts of ant cuticular hydrocarbons [24], [25], and experiments manipulating the rate of forager return [26], [27], [28], [29] show that how quickly an outgoing forager leaves on its next trip depends on its interactions with returning foragers. Foraging activity is more closely regulated when foraging rates are high, above a baseline rate at which foragers leave independently of the rate of forager return [29].
We developed a model that takes into account previous work on the regulation of foraging. We compared simulations using the model with new data, from field experiments, that show how the rate of at which outgoing foragers leave the nest changes in response to changes in the rate at which returning foragers go back to the nest.
Experiments manipulating forager return rate were performed in August 2009 and August–September 2010 at the site of a long-term study since 1985 of a population of P. barbatus near Rodeo, New Mexico, USA. In 2009 there were 33 trials in 9 colonies on 8 days, and in 2010 there were 29 trials in 8 colonies on 5 days, of which 4 were the same colonies as in 2009, for a total of 62 trials. All colonies were mature, more than 5 years old (ages determined by yearly census; methods in [21]).
Returning foragers were prevented from returning from the nest in minutes 4–7 of a 20-min observation; methods were the same as in [28], [29]. Rates of returning and outgoing foragers crossing an imaginary line along the trail were measured from video film using an image analysis system developed by Martin Stumpe (http://www.antracks.org). This image analysis system made it possible to measure foraging rates accurately on a shorter timescale than in previous work.
Most colonies use more than one foraging direction on a given day [21]. We filmed all trails and used the combined foraging rates for all trails. Foraging rates were calculated separately for the periods before (0–240 sec), during (240–430 sec), and after (500–1100 sec) the removal of returning foragers, as in previous work [28], [29]. The small interval between 430 and 500 sec allows for the time it takes ants passing the camera on the foraging trail to reach the nest. To correct for differences among trails in the distance between the camera and the nest, we found the average time, out of 5 observations, for an ant to travel back to the nest from the point where the foraging trail was filmed. To adjust for the distance between the camera and the nest, we then subtracted this travel time from the counts for outgoing foragers and added it to counts for returning foragers.
The average error in the accuracy of the image analysis software in counting foraging rate was 7.3%, estimated by comparing 66 counts made by observers from 500 frames (about 17 sec) of 44 video films with counts made by the image analysis software. Most errors were due to an extraneous object or shadow in the film at the point at which ants crossed the imaginary line where ants were counted. There was no bias toward counting more or fewer ants than actually crossed the line.
We first tested whether the return of foragers to the nest could be described as a Poisson process. To determine the distribution of intervals between the arrival of foragers at the nest, we used the data from the period before the removal of returning foragers (60–240 sec) in 39 trials. We calculated the fit with an exponential distribution of the empirical distribution of the interarrival times of the returning foragers, and measured the error using the total variation distance [30] between the two distributions. We found that the return of foragers to the nest can be described as a Poisson process: the distribution of intervals between returning foragers is exponential (Fig. 1) and independent of the spacing between foragers in the previous interval. We found the mean (SE) error for the 39 trials to be 0.056 (0.009) with a better fit at high foraging rates.
We began with a simple linear model in which the rate of outgoing foragers x(t) depends on the the rate of returning foragers plus a constant base rate:
where λb is a baseline rate of outgoing foragers, independent of the rate at which other foragers return; λac sets the number of outgoing foragers per returning forager; and f(t, τ) is the number of returning foragers between times t−τ and t.
For this initial, linear model, the most important parameter in predicting the rate of outgoing foragers is τ. Because this model integrates over all returns in time τ, fitting the model requires us to choose a value of t that gives an equally good fit to the observed data over a range of foraging rates. This is difficult because foraging rates vary greatly among colonies, days, and moment-to-moment changes in the conditions that foragers encounter [5], [27], [29]. We thus elaborated this model into a second one that avoids the integration of instantaneous arrivals over a time interval that is not uniform across different foraging conditions. Moreover, the model below captures the effect of a single returning forager and so explains the process at the level of individual ants, as for example in other non-linear models that describe pheromone trail foraging by ants [7].
The model operates in discrete time: returning foragers are observed in successive and equal time slots. We denote the rate of outgoing foragers as ‘α’, which increases by an amount c>0 for each returning food-bearing forager. Alpha decreases by an amount q>0 for each forager that leaves the nest, because the departure of each outgoing forager decreases the number of outgoing foragers in the queue at the nest entrance available to meet returning foragers. Alpha decays by an amount d>0 during each time slot, to reflect the lack of response to very low interaction rate [22]. Finally, α has a lower bound, α., to reflect the observation that outgoing foragers leave the nest at a fixed low rate even when no foragers return for a while [28], [29].
We assume that arrivals occur at the beginning of time slots and departures occur at the end of time slots. For n = 1, 2, …, let An denote the number of returning food-bearing foragers in the nth time slot, and let Dn denote the number of outgoing foragers leaving the nest. The rate at which ants leave the nest in the nth slot is αn, n = 1, 2, …. We assume that αn≥α>0 for n = 1, 2, …, where α is a parameter. The number of departures at the end of the nth time slot, Dn, was set equal to a Poisson random variable of mean αn. Given the αn, the Dn are statistically independent of each other and of An. The dynamics of αn are described by:
In developing the model, we sought to facilitate its future use to examine variation among colonies within this species, for example in the baseline rate at which ants leave the nest even when no ants return [29], and to take into account three features of the observed behavior of the ants in response to experimental manipulation of foraging rate [28], [29]. First, there is a lag in the recovery of the rate of outgoing foragers in response to recovery of the rate of returning foragers after a decline [28], [29] (Fig. 2). Second, we introduced q, the extent to which each departing forager empties the nest entrance of available foragers, because observations with a videoscope inside the nest show that outgoing ants are crowded in a small tunnel, so that once each outgoing ant departs it takes some time for the next outgoing ant to move to the top of the tunnel where it can meet returning foragers. Third, we include the decay parameter d because experimental results show that the response to returning ants is weaker after some time has elapsed since the last ant returned [24].
We compared the simulated output of the model with the data from field experiments on the response of outgoing foragers to a range of rates of returning foragers (Fig. 2). Using as input the data on the rate of returning foragers, we generated the simulated rate of outgoing foragers, adjusting one parameter and evaluating the resulting match with the observed rate of outgoing foragers.
The model has four parameters: α, c, q and d. We examined the fit between model and data for one parameter, c. We thus fixed α., q and d and varied c. As with any birth-death process, the ratio of c to e determines the distribution of {αn}. We set q to 0.05 to keep the range of values of αn within the range of observed foraging rates (0.15 to 1.2 ants per sec). We set d to 0 for the simulations reported here; however, empirical studies show that d may be an important parameter because it may vary by colony [29], or in response to variation in environmental conditions that could affect the rate of decay of chemical cues such as the cuticular hydrocarbons that ants assess by antennal contact [24]. Similarly, α, the baserate of foraging, was very small, equal to 0.01 ants per second [27].
To choose c for the given values of α, q, and d, we used for each of the 62 experimental trials the data on returning foragers and equations (3) and (4) to generate a simulated rate of outgoing foragers. We swept across values of c from 0.01 to 0.25, and found the relative root-mean-square error (RMSE) [31] between the simulated and observed rates of outgoing foragers in each time interval for each of 200 iterations. Each iteration produces a different output trace because of the independent Poisson random variable generated at equation (4). We chose the c with the lowest average RMSE, over the 200 iterations, between simulated and observed rates of outgoing foragers. The RMSE for the best c for each trial ranged from 0.237 to 4.9, and the mean (SE) RMSE for the 62 values chosen was 0.602 (0.077).
To evaluate how well our estimate of c captured, for a given colony, the effect of each returning forager on the rate of outgoing foragers, we compared the error among runs of the simulation, due to randomness in the departures at equation (4), with the error produced by varying c. To do this we compared the RMSE between observed and simulated rates of outgoing foragers among simulated runs with the RMSE for different values of c among trials of the same colony. We first found for each trial the mean range in RMSE, between observed and simulated rates of outgoing foragers, among 200 iterations of the simulation using the same value of c. To estimate the difference among runs of the simulation, we used each trace of returning foragers to produce 3 different traces of simulated outgoing foragers (A,B,C). We found the RMSE for A vs B and A vs C, and repeated this 200 times per trial at foraging rates ranging, as did the observed rates, from low (0.1 returning ants/sec) to high (1.2 returning ants/sec). We then found the mean RMSE when varying values of the best c among trials for the same colony. To do this we chose arbitrarily, for each of the 13 colonies, 2 of the 3 or 4 trials. For those 2 trials we found the average of the values of c, then found the RMSE as above [31] for simulated values of rates of outgoing foragers for each of the 1 or 2 remaining trials for that colony in the same year. We deleted one of the remaining trials in cases when foraging rates were much lower than for other trials with the same colony.
The results of this comparison indicate that our estimate of best c generates the observed outgoing forager rate within the same range of error as the error generated by randomness in forager departures. The mean (SE) change in RMSE among simulated runs was 13% (0.0495) at low foraging rates and 2.6% (0.01) at high foraging rates. The mean (SE) change in RMSE obtained by varying u among trials of the same colony was 15.5% (0.019).
We examined whether the correlation between rates of returning and outgoing foragers was at least as high in the simulation as in the data. To do this we compared the correlation between observed rates of returning and simulated rates of outgoing foragers, using the best value of c, with those between observed rates of returning and observed rates of outgoing foragers. We calculated the correlation between the observed rates of returning foragers and the simulated rate of outgoing foragers, in all 62 trials, by applying a moving-average rect filter [32] with a radius of 25 time slots, and then finding the empirical correlation coefficent between the smoothed traces [31]. We calculated in the same way the correlation between observed rates of returning and outgoing foragers. The simulated rates of outgoing foragers led to correlation coefficients higher than those between observed rates of returning and outgoing foragers (t-test, t = 8.98, p<0.001; Fig. 3).
Because previous work showed that the rate of outgoing foragers tracks the rate of forager return more closely when foraging rates are high, we examined how the correlation of rates of returning and outgoing foragers varies with foraging rate. The magnitude of the correlation coefficient between observed returning and simulated outgoing foragers increased with the mean rate of returning foragers (Spearmann's rank correlation, n = 62, z = −4.04, p = 0.0001, blue points in Fig. 3). The magnitude of the correlation coefficient for the observed returning and outgoing rates did not increase significantly with the mean rate of returning foragers (Spearmann's rank correlation, n = 62, z = −1.34, p = 0.18, red points in Fig. 3).
Our model provides a way to evaluate quantitatively the response of individuals to local interactions. It explains how ant colonies regulate foraging activity from moment to moment, in response to current food availability, without any central control or any spatial information on the location of food. Changes in a colony's foraging activity from moment to moment, and from day to day, show a predictable response to changes in forager return rate, despite considerable stochasticity (Fig. 2). Our results show that much of these changes in colony foraging activity can be explained by the effect of each returning forager on the probability that outgoing foragers leave the nest to search for food.
The model is analyzable because the distribution of returning foragers is well-approximated by a Poisson process at high foraging rates. Like cars on highways, returning foragers travel at different velocities and overtake each other; as [33] shows, this produces a Poisson process.
The simulated data provided by our model capture many of the features of a rich body of empirical results from a long-term study of the foraging behavior of harvester ant colonies. The model provides a simulated rate of outgoing foragers, in response to the rate of returning foragers, that is reasonably similar to the observed rate (Fig. 2), producing a close correlation between the rates of returning and outgoing foragers.
Another similarity between the model and observation is in the contrast between colony behavior when food availability is high, so that foragers find food quickly and the rate of forager return is high, and colony behavior when food availability is low, so that foragers find food slowly and the rate of forager return is low. Previous work on harvester ant foraging showed that rates of outgoing foragers are more closely adjusted to rates of returning foragers when foraging rates are high [29]. The same was true of our model. For example, Fig. 2 shows the data for two representative cases. When foraging rates are high (Fig. 2A), on a day when food availability is high and foragers find it quickly, the rate of returning and outgoing foragers is more closely matched than when foraging rates are low (Fig. 2B), so that foragers find food more slowly and return less frequently. The closer fit between outgoing and returning foragers at high foraging rates occurs because the range of values of αn tends to be much smaller and closer to α at low foraging rates than at high foraging rates, and this causes Dn to be close to zero and, hence, less correlated with An.
However, the simulation generally produces a closer correlation between the rates of returning and outgoing foragers than is observed in the data. The correlation coefficients for the observed rate of returning foragers and the simulated rate of outgoing foragers are higher than those for the observed rate of returning foragers with the observed rate of outgoing foragers (Fig. 3). In addition, the coefficient of correlation with the observed rate of returning foragers increased significantly with foraging rate for the simulated rate of outgoing foragers but does not increase significantly for the observed rate of outgoing foragers (Fig. 3).
The higher correlation in the simulation than in the data occurs because our model (equations 3 and 4) does not capture all of the factors that produce the actual rate at which ants leave the nest. For example, the structure of each nest probably affects the flow of ants in and out of the nest entrance, which in turn may affect the rate of interaction between outgoing and returning ants. A question for future work is whether nonlinear effects of nest structure influence the relation between overall foraging rate and the correlation of the rates of returning and outgoing foragers. Our model assumes the same relation between the rate of incoming and rate of outgoing foragers for all nests, and thus does not take into account the local influence of nest structure. Weather conditions also influence foraging activity, leading to day-to-day fluctuations in the foraging activity of a given colony [28].
The process described here is analogous to those operating in many other distributed networks, from computer networks to neural integrators, that regulate activity through the rate of interaction [34], [35]. Further work is needed to determine the details of the correspondence among these analogous systems; for example, in this system, the Poisson distribution of returning foragers is crucial.
The model presented here contributes to the study of the evolution of collective behavior in harvester ants, because it can be used to guide empirical measurement of differences among colonies in the regulation of foraging [29], by examining whether colonies tend to show characteristic parameter values. Harvester ant colonies differ in foraging behavior, and such differences persist from year to year as the colony grows older [21], [29]. Heritable variation among colonies in ecological relations, such as the regulation of foraging, is the source of variation in fitness [36]. Future work will examine differences among colonies in the response to interactions of returning and outgoing foragers. Small differences in the ants' response to local interactions may lead to ecologically important differences among colonies that shape the evolution of collective behavior.
|
10.1371/journal.pgen.1006430 | CLOCKWORK ORANGE Enhances PERIOD Mediated Rhythms in Transcriptional Repression by Antagonizing E-box Binding by CLOCK-CYCLE | The Drosophila circadian oscillator controls daily rhythms in physiology, metabolism and behavior via transcriptional feedback loops. CLOCK-CYCLE (CLK-CYC) heterodimers initiate feedback loop function by binding E-box elements to activate per and tim transcription. PER-TIM heterodimers then accumulate, bind CLK-CYC to inhibit transcription, and are ultimately degraded to enable the next round of transcription. The timing of transcriptional events in this feedback loop coincide with, and are controlled by, rhythms in CLK-CYC binding to E-boxes. PER rhythmically binds CLK-CYC to initiate transcriptional repression, and subsequently promotes the removal of CLK-CYC from E-boxes. However, little is known about the mechanism by which CLK-CYC is removed from DNA. Previous studies demonstrated that the transcription repressor CLOCKWORK ORANGE (CWO) contributes to core feedback loop function by repressing per and tim transcription in cultured S2 cells and in flies. Here we show that CWO rhythmically binds E-boxes upstream of core clock genes in a reciprocal manner to CLK, thereby promoting PER-dependent removal of CLK-CYC from E-boxes, and maintaining repression until PER is degraded and CLK-CYC displaces CWO from E-boxes to initiate transcription. These results suggest a model in which CWO co-represses CLK-CYC transcriptional activity in conjunction with PER by competing for E-box binding once CLK-CYC-PER complexes have formed. Given that CWO orthologs DEC1 and DEC2 also target E-boxes bound by CLOCK-BMAL1, a similar mechanism may operate in the mammalian clock.
| Circadian clocks control daily rhythms in animal, plant and fungal physiology, metabolism and behavior via transcriptional feedback loops. In Drosophila, the CLOCK-CYCLE (CLK-CYC) activator complex binds E-box regulatory sequences to initiate transcription of hundreds of effector genes including their own repressors, PERIOD (PER) and TIMELESS (TIM), which feed back to repress CLK-CYC until they are degraded, thus allowing another cycle of CLK-CYC activation. Although the repression process is critical for the stability and accuracy of circadian timekeeping, how PER-TIM complexes maintain a transcriptionally repressed state for many hours is not well understood. Here we demonstrate that the transcription factor CLOCKWORK ORANGE (CWO) antagonizes CLK-CYC E-box binding, thus enhancing the removal of CLK-CYC from E-boxes to maintain transcriptional repression. This process requires PER, which suggests that PER-TIM and CWO cooperate to maintain a transcriptionally repressed state by removing CLK-CYC from E-boxes. These results demonstrate that PER-TIM requires CWO to effectively repress circadian transcription, and given that circadian transcriptional regulators are well conserved, this mechanism may function to repress transcription in other animals including humans.
| Almost all organisms from Cyanobacteria to humans have internal circadian clocks that drive daily rhythms in physiology, metabolism and behavior, thereby synchronizing internal processes with the external environment. In eukaryotes, the circadian clock keeps time via one or more transcriptional feedback loops [1]. In Drosophila, a heterodimer formed by CLOCK (CLK) and CYCLE (CYC) binds E-box sequence activates transcription to initiate clock function. In the core loop, CLK-CYC activates period (per) and timeless (tim) transcription during mid-day, effecting a rise in per and tim mRNA levels that peaks during the early evening. PER and TIM proteins then accumulate, form a dimer, and move into the nucleus to bind CLK-CYC during the night, thereby inhibiting their transcriptional activity until PER and TIM are degraded early in the morning [2,3]. Another interlocked transcriptional feedback loop is also regulated by the core feedback loop. In this loop, CLK-CYC activates transcription of vrille (vri) and PAR-domain protein 1ɛ (Pdp1ɛ), which bind D-boxes to repress and activate transcription, respectively, and drive RNA cycling of Clk and other output genes in the opposite phase as per, tim, vri and Pdp1ɛ [4–6].
PER was previously found inhibit CLK-CYC binding to E-boxes in vitro [7], which suggests that the rhythmic transcription of CLK target genes are mediated by PER-dependent rhythms in E-box binding by CLK-CYC. Chromatin immunoprecipitation (ChIP) experiments using fly heads support this model, showing that CLK-CYC rhythmically bind E-boxes in the per circadian regulatory sequence (CRS) and the tim upstream sequence [8]. However, the mechanism by which CLK-CYC heterodimers are removed from E-boxes during repression is not well understood. PER is required for the rhythmic binding of CLK complexes, as CLK constantly binds to per and tim promoters in per01 flies [8], indicating that PER inhibits transcription by removing CLK-CYC from E-boxes. Interestingly, co-expression of another transcription factor, CLOCKWORK ORANGE (CWO), strongly enhanced PER-mediated repression in cultured Drosophila Schneider 2 (S2) cells [9], suggesting that PER is unable to efficiently remove CLK from DNA in the absence of other transcription repressors.
Previous studies demonstrated that CWO, a basic helix-loop-helix (bHLH)-ORANGE transcriptional factor [10], is a direct target of CLK-CYC [9,11,12]. In Drosophila Schneider 2 (S2) cells, overexpression of CWO reduces the basal transcription of per, tim, vri and Pdp1ɛ promoter-driven luciferase reporter genes [9,12,13]. Furthermore, in the presence of PER, CWO repress CLK mediated transcription 5–10 fold in S2 cells, indicating that CWO is a strong transcription repressor that can cooperate with PER to repress CLK-CYC mediated transcription [9]. In cwo mutants or cwo RNAi knockdown flies, the levels of per, tim, vri and Pdp1ɛ mRNAs are increased during the early to mid-morning [9,12]. These results suggest that CWO co-represses CLK-CYC activity along with PER during the end of a cycle [9,12]. However, the mechanism through which CWO represses CLK-CYC mediated gene transcription remains unknown.
In this study we demonstrate that CWO and CLK bind core clock gene E-boxes in a reciprocal pattern across the circadian cycle in vivo, which suggests that CWO competes with CLK to bind E-boxes. We also show that CWO acts to decrease CLK binding to tim E-boxes during early morning, when PER binds CLK-CYC to reduce its binding to DNA [8], but not during early night when CLK-CYC strongly binds E-boxes in the absence of PER. These results suggest a model for CWO function where CWO has low DNA binding affinity compared to CLK-CYC complexes during the activation phase, but has higher affinity compared to CLK-CYC-PER complexes, and is thus capable of removing CLK-CYC-PER complexes from E-boxes to consolidate and maintain repression. Constant high CWO binding to the tim promoter in Clkout flies (i.e. comparable to binding at ZT2 in wild-type) and constant low CWO binding in per01 flies (i.e. comparable to binding at ZT14 in wild-type) supports our model for CWO repression. As a whole, these results suggest that CWO co-represses CLK-CYC activity with PER by competing with CLK-CYC-PER complexes for E-box binding, therefore promoting the transition to off-DNA repression.
Earlier studies demonstrated that cwo mRNA cycles in phase with per, tim, vri, and Pdp1, but with a higher basal level, and thus lower amplitude [9,12–14]. To determine whether CWO protein levels also cycle, western analysis was carried out using head extracts from wild-type flies collected every 4 hours in a 12-h light/12-h dark (LD) cycle. We find that the levels of CWO do not change throughout an LD cycle (Fig 1), consistent with previous results [15]. Given that cwo mRNA levels cycle, it is possible that constant CWO levels result from post-transcriptional regulation or a long half-life.
CWO contains a bHLH domain, a structural motif that characterizes a family of E-box binding transcription factors [16–19], which suggests that CWO may regulate CLK-CYC target gene transcription via E-box binding. Previous ChIP-on-chip and gel-shift analyses in S2 cells demonstrated that CWO specifically binds to the E-box of core clock genes [12,13], however it is still unknown whether CWO binds those core clock genes in vivo, and whether the binding intensity changes throughout the day. To test these possibilities, ChIP assays were carried out on wild-type flies collected in the early morning (ZT2) and in the early night (ZT14) using CWO and CLK antisera. Fragments containing upstream E-boxes from tim, per, Pdp1 and vri, which are necessary for high-amplitude mRNA cycling in vitro or in vivo [4,5,20–25], were amplified from the immunoprecipitates and then quantified. In CWO immunoprecipitates, the tim, vri and Pdp1 E-box containing fragments were two to threefold more abundant at ZT2 than at ZT14 (Fig 2A), suggesting that CWO binding is time-dependent, though the dynamic binding of CWO on the per E-box fragment is less robust than the others. Importantly, this temporal binding pattern is antiphase to CLK binding, as CLK shows high binding intensity during the night at ZT14 and low binding during the daytime at ZT2 (Fig 2B), consistent with previous results [8,11].
The reciprocal binding pattern of CLK and CWO implies that these transcription factors compete for E-box binding. If so, both CLK and CWO must occupy the same E-boxes. To test this possibility, we determined how mutating E-boxes upstream of tim affected CLK and CWO binding. The circadian enhancer upstream of tim is comprised of two tandem E-boxes that are spaced seven nucleotides apart [24,26], a structure that is conserved among core clock genes in various species [27]. Both of these E-boxes were indispensable for tim mRNA expression in S2 cells [24], suggesting that these tandem E-box motifs are binding sites for both CLK and CWO. To determine if this is the case, a series of 136bp fragments from the tim promoter containing an E-box1 (E1) mutant (mE1-E2), an E-box 2 (E2) mutant (E1-mE2), an E1 and E2 double mutant (mE1-mE2) or a control with wild-type E-boxes (E1-E2) were generated, inserted into the pHPdestGFP vector [28], and targeted to the attP18 genomic site (Fig 3A).
To confirm that this promoter fragment is sufficient to drive rhythmic expression, we carried out quantitative reverse transcription-PCR (qRT-PCR) to monitor GFP mRNA levels in flies collected every 4-h during an LD cycle. Quantification of GFP mRNA levels in flies with WT tim promoter shows a ~10-fold diurnal rhythm with a peak at ZT14 and a trough at ZT2 to ZT6 (S1 Fig), consistent with timing and amplitude of per and tim mRNA cycling in wild-type flies [29,30]. However, even at the normal tim mRNA peak (ZT14), mE1-E2, E1-mE2 and mE1-mE2 flies express little or no eGFP mRNA (S1 Fig), indicating that both E1 and E2 are indispensable for expression of tim mRNA in vivo. This result is consistent with previous tim-luciferase reporter results in S2 cells [24].
We next carried out ChIP assays using CWO and CLK antisera on the same fly strains to test whether E1 and E2 are required for CWO and CLK binding. At ZT2, when CWO strongly binds to the tim promoter, CWO binding intensity was drastically reduced in mE1-E2, E1-mE2 and mE1-mE2 flies compared to WT (Fig 3B). Likewise, CLK binding intensity was drastically reduced in mE1-E2, E1-mE2 and mE1-mE2 flies compared to WT at ZT14, when CLK binding is strongest (Fig 3B). These results indicate that both E1 and E2 are indispensable for both CWO and CLK binding to the tim circadian enhancer. Given that CWO specifically targets E-boxes in S2 cells by Gel-shift analyses [13], we conclude that both CLK and CWO bind intact tandem E1-E2 motifs in vivo. In mice, CLK-BMAL1 dimers cooperatively bind tandem E-boxes in vitro [27,31], and this may be the case for CWO given the requirement for both E1 and E2 E-boxes.
Previous studies showed that increasing the level of CWO expression reduces per, tim, vri and Pdp1ɛ mRNA levels in S2 cells and that their trough mRNA levels are higher in cwo mutant or knockdown flies, indicating that CWO acts to repress CLK-mediated gene transcription in vitro and in vivo [9,12–14]. Given that CWO and CLK bind to the same E-box motif, we wondered whether CWO represses CLK-mediated transcription by inhibiting CLK binding. To test this possibility, ChIP assays were carried out using CLK antiserum on wild-type and cwo5703 flies at the trough (ZT2) and peak (ZT14) times of CLK-CYC target gene transcription and mRNA abundance in LD. Although cwo5703 mutants lengthen the period of activity rhythms by 2–3h in DD [9,13], the peak and trough phases of CLK-CYC target gene transcription and mRNA abundance are comparable in cwo5703 mutants and wild-type flies in LD [9,13]. We find that CLK binds tim E-boxes with a robust rhythm in wild-type flies and a lower amplitude rhythm in the cwo5703 mutant (Fig 4A). However, the intensity of CLK binding in cwo5703 is significantly increased at ZT2 compared to wild-type, indicating that CWO acts to reduce CLK-CYC binding at the trough of its binding cycle (Fig 4B). Given that CWO strongly binds tim E-boxes at ZT2 (Fig 2A), we propose that CWO inhibits CLK-CYC binding during the repression phase by antagonizing PER-CLK-CYC complexes to maintain off-DNA repression. There was no significant difference in CLK binding between cwo5703 and wild-type at ZT14 (Fig 4B), despite decreased peak levels of per, tim, vri and Pdp1ɛ mRNA at ZT14 in cwo mutant and RNAi knockdown flies [9,12–14], suggesting that CWO has little impact on CLK-CYC binding in the absence of PER.
Given that CWO suppresses CLK binding at ZT2 in the early morning but not at ZT14 during the early evening (Fig 4B), it is possible that PER is necessary for CWO to antagonize CLK E-box binding since PER accumulates to high levels in the nucleus around dawn and is at low levels in the cytoplasm around dusk [32]. Indeed, our results support a model developed previously to explain cooperation between CWO and PER to repress CLK-CYC mediated transcription in S2 cells [9]. In this model, CWO is proposed to compete with CLK-CYC heterodimers for E-box binding only when PER binds CLK-CYC, thereby reducing their affinity for E-box binding. To test this model, we performed ChIP assays using CWO antiserum on wild-type, Clkout and per01 flies collected at ZT2 and ZT14 in LD. In Clkout flies, which necessarily lack CLK-CYC heterodimers [33], CWO is bound to tim E-boxes at both ZT2 and ZT14 with binding signals comparable to the strong CWO binding in wild-type flies at ZT14 (Fig 5A). In contrast, in per01 flies, which lack PER-dependent repression of CLK-CYC activation [34], low binding signals of CWO were detected at ZT2 and ZT14, indicating that PER is indeed required for CWO to bind E-boxes (Fig 5A). Moreover, CWO binding was significantly increased in Clkout versus wild-type flies at ZT14, indicating that CLK-CYC binding at ZT14 reduces CWO binding. Likewise, a significant increase in CWO binding was also seen in wild-type versus per01 flies at ZT2, indicating that PER enhances CWO binding (Fig 5A).
To determine whether differences in CWO binding in Clkout and per01 flies were due to differences in CWO protein levels, we carried out western analysis using head extracts from these mutants collected at ZT2 and ZT14. Since cwo transcription is regulated in part by the transcriptional feedback loop, CWO protein levels are slightly lower in Clkout flies and slightly higher in per01 flies (Fig 5B and 5C). However, the lower levels of CWO in Clkout resulted in higher E-box binding, and higher CWO protein levels in per01 resulted in lower E-box binding. This result suggests that the differences in CWO-E-box binding are not due to altered CWO protein levels, but due to the relative DNA binding affinities of CWO and CLK in these mutants. These results, taken together, strongly support and extend the model described by Kadener et al., 2007, for CWO binding as it relates to CLK-CYC repression. When CLK-CYC targets are activated, CLK-CYC binds DNA with higher affinity than CWO, thus CLK binding is not altered in the presence or absence of CWO. When CLK-CYC targets are repressed, PER binds CLK-CYC complexes and decreases their DNA binding affinity, thereby favoring CWO binding to E-boxes and enhancing PER mediated removal of CLK-CYC-PER complexes from the DNA (Fig 6). Although we can’t exclude the possibility that PER enables CWO E-box binding independent of its interaction with CLK-CYC, the available evidence strongly supports the model proposed.
Rhythmic binding of CLK-CYC to E-boxes is essential for rhythmic transcription of the core circadian oscillator genes per and tim in Drosophila. CLK-CYC bind E-boxes upstream of per and tim in the late day and early night to activate transcription; and is released from these binding sites during late night [8,35,36]. Previous work demonstrated that CLK constitutively binds per and tim E-boxes in per01 flies, indicating that PER is essential for rhythmic binding of CLK-CYC, and is key to removing CLK-CYC from E-boxes [8]. In this study we report that CWO also contributes to removing CLK-CYC from E-boxes. In cwo5703 mutant flies, CLK binding intensity is significantly increased at the trough of its binding cycle, suggesting that repression is incomplete in the absence of CWO (Fig 4).
We find that CWO and CLK bind E-boxes upstream of tim in a reciprocal manner during a daily cycle, and that CLK shows significantly increased binding intensity at the trough of its binding cycle in cwo mutant flies, indicating that CWO acts to antagonize CLK-CYC binding. Given that both CWO and CLK are constitutively expressed (Fig 1; [8]), we believe that the key driver for the transition between dynamic CLK-CYC and CWO binding is the accumulation of PER, which alters the relative affinity of E-box binding by CLK-CYC. CWO shows low levels of tim E-box binding in per01 flies, in which CLK-CYC constantly bind E-boxes, but shows high levels of tim E-box binding in Clkout flies that lack CLK expression and E-box occupancy. These results suggest that CWO E-box binding affinity is lower than the CLK-CYC heterodimer and higher than the CLK-CYC-PER complex, which could account for the PER-dependent rhythms in CLK-CYC and CWO binding (Fig 6). During late day and early night, CLK-CYC binds E-boxes to activate transcription in the presence of CWO because CLK-CYC has higher DNA binding affinity. PER starts to accumulate in the nucleus during the night and interacts with CLK-CYC, which decreases CLK-CYC DNA interaction via reduced DNA binding affinity. Consequently, CWO displaces CLK-CYC-PER from E-boxes by binding with comparatively higher affinity. Once CLK-CYC-PER is removed, CWO occupancy on E-boxes prohibits CLK-CYC-PER from re-binding, thus maintaining transcriptional repression (Fig 6).
Unlike the constitutive CLK-CYC E-box binding in per01 flies [8], CLK-CYC binding is rhythmic in cwo5703 flies, but with a dampened amplitude due to elevated CLK binding at the trough (Fig 4A). This low amplitude rhythm in CLK binding may explain why a large proportion of cwo5703 flies show long period rhythms rather than losing rhythmicity entirely like per01 mutants [9,12–14]. We speculate that the long period phenotype is caused in part by a prolonged repression process. Based on the current model for repression of CLK-CYC transcription, PER-TIM complexes first bind CLK-CYC, thereby removing CLK-CYC from the E-boxes and inhibiting per and tim transcription, then PER and TIM degradation enables CLK-CYC binding to start another cycle of transcription [3]. Both of these steps could be delayed in a cwo mutant. In the absence of CWO it takes longer to remove CLK-CYC from the DNA; PER alone can repress CLK-CYC binding to some degree, but CLK-CYC-PER complexes still weakly bind E-boxes if CWO is absent, thus reducing CLK-CYC repression compared to wild-type flies. The outcome of incomplete repression of CLK-CYC E-box binding would be an increase in the trough levels of per and tim mRNAs, which is exactly what was observed in cwo mutant and RNAi knockdown strains [9,12–14]. Higher per and tim mRNA levels would in turn increase PER and TIM expression during the repression phase [14]. Higher levels of PER and TIM would not repress CLK-CYC binding efficiently in the absence of CWO, but would take longer to be degraded, thereby delaying the next cycle of transcriptional activation.
In addition to the increased trough levels of core clock gene mRNAs in cwo mutant and RNAi knockdown flies, the peak levels of these mRNAs are lower, particularly during DD [9,12–14]. Decreasing per mRNA levels also lengthen circadian period [37], thus making it difficult to determine the extent to which a lower mRNA peak or increased mRNA trough contributes to period lengthening in cwo mutant and RNAi knockdown flies. CLK binding at the peak of transcription is not significantly lower in cwo5073 than wild-type during LD (Fig 4B), which argues that CWO enhances CLK-CYC transcriptional activity independent of CLK-CYC E-box binding. Additional experiments will be needed to decipher the mechanism underlying this CWO dependent increase in CLK-CYC transcription.
PER dependent repression of CLK-CYC transcription is thought to occur in two stages. First, PER is recruited to circadian promoters by interacting with CLK to form PER-CLK-CYC complexes “on-DNA”, which inhibit CLK-CYC dependent transcription via an unknown mechanism. Subsequently, a decrease in the DNA binding affinity of PER-CLK-CYC complexes results in their release from DNA to initiate ‘‘off-DNA” phase of repression [35]. According to our model, CWO is critical for the transition to, and maintenance of, off-DNA repression. When PER-CLK-CYC complexes with low DNA affinity are formed, CWO promotes off-DNA repression by competing with CLK-CYC-PER complex for E-box binding. CWO occupancy on E-boxes then prevents PER-CLK-CYC from re-binding, thereby maintaining off-DNA repression.
In mammals, a similar pattern of antagonistic binding on E-boxes between transcription factors was recently reported; USF1 and a mutant form of CLOCK, CLOCKΔ19, bind to the same tandem E-boxes in a reciprocal manner. Wild-type CLOCK-BMAL1 complex binds E-boxes with much higher affinity than USF1, but CLOCKΔ19-BMAL1 binds E-boxes with a similar affinity to USF1, thus allowing USF1 to bind E-boxes [31]. Although this competitive binding is not thought to impact feedback loop function under normal circumstances, it demonstrates that other transcription factors can out-compete CLOCK-BMAL1 for E-box binding if the DNA binding affinity of CLOCK-BMAL1 is reduced. In this case CLOCK-BMAL1 binding is compromised by the ClockΔ19 mutation, but other mechanisms such as interactions with repressors and protein modifications could also reduce the binding affinity of CLOCK-BMAL1 or its orthologs.
As in Drosophila, rhythmic binding of CLOCK-BMAL1 to E-boxes drives circadian transcription in mammals (reviewed in [38]). Recent ChIP-seq analyses in mouse liver revealed time-dependent binding of CLOCK, BMAL1 and key negative feedback components including PER1, PER2, CRY1 and CRY2 [27,39–41]. The mechanism underlying the dynamic DNA occupancy of these transcription factors is not known, but previous work shows that the PER2-CLOCK interaction is required to initiate repression of CLOCK-BMAL1 dependent transcription [42], which suggests that CLOCK-BMAL1 may be removed from E-boxes by the same mechanism as CLK-CYC in Drosophila. A recent genome-wide nucleosome analysis in mouse liver revealed that rhythmic E-box binding by CLOCK-BMAL1 removes nucleosomes [43]. However, despite rhythmic CLOCK-BMAL1 binding, nucleosome occupancy on E-boxes is always well below surrounding sequences, even in Bmal1-/- mutant livers [43]. This result indicates that chromatin at CLOCK-BMAL1 target sites is not closed even when there is no CLOCK-BMAL1 binding, suggesting that other transcription factors may occupy these E-boxes when CLOCK-BMAL1 is absent. These results, taken together, suggest that rhythms in activator binding may be controlled by a common mechanism in Drosophila and mammals.
The mammalian orthologs of CWO, called DEC1 and DEC2 (and also SHARP2 and SHARP1, respectively), suppress CLOCK-BMAL1-induced activation [44–50]. Gel mobility shift and ChIP assays in vitro revealed that both DEC1 and DEC2 bind to E-box motifs targeted by CLK-BMAL1 [45–49], and the DNA-binding domain is required for DEC1 to regulate CLK-BMAL1-induced transactivation [48]. In addition, DEC1/2 shows synergistic activity to PER1 in the regulation of clock gene mRNA levels in the SCN, as exemplified by significant changes in the period of circadian activity rhythms when null mutants for Dec1, Dec2 or both Dec1 and Dec2 are combined with that for Per1 [44]. In contrast to the constant levels of CWO, DEC1 protein is rhythmically expressed in mouse liver, where DEC1 levels are high when PER-CRY complexes repress CLK-BLMAL1 transcription [51]. Taken together, these results raise the possibility that DEC1 and DEC2 may be a functional counterpart of CWO in competing with CLOCK-BMAL1 for E-box binding to repress CLOCK-BMAL1-mediated transcription.
DNA fragments containing wild-type or mutant E-boxes from the upstream tim circadian enhancer were used to construct GFP-reporter transgenes. These 136bp fragments extend from -578 to -714 relative to the tim transcription start site, and contain “E1-E2” E-box motifs that are wild-type (E1-E2), E1 mutant (mE1-E2), E2 mutant (E1-mE2) or E1-E2 mutant (mE1-mE2). These wild-type and mutant E-box fragments were generated by PCR amplification using the following primer sets: E1-E2, 5’-CACCTTTGGCAAATAAACGTGCGGCA-3’ and 5’-TGCCGGCGTTTGTGCGAA-3’; mE1-E2, 5’-CACCTTTGGCAAATAAACGTGCGGCACGTTGTGATTAAGATCTAGCCGAT-3’ and 5’-TGCCGGCGTTTGTGCGAA-3’; E1-mE2, 5’-CACCTTTGGCAAATAAGATCTCGGAGATTTGTGATTACACGTGAGCCGAT-3’ and 5’-TGCCGGCGTTTGTGCGAA-3’; mE1-mE2, 5’-CACCTTTGGCAAATAAGATCTCGGAGATTTGTGATTAAGATCTAGCCGAT-3’ and 5’-TGCCGGCGTTTGTGCGAA-3’. The PCR products were inserted into the pENTR/D-TOPO vector using pENTR Directional TOPO cloning kit (Invitrogen), and then subcloned into the pHPdesteGFP vector, which expresses Green Fluorescent Protein (GFP) according to the enhancer sequence inserted [28], using Gateway LR-Clonase System (Invitrogen). The nucleotide sequences of all transgenes were confirmed by sequencing. The resulting transgenes were injected into embryos (BestGene) for recombination into the attp18 genomic site via PhiC31-mediated transgenesis to yield tim circadian enhancer GFP (tim-CEG) flies [52–54].
Flies were entrained in a 12-h light/12-h dark (LD) incubator for at least 3 days, collected at the indicated time points, and frozen. Isolated frozen fly heads were homogenized in radioimmunoprecipitation assay (RIPA) buffer (20 mM Tris at pH 7.5, 150 mM NaCl, 1 mM EDTA, 0.05 mM EGTA, 10% glycerol, 1% Triton X-100, 0.4% sodium deoxycholate) containing 0.5 mM PMSF (phenylmethylsulfonyl fluoride), 10 μg/ml aprotinin, 10 μg/ml leupeptin, 2 μg/ml pepstatin A, 1 mM Na3VO4, and 1 mM NaF. This homogenate was sonicated 3 to 5 times for 10 s each time, using a Misonix XL2000 model sonicator at a setting of 3 and then centrifuged at 20,000 g for 10 min. The supernatant was collected as RIPA S extract, and protein concentration was determined by the Bradford assay. Equal amounts of RIPA S extract were run, transferred, and probed with guinea pig anti-CWO (GP-27), 1:5,000 and mouse anti-beta-actin (Abcom), 1:20,000. Horseradish peroxidase-conjugated secondary antibodies (Sigma) against guinea pig and mouse were diluted 1:5,000. Immunoblots were visualized using ECL plus (GE) reagent. Protein levels were measured by placing a rectangle of the same size over each CWO, ß-Actin or non-specific (NS) protein band on films used to visualize the immunoblots, and quantifying the signal within each rectangle via densitometric analysis using the ImageJ program. The levels of CWO were calculated as a CWO:ß-Actin or CWO:NS ratio, and CWO abundance at each time point was plotted relative to wild-type at ZT2.
Chromatin IP (ChIP) assays and qPCR quantification were performed as previously described [55]. CLK and CWO binding to E-boxes in the circadian enhancers upstream of tim, per, vri, and Pdp1 in wild-type flies and the circadian enhancer in tim-CEG flies were first quantified via qPCR, and the resulting values were corrected for nonspecific binding to an intergenic region on chromosome 3R (nucleotides 29576172 to 29576303). The primers used for qPCR were as follows: for tim E-boxes, 5’-ACACTGACCGAAACACCCACTC-3’ and 5’-GCGGCACGTTGTGATTACACG-3’; for per E-boxes, 5’-GGGTGAGTAATGCCGTTGCGAAAT-3’ and 5’-ATTTGCTGGCCAAGTCACGCAGTT-3’; for vri E-boxes, 5’-CTGGTGCCTCACATTCCACG-3’ and 5’- CAGCAGTCAAGTTATAGCAGCGC-3’; for Pdp1 E-boxes, 5’-GCACTCTCATTCTCTCTGTCGC-3’ and 5’-ACTTGGGGGACTGGAACTG-3’; for tim-CEG, 5’-GCCCCCTTCACCTTTGGCAAATA-3’ and 5’-TACAAGAAAGCTGGGTCGGCG-3’; and for the intergenic region, 5’-CAGGAGTCGVAGGACCAACC-3’ and 5’-GTCCTGAGAGGCTGAGAGGC-3’. PCR amplification using each pair of primers produced a single band of the expected size. The tim-CEG primers target vector sequences that flank the genomic tim E-box insert, and thus do not amplify endogenous tim genomic sequences.
Quantitative RT-PCR was performed as described [55,56], with some modifications, to measure GFP mRNA levels. Total RNA was isolated from frozen fly heads using Trizol (Invitrogen), and treated with a Turbo DNase DNA-free kit (Ambion) to eliminate genomic DNA contamination. DNA-free total RNA (1.0 μg) was reverse transcribed using oligo(dT) 12–28 primers (Invitrogen) and Superscript II (Invitrogen). The reverse transcription (RT) product was amplified with SsoFast qPCR Supermix (Bio-Rad) in a Bio-Rad CFX96 Real-Time PCR System using primers to GFP (5’-TACGGCAAGCTGACCCTGAAGT-3’ and 5’-CGCACCATCTTCTTCAAGGACG-3’) and ribosomal protein 49 (rp49) (5’-TACAGGCCCAAGATCGTGAA-3’ and 5’-GCACTCTGTTGTCGATACCC-3’). For each sample, mRNA quantity was determined using the standard curve for each gene analyzed. To determine the relative levels of GFP mRNA over a diurnal cycle, GFP mRNA levels were divided by rp49 mRNA levels for each time point and plotted as the GFP/rp49 mRNA ratio. To quantify GFP mRNA in different tim-CEG strains at the wild-type (E1-E2) peak, GFP/rp49 values were normalized to the E1-E2 value at ZT14.
|
10.1371/journal.pcbi.1006290 | Two critical positions in zinc finger domains are heavily mutated in three human cancer types | A major goal of cancer genomics is to identify somatic mutations that play a role in tumor initiation or progression. Somatic mutations within transcription factors are of particular interest, as gene expression dysregulation is widespread in cancers. The substantial gene expression variation evident across tumors suggests that numerous regulatory factors are likely to be involved and that somatic mutations within them may not occur at high frequencies across patient cohorts, thereby complicating efforts to uncover which ones are cancer-relevant. Here we analyze somatic mutations within the largest family of human transcription factors, namely those that bind DNA via Cys2His2 zinc finger domains. Specifically, to hone in on important mutations within these genes, we aggregated somatic mutations across all of them by their positions within Cys2His2 zinc finger domains. Remarkably, we found that for three classes of cancers profiled by The Cancer Genome Atlas (TCGA)—Uterine Corpus Endometrial Carcinoma, Colon and Rectal Adenocarcinomas, and Skin Cutaneous Melanoma—two specific, functionally important positions within zinc finger domains are mutated significantly more often than expected by chance, with alterations in 18%, 10% and 43% of tumors, respectively. Numerous zinc finger genes are affected, with those containing Krüppel-associated box (KRAB) repressor domains preferentially targeted by these mutations. Further, the genes with these mutations also have high overall missense mutation rates, are expressed at levels comparable to those of known cancer genes, and together have biological process annotations that are consistent with roles in cancers. Altogether, we introduce evidence broadly implicating mutations within a diverse set of zinc finger proteins as relevant for cancer, and propose that they contribute to the widespread transcriptional dysregulation observed in cancer cells.
| Recent large-scale cancer genomics initiatives have revealed that tumor cells typically have numerous genetic alterations, only a small subset of which play a functional role in cancer. Thus, discovering cancer-relevant mutations from tumor sequencing data is a major goal of cancer genomics. Here, we aimed to uncover functional mutations occurring in transcription factors, as gene expression dysregulation is frequently observed in human cancers. In particular, we analyzed Cys2His2 zinc finger genes, as they comprise the largest family of transcription factors in human. We aggregated somatic mutations within these genes by their positions within zinc finger domains, and found that two specific positions within zinc finger domains are recurrently mutated in three cancer types—Uterine Corpus Endometrial Carcinoma, Colon and Rectal Adenocarcinomas, and Skin Cutaneous Melanoma. The two heavily altered positions are known to influence whether and how the domains bind DNA, and thus mutations within them can substantially alter gene function. Further, these mutations, observed across a wide range of transcription factors, converge on at least two processes—chromatin remodeling and dysregulation of retroelements—that are increasingly being linked to human cancers. More generally, we propose that these uncovered mutations contribute to the widespread transcriptional dysregulation commonly observed in cancer cells.
| Recent cancer genomics efforts have sequenced the tumors of hundreds of individuals across tens of different cancer types [1, 2]. Analyses of these genomes have confirmed that cancer is a highly heterogeneous disease, with tumors of the same cancer type exhibiting numerous distinct somatic alterations. Uncovering which of these mutations play a functional role in oncogenesis or tumor progression is critical for furthering our understanding of cancers and for uncovering new therapeutic targets. While approaches based on identifying genes mutated across cancer samples more often than expected by chance have yielded both well-known and newly predicted cancer genes [3], the level of heterogeneity observed in cancers suggests that there are many genes that are mutated across smaller subsets of individuals but nevertheless play key roles in cancer progression.
Previously, analysis of protein structures revealed that many well-known cancer genes are enriched in mutations that affect protein stability or participation in interactions with nucleic acids, small molecules and peptides [4–6]. Thus, these types of somatic mutations are promising as cancer-relevant candidates, even if they occur infrequently across patient cohorts. Mutations of residues that affect the stability and specificity of DNA-binding domains are particularly noteworthy, as they can contribute to gene expression dysregulation, a widespread but highly varied phenomenon across tumors [7]. Indeed, mutations within DNA-binding proteins such as p53 [8], ARID1A [9], and GATA3 [10] are prevalent in cancers. However, the heterogeneity of gene expression profiles across cancers suggests that numerous less frequently mutated transcription factors may also be relevant for cancer.
To expand our knowledge of DNA-binding proteins that may play a role in human cancers, we analyzed somatic mutations found in Cys2His2 zinc finger (ZF) domains. ZF domains are major determinants of human regulatory networks, as they are contained in nearly half of human transcription factors [11]. Despite the frequency of ZF domains, the functions of most human ZF proteins remain largely mysterious. Nevertheless, individual ZF proteins have been implicated in a range of important biological processes, including apoptosis, cell differentiation, cell proliferation, and chromosomal organization [12]. Further, the largest subclass of human ZF proteins additionally contain Krüppel-associated box (KRAB) domains, and most of these proteins are thought to mediate transcriptional repression via interactions with chromatin-remodeling factors [13, 14] and many of them have been found to bind and repress retroelements [15–18]; these ZFs may be of particular interest as changes in chromatin and altered expression of transposable elements are both observed in cancers [19, 20].
The structurally and functionally important positions within ZF domains are well known [21], and this knowledge provides a unifying framework within which to evaluate somatic mutations—even if relatively infrequent—that are found across a heterogeneous set of ZF genes. In particular, we reasoned that since specific positions within the ZF domain are associated with distinct functional roles, such as stabilizing structure or influencing the interaction with DNA, individual positions may be under differing selective pressure in human cancers. Therefore, we aggregated somatic mutations across multiple proteins by the position they fell into within ZF domains, a method that has been previously proposed to identify mutation hotspots within domains [22]. We then assessed whether positions were mutated more often than expected when controlling for mutation rates within genes and tumor samples as well as for cancer-specific mutational signatures.
Our main finding is that two specific positions within ZF domains are mutated in three cancer types—Uterine Corpus Endometrial Carcinoma (UCEC), Colon and Rectal Adenocarcinomas (COAD/READ), and Skin Cutaneous Melanoma (SKCM)—significantly more often than expected. These two heavily altered positions are known via previous structural studies to influence DNA-binding activity [21, 23]. The uncovered mutations are found in numerous genes, and are enriched in those containing Krüppel-associated box (KRAB) repressor domains. We also demonstrate that genes affected by these mutations have high overall missense mutation rates, are expressed at levels comparable to those of known driver genes, and in aggregate have biological process annotations that are consistent with roles in cancers. Overall, our work implicates a diverse set of ZF proteins as functionally relevant for cancer, and we propose that mutations within these proteins contribute to the pervasive transcriptional dysregulation observed in cancer cells.
We identified 642 human genes containing 5483 “classic” Cys2His2 ZF domains, and determined the frequency with which each of the 21 positions in this domain was mutated across patient cohorts of 32 different cancer types (see Methods). Strikingly, we found that specific positions within ZF domains were recurrently mutated in three classes of cancers: Uterine Corpus Endometrial Carcinoma (UCEC), Colon and Rectal Adenocarcinomas (COAD/READ) and Skin Cutaneous Melanoma (SKCM) (Fig 1a). Position 9 (p9, numbered relative to the start of the α-helix that binds DNA) is mutated frequently in the UCEC and COAD/READ cohorts, largely in the form of an arginine to isoleucine mutation (R9I). Position 11 (p11) is mutated frequently in the SKCM cohort, almost exclusively as a histidine to tyrosine mutation (H11Y). These mutations affect a considerable fraction of tumors, with missense mutations at p9 found in 97 of 543 UCEC and 57 of 594 COAD/READ tumors, and missense mutations at p11 found in 204 of 470 SKCM tumors (Fig 1b). Further, a diverse set of ZF genes harbor these mutations, with p9 missense mutations in 367 and 228 genes in UCEC and COAD/READ, respectively, and p11 missense mutations in 199 genes in SKCM.
Positions 9 and 11 both play critical structural roles in ZF domains (Fig 1c and 1d). The histidine in p11 is one of four residues that coordinate zinc, and thereby is essential for stabilizing the domain; while this position can accomodate a cysteine and still coordinate zinc, substitutions of other residues lead to a loss of domain function [21, 25]. ZF proteins typically contain several ZF domains, with closely linked, adjacent ZF domains working together to bind DNA. Arginine in p9 stabilizes the docking of adjacent ZF domains via a contact with the backbone carbonyl or side chain at position -2 of the adjacent C-terminal ZF, and residues in p9 influence the orientation between these domains, an important factor in DNA recognition [21, 23]. Consistent with this functional role within arrays of ZF domains, ZF domains with R9I mutations have adjacent C-terminal ZF domains more often than expected, whereas this is not the case for ZF domains with H11Y mutations (S1 Fig). Given the key roles that p9 and p11 play in ZF structure and function, mutations at these positions are likely to alter or disrupt the DNA-binding properties of the proteins in which they are found.
Having observed recurrent mutations at two functionally crucial positions within ZF domains, we next sought to uncover whether these positions are altered more often than expected by chance. The numbers of p9 and p11 mutations vary substantially across individual tumors (Fig 2a). Further, since the number of these mutations observed in each tumor is positively correlated with the total number of missense mutations (Fig 2b), consideration of per-individual mutation rates is necessary to ascertain the significance of these two mutational peaks. Thus, we obtained a background model by performing trinucleotide context-preserving permutations of mutations within each of the 642 ZF genes (see Methods). The total actual number of missense mutations at p9 is 3.2 and 2.8 times the expected count for UCEC and COAD/READ, respectively, and at p11 is 1.9 times the expected count for SKCM (p < 0.0001, Fig 2c).
UCEC and COAD/READ tumors with mutations within the exonuclease domain of DNA polymerase ϵ (POLE) are ultramutated [26, 27]. To exclude the possibility that mutations at p9 arise solely from the ultramutator phenotype, we repeated the trinucleotide context-sensitive permutation analysis on the 1,074 UCEC and COAD/READ samples without missense mutations in the exonuclease domain of POLE. Even within this subset of individuals with lower overall mutational load, there are still (respectively) 2.1 and 1.5 times as many p9 missense mutations as expected (p < 0.0001). Notably, for all three cancers, other positions in the ZF domain have missense mutation frequencies that are roughly as expected using this permutation procedure (S2 Fig), indicating that after accounting for per-gene and per-individual mutational contexts, p9 and p11 are specifically enriched for mutation accumulation across these cancer cohorts.
Because different cancers exhibit distinct mutational signatures, we next considered ZF mutations in light of these cancer-specific mutational biases. Missense mutations at p9 cause arginine to isoleucine substitutions 67% and 66% of the time for UCEC and COAD/READ, respectively. Since arginine can only mutate to isoleucine via an AGA→ATA transversion, we compared the number of p9 AGA codons that are mutated in all ZF genes per tumor to what is expected based upon per-tumor rates of AGA trinucleotide mutations across all coding regions (see Methods). Remarkably, the total numbers of p9 AGA mutations across cohorts of UCEC and COAD/READ individuals are 9.4 and 9.9 times higher than expected (both p-values < 1e-10, Poisson binomial test).
UCEC and COAD/READ tumors with the POLE ultramutator phenotype show a significant increase in the G:C→T:A transversion rate, particularly when flanked by an A:T base pair [26–29]. Indeed, of the 55 UCEC and 15 COAD/READ tumors with R9I mutations, 57 have missense mutations within the exonuclease domain of POLE, and an additional 3 have missense mutations elsewhere in POLE. When restricting our analysis to the 63 UCEC and COAD/READ samples with missense mutations in the exonuclease domain of POLE, the numbers of p9 AGA mutations remain enriched as compared to what is expected based upon the per-exome background rates of AGA mutations in these cancers (10.8 and 12.1 times higher, p = 1.0e-12 and p = 3.6e-13, Poisson binomial test). Thus, while these arginine to isoleucine mutations are consistent with the ultramutator phenotype, they accumulate at p9 of ZF domains at an unexpectedly high rate.
In SKCM, the H11Y mutations involve a Cytosine to Thymine mutation in the first position of the histidine codon, and occur when the Cytosine is preceded by a pyrimidine; the mutational patterns seen with ultraviolet light exposure and in melanoma consist of frequent CC→CT and TC→TT mutations [30]. However, ZF H11Y mutations occur 13.5 times more frequently than expected (p = 4.4e-11, Poisson binomial test) when considering the per-exome rates of CC→CT and TC→TT mutations. Thus, as with the R9I mutations and the ultramutator phenotype in UCEC and COAD/READ, the H11Y mutations are in concordance with the mutational profile characteristic of skin cancers but occur significantly more frequently than expected.
Having demonstrated that p9 and p11 mutations are enriched in ZF domains, we next examined the overall missense mutation rates of the genes within which they are found, as cancer-relevant genes are more likely to be recurrently mutated. Genes with a missense mutation at p9 (UCEC and COAD/READ) or p11 (SKCM) harbor missense mutations anywhere in their sequences in a moderate but clinically relevant range of individuals (each gene is mutated in, on average, 3.7%, 1.6% and 3.2% of UCEC, COAD/READ and SKCM tumors respectively). We next calculated for these genes the rate of missense mutations per coding sequence base, averaged across all tumor samples for each gene, while excluding the missense mutations at these ZF positions. Missense mutation rates are significantly higher in these genes than in other genes (median values for the former are 16.8%, 33.5% and 19.0% higher than those of the latter, even though rates for the latter contain all missense mutations, with p-values 8.5e-15, 2.3e-14, and 1.6e-5 for UCEC, COAD/READ, and SKCM, respectively, Mann-Whitney U test).
We next excluded the possibility that these higher missense mutation rates reflect higher overall mutation rates in the genomic regions of these genes. For each cancer type, we divided the number of missense mutations in each gene by the total number of mutations that would lead to nonsynonymous changes in that gene, and likewise with synonymous mutations (see Methods). While the overall distribution of the ratios of these two values are affected by the trinucleotide mutation profiles of each cancer type, they are comparable across genes within each cancer type since they account for both gene length and codon composition. The ratios for ZF genes with at least one missense mutation at p9 (UCEC, COAD/READ) or p11 (SKCM) are higher than the ratios for all other genes with missense and synonymous mutations (one-sided Mann-Whitney p-values 2.4e-23, 1.1e-27, and 8.8e-28, respectively) (Fig 2d). We conclude that the set of ZF genes with mutations at p9 and p11 generally have higher missense mutation rates than expected based upon their synonymous mutation rates, thereby ruling out the explanation that these genes or their genomic locations are simply more mutable, and instead lending support to the relevance of these genes to cancers.
To confirm that the uncovered mutations are affecting genes that are expressed, we analyzed RNA-seq data from TCGA for all ZF genes containing a missense mutation in at least one individual at p9 in COAD/READ and UCEC and p11 in SKCM (Fig 3). For comparison, we also extracted gene expression values for genes that are identified in the abstracts of TCGA studies as significantly mutated in these cancers [28, 29, 31]. The expression levels of the mutated ZF genes overlap considerably with those of the previously implicated cancer genes. Further, for each ZF gene, the expression levels in samples where they are mutated in p9 (UCEC and COAD/READ) or p11 (SKCM) are similar to those where they are not, with 46% (UCEC), 53% (COAD/READ), and 50% (SKCM) of the values for affected samples falling within the interquartile range (i.e., the middle 50%) of the distribution of their respective genes. These results indicate that the ZF mutations are affecting genes that are expressed at levels sufficient to play an active role in cancer (Fig 3).
We found 425 distinct ZF genes that had missense mutations in either p9 in COAD/READ or UCEC or in p11 in SKCM. The sets of these genes mutated in each cancer type show considerable overlap (Fig 4). Within each cancer type, these mutations often occur in the same gene, either in separate individuals or in the same individual, with the highest counts occurring in ZNF721 in UCEC (17 mutations, 14 individuals) and COAD/READ (11 mutations, 8 individuals), and ZNF208 in SKCM (32 mutations, 17 individuals). To test whether p9 and p11 missense mutations disproportionately affect some ZF genes, we calculated the distribution of the number of genes that would be mutated in p9 or p11 by randomizing the mutations within all p9 or p11 sites while preserving trinucleotide context; these permutation tests account for both the number of ZF domains each gene contains and the nucleotide contexts in these positions of the domains (see Methods). The actual mutations are concentrated in fewer genes than expected; they occur in 367 genes in UCEC, 228 genes in COAD/READ and 199 genes in SKCM, and these values are 16.4%, 18.0% and 24.4% lower than the average number observed in randomizations (all three empirical p-values < 0.0001). In other words, p9 and p11 mutations tend to be preferentially found in subsets of ZF genes.
Given that p9 and p11 missense mutations are distributed unevenly across the entire set of ZF genes, we next considered whether particular subsets of ZF genes were disproportionately affected. We observed that of the mutated ZF genes, 68% contain a KRAB repressor domain, whereas only 51% of all ZF genes included in our study contain a KRAB domain. This proportion increases substantially when considering the 101 genes found to be mutated across all three cancer types, as 94 (93%) of them contain a KRAB domain (Fig 4). Further, the actual fraction of missense mutations that occur at these positions in KRAB-containing genes is higher than expected by chance (empirical p-values ≤ 0.0001 for all three cancers using the p9 and p11 permutations described in the previous section). Thus, KRAB-containing genes, and presumably their shared biological roles, are preferentially targeted by p9 and p11 mutations.
ZF genes with p9 and p11 missense mutations are largely understudied, with Gene Ontology (GO) biological process (BP) annotations unrelated to transcriptional regulation associated with only 28% of the p9 or p11 mutated ZF genes. Accordingly, GO functional enrichment on the sets of ZF genes with p9 or p11 missense mutations in each cancer type yielded only general terms, largely related to transcription and regulation. Only 12 mutated ZF genes are associated with KEGG pathways; remarkably, however, all but one of these are associated either with cancer pathways or with signaling pathways regulating pluripotency of stem cells. Further, when analyzing the proteins that interact [32] with mutated ZF genes (see Methods), the most significantly enriched KEGG pathways of the partners were Notch signaling pathway (UCEC, q = 2.1e-13 and COAD/READ, q = 0.12), a well-known cancer pathway [33] and viral carcinogenesis (SKCM, q = 0.16). Many of the most highly enriched GO terms among the partners were related to transcription and chromatin organization, as would be expected for KRAB-containing ZF genes. Interestingly, keratinization was the most significantly enriched BP term among partners in COAD/READ and SKCM and the third-most enriched in UCEC, largely due to numerous ZF protein interactions with keratins and keratin-associated proteins.
Position 9 and 11 missense mutations are found in several known cancer genes. These include 11 genes—BCL6, CTCF, PRDM1, BCL11B, KLF4, MECOM, SALL4, ZNF384, ZNF331, ZFHX3 and ZBTB16—which are included in the Cancer Gene Census (CGC) [34], a list of genes causally implicated in cancer. Additionally, there are several other p9 and p11 mutated ZF proteins that are not in the CGC but nevertheless have some support for a role in cancer. For example, PEG3, which had a p9 missense mutation in three UCEC tumors and one COAD/READ tumor and an H11Y mutation in one SKCM tumor, has been implicated in the tumor necrosis factor response pathway as part of a protein complex that activates NFκB [35]; it also plays a critical role in p53-mediated apoptosis [36]. ZNF382 had these mutations in six UCEC tumors and three COAD/READ tumors, and has been implicated as a tumor suppressor [37] that is silenced in multiple carcinoma cell lines, including colon, cervical and gastric. ZNF420, also known as APAK, had missense mutations at p9 in five UCEC tumors and one COAD/READ tumor, and at p11 in one SKCM tumor; this protein interacts with p53 and in normal cells suppresses p53-mediated apoptosis [38]. Further, other p9 and p11 mutated proteins are implicated in the p53 pathway, including ZNF273, ZNF677 and ZFP28, which are the three ZF proteins recently found to interact with p53 binding protein p53bp1 [17]. Three other genes with p9 or p11 missense mutations, ZNF281, ZNF148 (ZBP89), and ZEB1, have also been identified as ZF genes important in cancer onset and progression [39]. Overall, many of the known biological processes and pathways of mutated ZF genes support their roles in cancer.
We have found that two positions within ZF domains are recurrently mutated in three cancer types. Our findings are robust in a variety of settings, including the use of different mutation callers (see S1 Text), and considering subsets of patients both with and without mutations within POLE. A previous analysis by Miller et al. [22] found that 82 positions within a diverse set of domains are hotspots that accumulate mutations more often than expected assuming that mutations are uniformly distributed across each domain. One of these identified hotspots is position 7 of Pfam domain family zf-H2C2_2; we determined that it corresponds to p9 in the present study. That hotspot was not investigated further, nor was a hotspot corresponding to p11 reported. In contrast, we demonstrate that the mutations affect two functionally important positions within ZFs, and occur more often than expected when taking into account context-specific, per-gene, and per-patient mutation rates. Further, we establish that the genes affected by these mutations tend to be more highly mutated than other genes, and are expressed at levels comparable to other cancer-relevant genes.
Several factors seem to contribute to the enrichment of mutations seen at specific ZF domain positions in UCEC, COAD/READ, and SKCM. These cancer types have relatively high mutation rates, and the positions that are affected in these cancers have nucleotide contexts that are consistent with their overall mutational signatures. Indeed, there can be disparities in both the expected and observed mutation counts at functionally similar positions; for example, the trinucleotide contexts of the first codon position at p7 (which corresponds to the other histidine coordinating zinc) are not especially prone to mutation in SKCM as they are for p11. Conversely, other positions with similar contexts as those affected do not accumulate mutations—including p8 in UCEC and COAD/READ (S2 Fig)—suggesting the importance of the specific positions affected. While there do not appear to be specific, highly mutated positions within ZF domains when analyzing other cancer types (panel B in S5 Fig), the enrichment of missense mutations in relation to synonymous mutations in the set of affected ZF genes holds true for other cancer types as well (panel C in S5 Fig), suggesting that these ZF genes may be broadly affected across many cancer types.
The two altered positions we uncover are critical to the intended biological function of ZF proteins, as mutations in p11 abrogate zinc coordination and thus domain stability [25], while mutations in p9 alter the positioning of ZF domains with respect to DNA and thus affect DNA-binding preferences [23]. Nevertheless, while p9 and p11 mutations alter or even destroy the ability of a particular ZF domain to bind DNA, because ZF proteins bind DNA via multiple domains, the protein may still be able to bind DNA albeit with different binding preferences. Thus, the overall impact at the protein level may result in both the loss and gain of regulatory targets.
The mutated ZF genes, as well as ZF genes in general, are understudied and the pathways and functions that they participate in are largely unknown. Furthermore, ZF genes that have been characterized have been shown to take part in a diverse set of pathways, including differentiation, development, growth and metabolism [40]. We thus expect that a wide spectrum of functions may be affected by the observed mutations; this is consistent with the fact that gene expression dysregulation is rampant in human cancers, and typically hundreds of genes are differentially expressed between normal and tumor samples [7]. For these reasons, it is difficult to uncover downstream expression changes due to these mutated ZF genes; further experimental work would be helpful in revealing such effects.
Despite the diversity of mutated genes, at least two broad but overlapping regulatory functions are prevalent among them. First, KRAB-containing ZF proteins, which are enriched in our set of mutated genes, play a significant role in shaping chromatin. In particular, KRAB proteins can recruit their co-factor KAP1/TRIM28, which then serves as a scaffold for bringing together factors that induce heterochromatin; this can have long-range effects, with repression evident tens of kilobases away [41]. Given that recent ChIP experiments have revealed that KRAB-containing ZF proteins can have up to 15,000 binding sites across the human genome [17, 18], the mutations we observe can have a significant effect on widespread epigenetic changes. Indeed, proteins affecting chromatin organization are increasingly being implicated in oncogenesis [19]. Second, many KRAB-containing ZF proteins have been implicated in repressing retroelements, as well as the genes neighboring these elements, both in embryonic stem cells and in fully differentiated cells [15, 16, 42]. Two recent large-scale ChIP-seq and ChIP-exo experiments on ZF proteins have determined genomic binding profiles for 231 of the 425 genes with p9 (UCEC, COAD/READ) or p11 (SKCM) missense mutations [17, 18], and the ChIP binding peaks of 156 of them were found to overlap significantly with specific retroelements in at least one of these two studies. Thus, misregulation of retroelements—or perhaps their nearby genes—may result from the p9 and p11 somatic mutations we observe. Intriguingly, transposable element expression and insertions have been observed in cancers and have been proposed to provide a selective advantage for tumorigenesis [20, 43].
In conclusion, somatic mutations at specific positions are pervasive within ZF genes, the largest class of human transcription factors, in at least three cancer types. We propose that these mutations are key contributors to widespread transcriptional deregulation in the tumors in which they are found. The frequency, distribution and enrichment of these mutations across ZF domains strongly suggest that they confer a selective growth advantage to cancer cells. The specific ZF genes mutated vary across tumors, however, and while certain shared functions are likely involved, discovering the full range of downstream effects of these shared yet distinct mutations is an exciting avenue for future research.
Somatic mutations called by the MuTect2 software [44] were downloaded from the Genomic Data Commons portal. Point mutations from a total of 10,468 tumor samples were examined, spanning 32 cancer types after combining colon and rectal adenocarcinomas into a single cancer type (S1 Table). The TCGA data was processed as previously described [6], obtaining a set of reference gene transcripts and corresponding protein sequences onto which the mutations were mapped. If a gene had multiple isoforms, the one allowing the largest number of mutations to be mapped was retained. We also confirmed our analysis with mutations called by three other variant callers, and when run on a smaller set of stringently filtered samples (see S1 Text).
The HMMER function hmmsearch (versions 2.3.2 and 3.0) was run on Ensembl human protein sequences using 12 Pfam HMM profiles from the Cys2-His2 ZF clan: PF00096, PF12756, PF13912, PF12171, PF13913, PF13909, PF12874, PF12907, PF02892, PF06220, PF09237, and PF11931. Matches to these profiles were further required to have an E-value less than 0.1 and to match CX2CX9ΨX2HX3[H|C], where Ψ is a large, hydrophobic amino acid and the final amino acid can be H or C.
Mutations across individual ZF domains were aggregated according to the position in which they occurred within the domain. Mutations were only considered if the entire ZF domain exactly matched the protein sequence from RefSeq. These criteria only excluded a few mutations in each cancer type due to variations between Ensembl and RefSeq sequences.
For each cancer type, we considered all ZF genes, and computed for each position within the ZF domain the sum of all missense mutations observed across the samples at that position (panel A in S5 Fig). We visually observed ZF p9 and p11 mutational peaks for UCEC, COAD/READ and SKCM. For these three cancer types, we assessed the significance of the total number of mutations observed in each position using trinucleotide-preserving randomizations, adapted from the method of Hodis et al. [30]. Specifically, observed point mutations (missense and silent) were shuffled within each ZF gene’s coding sequence such that each observed mutation can only be moved to another position that involves the same nucleotide, and has the same flanking nucleotides on either side. Once randomizations were performed for all genes, the numbers of mutations at each position were recomputed. This process was repeated 10,000 times, and for each position, the actual number of mutations was compared with the distribution of mutations arising from the permutations. Empirical p-values were obtained by calculating, for each position, the fraction of permutations where the total number of mutations were at least as high as the actual number.
We performed a related but different permutation analysis to determine whether mutations are concentrated within particular subsets of ZF genes. In particular, we repeatedly performed context-sensitive randomizations of the locations of p9 mutations in UCEC and COAD/READ samples across all p9 codons in the ZF genes considered in our analysis, and likewise for p11 mutations in SKCM samples across all p11 codons. To obtain empirical p-values, we counted the number of permutations that had fewer genes with missense mutations in p9 (likewise, p11) than in the unpermuted data, as well as the number of permutations that had a higher fraction of p9 (likewise, p11) missense mutations occurring in KRAB-containing ZF genes.
We used the Poisson binomial distribution to determine whether the cumulative numbers of mutations affecting ZF p9 in COAD/READ and UCEC and p11 in SKCM were significantly higher than expected when taking into account per-patient context-dependent mutation rates. The parameters required to compute the Poisson binomial distribution are: (1) the number of trials, expressed as the total number of positions of interest (i.e., p9 or p11) across all ZF genes matching the context/s of interest (i.e., AGA in COAD/READ and UCEC and CCN/TCN in SKCM, on either strand); and (2) the per-individual mutation rates for the context/s of interest, computed across the whole exome. The poibin R package [45] was used to compute the probability of observing a number of mutations in the position of interest greater than or equal to the observed value.
For each cancer type, we computed the per-gene missense mutation rate as the total number of missense mutations observed in a gene, divided by the number of nonsynonymous sites in the gene, and likewise with synonymous mutations and sites. The nonsynonymous/synonymous sites are proportional to the outcomes of the possible mutations per position, and together sum to the length of the coding sequence.
Gene Ontology (GO) and KEGG functional enrichments on mutated gene sets were determined using the hypergeometric distribution, with the background set of all annotated genes, and with q-value correction.
To test enrichment of annotation sets in the interaction partners of the mutated ZF genes, we used the undirected form of the NEAT method [46], which is based on the hypergeometric distribution and tests for the enrichment of interactions between two gene sets. We performed the test between each set of ZF genes in question and each annotation set, with the modification that we computed one-sided p-values using hypergeometric tests and then computed q-values [47] across ontology terms or pathways within each cancer type.
|
10.1371/journal.pgen.1002288 | Novel Interactions between Actin and the Proteasome Revealed by Complex Haploinsufficiency | Saccharomyces cerevisiae has been a powerful model for uncovering the landscape of binary gene interactions through whole-genome screening. Complex heterozygous interactions are potentially important to human genetic disease as loss-of-function alleles are common in human genomes. We have been using complex haploinsufficiency (CHI) screening with the actin gene to identify genes related to actin function and as a model to determine the prevalence of CHI interactions in eukaryotic genomes. Previous CHI screening between actin and null alleles for non-essential genes uncovered ∼240 deleterious CHI interactions. In this report, we have extended CHI screening to null alleles for essential genes by mating a query strain to sporulations of heterozygous knock-out strains. Using an act1Δ query, knock-outs of 60 essential genes were found to be CHI with actin. Enriched in this collection were functional categories found in the previous screen against non-essential genes, including genes involved in cytoskeleton function and chaperone complexes that fold actin and tubulin. Novel to this screen was the identification of genes for components of the TFIID transcription complex and for the proteasome. We investigated a potential role for the proteasome in regulating the actin cytoskeleton and found that the proteasome physically associates with actin filaments in vitro and that some conditional mutations in proteasome genes have gross defects in actin organization. Whole-genome screening with actin as a query has confirmed that CHI interactions are important phenotypic drivers. Furthermore, CHI screening is another genetic tool to uncover novel functional connections. Here we report a previously unappreciated role for the proteasome in affecting actin organization and function.
| Individuals inherit two copies of each gene, one from each parent, and frequently the two copies are different from each other. Sometimes one copy is completely defective, but since there is one normal copy there may be no negative consequences. Our research is focused on understanding the consequence of inheriting one bad copy for two or more different genes. Geneticists refer to inheriting one bad copy of a gene as a haploinsufficiency and inheriting one bad copy of multiple genes as a complex haploinsufficiency. Using yeast as a model system, we have addressed, in a systematic way, what occurs if a cell inherits one bad copy of the gene called actin and one bad copy of one of the ∼1,000 essential genes in the yeast genome. We discovered that complex haploinsufficiency between actin and one of 60 different essential genes leads to reduced cell viability. These 60 genes are highly enriched for certain functional groups, including those involved in protein degradation and gene expression. It is expected that approaches such as these in model organisms will be applicable to understanding the potential for deleterious complex haploinsufficient interactions in the human genome.
| There has been an increasing interest in multi-genic influences on human disease, perhaps in part because most genetic disorders with simple, monogenic origins have been described but also because some very prevalent and devastating diseases, such as cancer and psychiatric disorders, are now recognized to have complex genetic influences [1]. Extensive efforts are being made to identify the genes involved in complex genetic disorders, but this frequently amounts to trying to find a needle in a haystack of genetic diversity. An admittedly large chasm exists between these top-down approaches (disease-to-gene) and the bottom-up approaches (gene-to-phenotype) being leveraged in model systems, but an argument can be made that the bottom-up approaches are having a broader impact in how we think about genome dynamics and hold the promise for identifying conserved genetic vulnerabilities. For example, the large datasets generated by yeast genomic and proteomic approaches have been critical to the field of network biology and have helped lead to an understanding of system robustness, interconnectedness and the evolutionarily conserved aspects of biological networks [2]. The relatively high number of connections between genes and proteins in yeast biological networks suggests that systems can often adjust to the loss of a single component, a prediction that is born out by the simple observation that most yeast genes (∼5,000 out of ∼6,000) are not essential [3]. Secondly, yeast networks show short path lengths between all genes/proteins in the network, indicating a high level of interconnectedness between cellular functions [4]. Clearly the greatest contributions to understanding genetic interaction networks has come from the labs of Charles Boone and colleagues in their efforts to define the entire constellation of binary synthetic lethal interactions by synthetic genetic array (SGA) analysis in Saccharomyces cerevisiae. Besides contributing to our understanding of biological network topology, the extensive amount of genetic interaction data generated by SGA synthetic lethal screening has allowed for understanding epistatic relationships and how they can be interpreted to infer functional relatedness, and in particular whether sets of genes function in the same, parallel, or unrelated pathways [4]–[5].
Our lab has more recently become interested in digenic influences on phenotype, specifically in diploid cells and organisms. Our approach is distinguished from that of SGA analysis because it allows for complications resulting from heterozygosity at two loci. Specifically, we examine the phenotypic consequences of being hemizygous (one loss of function allele/one wild type allele) at two loci; if the negative consequence is greater than that expected based on single hemizygosity, we define the gene pair as displaying a complex haploinsufficient interaction (CHI). Recent population-scale sequencing of human genomes suggests to us that CHI interactions may be very important phenotypic drivers and contributors to genetic disorders in human populations. The 1000 Genomes Project Consortium recently published the results from whole-genome sequencing of 179 individuals and remarkably discovered that on average, human individuals inherit 250–300 loss-of-function alleles [6]. This is a strikingly high number in comparison to previous estimates and suggests that every person inherits a unique combination of over 44,000 potentially deleterious CHI gene pairs (300-choose-two is a binomial coefficient determined by the factorial equation N!/K!(N-K)! where N is the number of elements and K the subsets of N). A critically important question is how many complex hemizygosities are deleterious. We previously addressed this question using the yeast model and a null allele of the essential actin gene to screen for CHI interactions with the ∼4,800 non-essential gene knock-out alleles. We have found that null alleles for ∼240 non-essential yeast genes display deleterious CHI interactions with the actin null allele ([7]. Although we admit that actin is a particularly important and multi-functional gene, these results strongly argue that the landscape of potentially deleterious CHI interactions is likely to be quite large.
As has been observed in the synthetic lethal data, our network of CHI genes is highly enriched for certain functional groups including several functions and complexes known to be involved in actin function. In addition, we uncovered unanticipated genetic vulnerabilities reflecting novel connections to the actin cytoskeleton that were not predicted by more traditional approaches such as protein interaction studies [7]. What this suggests to us is that efforts to map the causes of multi-genic human disorders is going to continue to be problematic as it may be difficult to predict which genes neighboring a chromosomal region linked to disease inheritance are responsible for the phenotype of interest. This is likely to be made even more complicated by the influences of several hundred additional loss-of-function alleles. We argue that knowing the landscape of binary CHI gene interactions in a simple eukaryotic may be very useful in predicting multi-genic influences in human genetic disorders, in particular when the affected genes disrupt core cellular functions. At this time we do not have clear examples of CHI interactions causing human disease. However, single gene haploinsufficiencies in 32 different transcription factors causes a diverse array of human genetic disorders [8], haploinsufficiency of 23 different tumor suppressor genes has been shown to contribute to tumorigenesis [9], and complex haploinsufficiency has been shown in mouse models to contribute to early aging [10] and tumorigenesis [11]–[12]. We suspect that examples of CHI interactions in humans are lacking due to the difficulties associated with detecting these interactions and not because they aren't consequential.
Since our first CHI screens were performed on the non-essential gene set, many core and conserved cellular processes that are facilitated largely by essential genes were missed. Here we report an adaptation of this approach that enables the analysis of complex heterozygous combinations of essential gene deletions and its use to expand the actin CHI network to include 60 essential genes. There is extremely significant functional enrichment in the actin CHI essential gene network, including the inclusion of many genes for proteasome components. The proteasome is known for its ability to degrade ubiquitinated proteins. However, our preliminary analysis of its role in actin regulation suggests it may play a structural role in the actin cytoskeleton.
In a previous report [7], we presented a genome-wide analysis of complex haploinsufficiency (CHI) interactions arising in diploid strains deleted for one copy of the unique and essential actin gene (ACT1) and for one copy of a non-essential gene. This was performed by mating a haploid strain deleted for ACT1 (kept alive by a counter-selectable ACT1-containing plasmid) with the EuroScarf haploid non-essential gene deletion mutant array. Diploids that were impaired for growth without the ACT1 plasmid indicated that the double hemizygosity was incapable of supporting (robust) growth. While we have uncovered ∼240 such detrimental combinations ([7]; and Haarer, Viggiano and Amberg, unpublished), an argument can be made that perhaps more relevant or stronger CHI interactions might occur when both hemizygous mutations are in essential genes. Furthermore, many fundamental processes are executed by essential genes and these would have been missed by the previous screening method.
In order to carry out a similar screen against essential genes, we took advantage of the observation that spores carrying deletions of many essential genes are still competent to mate prior to dying [13], presumably due to a sufficient complement of maternal factors to support germination and limited growth. Therefore, we first sporulated the hemizygous diploids from the EuroScarf (http://web.uni-frankfurt.de/fb15/mikro/euroscarf/) collection of S. cerevisiae strains deleted for essential genes. These spores were then directly mated to our act1Δ query strain (carrying ACT1 on a counter-selectable plasmid) and diploids carrying both the NatR-marked act1Δ allele and G418R-marked gene deletions were selected and analyzed for growth defects on medium that selected for cells that had lost the ACT1-bearing plasmid. In this manner we were able to perform complex haploinsufficiency screens against these essential genes.
As with CHI screening against non-essential genes [7], our essentials-only CHI screens generated substantial numbers of interactions that varied in severity from lethal to moderately slow growing. Overlap between screens varied, ranging from 11 to 49%. From four separate screens against the EuroScarf collection of 1095 essential-gene deletion strains, we obtained a total of 212 potential CHI interactions, of which 39 were hit twice, 14 were hit three times, and 6 were hit in all four screens.
We have previously found that screening robotically results in a large number of false positives necessitating manual confirmation or “hand testing” which involves manually mating the strains on plates and streaking for single colonies on selective plates. This allows us to examine the growth characteristics of several clonal isolates for each CHI cross. Hand retesting of the 212 potential CHI interactions allowed us to confirm 48 authentic act1Δ CHI interactions; the other 164 failed to confirm by the more rigorous hand tests and are therefore false positives. From the primary screen data, we identified related functional groups (see below) that implicated another 24 genes; of these, 12 of the genes' null alleles tested positive for a CHI interaction in subsequent hand tests. This combined network of 60 genes is shown in Figure 1; note that the color coding reflects the major functional gene ontology terms assigned to the genes as indicated in the key. This number of interactions is probably underestimated, as we continue to see some variability or “escape” from the FOA screen (see below and Figure 2A). We believe that escape can occur through multiple mechanisms, including conversion or duplication at either mutant locus, mutational inactivation of the plasmid-borne URA3 marker, or chromosome non-disjunction leading to trisomies for either the chromosome bearing the CHI gene or a chromosome for some other gene capable of suppressing the CHI interaction. We believe the latter is most likely, as the reported frequency of spontaneous aneuploidies in S. cerevsiae is as high as 1×10−4 [14]–[15]. Notably, the number of confirmed hits for both the essential and non-essential deletion collections is ∼5% of the total number of genes represented in each collection, which would seem to suggest that the essential genes are no more predisposed to CHI interactions than are the non-essential genes. However, it may be that that the false negative rate is higher for the essential gene screens possibly as a result of the sporulation step; if this is true then essential genes may, on average, have more CHI interactions than non-essential genes as has been observed for synthetic lethal screens.
To obtain a more quantitative measure of CHI interactions, we sought to determine growth rates of double hemizygotes upon loss of the ACT1 plasmid. For this purpose we adopted a small-scale liquid media approach and utilized a TECAN microplate reader/shaking incubator. Several notable observations were made while carrying out these liquid media growth tests: 1) The restrictive temperature for actin hemizygotes is significantly lower in liquid media. While 34–35° (our typical “high temperature” CHI screening mark) is permissive for growth of ACT1/act1Δ strains on solid media, we found that growth above 30° was severely impaired for such strains in liquid media. 2) Only the strongest CHI hits (generally scored as 1 = inviable or 2 = severe growth defects) exhibited detectably reduced growth rates by this method (see Figure 2B). Presumably, expansion of the CHI culture allows for greater opportunities for suppression events caused by genetic instability such as the generation of trisomies by chromosome non-disjunction. As mentioned above, even for the strongest CHI interactions, such as the interaction with tub1Δ, escape events are not infrequent. This can be seen fairly clearly in Figure 2A where the ACT1/act1Δ tub1Δ/TUB1 complex hemizygote culture traced with yellow triangles had an escape event while the culture traced in red squares did not. Interestingly, the liquid growth behavior of CHI strains manifested as delays in the onset of logarithmic growth and/or differences in the slopes during logarithmic growth (see Figure 2A and 2B). For the strongest CHI interactions, such as with tub1Δ and taf5Δ, long delays are followed by low slopes in the growth curve. For CHI interactions of more modest strength, such as with pfy1Δ, cct4Δ, and arc19Δ, modest delays are followed by fairly small differences in slope from the control cultures. Curiously, the rpn6Δ culture shows a delay over the control culture but no difference in slope once log-phase growth is entered (Figure 2B). Although it is difficult to obtain a strict quantitative measurement of a CHI interaction using TECAN growth curve analysis, it does provide another means to compare the strengths of the interactions over mere colony size.
A normal wild type cell typically has 4–5 actin cables that run along the long axis of the cell, terminating at the bud neck or extending into the bud. Actin cables are used by myosin motors to facilitate the polarized transport of materials into the bud. The other prominent actin structures are actin cortical patches, which in a wild type cell are polarized in the bud for most of the cell cycle and become polarized to the bud neck near the end of the cell cycle. Actin cortical patches are sites of endocytosis. For an example of normal actin organization, see “WT” in Figure 5C. Actin-specific phenotypes are relatively normal for most hemizygotes of the act1 CHI set (data not shown). We had previously reported that the non-essential genes whose null alleles are CHI with actin displayed a surprising frequency of actin organization defects as haploids [7]. A comparable analysis cannot be performed with the essential gene null alleles that are CHI with actin, instead we examined the actin cytoskeleton in some of the complex hemizygotes. This analysis was complicated by the fairly severe actin organization defects in the act1 hemizygote alone, which has abundant but depolarized actin cortical patches and poor actin cable formation and polarization (see Figure 3B). We did note some actin and morphological defects beyond those of the act1 hemizygote (Figure 3A) in several complex hemizygotes, including those carrying pfy1 (Figure 3B), rpn5 (Figure 3C), and tub1 (Figure 3D), while others had few, if any, additional defects. Lowering actin levels appears to affect actin cable formation more seriously suggesting that the patch F-actin nucleation machinery (the Arp2/3 complex) is less sensitive to low actin concentrations than the cable nucleation machinery (the formins). The larger, rounder cells of the act1/ACT1wt pfy1Δ/PFY1wt complex hemizygote and extreme patch depolarization (Figure 3B) are reminiscent of various pfy1 point mutants and represent a much less severe phenotype than that of barely viable pfy1Δ haploid cells [14]. Many of the complex hemizygotes with deletions of proteosome component genes showed an elongated cell and bud phenotype, an example of which is shown for rpn5 in Figure 3C. The ACT1/act1 tub1/TUB1 hemizygotes displayed severe and heterogeneous defects in actin organization and cell morphology. In many cases actin assembly was overly elaborate and completely polarized in the bud (Figure 3D) which is consistent with our previous observation that stabilization of filamentous actin is detrimental when actin is limiting [7].
As in our previous study [7], we noted that several genes whose null alleles are CHI with act1Δ express proteins that fall into common pathways or complexes. Hand testing of strains bearing null alleles for other (previously not hit) members of these pathways or complexes resulted in the identification of 12 additional CHI interactions with act1Δ. This observation alone argues that there is functional enrichment in the network. Some functional overlap was observed between this and our previous screens of the non-essential knock-outs. For example, here we found interactions with null alleles for two essential components of the Arp2/3 complex (ARP3 and ARC19) and a null allele for the Arp2/3 activator LAS17. We also found an interaction with a null allele for one of the few (nearly) essential actin binding protein genes (yeast profilin; PFY1), and we found interactions with null alleles for two genes for the essential Tcp1p chaperone complex, which helps to fold actin and tubulin subunits; the previous screen saw interactions with null alleles for several components of the separate and non-essential prefoldin chaperone complex that also helps to fold both actin and tubulin subunits [15].
Our original CHI screen against the non-essential genes missed many cellular functions that are executed by essential genes. Although the screens against the non-essential knock-outs identified many genes involved in endocytic processes, most of the exocytic genes that had been frequently found in previous classical genetic screens with actin (e.g. suppression screens; [16]) were missed because these are almost entirely essential. Here we saw an interaction with the null allele for the late secretory vesicle small GTPase Sec4 and null alleles of genes for two components of the coatamer complex (RET2 and COP1) that facilitate retrograde transport from the golgi to the ER. Also notable are the interactions with the tubulin cytoskeleton, in particular tub1Δ whose wild type allele encodes for alpha-tubulin, tub4Δ whose wild type allele codes for gamma-tubulin that nucleates microtubules at the spindle pole body, and spc42Δ whose wild type allele codes for a core component of the outer plaque of the spindle pole body. Since actin has been shown to be involved in nuclear positioning and segregation during mitosis ([17] and references cited within) the interactions with tubulin and spindle pole body genes may reflect catastrophic defects in nuclear positioning under conditions of limiting actin concentration. Statistical analysis does confirm that the network is enriched for genes annotated to the GO component term “cytoskeleton” with a p value = 6.62×10−13. This includes the genes: PFY1, TUB1, LAS17, CCT4, TUB4, ARC19, SPC42, SEC4, CCT8, and ARP3 [18].
Most striking were the extensive interactions with null alleles of genes for the TFIID complex and the proteasome. Knock-out alleles of nearly all of the TAFs that make up the proteins of the TFIID complex were found to be CHI with act1Δ; statistical analysis for GO-term enrichment gave a p-value of 1×10−14 for the GO Biological Process term “transcriptional pre-initiation complex assembly” [19]. This could be the result of poor expression of actin and/or actin binding proteins, a hypothesis that is supported by CHI interactions between act1Δ and deletion alleles for three components of the RNA polymerase II complex (RPB7, RPB3, and RPC10). However, since actin has been found to be a core component of the RNA polymerase II pre-initiation complex [20] and to be involved in transcriptional elongation [21], these interactions may reflect a more direct role of actin in transcription. For example, stoichiometry suggests that actin in the RNA polymerase II complex is in the monomer form [19] and we have shown that the actin hemizygote is particularly limited for G-actin [7].
Particularly surprising are the extensive CHI interactions with null alleles for genes of the proteasome. Of the 33 genes that encode for all components of the 26S proteasome, we found that the knock-outs for 15 were CHI with act1Δ. The p-value for enrichment of the GO biological process term “ubiquitin-dependent protein catabolic process” is a remarkable 6.5×10−13. The genetic interaction between actin and proteasome genes could also be the result of indirect effects such as alterations in the protein levels of actin or key actin regulators. We have not performed an extensive analysis of the levels of all actin regulators but have looked at actin itself and its levels are unchanged in proteasome mutants. Alternatively, it could reflect a direct role of the proteasome in the regulation of the actin cytoskeleton or vice versa.
The proteasome is a large macro-molecular complex that executes much of the protein degradation activity of the eukaryotic cell (see Figure 4). It consists of a 20S core particle (CP) that is capped at one or both ends by a 19S regulatory particle (RP; see Figure 4A). The CP consists of two central, heptameric ß-rings composed of 7 different ß-subunits that are sandwiched between two α-rings similarly composed of 7 α-subunits. The CP contains the proteolytic sites while the RP is involved in substrate recognition and protein unfolding. The RP can be further subdivided into a lid complex that is coupled to the base complex by the Rpn10 protein [22]. Interestingly, nearly all proteasome components (with the exception of the α-3 subunit Pre9p and the cap coupling subunit Rpn10p) are essential and presumably required for proteasome activity, and yet not all knock-outs of proteasome component genes display a CHI interaction with act1Δ (see Figure 4A). Concerning the RP, we observed that null alleles for all genes of the lid complex components except SEM1 were CHI with act1Δ. In contrast, a knock-out for only one gene (RPT2) of the base complex was CHI with act1Δ and a null allele for the Rpn10p linker was also not CHI with act1Δ. With respect to the 20S core, we observed that null alleles for only three of the alpha subunit genes (PRE5, PRE10, and SCL1) and three of the ß-subunit genes (PUP1, PRE1, and PRE2) were CHI with act1Δ. Note that three ß-subunit genes encode for the three proteolytic activities of the proteasome and two of the three are CHI with actin; PUP1 encodes for the subunit with trypsin-like activity and PRE2 encodes for the chymotryptic activity [23]–[24]. Interestingly, only the chymotryptic activity of the proteasome is essential [24].
The crystal structure of the proteasome 20S core indicates that each of the alpha and beta subunits occupies distinct locations within the two α– and ß-rings [25]. Figure 4B and 4C shows surface renderings of the structure of the 20S proteasome core with the CHI components colored and labeled. Strikingly, the α-ring CHI subunits are adjacent to each other and generally do not overlap with the ß-ring CHI subunits, such that α- and ß-subunits lie on opposite surfaces of the barrel-shaped core. The asymmetry of CHI interactions for proteasome subunit genes could reflect that these components are limiting and that reductions in their levels leads to defects in proteasome assembly. For example, the juxtaposition of CHI and non-CHI subunits in the 20S core particle may reflect the roles of co-templated interactions between subunits of the α- and ß-rings during the assembly of the intermediate, 20S CP ½ ring complex [26].
To examine if there may be a functional connection between the proteasome and actin, we obtained several Ts− alleles of proteasome component genes ([27] and gift of C. Boone) and examined their growth characteristics on both plate media and in liquid cultures after a shift to 37°C. We noted that the proteasome mutants varied in their extent of growth defects at 37°C. Interestingly, some proteasome mutants failed to form colonies at 37°C but appeared to grow relatively normally in liquid cultures at 37°C. Typically we have observed that strains with Ts− alleles tend to behave similarly on both plate and in liquid cultures (with the exception of the ACT1 hemizygote) and so the discrepancies observed for proteasome mutants may reflect something special about these genes or the proteasome in general. The growth behavior of the proteasome mutants is noted in Table 2.
To examine the actin cytoskeleton in proteasome mutants, we stained log-phase cultures of proteasome mutant strains with rhodamine-phalloidin following a 2 hr shift to 37°C and examined actin organization by fluorescence microscopy. Interestingly, there was a wide range of phenotypes observed, from completely wild type actin organization to very severe defects in actin organization (see summary in Table 2 and examples of actin staining in Figure 5B). We scored a mutant as having severe defects when virtually all the cells in the culture were mutant: generally very large cells usually lacking cables and with numerous depolarized patches. Moderate defects indicate cultures in which most of the cells have morphology defects, generally large and sometimes hyper-elongated cells with excessive cables. Actin defects did not strictly correlate with whether the null allele for the proteasome mutant is CHI with actin, whether the subunit it encodes is in a particular sub-complex of the proteasome, or the extent of temperature sensitivity of the mutant strain. In Figure 5 we highlight the results from a set of mutants that have largely equivalent Ts− phenotypes in liquid culture as shown by growth curves generated using a TECAN microplate reader (Figure 5A). This included alleles for three lid complex genes (rpn5-1, rpn11-14, and rpn12-1) and alleles for two base complex genes (rpn1-821 and rpt1-1). As can be seen in Figure 5B, the three lid complex mutants have severe actin and cell morphology defects that are notable at permissive temperature (25°C) and get even worse at 37°C. At 25°C the cells are large, frequently elongated with elongated buds, the actin cortical patches are well polarized, and the actin cables are present and polarized. At 37°C, the cells became extremely large with depolarized cortical patches and if actin cables are present they are largely disorganized. This phenotype was typical for those proteasome mutants that display actin organization defects and although these strains are presumably arresting their growth due to defects in proteolysis, the actin defects alone are severe enough to possibly cause inviability. In stark contrast, the rpn1-821 and rpt1-1 mutant cells were completely normal for both cell morphology and actin organization at both 25°C and 37°C (Figure 5C). The kinetics of growth arrest at 37°C for all strains was similar and yet only some of the mutants arrest with actin defects. Although we do not know the reasons for this phenotypic variation, we suspect it may relate to differences in how the mutants affect proteasome assembly and function, in particular a possible actin specific function.
The actin organization defects observed in the rpn5, rpn11, and rpn12 Ts− strains (and others) were expected to be attributable to loss of proteolysis activity of the proteasome complex. The lactocystin derivative MG132 is a commonly used proteasome inhibitor that covalently modifies the active site threonine of the chymotryptic center of the Pre2p subunit [25]–[26], [28]. Although MG132 specifically targets the chymotryptic activity of the proteasome, it significantly reduces the proteolytic activity of the other two ß-subunits as well [28]. This and related derivatives have been used extensively to study mammalian cell proteasome function but its utility in studying yeast proteasome function is limited by cell permeability issues. However, an ergosterol biosynthesis mutant (erg6) has been shown to increase cell permeability to a number of drugs, in particular an erg6 mutation has been shown to render yeast cells sensitive to MG132 [29]. We grew an erg6Δ strain to mid-log, treated it +/−100 µM MG132 for 2 hr and stained the actin cytoskeleton with rhodamine-phalloidin. A comparable treatment of an erg6 mutant strain had previously been show to cause an ∼80% inhibition of in vivo proteasome activity [29]. Despite an assumed comparable inhibition in our hands, we saw absolutely no effect of MG132 treatment on actin organization (Figure 6A). Deletion of the gene for the ABC, multi-drug transporter Pdr5p has also been shown to render yeast cells sensitive to 50 µM MG132 [30]. However, as similarly observed for the erg6Δ strain, treatment of a pdr5Δ strain with 100 µM MG132 for 2 hr had no effect on actin organization (Figure 6A). To confirm that the MG132 was in fact inhibiting the proteasome, we isolated proteins from cells treated with 50 µM MG132, separated the protein samples by SDS-PAGE and Western blotted with anti-ubiquitin antibodies (Figure 6B). Both the erg6Δ and pdr5Δ cells as well as the wild-type control accumulated ubiquitinated proteins during the 2 hr treatment with MG132. However, we noted that both erg6Δ and pdr5Δ strains continue to grow in the presence of 100 µM MG132 (see Figure 7A; GAC201 is a pdr5Δ strain) indicating that the residual activity of uninhibited proteasomes is sufficient to maintain cell growth and viability despite the accumulation of significantly high levels of ubiquitinated proteins.
It has been shown that yeast can survive without the trypsin-like activity of Pup1p and the caspase-like activity of Pre3p [23]. More recently a strain was developed in which the active site threonines of both Pup1p and Pre3p were mutated, rendering the strain completely reliant on the chymotrypsin activity of the Pre2p subunit. This strain also carries a pdr5Δ allele, rendering it completely sensitive to MG132 for cell growth and viability [31]. We confirmed the sensitivity of this strain to MG132 by performing TECAN growth curves (Figure 7A), note that the hyper-sensitized GAC202 strain arrests growth immediately following the addition of MG132 (Time 0). We treated this strain, and a control strain bearing wild type alleles of PUP1 and PRE3, +/−100 µM MG132 for 2 hr followed by rhodamine-phalloidin staining to visualize the actin cytoskeleton (Figure 7B). Strikingly, the MG132 sensitive strain treated with MG132 displayed a high percentage of cells with F-actin accumulation at the tips of long and narrow buds. We quantified the percentages of cells with these hyper-elongated buds and found that in cultures of the MG132-sensitive strain grown in MG132, nearly 20% of the cells displayed this phenotype as compared to at most 3% in control cultures (see Table 3). Since the control strains carry plasmids encoding the wild type copies of PUP1 and PRE3, we assume the low percentage of cells with elongated buds observed in the MG132-treated control strain are the result of plasmid loss. Interestingly, this elongated bud phenotype is distinct from that observed in proteasome subunit Ts− strains (see Figure 5B). Elongated buds are observed in some Ts− proteasome mutants but they tend to be broader, particularly in the neck region. The overwhelmingly common phenotype at 37°C, for proteasome mutants that do show actin defects, is large cells with a depolarized actin cytoskeleton. In contrast, the arrest phenotype observed in the MG132 hypersensitive strain is actin hyper-polarization and is more reminiscent of a cell cycle defect, likely reflecting the well-established role of the proteasome in degradation of cell cycle regulators [32].
The diversity of actin organization defects observed in some proteasome mutants suggested that perhaps the proteasome could be regulating actin via a direct interaction. To test this possibility, we purified assembly competent actin and the 19S proteasome RP from yeast. Actin filament pelleting assays were used to assay for 19S RP association specifically with actin filaments. When the actin is kept in the monomeric form by incubating in low salt G-buffer and submitted to high-speed centrifugation at 190,000× g, nearly all of the actin remains in the supernatant phase (Figure 8A). When the actin was incubated in the presence of ∼65 nM yeast 19S RP, the actin remained in the supernatant in G-buffer and none of the 19S RP was found in the pellet fraction (Figure 8A). However, when actin polymerization was stimulated with salt and magnesium in F-buffer, most of the actin is found in the pellet phase (see Figure 8B) and most of the 19S RP proteins pellet with the actin filaments. Note that the 19S RP is not found in the supernatant fraction as this fraction is taken from the very top of the reaction; we have sampled down the column of F-buffer centrifugation reactions and consistently find that the 19S RP migrates to a fraction above the pellet (data not shown). To confirm the apparent association of the 19S RP with actin filaments, we analyzed both the 19S preparation and the F-actin plus 19S pellet by mass-spectrometry, specifically reverse phase chromatography in line with tandem MS using an LTQ Orbitrap XL. The 19S prep was confirmed to contain multiple high confidence peptides for all of the lid and base components, for the Rpn10p linker, and for the Rpn14p proteasome assembly factor. In the F-actin plus 19S pellet a large number of actin peptides were identified, multiple peptides for all 9 base components, multiple peptides for 7 of 8 lid components but only 1 peptide for lid component Rpn5p (data not shown). It is likely that the large concentration of actin peptides in the pellet fraction obscured the identification of some 19S RP peptides. Regardless, the mass-spectrometry results provide strong support for the physical association of the full 19S RP with actin filaments in our pelleting assays.
To examine for a potential association between the 20S core proteasome and F-actin, pelleting assays were repeated with preparations of 20S core purified from a strain expressing Flag-tagged Pre1p [33]. As was observed for the 19S RP (Figure 8A), when the actin was incubated in the presence of 50–100 nM yeast 20S proteasome, the actin remains in the supernatant fraction (Figure 8C). When actin polymerization was stimulated in F-buffer, most of the 20S CP proteins pelleted with the actin filaments (see Figure 8D). Sampling of the entire fractionation consistently showed that the 20S proteasome, like the 19S RP, migrated to a fraction above the pellet (data not shown). The strong association of the 19S RP with F-actin suggests that 19S RP contamination in the 20S prep could be bridging the F-actin/20S association. To determine if 19S RP is present in the 20S preps, we performed the mass-spec analysis of the 20S preps and did find that the 20S preparation is contaminated with 19S RP proteins. However, the converse was not true; no 20S contamination was found in the 19S preparations (data not shown).
Although DNA replication and repair occur with impressive fidelity, the sheer amount of information that must be copied ensures that every time a genome is replicated, heritable mistakes are made. Over time, these errors and larger scale changes in chromosomes such as transposon movement, repeat expansion and contraction, and chromosome rearrangements have led to significant genetic diversity in populations. These genetic alterations have resulted in complex heterozygosity at many loci that are re-assorted every generation, resulting in a moving target of complex genetic interactions whose influence on phenotype is difficult to ascertain. Many, but not all, mutations are silent but it has been estimated from deep sequencing of 179 human individuals, that we all inherit 250–300 loss-of-functional alleles. Cancer cells represent perhaps the most extreme case of genetic diversity, as they frequently lose whole chromosomes or whole sections of chromosomes, resulting in the potential for thousands of complex haploinsufficient interactions. In our work we have been attempting to model CHI interactions in a simple eukaryote and specifically using the single essential actin gene, which plays critical roles in a large number of cellular functions. Our results confirm that there is a tremendous potential for deleterious CHI interactions in the genome. However, actin is so centrally important, we cannot say that all genes will be so genetically promiscuous.
Genetic interaction analysis has been used extensively in yeast to attempt to identify functionally related genes, the well supported assumption of course being that a binary gene interaction suggests relatedness. Such screens have led to a fairly anecdotal lore about what a particular type of interaction means, e.g. suppressor screens tend to identify gene products that physically interact or at least operate in the same pathway. The tremendous amount of synthetic lethal data generated by the Boone laboratory has pushed the ability to formalize rules about synthetic lethal interactions and how patterns of such interactions can be accurately used to determine function in same or parallel pathways [4]–[5]. Although very few CHI screens have been performed at this time, we expect that many of the lessons learned from synthetic lethality will hold true.
It is not surprising that a null allele for actin would have such a large number of CHI interactions; the actin cytoskeleton is centrally important to many cellular functions. Further, actin can exist in many highly dynamic forms that are regulated in complex spatio-temporal ways by a large number of associated proteins. Interpretation of any single CHI interaction will need to take into account these complexities. One thing that is clear is that CHI screens, like synthetic lethal screens, identify groups of functionally related genes and, with respect to actin, we can conclude that some very important cellular processes are hyper-sensitive to reductions in actin expression.
Since actin itself is haploinsufficient, we may have expected our screen to be overwhelmingly biased toward other genes that also display simple haploinsufficiency. However, this does not appear to be the case, thus bolstering the argument that, like other genetic interactions, a CHI interaction reflects functional specificity. Another possible explanation for CHI interactions could be cumulative reduced biosynthetic capacity [34], but our original CHI screen against the non-essential genes was not overly enriched for genes involved in biosynthesis, with the exception of the ribosome. The CHI screen against the essential genes presented here did identify a number of genes whose products are components of the proteasome and the core transcriptional apparatus, but we could not detect significant changes in the levels of actin, or the actin binding proteins Aip1p and cofilin in strains hemizygous for RPN5, RPB3, RPC10, TAF5, or RPS5 (data not shown) suggesting these interactions may reflect true functional connections.
A common interpretation of synthetic lethal interactions between null alleles is that the two genes operate in parallel pathways that impinge upon an essential function [4]. A fundamental difference of a CHI interaction is that since activity from both genes is reduced by at most 50% and not 100%, the pathways the genes function in are not blocked entirely. Therefore, we might expect that a CHI interaction may frequently be indicative that the two genes/gene products function within the same pathway or structure; that pathway function is compromised by loss of flux due to constrictions at two steps. Given the striking enrichment for proteasome genes and the novelty of a proteasome-actin connection, we chose to investigate a potential functional connection between the actin cytoskeleton and the proteasome.
Analysis of Ts− proteasome mutants showed that many, but not all, have severe defects in actin organization. In some cases these defects are so severe that this aspect of the phenotype alone might be expected to cause cell death. The data suggest, from two perspectives that the actin phenotypes cannot be purely explained by losses in proteasome proteolytic activity. First, treatment of sensitized cells with the proteasome proteolytic activity inhibitor MG132 resulted in very distinct defects in cell polarity and actin polarization that differs from that observed in Ts− proteasome mutants. Second, several Ts− proteasome mutants that cease growth entirely at 37°C arrest without any actin organization defects. We hypothesize that the phenotypic diversity observed in the proteasome mutants likely reflects differential effects of the alleles on multiple proteasome activities. For example, loss of proteolytic activity may indirectly affect actin and cell polarity. However, we hypothesize there is a second function that is evident when there are defects in a direct interaction between actin and proteasomes.
This hypothesis is supported by actin filament pelleting assays suggesting a direct physical interaction between the proteasome and actin filaments, although we cannot rule out possible non-specific interactions between two very large complexes. However, we failed to see pelleting of a BSA control with actin filaments indicating that actin is not merely “sticky” under these assay conditions (data not shown). Nonetheless, the data suggest that there are F-actin binding sites on both the 19S RP and the 20S CP. However, since mass-spectrometry identified 19S proteins in the 20S preparations, we cannot preclude the possibility that the 19S RP was bridging an interaction between F-actin and the 20S core in our assays. In either case, either double-capped 26S proteasome, or the 20S CP could be a bivalent actin filament cross-linking protein. This model agrees well with EM images showing a strikingly ladder-like cross-linking of rabbit muscle actin filaments by rabbit reticulocyte proteasomes [35]. These data suggest that the actin-proteasome interaction is conserved, which is supported by co-localization of proteasomes with the actin cytoskeleton in many different cell types including Xenopus oocytes [36], epithelial cells and fibroblasts [35], [37], myoblasts [38], and in the sarcomeres of muscle cells [37], [39]–[40].
One curious aspect of our proteasome data is that not all proteasome genes are CHI with actin and not all Ts− alleles in proteasome genes cause actin organization defects. In our previous work on CHI interactions between actin and the non-essential gene knock-outs [7], we uncovered functional divergence between ribosomal paralog genes for actin-related defects. It has been shown that proteasomes with an alternative composition can be assembled in certain genetic backgrounds [41]. However, we do not believe that our data suggest the existence of a proteasome of alternative composition that has actin-specific functions. Instead, given that proteasome assembly has been shown to occur in a series of ordered steps [26], , we hypothesize that subunit limitation in hemizygous strains or mutation of certain proteasome subunits leads to the accumulation of assembly intermediates that selectively affect actin-specific functions of the proteasome. In regards to the phenotypic differences between Ts− mutants of proteasome component genes, this likely reflects differences in the defects caused by the alleles. For example, some Ts− mutants may affect function without affecting structure while others may affect both function and structure. Given our results with MG132 and actin organization, we hypothesize that proteasome structure is required for proper actin organization but catalytic activity may not be required for proper actin organization.
In summary, our results clearly prove the utility of CHI screening for making novel connections between cellular subsystems, for predicting gene function, and for mapping the landscape of binary gene interactions that are likely to be relevant to system collapse in more complex organisms including multigenic influences in human genetic disorders. In particular, extensive CHI interactions between actin and proteasome genes predicted a novel role for the proteasome in affecting actin cytoskeleton organization.
Complex haploinsufficiency screening was carried out as described previously [7], with the following changes: the EuroScarf (http://web.uni-frankfurt.de/fb15/mikro/euroscarf/) collection of heterozygous diploid strains carrying essential gene deletions was first pinned in 384-colony format to four YPD+G418 plates, incubated for two days at 25–30°, then transferred to sporulation plates. After 7–12 days at 25°, sporulation of randomly selected strains was confirmed (sporulation frequencies were generally between 1–5%) and these sporulated strains were mated on YPD to the act1Δ query strain BHY336, which contains ACT1 on a URA3-based, centromere plasmid. Subsequent steps of diploid selection and CHI analysis were as described [7]. For some screens, the 384-colony plates were converted to 1536-colony format, equivalent to four-fold coverage per plate. Hand retesting was also as described previously, but with the additional step of first sporulating the heterozygous diploid strains prior to mating to BHY336.
For TECAN (Infinite F200)-based growth curves of CHI interacting strains, we picked colonies directly from diploid selection plates into liquid SCmsg+G418+Nat+/−FOA and followed growth at 30° for ∼48 hours. Data were imported into Microsoft Excel for analysis and generation of growth curves. For TECAN (Infinite F200)-based growth curves of proteasome mutants strains, 25°C cultures in YPD medium at ∼2×107 cells/ml were placed into wells of microtiter dishes and shifted to 37°C in the TECAN incubator/reader and cell density was monitored for 24 hr.
Rhodamine-phalloidin staining of actin was as described previously [43]. For rhodamine-phalloidin staining of complex hemizygotes, the haploid act1Δ query covered with the URA3-marked ACT1wt plasmid was mated to sporulations of the essential heterozygous diploid strains, the double hemizygotes were selected on medium containing NAT and G418, and plasmid loss was subsequently selected on 5-FOA medium. The resulting double hemizygotes were grown to mid-log in YDP medium and stained with rhodamine-phalloidin.
Yeast strains were grown to mid-log phase (∼2×107 cells/ml) and MG132 (A.G. Scientific Inc, San Diego, CA) dissolved in DMSO at 50 mM was added to a final concentration of 100 µM; an equal volume of DMSO was added to control cultures. The cells were incubated for 2 hr at 30°C, fixed in formaldehyde, and stained with rhodamine-phalloidin (see above).
Strains BY4742, erg6Δ and pdr5Δ were grown to mid log phase at 30°C and then treated with 50 µM MG132 (diluted from a 50 mM stock solution in DMSO) for 2 hours. 9 ml samples were harvested at time 0 min (before MG132 addition) and 30 min, 75 min and 120 min after MG132 addition. The cells were pelleted, the pellets were flash-frozen on dry ice and stored at −80°C. The pellets were thawed in 70 µl TBS (150 mM NaCl, 50 mM Tris-HCl, pH 7.4) supplemented with 1 mM PMSF (phenylmethanesulfonylfluoride) plus a 1∶250 dilution of Calbiochem Protease Inhibitor Cocktail IV (Calbiochem, La Jolla, CA) and lysed by vortexing with glass beads at 4°C. The protein concentrations of the cell lysates were determined by Bradford assay. The samples were diluted in an equal volume of urea-supplemented 2× sample buffer (50 mM Tris-HCl, pH 6.8, 8 M urea, 5% SDS, 1 mM EDTA, plus a 1∶20 dilution of a 1% bromophenol blue stock and a 1∶20 dilution of β-mercaptethanol), heated to 95°C for 5 min and ∼45 µg of total protein per well were separated by SDS-PAGE followed by transfer to nitrocellulose. Poly-ubiquitinated species were detected by Western blotting with anti-ubiquitin antibody clone Ubi-1 (Millipore, Billerica, MA) diluted 1∶500 in TBST (TBS plus 0.05% Tween 20) plus 5% dehydrated milk, followed by incubation in a secondary, HRP-conjugated goat anti-mouse IgG antibody (Sigma-Aldrich, St. Louis, MO) diluted 1∶1000 in TBST plus milk and detection with Supersignal (Thermo Scientific, Rockford, IL).
Yeast actin was purified by a modified [44] DNaseI affinity purification procedure [45]. Briefly, 100 mg DNaseI (Roche Diagnostics, Indianapolis, IN) was coupled to 3 g of swelled Sepharose 4B (Sigma-Aldrich, St. Louis, MO). The coupled beads were packed into two disposable polypropylene columns of 5 ml maximum capacity (PIERCE/ThermoScientific, Rockford, IL). Each column was washed with 25 ml 0.2 M NH4Cl in G-buffer (10 mM Tris, pH 7.5, 0.2 mM CaCl2, 0.5 mM ATP and 0.2 mM DTT) followed by 25 ml G-buffer with 0.1 mM phenylmethylsulfonyl fluoride (PMSF). ∼100 g of yeast pellet (Red Star Yeast from an ∼400 g brick) were thawed in ∼100 ml G-buffer with 0.1 mM PMSF plus Calbiochem protease inhibitor cocktail IV diluted 1∶500. The cells were passed 8 times through a micro-fluidizer (Microfluidics, Model 110 L, Newton, MA). The lysate was clarified in a Beckman JA-20 rotor at 12,000 rpm for 30 min at 4°C and then in a Beckman Ti70 rotor at 50,000 rpm for 50 min at 4°C, followed by filtration through common coffee filters. The filtered supernatant was split between the two columns and loaded at a flow rate of 1–2 ml/min. Each column was washed with 25 ml of 10% deionized formamide in G-buffer, 25 ml of 0.2 M NH4Cl in G-buffer and 25 ml of G-buffer. The actin was eluted with 50% deionized formamide in G-buffer. The protein was dialyzed overnight in dialysis tubing (diameter 11.5 mm) with a molecular weight cut-off of 3,500 Da (Spectrum Laboratories, Rancho Dominiguez, CA) against 2 liters of G-buffer (25 µM ATP). The actin was further purified by FPLC on a 1 ml MonoQ column (Pharmacia Biotech, Piscataway NJ), washed with 10 ml of G-buffer at a flow rate of 0.5 ml/min, and then eluted with a linear, 0–300 mM KCl gradient in G-buffer. The actin eluted above 200 mM KCl; peak fractions (as estimated by the chromatograph and verified by SDS-PAGE gel) were pooled, distributed in 11×34 mm Beckman ultracentrifuge tubes (1 ml per tube) and polymerized in F-buffer (G-buffer+ATP to 1 mM+1 mM EGTA+4 mM MgCl2) for 20 min at room temperature. The KCl concentration was then brought up to 600 mM and the samples were incubated for 1 h at room temperature. Polymerized actin was pelleted by ultracentrifugation in a Beckman TLA100.2 rotor at 90,000 rpm for 30 min. Pellets were re-suspended in G-buffer to 1–2 mg/ml (as estimated by PAGE), incubated on ice for 2–4 hours and dialyzed overnight in 1 liter G-buffer. Protein concentration was determined by Bradford assay.
Yeast 20S core proteasomal subunits were purified by an affinity purification procedure [33] as follows. ∼10 liters of strain RJD1144 [33] expressing flag-tagged Pre1p were grown overnight in synthetic medium, harvested, resuspended in 1 pellet volume lysis buffer (50 mM Tris, pH 7.5, 150 mM NaCl, 5 mM MgCl2), and frozen at −20°C. The cells were thawed in a room temperature water bath and then passed through a microfluidizer (Microfluidics, Model 110 L, Newton, MA) 10 times. The lysate was clarified in a Beckman JA-20 rotor at 17,000 rpm for 20 min and loaded onto a column packed with 1 ml anti-Flag M2 agarose beads (Sigma, St. Louis, MO). The column was washed with 50 ml lysis buffer+0.2% Triton, followed by two 50 ml washes with lysis buffer. The protein was eluted with elution buffer [25 mM Tris, pH 7.5, 150 mM NaCl, 50 mM MgCl2, 100 µg/ml FLAG peptide (Sigma, St. Louis, MO)], four 3 ml elutions were collected, the column was washed three times with 3 ml elution buffer minus the peptide and these fractions were collected as well. Protein concentrations were measured by Bradford assay, the fractions were pooled and the sample was dialyzed in dialysis tubing (diameter 11.5 mm) with a molecular weight cut-off of 3,500 Da (Spectrum Laboratories, Rancho Dominiguez, CA) against 1 liter G-buffer. The sample was then concentrated in a Vivaspin 15R concentrator (Sartorius Stedim Biotech, Bohemia, NY) and protein concentration was measured by Bradford assay.
For purification of the 19S proteasomal cap the same procedure was followed using strain RJD1171 (Table 1, [33]), which expresses flag-tagged Rpt1p.
To evaluate F-actin-proteasome interactions, 4 µM yeast actin was left for 1 h at room temperature to reach equilibrium in G- (5 mM HEPES, pH 7.5, 0.2 mM CaCl2, 0.2 mM ATP, 0.5 mM DTT) or F-buffer (27.5 mM HEPES, pH 7.5, 0.2 mM CaCl2, 25 mM KCl, 1 mM EGTA, 4 mM MgCl2, 0.7 mM ATP, 0.5 mM DTT) in the presence of either ∼65 nM 19S regulatory particle or 50–100 nM 20S core particle in 125 µl total volume. The reactions were submitted to high-speed centrifugation (70,000 rpm for 30 min in a TL100 ultracentrifuge), followed by SDS-PAGE and SYPRO Ruby (Bio-Rad, Hercules CA) staining of the total reaction (T), supernatant (S) and pellet (P). Note that the supernatant fractions were taken from the top of the centrifugation reactions.
|
10.1371/journal.pbio.1002305 | Mapping the Free Energy Landscape of PKA Inhibition and Activation: A Double-Conformational Selection Model for the Tandem cAMP-Binding Domains of PKA RIα | Protein Kinase A (PKA) is the major receptor for the cyclic adenosine monophosphate (cAMP) secondary messenger in eukaryotes. cAMP binds to two tandem cAMP-binding domains (CBD-A and -B) within the regulatory subunit of PKA (R), unleashing the activity of the catalytic subunit (C). While CBD-A in RIα is required for PKA inhibition and activation, CBD-B functions as a “gatekeeper” domain that modulates the control exerted by CBD-A. Preliminary evidence suggests that CBD-B dynamics are critical for its gatekeeper function. To test this hypothesis, here we investigate by Nuclear Magnetic Resonance (NMR) the two-domain construct RIα (91–379) in its apo, cAMP2, and C-bound forms. Our comparative NMR analyses lead to a double conformational selection model in which each apo CBD dynamically samples both active and inactive states independently of the adjacent CBD within a nearly degenerate free energy landscape. Such degeneracy is critical to explain the sensitivity of CBD-B to weak interactions with C and its high affinity for cAMP. Binding of cAMP eliminates this degeneracy, as it selectively stabilizes the active conformation within each CBD and inter-CBD contacts, which require both cAMP and W260. The latter is contributed by CBD-B and mediates capping of the cAMP bound to CBD-A. The inter-CBD interface is dispensable for intra-CBD conformational selection, but is indispensable for full activation of PKA as it occludes C-subunit recognition sites within CBD-A. In addition, the two structurally homologous cAMP-bound CBDs exhibit marked differences in their residual dynamics profiles, supporting the notion that conservation of structure does not necessarily imply conservation of dynamics.
| Cyclic adenosine monophosphate (cAMP) is a messenger molecule produced within cells to control cellular metabolism in response to external stimuli. Protein Kinase A (PKA) is the major receptor for cAMP. cAMP binds to tandem cAMP-binding domains (CBD-A and -B) within the regulatory subunits of PKA (R), unleashing the activity of the catalytic subunit (C). While CBD-A is required for C-subunit inhibition and activation, in RIα CBD-B functions as a “gatekeeper” domain that modulates the control exerted by CBD-A. However, it is not currently clear how ligand binding and dynamics of CBD-B mediate its gatekeeper function. We comparatively analyzed by Nuclear Magnetic Resonance (NMR) a two-domain construct of the regulatory subunit RIα with no ligand, with cAMP2 bound, and the C-bound form. These data show that both CBDs can exist in a system of uncorrelated conformational selection as both can independently sample activated and inactivated states (in what is known as a nearly degenerate free energy landscape). This explains why both RIα CBDs exhibit a higher cAMP-affinity than other cAMP receptors. Once cAMP has bound, the degeneracy is lost and dissociation of the kinase subunit is promoted through a combination of intra-domain conformational selection and changes in inter-CBD orientation. The proposed model—a double-conformational selection model—provides a general framework to interpret the effect of PKA mutations that have been reported in rare human disorders such as Carney complex and Acrodysostosis.
| Cyclic adenosine monophosphate (cAMP) is an ancient secondary messenger, and in higher eukaryotes, Protein Kinase A (PKA) is the major receptor for cAMP. The cAMP-dependent activation of PKA is utilized by a wide variety of extracellular stimuli to control the respective intra-cellular responses, such as regulation of the immune system and cell proliferation [1–4]. In the resting state, PKA exists as an inhibited tetramer formed by a dimeric regulatory subunit (R2), with each R-protomer binding and inhibiting one equivalent of catalytic subunit (C). Upon binding of four equivalents of cAMP to R2, the R2C2 tetramer at least partially dissociates, releasing the active C-subunit to phosphorylate downstream substrates [5–10]. The R-subunit of PKA is a multi-domain protein (Fig 1A), starting from a dimerization-docking domain followed by a flexible linker, which includes an inhibitory site for C and is in turn followed by two tandem cAMP-binding domains (CBD-A and -B) that provide additional contact sites for binding the C-subunit [11,12]. The RIα (91–379) construct spans the inhibitory site and both CBDs (Fig 1A) and recapitulates most of the features that are associated with both C-inhibition and cAMP-dependent activation of full-length PKA [13–15]. The structures of this construct have been solved in both the cAMP2-bound (i.e., active wild type [WT]) [14] and C-bound forms (i.e., inactive R333K mutant) [13], revealing major conformational changes that underlie the cAMP-dependent activation of PKA (Fig 1C). Furthermore, the dynamics of apo CBD-A have been recently shown to be another central determinant of the allosteric control of PKA activation by cAMP [16]. However, less is known about the dynamics of CBD-B and how they relate to the physiological function of this domain.
The CBD-B serves a pivotal function in both PKA inhibition and activation. The structure of C-bound RIα (91–379) has revealed that CBD-B contributes to the effective inhibition of the C-subunit, as the N-terminal helix of CBD-B provides a key site for binding C [13]. The structure of cAMP2-bound RIα (91–379), as well as mutagenesis and Markov state models, has shown that CBD-B also facilitates the cAMP-dependent release of C from R by functioning as a “gate-keeper” domain [10,17,18] that controls the availability of a Trp indole (i.e., W260), which stacks against the adenine of cAMP bound to CBD-A and acts as a lid for CBD-A [10,14,15,18,19]. However, as in the case of CBD-A [16], the cAMP- and C-bound structures of CBD-B are not sufficient for a full thermodynamic dissection of the central function of CBD-B, which requires also the investigation of apo CBD-B.
The currently available evidence suggests that apo CBD-B is highly dynamic. The structure of RIα (91–379) with cGMP bound to site A and an apo site B revealed overall elevated B-factors for CBD-B relative to CBD-A [20]. Furthermore, a small-angle X-ray scattering (SAXS) investigation of the R:C complex indicated that the orientation of CBD-B relative to CBD-A and the C-subunit is dynamic [21,22]. In addition, molecular dynamics (MD) simulations confirmed that the helix connecting CBD-A and -B is flexible in the absence of cAMP [23]. Overall, these initial results suggest that dynamics of nucleotide-free CBD-B are functionally critical, although currently the relationship between apo CBD-B dynamics and its function is not fully understood [24,25].
Here, we hypothesize that apo CBD-B pre-samples both active and inactive conformations with the binding of cAMP or the C-subunit selecting for the former or the latter, respectively. In order to test this hypothesis about conformational selection in CBD-B and to understand how the intra-CBD dynamics affect inter-CBD interactions, we comparatively analyzed by Nuclear Magnetic Resonance (NMR) the two-domain construct RIα (91–379) in its apo, cAMP2, and C-bound forms. The latter is the most technically challenging form of RIα (91–379) both in terms of sample preparation and NMR data acquisition [26–31], given the high MW (i.e., ~73 kDa), and was lacking in previous NMR investigations [32,33]. However, C-bound RIα (91–379) is included here because it plays a critical role in our comparative NMR analysis scheme as an essential reference term for the inactive state of R. In addition, the present comparative analyses include mutations designed to specifically probe inter-domain interactions [34,35] and provide an opportunity to re-assess previous results on CBD-A [16,36] in the context of a more complete construct with both tandem CBD-A and -B.
Our comparative NMR analysis supports a model of double conformational selection in which each apo CBD populates a nearly degenerate free energy landscape independently of the adjacent CBD. Such degeneracy is unusual for allosteric systems, which typically sample highly skewed conformational equilibria in their apo form [37–42], but it explains why even weak interactions with the C-subunit are sufficient for conformational selection, as well as why both RIα CBDs bind cAMP with higher affinity than other eukaryotic CBDs. Binding of two equivalents of cAMP eliminates this degeneracy by selecting the active conformation within each CBD and stabilizing the inter-CBD contacts. We also show that the CBD-A/B interface requires both cAMP and the indole of W260 to form, and, although it is not required for the intra-domain conformational selection, it contributes to the activation of PKA by occluding C-subunit recognition sites within CBD-A. Hence, the proposed model explains how CBD-B contributes to PKA inhibition and activation. In addition, our investigation revealed that cAMP-binding does not fully quench the dynamics of the CBDs, with major loci of residual dynamics found both at the inter-CBD interface and within each tandem CBD. Surprisingly, the residual dynamics profiles of the structurally homologous CBD-A and -B are different. This unexpected finding supports the notion that conservation of structure does not necessarily imply conservation of dynamics and opens new opportunities in the selective targeting of CBDs for therapeutic purposes.
As a first step towards mapping the functional conformational equilibria of apo PKA RIα (91–379), we compared it to two reference states, i.e., C-bound RIα (91–379), which is assumed to trap the inhibitory (“inactive”) conformation of RIα (91–379), and cAMP2-bound RIα (91–379), which is expected to represent the uninhibited (“active”) form of PKA-R. The comparative spectral analysis of apo versus C- versus cAMP2-bound RIα (91–379) is illustrated in Fig 2. Panels 2A and 2B focus on two representative residues of CBD-A and -B, i.e., G223 and S297, respectively, which are sufficiently removed from the cAMP and C-subunit binding interfaces to report primarily on the conformational equilibria of RIα (91–379). For both residues, the cross-peaks arising from the apo, C, and cAMP2-bound forms define a clear linear pattern, which is indicative of apo RIα (91–379) sampling two states (i.e., the active and inactive conformations) through an exchange that is fast in the chemical shift NMR time scale at the field of data acquisition (i.e., 700 MHz).
Given the fast regime for the active versus inactive exchange, the apo cross-peak position is simply a population-weighed linear average of the chemical shifts for the active and inactive conformations. Hence, the chemical shifts encode directly the relative populations of the active versus inactive equilibrium. For example, the intermediate position of the apo peaks of G223 and S297 relative to those of the C and cAMP2-bound forms (Fig 2A and 2B) suggests that in each apo CBD the populations of active and inactive states are comparable. In order to verify that this result does not depend on the specific choice of reporter residues (e.g., G223 and S297), the apo CBD active versus inactive populations were also assessed through the CHEmical Shift Projection Analysis (CHESPA) (Fig 2C) [37] and through complementary chemical shift correlations (Fig 2D and 2E). Both approaches rely on a wider collection of conformational equilibria-sensing residues as opposed to only one residue per CBD, as shown in Fig 2A and 2B.
The CHESPA analysis (Fig 2C) shows that, if C- and cAMP2-bound RIα (91–379) represent the inactive and active states of PKA, the average fractional activation observed for CBD-A is ~50%, with slightly higher values measured for CBD-B. The differences in average fractional activations between CBD-A and -B are not major as they are within one standard deviation of both distributions (Fig 2C). Similar conclusions are obtained through chemical shift correlation analyses (Fig 2D and 2E). Panels 2D and 2E report the (δApo − δcAMP) versus (δC − δcAMP) chemical shift correlations for apo CBD-A and -B, respectively. The linearity of the correlations in Fig 2D and 2E confirms that each CBD samples a dynamic two-state equilibrium of active and inactive conformations. In addition, the slopes of the correlations provide the molar fraction of the inactive state, i.e., ~50% and ~70% for CBD-A and -B, respectively, corroborating that each apo CBD accesses significant populations of both inactive and active conformations, although the former appears slightly higher for CBD-B than CBD-A. However, the somewhat higher population of inactive state in apo CBD-B versus apo CBD-A may also reflect the fact that in the R:C complex the CBD-B does not interact with the C-subunit as tightly as CBD-A. The transient nature of the CBD-B interactions with the C-subunit may shift the R:C cross-peaks of CBD-B towards the apo cross-peaks, resulting in an increase in the apparent fractional inactivation measured for apo CBD-B. In this respect, the actual population of inactive state in apo CBD-B is possibly lower than the measured 70%, confirming that the CBD-B samples significant (i.e., >30%) populations of active state even prior to cAMP binding.
It should be noted that even if we knew the exact value for the fractional inactivation of CBD-B, the data in Fig 2 alone would not be sufficient to gauge the degree of correlation between the active–inactive conformational transitions in the two tandem domains. For example, if the actual fractional inactivation of each CBD is 50%, two markedly different distributions of state populations are still possible: (1) 50% of active-active state, 50% of inactive-inactive state, and no mixed (inactive-active and active-inactive) states, or (2) 25% of each possible state. In the two scenarios the degree of correlation between the active-inactive transitions in the two domains is clearly different, but in both cases the fractional inactivation for each CBD is 50%. Hence, it is clear that understanding how the inactive versus active intra-CBD equilibria are coupled to each other requires an independent assessment of the CBD-A/B interactions in the context of a combinatorial analysis of CBD states (i.e., active-active, inactive-active, active-inactive, and inactive-inactive CBD-CBD states). For this purpose, we dissected the state-specificity of inter-domain interactions by probing the effect of CBD-B deletion on CBD-A, i.e., by comparing the RIα (91–379) versus RIα (91–244) constructs (Fig 3).
Fig 3 reports on the comparative NMR analysis of the RIα (91–379) versus RIα (91–244) constructs in their apo versus C- versus cAMP2-bound forms. Fig 3A shows that the cross-peak positions of the RIα (91–244):C complex do not change significantly relative to RIα (91–379):C, which includes the CBD-B as well, suggesting that inter-CBD interactions are negligible in the C-bound state. This conclusion is also supported by the chemical shift similarity between the RIα (91–244):C and RIα (91–379):C complexes (Fig 3B). Considering that the C-subunit of PKA is selective for the inactive-state of both CBDs, panels 3A and 3B point to negligible inter-CBD interactions when both CBDs adopt the inactive conformation (“inactive-inactive” state).
Fig 3C and 3D re-examines the RIα (91–379) versus RIα (91–244) comparison shown in panels 3A and 3B, but in the absence of both C-subunit and cAMP, i.e., for the apo forms of both constructs. Again the majority of the residues of CBD-A do not experience significant chemical shift changes upon deletion of CBD-B (Fig 3C and 3D), suggesting negligible inter-CBD interactions in apo RIα (91–379). Considering that each CBD in apo RIα (91–379) samples comparable populations of the respective active and inactive states, the absence of detectable inter-CBD contacts in apo RIα (91–379) suggests that in the absence of cAMP inter-CBD interactions are negligible not only in the “inactive-inactive” state but also in the remaining three combinations, i.e., the “active-inactive,” the “inactive-active,” and the “active-active” states. Overall, our data indicate that CBD-A and -B are largely independent of each other when cAMP is absent.
Unlike the cases of apo or C-bound RIα (91–379) (Fig 3A–3D), in the presence of excess cAMP the deletion of CBD-B:cAMP causes extensive chemical shift changes in CBD-A:cAMP (Fig 3E and 3F), pointing to the presence of significant inter-domain interactions. Interestingly, the residues in CBD-A:cAMP experiencing chemical shift variations upon deletion of CBD-B:cAMP (Fig 3G) are similar to those subject to solvent accessible surface area (SASA) variations upon deletion of RIα (226–379), which includes CBD-B (Fig 3H). The similarity of the residue maps in panels 3G and 3H suggests that the chemical shift differences observed in Fig 3E and 3F reflect primarily the disruption of inter-domain contacts upon CBD-B:cAMP deletion as opposed to possible changes in the intra-CBD active versus inactive equilibria. Based on this result, we hypothesized that the contribution of domain-domain interactions to the conformational selection of the active state in both CBDs is negligible compared to the contribution of cAMP. In order to test this hypothesis, we further probed the inter-domain interactions within RIα (91–379):cAMP2 through the W260A point mutation. W260 belongs to CBD-B but, through aromatic stacking, contacts cAMP bound to CBD-A (Fig 1C). Hence, W260 is expected to play a pivotal role in the cAMP-dependent CBD-A/B interactions [13].
The W260A mutation causes major chemical shift changes in the presence of excess cAMP (Fig 4A, green bars) and these changes mimic those caused by the deletion of the whole CBD-B (Fig 4B). These chemical shift patterns suggest the single point W260A mutation is sufficient to disrupt the CBD-A/B interface. The disruption of the CBD-A/B interface caused by the W260A mutation was also independently confirmed through 15N-relaxation measurements, which were analyzed in terms of reduced spectral densities (Fig 5). The W260A mutation causes a decrease in the average J(0) value and a concurrent increase in the average J(ωN) value (Fig 5A and 5B, red vertical arrows; Table 1), which are indicative of a reduction of the effective correlation time for overall tumbling, as expected upon de-correlation of the tumbling motions of the two adjacent CBDs, i.e., the W260A mutant conforms better than WT to a model of two domains joined by a flexible linker. This result is independently confirmed by the observation that W260A enhances the flexibility of the helical region connecting the two tandem CBDs, as supported by the fact that several residues in this region experience a reduction of J(0) values (Fig 5A, red dashed oval) and an enhanced solvent exposure (Fig 5A, green arrows) due to the W260A mutation. Hence, the changes in dynamics caused by W260A corroborate the chemical shift perturbation results (Fig 4A and 4B), indicating that the W260A mutation is sufficient to disrupt the CBD-A/B interface.
The observation that a single point mutation (i.e., W260A) compromises the integrity of a whole inter-CBD surface is suggestive of weak inter-CBD interactions, even in the presence of cAMP. In order to further investigate this hypothesis, we inspected the dynamics of the B-C hinge helices, which join CBD-A to CBD-B. As shown in Fig 5A, these helices exhibit J(0) values that are higher than other regions of the two-domain construct, suggesting the presence of residual ms-μs dynamics or diffusional anisotropy in the inter-domain hinge region even in the presence of excess cAMP. Diffusional anisotropy contributions arise in principle from either the main conformation, represented by the structure of cAMP2-bound RIα [14] with the two CBDs in close contact (Fig 1C), or possibly from minor populations of less compact and highly anisotropic conformers in which the CBD contacts are lost, as in the case of the dumbbell-shaped structure of C-bound RIα [13].
The diffusional anisotropy of both types of conformations was assessed through hydrodynamic simulations (Fig 5A, orange and blue traces) and in both cases a negligible contribution of diffusional anisotropy was determined for the elevated J(0) values observed in the B-C hinge helices connecting CBD-A to CBD-B. The observed J(0) values in the CBD-A hinge region are higher than the J(0) values predicted by hydrodynamic modeling based on the structure of the cAMP2-bound two-domain construct in the absence of internal motions (Fig 5A, orange trace; Table 1). Hence, the anisotropy of this compact conformer cannot account for the J(0) values observed in the hinge:A region (Fig 5A). Furthermore, although the diffusional anisotropy is amplified in the dumbbell-shaped structure observed for the C-bound R-subunit (Fig 5A, blue trace), contributions from this type of conformation to the elevated J(0) values observed for the hinge region of CBD-A are unlikely, as other regions (e.g., β8 of CBD-B) should then also display elevated J(0) values, which is not the case (Fig 5A). These considerations confirm that the elevated J(0) values observed for the B-C hinge helices connecting CBD-A to CBD-B reflect primarily ms-μs residual dynamics, which remain even after cAMP-binding to both domains (Fig 5A). The dynamic nature of the B-C helices joining CBD-A and -B calls for an assessment of the role of inter-domain interactions on the intra-CBD conformational equilibria.
The effect of inter-domain interactions on the conformational equilibria of each CBD was probed using the W260A mutation to selectively disrupt the CBD-A/B interface without significantly affecting the structures of the two CBDs, which are largely unperturbed by the mutation (Fig 4E). Specifically, we measured the degree of correlation between the W260A:cAMP2 versus WT:cAMP2, and the WT:C versus WT:cAMP2 chemical shift differences (Fig 4C and 4D). Similarly to the correlations shown in Fig 2D and 2E, this plot was restricted to residues that are sufficiently removed from the cAMP-dependent interfaces (e.g., R:C, R:cAMP, and CBD-A:CBD-B) to report primarily on the active versus inactive equilibrium of each CBD. However, unlike Fig 2D and 2E, no significant correlation was observed in either CBD-A or -B (Fig 4C and 4D), confirming the hypothesis that the contribution of inter-CBD interactions to the intra-CBD conformational selection is marginal compared to the contribution of cAMP (Fig 2). This hypothesis is also independently confirmed by previous results based on the RIα (91–244) CBD-A construct, showing linear cross-peak patterns with the apo form found in a central position, as in Fig 2A for CBD-A in the two-domain construct [16]. Overall, both the W260A mutation and the CBD-B deletion consistently indicate that cAMP-binding alone is sufficient to almost quantitatively stabilize the active-state of each CBD, even in the absence of CBD-CBD contacts. However, cAMP-binding does not result in a full quenching of dynamics and the profile of residual dynamics after cAMP binding defines critical differences between the two structurally homologous CBDs.
The ms-μs residual dynamics detected for the B-C helices is a unique feature of CBD-A, as no noticeable internal dynamics were observed for the B-C helices of CBD-B (Fig 5A). In fact, both J(0) and J(ωH+ωN) spectral densities measured for residues in the B-C helices of CBD-B appear in overall good agreement with the values predicted based on hydrodynamic simulations starting from the cAMP2-bound structure of the R-subunit in the absence of internal motions (Fig 5A and 5C). Hence, the hinge helix dynamic profile appears to be highly domain-specific. Another region subject to a domain-specific dynamic profile is the base-binding region (BBR). The BBR of CBD-B exhibits elevated J(ωH+ωN) values relative to other regions of CBD-B, with the exception of the C-terminal tail (Fig 5C and Table 1), indicating that the BBR:B is flexible in the ps-ns time-scale. Unlike BBR:B, for BBR:A no internal dynamics were detected, as both J(0) and J(ωH+ωN) spectral densities for BBR:A residues appear in agreement with the values expected based on rigid-body hydrodynamic modeling (Fig 5A and 5C; Table 1). Overall, distinct dynamic profiles emerge for CBD-A and -B, in spite of their structural homology. The B-C helices exhibit ms-μs flexibility in CBD-A but not in CBD-B, while the BBR region is subject to ps-ns dynamics in CBD-B but not in CBD-A (Fig 5 and Table 1).
The comparative NMR analyses of the dynamic conformational equilibria of CBD-A and -B presented here support a “double conformational selection” model for the cAMP-dependent activation of PKA RIα (Fig 6A). In the absence of cAMP the PKA-RIα region spanning the two tandem cAMP-binding domains (CBD-A and -B), denoted as RAB, samples four states populating a nearly degenerate free-energy landscape, in which each CBD accesses both C-binding competent (“inactive”) and cAMP-binding incompetent (“active”) conformations with comparable populations, as shown in Fig 2. In the absence of cAMP, such active versus inactive sampling of each CBD occurs independently of the adjacent CBD, due to the absence of significant inter-domain interactions, as shown in Fig 3A–3D. Out of these four nearly degenerate states, the “inactive-inactive” state exhibits the highest affinity for the C-subunit, as both domains are primed for binding C and the inter-domain helical region is also available for interacting with C. The CBD-B affinity for the C-subunit is known to be lower than that of CBD-A, but weak interactions between CBD-B and the C-subunit are still sufficient to drive the conformational equilibria towards the inactive state due to the apo RAB near degeneracy.
The free energy near-degeneracy of apo RAB is eliminated upon binding of two equivalents of cAMP, which select the active state in each CBDs (“active-active” state) and promote CBD-CBD interactions through the lid capping exerted by W260, which is essential to preserve the integrity of the whole CBD-A/B interface, as show in Fig 4A and 4B. The CBD-A/B interface stabilized by cAMP contributes to the dissociation of the R:C complex, because selected CBD-A/B contacts overlap with the R:C interface, as revealed by the SASA difference analysis of Fig 6B and 6C (dashed rectangles). In addition, the inter-domain interactions stabilized by cAMP result in bending of the C-helix, thus further reducing the affinity of RAB for the PKA catalytic subunit (C), which prefers an elongated unbent C-helix [13]. Overall, the emerging evidence suggests that cAMP contributes to the R:C dissociation through three distinct but concurrent mechanisms: (a) selection of active conformation of CBD-A, which exhibits low affinity for C; (b) selection of active conformation of CBD-B, which further reduces the affinity for C; (c) stabilization of inter-domain interactions that are incompatible with the R:C interface.
Mechanism (a), i.e., intra-CBD-A conformational-selection, significantly contributes to the activation of PKA by cAMP, given the high affinity for C of CBD-A in the inactive state [43,44]. However, the contributions of (b) and (c) are essential to explain why CBD-B significantly lowers the Ka for the cAMP-dependent activation of PKA by more than order of magnitude from 1 μM to 80 nM [36]. Although the deletion of CBD-B does not appreciably perturb the affinity of R for C [43], mechanism (b), i.e., intra-CBD-B conformational-selection to stabilize the active state of CBD-B, is still pivotal to promote the inter-CBD interactions, which contribute to the release of C through mechanism (c). Interestingly, even in the presence of cAMP the free energy of inter-domain interaction is marginal compared to the free energy of cAMP binding, as CBD-B deletion does not significantly reduce the free energy of unfolding of CBD-A [36,45–47] or of cAMP-binding to CBD-A [16,18,48,49]. Hence, the “closed” inter-domain topology of RAB:cAMP2, in which W260 from CBD-B caps cAMP in CBD-A, exists in a dynamic equilibrium with minor populations of an “open” active-active state, in which inter-domain coupling is negligible (Fig 6A). Since intra-domain conformational selection in RAB relies primarily on cAMP binding with only negligible contributions from domain-domain interactions (Fig 4C and 4D), in such an “open” topology of RAB:cAMP2, which is mimicked by the W260A mutant, the active state selection within each CBD is not compromised, i.e., mechanisms (a) and (b) are still effective, but mechanism (c) is not activated.
A critical feature of the double conformational selection model proposed here to explain the functional role of CBD-B (Fig 6A) is the near free energy degeneracy of the states sampled by apo RAB. This degeneracy is quite unique of PKA RIα, as it was not observed for other structurally homologous CBDs, such as those of the hyperpolarization and cyclic nucleotide activated (HCN) ion channels, for which conformational selection relies on apo equilibria that are skewed towards the inactive (auto-inhibited) state with typically low affinity for cAMP [39,50]. This observation provides a possible explanation as to why both CBDs of PKA RIα bind cAMP with higher affinity than HCN, in spite of the structural homology among eukaryotic CBDs. The higher cAMP-affinity exhibited by PKA versus HCN provides a basis to understand the reduced off-rate for cAMP, explaining why cAMP release from PKA RIα would become the rate-determining step of the kinetics of cAMP signal termination by phosphodiesterases (PDEs) in the absence of direct PDE-PKA RIα interactions [51–53].
Furthermore, our investigation revealed marked differences in dynamics also among homologous CBDs in their cAMP-bound forms. For example, in both CBD-A and -B of PKA RIα the adenine base of cAMP is sandwiched by β-strands 4 and 5 (Base Binding Region or BBR) on one side and on the other side by an aromatic side chain lid (i.e., W260 in CBD-A and Y371 in CBD-B), which is part of hinge helices that are C-terminal to the β-subdomains of the respective CBDs. Despite the fact that this structural pattern of cAMP recognition through the BBR and hinge helix-lid motifs is conserved in both CBDs, the BBR:B is significantly more dynamic in the ps-ns timescale than the BBR:A, while the hinge:A is significantly more dynamic in the ms-μs timescale than the hinge:B (Fig 6A). The hinge:A dynamics explain why CBD-A in RIα tolerates, without major affinity losses, cAMP analogues with bulky substituents at the adenine C8 better than anticipated based on the narrow pocket observed in the static structure [54]. In general, these observations corroborate the notion that structurally homologous domains may still exhibit marked differences in dynamic profiles that affect affinities and recognition [55], opening new opportunities for the design of selective ligands [56] that target specific eukaryotic CBDs.
The proposed model (Fig 6) also provides an initial framework to dissect the molecular mechanism underlying disease related PKA RIα mutants. Several mutations reported for PKA RIα CBDs have been linked to the Carney complex and to Acrodysostosis and result in either de-regulation or over-regulation of PKA kinase activity [57–59]. Based on the model of Fig 6, it will be critical to evaluate how such mutations perturb the apo and the cAMP-bound inhibitory equilibria of both CBDs. Perturbations in the apo active versus inactive equilibria lead to changes in cAMP-affinities up to three-orders of magnitude [60], while perturbations in the holo inhibitory equilibria will modulate the inter-CBD interaction which is formed preferentially in the holo/active-holo/active state. Considering the competition between inter-CBD and R:C interactions (i.e., mechanism [c], mentioned above), the modulation of the inter-CBD interaction is expected to in turn alter the Ka for the cAMP-dependent activation of PKA. In general, it is anticipated that the model proposed in Fig 6 will assist in rationalizing how PKA dysfunctional dynamics results in dysregulation of PKA and disease.
All PKA RIα constructs as well as the PKA C-subunit were expressed and purified according to previously published protocols [16,32]. The W260A mutant was prepared by site-directed mutagenesis.
All NMR spectra were recorded at 306K, unless otherwise specified, using a Bruker AV 700 spectrometer equipped with a TCI cryo-probe and processed with NMRpipe [61] employing linear prediction, unless otherwise specified, and a resolution enhancing 60° shifted sine squared bell window function. All spectra were analyzed with Sparky [62] using Gaussian line-fitting. Assignments were obtained either through triple-resonance 3-D experiments (i.e., HNCO, HNCA, HN(CO)CA, CBCA(CO)NH and HNCACB, in a TROSY-version for high MW constructs) [63,64] and/or through spectral comparisons, if no ambiguities were present. The secondary structure probabilities were determined using the secondary chemical shifts via the PECAN software [65]. Other NMR experiments are discussed below.
Uniformly 2H,15N-labeled PKA RIα (91–379) and (91–244) were concentrated to 100 μM in the NMR buffer (50 mM MOPS, pH 7.0, 100 mM NaCl, 10 mM MgCl2, 5 mM DTT, with or without 1 mM cAMP, 0.02% sodium azide, and 5% 2H2O). The C subunit-bound RIα (91–379) and (91–244) complexes were prepared with 1 mM AMP-PNP in the NMR buffer as previously described [16]. Transverse-relaxation optimized spectroscopy (TROSY) 2-D experiments with 80 (t1) and 1,024 (t2) complex points and spectral widths of 31.82 ppm and 15.94 ppm for the 15N and 1H dimensions, respectively, were recorded with 12 scans and a recycle delay of 1.70 s. Sensitivity enhanced 15N-1H hetero-nuclear single quantum coherence (HSQC) spectra with 128 (t1) and 1,024 (t2) complex points and spectral widths of 31.82 ppm and 15.94 ppm for the 15N and 1H dimensions, respectively, were recorded with eight scans and a recycle delay of 1.0 s. The 1H and 15N carrier frequencies were set at the water resonance and at the centre of the amide region, respectively. The C-subunit and cAMP-bound RIα(91–379) forms served as the reference states in the CHEmical Shift Projection Analysis (CHESPA) to evaluate the position of the inhibitory equilibria in the apo RIα(91–379). The combined chemical shifts (CCS), fractional activation, and cos(θ) values were calculated as previously described [37].
Uniformly 15N-labeled PKA RIα (119–379) wild type and W260A mutant were concentrated to 100 μM in 50 mM MES (pH 6.5), 100 mM NaCl, 5 mM DTT, 2 mM EDTA, 2 mM EGTA, 2 mM cAMP, 0.02% sodium azide and 5% 2H2O. The R1 relaxation rates were measured with water flipback and sensitivity enhanced pulse sequences using relaxation delays of 100 (×2), 200, 300, 400(×2), 500, 600, 800, and 1,000 ms [38]. The R2 measurements were acquired using CPMG relaxation delays of 8.48, 16.96, 25.44, 33.92, 42.4, 50.88, 59.36, 76.32, and 93.28 ms with an offset and duty cycle compensated 15N R2 CPMG pulse sequence with an inter-180° pulse delay of 0.9 ms [38]. Both R1 and R2 experiments were acquired as pseudo 3-D datasets with recycle delays of 1.5 s. The NOE spectra were collected as a set of ten replicas. The acquired spectra were then co-added in the time domain prior to Fourier transformation. The R1 and R2 relaxation rates were determined by using cross peak fit heights in Sparky. The errors on the R1 and R2 rates were estimated from the Gaussian distributed random noise. The errors on the NOE values were gauged based on the standard deviation between fit heights in the replicate spectra. The 15N relaxation data were mapped into reduced spectral densities as previously described [66]. The reduced spectral density values in the absence of internal motions were computed through hydrodynamic simulation using the HydroNMR software [67,68] and the cAMP2- and C-bound structures (PDB Codes: 1RGS and 2QCS, respectively, but without C-subunit for comparison with the values from 1RGS). An atomic element radius of 2.8 Å was used in all hydrodynamic simulations. The 15N relaxation experiments were complemented by H/D exchange data based on HSQC spectra acquired as previously described [32] using the wild type and the W260A mutant RIα (119–379) construct. For this purpose, the protein was concentrated to 100 μM in 50 mM MES (pH 6.5), 100 mM NaCl, 5 mM DTT, 2 mM EDTA, 2 mM EGTA, 0.02% sodium azide, and 5% 2H2O. The cAMP bound sample was prepared by adding an excess 100 μM cAMP in the buffer.
|
10.1371/journal.pntd.0002841 | Easy Identification of Leishmania Species by Mass Spectrometry | Cutaneous leishmaniasis is caused by several Leishmania species that are associated with variable outcomes before and after therapy. Optimal treatment decision is based on an accurate identification of the infecting species but current methods to type Leishmania isolates are relatively complex and/or slow. Therefore, the initial treatment decision is generally presumptive, the infecting species being suspected on epidemiological and clinical grounds. A simple method to type cultured isolates would facilitate disease management.
We analyzed MALDI-TOF spectra of promastigote pellets from 46 strains cultured in monophasic medium, including 20 short-term cultured isolates from French travelers (19 with CL, 1 with VL). As per routine procedure, clinical isolates were analyzed in parallel with Multilocus Sequence Typing (MLST) at the National Reference Center for Leishmania.
Automatic dendrogram analysis generated a classification of isolates consistent with reference determination of species based on MLST or hsp70 sequencing. A minute analysis of spectra based on a very simple, database-independent analysis of spectra based on the algorithm showed that the mutually exclusive presence of two pairs of peaks discriminated isolates considered by reference methods to belong either to the Viannia or Leishmania subgenus, and that within each subgenus presence or absence of a few peaks allowed discrimination to species complexes level.
Analysis of cultured Leishmania isolates using mass spectrometry allows a rapid and simple classification to the species complex level consistent with reference methods, a potentially useful method to guide treatment decision in patients with cutaneous leishmaniasis.
| Cutaneous leishmaniasis is a disease due to a small parasite called Leishmania. This parasite causes disfiguring skin lesions that last for months or years. There are many different subtypes of Leishmania, each giving rise to lesions of different severity and responding to therapies in its own way. Treating physicians must know as soon as possible which subtype of Leishmania is involved to propose the best treatment. Because it is impossible to differentiate the Leishmania subtypes microscopically, the identification of the culprit subtype currently requires complex and expensive typing methods, the results of which are generally obtained several weeks after the diagnosis. Here, we have evaluated the ability of a new method using mass spectrometry to differentiate Leishmania subtypes. Our results were consistent with those provided by reference typing methods and were obtained rapidly after the parasite had been cultured in vitro. This new method may help physicians know very soon which Leishmania subtype is involved thereby facilitating treatment choice.
| Cutaneous leishmaniasis (CL) affects 1.5 million patients each year and displays a wide spectrum of clinical forms from small self-resolving papules to severe destructive mucosal lesions. The infecting Leishmania species influence the clinical presentation of CL [1] but lesion characteristics are not specific enough for a robust species determination in a given patient [2]–[4].While 2 species of the Leishmania subgenus - L. major and L. mexicana - are associated with frequent spontaneous cure within a few months [3], the 2 main species of the Viannia subgenus – Leishmania braziliensis and L. panamensis/guyanensis are associated with a 1–15% risk of delayed mucosal metastasis [5]. Considering the variable severity of CL, recent guidelines recommend using local therapy whenever possible and systemic therapy if local therapy fails or cannot be performed [3], [6], [7]. This step-wise decision process integrates not only lesion number and size, patients status (age and co-morbidities), but also the infecting species [8].
The influence of the infecting Leishmania species on treatment outcome is well established [4], [9], [10]. Thus, species identification is important to determine the clinical prognosis and to select the most appropriate therapeutic regimen. In current clinical practice, treatment decision is generally presumptive, the infecting species being suspected on epidemiological and clinical grounds [3] but this approach requires a specific clinical expertise and frequently updated knowledge of the geographic distribution of Leishmania species [3]. A simple, rapid method to type cultured isolates would facilitate an easier and more robust treatment decision based on confirmed species identification.
Available methods to type Leishmania in cultured isolates or directly in lesions are still complex and poorly standardized. At present, isolation of the parasite in culture is necessary for identification by multilocus enzyme electrophoresis (MLEE), which has long been the reference for Leishmania species identification [11] [12] [13]. Only a few specialized centers currently perform MLEE, the result of which is available several weeks after the isolation of the parasite in culture. These difficulties have led to the development of molecular methods for species identification, generally based on DNA amplification by PCR, followed by single or multilocus sequencing (MLST) or restriction fragment length polymorphism analysis [14] or single strand conformation polymorphism or sequencing of different targets including the heat shock protein 70 (hsp70) gene [15], [16]. Some of these methods can be applied directly to biological samples avoiding the parasite culture step [17]. However, these molecular methods lack inter-laboratories standardization and require the use of expensive reagents.
Matrix-assisted laser desorption ionization–time-of-flight (MALDI-TOF) mass spectrometry (MS) has emerged as a powerful tool for the identification of microorganisms. Using MALDI-TOF MS, the protein spectral “fingerprint” of an isolate is compared to a reference spectral database, yielding results within 1 hour [18]. Although spectrometers are still relatively expensive, the initial investment is justified by a broad use spanning a wide diversity of microbiological samples [19], and the cost of reagents is very limited. This approach has been applied with success to bacteria, yeasts and filamentous fungi, but to our knowledge, no study on direct identification of protozoans has been published yet [20]–[23]. We have investigated the value of mass spectrometry MALDI-TOF for the identification of Leishmania species in patients with CL.
From 2011 through 2013, data and samples were collected each time treatment advice was sought from an expert at our hospital for patients with CL. Diagnosis procedures were not modified by the process, expert treatment advice was part of normal medical care, data and sample collection was in the context of national health surveillance. Patients were informed of the process by their attending physician using a procedure common to all French National Reference Centers (NRC) (http://www.parasitologie.univ-montp1.fr/doc/Declaration_pub_2011.pdf). Data were obtained through the standard reporting form of the NRCL. This form is available online and is anonymous and the anonymisation process is irreversible. The following characteristics are provided on the form: age (children defined as <16 years), sex, clinical form, and for CL or MCL: number of lesions, the presumed place of infection. The collection of parasite isolates was performed in the context of this surveillance program.
Parasitological diagnosis was performed and analyzed as previously described by lesion scraping, biopsy or aspirate followed by direct examination of Giemsa-stained smears, histological analysis of HES- or Giemsa-stained tissue sections, culture or PCR [24]. To increase the robustness of the analysis, 10 New World isolates were obtained from the Tropical institute in Antwerpen (ITM, Belgium), all strains were re-suspended in 10% glycerol and stored in liquid nitrogen until use.
Needle aspirate of skin lesions was performed under local anesthesia then cultured in both Nicolle-McNeal-Novy (NNN) medium and Schneider medium supplemented with 20% fetal calf serum, penicillin and streptomycin [25]. Cultures were kept at 25°C and observed under an inverted microscope in search for motile promastigotes, twice a week for 1 month. Each week positive culture was expanded in 20 mL glass bottles with Schneider medium (20% fetal calf serum, penicillin and streptomycin), one part was frozen in liquid nitrogen −80°, and the other was used for species identification. Aliquots were thawed immediately at 37°, re-suspended in 5 ml of Schneider medium (20% FCS, penicillin streptomycin) and incubated à 27°C. Subcultures were counted daily and analysed at the end of 3-day growth period (growth period being defined à> = ×3 fold multiplication over 3 day), a growth curve was established for each isolate to perform the proteomic analysis during the exponential growth or early stationary phase. For Leishmania strains isolated only in NNN medium, promastigotes were concentrated by centrifugation (2500 rpm×10 minutes) and resuspended in 20% HS Schneider medium 24–72 hours before proteomic analysis.
Positive cultures were sent to the NRCL for confirmation and species identification using a multilocus sequence typing (MLST) approach based on the analysis of seven single copy coding DNA sequences [26]. Isolates from ITM had already been typed by hsp 70 sequencing [15], [16].
Promastigote suspensions from the expanded cultures were centrifuged 3000 g for 3 minutes and the supernatant removed before the pellet was washed twice in pure water, the pellet was then re-suspended in 300 µL of pure water before adding 900 µL of ethanol. After another round of centrifugation, 10 µL of 70% formic acid and 10 µL pure acetonitrile were added to the residual pellet and the subsequent solution was repeatedly and thoroughly vortexed before a final centrifugation. Each centrifugation step was performed at 10 000 g for 2 min at room temperature.
The supernatant was distributed (0.5 µl droplet) in duplicates on a MALDI ground steel sample slide (Bruker-Daltonics, Bremen, Germany) then air-dried. The α-cyano-4-hydroxy-cinnamic acid (CHCA) matrix (Bruker-Daltonics), prepared at a concentration of 50 mg/ml in 50% acetonitrile and 50% water with 2.5% TFA, was sonicated for 5 min before being spotted (0.5 µl) over the dried sample. A DH5-alpha Escherichia coli protein extract (Bruker-Daltonics) was deposited on the calibration spot of the sample slide for external calibration. MALDI analysis were performed on a BrukerAutoflex I MALDI TOF mass spectrometer with a nitrogen laser (337 nm) operating in linear mode with delayed extraction (260 ns) at 20 kV accelerating voltage. Each spectrum was automatically collected in the positive ion mode as an average of 500 laser shots (50 laser shots at 10 different spot positions). Laser energy was set just above the threshold for ion production. A mass range between 3,000 and 20,000 m/z (ratio mass/charge) was selected to collect the signals with the AutoXecute tool of flexcontrol acquisition software (Version 2.4; Bruker-Daltonics). Only peaks with a signal/noise ratio >3 were considered. Spectra were eligible for further analysis when the peaks had a resolution better than 600. For each cultivation condition, we collected mass spectra from 2 biological replicates and 4 technical replicates.
Data were processed with Biotyper version 1.1 (Bruker-Daltonics) and ClinProTools 3.0 (bruker-Daltonics) as described [18]. ClinProTools software was used to visualize all spectra as virtual gels and to calculate variability for each of the defined markers. The Biotyper software performs smoothing, normalisation, baseline subtraction, and peak picking using default parameters. Strain comparison was done by principal component analysis (PCA) [27]. Distance values were calculated using Biotyper to build score-oriented dendrograms. Based on these distance values, a dendrogram was generated using the according function of the statistical toolbox of Matlab 7.1 (The MathWorks Inc., USA), which was integrated into Biotyper 1.1. The clustering approach is based on similarity scores implemented in the software.
Reproducibility was evaluated by comparing spectra obtained from two independent experiments for each strain. The repeatability and stability of the profiles over generations was tested using a series of extracts obtained from subcultures. One strain was maintained in 2 separate cultures then analyzed in duplicate every 72 h over 5 weeks.
Of the 46 isolates analyzed, 25 (54%) were from the Old World and 21 (46%) from the New World (Table 1). L. major was predominant among Old World isolates (16 isolates/64%), followed by L. donovani (3/12%), L. tropica (3/12%), L. infantum (2/8%), and L. killicki (1/4%). One additionnal isolate from a patient with diffuse cutaneous leishmaniasis acquired in Martinique (French West Indies) related to a trypanosomatid described previously in a very small number of patients [28], [29] was also analyzed. L. braziliensis/L. peruviana were predominant among New World isolates (13 isolates/62%), followed by L. guyanensis/L. panamensis (5/24%), and L. mexicana/L. amazonensis (3/14%). Ten isolates were from ITM in Antwerpen (Belgium), 15 from the Parasitology laboratory at Pitié-Salpêtrière Hospital (Paris, France), 1 from the clinical laboratory at Institut Pasteur (Paris, France). Twenty were collected and analyzed blindly from French travelers (19 with cutaneous leishmaniasis and 1 with visceral leishmaniasis) at the Pitié-Salpêtrière Hospital as part of the routine diagnosis mission of the laboratory between July 2011 and June 2013. A retrospective analysis of our diagnostic activity from 2010 through 2013 showed that 48%, 68%, 82%, 97% of cultures from 47 positive clinical samples were positive after 3, 7, 10, 28 days of incubation, respectively.
Preliminary tests showed that stable spectra were obtained with 106 promastigotes but that an optimal discrimination of peaks was achieved with 107 promastigotes. Because the culture medium influences spectra (not shown), isolates growing better in NNN medium were transferred for one cycle (i.e., 48–96 hours) in Schneider medium before analysis. The growth kinetics of 2 L. tropica isolates was established over two consecutive cycles and spectra were obtained at three stages: exponential (24–72 hours), stationary (72–148 hours) and decay, showing that spectra were reproducible at the exponential and early stationary stages (data not shown). All spectra were then obtained from late exponential/early stationary stages. The reproducibility was further established by analyzing 16 replicates of the same samples for a L. (L) infantum and a L. (V) braziliensis isolate (Suppl Fig. S2 & 3), by analyzing samples from the same isolates of L. (L) infantum maintained in culture for several days (Suppl Fig. S4), by analyzing samples from the same isolates of L. (V) braziliensis analyzed at day 0 and a second culture of the same isolate frozen and thawed for subculture 6 months later (Suppl Fig. S5) and by analyzing 3 L. (L) infantum and 5 L. (V) braziliensis isolates (Suppl Fig. S6 & 7). Identification was accurate in all cases. The same approach was also performed with a L. major isolate maintained in culture in duplicate for 5 weeks. Spectra lead to the same species identification at all points (not shown).
Peaks 11121 (+/−7) and 7114 (+/−4) were both present in all 18 isolates belonging to the Viannia subgenus - 13 L. braziliensis, 5 L. guyanensis/L. panamensis –and absent in all 28 isolates of the Leishmania subgenus - 3 L.mexicana/L. amazonensis, 16 L. major, 5 L. donovani/L. infantum, 4 L. tropica/L. killicki (Table 1, Fig. 1). Conversely, peaks 6153 (+/−3) and 7187 (+/−5) were present in all isolates of the Leishmania subgenus and absent in all isolates belonging to the Viannia subgenus. Of note, none of these 4 peaks were present in the isolate identified as L. martiniquensis. The discriminating power of other peaks was then interpreted in the context of the 2 subgenera.
Within the Leishmania subgenus, peaks 5589 (+/−3) and 11180 (+/−6) were present only in L. mexicana/L. amazonensis isolates and absent in other isolates, identified by reference methods as L. killicki, L. tropica, L. major, L. infantum, L. donovani (Table 1 & Fig. 1B). Similarly, within this Leishmania subgenus, peaks 5630 (+/−2) & 5937 (+/−2), or 5753 (+/−3) were present in isolates considered by reference methods as L. major, L. tropica respectively. Peak 7875 (+/−5) was present in the 5 isolates allocated by MLST to the L. donovani complex (L. donovani & L. infantum). Peak 5726 (+/−6) was present in the 3 isolates identified as L. donovani by reference methods and absent in the 2 L. infantum isolates. All those species-defining peaks (Leishmania subgenus) were absent in the single L. killicki isolate (Table 1). Within the Viannia subgenus, the pair of peaks 5987 (+/−3) & 6173 (+/−3) was present in all L. braziliensis/L. peruviana isolates and absent in all L. guyanensis/L. panamensis isolates. All L. guyanensis/L. panamensis isolates expressed either the 6015 (+/−5) or the 6234 (+/2) peak that were both absent in all L. braziliensis/L. peruviana isolates. Slight variations for the value of the peaks were observed (Table 1) but – in this relatively limited set of isolates - did not jeopardize the manual, computer independent identification process.
Figure 2: shows mass spectrometry spectra from four different isolates of Leishmania included in the reference library (2 from the Viannia subgenus 2 from the Leishmania subgenus). The four peaks discriminating subgenera and several peaks discriminating species complexes are shown on these spectra, and are labeled with their respective molecular weights thus allowing an easy analysis based on the algorithm (Fig. 1B).
A cluster analysis based on a correlation matrix was performed for Old and New world Leishmania isolates, in order to assess the ability of the MALDI-TOF MS to generate a classification consistent with that obtained by reference methods. As depicted in (Fig. 1A), the resulting dendrogram for all Leishmania isolates showed separate clusters corresponding to the species typed by reference methods, falling appropriately into the 2 subgenera (Leishmania and Viannia). Isolates considered by MLST as L. major were located on one branch, clearly distinguished from isolates considered as L. donovani/L.infantum and L. tropica. The single L. killicki isolate analyzed to date was located in the Leishmania subgenus, close to isolates considered as L. donovani/infantum. The dendrogram built from PCA differentiated clearly L. guyanensis/L. panamensis from L. braziliensis/L. peruvianas species complexes but segregations between L. panamensis and L. guyanensis, L. braziliensis and L. peruviana were not possible at this stage. Isolates belonging to the L. mexicana complex fell into the Leishmania subgenus on a distinct branch. It appeared close to the single trypanosomatid isolate from the French West Indies (recently named L. martiniquensis, Desbois et al., personal data, Fig. 1).
Applied on a set cultured isolates spanning most Leishmania species of medical importance, MALDI-TOF MS generated a classification consistent with results obtained by reference methods (MLST or heat-shock protein 70 gene sequencing). This was achieved using a simple, database independent analysis of MALDI-TOF spectra based on the algorithm. The simplicity of the analytical procedure allowed a minute output of results as soon as fair parasite growth was obtained in monophasic medium. Taken together, these observations suggest that MALDI-TOF may be a useful tool to facilitate treatment decision in cutaneous leishmaniasis.
Treatment of CL should indeed be based on species identification [1], [3], [4], [9], [10], [30]–[32]. For example, systemic antimony is generally more effective in patients infected with L. braziliensis than in patients infected with L. guyanensis [33], [34] or L. mexicana [9]. Conversely pentamidine is frequently used to treat L. guyanensis/panamensis CL [35] but is poorly effective in L. braziliensis CL [36]. Many patients get infected in places where both species circulate and may therefore receive initially first course of a suboptimal treatment. In the Old World, L. major can be treated effectively and easily with a 3rd generation aminoglycoside ointment [37], [38] but the efficacy of this topical formulation in patients infected with L. tropica or L. infantum is as yet unknown [4]. In all these situations, rapid species identification should help adopt the most appropriate option in a majority of patients. Because more than 80% of cultures in our context are positive in the first 10 days after sampling, and because it takes a few hours to obtain the MALDI-TOF spectrum, in many instances time-to-identification is now short enough to with-hold treatment decision until species identification is available. Analysis of a higher number of isolates will be necessary to deliver a more solid dendrogram, particularly to determine whether MALDI-TOF MS can achieve a robust discrimination between L. braziliensis and L. peruviana, L. panamensis and L. guyanensis, L. donovani and L. infantum, L. mexicana and L. amazonensis. Fortunately, in current algorithms therapeutic decision in cutaneous leishmaniasis is not heavily impacted by these differentials [3], [6].
The approach presented here has no taxonomic ambition but was evaluated for potential use in medical practice. We limit our conclusions to the ability of MALDI-TOF MS to generate clusters congruent with those raised by reference methods. Because Leishmania species determination is complex, extension of explorations will be performed in the context of multinational networks such as the LeishMan consortium (http://www.leishman.eu) that merges information from several European countries and benefits from the presence of experts with strong expertise in Leishmania species identification [14]–[16]. Interpretation of complex spectra in very rare cases of co-infection with two Leishmania species will be attempted in this context. Optimization of pre-analytical steps, including culture conditions, parasite concentration in pellets and protein extraction was important for a robust interpretation of spectra. We selected a 72–96 h incubation period, corresponding to the exponential phase or early stationary phase of growth to limit variations in protein content. Schneider, an axenic medium (20% fetal calf serum, penicillin and streptomycin), was chosen as the reference because its supports rapid growth of most isolates and is associated with reproducible spectra. The need for a culture step is a weakness of mass spectrometry shared with several other typing methods. In rare instances, parasite isolation was difficult or slow. We partially circumvented this bottleneck by culturing clinical samples simultaneously on Schneider and NNN medium. Once adapted, NNN-dependent isolates were transferred for one cycle in Schneider medium then processed for MALDI-TOF analysis. In the long-term, the approach may be further simplified by using dipsticks developed from discriminating peaks. Sensitive protein detection using immunochromatography directly from a lesion scraping or aspirate – as currently developed for diagnosis- may indeed by-pass the culture step and further accelerate species identification.
Another relative limitation of mass spectrometry is that the method is currently handled by reference centers only. However, because mass spectrometers have a wide spectrum of medical applications in microbiology, prices are dropping and cheaper versions are emerging.
In the short term, we found that MALDI-TOF (MS) was a promising approach to generate spectra from Leishmania promastigotes with high identification at the species level consistent with the reference method. A limitation of the technique is the need for cultivation parasites. Nevertheless, as compared with molecular biology [39], this approach offers great advantages, in particular speed, simplicity, cost for isolate identification and was easy to integrate into the organization of a clinical laboratory. Not least, the intuitive interpretation of spectra was well-suited for allowing for close interactions between parasitologists and clinicians. These strengths should predictably facilitate rapid treatment decision in cutaneous leishmaniasis.
|
10.1371/journal.pgen.1004292 | Molecular Mechanisms of Hypoxic Responses via Unique Roles of Ras1, Cdc24 and Ptp3 in a Human Fungal Pathogen Cryptococcus neoformans | Cryptococcus neoformans encounters a low oxygen environment when it enters the human host. Here, we show that the conserved Ras1 (a small GTPase) and Cdc24 (the guanine nucleotide exchange factor for Cdc42) play an essential role in cryptococcal growth in hypoxia. Suppressor studies indicate that PTP3 functions epistatically downstream of both RAS1 and CDC24 in regulating hypoxic growth. Ptp3 shares sequence similarity to the family of phosphotyrosine-specific protein phosphatases and the ptp3Δ strain failed to grow in 1% O2. We demonstrate that RAS1, CDC24 and PTP3 function in parallel to regulate thermal tolerance but RAS1 and CDC24 function linearly in regulating hypoxic growth while CDC24 and PTP3 reside in compensatory pathways. The ras1Δ and cdc24Δ strains ceased to grow at 1% O2 and became enlarged but viable single cells. Actin polarization in these cells, however, was normal for up to eight hours after transferring to hypoxic conditions. Double deletions of the genes encoding Rho GTPase Cdc42 and Cdc420, but not of the genes encoding Rac1 and Rac2, caused a slight growth retardation in hypoxia. Furthermore, growth in hypoxia was not affected by the deletion of several central genes functioning in the pathways of cAMP, Hog1, or the two-component like phosphorylation system that are critical in the cryptococcal response to osmotic and genotoxic stresses. Interestingly, although deletion of HOG1 rescued the hypoxic growth defect of ras1Δ, cdc24Δ, and ptp3Δ, Hog1 was not hyperphosphorylated in these three mutants in hypoxic conditions. RNA sequencing analysis indicated that RAS1, CDC24 and PTP3 acted upon the expression of genes involved in ergosterol biosynthesis, chromosome organization, RNA processing and protein translation. Moreover, growth of the wild-type strain under low oxygen conditions was affected by sub-inhibitory concentrations of the compounds that inhibit these biological processes, demonstrating the importance of these biological processes in the cryptococcal hypoxia response.
| When Cryptococcus neoformans, an environmental fungal pathogen, enters the human host, it encounters a low oxygen condition. The well conserved Ras1 and Cdc24 proteins are known for their key roles in maintenance of the actin cytoskeletal integrity in eukaryotic cells. In this work, we show a unique role of RAS1 and CDC24 in the growth of C. neoformans in a low oxygen environment. Actin polarization, however, appeared normal in the ras1Δ and cdc24Δ strains under hypoxic conditions for up to eight hours. We show that PTP3 is required for hypoxic growth and it can rescue the hypoxic growth defect in ras1Δ and cdc24Δ. Genetic analysis suggested that RAS1 and CDC24 function linearly while CDC24 and PTP3 function parallelly in regulating hypoxic growth. RNA sequencing combined with analysis by small molecular inhibitors revealed that RAS1, CDC24 and PTP3 regulate several biological processes such as ergosterol biosynthesis, chromosome organization, RNA processing and protein translation which are required in the cryptococcal response to hypoxic conditions.
| Oxygen availability is critical for many biochemical reactions in eukaryotic cells and their ability to adapt to oxygen limitation is essential for survival. Most organisms are able to sense the change in environmental oxygen concentration and respond swiftly. For instance, the human fungal pathogen Cryptococcus neoformans is an obligate aerobic fungus that grows in an ecological niche with ambient air. Once inhaled by the susceptible host, C. neoformans has to adapt to suboptimal levels of oxygen in the lung and disseminates to the brain where oxygen levels are even lower than the lung. How aerobic environmental organisms adapt to suboptimal concentrations of oxygen in the host is an important issue for the understanding of cryptococcal pathogenesis.
Various mechanisms of oxygen sensing and responses to hypoxia in yeast and pathogenic fungi have been reviewed recently [1]–[4]. C. neoformans utilizes Sre1, the mammalian sterol regulatory element-binding protein (SREBP) homolog, to control sterol homeostasis, oxygen sensing, and virulence in mice [5], [6]. Sre1 is also the major regulator of the hypoxic response in Schizosaccharomyces pombe and it regulates expression of more than 100 genes [7]. However, there are no obvious SREBP orthologs in Saccharomyces cerevisiae or in any of the species belonging to the Candida clade. These yeast cells respond to limited-oxygen conditions by inducing expression of a large number of hypoxic genes, which encode oxygen-related functions in respiration and biosynthesis of heme, lipids, cell-wall and membranes [2], [4]. Studies using the hypoxia mimetic compound CoCl2 revealed that the ability of C. neoformans to grow in low oxygen conditions was linked to mitochondrial function, the ability of cells to respond to reactive oxygen species, and gene expression associated with ubiquitination as well as sterol and iron homeostasis [8].
Recent genetic studies have demonstrated that normal actin function and actin-binding proteins are important for the growth of C. neoformans in hypoxic conditions [9]. Actin rearrangements are regulated by small GTPases of the Ras and Rho subfamilies [10]. In mammalian cells, Ras transduces signals to multiple pathways that regulate the expression of nuclear genes as well as those required for rearrangement of the actin cytoskeleton [11], [12]. Ras proteins are upstream determinants of actin cytoskeletal integrity and cell stress response as exhibited in many organisms [13]–[15]. Rho-GTPases function as molecular switches by cycling between inactive GPD-bound and active GTP-bound forms. In mammalian cells, the Rho family of G proteins such as Cdc42, Rac, and Rho play complementary roles in the actin cytoskeleton organization [16]. Actin re-polarization at 37°C is delayed in ras1Δ mutant cells of C. neoformans [17]. Additionally, actin cytoskeletal architecture and cellular morphogenesis are controlled by elements of the Ras pathway in a temperature dependent manner [18], [19]. C. neoformans Ras1 also cooperates with conserved Rho-GTPases such as Cdc42 and Rac1 in controlling cell morphology under stressful conditions and regulate filamentous growth during mating [19]–[21]. C. neoformans contains two functional Cdc42 paralogs, Cdc42 and Cdc420, which are required for growth at high temperatures but are not required for viability under non-stress conditions [22]. Rac1 and Rac2 function downstream of Ras1 in C. neoformans and together with Ste20, which belongs to Cdc42p-activated signal transducing kinase and is a member of the PAK (p21-activated kinase) family, control high-temperature growth and cellular differentiation [20], [23].
Guanine nucleotide exchange factors (GEF), GTPase-activating proteins (GAP) and guanine nucleotide dissociation inhibitors (GDI) are proteins that enable and control the transition of Rho-GTPases between the GPD-bound and active GTP-bound form [24]. In C. neoformans, Cdc24, a Cdc42-specific GEF, is a Ras1 effector mediating the ability of this fungus to grow at high temperatures [19]. As in the case of RAS1, CDC24 is required to cause disease. Epistasis and yeast two-hybrid analysis indicated that the Cdc24 homolog of S. pombe, Scd1, forms a protein complex with Cdc42 and Ras1 supporting the functional relationship between Ras1 and Cdc24 proteins [25]. In this study, we show that growth of C. neoformans in low oxygen conditions requires RAS1 and CDC24. Interestingly, there was no clear evidence that actin polarization was compromised in ras1Δ and cdc24Δ strains under hypoxic conditions. Suppressor screening enabled the identification of Ptp3 as a downstream effector of Ras1 and Cdc24. PTP3 encodes a putative homolog of the phosphotyrosine-specific protein phosphatases and shares similarity with the S. cerevisiae Ptp3 that regulates the Hog1 mitogen-activated protein kinase [26]. By RNA sequencing analysis and using small molecular inhibitors, we have demonstrated that RAS1, CDC24 and PTP3 regulate cryptococcal response to hypoxia that involves several biological processes such as ergosterol biosynthesis, chromosome organization, RNA processing and protein translation.
It is known that Ras1-Cdc24 signal transduction pathway mediates actin polarization [19] and proper actin function is important for hypoxic response in C. neoformans [9]. We, therefore, examined the involvement of genes that function in the Ras1-Cdc24 signal transduction pathway in the cryptococcal response to hypoxia. Figure 1A shows that ras1Δ and cdc24Δ strains fail to grow at 1% O2 but grow comparably to the wild-type H99 in ambient oxygen which suggests an essential role of the Ras1-Cdc24 signaling pathway in hypoxia response. It has been proposed that Ras1 signals utilize multiple Rho-GTPases and Ste20 to coordinately regulate polar growth in C. neoformans [22]. We found that deletion of the individual genes encoding the two types of Rho-GTPase genes, including CDC42, CDC420, RAC1 and RAC2, did not affect the ability of C. neoformans to grow in low oxygen. Although growth of the cdc42Δcdc420Δ double deletant was slightly reduced in 1% O2 compared to the wild-type, growth of the rac1Δrac2Δ double deletant was comparable to the wild-type (Figure 1A). Furthermore, deletion of STE20 only caused a subtle reduction in growth at the same low oxygen condition. C. neoformans and S. pombe both employ the Sre1 pathway as part of their hypoxia response network suggesting the biological similarity between these two fungi [1]. However, the S. pombe mutants, scd1 (encoding a Cdc24 homolog), ras1, and efc25 (encoding a Ras1-GEF) did not display a hypoxia sensitive phenotype at 0% O2 (Figure 1B).
In C. neoformans, the Ras1 signaling pathway is known to regulate invasive growth and mating via the cAMP signaling pathway [18], [27]. The cAMP signaling pathway is also involved in response to stress from heavy metals and toxic metabolites [28]. To test the importance of the cAMP stress response pathway in hypoxic conditions, we examined the growth phenotype of mutants defective in the cAMP signaling pathway genes including gpa1, cac1, aca1, and pka1. We found that none of the mutants was susceptible to 1% O2 (Figure 1C) which suggested that the RAS1-CDC24-dependent hypoxia response is not directly related to the cAMP signaling pathway in C. neoformans.
As described previously [17], [19], log phase cells of ras1Δ and cdc24Δ strains grown in ambient oxygen were slightly larger in size than the wild-type strain (Figure 2). When these cells were shifted to 1% O2 and incubated for 8 h, the mutant cells were clearly larger than the wild-type cells (Figure 2). After 15 h at 1% O2, both ras1Δ and cdc24Δ cells ceased to multiply and a large portion of the cells were arrested as large unbudded cells suggesting a compromised budding process. This phenotype is similar to the terminal phenotype of ras1Δ and cdc24Δ incubated at restricted temperature [17], [19]. Most of the ras1Δ and cdc24Δ cells, however, were still viable at 1% O2 because when these mutants were first incubated at 1% O2 for 3 days and then transferred to 20% O2 for an additional 3 days, most of the arrested cells at 1% O2 resumed growth (Figure S1).
The ras1Δ cells manifest depolarized F-actin when grown at restricted temperature and visualized by rhodamine-conjugated phalloidin stain [17]. We examined the distribution of actin by expressing the Lifeact-RFP construct in a ras1Δ and a cdc24Δ strain. Lifeact contains a 17-amino-acid peptide of Abp140 from S. cerevisiae that has been used to visualize actin [29]. Interestingly, no clear difference of Lifeact-RFP signal was observed between the wild-type and either the ras1Δ or the cdc24Δ strain both at 20% O2 or 1% O2 for 8 h (Figure 2 middle panels). After 15 h in 1% O2, the Lifeact-RFP signal was not clearly visible in the ras1Δ and cdc24Δ cells due to the lack of budding cells. To quantitatively determine the ability of each strain in polarizing actin to the bud, the percentage of cells containing Lifeact-RFP staining in small to medium sized buds were determined. We observed that actin polarization was similar in wild-type, ras1Δ and cdc24Δ strains in 20% O2 (99%, 100% and 100%, respectively) as well as in 1% O2 for 2 h (87%, 87% and 90%, respectively). Similar results were observed in phalloidin stained cells under the same conditions (data not shown). These results indicate that ras1 and cdc24 mutations do not visibly affect actin polarization under hypoxic conditions for up to 8 hours.
To characterize the Ras1-Cdc24 dependent hypoxic growth pathway, we performed suppressor screens to identify genes whose multicopy presence suppressed the no-growth phenotype of cdc24Δ at 1% O2. We initially obtained 17 transformants of cdc24Δ that were able to grow at 1% O2. The inserts in the episomes of these transformants were subsequently PCR-amplified and sequenced. Among the successfully amplified PCR clones, three contained overlapping sequences of YPD1 (see Table 1 for gene designation of each suppressor and description of the function of S. cerevisiae homologs). Three clones contained the entire sequence of CDC24, PTP3, and ERJ5 genes, respectively. Two clones each contained HRD1 and GEF1 with portions of the 5′ end sequence missing. We cloned the entire genomic regions of HRD1 and GEF1 and used to confirm that both HRD1 and GEF1 genes suppressed the hypoxia phenotype of cdc24Δ. Since ERJ5 and GEF1 only weakly restored the phenotype of cdc24Δ at 1% O2 (Figure S2A), we conducted no further studies on these two genes. Figure 3A shows that extra copies of CDC24, YPD1, PTP3 and HRD1 are able to complement cdc24Δ growth defect at 1% O2.
It has been demonstrated that extra copies of CDC420 (GenBank ID: DQ991433) and STE20 suppress the temperature sensitive phenotype of cdc24Δ [19]. We introduced extra copies of each of these genes in cdc24Δ and examined its growth at 1% O2. Figure 3A shows that CDC42 fully restores the growth of cdc24Δ at 1% O2 and STE20 weakly complements the cdc24Δ hypoxic phenotype while CDC420 fails to suppress the phenotype. It has been shown that CDC24, CDC42, CDC420, and STE20 can suppress the temperature sensitive phenotype of ras1Δ [19]. We found that all four of these genes weakly suppress the hypoxic phenotype of ras1Δ (Figure 3B). Although RAC1 could restore thermotolerance in ras1Δ mutants [19], RAC1 and RAC2 did not suppress the hypoxic phenotypes of cdc24Δ or ras1Δ (data not shown). Interestingly, three newly identified cdc24Δ suppressors, HRD1, PTP3, and YPD1 could complement the growth defect of ras1Δ at 1% O2 (Figure 3B upper panels).
To determine if the newly identified cdc24Δ suppressors play a role in cryptococcal growth at 1% O2, we deleted PTP3 and HRD1 in the wild-type strain H99. Deletion of PTP3 caused hypersensitivity to 1% O2 but not for hrd1Δ (Figure 3C). YPD1 is required for viability of C. neoformans and can be only deleted in a background that carries the hog1 deletion [30]. We examined the phenotype of ypd1Δhog1Δ at 1% O2 and found that YPD1 may be not essential for growth in hypoxic conditions (Figure 3C). Furthermore, CnYpd1 is a homolog of a His-containing phosphorelay intermediate protein and the cryptococcal phosphorelay system consists of seven sensor histidine kinases (Tco1 to -7) and two response regulators (Ssk1 and Skn7) [31]. Among the seven sensor histidine kinases, TCO1 and TCO2 have been shown to play a role in hypoxia response in a specific strain derived from H99 [6]. However, the deletion of SSK1, SKN7, TCO1, TCO2, or TCO1/TCO2 double deletion in KN99a, the mating type a derivative of H99, did not affect growth in hypoxic conditions (Figure 3D). This suggests that the cryptococcal phosphorelay system is not essential for response to hypoxia. Additionally, deletion of seven known G protein-coupled receptors (GPCRs) did not affect growth in hypoxic conditions (Figure S2B) suggesting that the pathways linked to GPCRs are not by themselves required for cryptococcal growth in hypoxia.
PTP3 has not been characterized in C. neoformans. In S. cerevisiae, Ptp3 is involved in the inactivation of mitogen-activated protein kinase (MAPK) during osmolarity sensing and dephosporylation of Hog1 MAPK as well as regulation of its localization [26], [32]–[34]. Moreover, the S. cerevisiae Hog1 mediates a hypoxic response in S. cerevisiae [35]. To determine the importance of HOG signaling pathway in C. neoformans response to hypoxia, we examined the growth of hog1Δ and pbs2Δ, a MAPK kinase kinase of the HOG signaling pathway [36] at 1% O2. Figure 3D shows that although PTP3 is required for growth in hypoxic conditions, the HOG pathway genes, HOG1 and PBS2 are not. Since the cryptococcal Ptp3 may also be a negative regulator of Hog1, as in S. cerevisiae, deletion of PTP3 could cause hyperactivation of Hog1 leading to the inability of ptp3Δ to grow in hypoxic conditions. To test this possibility, we deleted HOG1 in ptp3Δ. The hog1Δ strain grew better than the wild-type in 1% O2 and ptp3Δhog1Δ double deletant grew comparable to the wild-type in 1% O2 (Figure 3E). We also deleted HOG1 in ras1Δ and cdc24Δ strains. Interestingly, both ras1Δhog1Δ and cdc24Δhog1Δ double deletants grew comparable to the wild-type in 1% O2. These data indicate that deletion of HOG1 compensated for the hypoxic growth defect of ras1Δ, cdc24Δ, and ptp3Δ suggesting existence of genetic interactions between HOG1 and these three genes. It is possible that cryptococcal Ptp3 regulates the activity of Hog1 by modulating the phosphorylation status of Hog1. Figure 4 shows that Hog1 was phosphorylated at normoxic conditions in H99 as previously described [36]. Interestingly, Hog1 phosphorylation in ras1Δ, cdc24Δ, and ptp3Δ was similar to the wild-type at normoxic conditions. Hog1 phosphorylation significantly increased when cells were transferred to 1% O2 for one hour in the wild-type strain but not in the ptp3Δ and cdc24Δ strains (Figure 4). Although there was a statistically significant increase of Hog1 phosphorylation in cdc24Δ, the amount of increase was not as high as in the wild-type. These results demonstrate that Hog1 was not hyperphosphorylated in response to hypoxic treatment in ras1Δ, cdc24Δ, and ptp3Δ.
Suppressor studies in hypoxic conditions suggested that RAS1 is epistatically upstream of CDC24, which is upstream of PTP3. To determine if these three genes function linearly or in parallel in regulating hypoxic growth, we generated ras1Δcdc24Δ and cdc24Δptp3Δ double deletants. The growth of ras1Δ and cdc24Δ was slightly slower than the wild-type strain at 37°C and ras1Δcdc24Δ double deletion strains grew slower than the parental single deletants (Figure 3F). These data indicate that the effect of RAS1 and CDC24 double deletions on growth was synthetic at elevated temperatures, suggesting that RAS1 and CDC24 may function in compensatory/distinct pathways in regulating thermal tolerance. The growth of ptp3Δ was similar to the wild-type at 37°C but the cdc24Δptp3Δ double deletant failed to grow at 37°C (Figure 3G). In addition, PTP3 failed to complement the growth defect of ras1Δ or cdc24Δ at elevated temperatures (Figure S2). These data suggested that the effect of CDC24 and PTP3 double deletions was synthetic at elevated temperatures and CDC24 and PTP3 may function in compensatory pathways in regulating thermal tolerance.
Since ras1Δ, cdc24Δ and ptp3Δ could not grow in 1% O2, we examined the hypoxic phenotype of double deletions at slightly higher oxygen conditions. The growth of ras1Δ and cdc24Δ was slower than the wild-type in 5% O2 (Figure 3F) and the growth retardation of ras1Δcdc24Δ double deletants in 5% O2 was no more severe than the parental single deletants. This result suggests that RAS1 and CDC24 may function in a linear pathway to handle low oxygen stress. Since cdc24Δptp3Δ failed to grow in the presence of 5% CO2 (Figure 3G), we examined the hypoxic phenotype of cdc24Δptp3Δ in 6% O2 without CO2 at 28°C. The growth of ptp3Δ and cdc24Δ was slightly slower than the wild-type in 6% O2 without CO2 at 28°C. In contrast, cdc24Δptp3Δ failed to grow at 6% O2 without CO2 (Figure 3G) demonstrating that the effect of these double deletions is additive in 6% O2 and suggesting that CDC24 and PTP3 may function in parallel to manage hypoxic stress.
After finding that several common stress response pathways proteins were not individually required for hypoxic response, we used RNA sequencing to uncover the pathways which are involved in cryptococcal response to hypoxic stress. We performed high-throughput sequencing of cDNA made from poly(A) RNAs obtained from 1% O2 grown cells of the wild-type, ras1Δ, cdc24Δ, ptp3Δ and cdc42Δcdc420Δ double deletant strains. Based on the results of suppressor and double deletants analysis, we speculated that there may be an overlap in the group of genes that are differentially expressed between the wild-type and each mutant of ras1, cdc24 and ptp3. We first focused our analysis of RNA sequencing data on the group of genes commonly affected by ras1Δ, cdc24Δ and ptp3Δ but not by cdc42Δcdc420Δ, since the cdc42Δcdc420Δ double deletant only displayed weak hypoxic phenotype. Among the statistical significance of the differentially expressed genes that had FPKM>10 (fragments per kilobase of transcript per million mapped reads), 441 genes belonging to this group were identified (Figure 5A).
To obtain a global picture of the changes in gene expression, we used DAVID Functional Annotation Tools [37] to search for the over-represented GO terms within this group of genes. We used the tool of functional annotation clustering in DAVID to cluster functionally related annotations into groups for related gene–term relationships [38]. Among the group of 441 genes that showed significant differential expression, we found 4 annotation clusters with an enrichment score greater than 1.4. These included ergosterol biosynthesis, chromosome organization, amino acid biosynthetic process and mRNA processing (Table 2). We used RT-PCR and chose 5 differentially expressed genes from the cluster with the highest enrichment score, ergosterol biosynthetic process, to confirm the expression levels. Figure 5B shows that the expression levels of ERG3, ERG4, ERG6, ERG7, and ERG11 in the ergosterol biosynthesis pathway are consistently lower in ras1Δ, cdc24Δ and ptp3Δ compared to that of the wild-type strain while the expression levels of these gens in cdc42Δcdc420Δ double deletant strain are similar or slightly higher than in the wild-type. Therefore, quantitative RT-PCR results support the finding from RNA sequencing analysis.
It is likely that the observed changes in transcripts levels for the genes in ergosterol biosynthesis may also reflect the changes in production of ergosterol and its biosynthetic intermediates. We analyzed the sterol profiles of the wild-type and mutants strains to explore this possibility. Cells were transferred from 20% O2 to 1% O2 and total sterols were extracted and analyzed at several time points after the transfer. We did not find any new ergosterol intermediate in any mutant (Figure S3 and data not shown). Furthermore, no significant difference of ergosterol was detected between the wild-type and all mutants except that ptp3Δ produced lower amounts of ergosterol in both normoxic and hypoxic conditions (Figure 6A). Interestingly, almost all the discernible ergosterol intermediates including squalene, eburicol, 4α-methyl fecosterol, ergost-7-enol, and ergost-7, 22-enol accumulated at significantly higher levels in cdc24Δ and ras1Δ compared to the wild-type in normoxic and hypoxic conditions (Figure 6A and Figure S4). In contrast, the ergosterol intermediates were not significantly different between the wild-type and ptp3Δ or cdc42Δcdc420Δ except that ptp3Δ accumulated significantly lower amounts of squalene and 31-noreburicol after 5 h in 1% O2 (Figure 6A and Figure S4). These results indicate that Ras1 and Cdc24 function differently compared to Ptp3 in regulating ergosterol biosynthesis and the production of sterols is similar between cdc42Δcdc420Δ and wild-type. Since extra copies of PTP3 enabled growth of cdc24Δ and ras1Δ at low oxygen conditions, it is possible that accumulation of the ergosterol intermediates in cdc24Δ and ras1Δ would be reduced in the PTP3 over-expressed strains. Interestingly, amounts of the major ergosterol intermediates, squalene, ergost-7-enol, and ergost-7, 22-enol, did not decrease but increased significantly in cdc24Δ+PTP3 and ras1Δ+PTP3 strains at most of the time points (Figure 6B). These data suggest that suppression of the cdc24Δ and ras1Δ growth deficiency by PTP3 at 1% oxygen was not due to a reduction in elevated levels of ergosterol intermediates. It also suggests that the accumulation of ergosterol intermediates in cdc24Δ and ras1Δ may not be the major cause for the inability of these strains to grow in hypoxic conditions.
Several biological processes were identified by GO-term analysis using data derived from RNA sequencing. It is possible that the identified processes play a critical role in regulating hypoxic growth in the mutants or the wild-type. Inhibitors known to affect these biological processes were chosen to evaluate their impact on growth under normoxic or hypoxic conditions. The drug concentrations sub-inhibitory to the wild-type strain in normoxic condition were used in all the analyses.
Ergosterol is an essential component of fungi and its synthesis can be blocked by fluconazole and fenpropimorph. Fluconazole targets the lanosterol 14α-demethylase and fenpropimorph inhibits C14-sterol reductase and C8-sterol isomerase [39]. We selected these two inhibitors to investigate the possible association of ergosterol biosynthesis with the observed phenotype. Figure 7A shows that ptp3Δ and cdc42Δcdc420Δ were hypersensitive to 8 µg/ml of fluconazole in normoxic conditions compared to the wild-type. In addition, treatment of fenpropimorph resulted in the hypersensitivity of ras1Δ, cdc24Δ, ptp3Δ and cdc42Δcdc420Δ in normoxic conditions. The hypersensitivity to ergosterol biosynthesis inhibitors suggested that all the mutants may have a disturbed ergosterol biosynthetic pathway and Ptp3 and Cdc42/Cdc420 may play different roles in this pathway compared to Cdc24 and Ras1. Furthermore, the wild-type strain was hypersensitive to fluconazole in 1% O2 but not to fenpropimorph, indicating that inhibition of the lanosterol 14α-demethylase activity in ergosterol synthesis affects the cryptococcal growth in hypoxic conditions.
It is well known that histone acetylation regulates many chromosomal functions. We chose two histone deacetylase inhibitors, trichostatin A (TSA) and sodium butyrate [40] [41] to investigate the possible association of mutant phenotypes with chromosome functions indicated by GO-term analysis. Figure 7B shows that ras1Δ, cdc24Δ, ptp3Δ and cdc42Δcdc420Δ double deletant all exhibit hypersensitivity at 64 µg/ml TSA and 128 mM sodium butyrate in normoxic conditions compared to the wild-type, although the level of sensitivity varied among the strains. Furthermore, the wild-type strain was hypersensitive to TSA and sodium butyrate in 1% O2 compared to the control YPD medium. To access the possible involvement of mRNA processing, amiloride and staurosporine were selected as inhibitors. Amiloride modulates oncogenic alternative RNA splicing [42] and selectively binds to an abasic site in RNA Duplexes [43]. Although staurosporine is a prototypical ATP-competitive kinase inhibitor that interacts with the ATP binding pocket with little selectivity [44], staurosporine is also known to inhibit yeast RNA splicing in vitro by blocking ATP binding to any of the DEXD/H-box RNA helicases involved in splicing [45]. We found that ras1Δ, cdc24Δ, ptp3Δ and cdc42Δcdc420Δ double deletant were hypersensitive to staurosporine as well as amiloride in normoxic conditions (Figure 7C). In addition, the wild-type strain grew poorly in the presence of these drugs in 1% O2. These results support the notion that chromosome organization and RNA processing are involved in the cryptococcal hypoxic response.
Since cdc42Δcdc420Δ displayed slight sensitivity to hypoxic conditions, we also applied DAVID Functional Annotation Tools to analyze the group of 489 genes coordinately expressed among ras1Δ, cdc24Δ, ptp3Δ and cdc42Δcdc420Δ (Figure 5A.). Four high enrichment score annotation clusters were identified including ribosome assembly, tRNA amino acylation, regulation of translation and chromosome organization suggesting that protein biosynthesis and/or translation are affected in these mutants (Table 3). We used cycloheximide and streptomycin to investigate the possible association of these genes in protein biosynthesis. Cycloheximide is a glutarimide antibiotic that binds to the 60 Sribosomal subunit and inhibits translation elongation [46]. Streptomycin is an aminoglycoside antibiotic that binds to the small ribosomal subunit of eukaryotic cells and inhibits ribosomal translocation as well as compromises translation fidelity [47]. In comparison to the wild type, growth of ras1Δ and cdc24Δ was slightly affected in the presence of cycloheximide while the growth of ptp3Δ and cdc42Δcdc420Δ double deletant was noticeably reduced (Figure 7D). Similarly, the growth of ras1Δ, cdc24Δ, and ptp3Δ was slightly reduced and the growth of cdc42Δcdc420Δ double deletant was considerably affected in media containing streptomycin. Furthermore, the wild-type H99 strain grew slower in the presence of streptomycin or cycloheximide in 1% O2 compared to the YPD control but not in normoxic conditions suggesting that hypoxic growth requires the fully functional protein translation machinery in C. neoformans.
The highly conserved Ras1 and Cdc24 proteins are well known for their importance in the maintenance of actin cytoskeletal integrity. However, the connection between Ras1-Cdc24 signaling pathway and the growth of organisms in a hypoxic environment has not been reported. This study is the first to report on the link of RAS1 and CDC24 with hypoxic growth in C. neoformans. The fission yeast, S. pombe and C. neoformans behave similarly in suboptimal levels of oxygen. In S. pombe, reduction in sterol synthesis caused by the lack of sufficient oxygen is sensed by Sre1, which is a transcriptional regulator of the hypoxia related genes [7]. However, neither RAS1 nor SCD1 (encoding Cdc24 homolog) is required for growth of S. pombe in low oxygen conditions (Figure 1). C. albicans does not contain a Sre1 homolog but both Ras1 and the adenylate cyclase Cdc35 are known to be associated with hyphal growth in hypoxic conditions [48]. Unlike the ras1Δ in C. neoformans, however, the C. albicans ras1Δ and cdc35Δ strains are able to grow in yeast form under hypoxic conditions. Thus, the involvement of Ras1-Cdc24 signaling pathway in hypoxic growth is not a universal fungal paradigm.
Which pathways do Ras1 and Cdc24 proteins utilize to regulate hypoxic growth in C. neoformans? Our data excludes a few possible candidates. First, although actin function is required for proper growth in hypoxic conditions [9], reorganization of the actin cytoskeleton did not appear to be different in ras1Δ and cdc24Δ strains compared to the wild-type strains for up to 8 hours. However, extra copies of CDC42 suppressed the growth defect of cdc24Δ and ras1Δ under hypoxia (Figure 3). It is possible that Cdc42 regulates processes other than actin polarization or that the failure of cdc24Δ and ras1Δ to proliferate in hypoxic conditions is in part due to otherwise unnoticeable malfunction in actin polarization. Second, genes from pathways involved in osmotic and genotoxic stress such as cAMP signaling, Hog1, and the two-components like phosphorelay, are not individually required for growth of C. neoformans in hypoxic conditions. These results parallel the observation that the Ras1-Cdc24 signaling pathway operates differently from pathways that are employed in response to osmotic and genotoxic stresses [28]. Furthermore, seven known G protein-coupled receptors (GPCRs) were not required for growth in hypoxic conditions suggesting that the cryptococcal cells do not utilize these receptors to sense hypoxic stress. Thus, Ras1 and Cdc24 proteins play an unusual role(s) in C. neoformans to overcome hypoxic stress.
We have noted that although growth of the cdc42Δcdc420Δ double deletant was only slightly susceptible to 1% O2, it was hypersensitive to all the tested inhibitors even in the normoxic conditions. Cdc42 and Cdc420 belong to the Rho-GTPase family and the Rho-GTPases in all eukaryotic cells are key regulators of the signaling pathways that control actin organization and morphogenetic processes (see recent reviews [49], [50]). Unlike the case in S. pombe and S. cerevisiae [51], [52], CDC42 was not essential in C. neoformans and the cdc42Δcdc420Δ double deletant strain was viable under normal growth conditions [22]. Hypersensitivity to all the tested inhibitors may be due to inhibition of growth in otherwise sick strains or, alternatively, it substantiates the importance of Cdc42 type Rho-GTPase for basic biological functions in C. neoformans.
It has been shown that a single Rho GTPase is regulated by more than one RhoGEF or RhoGAP [53]. At the same time, RhoGEFs and RhoGAPs regulate more than one Rho protein. In C. neoformans, the Rho GTPase genes CDC42, CDC420 and RAC1 can restore thermotolerance of the ras1Δ mutants while only CDC420 visibly suppresses the temperature sensitive phenotype of cdc24Δ [19]. We have observed that only CDC42 can restore the ability of cdc24Δ to grow at 1% O2 while CDC42 and CDC420 weakly suppress the hypoxic phenotype of ras1Δ. In addition, RAC1 and RAC2 did not suppress the hypoxic phenotypes of cdc24Δ or ras1Δ. Hence, the paralogs of these Rho-GTPases play different roles under different environmental conditions in C. neoformans. Additionally, we observed that single mutants of Rho-like GTPases did not impair growth in hypoxic conditions. It is possible that there is enough functional redundancy among the related Rho-like GTPases that single mutants (or even double mutants) may not display a strong phenotype in low oxygen. This possibility was to a certain extent supported by the reduced growth of cdc42Δcdc420Δ double deletant strain in 1% O2 but not by the rac1Δrac2Δ double deletant strain. The growth defect of cdc42Δcdc420Δ, however, was not as severe as ras1Δ or cdc24Δ at 1% O2 suggesting that other Rho-GTPases may be involved. Furthermore, triple deletants cdc42Δcdc420Δrac1Δ and cdc42Δcdc420Δrac2Δ grew comparably to the cdc42Δcdc420Δ double deletant strains in 1% O2 (data not shown). It is possible that the hypoxic growth phenotype of cdc42Δcdc420Δrac1Δrac2Δ quadruple deletant may provide more information but we have so far been unable to generate such a mutant. Alternatively, it is also possible that other types of Rho-GTPases may be involved in the regulation of growth in hypoxia. The detailed mechanisms of how Rho-GTPases and their regulators contribute to the hypoxia response remain to be elucidated.
Among the hypoxic mutants of C. neoformans thus far identified [5], [8], [9], [54], [55], ras1Δ and cdc24Δ displayed the clearest phenotype in 1% O2 and enabled us to perform suppressor screening. It is clear that 1% O2 or elevated temperature did not kill ras1Δ and cdc24Δ cells since they were able to resume growth once transferred back to ambient conditions (Figure S1). It is not clear, however, what the physiological state of ras1Δ and cdc24Δ cells is under hypoxic conditions. Because cdc24Δ and rasΔ1 behaved similarly in accumulation of ergosterol biosynthetic intermediates and sensitivity to many inhibitors, which was different from ptp3Δ, it is likely that the pathways controlled by Ptp3 are different from Ras1 and Cdc24. The genetic interactions among RAS1, CDC24 and PTP3 are complicated. PTP3 was able to suppress the hypoxic growth phenotype but not thermal intolerance of cdc24Δ and ras1Δ. CDC24 was able to suppress the hypoxic and thermal intolerance phenotype of ras1Δ but RAS1 could not suppress the hypoxic phenotype of ptp3Δ and cdc24Δ (Figure 3 and data not shown) [19]. Growth analysis of double mutants at elevated temperature suggests that RAS1, CDC24 and PTP3 may function in parallel to regulate thermal tolerance. In contrast, under hypoxic conditions RAS1 and CDC24 may function in a linear pathway to regulate hypoxic growth while CDC24 and PTP3 may function in parallel for such growth. Details of the mechanism as to how RAS1, CDC24 and PTP3 interact require further study.
Ptp3 shares similarity to the family of phosphotyrosine-specific protein phosphatases that are important in cell signaling [56], [57]. Oxidation of tyrosine phosphatases in hypoxia followed by re-oxygenation functionally links the tyrosine phosphatases to the hypoxia response in mammalian systems [58]. However, the involvement of Ptp3 in hypoxic response in lower eukaryotes has not been established. It is known that Hog1 in S. cerevisiae mediates the hypoxic response [35] and disruption of PTP3 in S. cerevisiae results in constitutive Hog1 tyrosine phosphorylation [26]. Therefore, it is possible that the deletion of cryptococcal PTP3 causes hyperactivation or constitutive activation of Hog1 that leads to the inability of ptp3Δ to grow in hypoxic conditions and deletion of HOG1 in ptp3Δ eliminates the Hog1 influence. However, we showed that deletion of cryptococcal PTP3 did not result in the elevation of Hog1 phosphorylation in normoxic conditions and Hog1 phosphorylation in hypoxic conditions was much lower in ptp3Δ compared to the wild-type strain. Thus, the inability of the ptp3Δ strain to grow in hypoxic conditions is not due to hyperactivation of Hog1. These results indicate that the interaction between Ptp3 and Hog1 is different between C. neoformans and S. cerevisiae. Interestingly, hog1Δ grew better than the wild-type in 1% O2 and ptp3Δhog1Δ, ras1Δhog1Δ and cdc24Δhog1Δ double deletants also grew as well as the wild-type in 1% O2. It is possible that the enhanced growth of hog1Δ somehow compensates for the growth deficiency of ptp3Δ, ras1Δ and cdc24Δ in hypoxic conditions thereby enabling the double deletants to grow in 1% O2. The detailed mechanisms of interactions between HOG1 and PTP3, RAS1, or CDC24 are yet to be elucidated.
We employed RNA sequencing to elucidate the putative pathways that are affected by mutations in RAS1, CDC24, and PTP3. Based on the results of the GO term functional clustering analysis, we identified several pathways that are commonly affected by these mutants that were verified by phenotypic studies using small molecular inhibitors. Hypersensitivity of the mutants to inhibitory compounds has been used extensively as a tool to dissect the gene function. Such approaches may have a disadvantage in that targets of some of the drugs may not be specific. For example staurosporine is a broad range kinase inhibitor and the observed phenotype may not be due to targeted effect. To partially circumvent such a caveat, two different inhibitors were chosen for each examined pathway. Hypersensitivity of the mutants to fluconazole and fenpropimorph suggests the association of Ras1, Cdc24 and Ptp3 in the ergosterol biosynthesis pathway. Time course analysis of sterol profiles confirms that ergosterol biosynthesis is indeed affected in these mutants. Nevertheless, correlation between the amounts of the ergosterol intermediates and transcript levels of the four selected genes involved in ergosterol biosynthesis pathway was not apparent. It is possible that higher amounts of ergosterol intermediates occurring in the stressed ras1Δ and cdc24Δ mutants resulted in a compensatory reduction in transcript levels of the four selected ERG genes. At least six ergosterol intermediates were accumulated to higher levels in ras1Δ and cdc24Δ, but not in ptp3Δ. In contrast, ptp3Δ produced lower amounts of ergosterol compared to the wild type while the amount of ergosterol did not change in ras1Δ and cdc24Δ suggesting that Ptp3 functions differently in regulating ergosterol biosynthesis compared to Ras1 and Cdc24. Furthermore, it has been shown that deletion of the gene encoding sterol regulatory element-binding protein, SRE1, causes a reduction of growth in hypoxic conditions [5], [6]. The sre1 mutants also displayed a decrease in ergosterol content and increase in several ergosterol intermediates but the patterns of the changes are different from the results of our mutants. Therefore, it is likely that the mechanism regulated by Sre1 is different form Ras1, Cdc24 and Ptp3.
Inhibitors studies show that ras1Δ, cdc24Δ and ptp3Δ all exhibited various degrees of hypersensitivity to inhibitors of chromosome organization, RNA processing and protein translation in a normoxic condition. Although it is possible that the hypersensitivity phenotype is due to inhibition of growth in otherwise already defective strains, it is more likely that these biological processes are perturbed in these mutants. The retarded growth of the wild-type strain in the presence of these drugs under hypoxia apparently suggest that C. neoformans requires these biological processes to be fully functional in the management of stress generated by hypoxic conditions. Post-translational modification of histone proteins is known to play a critical role in regulating chromatin structure (for a review see [59]). A recent study has demonstrated that the S. pombe histone H2A dioxygenase Ofd2 regulates gene expression during growth in hypoxia, which suggests that chromosome organization is important to the hypoxia response [60]. However, the exact modification of H2A by Ofd2 remains to be determined. We have also isolated a cryptococcal asc1 mutant that displayed a hypoxic phenotype (data not shown). Asc1 of C. neoformans shares similarity to S. cerevisiae Asc1 and mammalian RACK1, which is a conserved core component of the eukaryotic ribosome and functions in translational control [61]. Furthermore, previous functional analysis of C. neoformans mutants showed that the ability to respond to reactive oxygen species and mitochondrial function are important for growth under CoCl2 and low oxygen conditions [8]. Taken together, it is clear that C. neoformans overcomes hypoxic stress by employing many fundamental biological processes that had not been implicated in other organisms.
All strains used in this study were derived from the genome sequence strain H99 and are listed in Table S1. YEPD contains 1% yeast extract, 2% Bacto-peptone and 2% dextrose. YES medium contains 0.5% yeast extract, 3% glucose and 225 µg/ml each of uracil, adenine, leucine, histidine, and lysine. Low-oxygen conditions (5% CO2 and 1% O2) were maintained using an Invivo2 400 workstation at the indicated temperatures (Ruskinn, UK). For simplicity, we only mention the oxygen concentration without describing the CO2 concentration unless specified in our study. The GasPakPlus anaerobic system (BRL) was used to generate 0% O2 conditions. Spot assay was conducted by serial dilutions of cultures, spotted on agar plates, incubated at conditions as indicated for 3–6 days.
Zeiss Axiovert fluorescent microscope equipped with an AxioCam MRm digital camera was used to visualize fluorescent and differential interference contrast microscopy (DIC) images. Axiovision (version 4.0) was used to capture the images that were further processed by Adobe Photoshop CS4 software. Percentage of the polarized cells was determined as described [13].
cdc24Δ was transformed with a multicopy suppressor library made in a multicopy episomic plasmid, pYCC725, from LP2, a B-3501 derived strain [62]. Suppressor clones were selected by plating the transformants on a medium containing 100 µg/ml nourseothricin at 1% O2 and 28°C. Three separate screens were performed and 17 transformants were isolated. The plasmids in transformants were isolated and reconfirmed for their ability to restore the growth of cdc24Δ at 1% O2. The inserts in the episomes of the transformants were PCR-amplified using primers OYC725B and OYC725C and the amplified PCR fragments were sequenced as described [62]. The genomic content in the PCR amplified region was identified by a BLAST search of the B-3501 genome sequence. Episomal plasmids from each clone were rescued in E. coli and retransformed into the cdc24Δ strain to confirm the phenotype. For HRD1 and GEF1, the entire gene region was cloned by PCR and reconfirmed for its ability to suppress the hypoxia phenotype of cdc24Δ.
Gene deletion was carried out via homologous recombination by biolistic transformation [63]. Overlapping PCR technique was used to generate deletion constructs [64]. PCR and Southern hybridization was used to confirm homologous integrations. Wild type genes were PCR amplified from the strain H99, cloned and sequenced as described [9]. To generate complementation construct, each gene was either inserted in the multiple cloning site of pYCC744 which contained the NAT gene as a selectable marker [9] or cloned by overlapping PCR with Hygromycin gene as selectable maker.
Cells were grown in YPD medium to early log phase in normoxic conditions, spun down and transferred to 1% O2 for 1 hr. Equal volume of ice-cold stop buffer (0.9% NaCl, 1 mM NaN3, 10 mM Na-EDTA, 50 mM NaF) was added to the culture. Cells were spun down, washed once in ice-cold stop buffer and lyophilized. Lyophilized cells were disrupted with 1-mm zirconia/silica bead using FastPrep-24 (MP Biomedicals, CA) and resuspended in lysis buffer as described [36]. Cell lysates were spun down at 4°C and protein concentrations were determined using Bio-Rad Bradford Protein assay reagent (Richmond, CA). An equal amount of protein (20 µg) was loaded on the Any kD Criterion TGX Stain-Free gel (Bio-Rad, Richmond, CA) and proteins were transferred to a PVDF membrane. The western blot was incubated with a rabbit phospho-p38-MAPK antibody (Cell Signaling, Beverly, MA) and with a secondary anti-rabbit horseradish peroxidase–conjugated antibody. The blot was developed using the Clarity Western ECL (Bio-Rad, Richmond, CA). The signal was quantitated using ChemiDoc MP imaging system (Bio-Rad, Richmond, CA). The blot was stripped and used for detection of Hog1 with a rabbit polyclonal anti-Hog1 antibody (Santa Cruz Biotechnology, Santa Cruz, CA). The ratio of the signal intensity between phosphorylated Hog1 and total Hog1 was calculated and expressed as relative phosphorylation levels of the wild-type control.
To isolate RNA from 1% O2 grown cells, overnight cultures of wild-type and each deletant strain were refreshed in fresh YPD media for 5 h in ambient air at 28°C and shifted to 1% O2 for 2 h at 28°C. RNA was extracted from cryptococcal cells using Trizol (Invitrogen, Carlsbad, CA) and purified with RNeasy MinElute cleanup kit (Qiagen, Valencia, CA). The RNAseq was performed at RML Research Technologies Section, NIAID, NIH. The Illumina TruSeq RNA Sample Preparation Kit (Illumina, San Diego, CA) and its workflow were used for the preparation of barcoded RNA-Seq libraries. Final library products were quantified by qPCR using a KAPA Illumina GA Library Quantification Kit (KAPA Biosystems, Boston, MA) and sequenced on a HiSeq 2000 (Illumina) to produce paired 100 bp reads. Each of the 12 RNAseq libraries were given a unique barcode and pooled for clustering. We used 4 lanes on the HiSeq for sequencing, each lane containing the 12 libraries. Initial processing was performed using Illumina pipeline (CASAVA 1.8). TruSeq adapters were then trimmed and the reads were quality filtered with the FastXToolkit. All reads were mapped, using TopHat vl. 3.0 [65], to the C. neoformans H99 reference genome provided by the Broad Institute. The total number of sequence reads ranged from 78 to 102 million pairs of which nearly 87.6%–88.9% were uniquely aligned and properly paired to the reference genome assembly. The mapped reads were then used downstream by Cufflinks software packages [65] for transcript assembly, differential expression, and for gene model construction for RNA-Seq. A union of transcripts from all four samples was produced by Cuffcompare from individually assembled transcripts of each sample which yielded 6967 transcripts. Cuffdiff was used to determine statistically significant differences between strains on assembled transcripts. Gene expression was normalized using the number of fragments per kilobase of transcript per million mapped reads (FPKM). Venn's diagram was produced by taking the statistically significant differentially expressed genes with FPKM cut-off of >10. Gene ontology analysis was performed using DAVID Functional Annotation Tools and with the tool of functional annotation clustering and GO Fat database (GOTERM_BP_FAT) [37]. Based on the homology between H99 and JEC21, each of the H99 gene identification designated by Broad institute was converted to the JEC21 gene identification before using DAVID since DAVID did not recognize the Broad institute identification. The analyzed RNA-Seq data has been submitted to the Sequence Read Archive (SRA) at NCBI and can be viewed under accession number of experiment SRX347687.
Log phase grown cells were spun down and transferred to 1% O2 for 1, 3 or 5 hrs or 20% O2 for 2 hrs in YPD medium. Cells were harvested for sterol extraction. Sterol extraction and analysis was carried out as described [31]. Briefly, 10 mM of sodium azide was added to the culture medium immediately before harvest. Cells (1×108 total) were harvested by centrifugation and resuspended in 9 ml methanol and 4.5 ml 60% (wt/vol) KOH together with 5 µg cholesterol which was used as an internal recovery standard. Samples were heated to 75°C in a water bath for 2 h to complete the saponification and the sterols were then extracted with hexane. The extracted sterols were analyzed and characterized by GC-MS (gas chromatography-mass spectrometry) using an ISQ mass spectrometer from Thermo Electron, coupled to a Trace GC Ultra chromatograph, from Thermo Electron, in the EI (electron impact ionization) mode. This GC-MS instrument used a Restek 5MS fused silica column (30 m length, 0.25 mm i.d., 25 µm film thickness) with a program temperature from 200°C (1 min) to 300°C at a rate of 10°C/min. Cholesterol and ergosterol were identified by comparison with standards. The mass spectral data used in combination with a database (NIST/EPA/NIH Mass Spectral Library version 2.0) and published spectra from C. neoformans [66] allowed the identification of several peaks for which there were no standards available. The nomenclature of the discernable ergosterol intermediates was followed according to the C. neoformans study [66]. Relative amounts of each ergosterol intermediate were determined by comparing the area under the peaks in the chromatogram versus the area under the cholesterol peak to correct for recovery.
Isolation and analysis of genomic DNA was carried out as described previously [67]. RNA was treated with RNAse-free DNAse (Ambion, Austin, TX) to remove genomic DNA before quantitative real time reverse transcription PCR (qRT-PCR). cDNA was synthesized using high-capacity cDNA archive kit (Applied Biosystems, Foster City, CA). The qRT-PCR was performed using 20 µl triplicate reactions with SYBR select master mix from two biological replicates and the ABI PRISM 7500 sequence detection system (Applied Biosystems, Foster City, CA). The PCR efficiency and CT determination was performed using the algorithm as described [68]. The primers used for RT-PCR are listed in Table S2. Data were normalized with ACTIN1 level and expressed as the amount in each deletant strain relative to that of H99.
|
10.1371/journal.pgen.1002162 | Interactions between Glucocorticoid Treatment and Cis-Regulatory Polymorphisms Contribute to Cellular Response Phenotypes | Glucocorticoids (GCs) mediate physiological responses to environmental stress and are commonly used as pharmaceuticals. GCs act primarily through the GC receptor (GR, a transcription factor). Despite their clear biomedical importance, little is known about the genetic architecture of variation in GC response. Here we provide an initial assessment of variability in the cellular response to GC treatment by profiling gene expression and protein secretion in 114 EBV-transformed B lymphocytes of African and European ancestry. We found that genetic variation affects the response of nearby genes and exhibits distinctive patterns of genotype-treatment interactions, with genotypic effects evident in either only GC-treated or only control-treated conditions. Using a novel statistical framework, we identified interactions that influence the expression of 26 genes known to play central roles in GC-related pathways (e.g. NQO1, AIRE, and SGK1) and that influence the secretion of IL6.
| Glucocorticoids (GCs) are steroid hormones produced by the human body in response to environmental stressors. Despite their key role as physiological regulators and widely administered pharmaceuticals, little is known about the genetic basis of inter-individual and inter-ethnic variation in GC response. As GC action is mediated by the regulation of gene expression, we profiled transcript abundance and protein secretion in EBV-transformed B lymphocytes from a panel of 114 individuals, including those of both African and European ancestry. Combining these molecular traits with genome-wide genetic data, we found that genotype-treatment interactions at polymorphisms near genes affected GC regulation of expression for 26 genes and of secretion for IL6. A novel statistical approach revealed that these interactions could be distinguished into distinct types, with some showing genotypic effects only in GC-treated samples and others showing genotypic effects only in control-treated samples, with differing phenotypic and molecular interpretations. The insights into the genetic basis of variation in GC response and the statistical tools for identifying gene-treatment interactions that we provide will aid future efforts to identify genetic predictors of response to this and other treatments.
| Glucocorticoids (GCs) are steroid hormones that mediate homeostatic responses to environmental stressors through the regulation of critical physiological processes (e.g. immune response, energy metabolism and blood pressure (reviewed in [1])). Owing to early observations of the anti-inflammatory properties [2] of cortisol (i.e. the endogenous GC in humans), synthetic GCs are widely used as pharmaceuticals for inflammatory and autoimmune diseases (e.g. asthma [3] and rheumatoid arthritis [4]). GCs are also used in the treatment of several types of cancer [5], most notably lymphoid malignancies [6], due to their pro-apoptotic activities and for symptomatic relief. While there is evidence for a substantial genetic contribution [7]–[12], and for inter-ethnic differences in drug response [13], [14], little is known about the genetic architecture of variation in GC response within and between human populations.
Genetic effects on GC action could provide a mechanism for a vast array of gene-environment interactions, which could have major implications for human phenotypic variation. In fact, evidence of such interactions has been observed for numerous traits relevant to GCs including obesity [15], cardiovascular disease [16] and asthma [17]. With few exceptions (e.g. a regulatory polymorphism in the promoter of IL6 [18]), little is currently known about the mechanisms that underlie gene-environment interactions. If not properly accounted for, these interactions can complicate efforts to identify genetic and environmental factors associated with disease risk. Furthermore, identifying genetic variation that interacts with pharmaceutical treatments like GCs, which are a specific subset of environmental factors, is of particular interest from a clinical perspective and constitutes the primary goal of pharmacogenetics.
As GCs act largely by inducing changes in the expression of target genes [19], regulatory polymorphisms are likely to contribute to variation in response. The initial steps of the GC response pathway are mediated by the GC receptor (GR) and interacting transcription factors. GC binding allows the GR to translocate from the cytoplasm to the nucleus, where it regulates gene expression through at least two distinct mechanisms. The GR can either drive the assembly of novel transcriptional regulatory complexes at target genes, or inhibit regulatory complexes, such as NFκB [20], that are already active at target genes. Some direct GR target genes are, in turn, transcription factors that regulate downstream target genes.
Here, we provide an initial view of the genetic architecture of variation in the GC-mediated regulation of transcription and protein secretion. To accomplish this, we measured the expression of 13,232 genes and the secretion levels of 10 proteins in paired aliquots, one treated with the synthetic GC dexamethasone (dex) and one treated with the vehicle for dex (EtOH) as a control, in a panel of 114 densely genotyped HapMap B-lymphocytes transformed with Epstein-Barr Virus (EBV), commonly known as lymphoblastoid cell lines (LCLs). This panel included 57 Yoruba (YRI) from Nigeria and 57 Toscani (TSI) from Italy. EBV transformation proceeds, in part, by mimicking CD40 activation and ultimately leads to cellular proliferation through a variety of mechanisms, including the activation of the NFκB signaling pathway [21]. Given their activated state, LCLs are a suitable system for studying the immunorepressive effects of GCs. Additionally, some regulatory variants that affect GC response in LCLs may be shared with other cell types, as observed for baseline expression [22]–[24].
We found that 4,568 genes were differentially expressed, at a FDR<0.01 (p<0.003), following treatment with GCs (8 h, 1 uM dexamethasone), corresponding to ∼32% of the expressed genes. This number is similar to that observed in a recent study of equivalent sample size in osteoblasts treated with GCs [25], but larger than previous studies that used much smaller samples (often a single cell line; e.g. [26]). This suggests that large sample sizes are necessary to identify many GC target genes. Accordingly, we found that sub-sampling data from our full panel of LCLs dramatically reduced the number of differentially expressed genes, especially at genes with inter-individual variation in transcriptional response (Figure S1). It should be noted that tests of differential expression rely on magnitude of transcriptional response and its consistency across individuals. Because our main goal is to identify the genetic basis of variation in response, we did not limit our mapping analyses (see below) to the differentially expressed genes.
Among the differentially expressed genes in LCLs, we found roughly equal numbers of up and down-regulated genes. Up-regulated genes were enriched for GC-related biological processes including cellular response to stimulus (p = 4.1×10−6, FDR = 7.5×10−5) and cell cycle (p = 1.4×10−5, FDR = 2.6×10−4), consistent with GC regulation of lymphocyte proliferation. Down-regulated genes were enriched for immune response genes (p = 1.1×10−10, FDR = 4.6×10−9) and for genes involved in the positive regulation of I-kappaB kinase/NF-kappaB cascade (p = 3.3×10−5, FDR = 3.3×10−4), consistent with the immunorepressive role of GCs.
To explore the extent of tissue-specificity in the transcriptional response to GCs, we compared our data to the results in osteoblasts [25]. We found a significant overlap between the genes differentially expressed in LCLs and in osteoblasts (p = 4.8×10−13), but only 28% of genes differentially expressed in our study are differentially expressed (p<0.05) in osteoblasts. This likely reflects some amount of tissue specificity, although other factors are likely to contribute (e.g. incomplete power [23], differences in duration of treatment).
We measured and corrected for multiple factors related to EBV-transformation that have been previously shown to be associated with gene expression patterns at baseline [27], [28] (e.g. EBV copy number). Unlike baseline expression, these factors showed hardly any evidence for an effect on transcriptional response (see Table S1); nonetheless, we corrected for them in all subsequent analyses.
Many of the proteins involved in the GC-mediated regulation of transcription are well characterized (i.e. GR and interacting transcription factors). Genetic variants that impact the function of these regulatory proteins are likely to influence transcriptional response at several, and potentially many, downstream genes. Consequently, the genes that encode these proteins are candidate expression quantitative trait loci (eQTLs) acting in trans to modulate the transcriptional response to GCs. However, genome-wide tests for trans eQTLs suffer from a tremendous multiple testing burden. Therefore, to reduce the number of tests being performed, we first examined only these candidate genes for response eQTLs. We used simple linear regression to test for an association between log fold change in expression at each expressed gene in the genome and genotype at all HapMap SNPs within 100 kb of the gene that encodes the GR (NR3C1), and found no significant evidence of association at a FDR<0.2 (Figure 1a, Figure S4a-S4b). Similarly, as the GR interacts with other transcription factors in the regulation of target gene transcription, we also tested all HapMap SNPs within 100 kb of 34 genes that encode transcription factors known to interact with the GR (listed in Materials and Methods [29]). Here again, we found no evidence for an effect of genetic variation at these loci on the transcriptional response to GCs at a FDR<0.2 (Figure 1b, Figure S4c-S4d).
We then performed an unbiased, genome-wide scan for genetic variation associated with GC response. Specifically, we tested for an association between every HapMap SNP and log fold change at every gene. While this analysis did not reveal any significant associations at a FDR<0.2, we found that the top association was between log fold change at C1orf106 and genotype at an intronic SNP (rs4915463, p = 8.4×10−11, FDR<0.67). Given the proximity of the associated SNP to the C1orf106 locus and work by others highlighting the impact of cis-acting regulatory polymorphisms on baseline expression [30], [31], we then focused our analyses on HapMap SNPs near each of the 12,619 expressed, autosomal genes. We found the strongest signal when we tested for an association between log fold change at each gene and all SNPs within 100 kb (compared to either a genome-wide scan or testing SNPs within 500 kb of each gene, Figure 1c). This analysis revealed local response eQTLs for 8 genes at a FDR<0.1 (Figure 2). These included genes previously shown to play important roles in GC-related biological processes, including regulation of immune response (MT1X [32] and MFGE8 [33]) and cell cycle progression (e.g. BIRC3 [34]). These also included NQO1, a gene previously shown to affect variation in response to GC pharmaceutical treatment [35].
Visual examination of the genes in Figure 2 indicates that different genes show qualitatively different patterns. For some genes, a genotypic effect is evident either in only the GC-treated condition (C1orf106, NQO1, C9orf5, MFGE8, and BIRC3) or only the control-treated condition (MT1X). For others, an effect is evident in both, but differs between the two conditions (DNAJC5G and MS4A7). These different patterns may have different mechanistic and phenotypic interpretations, but are not distinguished by the test of log fold change, and so researchers have previously been forced to identify such patterns post hoc (e.g. [36]). To address this, we developed a novel statistical framework that explicitly compares and identifies these different patterns of interaction.
In brief, our method explicitly compares five different models relating each SNP to phenotypic measurements in the two treatment conditions (GC and control):
For each SNP, we computed a likelihood ratio, or Bayes Factor (BF) that measures the relative support in the data for each model 1–4. These BFs take account of the paired nature of the data, and the correlations between measurements in the same LCL in different conditions. We used a hierarchical model [37] to combine information both across SNPs in each gene region, and across genes, ultimately computing a posterior probability for each gene that it follows each of the models 1–4, i.e. that it is affected by a polymorphism that follows that model. We used these posterior probabilities both to identify high-confidence eQTLs of each type, and to estimate false discovery rates among eQTLs exceeding any given posterior probability threshold. This method is broadly applicable to the study of any gene-environment interactions with paired phenotype measurements.
Using this novel framework we identified 26 genes with high-confidence interactions (posterior probability of interaction>0.7, FDR = 0.10) between GC treatment and eQTLs. These interaction eQTLs included 7 of the 8 response eQTLs identified by mapping log fold change. The remainder generally showed strong, but not genome-wide significant (FDR<0.10), association with log fold change (see Table S2 and Figure S2). The larger number of interactions identified compared with mapping log fold change (26 versus 8), therefore, reflects an increase in power that comes from explicitly considering different plausible interaction scenarios. Of these, the majority (18 of 26) showed strongest support for GC-only interactions, with the remainder (8 of 26) showing strongest support for control-only eQTLs. Only one interaction between treatment and genotype identified through mapping the log fold change was not identified by the Bayesian hierarchical model (DNAJC5G). This interaction was the least significant of the 8 identified by mapping log fold change, and although it also shows some signal in the Bayesian analysis (BF for general interaction versus null = 4.9×102, BF for general interaction versus no-interaction = 6.4×103), the signal was insufficient to outweigh the low prior probability of a general interaction estimated by the hierarchical model (prior = 0.001, see Table S4).
The Bayesian hierarchical model revealed eQTLs at genes with clear biological relevance to GC-related biological processes that were not identified through mapping the log fold change. These include additional genes involved in the regulation of immune response (e.g. CST7 [38] or NLRP2 [39]) and cell cycle progression (e.g. PDGFRL [40]), well-established GC targets, such as serum and glucocorticoid regulated kinase 1 (SGK1 [41]), and previously unknown GC target genes. For example, we found a control-only eQTL for multiple coagulation factor deficiency 2 (MDCF2), which is involved in the production of pro-coagulation factors [42]. The effect of GCs on coagulation is controversial [43], but has been suggested to play a role in their therapeutic effects on diseases such as asthma [44].
Given that cortisol regulates a variety of physiological processes relevant to numerous diseases, we compared our eQTL results to those from genome-wide association studies collected as a part of the GWAS catalog [45]. We found that a GC-only eQTL for AIRE (rs762421) was associated with risk of the Crohn's disease [46]. AIRE encodes a potent repressor of autoimmunity and can cause severe autoimmune disease when mutated [47]. In addition to its role in removing autoreactive T cells in the thymus, AIRE also plays a role in B-cell mediated immune response [48]. We found that the putative risk allele (rs762421-G) is associated with the down-regulation of AIRE expression by GCs. This allele may confer increased susceptibility to this autoimmune disease by allowing GCs to decrease AIRE expression.
In addition to these interacting polymorphisms, our analysis identified a much larger number of genes (6,813 genes) affected by no-interaction eQTLs (posterior probability>0.7; FDR = 0.16). In other words, transcript levels at these genes depend on the eQTL genotype, but the magnitude of transcriptional response does not (i.e. model 2, see Figure S10). Our observation that the vast majority of cis-acting regulatory polymorphisms with identical genotypic effects across treatment conditions is consistent with findings in osteoblasts treated with GCs [25] and in yeast [49], suggesting that this may reflect a general biological trend, rather than a feature specific to our treatment and experimental system. We compared the distribution of minor allele frequencies between the eQTLs following these three models and did not observe any significant differences (Figure S9).
The results reported above come from using a hierarchical model, which combines information across SNPs within each gene. One limitation of this hierarchical model is that it allows at most one eQTL per gene. This may cause it to miss interacting SNPs in genes that contain both interacting and non-interacting eQTLs, and for this reason the probabilities on interacting models may be underestimated. (More generally this feature could cause apparent discrepancies between the results from the hierarchical model and the log fold change analysis, although this does not seem to be the case in the results above.) To assess whether this limitation might have led us to miss some strong interaction signals we also performed a SNP-level analysis using the BF (for interaction models 2–4 vs non-interaction models 0–1) computed for each SNP. This analysis identified 247 SNPs, in 120 distinct genes, with BF exceeding 103, although none exceeding 105, that are candidates for being interacting eQTLs (Table S5).
To determine whether GC-only and control-only eQTLs represented regulatory polymorphisms with treatment-specific genotypic effects, we assayed treatment-dependent allelic imbalance using quantitative real time PCR in heterozygotes. This assay also asks whether local eQTLs act in cis, as alleles at cis-regulatory polymorphisms, by definition, affect target gene transcription only on the chromosome on which they reside. Among the 26 interaction eQTLs, we chose five at random among those for which a common coding SNP could be reliably genotyped. We assayed three genes with evidence of GC-only eQTLs (C9orf5, LSG1, and MFGE8). We found significant allelic imbalance, with allelic effects in the same direction as predicted by the eQTL mapping results, in GC-treated samples, but not in control-treated samples, for all of them (Table 1). We also performed allelic imbalance assays on 2 of the 8 control-only eQTLs (SRD5A2, C12orf45). We found significant allelic imbalance at C12orf45 only in the control-treated samples. While not significant at p<0.05, we observed a pattern consistent with a control-only eQTL at SRD5A2. Our failure to fully validate all 5 assayed eQTLs by allelic imbalance could reflect some level of false positive identifications of eQTL interactions, but may also reflect incomplete power of the allelic imbalance assay.
We compared our results with those from an independent GC response eQTL mapping study in LCLs derived from asthma patients (W. Qui and K. Tantisira, personal communication). We found that 4 of the 9 interaction eQTLs that we identified, and that were tested in both studies, showed significant associations with log fold change in this independent dataset (p<0.05, C1orf106, LSG1, CST7, and MS4A7), and an additional 2 showed suggestive associations (p<0.1, SYT17 and BIRC3). This overlap is highly significant (p = 8.5×10−4). Importantly, the overlap for single-treatment eQTLs is much greater than that for response eQTLs: all of the top 10 eQTLs identified by Qiu et al (2011) in each treatment condition were replicated in our data (p<0.05), while only 1 of the top 10 eQTLs for log fold change was replicated. This contrast highlights the known statistical challenge of mapping gene-environment interactions.
We also tested 15 of our interaction eQTLs (i.e. all eQTLs tested in both studies) for an association with response to GC therapy in 172 asthma patients (W. Qui and K. Tantisira, personal communication). We found that a GC-only eQTL for TNIP1 was significantly associated with patient response (rs6870205, p = 2.5×10−3, Bonferroni-corrected p = 0.037). TNIP1 has an established role in the immune response, as it encodes a protein that inhibits NFκB [50] and contains polymorphisms that have been associated with risk of systemic lupus erythematosus [46].
We observed substantial allele frequency differences between populations at many of the putative interaction expression quantitative trait nucleotides (eQTNs), defined as the most strongly associated SNP for each gene. Furthermore, differences in allele frequency at these eQTNs were predictive of differences in average transcriptional response between populations (r2 = 0.33, p = 5.3×10−3, Figure 3a). This demonstrates that these eQTNs contribute to differences in response between populations, and so may also contribute to inter-ethnic disparities in GC-related diseases and in drug response. It also provides independent supporting evidence that these eQTNs interact with GC treatment.
In some cases, allele frequency differences may explain why genes respond to GC treatment only in individuals of one population. For example, we observed that the GC-only eQTL allele associated with up-regulation of the detoxification enzyme NAD(P)H:quinone oxidoreductase 1 (NQO1) was extremely rare outside equatorial African populations (Figure 3b), likely causing the observed lack of NQO1 response in TSI LCLs, and the strong up-regulation in many YRI LCLs (Figure 3c). This result may be of particular relevance to ethnic disparities in leukemia patient response to GCs, as alleles that reduce NQO1 enzymatic activity have been associated with decreased response to a chemotherapy regime that included GCs in patients with acute lymphoblastic [51], [52] and acute myeloid leukemia [53].
In an effort to identify additional genes with differences in average transcriptional response between populations, we applied the same statistical framework described above to test for interactions between population (rather than genotype) and GC treatment. Using this approach, we identified 258 genes with differences in transcriptional response (posterior>0.7, FDR = 0.128) between populations; of these, 130 were up-regulated by GC treatment while 128 were down-regulated. We found a consistent pattern across genes, with a tendency for stronger up-regulation in YRI LCLs at 78% of up-regulated genes with population differences in response (Figure S3). Interacting eQTLs are enriched among genes with population differences in response compared to all expressed genes (odds ratio = 6.0, p = 5.4×10−3) while no-interaction eQTLs are not enriched (odds ratio = 0.99).
The attenuation of the immune response by GCs is partially mediated by decreased secretion of pro-inflammatory molecules. We measured the secreted levels of 9 pro-inflammatory proteins (IL1α, IL6, IL8, IP10, MDC, Rantes, TNFα, TNFβ) and 1 anti-inflammatory protein (IL10). Five pro-inflammatory proteins showed significant differential secretion in response to GCs in LCLs (TNFα, TNFβ, Rantes, IP10 and IL1α –Table S3); all five showed lower secretion levels in the presence of GC, consistent with the immune-repressive role of GCs. To identify genetic variation that influences GC-mediated regulation of protein secretion, we tested HapMap SNPs for association with log fold change in secretion at each protein. Similar to our eQTL results, we found significant associations (at a FDR<0.2) only when we limited our search to SNPs near the genes that encode each protein (i.e. we found no significant associations in genome-wide or a candidate gene analysis). Testing SNPs within 100 kb of each cytokine, we found a significant association between secretion response at IL6 and genotype at a SNP ∼56 kb downstream (rs10225286, p = 1.9×10−4, FDR = 0.1, Figure 4). Because this SNP did not show strong evidence of an effect on IL6 transcriptional response, we propose that it affects secretion through a mechanism independent of mRNA levels or that it affects transcriptional response at a different treatment time point.
Here, we report a genome-wide scan for genetic variation that influences the GC-mediated regulation of transcription and protein secretion. The cellular response to GCs depends on a well-characterized set of regulatory proteins (i.e. the GR and interacting proteins). This provided us with a set of strong candidate loci to perform trans-eQTL mapping tests. Despite this, we found no evidence for trans-acting factors. In contrast, the strongest signal from an unbiased genome-wide scan was a SNP associated with transcriptional response at a nearby gene, and even more eQTLs were revealed when we limited our analysis to SNPs within 100 kb of each gene. Numerous studies have tested genetic variation within or near the GR and interacting transcription factors for association with patient response to GC treatment. These studies have found mostly rare functional polymorphisms that are unlikely to explain most heritable variation in GC response (reviewed in [54]). Furthermore, rare polymorphisms in the GR have dramatic phenotypic effects (e.g. extreme hypoglycemia and hypertension [55]), as expected for a master regulator that influences all downstream processes. Instead of genetic variants in master regulators, our results suggest that cis-regulatory polymorphisms that interact with GC treatment at target genes could play an important role in GC response, as first suggested based on observations at the SGK1 gene [56]. These findings suggest that future attempts to identify genetic variation associated with clinical response to GCs may benefit from focusing on likely cis-regulatory polymorphisms that impact response at individual GC target genes, instead of testing master regulators of the GC response pathway.
We found that associations between genotype and transcriptional response could be discriminated into distinct categories based on the configuration of genotypic effects across treatment conditions. These categories likely correspond to specific genetic mechanisms. GC-only eQTLs may reflect polymorphisms that influence the binding of transcription factors that are only active in the presence of GC treatment (e.g. the GR and interacting transcription factors). In support of this hypothesis, we found that GC-only eQTLs tended to affect up-regulated genes (13 of 18). Although the causative polymorphism may not be among the genotyped SNPs, we found examples of GC-only eQTLs where most of the signal centered on a SNP that disrupts a predicted GR binding site, such as the eQTN for C9orf5 (rs10816772, p for motif match = 6.8×10−3).
Control-only eQTLs are compatible with a variety of mechanisms. For example, they may reflect polymorphisms that disrupt the binding of regulatory complexes, like NFκB, that are directly inhibited by the GR (e.g. through protein-protein interaction). Consistent with this, we found examples of control-only eQTLs where most of the signal centered on a SNP that disrupts a predicted binding site for a transcription factor directly inhibited by GR, such as the eQTN for FBXL6 (rs10448143, matrix and core similarity for NFkB>0.9). Direct inhibition of transcription factors by GR generally leads to down-regulation of target genes. However, we found equal numbers of control-only eQTLs affecting up-regulated and down-regulated genes (4 of each), so additional mechanisms must explain some fraction of control-only eQTLs. These may include genetic effects on regulatory elements that are indirectly inhibited by GC treatment (e.g. through GR competition for access to DNA by another transcription factor) or polymorphisms that affect transcriptional response at secondary targets.
The different categories of interactions identified by our method may also have distinct phenotypic interpretations. Polymorphisms with GC-only effects on expression are likely to directly affect the action of the GC-activated regulatory machinery. In contrast, polymorphisms with control-only effects have no impact on the cellular processes in the presence of GCs, but may still affect phenotype by influencing variation in a ‘pre-treatment’ state. For example, genetic effects on pro-inflammatory cytokine levels prior to GC exposure could affect the amount of time cells take to reach the optimal, lower levels required to effectively suppress inflammation. In summary, control-only QTLs may contribute more to variation in underlying disease mechanisms, while GC-only QTLs may contribute to variation in GC pharmacodynamics. However, we also note that, given their lower rates of validation and replication, there may be a higher false positive rate for control-only eQTLs.
Inter-ethnic differences in GC response have been observed clinically [13], [14], and the prevalence of many GC-regulated physiological traits differs across human populations [57]. By combining association mapping with comparisons between populations, our study allowed a direct assessment of the genetic basis of population differences in the cellular response to GCs. We found that ancestry had substantial and systematic effects on the transcriptional response to GCs, with a tendency for stronger up-regulation after GC treatment in YRI LCLs. Possible causes of such patterns include: non-genetic ‘confounders’ (e.g. differences in immortalization procedure [58]), trans-acting alleles that increase response and are at higher frequency in YRI, or multiple, independent cis-acting alleles that increase response in YRI at up-regulated genes. Our data favor the last explanation. It seems unlikely that non-genetic ‘confounders’ explain all or most of the population differences, as we found that the measured ‘confounders’ showed limited evidence of effects on transcriptional response or differences between populations (Figure S5). Although we cannot exclude the possibility that population differences reflect a trans-acting eQTL with differences in allele frequency, we found little support for this explanation. Instead, we found evidence suggesting that population differences may reflect differences in allele frequency at cis-regulatory polymorphisms, as genes with population differences in response were more likely to have local interaction eQTLs. The possibility that a stronger response in YRI reflects differences in allele frequency at cis-regulatory polymorphisms is particularly interesting from an evolutionary perspective, as differences in allele frequency acting in a consistent direction (i.e. increasing GC responsiveness) across multiple independent QTLs are usually interpreted as evidence of polygenic adaptation [59]–[61].
In addition to these biological insights, we contribute novel statistical methodology for mapping response phenotypes and identifying gene-environment interactions. These methods are applicable for any setting contrasting genotypic effects between two conditions (with paired measurements), including pharmacogenetic studies of clinical response to drug therapy (e.g. [62]) and, especially, functional genomic studies of genetic effects on treatment response similar to the one presented here. These methods provide a more powerful alternative to mapping a measure of response (e.g. log fold change), which fails to distinguish among different types of interactions, or comparing results from mapping separately in each condition, which ignores the paired nature of the data.
In summary, this study provides an initial characterization of the genetic basis of variation within and between human populations for a key physiological regulator and commonly administered pharmaceutical. The biological insights and statistical tools presented here extend our current understanding of the genetic basis of variation in response to GCs, and will aid future efforts to characterize the genetics of response to this and other treatments.
All cellular experiments described were conducted in lymphoblastoid cell lines (LCLs), B lymphocytes transformed with Epstein-Barr virus, that were collected as a part of the International HapMap project. LCLs were thawed and passed once in RPMI media supplemented with 15% fetal bovine serum, then washed twice with phosphate-buffered saline and moved to RPMI media supplemented with 15% charcoal-stripped fetal bovine serum. After one passage in media with charcoal-stripped fetal bovine serum (corresponding to a minimum culturing time of 5 days), LCLs were seeded in the evening at a density of 5×105 cells/ml. After an overnight incubation, LCLs were treated with 10−6 M dexamethasone, and an equal amount of vehicle solution (solution composed of 1% ethanol and 99% cell culture media) as a negative control for treatment. For each LCL, one set of dex and control aliquots was treated for 8 hours (to quantify mRNA abundance) and the other for 24 hours (to assay inflammatory protein secretion). The study design is depicted in Figure S6. LCLs were thawed, cultured and treated in batches completely balanced by treatment, population, technician and time of day. For quality control purposes, biological replicates were performed for one batch of four cell lines and both expression and treatment response were highly replicable (Figure S7). Collection of all samples took 4 months.
For each expression study described in the preliminary data, total RNA was extracted from each cell culture sample using the QIAgen RNeasy Plus mini kit, and was found to be of high quality. RNA was extracted from all 240 samples over the course of 5 days. Total RNA was then reverse transcribed into cDNA, labeled, hybridized to Illumina HumanHT-12 v3 Expression BeadChips and scanned at the Southern California Genotyping Consortium (SCGC: http://scgc.genetics.ucla.edu/) at the University of California at Los Angeles. Each RNA sample was hybridized to two separate arrays (i.e. in two technical replicates). To avoid batch effects on RNA measurements, all 480 microarrays were hybridized within 7 days. Summary data (e.g. mean intensity of each probe across within-array replicates) were obtained using the BeadStudio software (Illumina) at the SCGC. The microarray data has been deposited in the Gene Expression Omnibus (GEO), www.ncbi.nlm.nih.gov/geo, under accession number GSE29342.
Low-level microarray analysis was performed using the Bioconductor software package LUMI [63] in R (http://www.r-project.org). We used applied variance stabilizing transformation [64] to all arrays, removed probes with intensities indistinguishable from background fluorescence levels in all samples (leaving 23,700 expressed probes), and performed quantile normalization across all arrays. Probes were annotated by mapping to the RNA sequences from RefSeq using BLAT. To avoid ambiguity in the source of a signal due to cross-hybridization of similar RNA species, probes that mapped to multiple genes were excluded from further analyses. Probes that contained one or more HapMap SNPs were also removed from further analyses to avoid spurious associations between expression measurements and SNPs in linkage disequilibrium.
To avoid spurious results and to reduce noise due to potential confounders, we measured several covariates relevant to LCL biology including: EBV genome copy number, growth rate and mitochondrial genome copy number. EBV and mitochondrial genome copy number were assessed using Taqman Gene Expression Assays (Assay # Hs02596867_s1 for mitochondria and Pa03453399_s1 for EBV). RNaseP was used as an endogenous control for both assays. We then used linear regression to remove the effects of these potential confounders at each gene and confounder-corrected data were used in all subsequent analyses.
In order to identify genes that, on average across individuals, changed expression levels upon treatment with GCs, we performed multiple linear regression at each gene with treatment as the covariate of interest while taking other measured covariates into account. To reduce the effects of outliers, microarray intensity values were quantile normalized to a N(0,1) distribution across all samples (treated and untreated). We used the distribution of p-values observed when sample labels are permuted (ten permutations were used), an empirical estimate of the p-value distribution under the null, to estimate the false discovery rate (FDR). We used the online tool DAVID [65], [66] to identify biological categories enriched among differentially expressed genes, using all genes expressed in LCLs (based on microarray data) as a background.
We used all HapMap SNPs for all mapping experiments described. As TSI LCLs were only typed for phase III SNPs, we used the CEU population sample to impute genotypes at all HapMap phase I and II SNPs. Similarly, we imputed SNPs for phase III YRI LCLs based on the YRI LCLs included in phase I and II. Imputation was performed using BIMBAM [67], which infers missing genotypes based on correlations between missing and typed genotypes observed in samples where all genotypes are typed. QTL mapping results were not qualitatively different if using imputed or genotyped SNPs.
We tested for association between all HapMap SNPs and transcriptional response at each gene, using log fold change in expression (GC-treated over control-treated expression) as a measure of response. For our candidate gene-based scan for trans-acting eQTLs that influenced response, we tested all HapMap SNPs within 500 kb and 100 kb (in two separate sets of analyses) of genes encoding the GR and transcription factors that interact with the GR. Interacting transcription factors include the genes that encode the components of the NFkB complex, AP1, Oct1, Oct2, CREB, ETS1, STAT3, STAT5, STAT6, C/EBP, TFIID, T-bet, PU.1/Spi-1, Smad3, Smad4, Smad6, COUP-TFII, IRF3, STIP1, Hic5/Ara55, and nTrip6 [29]. P-values calculated with permutated genotype labels were used as an empirical null distribution. In order to maintain the correlation structure across genes, the same permutation seed was used for all genes in both candidate gene tests and the genome-wide scan. Ten permutations were performed for the test of variation within 500 kb, 100 permutations were used for the test of variation within 100 kb and 3 permutations were used for the genome-wide scan. For mapping log fold change at SNPs within 500 kb or 100 kb of each gene, permutation seeds were set separately at each gene. Association tests were performed using a combination of Python, the R statistical package and the genetic association mapping program PLINK.
We developed a novel Bayesian statistical framework for genetic association analysis in settings where measurements are available on the same individuals in two different conditions (in our case, GC-treated and control-treated). Our methods extend and improve the methods from Barber et al. (2009) to explicitly consider “qualitative interaction” models where genetic variants are associated with measurements in only one of the two conditions. Our method takes into account both sample pairing and the intra-individual correlation of measurements under the two conditions. We describe our method in greater detail in Text S1. These methods are implemented in software called BRIdGE (Bayesian Regression for Identifying Gene-Environment interactions), which is available on the Stephens and the Di Rienzo laboratories' web pages (http://stephenslab.uchicago.edu/software.html, http://genapps.uchicago.edu/labweb/index.html).
We used TaqMan quantitative genotyping assays to test for allelic imbalance at coding SNPs in LD with eQTLs that interacted with GC treatment. Imbalanced expression of the two coding alleles is an independent line of evidence for a cis-acting regulatory polymorphism and for the configuration of the effect in the two treatment conditions (i.e. the interaction model). Total RNA from an aliquot of the same culture samples used to hybridize microarrays (this was a separate RNA extraction as that used to hybridize microarrays) was synthesized into cDNA using the High-Capacity cDNA Reverse Transcription Kit (Applied Biosystems, Foster City, CA) according to the manufacturer's protocol. Taqman SNP Genotyping Assays were used to quantify relative mRNA abundance of each allele on an ABI PRISM 7900HT Sequence Detection System. To account for differences between the two fluorochromes, a standard curve was built for each of the two alleles using serial dilutions of a genomic DNA from an individual that was heterozygous at the coding SNP. For each assay, we calculated the natural log-ratio between the two different alleles. The numerator of this ratio was always the allele associated with increased expression in the corresponding treatment condition. Within each treatment, we quantile normalized allelic log-ratios and used a one-tailed t-test to identify significant differences in average allelic log-ratios between heterozygotes and homozygotes (as an empirical null distribution of allelic log-ratios) at the eQTL.
We compared our eQTL results to multiple genetic association studies including Qiu et al. (2011) and those in the GWAS catalog. For each interaction eQTL, we compared evidence at the most associated SNP in our data when it was tested in both studies. When the most associated SNP was not tested in the comparison dataset, we identified the best proxy SNP for each eQTL among those tested in both studies. To ensure that the best proxy SNP captured the pattern at the original eQTL, we required the proxy SNP to show strong evidence of association for the same eQTL model as the original eQTL (BF for association>500 and posterior probability for model>0.5).
We contrasted the transcriptional response to GCs between YRI and TSI LCLs. Differences in transcriptional response between populations will result in differences in average expression levels that differ depending on treatment, as opposed to GC-independent population differences that will be identical in both treatments. As this is analogous to gene-environment interactions, we used the same statistical framework to identify genes with differences in transcriptional response between populations (see Bayesian regression for identifying genetic associations and interaction with treatment in Text S1). Covariate-corrected expression levels were quantile normalized across individuals (both YRI and TSI) for each gene to reduce the effect of outliers. As population differences at the phenotypic level may reflect population differences in response following a consistent pattern across many genes, we identified the direction of population differences at each gene in terms of log-fold change.
A multianalyte ELISA assay (Millipore) was performed on the culture medium of the cell aliquots treated for 24 hours. The assay was performed at the Flow Cytometry Facility at the University of Chicago, according to the manufacturer instructions. Two technical replicates were run for each sample. Samples were assayed in batches balanced by treatment and population. For each analyte, the average quantity across technical replicates was calculated and used for all subsequent analyses. The correlation structure between paired aliquots for each sample (GC and control) was visually inspected (Figure S8). A small subset of samples with low quantity detected showed no correlation between GC and control aliquots because of noise in the measurement at low concentrations. Consequently, these samples were excluded from downstream analyses. Secretion levels were highly correlated across proteins, likely representing a latent factor that generally affects secretion levels. To remove the effect of this latent factor, we used linear regression to correct secretion levels at each protein by secretion levels at all other measured proteins.
|
10.1371/journal.ppat.1007532 | Interferon-induced transmembrane protein 3 blocks fusion of sensitive but not resistant viruses by partitioning into virus-carrying endosomes | Late endosome-resident interferon-induced transmembrane protein 3 (IFITM3) inhibits fusion of diverse viruses, including Influenza A virus (IAV), by a poorly understood mechanism. Despite the broad antiviral activity of IFITM3, viruses like Lassa virus (LASV), are fully resistant to its inhibitory effects. It is currently unclear whether resistance arises from a highly efficient fusion machinery that is capable of overcoming IFITM3 restriction or the ability to enter from cellular sites devoid of this factor. Here, we constructed and validated a functional IFITM3 tagged with EGFP or other fluorescent proteins. This breakthrough allowed live cell imaging of virus co-trafficking and fusion with endosomal compartments in cells expressing fluorescent IFITM3. Three-color single virus and endosome tracking revealed that sensitive (IAV), but not resistant (LASV), viruses become trapped within IFITM3-positive endosomes where they underwent hemifusion but failed to release their content into the cytoplasm. IAV fusion with IFITM3-containing compartments could be rescued by amphotericin B treatment, which has been previously shown to antagonize the antiviral activity of this protein. By comparison, virtually all LASV particles trafficked and fused with endosomes lacking detectable levels of fluorescent IFITM3, implying that this virus escapes restriction by utilizing endocytic pathways that are distinct from the IAV entry pathways. The importance of virus uptake and transport pathways is further reinforced by the observation that LASV glycoprotein-mediated cell-cell fusion is inhibited by IFITM3 and other members of the IFITM family expressed in target cells. Together, our results strongly support a model according to which IFITM3 accumulation at the sites of virus fusion is a prerequisite for its antiviral activity and that this protein traps viral fusion at a hemifusion stage by preventing the formation of fusion pores. We conclude that the ability to utilize alternative endocytic pathways for entry confers IFITM3-resistance to otherwise sensitive viruses.
| Expression of interferon-induced transmembrane proteins (IFITMs) in target cells potently inhibits fusion of many unrelated enveloped viruses, including the Influenza A virus, whereas arenaviruses, such as the Lassa fever virus, are resistant to these factors. The mechanism by which IFITMs interfere with the viral fusion step and the mechanism of virus escape from these restriction factors are poorly understood. Here, we tagged the late endosome-resident IFITM3 with fluorescent proteins and visualized single virus entry and fusion with endosomes in living cells expressing these constructs. Single virus and endosome tracking experiments demonstrate that the sensitive Influenza A virus is trapped within acidic IFITM3-positive endosomes that are not permissive for viral fusion. In contrast, the resistant Lassa virus consistently enters and fuses with endosomes lacking IFITM3. Our results imply that accumulation of IFITM3 in virus-carrying endosomes is a prerequisite for blocking fusion of diverse enveloped viruses and that viruses insensitive to this protein escape restriction by entering through distinct endosomal trafficking pathways that do not converge with IFITM3-positive compartments.
| Fusion of enveloped viruses with the host cell membrane is a key step leading to infection. Viral fusion is initiated upon interactions between virus surface glycoproteins and cellular receptor(s) and/or upon the reduction in pH that follows endocytosis (reviewed in [1–3]). The extensive conformational changes that ensue in the viral glycoproteins promote fusion between viral and cellular membranes [4–6]. There is strong evidence that viral fusion—and membrane fusion in general—proceeds through a hemifusion intermediate defined as a merger of two contacting membrane leaflets without additional merger of distal leaflets that results in the formation of a fusion pore [6–8]. Accordingly, hemifusion is manifested as lipid mixing between viral and host membranes without viral content release, while full membrane fusion entails mixing of distinct aqueous contents delimited by the two membranes [4–6]. It has been demonstrated that sub-optimal conditions for membrane fusion including low density of viral glycoproteins, reduced temperature, and—where applicable—insufficiently acidic pH, favor dead-end hemifusion that does not progress to full fusion [9–12]. Thus, the progression to full viral fusion that culminates in the release of nucleocapsid into the cytoplasm is largely dependent on local conditions.
The viral envelope glycoproteins responsible for mediating membrane fusion are targets for neutralizing antibodies and virus entry inhibitors. In addition, new innate restriction factors inhibiting virus fusion have been discovered in recent years [13–15]. Among these factors is the family of small interferon-induced transmembrane proteins (IFITMs) that exhibits broad-range of antiviral activity [15–17]. This family includes IFITM1, which localizes predominantly at the plasma membrane, as well as IFITM2 and IFITM3, which contain an endocytic signal in their cytoplasmic N-terminal domain and thus localize to late endosomal and lysosomal membranes [18–21]. IFITMs effectively block entry of many unrelated enveloped viruses, including orthomyxoviruses (influenza A virus, IAV), paramyxoviruses (Respiratory Syncytial Virus, RSV), flaviviruses (Dengue, West Nile), filoviruses (Marburg, Ebola), and coronaviruses (SARS) [15–17, 20, 22–25]. IFITM3 alone is responsible for the bulk of antiviral effects of interferon in cell culture [15]. Importantly, mice lacking the ifitm3 gene more readily succumb to IAV and RSV infection than control mice [26, 27]. There are, however, viruses that are resistant to IFITM-mediated restriction. Murine Leukemia Virus (MLV), Old and New World arenaviruses (Lassa Virus and Junin Virus, respectively), as well as several enveloped DNA viruses, are not affected by IFITMs [15, 28, 29].
The mechanism by which IFITMs inhibit fusion of most viruses, while sparing others, is not understood. We and others have shown that IFITM expression does not elevate the overall endosomal pH [15–19, 22, 30, 31] and, thus, should not block acid-triggered refolding of viral fusion proteins that initiate membrane fusion. Clues regarding the antiviral mechanisms of IFITMs come from their subcellular distribution which tend to correlate with IFITM’s potency against different viruses. IFITM2 and -3 better restrict viruses entering from late endosomes, while IFITM1 tends to be more effective against viruses that are thought to fuse with the plasma membrane or with early endosomes (reviewed in [17]). Indeed, expression of an IFITM3 mutant that redistributes the late endosome/lysosome-resident protein to the cell surface abolishes antiviral activity against IAV [32]. There are, however, exceptions to this rule. The fact that IFITM1 outperforms IFITM3 in restricting EBOV fusion [25] highlights the importance of cellular trafficking, as opposed to the steady state distribution, for antiviral activity. Also, a relatively weak IAV restriction exhibited by an IFITM1 chimera containing the N-terminal domain of IFITM3 that localizes to late endosomes suggests a role for other factors in addition to appropriate subcellular localization [21].
The most popular view of the mechanism of IFITM’s antiviral activity is that these proteins create “tough membranes” that are not conducive to fusion [17, 18, 22]. Two principal models for membrane stiffening by IFITMs have been proposed–a direct effect on the membrane in the immediate proximity of these proteins [19, 25, 33–35] that could involve changing the membrane fluidity and/or curvature [22, 33, 35], and an indirect effect through altering the lipid composition of endosomes [18]. Several lines of evidence support the proximity-based antiviral activity of IFITMs. First, as discussed above, there is a general correlation between the subcellular localization of IFITMs and their potency against viruses entering from distinct cellular compartments (reviewed in [17]). Second, IFITM3-mediated restriction, but not restriction by the plasma membrane-resident IFITM1, can be bypassed by forcing virus fusion with the plasma membrane [25, 30]. Third, IFITM incorporation into the viral membrane effectively inhibits fusion/infectivity [34, 36–38]. On the other hand, IFITM3 has been reported to bind to and inhibit the function of vesicle-associated membrane protein-associated protein A (VAPA) [18], the master regulator of endosome-ER lipid transport. While this model has been disputed by several groups [30, 35], a recent study provided evidence for the antiviral effect of cholesterol accumulation in late endosomes/lysosomes and confirmed accumulation of cholesterol in these compartments upon IFITM3 expression [39]. It thus remains unclear whether IFITMs must be present at the sites of virus fusion to block virus entry or affect fusion indirectly, by dysregulating lipid transport or metabolism.
We have previously shown that IFITM3 does not restrict the lipid-mixing (hemifusion) stage of viral fusion, but rather inhibits the formation of a fusion pore [30]. However, the inability to directly visualize IFITM3 in the context of virus entry into live cells precluded us from assessing whether this factor blocks fusion through a proximity-based mechanism. Here, we overcame this limitation by constructing a functional fluorescent IFITM3 protein and imaging virus co-trafficking and fusion with endosomal compartments in cells expressing this protein. Comparison of entry and fusion of IFITM3-sensitive (IAV) and–resistant (LASV) viruses by 3-color live cell imaging revealed that IAV enters into and remains trapped within endosomes enriched in fluorescent IFITM3, where viruses underwent hemifusion but failed to complete the fusion reaction. In contrast, LASV particles entered and fused with endosomes devoid of IFITM3, implying that LASV escapes restriction by utilizing an endocytic pathway distinct from that employed by IAV. Collectively, our results provide strong support to a proximity model by which presence of IFITM3 at the preferred sites of virus entry restricts viral fusion.
To assess the co-distribution of viruses with IFITM3 at the time of fusion in live cells, we generated fluorescently-tagged IFITM3 protein. Linear N- or C-terminal fusions of EGFP (or similar fluorophores) with IFITMs render IFITMs nonfunctional. However, we found that the coding sequence of EGFP inserted into the N-terminal region of IFITM3, predicted to reside in the cytoplasm [40–42], generates a functional protein (Fig 1). First, to verify correct subcellular distribution of the fluorescent construct, we co-expressed IFITM3-iEGFP (iEGFP stands for “internal” EGFP) with an N-terminally myc-tagged IFITM3 in HeLa cells and confirmed their colocalization by fixing and immunostaining the cells (Fig 1B). Extensive colocalization between IFITM3-iEGFP and myc-IFITM3 suggests that subcellular localization of IFITM3 is not perturbed by incorporation of an EGFP tag. To test the functionality of the fluorescent construct, control A549 cells transduced with an empty vector (Vector) and cells transduced with IFITM3-iEGFP were infected with varied doses of influenza A/WSN/33 virus and the resulting infection measured by immunostaining cells for HA antigen. A549 cells were selected because they express very low endogenous levels of IFITM3 [15, 30, 31]. Cells expressing the EGFP-labeled IFITM3 were consistently much more resistant to influenza infection than control cells (Fig 1C). Microscopic analysis revealed that cells expressing intermediate to high levels of IFITM3-iEGFP did not stain for HA antigen (Fig 1D), demonstrating that IFITM3-iEGFP protects cells from IAV infection, similar to unlabeled IFITM3 [15, 30].
To facilitate multi-color live cell imaging of IFITM3 with fluorescently labeled viruses, we replaced the internal EGFP tag with cyan mTFP1 or bright green mNeonGreen protein. We next examined the ability of fluorescent IFITM3 constructs to inhibit IAV fusion using HIV-1 particles pseudotyped with influenza HA and NA proteins from the H1N1 A/WSN/33 strain (designated as IAVpp) and carrying the β-lactamase-Vpr (BlaM-Vpr) chimera, as described in [30]. A549 cells transduced with an empty vector, unlabeled IFITM3, IFITM3-imNG, or IFITM3-imTFP1 were inoculated with IAVpp, and the extent of viral fusion was measured after 2 h at 37°C based on the resulting cytosolic BlaM activity [43, 44]. Compared to the Vector control, IFITM3 severely restricts IAVpp fusion, as shown previously [30]. IFITM3-imNG and IFITM3-imTFP1 proteins were expressed in A549 cells at levels comparable to that of untagged IFITM3, as determined by Western blot analysis of respective cell lysates (Fig 1F), and also potently inhibited IAVpp fusion (Fig 1E). Comparable expression of the two fluorescent constructs is further supported by live cell fluorescence microscopy analysis (S1 Fig). Taken together, our results demonstrate both the appropriate subcellular localization and antiviral activity of fluorescently labeled IFITM3 constructs.
Lipid mixing between two membranes is a necessary but not sufficient condition for complete fusion; lipids can diffuse through a hemifusion intermediate, without opening of a fusion pore (e.g., [11, 30, 45, 46]). We have previously shown that IFITM3 expression does not inhibit lipid mixing (hemifusion) between IAV and host endosomal membranes, but rather interferes with the formation of fusion pores [30]. The ability to visualize the distribution of functional fluorescent IFITM3 in living cells enables spatiotemporal analysis of viral fusion restriction. To test lipid mixing activity in the context of endosomes containing fluorescent IFITM3, we co-labeled infectious IAV (A/PR/8/34 H1N1) with a self-quenching concentration of the lipophilic dye SP-DiI18 and with the amine-reactive far-red Alexa Fluor647 dye (AF647) that labels surface glycoproteins of the virus [30]. As shown in the schematic diagram, single-virus hemifusion with an endosome can be detected by the appearance of bright SP-DiI18 spots resulting from dilution of this dye within an endosomal membrane (Fig 2A). Double-labeled IAV were incubated with cells at 4°C for 20 min, followed by the addition of pre-warmed Live Cell Imaging Buffer (LCIB), and imaging continued at 37°C. Representative snapshots from the time-lapse movie illustrate the single IAV SP-DiI18 (colored green) dequenching event around 25 min post-infection of Vector cells, indicating redistribution of the dye into an endosomal membrane (Fig 2B, S1 Movie). Fluorescent traces obtained by single IAV tracking show an increase in SP-DiI18 intensity over time, whereas the reference AF647 signal (red) remains relatively constant (Fig 2C). From these traces, the time required for complete dequenching (Δt) and the extent of dequenching (ratio of the initial and final mean intensities If/Ii) can be determined (Fig 2C).
To evaluate whether lipid mixing between IAV and IFITM3-containing compartments occurs, AF647 and SP-DiI18 labeled viruses were incubated with A549 cells expressing IFITM3-imNG, and virus entry/fusion monitored by three-color live cell microscopy (Fig 2D). As expected, fluorescent IFITM3 primarily localized to endosomes that exhibited retrograde and anterograde movement in living cells. The dynamics of IFITM3-imTFP1 intracellular transport is illustrated by S2 Movie. Snapshots of a single IAV show the hemifusion event that occurs within an IFITM3-imNG endosome (IFITM3), with the onset of SP-DiI18 dequenching detected at around 32 min (Fig 2E, S3 Movie). The overall kinetics of onset of lipid mixing was not affected by IFITM3 expression (S2A Fig). Fluorescence traces with levels of fluorescence intensity from IFITM3 (blue), SP-Dil18 (green), and AF647 (red) are shown with the time to SP-DiI dequenching (Δt) and dequenching ratio of the final and initial mean intensities (Fig 2F). It is clear that IAV HA-mediated lipid mixing is not inhibited by accumulation of IFITM3 in endosomes.
We next asked whether the presence of fluorescent IFITM3 affects the rate or extent of lipid exchange between IAV and the limiting membrane of a vesicle. The fold of SP-DiI18 dequenching was measured by calculating mean ratios of single IAV SP-DiI18 signals after and before dequenching (If/Ii) in A549 Vector cells and in A549-IFITM3-imNG cells occurring within IFITM3+ punctae (as in Fig 2F). The extent of SP-DiI18 dequenching was independent of virus colocalization with IFITM3-imNG endosomes at the time of lipid mixing (Fig 2G). Since the extent of dequenching of lipophilic dyes is proportional to their fold-dilution (e.g., [47, 48]), this result indicates that the average size of recipient endosomes is the same in control and IFITM3-expressing cells. However, the SP-DiI18 dequenching time (Δt) was significantly longer in IFITM3+ endosomes compared to Vector cells (Fig 2H), consistent with a hemifusion connection that is more restrictive for lipid diffusion in IFITM3+ endosomes compared to control cells. The slower redistribution of SP-DiI18 to endosomes enriched in IFITM3-imTFP1 is consistent with our previous conclusion that these events represent IAV hemifusion but not full fusion [30]. Indeed, lipid diffusion through both leaflets of a fusion pore is expected to be faster than diffusion through contacting leaflets of a hemifusion intermediate [49]. In another example, IAV SP-DiI18 dequenching can also occur with a bi-phasic increase in intensity, suggesting either transient fusion pore closure or representing the transition from a restrictive hemifusion structure that attenuates lipid diffusion to a fusion pore (S2 Fig).
To visualize single IAV fusion in A549 cells, we pseudotyped the HIV-1 core with H1N1 HA and NA glycoproteins, as previously described [30]. IAVpp was labeled with a bi-functional mCherry-2xCL-YFP-Vpr construct, with a 2xCL tandem cleavage site for the viral protease that is cleaved during virus maturation, generating a free mCherry and a core-associated YFP-Vpr [50]. Virus fusion is detected based upon the release of mCherry into the cytoplasm through a fusion pore, while the YFP-Vpr marker remains in the viral core (Fig 3A). Labeled IAVpp were bound to A549 Vector cells in the cold by spinoculation (see Methods). Virus fusion was synchronously initiated by adding pre-warmed LCIB and visualized by two-color live cell microscopy for 2 hours with mCherry and YFP signals acquired every 6 seconds. Single IAVpp entered and fused with A549 cells, in agreement with our published data [30]. Time series images show the initial trafficking of a representative mCherry/YFP-Vpr labeled virus until fusion occurs, as evidenced by the loss of mCherry fluorescence, while the YFP-Vpr signal remains relatively constant (Fig 3B and 3C, S4 Movie).
Considering that the overwhelming majority of the IAV lipid mixing events occur in IFITM3-containing endosomes (Fig 2) and based upon our previous observation that IFITM3 expression inhibits single IAVpp fusion [30], we hypothesized that a sufficiently high local IFITM3 concentration is required for restriction of IAV fusion. To test this hypothesis, we synchronized IAVpp entry into A549-IFITM3-imTFP1 cells, as described above, and monitored mCherry/YFP-Vpr-labeled viral particles using three-color live cell imaging. Time series images show the IAVpp entering the IFITM3+ compartment and co-trafficking without undergoing fusion (Fig 3D). The fluorescence traces obtained by single virus tracking confirm entry (indicated by the red arrow) into an IFITM3+ endosome. At a later time, the virus-carrying endosome encounters and merges with another IFITM3+ compartment, as indicated by the second red arrow around 36 min (Fig 3E, S5 Movie). This pseudovirus did not fuse (release mCherry) for as long as the time-lapse imaging was performed. Co-trafficking of IAVpp with an IFITM3+ compartment can be visualized by examining 3D trajectories of the virus and relevant endosomes, which shows IAVpp co-trafficking with the first and then the second endosomal compartments containing IFITM3-imTFP1 (Fig 3F). Importantly, analysis of 6309 particles did not reveal a single viral fusion event occurring after extensive IAVpp co-trafficking with an IFITM3+ compartment.
In addition to tracking IAV particles that co-traffic with IFITM3+ compartments for an extended period of time, transient encounters with IFITM3+ compartments that did not inhibit subsequent IAVpp fusion were also observed. An example trace of an IAV pseudovirus shows a brief (~40 sec) apparent co-localization with an IFITM3+ compartment at around 23 minutes (Fig 3G and 3H, S6 Movie). Examination of 3D trajectories demonstrate that the particle does not significantly colocalize/co-traffic with the IFITM3+ endosome and that fusion occurs at around 31 minutes with no above-background IFITM3-imTFP1 signal (Fig 3H and 3I). Another example of false colocalization of IAVpp with an IFITM3+ endosome is shown in S3 Fig and S7 Movie. The viral particle appears to transiently colocalize with an IFITM3+ endosome when visualized in 2D, but particle tracking performed in 3D shows that the viral particle and the IFITM3+ endosome traffic in different Z-planes. These observations suggest that a transient and chance encounter of a virus carrying endosome with an IFITM3+ endosome is not sufficient to restrict fusion. In contrast, fusion is blocked after sustained and prolonged IAVpp co-trafficking with an IFITM3+ endosome, which strongly implies that the virus is being carried by an IFITM3-enriched compartment.
Further analysis of time-lapse acquisitions performed in at least 15 independent experiments shows that, on average, 2.2% of IAVpp fuse in Vector cells and none of the 6309 analyzed particles in IFITM3+ endosomes underwent fusion (Fig 3J). Our results thus demonstrate, for the first time, that the presence of IFITM3 in the endosomes carrying the virus is key to restriction of IAV fusion. The unimpeded lipid mixing between IAV and IFITM3+ endosomes, together with the lack of viral content release, strongly imply that IFITM3 traps the IAV fusion at a hemifusion stage by blocking the formation of a small fusion pore (in agreement with our previous study [30]).
Inhibition of IAVpp fusion after sustained co-trafficking with IFITM3-imTFP1-enriched endosomes may occur through a direct block of viral fusion by the restriction factor. Alternatively, IFITM3 may indirectly interfere with IAV fusion by altering the properties of endosomes, such as the luminal pH. The overall acidity of IFITM3+ endosomes was assessed by loading A549-IFITM3-imTFP1 cells with the acidic compartment marker, LysoTracker Red (S4 Fig). Imaging of fixed cells shows strong cytoplasmic colocalization between IFITM3+ endosomes and LysoTracker Red positive compartments, with only a small fraction of peripheral IFITM3+ endosomes lacking a detectable LysoTracker Red signal (S4A Fig, Inset). Analysis of multiple fields of view confirms that most IFITM3+ endosomes, with the exception of a small number of peripheral endosomes, accumulate the lysosomal marker (S4B Fig). This finding is consistent with progressive acidification of early IFITM3-containing endosomes through a maturation process and thus supports the notion that virus-carrying late IFITM3+ endosomes are acidic.
To further test whether IFITM3+ compartments are otherwise permissive for viral fusion, we sought to render IFITM3-imTFP1 inactive by pretreating cells with Amphotericin B (AmphoB), which is known to antagonize the antiviral activity of IFITM3 ([35] and Fig 4A). We also used the inactive oligomerization-defective IFITM3 mutant, with alanine substitutions at F75 and F78 (denoted 2M) [21]. As expected, the 2M-IFITM3-imTFP1 mutant did not inhibit IAVpp fusion (Fig 1E), in spite of being expressed at a level comparable to IFITM3 and IFITM3-imTFP1 (Fig 1F).
We next probed the ability of single IAVpp to fuse with IFITM3+ compartments under conditions that rescue bulk IAVpp fusion. A549-IFITM3-imTFP1 cells were infected with IAVpp labeled with mCherry-2xCL-YFP-Vpr, as above, in the presence of 1 μM AmphoB. As seen with a bulk fusion assay (Fig 4A), single IAVpp fuses with IFITM3+ endosomes in the presence of AmphoB. Representative single virus images show the entry and subsequent co-trafficking of an IAV particle with an IFITM3+ compartment and fusion within the compartment, as indicated by the sudden loss of mCherry (Fig 4B and 4C, S8 Movie). Of the 23 total fusion events that occur in IFITM3-imTFP1 cells treated with AmphoB, 7 particles co-traffic and fuse with IFITM3+ compartments (Fig 4G). In contrast to IFITM3-imTFP1 expressing cells in the presence of AmphoB, IAVpp exclusively fused at sites devoid of the mutant IFITM3 in 2M-IFITM3-imTFP1 cells (Fig 4D and 4E, S9 Movie). None of the IAVpp fusion events of the total 4442 particles annotated in 2M-IFITM3 cells co-traffic with IFITM3+ compartments. Fig 4D and 4E illustrates this phenomenon, whereby an IAVpp particle fuses within a 2M-IFITM3-imTFP1 cell but does not co-traffic with appreciable local 2M-imTFP1 maxima. These results suggest that loss of antiviral activity of the F75/78A IFITM3 mutant may be due to its altered subcellular distribution that prevents co-trafficking with IAV. This is in contrast to AmphoB, which renders wild-type IFITM3-imTFP1 inactive without affecting its trafficking pathways. Of note, both conditions that rescued the IAVpp fusion delayed the fusion kinetics relative to untreated cells expressing IFITM3-imTFP1 (S5 Fig), indicating a global effect on the rate of virus endocytosis and entry into permissive compartments. Together, the above results support the notion that IFITM3 inhibits IAV fusion through a proximity-based mechanism–by co-trafficking with the virus and accumulating in compartments that are otherwise permissive for IAV fusion.
To visualize single LASV entry and fusion, which is not restricted by IFITM3 in A549 cells [15], we pseudotyped the HIV-1 core containing the bi-functional mCherry-2xCL-YFP-Vpr marker with the LASV GPc envelope glycoprotein complex to generate LASV pseudoparticles (LASVpp). LASVpp imaging in A549 cells confirmed the ability to track single particles and detect their fusion (release of mCherry) in late endosomal compartments (Fig 5B, S10 Movie). Interestingly, LASVpp fusion exhibited a unique feature rarely seen for other viruses, including IAV. The YFP-Vpr fluorescence, which is markedly decreased at mildly acidic pH [51, 52], was consistently quenched at some point prior to viral fusion, demonstrating acidification of intraviral pH [53, 54] (schematized in Fig 5A). Single frame images show that YFP-Vpr signal quenched for ~10 min before viral fusion, which is observed as the loss of mCherry signal (red) and concomitant reappearance of the YFP-Vpr signal (Fig 5B and 5C, arrow). The dequenching of YFP fluorescence can be attributed to the re-neutralization of the virus’ interior through a fusion pore connecting it to the cytoplasm [52, 54, 55]. Based on the differences in IAVpp and LASVpp fusion, we classified single virus fusion events into “Type I”, in which mCherry signal is lost without acidification of the virus interior (YFP quenching), as observed in IAV fusion, and “Type II” events, in which acidification of the virus interior occurs prior to fusion (mCherry release), as observed in LASV fusion. In A549 cells, only 9% of LASVpp fusion events are Type I, while most particles—91%—undergo Type II fusion (Fig 5D). Of note, YFP-Vpr quenching occurs for most particles not undergoing fusion at later times after infection due to a non-specific acidification of the viral interior in late acidic compartments. These events representing a non-productive entry of LASVpp were excluded from analysis.
Additional experiments to confirm single LASVpp fusion in A549 cells were also performed. Control cells were treated with a broad-spectrum arenavirus entry inhibitor, ST-193 [56], which abrogated single LASVpp fusion events (Fig 5D). A total of 90 Type II events were observed in at least 24 independent experiments, with 6430 viral particles in Vector control cells and 5264 viral particles observed in cells treated with ST-193. LASVpp fusion with A549 cells measured by a bulk BlaM assay also demonstrated potent inhibition of viral fusion in the presence of 10 μM ST-193 or upon raising the endosomal pH by 40 mM NH4Cl (Fig 5E). Thus, the observed changes in fluorescent signals faithfully represent single LASVpp fusion.
Consistent with the lag between YFP quenching and fusion (Fig 5B and 5C), the kinetics of LASVpp fusion lagged behind the YFP quenching events (Fig 5F). The average lag time between YFP-Vpr quenching and fusion for LASVpp in A549 cells is 14.1 minutes (Fig 5F, inset). To test if the observed lag was due to the requirement for further virus trafficking to fusion-permissive compartments, we asked if it depended on how long a virus trafficked prior to YFP-quenching (Fig 5G). The lag between quenching and fusion does not appear to be correlated with the waiting time for quenching (R2 = 0.0263), suggesting that LASVpp fusion following the YFP quenching is a stochastic event that does not depend on the virus trafficking history. Although acidification of the virus interior does not directly report the time of acidification of endosomal lumen, the above results demonstrate that LASV GPc retains fusion-competence under acidic conditions for a considerable time before it fuses with permissive late endosomes, perhaps after binding to LAMP1 [57, 58].
We next assessed the basis for LASV resistance to IFITM3 restriction. A549-IFITM3-imTFP1 cells were infected with LASVpp labeled with mCherry-2xCL-YFP-Vpr, as above. Single particle imaging revealed that LASVpp did not co-traffic with IFITM3-imTFP1-positive endosomes and that subsequent viral fusion occurred at sites devoid of this fluorescent restriction factor (Fig 6A, S11 Movie). Analysis of single LASVpp fluorescence intensities in A549-IFITM3-imTFP1 cells shows a typical Type II fusion event which occurs within an endosome lacking above-background amounts of IFITM3-imTFP1 (Fig 6B).
Live cell imaging experiments were performed at least 6 times independently, and on average, 1.64% and 1.87% of double-labeled LASVpp particles bound to cells fused in A549 Vector and A549-IFITM3-imTFP1 cells, respectively (p = 0.551) (Fig 6C). These data confirm previous reports that the expression of IFITM3 does not affect LASV fusion [15, 30]. In addition, the kinetics of LASVpp fusion in control and A549-IFITM3-imTFP1 cells were not significantly different (S6A Fig). LASVpp fusion kinetics were the same regardless of IFITM3-imTFP1 expression, as observed using the BlaM assay and stopping fusion at varied times by NH4Cl addition (S6B Fig).
We note that in one or two rare examples, LASVpp fusion appears to occur in an endosome containing detectable IFITM3-imTFP1 signal. Representative images and fluorescence traces show co-trafficking of a LASV particle within an IFITM3+ endosome until fusion occurs around 14 min post-infection (Fig 6E and 6F and Inset, S12 Movie). This unique event is atypical of the majority of tracked particles due to several reasons: (1) there is an apparent colocalization with IFITM3+ beginning at time 0; and (2) LASVpp rarely fuse as early as 14 min post-infection. Most importantly, the fusion event in Fig 6E and 6F appears to represent transient fusion pore opening, as indicated by the sudden re-quenching of YFP, or re-acidification of the virus interior following pore closure (Fig 6E–6G). We report this instance to illustrate that, while LASVpp typically avoids IFITM3+ endosomes, miniscule levels of fusion may occur within IFITM3+ compartments. Overall, LASVpp exhibit a strong tendency to bypass IFITM3+ endosomes and this important feature likely represents the mechanism by which this virus escapes restriction.
Analysis of single IAVpp and LASVpp co-trafficking with IFITM3-imTFP1 endosomes in the context of fusion (Figs 3–6) suggests that IAVpp would be trapped in IFITM3-positive endosomes/multivesicular bodies, while LASVpp would not. To test this notion, we followed the bulk virus uptake and transport in live A549 cells expressing IFITM3-imNG. Cells were incubated in the cold for 1.5 hr with IAVpp or LASVpp labeled with an internal fluorescent marker, Gag-mCherry, to allow virus binding. Cells were then incubated for indicated times (0, 15, 30, and 60 min) at 37°C, fixed and imaged at high spatial resolution. Representative images of IAVpp and LASVpp co-localization with IFITM3 containing vesicles at different time points are shown in Fig 7A. Quantification of virus colocalization with IFITM3-imNG over time shows that IAVpp increasingly co-localizes with IFITM3-imNG compartments, while LASVpp does not show a significant increase in colocalization up to 1 hr post-infection (Fig 7B). These data support the hypothesis that restriction-sensitive viruses (as is the case for IAV) co-traffic with IFITM3, while resistant viruses (like LASV) are transported through distinct endosomal compartments devoid of this restriction factor.
To further test whether the presence of IFITM3 is necessary and sufficient to restrict IAV fusion, we generated control IAV particles that incorporated IFITM3 through co-expression in virus-producing cells. These pseudoviruses contained BlaM-Vpr to assess their fusion-competence. IFITM3 incorporation into virions and its possible effects on HIV-1 maturation or the influenza HA incorporation into pseudoviruses were verified by Western blotting (Fig 8A). IFITM3 was present in pseudoviruses prepared in producer cells expressing IFITM3 but not in the Vector control. Furthermore, the amount of p24 protein and influenza HA were the same in the two preparations (Fig 8A), indicating that IFITM3 incorporation does not perturb the expression or proteolytic processing of HA.
To probe the fusion activity of IFITM3-containing pseudoviruses, A549 cells were incubated with IAVpp/IFITM3 (or control viruses lacking IFITM3) at 4°C for 30 min, followed by incubation for 2 hours at 37°C in either drug-free medium or medium supplemented with AmphoB (which rescues IAV fusion in A549-IFITM3-imTFP1 cells, Fig 4A and [35]). Compared to the control IAVpp, fusion of IAVpp containing IFITM3 was potently inhibited (p<0.001, Fig 8B). Interestingly, AmphoB rescued IAVpp/IFITM3 fusion (Fig 8B), suggesting a direct effect of this antibiotic on IFITM3 or the viral membrane that is independent of cellular processes, including endocytic transport. The diminished ability of IAVpp produced in the presence of IFITM3 to fuse with target cells was not caused by IFITM3-containing extracellular vesicles present in the viral preparations, as have been suggested in [59]. Viral fusion was not significantly diminished when A549 cells were infected with a mixture of control IAVpp and extracellular medium from cells transfected only with an IFITM3-expressing plasmid (S7A Fig). Although IFITM3-containig extracellular vesicles were effectively concentrated by our virus concentration protocol using LentiX (S7B Fig), these vesicles did not modulate the fusion activity of IAVpp under our experimental conditions involving a brief exposure of target cells to the virus and vesicles. Taken together, these results suggest that the presence of IFITM3 is necessary and sufficient to restrict IAV fusion, irrespective of whether or not IFITM3 is expressed in the target or viral membrane.
Further evidence supporting the proximity-based antiviral activity of IFITM3 was obtained by measuring LASV GPc-mediated cell-cell fusion. In this model, cell fusion is triggered by exposure to low pH, which bypasses the need for endocytic trafficking that may sort LASVpp away from IFITM3+ compartments. Cos7 cells transiently expressing LASV GPc were brought in contact with 293T cells stably expressing IFITM1, IFITM2 or IFITM3 [22] or, in control experiments, with parental 293T cells. Cos7 and 293T cells were pre-loaded with different cytosolic fluorescent dyes to monitor fusion initiated by an acidic buffer, as previously described [22]. In stark contrast to LASVpp fusion with IFITM3-expressing cells (Fig 6), GPc-mediated cell-cell fusion was markedly inhibited by all three IFTIMs expressed in target cells (Fig 8C). This result supports the notion that LASV GPc is not inherently resistant to IFITM restriction and that the reason LASV is insensitive to IFITM3 expression is through its usage of trafficking pathways that are distinct from those used by IFITM3.
A remarkable breadth of enveloped viruses that are restricted by IFITM proteins suggests a universal mechanism for antiviral activity that likely involves altering the properties of the host cell membranes in a way that precludes viral fusion. It remains unknown how IFITMs exert their antiviral effects and, equally importantly, how arenaviruses and MLV escape restriction. In this study, we addressed a critical question of whether IFITMs work by a proximity mechanism, which requires their presence at the sites of virus entry and whether the lack of local IFITMs is a major determinant of virus resistance.
Through constructing a functional fluorescently tagged IFITM3 protein, we were able to visualize its dynamic distribution in living cells, in the context of single virus entry and fusion. Imaging experiments demonstrate that IAV restriction involves virus co-trafficking with IFITM3-containing endosomes that can culminate in lipid mixing (hemifusion) but does not progress to complete fusion (viral content release). This important finding, along with our previous work [17, 30, 35], strongly supports a proximity model for virus restriction, as opposed to alternative models that involve, for example, dysregulation of cholesterol transport from late endosomes [18, 39]. Also importantly, we documented the “avoidance” mechanism of LASVpp escape from IFITM3 restriction through virus trafficking and fusion with endosomes lacking this restriction factor. Consistently, LASV GPc-mediated cell-cell fusion is sensitive to IFITM proteins expressed on the surface of target cells. These results highlight the importance of regulation of IFITM trafficking for antiviral activity and offer important clues regarding the determinants of virus resistance to restriction. Of note, the presence of IFITM3 at the sites of IAV fusion does not rule out the possibility that the antiviral effect is due to recruitment of downstream effector proteins, such as ZMPSTE24 [60]. IFITMs have the propensity to hetero-oligomerize [21] and interact with a number of other proteins [61], so it is possible that IFITM-driven protein complexes alter the membrane properties and disfavor viral fusion (see below).
Single particle tracking revealed that IAV fusion was inhibited in compartments that accumulated substantial amounts of fluorescent IFITM3. Due to the relatively high and variable background fluorescence in cells expressing fluorescent IFITM3, it is difficult to quantitatively assess whether there is a threshold density of this protein below which viruses are not restricted. In other words, it is unclear whether inhibition of IAV fusion by IFITM3 occurs through an all-or-none mechanism or there is an inhibition “gradient” whereby the probability of fusion is inversely proportional to the IFITM3 signal. Future studies using improved fluorescence labeling techniques and controlled IFITM3 expression levels will help distinguish between these modes of action. The existence of distinct domains within the highly dynamic endosomal membranes (e.g., [62–66]) adds an additional layer of complexity when interpreting the IFITM3 restriction results. It is possible that the extremely rare single LASVpp fusion events that appear to colocalize with IFITM3 occur with IFITM3-free domains within the limiting membrane of an endosome.
We have previously documented unimpeded IAV lipid mixing activity in IFITM3-expressing cells [30]. Analyses of lipid dye dequenching, irrespective of colocalization with IFITM3, did not reveal significant differences in the rate or extent of lipid mixing. This finding is in disagreement with the reduced IAV lipid dequenching in IFITM3-expressing cells reported in [39]. The reason for discrepant results is likely related to the use of a bulk lipid dequenching assay in [39], as compared to the real-time single IAV tracking in our experiments. Importantly, in the present study, we were able to show the markedly slower rate of lipid redistribution to IFITM3-containing endosomes by focusing on events occurring in these compartments compared to lipid mixing in control cells. It should be noted that, in spite of exogenous incorporation of DiI into the viral membrane in the commonly used labeling protocol (e.g., [30, 67]), the dye readily redistributes to both membrane leaflets, as we have demonstrated previously [68]. Thus, a slower lipid mixing between IAV and IFITM3-positive endosomes is consistent with a more restricted dye diffusion through the merged contacting leaflets of hemifused membranes, as compared to diffusion through both leaflets of a fusion pore.
The exact mechanism by which IFITM3 inhibits the transition from hemifusion to fusion is not clear. A large body of work demonstrates a critical role of lipid composition, and specifically of mechanical properties of lipid membranes, in protein-mediated membrane fusion (reviewed in [69]). Bending energies of highly curved lipid intermediates that form and resolve during merger of lipid bilayers are key determinants of the fusion process (reviewed in [69–71]). In addition, hemifusion and the formation of a fusion pore within a hemifusion diaphragm are associated with changes in areas of contacting and distal monolayers, respectively. Thus, viral fusion pore opening could be blocked by: (1) increased membrane bending modulus; (2) increased negative curvature of the cytoplasmic leaflet that disfavors the formation of a net positive curvature fusion pore [69, 72]; (3) expansion of the hemifusion diaphragm to a size beyond that permissible for fusion pore formation [73]; or (4) reduced “fluidity” (lateral diffusion) of the cytoplasmic leaflet, which can be caused by IFITM homo/hetero-oligomerization [21]. The latter effect is expected to disfavor the fusion pore opening due to inability to quickly remove excess lipid from the hemifusion site. IFITMs have been reported to alter membrane fluidity [21, 35], and to increase the lipid order and confer positive spontaneous curvature [22, 33]. It is thus possible that individual effects of IFITMs on lipid membranes or their combination are responsible for the fusion block. Importantly, a recent study demonstrated that mutations in distinct regions of IFITM3 regulate its inhibitory vs enhancing activity against infection by different coronaviruses [74]. The ability to switch between inhibition and promotion of coronavirus fusion by introducing point mutations in IFITM3 further supports the proximity-based mechanism of virus restriction.
We have previously proposed an alternative mechanism of IFITM3-mediated virus restriction referred to as a “fusion decoy” model [30]. According to this model, viruses are redirected into multivesicular endosomes where unrestricted fusion with intraluminal vesicles, as opposed to fusion with the limiting membrane of an endosome, does not allow viral capsid release into the cytoplasm. The single virus content (mCherry) release assay would not detect IAVpp fusion with intraluminal vesicles, as the content marker will remain contained within the same endosome. The similar extent of lipid dye dequenching upon single IAV fusion with control and IFITM3-positive endosomes (Fig 2) appears compatible with virus hemifusion to the limiting membrane, but the slower dequenching rate in IFITM3 compartments could be due to multiple rounds of hemifusion with intraluminal vesicles. Therefore, additional experiments are needed to test the validity of a “fusion decoy” model.
Recent studies have documented the ability of IFITMs to interfere with viral fusion when incorporated into the viral membrane [34, 36–38]. In fact, IFITMs appear to more potently inhibit HIV-1 infection when incorporated into virions, as compared to their expression in target cells [38]. It is tempting to assume that the same mechanism of the IFITMs’ antiviral activity is functional in both cellular and viral membranes, but this notion has not been explicitly tested. The ability of IFITM3 to inhibit IAV fusion irrespective of whether it is expressed in the target or viral membrane (Fig 8B) supports the universal mechanism of IFITM3-mediated restriction that involves altering the properties of lipid membranes, as opposed to interacting with viral or cellular proteins. Our finding that AmphoB rescues the fusion-competence of IAVpp containing IFITM3, similar to its antagonistic effect on the cell-expressed IFITM3 [35], is also consistent with the common mechanism of virus restriction. Moreover, the static nature of the viral membrane, which is in stark contrast to the highly dynamic cell membranes, supports a direct effect of AmphoB on the viral membrane, perhaps through alterations of membrane fluidity [22, 35]. Thus, virions containing IFITMs in their membranes could provide a tractable model for mechanistic studies of these proteins.
Our study focused on IFITM3 protein, which shares a relatively high sequence homology and subcellular distribution with IFITM2. Although we have not addressed the mechanism of action of the plasma membrane-resident IFITM1, the published literature and our findings support the notion that this protein also acts by a proximity-based mechanism. We thus speculate that all members of the IFITM family accumulate at the sites of fusion of sensitive viruses and block the formation of a fusion pore. The current study provided strong evidence that LASV escapes IFITM3 restriction by entering through alternative endocytic pathways, but has not addressed whether other IFITM-resistant viruses, such as Junin virus or MLV, employ the same strategy to infect IFITM-expressing cells. Future studies addressing this question will help generalize the escape mechanism discovered in this work and may suggest strategies to increase the potency of IFITMs by modulating their intracellular trafficking.
We obtained HEK 293T/17, Cos7 and human lung epithelial A549 cells from ATCC (Manassas, VA). TZM-bl cells were obtained from NIH AIDS Research and Reference Reagent Program. 293T cells stably expressing IFITM1, IFITM2 or IFITM3 were a gift from Dr. Shan-Lu Liu, Ohio State University [22]). Cells were maintained in Dulbecco’s Modified Eagle Medium (DMEM, Cellgro, Mediatech, Masassas, VA) containing 10% heat-inactivated Fetal Bovine Serum (Hyclone Laboratories, Logan, UT or Atlanta Biologicals, Flowery Branch, GA) and 1% penicillin/streptomycin from Gemini Bio-products (West Sacramento, CA). For HEK 293T/17 cells the growth medium was supplemented with 0.5 mg/ml G418 (Genesee Scientific, San Diego, CA). DMEM without phenol red was purchased from Life Technologies (Grand Island, NY). Live Cell Imaging Buffer (LCIB) and FluorobriteTM DMEM were purchased from Life Technologies (Grand Island, NY). Stable cell lines expressing fluorescently-tagged IFITM3 used for imaging analysis were obtained by transducing with VSV-G pseudotyped viruses encoding wild-type or the 2M mutant (F75/78A) IFITM3 [21] or with the Vector pQCXIN (Clontech) and selecting with 1 mg/ml G418. Following selection, cells were maintained in 0.5 mg/ml G418. Stable cell lines expressing unlabeled and fluorescently-tagged IFITM3 used for bulk fusion assays were obtained by transducing with VSV-G pseudotyped viruses encoding wild-type IFITM3 with the Vector pQXCIP (Clontech) and selecting with 1.5 μg/ml puromycin.
The pR8ΔEnv, pR9ΔEnv, BlaM-Vpr, pcRev, pMDG-VSV-G, MLV-Gag-Pol, HIV-1 Gag-mCherry (encoding an uncleavable mCherry fluorescent tag used in fixed cell co-localization experiments), IFITM3/pQCXIP, F75/78A IFITM3/pQCXIN expression vectors were described previously [15, 21, 43, 75]. The mCherry-2xCL-YFP-Vpr (mCherry fused to YFP-Vpr through a cleavable linker containing two HIV protease cleavage sites– 2xCL), as described previously [50], was used for single particle tracking of fusion events in live cells. The pCAGGS vectors encoding influenza H1N1 WSN HA and NA were provided by Drs. Donna Tscherne and Peter Palese (Icahn School of Medicine, Mount Sinai) [76]. The LASV GPc plasmid was a gift from Dr. F.-L. Cossett (Université de Lyon, France) [77].
To internally label human IFITM3 (accession NM_021034) with EGFP, sites likely permissive to insertions were identified based on sequence alignments of human and mouse IFITM family proteins. An EGFP cassette flanked by two linkers was created by PCR (forward primer TCAAGGAGGAGCACGAGGTGGCTGTGCTGGGGGCGCCCCACAACCCTGCTCCCGGCGGAGGAAGCGGCGGAGTGAGCAAGGGCGAGGAGC; reverse primer GACGACATGGTCGGGCACGGAGGTCTCGCTGCGGATGTGGATCACGGTGGATCCGCCTCCGCTTCCGCCCTTGTACAGCTCGTCCATGCC) and inserted using Gibson assembly into KasI/BsaBI-cut IFITM3 cDNA. The resulting protein has amino acids 41 (P) and 42 (T) of wild-type IFITM3 removed. The cloning of IFITM3-imNeonGreen (IFITM3-imNG), and IFITM3-imTFP1 (IFITM3-imTFP1) into pQCXIN (Clontech) and pQXCIP (Clontech) retroviral expression vectors were done in two steps. In the first step, the EGFP in IFITM3-iEGFP/pLVX Tet on construct was replaced either with mNeonGreen or mTFP1 by overlapping PCR. The IFITM3-iEGFP/pLVX Tet on construct contain an EcoRI restriction site in the 5’ of the IFITM3-iEGFP cDNA and a BsrgGI in the 3’ of GFP cDNA that facilitated the replacement of GFP. The 5’ of IFITM3 cDNA was amplified by PCR using forward primer P1 (containing EcoRI restriction site) TACCACTTCCTACCCTCGTAAAGAATTCGCCACCATGAATCACACTGTCCAAACCTTC, and reverse primer P2: TGTGGTCTCCTCGCCCTTGCTCACTCCGCCGCTTCCTCCGCCGGGAGC. The mTFP1 fragment was amplified using forward primer P3: GCTCCCGGCGGAGGAAGCGGCGGAGTGAGCAAGGGCGAGGAGACCACA (complementary to P2), and the reverse primer P3 (containing BsrgGI restriction site) GATCCGCCTCCGCTTCCGCCCTTGTACAGCTCGTCCATGCCGTCGGTGGAATT. The fragments were purified, mixed and the overlapping PCR was perfomed using the forward primer P1 and the reverse primer P3. The final PCR fragment and IFITM3-iEGFP/pLVX Tet on were digested with EcoRI and BsrgGI restriction enzymes, purified and ligated. In the second step, the IFITM3-imTFP was amplified by PCR with forward primer P4 (containing AgeI restriction site) GCAGGAATTGATCCGCGGCCGCACCGGTAGGCCACCATGAATCACACTGTCCAAACCTTC, and reverse primer P5 (containing EcoRI restriction site) AGGGGTGGGGCGGGGGGGGGCGGAATTCTTAGTGATGGTGATGGTGATGGCCTTG, digested with AgeI and EcoRI restriction enzymes, purified and ligated into pQCXIN or pQXCIP vectors. For IFITM3-imNeonGreen construct the overlapping PCR was done using IFITM3-imTFP/pQCXIN construct as template and the AgeI and BamHI restriction sites. The 5’ of IFITM3 cDNA was amplified by PCR using forward primer P5 (containing AgeI restriction site) GCAGGAATTGATCCGCGGCCGCACCGGTAGGCCACCATGAATCACACTGTCCAAACCTT, and reverse primer P6: ATCCTCCTCGCCCTTGCTCACCATTCCGCCGCTTCCTCCGCCGGGAGC. The mNeonGreen cDNA was amplified with forward primer P7: GCTCCCGGCGGAGGAAGCGGCGGAATGGTGAGCAAGGGCGAGGAGGAT (complementary to P6), and reverse primer (containing BamHI restriction site) CGCTGCGGATGTGGATCACGGTGGATCCGCCTCCGCTTCCGCCCTTGTACAGCTCGTCCATGCCCA. Purified PCR fragments were mixed and the overlapping PCR was done using the forward primer containing AgeI restriction site and the reverse primer containing BamHI restriction site. The fragment and IFITM3-imTFP1 plasmid were digested with AgeI and BamHI, purified and ligated. The F75/F78A IFITM3-imTFP1 (2M-IFITM3-imTFP1) mutants was obtained by Quick-change site-directed mutagenesis (Stratagen, La Jolla, CA) using IFITM3-imTFP1/pQCXIN as template.
Alexa Fluor 647-NHS ester (AF647) and the lipophilic dye SP-DiIC18 (1,1'-Dioctadecyl-6,6'-Di(4-Sulfophenyl)-3,3,3',3'-Tetramethylindocarbocyanine) were purchased from Invitrogen/Life Technologies (Grand Island, NY). The LASV fusion inhibitor ST-193 was purchased from Aurum Pharmatech (Franklin Park, NJ). Amphotericin B (AmphoB) was obtained from Quality Biological (Gaithersburg, MD). Primary antibodies used were rabbit directed at the N-terminus of IFITM3 (Abgent, San Diego, CA), mouse anti-tubulin from Sigma (St. Louis, MO), HIV-1 IG serum (NIH AIDS Research and Reference Reagent Program), and rabbit anti-WSN Influenza R2376 (a generous gift from Dr. David Steinhauer, Emory University). Secondary antibodies used were rabbit anti-mouse IgG(H+L)-HRP (EMD Millipore), mouse anti-rabbit IgG(H+L)-HRP (EMD Millipore), and goat anti-human IgG-HRP (H+4) (Thermo Scientific).
Pseudoviruses were produced by transfecting HEK 293T/17 cells with JetPRIME transfection reagent (Polyplus-transfection, Illkirch-Graffenstaden, France). For all pseudovirus productions the transfection reagent/DNA containing medium was replaced with fresh phenol red-free medium after ~14 hrs. Viruses were harvested ~48 hrs post-transfection, and cellular debris were removed by centrifuging at 230xg for 10 min. The collected viruses were passed through a 0.45 μm polyethersulfone filter (PES, VWR) to further clear cellular debris and virus aggregates, aliquoted and stored at -80°C. The infectious titers (~106 IU/ml) were determined using serial dilutions of the inoculum in TZM-bl cells using a β-galactosidase assay. To produce the pseudoviruses for co-localization analysis, HEK293T/17 cells were grown to ~60–70% confluency in a 6-well culture dish and transfected with 0.8 μg pR8ΔEnv, 0.4 μg pcRev, 0.5 μg HIV-1 Gag-mCherry (Not cleaved by protease) and 0.8 μg GPc-Lassa or 0.4 μg each of WSN HA- and NA-expressing plasmids, respectively. For single viral fusion experiments in live cells, dual-labeled LASVpp was made by transfecting the HEK293T/17 cells grown to ~60–70% confluency in a 6-well culture dish using 0.8 μg pR9ΔEnv, 0.2 μg pcRev, 0.2 μg mCherry-2pxCLYFP-Vpr and 1 μg GPc-Lassa plasmids. Similarly, dual-labeled IAVpp were produced using 4 μg pR9ΔEnv, 1 μg pcRev, 1 μg mCherry-2xCL-YFP-Vpr and 2.5 μg each of WSN HA- and NA-expressing plasmids for transfection of ~60% confluent cells in a 100 mm dish. The purified viruses were diluted 10-fold in PBS without calcium or magnesium (PBS-/-, Cellgro, Mediatech), bound to poly-L-lysine coated 8-well chamber cover slips (LabTek, MA), and imaged to estimate the co-labeling efficiency which was over 90% for all pseudoviruses used for this study.
To generate VSV-G pseudotyped viruses encoding fluorescently-tagged IFITM3, HEK293T/17 cells grown in 6-well plate were transfected with 0.3 μg of VSV-G plasmid, 0.6 μg MLV-Gag-Pol plasmid, and 1.1 μg of either an empty pQXCIN or pQXCIP vector, or containing IFITM3-imNG, IFITM3-imTFP1, or 2M-IFITM3-imTFP1.
For intraviral IFITM3 pseudovirus production, HEK293T/17 cells grown in 100-mm dishes were transfected with 2 μg of pCAGGS H1N1 HA/NA, 3 μg pR9ΔEnv, 1.5 μg BlaM-Vpr, 0.5 μg pcRev, and 5 μg of either empty Vector pQCXIP, IFITM3, or 2M-IFITM3 using JetPRIME reagent. The viral supernatants cleared of cellular debris as described above, were concentrated 10x, using Lenti-X Concentrator (Clontech, Mountain View, CA). Following overnight concentration with Lenti-X, virus was precipitated by centrifuging at 1439xg for 45 min, 4°C, resuspended in DMEM without phenol red or FBS, and stored at -80°C.
For lipid mixing (hemifusion) experiments, influenza virus surface proteins and membrane were co-labeled with AF647 and with the lipophilic dye SP-DiIC18, respectively. Briefly, 100 μg of the purified IAV A/PR/8/34 virus (2 mg/ml, Charles River, CT) was mixed with 50 μM AF647 in 150 mM freshly prepared sodium bicarbonate buffer, pH 9.0. The labeling reaction was allowed to proceed at room temperature with tumbling in the dark for 30 min. Next, 5.8 μL of 1.75 mM SP-DiI18 was added to this reaction, while gently vortexing and viruses further incubated at room temperature in the dark for 1 hr with shaking. The AF647 was quenched by adding 2 μL of 1 M Tris-buffer, pH 7.0. The labeled viruses were purified from excess dyes on a Nap-5 gel filtration column (GE Healthcare) that was equilibrated with 50 mM HEPES, pH 7.4, 145 mM NaCl at room temperature. The fractions containing labeled viruses were passed through a 0.45 μm filter to remove any large lipid and/or virus aggregates. The purified viruses were bound to poly-L-lysine coverslips and imaged to quantify their co-labeling efficiency which was at least 55%, determined as the percentage of AF647 labeled viruses that showed detectable signal (under the high self-quenching concentrations) of SP-DiI18. The viruses were aliquoted into tubes, flash-frozen, and stored at -80°C until use.
The β-lactamase (BlaM) assay for virus-cell fusion were performed as described previously [30, 43]. Briefly, pseudoviruses containing a β-lactamase-Vpr chimera (BlaM-Vpr) were bound to target cells by centrifugation at 4°C for 30 min at 1550xg. Unbound viruses were removed by washing with DMEM without phenol red supplemented with 20 mM HEPES (GE Healthcare Life Sciences). Fusion was initiated by shifting to 37°oC for 2 hours, after which cells were placed on ice and loaded with the CCF4-AM substrate (Life Technologies) and incubated overnight at 11°C. The cytoplasmic BlaM activity (ratio of blue to green fluorescence) was measured using a SpectraMaxi3 fluorescence plate reader (Molecular Devices, Sunnyvale, CA).
The p24 content of viral stocks was determined by ELISA, as described previously [78]. Whole cell lysates were harvested in RIPA Buffer (Sigma) supplemented with protease inhibitors (Complete Protease Inhibitor Cocktail, Roche), incubated on ice for 10 min, and cleared by centrifugation at 16,000xg for 5 min. Total protein was measured using a bicinchoninic acid assay (BCA, Pierce) and normalized protein was loaded onto 4–15% polyacrylamide gels (Bio-Rad, Hercules, CA). Precision Plus Protein Standards (Kaleidoscope Bio-Rad) were used as molecular weight markers. Proteins were transferred onto a nitrocellulose membrane, blocked in 10% Blotting-grade Blocker (Bio-Rad) in PBS-T (phosphate buffered saline with 0.1% Tween-20) for 30 min at room temperature. Membranes were incubated in primary antibodies overnight at 4°C in 5% Blotting-grade Blocker with gentle shaking: rabbit anti-IFITM3 (1:500), mouse anti-tubulin (1:3000), human HIV-IG) (1:2000), and rabbit anti-WSN Influenza R2376 (1:100). After washing membranes with PBS-T at room temperature, Horseradish peroxidase-conjugated (HRP) goat anti-rabbit, rabbit anti-mouse, and goat anti-human secondary antibodies were added in 5% Blotting-grade Buffer for 1 h at room temperature with gentle shaking. Following PBS-T washing of membranes, ECL Prime chemiluminescence reagent (GE Healthcare) was used for protein detection.
The effector Cos7 cells were transfected with the Lassa virus GPc expression vector. Briefly, cells were grown on 35 mm culture dishes to ~60% confluency and transfected with 4 μg GPc expression vector using a calcium-phosphate protocol [22]. After 48 hours following transfection, cells were loaded with 1.3 μM of the green cytoplasmic Calcein-AM dye (Invitrogen). In parallel, 293T cells or their derivatives stably expressing human IFITM1, IFITM2 or IFITM3 [22] were labeled with 30 μM of the blue cytoplasmic dye CMAC (Invitrogen). Effector and target cells were washed, detached from the culture dishes using a non-enzymatic solution, resuspended in PBS++, mixed at a 1:1 ratio and co-plated onto 8-well chamber slides. After incubating for 30 min at room temperature, cells were exposed to a pH 5.0 buffer at 37°C for 20 min, and the resulting cell-cell fusion was measured by visual inspection under a fluorescent microscope, as described in [22]. Ten fields of view each containing 10–12 heterologous cell pairs were examined in each well.
A549 cells expressing IFITM3-imNG were cultured on collagen-coated 8-chamber coverslips (Lab-Tek, NY #1.5 glass) in Fluorobrite DMEM to ~60–80% confluency. The cells were chilled by placing on ice for 10 min, followed by aspirating the media and washing with cold phosphate buffered saline with calcium and magnesium (PBS+/+). Cells were inoculated with a 5-fold dilution of the mCherry-labeled IAV or LASV pseudoviruses in 100 μL of cold LCIB supplemented with 2% FBS, and viruses were allowed to bind to cells by incubation at 4°C for 90 min. Unbound viruses were removed by washing with cold PBS+/+, and virus entry was initiated by adding 200 μL pre-warmed (37°C) LCIB. The slides were incubated at 37°C for varied times followed by fixation with 4% paraformaldehyde (PFA) in PBS-/- for 10 min at 37°C. For the zero time-point, cells were fixed with PFA immediately following the initial virus binding step at 4°C. After fixation, PFA was washed away with PBS-/- a few times and cells were imaged.
For co-localization analysis, the fixed cells were imaged on DeltaVision Elite (GE Healthcare) widefield microscope, using an Olympus PlanApoN 60x/1.42 NA oil immersion objective. Multiple Z-stacks with a spacing of 0.1 μm covering the entire thickness of the cells were acquired using a standard GFP/Cherry filter set. Deconvolution of the Z-stacks was done post acquisition to improve the signal-to-background ratio of the IFITM3-mNeonGreen vesicles and of mCherry labeled viruses using SoftWorX (DeltaVision, GE Healthcare). The deconvolved Z-stacks were used for quantitative volume based (voxel based) co-localization analysis using a custom protocol in the image analysis program Volocity (Perkin Elmer, Waltham, MA). Briefly, after background subtraction, the IFITM3 containing vesicles were identified as objects. mCherry virus particles with a size/volume threshold of 0.512 μm3 (corresponding to a 2x2x2 voxel) and that were associated with IFITM3-expressing cells were identified. A virus particle was considered co-localized with IFITM3 when at least 50% of its volume overlapped within an IFITM3 expressing object. For every time point, at least 4 different fields of view containing multiple cells were imaged and the mean values of the % co-localization calculated.
A549 IFITM3-imNG, IFITM3-imTFP1, or 2MIFITM3-imTFP1 cells were seeded onto 35 mm collagen-coated glass-bottom Petri dishes (MatTek, MA) one day prior to imaging and cultured in DMEM without phenol red supplemented with 10% FBS, penicillin, and streptomycin. Following a wash with room temperature PBS, cells were fixed in 3.5% paraformaldehyde in PBS for 10 min. For imaging with LysoTracker™, cells were seeded as above and incubated with 30 nM LysoTracker™ Red DND-99 (ThermoFisher) diluted in pre-warmed LCIB supplemented with 2% FBS for 30 min prior to fixation. Images were acquired on a DeltaVision microscope using an Olympus UPlanFluo 40x/1.3 NA oil immersion objective (Olympus, Japan). Multiple Z-stacks with a spacing of 0.1 μm covering the entire thickness of the cells were acquired and deconvolved.
A549 cells were seeded onto 35 mm collagen-coated glass-bottom Petri dishes (MatTek, MA) one day before imaging and cultured in Fluorobrite DMEM supplemented with 10% FBS, penicillin, streptomycin and L-glutamine. Before imaging, the cells were pre-chilled on ice and washed with ice-cold PBS+/+. A small amount of viral suspension (~1 μL) diluted in 60 μL of cold LCIB was added to the cells and spinoculated at 1550xg at 4°C for 20 min. After spinoculation, the cells were washed twice with PBS +/+ to remove any unbound viruses and a small volume (~150 μL) of ice-cold LCIB was added to the cells. Viral entry was initiated by adding 2 mL of pre-warmed (37°C) LCIB containing 2% FBS, and cells were imaged immediately on DeltaVision microscope equipped with a temperature and humidity-controlled chamber. Every 6–8 sec, at least three Z-stacks spaced by 1.5–2 μm were acquired to cover the thickness of cells using Olympus 40x UPlanFluo 40x/1.3 NA oil objective (Olympus, Japan). The three-color viral fusion experiments were done with A549 cells expressing IFITM3-imTFP1 or 2M-imTFP1 using a standard CFP/YFP/Cherry filter set (Chroma, VT), while a TRITC/FITC/Cy-5 filter set was used for the three-color lipid mixing experiments with IFITM3-mNeonGreen expressing cells. LCIB in all the experiments was supplemented with 2% FBS.
The time-lapse Z-stack movies are visually inspected as maximum intensity projections, using ImageJ, and the ROI manager tool was used to annotate the single fusion or hemifusion events (observed as color change or DiI intensity increase). The sets of waiting times for hemi/fusion obtained from multiple movies are combined from experiments done on different days, sorted and plotted as cumulative probability curves that show the kinetics of the event. Acquired image series were converted to maximum intensity projections and annotated particles were tracked using either Volocity (GE Healthcare) or ICY image analysis software (icy.bioimageanalysis.org). Fluorescently labeled viral particles were identified using the spot detection algorithm and tracked in 3D to determine fluorescence intensities at every time point. With three-color imaging to track double-labeled viral particles that co-traffic with fluorescently-labeled IFITM3 compartments, single Z-planes in which the viral particle trafficked were used so that background subtraction could be performed using the ICY spot tracking algorithm. The local background was determined by dilating the identified objects corresponding to viral particles by two pixels. The difference between the integrated intensities of the particle and the dilated surrounding gave the intensity surrounding the particle, from which an average per pixel local background was calculated. This was used to obtain the background-corrected intensity of the particle at every time point, which is plotted in the time traces as shown in the figures.
For the lipid mixing experiments, SP-DiI18 dequenching curves for the individual viruses were obtained by tracking particles, either using the AF647 channel or the DiI channel. The dequenching ratios and times were manually obtained from time traces that show at least 4-fold increase in SP-DiI18 intensity. The dequenching ratio was calculated as ratio of the mean SP-DiI18 intensity before the rise of the signal and the mean intensity after completion of dequenching. The dequenching time was measured as time taken to reach an intensity plateau from the time of initial rise in intensity. The traces showing multi-phase increase in SP-DiI18 signal in IFITM3 expressing cells were excluded from the calculation of dequenching ratios and times.
|
10.1371/journal.pcbi.1000843 | Intergenic and Repeat Transcription in Human, Chimpanzee and Macaque Brains Measured by RNA-Seq | Transcription is the first step connecting genetic information with an organism's phenotype. While expression of annotated genes in the human brain has been characterized extensively, our knowledge about the scope and the conservation of transcripts located outside of the known genes' boundaries is limited. Here, we use high-throughput transcriptome sequencing (RNA-Seq) to characterize the total non-ribosomal transcriptome of human, chimpanzee, and rhesus macaque brain. In all species, only 20–28% of non-ribosomal transcripts correspond to annotated exons and 20–23% to introns. By contrast, transcripts originating within intronic and intergenic repetitive sequences constitute 40–48% of the total brain transcriptome. Notably, some repeat families show elevated transcription. In non-repetitive intergenic regions, we identify and characterize 1,093 distinct regions highly expressed in the human brain. These regions are conserved at the RNA expression level across primates studied and at the DNA sequence level across mammals. A large proportion of these transcripts (20%) represents 3′UTR extensions of known genes and may play roles in alternative microRNA-directed regulation. Finally, we show that while transcriptome divergence between species increases with evolutionary time, intergenic transcripts show more expression differences among species and exons show less. Our results show that many yet uncharacterized evolutionary conserved transcripts exist in the human brain. Some of these transcripts may play roles in transcriptional regulation and contribute to evolution of human-specific phenotypic traits.
| Phenotypic differences between closely related species, such as humans and chimpanzees, might be determined to a large extent by differences between their transcriptomes. Recent studies using microarray and high-throughput sequencing technologies have demonstrated that beside annotated genes, a large proportion of the human genome can be transcriptionally active. Little is known, however, about the extent and the conservation of human brain transcripts located outside of the known genes' boundaries. Here, we use high-throughput transcriptome sequencing to characterize the non-ribosomal transcriptome of the human cerebellum and compare it to the transcriptomes of chimpanzee and rhesus macaque. Our results show that close to 40% of all transcripts expressed in the human brain map within repetitive elements. By contrast, less then 10% of the human brain transcriptome corresponds to non-repetitive intergenic regions. Nonetheless, within these regions we identify more than a thousand novel highly transcribed evolutionary conserved locations. Some of the intergenic transcripts show distinct human-specific expression and may have contributed to evolution of human-specific phenotypic traits.
| Transcriptome studies conducted by various methodologies, such as conventional sequencing, tiling arrays, and, most recently, high-throughput sequencing, have consistently indicated that a large proportion of transcription takes place outside known gene boundaries (see [1], [2] and references therein). Among human tissues, the brain transcriptome is one of the most complex [3], [4]. Changes in expression of brain transcripts have been suggested to play an essential role in evolution of the human phenotype [5]. Indeed, expression of protein-coding genes differs greatly between humans and one of our closest relatives [6], [7], [8]. Furthermore, comprehensive analysis of approximately 1% of the human and chimpanzee brain transcriptomes using tiling arrays found multiple instances of differential expression outside annotated gene regions [9].
To systematically characterize the transcriptome in a particular brain region, cerebellar cortex, and identify its human-specific features, we performed high-throughput sequencing using the Illumina platform to analyze transcripts expressed in ten humans, four chimpanzees, and five rhesus macaques. All individuals are adult males (Table S1). While most previous studies [1], [2], [4], [10] have focused on the RNA fraction carrying poly(A) tails, we sequenced all transcripts present in the total RNA, excluding ribosomal RNA (rRNA) and depleting RNA transcripts shorter than 200 nucleotides (nt). Our experimental strategy is similar to the strategy used to characterize total transcriptome of HeLa cells [11], [12], with the difference that these studies either focused on the 3′-region of the transcript or sequenced mixture of polyA+ and polyA− transcripts with predominant part of polyA-enriches ones. To reduce within-species variation, we pooled the total RNA from brains of four to five individuals of the same species into one sample. To estimate technical as well as remaining within-species biological variation, we sequenced two independent human samples, each comprising total RNA from five individuals.
For each sample, we obtained an average of ∼10,000,000 sequence reads of 36 nt corresponding to ∼7,200,000 unique sequences. From these reads, we can map on average 51% to the corresponding reference genomes and annotated exon junctions (Table S2). Excluding the remaining sequences mapping to rRNA, we find that in humans 26% of the reads map to annotated exons and exon junctions, 2% - to mitochondrial genes, and less than 1% - to annotated non-coding RNA (ncRNA) (Figure 1A). Although these proportions are much greater than the corresponding genomic fractions (Figure 1A), they still represent less than a third of the total non-ribosomal human brain transcriptome. The remaining reads map within introns and intergenic regions (49% and 23% of the transcriptome, respectively). Such a distribution of transcriptome reads is not unique to humans, but shared among the three primate species studied (Figure S1).
Within intronic and intergenic regions, more than half of transcription originates from repetitive sequence elements, occupying in total ∼42% of the entire transcriptome (Figure 1A). This proportion is substantially greater than that reported in the human brain using cap-selected transcript tags (∼10%) [13]. For most of the repeat families, the expression is proportional to the genome fraction occupied (Figure 1B). Still, for some, such as simple and low complexity repeats, as well as repeat families derived from functional ncRNA, such as snRNA, snpRNA and 7SK RNA, the expression level is higher than expected from the repeat family size alone in all three species studied (Figure S2).
More than 90% of repeats present in the human genome result from transposable element (TE) activity taking place over hundreds of millions of years. Estimating the transcriptional activity of different TE families, we find that the most recently expanded ones, the Alu elements, show elevated transcriptional activity per genomic fraction occupied by the family (Figure 1C). The effect is more obvious when normalizing by the genomic fraction occupied by repeat elements actually expressed in brain (Figure 1D). We find the same effect in the other two species (Figure S3), indicating that elevated expression of certain Alu elements in brain might be widespread among primates.
Excluding repeats, intergenic regions contain 7% of all non-ribosomal human brain transcriptome sequences. These sequences are not distributed evenly, but concentrate within distinct regions (Figure 2A, 2B). Notably, the expression levels of such intergenic highly transcribed regions (igHTR) are comparable and, frequently exceed the expression levels of annotated exons (Figure 2C). We used two parameters to define igHTR: the maximum spacing between two neighboring reads and the minimum number of mapped sequence reads within the genomic regions. For convenience, we set these parameters to 150 nt and 10 reads for most of the analysis. In the two human samples, we find 883 and 790 of such highly transcribed intergenic regions (igHTR) not overlapping with any annotated human transcripts (Materials and Methods, Table S3). Out of these igHTR, 580 (66% and 73% for the two human samples) overlap between the samples, with the majority of igHTR overlapping by more than 80% of their length (Figure S4), while less than 1% would be expected to overlap by chance (simulation, p<0.01). Further, for all 1,093 igHTR identified in at least one of the two human samples, the expression levels correlated well between the samples (Spearman correlation, rho = 0.90, p<10−15) (Figure S5), even when the corresponding region did not pass the igHTR definition cutoff in one of the samples. Using different igHTR definition cutoffs, we get principally the same results throughout the analysis (e.g. Figure S6). Finally, using human brain expressed sequence tag (EST) libraries, we find further support for 48% of 1,093 igHTR found in at least one of the two human samples, significantly more than expected by chance (simulation, p<0.01) (Figure 2B).
Similar to humans, we can identify igHTR in chimpanzee and rhesus macaque brain transcriptomes. Expression levels of individual igHTR show significant positive correlation between the two human samples and among the three species (Spearman correlation, rho>0.7, p<10−15) (Figure S5). Thus, igHTR expression is largely conserved across the three primate species. To test whether igHTR are conserved at the DNA sequence level, we used PhastCons scores based on nucleotide conservation among 18 placental vertebrates genomes [14]. We find that igHTR show significantly greater conservation than randomly chosen intergenic regions or annotated genic regions including both exons and introns, but are less conserved than exons alone (Figure 2D). Further, DNA sequence conservation correlates positively with igHTR expression level (Figure S6). Thus, although both expression level and DNA sequence conservation do not prove functionality, it is likely that at least some of the identified igHTR represent functional transcripts.
Do igHTR represent extensions of known genes or independent coding and/or non-coding transcripts? The size distribution of transcription clusters shows two distinct peaks: a minor one at 45 nt and a major one at 500 nt (Figure 2E). Although 500 nt is longer than the average exon size in humans, the definition of igHTR boundaries by our method is not precise. When we define exons using the same criteria as igHTR, we find a similar length distribution for both exons and long igHTR (Figure 2E). More than half of all igHTR (65%) cluster within intergenic regions, with an average of four igHTR per group. Notably, the distances between igHTR within such clusters are similar to an average intron length (Figure 2F). Furthermore, within clusters, individual igHTR are expressed at similar levels, resembling expression of exons within a gene (Figure S7). Finally, 53 individual igHTR within clusters can be connected by at least one EST sequence, while less than 5, on average, are expected by chance (simulation, p<0.01) (Figure 2G, S8). Thus, more than one half of igHTR appear to form long transcripts with exon-intron structure closely resembling annotated protein-coding genes.
With respect to the genomic location, igHTR tend to be situated within gene-rich regions, with 49% of human igHTR located within 10 kb of the nearest gene (simulation, p<0.01). Interestingly, 84% of these igHTR are close to the 3′-end, rather than 5′-end of the nearest gene (Figure 2B). Expression levels of these igHTR correlate positively with expression of the adjacent genes (Figure S9). Further, a total of 70 out of the 452 igHTR and igHTR clusters located within 10 kb from 3′-end of the nearest gene in at least one human sample can be connected to the gene by 263 EST sequences (simulation p<0.01) (Figures 2G, S10). Notably, within these igHTR, we find a significant excess of conserved microRNA (miRNA) binding sites, one of the characteristic features of 3′-UTRs of annotated transcripts (Figure S11). Thus, these igHTR may represent alternative or extended 3′-UTR of annotated genes, potentially contributing to microRNA-directed expression regulation in the primate brain.
With respect to function, 251 genes that contain igHTR within 10 kb from the gene boundaries (204 of them are situated downstream for gene and may correspond to 3′-UTR extensions) show significant enrichment among GO terms [15] and KEGG pathways [16] (Table S4, S5). Notably, these genes are mainly involved in neural functions, such as signal transduction, regulation of synaptic plasticity, learning, glutamate signaling pathway and long-term potentiation pathway, as well as two major pathways associated with lifespan duration: insulin signaling and mTOR signaling.
With respect to protein coding capacity, as determined by codon substitution frequencies (CSF) [17], igHTR scored lower than known protein coding genes, but still significantly higher than known non-coding RNAs (ncRNAs) (Wilcoxon test, p<2.2e-16) (Figure S12). Based on the chosen CSF cutoff, approximately 10% of all human igHTR may have protein-coding capacity. The remaining igHTR may represent as yet unannotated ncRNA. Supporting this suggestion, we find significant overlap (Figure S13, simulation, p<0.01) between igHTR and large intergenic non-coding RNA (lincRNA) identified in mouse and human cell lines [18], [19], involving 19% of all identified human igHTR. An additional 10% of human igHTR overlap with ncRNA predictions based on secondary structure and folding potential score determined by EvoFold [20] (Figure S14, simulation, p<0.01) (Figure 2B).
To determine the extent of expression divergence between human, chimpanzee, and rhesus macaque brain transcriptomes, we first tested whether expression of known protein coding genes could separate species according to their phylogenetic relationship. Based on expression of 13,832 genes detected in at least two out of four samples in our dataset, we found that in agreement with previously reported results based on microarray data, gene expression differs significantly among the three species (Figure 3A, S15). Furthermore, expression divergence among species increases with the time of species divergence, independent of normalization procedures and distance measures used (Materials and Methods, Figures 3B, S16).
Next, we identified genes with species-specific expression using a Bioconductor package for differential expression analysis of digital gene expression data, “edgeR” [21]. Following this methodology, we first used the variation between two human samples to build a null model of changes in read counts across all loci studied and then used this null model to identify expression differences between species. Further, we used Benjamini-Hochberg multiple testing correction to set the false discovery rate below 5% (Materials and Methods). Following this procedure, we identified 118 genes with human-specific expression in both human samples (Table S6). To test whether these expression differences are reproducible, we compared them with published expression differences measured between three human and three chimpanzee cerebellar samples using microarrays [22]. For 34 genes present in both datasets (Materials and Methods), we find significant positive correlation of human-chimpanzee expression differences (Pearson correlation r = 0.68, p = 0.0001; Spearman correlation rho = 0.55, p = 0.0008).
Functional analysis of the 118 genes with human-specific expression did not yield significant results, but showed an enrichment trend among genes involved in transcriptional regulation (Table S4). This finding is consistent with previous studies, suggesting transcriptional regulation may play an important role in human brain evolution [23], [24], [25]. Further, in terms of amino acid divergence between humans and chimpanzees or between humans and mice, as well as promoter sequence divergence, 118 genes showed tendency for greater conservation than all genes expressed in at least one of our four samples (Table S7). Thus, observed gene expression changes are not likely to reflect relaxation of selective constraint.
In addition to gene expression differences, we compared the extent of expression divergence among the three species for different types of transcripts: exonic, intronic, intergenic, and repeats. To compare expression divergence of these different transcript types on the same basis, we used two approaches. In the first approach, in addition to igHTR, we identified all other highly transcribed regions (HTR) present in human, chimpanzee, and rhesus macaque brain transcriptomes and compared their expression levels across species. From a total of 16,159 HTR found among the three species, 10,654 (65.9%) correspond to exons, 904 (5.6%) to introns, 528 (3.3%) to intergenic regions, 3,007 (18.6%) and 1,066 (6.6%) to intronic and intergenic repeats, respectively. To identify the HTR with species-specific expression, we applied the methodology described above, based on the edgeR package. Following this approach, 24 HTR (11% in all species-specific HTR) can be classified as human-specific, 32 (15%) as chimpanzee-specific, and 159 (74%) as specific to rhesus macaque in the three species comparison (Table S8). Intriguingly, for humans, we find a slight but significant excess of HTR with species-specific expression within intergenic regions (one-sided binomial test, p<2.2e-16) (Figures 3C, S17, Table S9).
In the second approach, we identified regions showing extreme species-specific divergence by comparing transcriptome coverage in a sliding window over the entire human-chimpanzee-macaque (HCM) genome alignment (Figure 3D). Windows were defined to contain the same total number of sequence reads (N = 50) summing over all three species. Using the described above approach to identify species-specific genomic windows (GW) (Table S10), we find a strong excess of intergenic region representation in all three species (one-sided binomial test, p<2.2e-16) (Figure 3C lower bars, Figure S18, Table S11). We obtain the same result using both species-specific and human-based annotations (Figure S18, Table S11). Further, the result did not depend on recent duplication events or alignment problems, as determined by allowing multiple-location mapping, use of alternative reference species in alignment construction and visual inspection of all species-specific widows. Thus, in the three primate species studied, genomic regions with extreme species-specific expression patterns are more than twice as likely to originate within intergenic regions than expected by chance (Table S11).
Our study, although based on a few samples, uncovers basic features of the brain transcriptome that are shared among the three primate species and identifies the most divergent expression patterns specific to the human brain. Among shared features, we find that exons alone contribute approximately a quarter of the total non-ribosomal transcriptome, while exons and introns together contribute three-quarters. Previously published human brain transcriptome sequencing data based on polyadenylated transcripts contains a higher proportion of exonic and a lower proportion of intronic transcription (54% and 24%, respectively, Figure 1A) [26]. Thus, many of the intronic transcripts detected in our study may represent unprocessed non-polyadenylated precursors of mature mRNA. Non-repetitive intergenic transcripts, however, occupy similar proportions (7%) in both poly(A)-enriched and the total human brain transcriptomes.
While 42% of the human brain transcriptome originate within repetitive elements, most of the repeat expression is directly proportional to the occupied genomic length and, therefore, might represent “transcriptional background”. Some of the repeat families, however, are transcribed above the background level. While some of these families, such as snRNAs, snpRNAs and 7SK RNA that derived from functional ncRNA might be actively transcribed, high expression of simple and low complexity repeats is unusual. Notably, analysis of cap-selected mouse and human transcript tags across 12 tissues shows that simple and low complexity repeats have distinct tissue-specific expression profiles and are highly expressed in brain in both species [13]. Similarly, elevated expression of Alu elements from the most recently expanded subfamilies is unusual and indicates that these elements might be transcribed actively.
Besides repeats, intergenic transcription is highly non-uniform, containing distinct highly transcribed regions conserved between species both in terms of their expression and DNA sequence. A substantial proportion of these regions (23%) may represent alternative or extended 3′-UTR of known genes, enriched in conserved microRNA binding sites. In mouse brain, 3′-UTR extensions containing miRNA binding sites were found in microRNA-Argonaute complexes, indicating their role in miRNA-directed expression regulation [27]. Further, changes in 3′-UTR length have been shown to play a role in miRNA regulation of cell proliferation and mouse embryonic development [28], [29]. Thus, identified novel 3′-UTR may play an important role in microRNA-directed regulation in the primate brain.
Another substantial proportion of identified intergenic transcripts (29%) overlap recently identified lincRNA and ncRNA predicted by EvoFold. Since our analysis is limited to highly expressed transcripts, most of them are expressed at higher levels than protein-coding genes. This indicates that at least some of these intergenic transcripts represent novel ncRNA functioning in the primate brain. We have to note, however, that these transcripts represent a small fraction of all identified lincRNA and ncRNA predicted by EvoFold: 1.7% and 0.3%, respectively. Thus, the vast majority of lincRNA and ncRNA predicted by EvoFold are not expressed in human cerebellum, or are expressed at levels below our igHTR detection threshold.
With respect to evolutionary features, the extent of expression divergence increases with greater species' phylogenetic divergence time. In our study, we do not observe an excess of expression divergence on the human lineage, previously reported in another brain region, cerebral cortex [6], [7]. Thus, in different brain regions, the transcriptome may have evolved at different rates during human evolution. It has to be noted, however, that our study does not provide intra-species variation estimates, and cannot be directly compared with the previous studies. Further work is needed to investigate this question. Notably, we find that the most extreme human-specific expression patterns, as well as extreme expression patterns characteristic for the other two primate species, show greater than expected enrichment within intergenic regions. Thus, further characterization of intergenic transcription will be necessary for understanding regulatory evolution in primates and identification of the molecular mechanisms underlying the evolution of the human-specific phenotype.
Informed consent for use of the human tissues for research was obtained in writing from all donors or the next of kin. All non-human primates used in this study suffered sudden deaths for reasons other than their participation in this study and without any relation to the tissue used.
We dissected postmortem cerebellar cortex samples from ten male humans (8–54 years old), four male chimpanzees (8–40 years old), and five male rhesus macaques (4–20 years old). All human postmortem brain tissue samples were obtained from the NICHD Brain and Tissue Bank for Developmental Disorders (NICHDBB) (Baltimore, MD, USA). Forensic pathologists at the NICHDBB defined all subjects as normal controls. No subjects with prolonged agonal state were used. Chimpanzee samples were obtained from the Yerkes Primate Center (Atlanta, GA, USA), the Anthropological Institute & Museum of the University of Zürich-Irchel, (Zürich, Switzerland), and from the Biomedical Primate Research Centre (Rijswijk, Netherlands). The rhesus macaque samples were obtained from the SuZhou Experimental Animal Center (SuZhou, China). All samples contained RNA of comparable and high quality (Table S1).
Total RNA was extracted from dissections by Trizol reagent (Invitrogen, Carlsbad, CA) and treated for 30 min at 37°C with RNase free DNase I (Ambion, Austin, TX). RNA was purified with the RNeasy MinElute Kit according to the manufacturer's instructions (Qiagen, Valencia, CA). This procedure depletes RNA molecules with length shorter than 200 nt. Resulting RNA samples from five human, five macaque, or four chimpanzee individuals was mixed in equal proportions within species, resulting in two human, one chimpanzee, and one rhesus macaque pooled samples (Table S1). 10ug RNA was treated with two rounds of RiboMinus kit (Invitrogen) to remove most of the Ribosome RNA. The cDNA libraries were prepared starting from 2ug of rRNA-Reduced total RNA per sample and using random hexamer primers (Invitrogen, Cat. No. 48190-011). It has to be noted that the resulting double-stranded cDNA fragments do not preserve information about the strand specificity of the original transcript. The Illumina sequencing libraries were prepared according to the single-end sample preparation protocol (http://www.illumina.com). The libraries were sequenced using the 1G Illumina Genome Analyzer. The sequencing products were the single-end 36 nucleotides (nt) long sequence reads. All sequence data including quality scores is deposited into the NCBI's Short Read Archive, accession number SRA011534.
All raw sequencing reads were mapped to the corresponding reference genomes (hg18, panTro2, and rheMac2), allowing a maximum of four mismatches, using Short Oligonucleotide Alignment Program (SOAP, version 1) [30]. Using a smaller number (two or three) of allowed maximum mismatches did not affect the analysis (data not shown). Only the reads that mapped uniquely were included in the analysis, unless indicated otherwise. For the three species, all uniquely mapped reads were annotated based on the species-specific gene annotation from Ensembl (release 50) provided by BioMart (http://www.biomart.org/) [31] or based on the human annotation (see below). Throughout the analysis, exons and intron categories are based on all exons, both coding and non-coding, of protein-coding genes according to Ensembl (release 50) annotation. Reads mapping to rRNA (both uniquely and allowing multiple mapping) were excluded from the analysis. To ensure complete and unbiased exclusion of rRNA sequences, for each species we mapped reads to all rRNA sequences annotated in the three species. In each sample, 36–39% of all mapped reads mapped to rRNA. For uniquely mapped reads, 1–2% mapped to rRNA. The repeat annotation was taken from the RepeatMasker table provided by the UCSC table browser (http://genome.ucsc.edu/) [32]. The genomes were separated into 7 categories: exons, intronic repeats, introns, intergenic repeats, mitochondrial chromosome, non-coding RNA, and intergenic regions. This order is further used as a category hierarchy for sequence reads annotation, from the highest to the lowest level, respectively. A sequence read was assigned to a category if at least one nucleotide of the read mapped to the category's genomic region according to the above hierarchy and independent of the strand orientation, as strand information was lost during sequence library preparation. Further, sequence reads mapped to exon junctions were assigned to exons. Although our approach biases annotation to the categories high in the hierarchy, such as repetitive elements, this effect is not large. Specifically, we find that in humans only 7% of all sequence reads we assign to repeats do not map completely within repetitive elements and, therefore, can be assigned to other categories. Further, for only 2% of all sequence reads we assign to repeats, less than half of a read sequence is contained within repetitive elements. The distribution of mapped reads shown in (Figure 1A) and (Figure S1A) is based on counting the number of sequence reads mapped to both unique and multiple (≤100 locations) positions in the genome. We obtain similar results considering only sequence reads mapped to unique genomic positions (Figure S1B).
We estimated the expression levels of repeat families based on uniquely mapped sequences. Including sequence reads mapped to multiple positions increased the total number of reads mapped to repeat regions by approximately 10%, but did not affect the results qualitatively. To normalize the expression by the lengths of unique DNA in each repeat family, we calculated the numbers of potential positions in repeat elements that can be mapped uniquely, then we summed up these numbers of all the elements and that of the expressed elements separately. This length calculation was done for both the analysis of repeat expression level vs. repeat length (Figure 1B) and the analysis of repeat transcriptional activity vs. repeat age (Figure 1C, 1D).
Pair-wise genome alignments of human-chimpanzee and human-macaque were downloaded from UCSC genome browser (genome versions: hg18, panTro2 and rheMac2). Based on these alignments, Human-Chimpanzee-Macaque (HCM) three-way alignment was constructed Using Multiz software package [33]. The human genome was selected as reference during the construction unless indicated otherwise. The regions in the HCM alignment were also considered as 3 species consensus regions (HCM consensus regions).
We used two parameters to determine whether a region is a HTR. The first is the maximum spacing (maxspacing) between two neighboring reads (from 5 to 3′ on the forward strand). The second is the minimum number of mapped sequences (minhits) within the regions. For convenience, we use maxspacing = 150 nt and minhits = 10 for all HTR analysis shown in the paper, except Figure S5. The chosen parameters are conservative, as we only select genomic regions with unusually high expression levels (Figure 2B). Using other parameter sets did not affect the results. igHTR were defined to be located entirely within intergenic regions according to Ensembl (release 50) annotation of protein-coding genes, non-coding genes, and pseudogenes. Further, they did not overlap with RefSeq and VEGA transcript annotation (downloaded from UCSC table browser, http://genome.ucsc.edu) [32]. To identify HTR in humans, chimpanzee, and macaque on an equal basis, total numbers of mapped sequences were equalized among the samples by random sub-sampling of 1,500,000 mapped sequences for each sample. This number was based on the read number in the sample with the lowest coverage. HTR were aligned across species based on HCM alignment. HTR genomic boundaries were defined based on the 5′-most and 3′-most coordinates found among the four samples.
To calculate the expression correlation of individual igHTR in the three species, we unified igHTR identified in the four samples. We mapped igHTR identified in chimpanzee and macaque onto the human genome using the LiftOver tool from UCSC genome browser (http://genome.ucsc.edu/cgi-bin/hgLiftOver).
All simulation tests were done by randomly selecting the same number of genomic regions with the same length distribution as the actual igHTR 1,000 times. The sample genomic regions differed depending on the tested variable and are described specifically in each case (see Supplementary Information for details).
Sequence conservation analysis was based on the sequence conservation measures provided for each nucleotide position by the PhastCons conservation scores for 18-way multiple alignments between the human genome and 17 other placental mammalian species [34]. Conservation was determined for nucleotides within human igHTR, as well as for the entire human intergenic regions, genic regions (including both exons intron), and exons by randomly sampling the same number of nucleotides from these regions 1,000 times.
We tested protein-coding potential of human igHTR by determining the maximum CSF (codon substitution frequency) score observed across the entire genomic locus, following [17]. Briefly, we used a scoring matrix built from human-mouse alignment and computed the CSF scores across sliding windows of 90 nucleotides. We then scanned all 6 possible reading frames in each window, since we have the strand information. After computing a score for each window, we defined the “max CSF score” for a cluster to the maximum observed score across the region. Then, we chose CSF cutoff that discriminates well between coding and non-coding regions based on the CSF distributions of known protein-coding and non-coding regions. We chose cutoff at CSF = 2, which gives specificity (97.9%) and sensitivity (93.2%) (Figure S12). Finally, we applied this cutoff to the CSF distributions of igHTR to estimate the proportion of potential protein-coding regions.
For overlap between lincRNA (large intergenic non-coding RNA) and igHTR, we used published lincRNA identified in mouse [18] and human [19]. We downloaded the lincRNA tables provided by these two papers and identified the human orthologs of the mouse lincRNA as described in [19]. Next, we combined the human lincRNA and the human orthologs with mouse lincRNA for the analysis.
For overlap between EvoFold predictions and igHTR, we download a total of 47,510 predicted RNA from UCSC browser [20]. As many of these predictions are short (∼20 nt), we assume that they originate from a longer precursor and extend the predicted locations by 1 kb at both ends for the analysis.
Among all annotated human protein-coding genes (Ensembl release 50), 18,391 can be matched between the three species based on HCM alignment. Out of these genes, 13,832 expressed in at least two of the four samples were used in this analysis. The gene expression levels were calculated as the number of sequence reads uniquely mapped in exons, normalized by the gene's exonic region length. Reads mapped to exon junctions were not counted here, because some exon boundaries might not been matched accurately between genomes based on HCM genome alignment. The expression levels were normalized across samples using quantile normalization (normalize.quantiles function in R) [35]. Divergence between samples was estimated based on Euclidean distance, Manhattan distance, and 1-rho (Spearman correlation coefficient) (Figures S15, S16). Further, to remove influence of expression level on divergence calculation, we Z-transformed expression levels before the expression distance calculation: the expression level of each gene was set to mean = 0 and standard deviation = 1 across the four samples (Figures S15, S16). The UPGMA trees (Figure S15) were constructed using R-package ape and phangorn.
To identify species-specific expression of genes, HTR, or GW, we used a Bioconductor package for differential expression analysis of digital gene expression data, “edgeR” [21]. This package models the digital expression data using a negative binomial (NB) distribution with parameters estimated from the actual data. First, we estimated the dispersion parameter in the NB model by comparing expression in two human samples (function estimateCommonDisp in edgeR package). This estimated common dispersion was then used in an exact test (function exactTest) analogous to the Fisher's exact test to detect differential expression between any two species. The resulting p-values were adjusted with Benjamini-Hochberg multiple testing correction to control the false discovery rate to be below 5%. Species-specific expression was identified separately in two groups of samples. Group one (G1) contained Human1, Chimpanzee, and Macaque samples. Group two (G2) contained Human2, Chimpanzee, and Macaque samples. Genes, HTR, or GW with significant expression difference in human-chimpanzee and human-macaque comparisons, but not in chimpanzee-macaque comparison, in both G1 and G2 were classified to have human-specific expression. Similarly, we identified genes, HTR, or GW with chimpanzee-specific and rhesus macaque-specific expression (Tables S6, S8, and S10).
HTR were determined over the entire HCM alignment using standard parameters (maxspacing = 150 nt and minhits = 10) and assigned to the annotation categories according to the hierarchy mentioned above (Materials and Methods: Read mapping and annotation). We defined GW as HCM alignment regions containing a total of 50 sequence reads in the three species.
For 118 genes with human-specific expression, 251 genes containing igHTR (within 10 kb from the gene boundaries in both directions in the human samples), and for 204 (of 251) genes with igHTR near 3′-UTR, we performed GO-term/KEGG-pathway enrichment analysis using 15,263 genes expressed in at least one out of four samples as background. For the GO function enrichment analysis, we downloaded the Ensembl gene-GO annotation from the Ensembl database [31]. We then used the func_hyper program of the package FUNC to test for category enrichment. The program generates raw enrichment p-values for each category based on hypergeometric distribution, then performs permutations of genes to determine whether the detected enrichment is greater than expected by chance, generating a global enrichment p-value [36]. For KEGG pathway enrichment analysis, we downloaded Ensembl gene-KEGG annotation from the KEGG database, and use in-house code written in R-language (supplied on request) that uses the same strategy as func_hyper. The resulting GO terms from “biological process” taxonomy and KEGG pathways with raw enrichment p-value<0.05 are listed in Tables S4 and S5.
To compare human-chimpanzee expression differences, we used expression data measured using Affymetrix arrays in three human and three chimpanzee adult cerebellar samples [22]. Provided expression levels of 6,645 genes were quantile normalized and log2 transformed. Based on these data, for each gene we calculated human-chimpanzee difference as the difference between mean expression levels in the two species. In our current RNA-Seq data, 14,959 genes are expressed in at least one of the three samples. For these genes, we quantile normalized the expression levels across three samples, log2-transformed, and calculated human-chimpanzee difference as the difference between mean expression levels in the two species. Out of 118 genes with human-specific expression in RNA-Seq experiment, 34 were present in both data sets.
We compare selective constrains in 118 genes with human-specific expression to that of 15,263 genes expressed in at least one out of four samples based on three measures: (1) Ka/Ks between human and mouse: the data was downloaded from Ensembl (release 50) [31] via Biomart and only considering 1∶1 orthologs between human and mouse. (2) Ka/Ki between human and chimpanzee: this data was downloaded from [37]. (3) Promoter sequence divergence (Kp) between human and chimpanzee: this data was downloaded from [38]. The results are shown in Table S7.
|
10.1371/journal.pgen.1005581 | A Hereditary Enteropathy Caused by Mutations in the SLCO2A1 Gene, Encoding a Prostaglandin Transporter | Previously, we proposed a rare autosomal recessive inherited enteropathy characterized by persistent blood and protein loss from the small intestine as chronic nonspecific multiple ulcers of the small intestine (CNSU). By whole-exome sequencing in five Japanese patients with CNSU and one unaffected individual, we found four candidate mutations in the SLCO2A1 gene, encoding a prostaglandin transporter. The pathogenicity of the mutations was supported by segregation analysis and genotyping data in controls. By Sanger sequencing of the coding regions, 11 of 12 other CNSU patients and 2 of 603 patients with a diagnosis of Crohn’s disease were found to have homozygous or compound heterozygous SLCO2A1 mutations. In total, we identified recessive SLCO2A1 mutations located at seven sites. Using RT-PCR, we demonstrated that the identified splice-site mutations altered the RNA splicing, and introduced a premature stop codon. Tracer prostaglandin E2 uptake analysis showed that the mutant SLCO2A1 protein for each mutation exhibited impaired prostaglandin transport. Immunohistochemistry and immunofluorescence analyses revealed that SLCO2A1 protein was expressed on the cellular membrane of vascular endothelial cells in the small intestinal mucosa in control subjects, but was not detected in affected individuals. These findings indicate that loss-of-function mutations in the SLCO2A1 gene encoding a prostaglandin transporter cause the hereditary enteropathy CNSU. We suggest a more appropriate nomenclature of “chronic enteropathy associated with SLCO2A1 gene” (CEAS).
| Advanced diagnostic innovations such as capsule endoscopy and balloon endoscopy have provided better understanding of endoscopic findings of small bowel diseases. However, it remains difficult to diagnose small intestinal diseases such as Crohn’s disease, intestinal tuberculosis, and nonsteroidal anti-inflammatory drug-induced enteropathy by the endoscopic findings alone. We previously reported a rare autosomal recessive inherited enteropathy characterized by persistent blood and protein loss from the small intestine. This enteropathy has an intractable clinical course with ineffectiveness of immunosuppressive treatment. In this study, we identified recessive mutations in the SLCO2A1 gene, encoding a prostaglandin transporter, as causative variants of this disorder by exome sequencing of four families, and showed that this disease is distinct from Crohn’s disease. We also showed that the mutations found in the patients caused functional impairment of prostaglandin E2 uptake within cells. The present findings suggest that genetic analysis together with detailed clinical information is invaluable for diagnosis of the disease, and that there may be a concept of enteropathy referred to as “prostaglandin-associated enteropathy”, irrespective of ethnic background.
| The use of capsule endoscopy and balloon endoscopy has provided a better understanding of the features of small bowel ulcers among various gastrointestinal disorders, such as Crohn’s disease (CD), intestinal tuberculosis, and nonsteroidal anti-inflammatory drug (NSAID)–induced enteropathy [1,2]. Previously, we proposed a rare clinicopathologic entity characterized by multiple small intestinal ulcers of nonspecific histology and chronic persistent gastrointestinal bleeding as chronic nonspecific multiple ulcers of the small intestine (CNSU) [3,4]. The macroscopic findings of CNSU are characterized by multiple thin ulcers in a linear or circumferential configuration and concentric stenosis, and apparently mimic those of NSAID-induced enteropathy [4–7]. CNSU predominantly occurs in females and the symptoms, such as general fatigue, edema, and abdominal pain, appear during adolescence. The clinical course of the disease is chronic and intractable with reduced effects of immunosuppressive treatment including prednisolone and azathioprine.
Although CNSU predominantly occurs in females, it also appears to be an autosomal recessive inherited disorder because of frequent parental consanguinity [8]. To identify the causative gene for this disorder, we performed whole-exome sequencing and identified recessive mutations in the SLCO2A1 gene, encoding a prostaglandin transporter, as causative variants. Furthermore, we replicated our findings in other patients with CNSU and established a genetic cause for this inherited disease.
We performed whole-exome sequencing in five affected females with CNSU (A-V–2, B-IV–3, C-IV–3, D-II–4, and D-II–5) and one unaffected individual (A-V–3) (Figs 1 and 2). Parental consanguinity was noted in families A, B, and C. The average depth of sequence coverage in the whole-exome sequencing data was 68.9× (S1 Table). We identified a total of 368,403 variants, of which 20,271 were non-synonymous or splice-site mutations.
By filtering the data with dbSNP135, we found 2,406 variants located in 1,578 genes. Based on the parental consanguinity of the patients, we focused on the shared genes with homozygous variants among three affected individuals (A-V–2, B-IV–3, and C-IV–3) and found nine candidate genes, PCSK9, ASPM, DAG1, SLCO2A1, MCPH1, EFEMP2, DDHD1, PKD1L3, and SYNGR1. After consideration of the results for the unaffected individual (A-V–3) and another two sibling patients (D-II–4 and D-II–5), only SLCO2A1 remained as a candidate gene. The three patients with parental consanguinity (A-V–2, B-IV–3, and C-IV–3) had a homozygous splice-site mutation in the SLCO2A1 gene, c.1461+1G>C or c.940+1G>A (Fig 1 and Table 1). The two sibling patients had compound heterozygous mutations, c.664G>A (p.Gly222Arg) and c.1807C>T (p.Arg603X). All four mutations were predicted to be loss-of-function or damaging mutations by SIFT and PolyPhen–2.
The four identified SLCO2A1 mutations were confirmed to be present in five affected individuals (A-V–2, B-IV–3, C-IV–3, D-II–4, and D-II–5) by Sanger sequencing (Fig 1). Segregation analysis of patient A-V–2 revealed that her unaffected parents, sister, brother, daughter, and son carried the heterozygous c.1461+1G>C mutation (Fig 1A).
To compensate for bias in our analysis, such as the possibility of ethnic-specific variants, we genotyped the four candidate variants in 747 unaffected Japanese subjects from our previous genome-wide association study [9]. All mutations except for the c.940+1G>A mutation were absent in controls (Table 2). The c.940+1G>A mutation was found in the heterozygous state in 3 of 747 controls, showing a similar allele frequency of 0.0022 to the public exome database for the Japanese population (HGVD database).
Subsequently, we screened all 14 coding exons and intron-exon boundaries using Sanger sequencing in 12 other CNSU patients, and identified two novel mutations, c.421G>T (p.Glu141X) and c.1372G>T (p.Val458Phe) (Table 1). Eleven of the 12 patients were found to have homozygous (nine patients) or compound heterozygous (two patients) SLCO2A1 mutations that were rare and predicted to be deleterious by SIFT, PolyPhen–2, and PROVEAN (Table 1). The remaining patient (66-year-old female), who did not have an SLCO2A1 mutation, was diagnosed as CNSU because of multiple ulcerations in the duodenum and jejunum. Although anti-tumor necrosis factor-α antibody therapy was ineffective, clinical improvement was achieved by enteral nutrition.
Because CNSU can be misdiagnosed as CD in some cases, we searched for the six identified mutations of SLCO2A1 in CD patients to identify concealed CNSU patients. Among 603 patients previously diagnosed as CD [10], two individuals (patients 17 and 18 in Table 1) were found to carry a pair of compound heterozygous SLCO2A1 mutations, c.940+1G>A/c.547G>A and c.940+1G>A/c.421G>T, respectively. The c.547G>A mutation (p.Gly183Arg), was a novel mutation at a highly conserved site and predicted to be deleterious by in silico analysis. The clinical information for the two individuals was reviewed retrospectively, and the diagnosis of CNSU was confirmed. In total, we identified seven deleterious SLCO2A1 mutations in 18 patients (Table 2).
In total, we found 18 patients with CNSU confirmed by genetic analysis (Table 1); 14 of them were female. In all patients, the ulcers occurred in the ileum (Fig 2A–2D). The stomach and duodenum were affected in five (27.8%) and eight (44.4%) patients, respectively.
Because mutations in the SLCO2A1 gene, encoding a prostaglandin transporter, have been reported to be the pathogenic cause of primary hypertrophic osteoarthropathy (PHO; OMIM 614441) [11–13], we investigated whether CNSU patients had clinical manifestations of PHO. Although no patients had any clinical manifestations of PHO requiring treatment, mild digital clubbing or periostosis was present in seven of 18 patients (S2 Table). Moreover, three male patients (patients 12, 16, and 17) fulfilled the major clinical criteria for PHO, having digital clubbing, periostosis, and pachydermia. There were no female patients who fulfilled the major clinical criteria (Fig 2E and 2F).
Among the identified SLCO2A1 mutations, a splice-site mutation of intron 7 (c.940+1G>A) was the most frequent, and nine of the 18 patients were homozygous for this mutation. There were no obvious correlations between the types of mutations and the clinical phenotypes.
Because the SLCO2A1 gene encodes a prostaglandin transporter that mediates the uptake and clearance of prostaglandins, the urinary levels of prostaglandin E metabolite (PGE-M) were measured. The urinary PGE-M levels in CNSU patients were significantly higher than those in unaffected individuals (p = 0.00013; S1 Fig).
Using RT-PCR, we demonstrated that splicing of the SLCO2A1 mRNA, derived from biopsy specimens of the small intestine, was altered in affected siblings with the homozygous c.940+1G>A mutation (patients 6 and 7) compared with a control individual (Fig 3A). Sequencing of the RT-PCR products revealed deletion of the whole exon 7 of SLCO2A1, leading to a frameshift at amino acid position 288 and introduction of a premature stop codon after six amino acid residues (p.R288Gfs*7). Sequencing of the RT-PCR products of the transcripts in peripheral blood mononuclear cells from patient A-V–2 revealed that the homozygous c.1461+1G>C mutation led to a 23-bp frameshift insertion into intron 10, resulting in a premature stop codon (p.I488Lfs*11) (S2 Fig).
For functional analysis of the intact and truncated SLCO2A1 proteins, we investigated the 3H-labeled prostaglandin E2 (PGE2) transport ability in HEK293 cells transfected with intact SLCO2A1 and mutant SLCO2A1 proteins for each identified mutation (c.940+1G>A, p.Gly222Arg, p.Arg603X, p.Glu141X, p.Val458Phe, and p.Gly183Arg). HEK293 cells transfected with intact SLCO2A1 showed the ability for PGE2 transport. In contrast, HEK293 cells transfected with the mutant SLCO2A1 proteins were unable to uptake 3H-labeled PGE (p < 0.0001; Fig 3B). These findings demonstrated that the mutant SLCO2A1 proteins identified in patients lost their functional ability as a PGE transporter.
In control sections of normal small intestinal mucosa, SLCO2A1 was expressed on the cellular membrane of vascular endothelial cells in the small intestine, as evaluated by immunohistochemistry and immunofluorescence staining with a specific anti-SLCO2A1 antibody recognizing the fifth extracellular domain coded by exons 9–11 of the SLCO2A1 gene (Fig 3C). We then analyzed the expression of SLCO2A1 in the small intestine of affected individuals with the homozygous c.940+1G>A mutation (patients 6 and 7) by immunofluorescence staining. However, the immunofluorescence staining did not detect any SLCO2A1 protein in the vascular endothelial cells of the patients (Fig 3C). These results indicated that the entire SLCO2A1 protein was unexpressed in affected individuals with the homozygous c.940+1G>A mutation, consistent with the results of the mRNA transcript sequencing.
To investigate the subcellular localization of SLCO2A1 and the truncated SLCO2A1 protein (ΔSLCO2A1) corresponding to the c.940+1G>A mutation, we constructed expression vectors for GFP-SLCO2A1 and GFP-ΔSLCO2A1 fusion proteins and transfected them into HEK293 cells. GFP-SLCO2A1 was localized on the cellular membrane (Fig 3D, arrows) as well as in the cytoplasm of transfected HEK293 cells, while GFP-ΔSLCO2A1 did not accumulate on the cellular membrane (Fig 3D).
In this study, we performed whole-exome sequencing in five Japanese patients with CNSU and one unaffected individual, and identified the SLCO2A1 gene as the candidate for this disorder. We further confirmed that SLCO2A1 gene mutations were involved in the pathogenesis of CNSU by genotyping of control subjects and other CNSU patients. Moreover, a genetic analysis of 603 patients previously diagnosed as CD revealed that two CNSU patients had been included in this disease group. In total, we identified seven different mutations in the SLCO2A1 gene, comprising two splicing-site mutations, two truncating mutations, and three missense mutations, as the causative gene defects for CNSU. Therefore, we propose a more appropriate nomenclature, “chronic enteropathy associated with SLCO2A1 gene” (CEAS), for this disease.
The SLCO2A1 gene encodes a prostaglandin transporter that may be involved in mediating the uptake and clearance of prostaglandins in numerous tissues [14–16]. This gene has already been reported as a causative gene for a subtype of PHO [11]. In fact, three of the seven identified mutations, c.664G>A, c.940+1G>A, and c.1807C>T, have also been reported as causative mutations for PHO [11–13,17,18]. We found that three male patients with CEAS had all of the major clinical features of PHO, such as digital clubbing, periostosis, and pachydermia. Moreover, either digital clubbing or periostosis was present in seven of 18 patients. These findings indicate that CEAS and PHO share a causative gene and that their clinical features are profoundly influenced by other modifiers. Taken together with the facts that that CEAS predominantly occurs in females and PHO predominantly occurs in males [8,17], a sex-influenced gene or hormone may be the main disease modifier. Zhang et al. [17] reported that two female family members of a PHO patient had no clinical features of PHO, despite having a homozygous SLCO2A1 mutation. Moreover, it is interesting to note that these two siblings both had anemia and hypoalbuminemia, suggesting that they had CEAS.
PHO is also known to be caused by mutations of HPGD, encoding 15-hydroxyprostaglandin dehydrogenase (15-PGDH), as well as SLCO2A1 [19]. The transmembrane prostaglandin transporter encoded by the SLCO2A1 gene delivers prostaglandins to cytoplasmic 15-PGDH, resulting in their degradation [14,20]. Because 15-PGDH is the main enzyme for prostaglandin degradation, systemic PGE2 levels are increased in patients with HPGD deficiency. Consistent with the findings in our present investigation, Zhang et al. [11] reported that the urinary levels of PGE2 and PGE-M in SLCO2A1-deficient individuals with PHO are significantly higher than those in controls. In fact, the clinical features of PHO were assumed to be caused by excessive PGE2. Meanwhile, although elevated levels of PGE2 in gastrointestinal tissues are commonly known to protect against mucosal inflammation via the prostaglandin receptor EP3/EP4 [21–23], multiple intestinal ulcers occur in CEAS. This discrepancy and the pathogenesis of intestinal ulcers need to be clarified in future studies.
Although CEAS is presumed to be unaccompanied by immunological inflammation in its pathogenesis, a portion of CEAS patients can be misdiagnosed as CD because of the shared common clinical features, such as multiple small intestinal ulcers, anemia, and hypoalbuminemia. In this study, two of 603 patients previously diagnosed as CD were found to be affected with CEAS by genetic analysis. Because corticosteroid and anti-tumor necrosis factor-α antibody therapies are ineffective for CEAS, recognition and precise diagnosis of CEAS are critical to avoid unnecessary therapies. The findings of our investigation lead us to conclude that genetic analysis in addition to detailed clinical information including digital clubbing, blood tests, and gastrointestinal examinations are invaluable for distinguishing CEAS from CD.
Cases of a similar enteropathy referred to as cryptogenic multifocal ulcerous stenosing enteritis (CMUSE) have been reported in Western populations [24–26]. This enteropathy has been shown to be an autosomal recessive inherited disease caused by mutations in the PLA2G4A gene [27]. CEAS and CMUSE share common clinicopathologic features with respect to age of onset, chronic and recurrent clinical course, and nonspecific stenosing small intestinal ulcers [4,25]. However, the sex predominance, response to steroid therapy, and lesion sites are obviously different between the two conditions. The PLA2G4A gene encodes cytoplasmic phospholipase A2-α (cPLA2α), which catalyzes the release of arachidonic acid from membrane phospholipids. CMUSE patients with compound heterozygous mutations of PLA2G4A have been reported to show globally decreased production of eicosanoids such as PGE2 and thromboxane A2, resulting in multiple ulcers of the small intestine and platelet dysfunction [27,28]. Because impaired prostaglandin use underlies CEAS, CMUSE, and NSAID-induced enteropathy, we propose a new entity of gastrointestinal disorders, namely “prostaglandin-associated enteropathy”.
In conclusion, we have identified loss-of-function mutations in the SLCO2A1 gene as the cause of CEAS. The present findings clearly indicate that CEAS is a genetically distinct entity independent of other gastrointestinal disorders including CD, NSAID-induced enteropathy, and CMUSE. Further studies are needed to elucidate the pathogenesis of CEAS and identify new therapeutic molecular targets for “prostaglandin-associated enteropathy”.
Written informed consent for genetic studies was obtained from each individual. The study was approved by the institutional review board at each collecting site in accordance with the Declaration of Helsinki Principles.
We obtained blood samples and family pedigrees from 17 Japanese patients with CNSU and eight unaffected family members in 15 families. The diagnosis of CNSU was based on the published clinical criteria and clinical courses (S3 Table) [8,29]. Genomic DNA samples from 747 participants in our previous genome-wide association study for ulcerative colitis [9] and 603 patients with CD [10,30] were used after excluding subjects who recalled their consent.
DNA was extracted from peripheral blood using standard methods. Whole-exome sequencing in five affected individuals (A-V–2, B-IV–3, C-IV–3, D-II–4, and D-II–5) and one unaffected individual (A-V–3) was performed to identify candidate genetic variants (Fig 1). Genomic DNA was enriched using a TruSeq Exome Enrichment Kit (Illumina, San Diego, CA, USA) according to the manufacturer’s instructions, and paired-end sequencing was carried out with an Illumina HiSeq 2000 instrument. Reads were aligned to the human genome reference sequence (hg19 NCBI build 37.1) and decoy sequences using BWA software [31]. Duplicate reads were removed with the Picard program (http://picard.sourceforge.net/). Recalibration and realignment of the data were accomplished with Genome Analysis Toolkit (GATK) [32,33]. Single nucleotide variants and small insertions and deletions (indels) were identified by GATK Unified Genotyper. The effect of each missense mutation was predicted using SIFT (http://sift.jcvi.org/) [34], PolyPhen–2 (http://genetics.bwh.harvard.edu/pph2/) [35], and PROVEAN (http://provean.jcvi.org/) [36].
To compensate for bias in our analysis, such as the possibility of ethnic-specific variants, we genotyped the four candidate variants identified by exome sequencing in 747 unaffected Japanese subjects by Sanger sequencing and restriction fragment length polymorphism analysis (S4 Table). For further validation, Sanger sequencing of all exons of the SLCO2A1 gene in other CNSU patients was performed using standard protocols. Finally, we genotyped the six identified mutation sites in the SLCO2A1 gene in clinically diagnosed CD patients, because CNSU can be misdiagnosed as CD.
Urine samples were collected from 15 CNSU patients and 13 unaffected individuals. The PGE-M levels were measured in duplicate using competitive enzyme-linked immunosorbent assays (Cayman Chemical, Ann Arbor, MI, USA).
We analyzed the exon 7 and exon 10 boundary mutations using RT-PCR to examine the effects of the splice-site mutations on SLCO2A1 transcription. Total RNA was extracted from biopsy specimens of the small intestine and peripheral blood mononuclear cells using a NucleoSpin RNA Kit (Macherey-Nagel, Düren, Germany) or PAXgene Blood RNA Kit (Qiagen, Hilden, Germany). cDNA was synthesized using a PrimeScript First Strand cDNA Synthesis Kit (TaKaRa, Otsu, Japan). The PCR products obtained from the cDNAs were sequenced (S5 Table).
A full-length cDNA (NM_005630) expression vector and C-terminally GFP-tagged cDNA expression vector were purchased from OriGene Technologies (Rockville, MD, USA). To construct vectors carrying a mutated cDNA, a KOD -Plus- Mutagenesis Kit (Toyobo, Osaka, Japan) was used according to the manufacturer’s instructions. The expression vectors were amplified by inverse PCR with specific primer sets (S6 Table). The PCR products were self-ligated, and transformed into Escherichia coli chemically competent DH5α cells. To correct a frameshift in the downstream of exon 7, a C-terminally GFP-tagged cDNA expression vector with deletion of exon 7 was amplified again.
On the day before transfection, HEK293 cells were trypsinized, counted, and plated onto 12-well plates at a density of 4×105 cells/well. The cells were transfected by adding a premixed solution containing 0.4 μg of expression vectors and 2 μl of ScreenFectA (Wako, Osaka, Japan). After 24 hours of incubation, the medium was exchanged twice with warmed Waymouth’s medium (Life Technologies, Carlsbad, CA, USA), and the cells were incubated for 30 minutes at 37°C in uptake medium containing [5,6,8,11,12,14,15-3H(N)]-PGE2 (PerkinElmer, Waltham, MA, USA) at 0.6 nM. The cells were washed four times with cold Waymouth’s medium, and lysed with 200 μl of RIPA Buffer (Thermo Fisher Scientific, Hemel Hempstead, UK) containing a protease inhibitor (Roche, Basel, Switzerland). The total protein concentration was quantified using a BCA Protein Assay Kit (Thermo Fisher Scientific). Next, 150 μl of cell lysate was mixed with 5 ml of MicroScint–20 (PerkinElmer), and scintillation counting was performed in a Tri-Carb 3100TR liquid scintillation spectrometer (PerkinElmer).
Formalin-fixed paraffin-embedded tissues were sectioned at 3-μm thickness. After antigen unmasking in 10 mM sodium citrate buffer (pH 6) for 15 minutes at 121°C, the sections were blocked with Protein Block Serum-Free (Dako, Glostrup, Denmark) for 30 minutes at room temperature. The sections were then incubated with a diluted anti-SLCO2A1 antibody (HPA013742; Sigma-Aldrich, St. Louis, MO, USA; antigen sequence: PSTSSSIHPQSPACRRDCSCPDSIFHPVCGDNGIEYLSPCHAGCSNINMSSATSKQLIYLNCSCVTGGSASAKTGSCPVPCAH) overnight at 4°C, followed by MAX-PO (MULTI) (Nichirei, Tokyo, Japan) for 30 minutes at room temperature. DAB solution (Nichirei) was applied for color development. After the immunocytochemistry, the sections were counterstained with Mayer’s hematoxylin solution (Nichirei). For immunofluorescence, the sections were incubated with a primary antibody mixture of the anti-SLCO2A1 antibody (HPA013742) and an anti-VE-cadherin antibody (LS-B3780; LifeSpan BioSciences, Seattle, WA, USA) overnight at 4°C, followed by a secondary antibody mixture of Alexa Fluor 568-conjugated goat anti-rabbit IgG (H&L) antibody and Alexa Fluor 488-conjugated goat anti-mouse IgG (H&L) antibody (Life Technologies) for 30 minutes at room temperature. The stained sections were analyzed using an ECLIPSE TE2000-U (Nikon, Tokyo, Japan).
For observation of HEK293 cells expressing GFP-fusion proteins, cells were fixed with 4% paraformaldehyde phosphate buffer solution (Wako) for 20 minutes, and then permeabilized with 0.1% Triton X–100 (Sigma-Aldrich) in D-PBS(-) solution (Wako) for 20 minutes. Nuclei were stained with 16.2 μM Hoechst 33342 (Life Technologies) in D-PBS(-) solution for 5 minutes. GFP and nuclei were visualized using a 40× objective on an LSM710 Laser Scanning Confocal Microscope (Carl Zeiss, Oberkochen, Germany).
The chi-square test and Fisher’s exact test, where appropriate, were used to analyze categorical data. Student’s t-test was used to compare quantitative data between two groups. Dunnett’s method was used for multiple comparisons with a control group. The analyses were performed using JMP Pro statistical package 11.2.0 (SAS Institute, Cary, NC, USA). Values of p < 0.05 were regarded as statistically significant.
|
10.1371/journal.pntd.0001994 | Cerebral Toxoplasmosis Mimicking Subacute Meningitis in HIV-Infected Patients; a Cohort Study from Indonesia | HIV-associated subacute meningitis is mostly caused by tuberculosis or cryptococcosis, but often no etiology can be established. In the absence of CT or MRI of the brain, toxoplasmosis is generally not considered as part of the differential diagnosis.
We performed cerebrospinal fluid real time PCR and serological testing for Toxoplasma gondii in archived samples from a well-characterized cohort of 64 HIV-infected patients presenting with subacute meningitis in a referral hospital in Indonesia. Neuroradiology was only available for 6 patients.
At time of presentation, patients mostly had newly diagnosed and advanced HIV infection (median CD4 count 22 cells/mL), with only 17.2% taking ART, and 9.4% PJP-prophylaxis. CSF PCR for T. Gondii was positive in 21 patients (32.8%). Circulating toxoplasma IgG was present in 77.2% of patients tested, including all in whom the PCR of CSF was positive for T. Gondii. Clinically, in the absence of neuroradiology, toxoplasmosis was difficult to distinguish from tuberculosis or cryptococcal meningitis, although CSF abnormalities were less pronounced. Mortality among patients with a positive CSF T. Gondii PCR was 81%, 2.16-fold higher (95% CI 1.04–4.47) compared to those with a negative PCR.
Toxoplasmosis should be considered in HIV-infected patients with clinically suspected subacute meningitis in settings where neuroradiology is not available.
| If HIV-infected patients present with seizures, focal neurological symptoms or confusion, a CT-scan or MRI of the brain is normally made. If mass lesions are found (and the CD4 cell count is sufficiently low), cerebral toxoplasmosis is suspected, and often treated empirically. However, some of the symptoms of cerebral toxoplasmosis may mimic those of subacute meningitis. Therefore, in settings where no cerebral imaging can be performed, HIV-associated cerebral toxoplasmosis may be under-diagnosed. We retrospectively looked for toxoplasmosis in a cohort of HIV-infected patients presenting with subacute meningitis in an Indonesian hospital, where neuroradiology was not available for most patients. Patients mostly came with newly diagnosed and advanced HIV infection and few were on HIV-treatment or PJP-prophylaxis. Molecular testing of cerebrospinal fluid (CSF) was positive for Toxoplasma gondii in 32% of patients, serology was positive in 78%. Clinically, in the absence of neuroradiology, toxoplasmosis was difficult to distinguish from tuberculosis or cryptococcal meningitis. A positive CSF T. gondii PCR was associated with a two-fold increased mortality. We conclude that toxoplasmosis should be considered in HIV-infected patients with clinically suspected subacute meningitis in settings where neuroradiology is not available.
| In settings of Africa and Asia, the most common cause of subacute meningitis in patients with advanced HIV infection is either tuberculous or cryptococcal infection [1], [2]. However, in many patients, the etiology of subacute meningitis cannot be established [1], [3]. In line with a large retrospective cohort of adult meningitis patients in South Africa, where 52.8% had no definite diagnosis despite extensive microbiological testing [1], we could not identify the causative pathogen in 48.9% of HIV-infected meningitis patients in an Indonesian setting [4].
Toxoplasmosis is a common and serious central nervous system (CNS) infection in patients with advanced HIV infection [5]–[8], although its incidence has decreased with introduction of antiretroviral treatment (ART) [6], [9]. Cerebral toxoplasmosis mostly presents as cerebral mass lesions with headache, confusion, fever, lethargy, seizures, cranial nerve palsies, psychomotor changes, hemiparesis and/or ataxia [10]. Some of these symptoms may also mimic meningitis, but cerebral toxoplasmosis is generally not considered as a differential diagnosis of subacute meningitis in HIV-infected patients. This is especially the case in low-resource settings where no CT or MRI can be performed. We have therefore examined if toxoplasmosis can be diagnosed in HIV-infected patients presenting with subacute meningitis of unknown origin in Indonesia, using cerebrospinal fluid (CSF) PCR for T. gondii.
Anonymized CSF and blood samples were used from an already-existing hospital collection, from a cohort of patients collected as part of a project ‘Optimization of diagnosis of meningitis’, approved by the Ethical Committee of Hasan Sadikin Hospital/Medical Faculty of Universitas Padjadjaran, Bandung, Indonesia (No. 85/FKUP-RSHS/KEPK/Kep/EC/2006). As this study was done using already existing sample collection, no separate consent was asked for this study. HIV testing is done routinely with oral informed consent for all patients with suspected meningitis in Hasan Sadikin hospital, after 24% were found HIV-positive in a previous cohort study of 185 patients in the same hospital [4]. Consent is obtained from closest relatives (husband/wife or parents) for those patients who are unstable or unconscious at time of presentation. With approval from the ethical committee HIV testing was done anonymously afterwards for those who had died before consent could be obtained.
We included adult patients presenting with suspected meningitis at Hasan Sadikin Hospital, the top referral hospital for West Java Province, Indonesia, between December 2006 and October 2010. Clinical data including outcome was recorded in individual case report form. Definite TB meningitis was diagnosed if CSF culture or real time PCR were positive for M. tuberculosis, cryptococcal meningitis if either CSF India Ink examination, culture or cryptococcal antigen testing were positive, and toxoplasmosis if CSF T. gondii PCR was positive. HIV testing is done routinely for patients presenting at this hospital, but cerebral CT-scanning is rarely done in this setting and is not covered by the government health insurance for the poor.
CSF cell count and differentiation, protein and glucose were measured. CSF microscopy was done for cryptococci, acid-fast bacilli and bacterial pathogens. CSF was cultured for Mycobacterium tuberculosis (solid Ogawa and liquid MB-BacT, Biomerieux), bacterial pathogens (blood agar, chocolate agar, and brain-heart infusion) and fungi (Sabouraud). Cryptococcal antigen (CALAS, Meridian Diagnostics) testing was done on CSF samples following the manufacturer's instructions. Five to 7 ml CSF samples were used for molecular testing. After centrifugation of CSF samples at 3000×g for 10 minutes, DNA was extracted from 200 µl of CSF sediment by using QIAmp DNA mini kit (Qiagen, USA). CSF M. tuberculosis real time PCR was done using IS6110, a repeated insertion sequence specific for M. tuberculosis, as a target [11]. Measurement of CD4-cell count for HIV-patients only became available during the time of the study and was measured only for those who survived for more than 4 days. Real time PCR for T. gondii, using the multicopy B1 gene of the T. gondii as the target as described elsewhere [12], was performed to archived CSF samples at Radboud University Nijmegen Medical Centre. CSF specimens from 22 HIV-negative meningitis patients (16 with definite TB meningitis, 2 with bacterial meningitis, and 4 with no definite diagnosis), and nine patients with non-infectious CNS diseases, all recruited at Hasan Sadikin Hospital, were used as controls for T. gondii PCR. These samples were collected during the study period over a similar time scale compared to the case CSF samples. Toxoplasma immunoglobuline G (toxoplasma IgG) were measured by electro chemiluminescent assay (ECLIA, Elecsys, Roche) in archived serum samples of patients included in the study.
Characteristics of patients with definite tuberculosis, cryptococcosis and toxoplasmosis were compared using Chi-square test for proportions and Mann-Whitney U test for continuous variables. Progression to death using 2-month mortality data was examined by Kaplan–Meier estimates.
During the period, 401 patients presented with clinical meningitis, 76 were diagnosed with HIV infection, and 64 had archived CSF samples and were included in this study. Patients included in the study presented after a median 7 days, with meningismus (86.0%), headache (80.8%), lowered consciousness (33.3%), fever (28.8%), hemi- or tetraparesis (28.6%), cranial nerve palsies (12.5%), and seizures (10.9%). HIV was newly diagnosed in 53 patients (82.8%). All 11 patients previously diagnosed with HIV were taking ART, and 6 were using co-trimoxazole as Pneumocystis jiroveci (PJP) prophylaxis at time of presentation. The median CD4 cell count was 22 cells/mL, and less than 200 cells/mL in 22 out of 23 patients tested (96%).
CSF T. gondii PCR was positive in 21 of 64 HIV-infected patients (32.8%), with a median Ct-value of 36.0 (IQR: 34.2–39.3). None of the 22 HIV-negative control and 9 non-infectious CNS disease patients had a positive T. gondii PCR. Archived serum sample was not available in 14 patients. Toxoplasma IgG was positive in 78% of patients tested, including all patients with positive CSF T. gondii PCR. Toxoplasma IgG titers were higher among patients with a positive CSF T. gondii PCR (p = .017). A definite diagnosis of TB meningitis was established in 21/64 patients (32.8%). Out of 21 patients with positive T. gondii PCR, five had combined tuberculosis and toxoplasmosis. Cryptococcosis was diagnosed in 15/64 patients (23.4%), including two who were also diagnosed with tuberculosis. In 14 patients (21.9%) no causative pathogen was isolated.
Neck stiffness, headache and fever, the classical signs of meningitis, were equally common in patients diagnosed with toxoplasmosis, cryptococcosis and tuberculosis, as were most other signs and symptoms, except hemiparesis (Table 1). None of the patients with toxoplasmosis had received ART or co-trimoxazole prophylaxis prior to admission with meningitis. CT scans were available for 6 patients, including 4 with a positive T. gondii PCR. Three showed signs of hydrocephalus, one a hypodense lesion that showed no enhancement using contrast, and two were normal. No mass lesions typical for cerebral toxoplasmosis were seen.
CSF cell count and protein were normal or mildly elevated in patients with toxoplasmosis, and hypoglycorrhachia was less common compared with tuberculosis or cryptococcosis (Table 1). CD4 counts, missing in two-thirds of patients due to early death or the unavailability of CD4 cell testing during the initial phase of the cohort study, were low in all but one patient. Table 2 lists the CSF findings of individual patients, Figure 1 is a graphic representation of the CSF cell count, protein and glucose ratio, showing the overlap in CSF findings between patients with toxoplasmosis, cryptococcal and tuberculosis CNS infection.
Patients with confirmed cryptococcosis received amphotericine B, followed by fluconazole; all others received empiric tuberculosis treatment combined with adjunctive corticosteroids [13]. No toxoplasmosis treatment was given, as T. gondii PCR was performed retrospectively and was not available at time of presentation. Eight patients were lost to follow up and were not included in Kaplan Meier analysis. Mortality among those with positive CSF T. gondii PCR was 2.16-fold (95% CI 1.04–4.47) higher compared to those who had a negative PCR result; median survival was 7 days for toxoplasmosis, 7 days for tuberculosis meningitis, 110 days for cryptococcosis, and 32 days for patients with an unknown cause of meningitis (Figure 2).
In our cohort of HIV-infected patients presenting with clinical signs and symptoms of CNS infection, CSF T. gondii PCR was positive in 32.8% of patients, sometimes in conjunction with tuberculosis. In the absence of CT or MRI of the brain, toxoplasmosis could not be distinguished from tuberculosis or cryptococcosis. Mortality in this cohort of newly diagnosed and advanced HIV infection was extremely high and associated with a positive T. gondii PCR.
Cerebral toxoplasmosis typically causes space occupying lesion(s), leading to subacute or acutely developing confusion, with or without focal neurological deficits [14]. In the absence of CT or MRI of the brain, common findings like headache, fever, hemiparesis and decreased level of consciousness [10] may mimic those of meningitis [4], [15]–[17]. In previous series of cerebral toxoplasmosis [18], [19], meningeal signs have been reported in 3 to 16% of the patients, although in many reports neck stiffness is not mentioned [14]. Although rare, cases of spinal cord toxoplasmosis have also been reported [20]. No typical mass lesions were found in 6 patients with an available CT scan. This is not surprising, as this study depended on the availability of CSF samples, that would not have been obtained if typical mass lesions had been found.
We used T. gondii PCR for diagnosis of cerebral toxoplasmosis. In previous studies CSF T. gondii PCR had a sensitivity of 50–60% to confirm cerebral toxoplasmosis in HIV-infected patients [21], [22]. The sensitivity is possibly higher among patients with meningoencephalitis compared to those with space-occupying lesions only, but this has not been examined. Specificity of T. gondii PCR is high, between 97 and 100% [22]–[24]. The positivity rate of 32.8% in our study might therefore be an underestimate, especially in the category of patients in whom no other pathogen was isolated despite extensive microbiological testing.
In our cohort, toxoplasmosis could not be distinguished clinically from tuberculosis and cryptococcosis. From our previous series [4], CSF samples were available for the current study for 36/47 HIV-infected patients. Ten out of 17 patients who were diagnosed with ‘probable TB meningitis’ and ‘unknown’ in the previous study were found to have a positive T. gondii PCR (and no bacteriological confirmation of tuberculosis) in the current study. Diagnosis of cerebral toxoplasmosis is usually based on clinical findings and CT or MRI of the brain. However, if cerebral imaging is lacking, toxoplasmosis may not be considered. Positive toxoplasma serology, which has a high sensitivity but very poor specificity, is helpful to exclude but not to confirm cerebral toxoplasmosis, although some reports suggest that high toxoplasma IgG titers are only found in patients with symptomatic toxoplasmosis [25]. Indeed, in our study, patients with a positive T. gondii PCR had a higher IgG titers compared to those who had a negative PCR. An autopsy study from India provides further support for the notion that cerebral toxoplasmosis is not always considered; among 233 HIV patients, toxoplasmosis accounted for 6.8% of deaths, but in none of these cases toxoplasmosis had been suspected clinically [26].
The incidence of cerebral toxoplasmosis varies between countries [14] and is related to the seroprevalence of toxoplasmosis in the general population [19], [25]. In the United States, toxoplasma seroprevalence varies from 3% to 30%, whereas in France 73%–90% of the population is infected [10]. Reported seroprevalence rates were varied from 13–31% in the general population, and 45–68% in HIV patients in studies from several developing countries [8], [27], [28]. In our study, 78% of patients had detectable toxoplasma IgG, but this does not reflect the seroprevalence in the general population or among unselected HIV-infected patients, as only meningitis patients were examined.
Mortality in this cohort of patients was very high, higher compared to reported rates in other series [8], [9], [29]. One explanation is that patients mostly presented with advanced and untreated HIV infection. In addition, no toxoplasmosis treatment was provided, as PCR was done retrospectively on archived samples. In our previous study, HIV infection was associated with a 2.5-fold increased mortality among patients presenting with meningitis [4]. Data from the current study suggests that this may is at least in part attributable to a high prevalence of (unrecognized and untreated) toxoplasmosis. Future studies should examine the benefit of timely diagnosis and/or empiric treatment of toxoplasmosis for patients in settings like ours. Empiric treatment for subacute meningitis in HIV-infected patients should probably also include tuberculosis, which is difficult to exclude as culture is slow and microscopy and commercial PCR assays have insufficient sensitivity [30].
Our study suffers from several limitations. Most importantly, no CT or MRI of the brain could be performed. In addition, clinical data, CD4 cell counts and other laboratory parameters were missing in a number of patients. Despite these limitations the data strongly suggest that toxoplasmosis should be included in the differential diagnosis of HIV-infected with clinically suspected subacute meningitis, and that molecular testing or empiric treatment for toxoplasmosis should be considered in these patients, especially if no CT or MRI can be performed. Obviously, timely diagnosis and treatment of HIV will help prevent this severe opportunistic infection.
|
10.1371/journal.pgen.1002425 | Microenvironmental Regulation by Fibrillin-1 | Fibrillin-1 is a ubiquitous extracellular matrix molecule that sequesters latent growth factor complexes. A role for fibrillin-1 in specifying tissue microenvironments has not been elucidated, even though the concept that fibrillin-1 provides extracellular control of growth factor signaling is currently appreciated. Mutations in FBN1 are mainly responsible for the Marfan syndrome (MFS), recognized by its pleiotropic clinical features including tall stature and arachnodactyly, aortic dilatation and dissection, and ectopia lentis. Each of the many different mutations in FBN1 known to cause MFS must lead to similar clinical features through common mechanisms, proceeding principally through the activation of TGFβ signaling. Here we show that a novel FBN1 mutation in a family with Weill-Marchesani syndrome (WMS) causes thick skin, short stature, and brachydactyly when replicated in mice. WMS mice confirm that this mutation does not cause MFS. The mutation deletes three domains in fibrillin-1, abolishing a binding site utilized by ADAMTSLIKE-2, -3, -6, and papilin. Our results place these ADAMTSLIKE proteins in a molecular pathway involving fibrillin-1 and ADAMTS-10. Investigations of microfibril ultrastructure in WMS humans and mice demonstrate that modulation of the fibrillin microfibril scaffold can influence local tissue microenvironments and link fibrillin-1 function to skin homeostasis and the regulation of dermal collagen production. Hence, pathogenetic mechanisms caused by dysregulated WMS microenvironments diverge from Marfan pathogenetic mechanisms, which lead to broad activation of TGFβ signaling in multiple tissues. We conclude that local tissue-specific microenvironments, affected in WMS, are maintained by a fibrillin-1 microfibril scaffold, modulated by ADAMTSLIKE proteins in concert with ADAMTS enzymes.
| The microenvironment is specified by cell-surface molecules, growth factors, and the extracellular matrix. Here we report genetic evidence that implicates fibrillin-1, a ubiquitous extracellular matrix molecule that sequesters latent growth factor complexes, as a key determinant in the local control of musculoskeletal and skin microenvironments. A novel mutation in fibrillin-1 demonstrates that modulation of the fibrillin microfibril scaffold can influence tissue microenvironments and result in the clinical features of Weill-Marchesani syndrome (WMS), including thick skin, short stature, and brachydactyly. Dysregulated WMS microenvironments diverge from Marfan pathogenetic mechanisms, which lead to broad activation of TGFβ signaling in multiple tissues.
| Mutations in fibrillin-1 cause the pleiotropic features of the Marfan syndrome (MFS, OMIM#154700). MFS is recognized by its unique combination of skeletal, cardiovascular, and ocular features (long bone overgrowth, aortic root dilatation and dissection, and ectopia lentis). More than a thousand different mutations in FBN1, the gene for fibrillin-1, are known to cause MFS, suggesting that the same general pathogenetic mechanisms are initiated by each of these distinct mutations. In contrast, Weill-Marchesani syndrome (WMS, OMIM #608328) is a rare disorder described as “opposite” to MFS [1]. WMS, one of several types of acromelic chondrodysplasias, is characterized by short stature, brachydactyly, thick skin, and ectopia lentis. Previous studies reported that the autosomal dominant form of WMS is caused by mutations in FBN1 [2], [3], while mutations in ADAMTS10 were shown to cause recessive WMS [4], [5]. Since the clinical features of WMS and MFS may sometimes overlap [6], it is not certain how rare mutations in FBN1 can bring about WMS instead of MFS. Additional investigations are required in order to clearly establish the role of fibrillin-1 in causing WMS.
A role for fibrillin-1 in skin fibrosis was first suggested when a mutation in Fbn1 was identified in the tight-skin (tsk) mouse [7]. More recently, mutations in FBN1 were found in Stiff Skin Syndrome (SSKS, OMIM #184900), a rare disorder characterized by hard, thick skin and joint contractures [8]. Both the tsk and SSKS phenotypes are caused by heterozygous mutations. However, the tsk mutation is a large in-frame gene duplication, while SSKS mutations are missense mutations confined to exon 37. The molecular mechanisms by which fibrillin-1 regulates skin fibrosis are obscure. Why the tsk and SSKS mutations do not cause MFS is also obscure.
Fibrillin-containing microfibrils are small diameter fibrils that are usually found in bundles or in association with elastic fibers. Individual fibrillin microfibrils are long and can be extended in vivo when tissues are under tension [9]. In humans and mice with MFS, fibrillin microfibril bundles were fragmented in the skin [10], [11]. In contrast, fibrillin microfibrils in human scleroderma skin were disorganized, when labeled and examined by immunoelectron microscopy [12]. The latter observation was extended by ultrastructural studies of fibrillin microfibrils in SSKS [8]. In SSKS, fibrillin microfibrils were found in large aggregates within which individual microfibrils appeared to be short [8]. These electron microscopic observations suggest that structural abnormalities in fibrillin microfibrils may underlie the differences between MFS and SSKS disease pathologies.
Another possibility is that mutations causing SSKS or WMS perturb growth factor signaling, since fibrillin-1 targets and sequesters the large latent Transforming Growth Factor β (TGFβ) complex [13], [14] as well as multiple Bone Morphogenetic Proteins (BMPs) and Growth and Differentiation Factor-5 (GDF-5) [15]–[17]. In MFS, abnormal activation of TGFβ signaling contributes to phenotypes in the lung [18] and aorta [19], but TGFβ signaling may not be abnormally activated in the skin. In SSKS or WMS skin, overproduction of collagen may be predicted to be due to abnormal activation of TGFβ. But, it is unclear why abnormal activation of TGFβ signaling would be limited to the skin in SSKS or WMS and alternatively to the lung and aorta in MFS. Different mechanisms may be involved in the activation of TGFβ signaling in these disorders and/or important unknown factors may limit the effects of mutations in fibrillin-1 to specific tissues.
Here we identify a novel mutation in fibrillin-1 in a family with WMS. In order to reveal the complex mechanisms by which fibrillin-1 differentially regulates connective tissues, we replicated this mutation causing human WMS in the mouse. We show that this mutation does not cause MFS, since the WMS mouse survives normally and does not display major features of MFS, even in homozygosity. Instead, WMS mutant mice develop skin fibrosis associated with distinctive ultrastructural abnormalities in fibrillin microfibrils. In addition, WMS mice demonstrate retardation of long bone growth. Therefore, WMS mice recapitulate cardinal features of human WMS. To elucidate molecular mechanisms of WMS, we provide evidence that the WMS mutation abolishes the binding site in fibrillin-1 for a novel family of proteins, the ADAMTSLIKE (ADAMTSL) proteins. We also connect ADAMTSL proteins with ADAMTS-10 and propose that these proteins form a complex with fibrillin-1. These findings implicate ADAMTSL proteins, together with ADAMTS-10, in the regulation of fibrillin microfibril structure. Insights gained from these studies are relevant to MFS and to the expanding genetic diseases that constitute the Marfan-related disorders. Furthermore, our results point out the importance of fibrillin-1 in the local regulation of tissue-specific microenvironments.
A family with autosomal dominant WMS was previously described, and linkage analysis identified FBN1 on chromosome 15q21.1 as the disease locus [2]. This family includes affected individuals in three generations (Figure 1a). Affected individuals exhibited characteristic features of WMS including microspherophakia, ectopia lentis, glaucoma, brachydactyly, short stature, and thickening of the skin, as previously documented for individuals 5016, 4084, and 5010 [2]. We further examined individuals 5010 (45 years of age), 5011 (15 years of age), and 6013 (10 years of age) and unaffected family members. Early removal of ocular lenses and short stature were common. Mild brachydactyly of the toe and some limitation of small joints were found in 6013. 5011 showed brachydactyly, more pronounced limitation of large and small joints, and thickened skin. 5010 showed severe limitation of small and large joints with pain and loss of dexterity and thick forearm skin without striae. In addition, all three individuals showed increased truncal and axial muscle bulk. No history or clinical evidence of valvular cardiac or aortic disease was found in this family.
Southern blotting of genomic DNA and PCR followed by DNA sequencing revealed a heterozygous 7895 nt genomic deletion in FBN1 (Figure S1a and Figure 1b) with boundaries in introns 8 and 11. PCR results from unaffected and affected family members demonstrated that the mutation segregated with the disease (Figure 1b). Transcripts from the mutant allele lacked exons 9–11 (Figure 1b), predicting in-frame translation of fibrillin-1 molecules in which the first 8-cysteine domain, the proline-rich region, and EGF-like domain 4 are missing (Figure 1c). No similar mutations have been reported in the FBN1 mutation database, where only 23 of 1,013 mutations were deletions or insertions [20]. By using ribonuclease protection assay, we were able to show that the wildtype and the mutant FBN1 allele were equally expressed (Figure S1b). Using a quantitative sandwich ELISA, we found that wildtype and mutant fibrillin-1 proteins were equally secreted by affected WMS fibroblasts (Figure 1d). Immunofluorescence of skin from an affected individual (5011) showed fibrillin fibrils that appeared normal and not fragmented (Figure 1e), unlike the fragmented fibrils observed in MFS skin [10].
In contrast to the numerous different mutations in FBN1 known to cause MFS, there is only one report of an FBN1 mutation in a family with autosomal dominant WMS [3]. There are also reports of individuals with WMS that overlap with MFS [6]. In order to test whether the three-domain deletion in fibrillin-1 found in our family with WMS causes WMS and not MFS, we replicated the mutation in a mouse model (WMΔ) using a gene targeting strategy (Figure 2a). WMΔ heterozygous (WMΔ/+) and homozygous (WMΔ/WMΔ) mice breed well and are viable. Both heterozygous and homozygous mutant mice live longer than 1.5 years with no signs of aortic disease typical of MFS. Aortic root morphology in heterozygous and homozygous mutants is normal, even at 10 months of age (Figure 2b and Figure S2), in contrast to heterozygous and homozygous mutant mouse models of MFS [11], [19], [21]–[23]. In addition, with the exception of the mutant mgR/mgR [21], which is hypomorphic for normal Fbn1 and dies during early adulthood, homozygous mutant mouse models of MFS die in the early postnatal period [11], [22], [23]. By these two major criteria for MFS in mice—aortic disease and early death of homozygotes—WMΔ mice do not model MFS.
Brachydactyly and short stature are features of WMS, while arachnodactyly and tall stature are characteristic of MFS. Therefore, long bones in the WMΔ mutant mice were measured. Growth of long bones appeared to be normal in the first two weeks of postnatal life but was reduced by 3–4 weeks of age in homozygous mice (WMΔ/WMΔ). Measurements of the long bones at 1 month of age in the WMΔ/WMΔ mice using μCT showed a statistically significant (P<0.05) reduction in lengths of the radius, ulna, and tibia of 6–10% (Figure 2c) compared to age and gender matched wildtype controls. At 3–4 weeks of age, length measurements of metacarpals and proximal and distal phalanges in fore- and hindpaws were also reduced between 2–23% in the WMΔ/+ and WMΔ/WMΔ mice (Figure 2d). These findings are consistent with the WMS phenotype. However, by 5 months of age, these differences in length were normalized.
Gross examination suggested a thickened, less elastic skin in WMΔ mutant mice (Figure 3a). Histology of skin biopsies from WMΔ mice showed excessive collagen deposition in the dermis starting at 1 month of age. Hematoxylin and eosin or Masson's Trichrome stains (Figure 3b, 3c) revealed a widened dermal layer with decreased hypodermal fat, and thicker, more densely packed collagen fibers in mutants compared to wildtype mice. qPCR analyses demonstrated upregulated expression of collagen genes in the skin from mutant mice (Figure 3d). Skin thickness, as determined by histological stains and detected by touch by 7 months of age, persisted through old age in the mutant mice.
Electron microscopy after immunogold labeling with anti-fibrillin-1 antibodies showed alterations in fibrillin microfibril ultrastructure in WMΔ/+ and WMΔ/WMΔ skin: large bundles of microfibrils as well as microfibrils around elastin cores showed reduced periodicity of immunogold labeling in the mutants (Figure 4a,4b). In addition, large accumulations of microfibrils were prominent in WMΔ/+ and WMΔ/WMΔ skin (Figure 4a), and elastic fibers appeared moth-eaten compared with wildtype littermates (Figure 4b). The disorganized appearance of microfibrils is better visualized in the three-dimensional aligned tilt series of immunolabeled microfibrils from WMΔ/WMΔ and wildtype skin samples supplied Videos S1 and S2.
While some areas of the skin appeared normal, electron microscopic examination of skin from an 18 year old individual with WMS (unrelated to the WMS family described above) revealed unusually large abnormal aggregates of microfibrils (visible at low magnification in Figure 4c, left panel, arrows). Elastic fibers also appeared moth-eaten (data not shown). Similar to observations of WMΔ mutant mouse skin (Figure 4a), immunogold labeling with antibodies specific for fibrillin-1 demonstrated both irregular labeling of the microfibril aggregates and periodic labeling of apparently normal microfibrils (Figure 4c, middle panel). Based on these immunolocalization results with antibodies specific for fibrillin-1, we conclude that the small and large aggregates are composed of abnormal bundles of fibrillin-1 microfibrils. Large microfibril aggregates similar to those in human WMS (Figure 4c, left and middle panels) were found in skin from older (11–20 month old) WMΔ/WMΔ mice (Figure 4c, right panel). These results provide evidence for a common pathogenetic mechanism for fibrosis in human WMS and in this mouse model of WMS.
We previously showed that ADAMTSL-6 interacts with the N-terminal half of fibrillin-1 with high affinity (KD = 80 nM) [24]. In order to test whether other ADAMTSL family members also bind to fibrillin-1 and whether the binding site utilized by ADAMTSL-6 is perturbed by the WMS three-domain deletion in fibrillin-1, we generated recombinant fibrillin-1 polypeptides, rF84 and rF84WMΔ (Figure S3) and rF90 and rF90WMΔ (Figure 1c), as well as recombinant human ADAMTSL-1, -2, -3, and mouse papilin polypeptides (Figure S3). Surface plasmon resonance (SPR) technology was employed to measure interactions between fibrillin-1 and ADAMTSL polypeptides. Similar to ADAMTSL-6 [24], ADAMTSL-2, -3, and papilin polypeptides interacted with the N-terminal half of fibrillin-1, while ADAMTSL-1 did not. Binding to the C-terminal half of fibrillin-1 was negative for all ADAMTSL proteins tested (data not shown). SPR sensorgrams are shown for ADAMTSL-2 binding to fibrillin-1 and for ADAMTSL-3 binding to fibrillin-1 (Figure 5a). ADAMTSL-2, -3, and -6 and papilin polypeptides did not bind to recombinant fibrillin-1 polypeptides with the WMS three-domain deletion (Figure 5a and Table S1). Binding constants for all of these interactions were calculated from the SPR data (Table S1). We conclude that the fibrillin-1 domains consisting of the first 8-cysteine domain, the proline-rich region, and the 4th generic EGF-like domain contain the ADAMTSL binding site(s).
Because recessive WMS is caused by mutations in ADAMTS10 [4], [5], FBN1 and ADAMTS10 share a genetic pathway. Therefore, we hypothesized that fibrillin-1, ADAMTS-10, and some ADAMTSL proteins form protein complexes. In a pull-down assay with Ni-NTA as a resin, we showed that full-length, His6-tagged ADAMTS-10 in media of stably transfected EBNA 293 cells bound to the N-terminal half of fibrillin-1 (Figure 5b). From SPR interaction studies, a KD of 450 nM was calculated for the binding of the N-terminal fibrillin-1 polypeptide to the C-terminal end of ADAMTS-10 (data not shown). The C-terminal recombinant ADAMTS-10 polypeptide used in the SPR studies represents the noncatalytic region of ADAMTS-10, a region composed primarily of Tsp1 repeats (Figure S3). SPR also showed that the C-terminal end of ADAMTS-10 interacted with the C-terminal end of ADAMTSL-3 with high binding affinity (KD = 2 nM) (Figure 5c). However, neither ADAMTSL-2 nor -1 bound to ADAMTS-10, indicating that ADAMTS enzymes may partner only with specific ADAMTSL proteins. Taken all together, these results suggest that direct interactions between fibrillin-1, ADAMTS-10, and specific ADAMTSL proteins are involved in the pathogenesis of WMS.
Based on in vitro studies, we hypothesized that the mutant WMS fibrillin-1 cannot interact properly in vivo with certain members of the ADAMTSL family of proteins. To test if localization of ADAMTSL proteins is altered in WMΔ mutant mice, we stained skin with antibodies specific for ADAMTSL-6 [24]. Results showed a reduction in ADAMTSL-6 immunofluorescence in skin from WMΔ/+ and WMΔ/WMΔ mutant mice compared to wildtype littermate skin (Figure 6a). Fibrillin-1 immunofluorescence was equal in pattern and abundance in WMΔ mutant and wildtype mice (data not shown). Since antibodies specific for ADAMTSL-2 and -3 are not yet available, we were unable to determine whether these proteins also colocalize with fibrillin-1 in skin and whether these are also reduced in WMΔ mutant mice.
ADAMTSL-6 and ADAMTS-10 promote fibrillin-1 fibril formation [24], [25]. Therefore, we examined human and mouse fibroblasts for defects in fibrillin-1 fibril formation. The Marfan cell culture assay [10] was used. In this assay, control fibroblasts assemble a matrix containing abundant immunofluorescent fibrillin-1 fibrils, while MFS fibroblasts assemble only few or no fibrils [10]. Unlike MFS fibroblasts, fibroblasts from a member (5010) of the WMS family described here showed abundant immunofluorescent fibrillin-1 fibrils (Figure 6b). However, these fibrils appeared to be much thinner and less bundled than control (CRL2418) fibrillin-1 fibrils. Fibroblasts from wildtype, heterozygous and homozygous WMΔ littermates also showed abundant immunofluorescent fibrillin-1 fibrils (Figure 6c). Immunofluorescence staining of WMΔ/+ and WMΔ/WMΔ cultures suggested thicker bundles of fibrillin-1 fibrils than those in wildtype cultures (Figure 6c). These results underscore the conclusion that the WMS mutation in fibrillin-1 works mechanistically differently than other FBN1 mutations that cause MFS. In addition, results suggest that the mutant WMS fibrillin-1 causes defects in fibrillin-1 fibril aggregation or bundling, but observed differences between the human WMS and the mouse WMΔ cultures cannot currently be explained.
Another potential mechanism contributing to pathogenesis of WMS is abnormal activation of TGFβ signaling. Our findings of upregulated collagen genes in the skin of WMΔ mice and increased Trichrome staining of WMΔ dermis (Figure 3) suggested increased TGFβ signaling. However, Western blotting for pSmad 2/3 showed no differences between wildtype and WMΔ skin at multiple time points (data not shown), and qPCR quantitation of TGFβ-responsive genes such as Ctgf, Pai1, and Postn also demonstrated no differences at multiple time points (data not shown). When total (Figure 7a) and active TGFβ (Figure 7b) were measured in the medium of human cultured fibroblasts, no significant difference was found between control and WMS fibroblasts. Skin samples from heterozygous and homozygous WMΔ mice of different ages were examined for the presence of myofibroblasts. Staining with antibodies specific for α-smooth muscle actin did not reveal increased numbers of myofibroblasts in mutant WMΔ mice (Figure 7c).
We also tested interactions between Latent TGFβ Binding Proteins (LTBPs) and ADAMTSL proteins, since an interaction between ADAMTSL-2 and the middle region of LTBP-1 was previously identified [26]. SPR binding studies (summarized in Table S2) showed no interaction between ADAMTSL-2 and the middle region of LTBP-1 (rL1-M [13]). However, binding between ADAMTSL-2 and the C-terminal domains of LTBP-1 present in rL1-K [13] was detected. Binding between ADAMTSL-3 and the C-terminal domains of LTBP-1 was also detected, but neither ADAMTSL-2 nor -3 interacted with LTBP-4.
WMS is considered to be clinically homogenous, even though the genetic basis of WMS is heterogeneous [27]. Both recessive and dominant forms of WMS present equally with myopia, glaucoma, cataract, short stature, brachydactyly, thick skin, and muscular build. There may be significant differences in incidence of microspherophakia, ectopia lentis and joint limitations between the recessive and dominant forms [27], but these features are also common to both. Currently, there is a single report of a mutation in FBN1 in a family with dominant WMS [3].
Results presented here identify a novel mutation in FBN1 in a family with dominant WMS, which was previously linked to FBN1 [2]. The mutation is predicted to result in fibrillin-1 molecules that lack the first 8-cysteine domain, the proline-rich region, and the adjacent EGF-like domain. Replication of this mutation in mouse Fbn1 clearly demonstrated that the mutation reproduces at least one cardinal feature of WMS—thick skin—and does not cause the clinical equivalent of MFS. Reduced long bone growth was also found in WMΔ mice, consistent with the human WMS traits of short stature and brachydactyly. However, by 5 months of age, long bone growth was normalized, perhaps reflecting differences between humans, in whom growth plate closure occurs at skeletal maturity, and rodents, whose growth plates fail to close [28], and who grow for a longer period of time than humans using cellular processes which are not active in adult humans [29]. In addition, hypermuscularity, a feature of human WMS, is also found in WMΔ mutant mice (data not shown).
Biochemical studies comparing wildtype and WMS mutant fibrillin-1 revealed that the WMS mutation abolished a specific binding site in fibrillin-1 for ADAMTSL-2, -3, -6 and papilin. Further biochemical studies suggested that specific ADAMTSL proteins may interact with ADAMTS-10 and that ADAMTS-10 binds to fibrillin-1. Our results are consistent with previous studies of papilin, the first of the ADAMTSL proteins to be described, and procollagen N-proteinase (now called ADAMTS-2), which indicated that the “papilin cassette” (homologous to the noncatalytic regions of ADAMTS enzymes) may interact with ADAMTS metalloproteinases [30]. In addition, binding between ADAMTS-10 and fibrillin-1 was recently demonstrated [25]. Therefore, we propose that a ternary complex of ADAMTSL, ADAMTS-10, and fibrillin-1 may be formed. Such a ternary complex is depicted in Figure 8, showing how ADAMTSL-3 and ADAMTS-10 may bind as a complex to fibrillin-1 in microfibrils.
Mutations in ADAMTSL2 cause geleophysic dysplasia [26]. Mutations in ADAMTS17 cause a Weill-Marchesani-like syndrome [31], and mutations in ADAMTSL4 cause autosomal recessive isolated ectopia lentis [32]. Both geleophysic dysplasia and WMS are acromelic dysplasias sharing features of short stature, brachydactyly, thick skin, limited joint mobility, and hypermuscularity. Ectopia lentis is a common feature of MFS and WMS. All together, these related genetic disorders suggest that specific ADAMTSL (at least ADAMTSL-2 and -4) and ADAMTS (ADAMTS-10 and -17) proteins modulate fibrillin-1 function in the skeleton, skin, joints, muscle, and eye. Our biochemical data also implicate ADAMTSL-3 and -6 in these pathways. Whether all members of the ADAMTSL/ADAMTS family perform similar roles in the modulation of fibrillin-1 function is unknown. However, if similar functions are performed, differences in temporal and spatial regulation of the expression of these genes could account for tissue-specific variation in these related disorders.
An in-frame deletion of 24 nucleotides was found in FBN1 to cause autosomal dominant WMS [3]. This mutation (5074_5097del) deletes 8 amino acid residues (R1692 – Y1699) from the fifth 8-cysteine domain (also called TB5) in fibrillin-1. When fibrillin-1 is modeled within microfibrils [33], the fifth 8-cysteine domain in one molecule of fibrillin-1 is close to the ADAMTSL binding site in an adjacent fibrillin-1 molecule (Figure 8). Recently, 16 novel heterozygous mutations in FBN1 causing geleophysic dysplasia or acromicric dysplasia were also identified in the fifth 8-cysteine domain [34]. WMS, geleophysic dysplasia, and acromicric dysplasia are members of the acromelic group of dysplasias with similar as well as distinctive clinical features. In our model, the clustering of fibrillin-1 domains associated with acromelic dysplasias and with binding sites for ADAMTSL proteins involved in acromelic dysplasias may specify a new microenvironment controlling thick skin and musculoskeletal growth.
It is interesting that, when fibrillin-1 is modeled within microfibrils [33], the single RGD-containing domain in fibrillin-1 is close to the ADAMTSL binding site in fibrillin-1 (Figure 8). Integrin binding to RGD sites is known to perform important roles in matrix assembly [35] and to be critically dependent on the surrounding sequences, which can silence RGD function [36]. SSKS is caused by missense mutations in FBN1 exon 37, which encodes the domain containing the RGD site [8]. Therefore, it can be speculated that SSKS mutations in FBN1 lead to diminished integrin activity. Because dermal fibrosis and the abnormal ultrastructural appearance of fibrillin microfibrils were similar in both SSKS [8] and WMS (Figure 4), it seems likely that integrin interactions with fibrillin-1 are perturbed in both SSKS and WMS. Furthermore, the proximity of the RGD-containing domain to the ADAMTSL binding site in fibrillin-1 suggests that integrins may cooperate with ADAMTSL proteins and ADAMTS-10 in modulating fibrillin microfibril [24], [25] aggregation and organization. However, the molecular mechanisms of this cooperation remain unknown.
Abnormal TGFβ signaling may play a role in these disorders, since fibrillin-1 microfibrils target and sequester the large latent TGFβ complex [13]. We found upregulated collagen gene expression and increased Trichrome staining in the skin of WMΔ mutant mice, results that are consistent with activated TGFβ signaling. In addition, molecular interactions of LTBP-1 with both ADAMTSL-2 and ADAMTSL-3 were determined, suggesting that loss of the ADAMTSL binding site in WMΔ mutant mice might render the large latent TGFβ complex more susceptible to activation. However, if activation of TGFβ signaling underlies the dermal fibrosis in WMΔ mutant mice, this activation of signaling did not manifest detectable differences in other conventional TGFβ signaling readouts (e.g., increased α-smooth muscle actin positive cells). It has been recently speculated that mechanical forces may be required to activate the latent TGFβ complex [37]. Therefore, it is possible that local changes in the fibrillin microfibril matrix could influence force-dependent activation of latent TGFβ, perhaps leading to a local increase in signaling.
Activation of TGFβ signaling has been shown for geleophysic dysplasia [26], [34], acromicric dysplasia [34], and for MFS [18], [19]. We propose that, in consort with the different tissue-specific manifestations of disease in WMS and MFS, activation of TGFβ signaling in these diseases may be limited (WMS) or more global (MFS) in scope. In MFS, the broad activation of TGFβ signaling in multiple tissues matches the pleiotropic features of the disease and the requirement for general pathogenetic mechanisms initiated by multiple different disease-causing mutations. In the case of WMS, we propose that fibrosis limited to the skin is due to the dysregulated interaction between abnormally organized microfibrils and the large latent TGFβ complex within dermal microenvironments. Our data suggest that direct interactions between ADAMTSL proteins, fibrillin-1, and LTBP-1 (Figure 8) may be dysregulated in WMS, leading to concomitant structural and signaling abnormalities within local spaces. However, the discrepancy in the data measuring TGFβ activity between our WMS fibroblast cultures and geleophysic and acromicric dysplasia fibroblast cultures [26], [34] is not yet understood. Although further investigations are required in order to determine the roles of ADAMTSL/ADAMTS-10 complexes, integrins, and LTBP-1 in the fine regulation of TGFβ signaling in WMS skin fibrosis, we conclude that these molecular pathways work locally—in microenvironments—to control skin fibrosis. While the importance of the microenvironment is appreciated in development and cancer [38], this is to our knowledge the first evidence for microenvironmental regulation by fibrillin-1.
In summary, our results suggest an improved concept of the architectural and regulatory functions of fibrillin-1. Previously, the microfibrils of elastic, distensible tissues were thought to function mechanically only as a limiting component for a cross-linked, isotropic elastin matrix. Subsequently, the attachment of LTBPs and BMPs demonstrated that fibrillin microfibrils participate in the storage and release of growth factors. Now we show that fibrillin-1 also selectively binds the metalloproteinase ADAMTS-10 and some non-enzymatic ADAMTSL proteins, enabling a clustering of these protein complexes in the vicinity of the fibrillin-1 RGD site and suggesting the potential for integrin involvement in ADAMTS/ADAMTSL/fibrillin functions. The established mutual affinity of the protein components of this cluster opens varied biochemical pathways that need to be explored in the future. The genetic evidence in humans and mice shows that perturbation of such biochemical pathways can lead to significant pathobiological consequences. In addition, the genetic evidence clearly demonstrates that fibrillin-1 microfibrils, although ubiquitous structural elements in the connective tissue space, perform local functions to support tissue microenvironments.
From the perspective of normal development and tissue homeostasis, we propose that the fibrillin microenvironment may enable two-way communication between a cell and its surroundings. The extended fibrillin fibril may function as a sensor for mechanical distortion of the matrix, signaling the cell when there is a need for additional, reinforcing structural components like collagens. The presence of large latent TGFβ complexes within the fibrillin microenvironment conveniently couples matrix mechanics with available signals for upregulation of collagens. Installation of new structural materials into pre-existing matrices likely requires remodeling enzymes like metalloproteinases. ADAMTS enzymes, localized to the fibrillin microenvironment as well, possibly with the help of ADAMTSL proteins, could serve such purposes, or they might participate in the activation of nearby latent growth factors. Understanding pathogenetic mechanisms underlying WMS and MFS will elucidate the local, fine adjustments required for growth, homeostasis, and repair.
Clinical studies were performed with informed consent and local OHSU Internal Review Board approval. All mouse work was approved by the OHSU IACUC committee.
The family pedigree shown in Figure 1 has been previously described [2]. Dermal fibroblast cultures were established from punch biopsies obtained with informed consent and local OHSU IRB approval. All participants were evaluated for myopia, glaucoma and dislocated lenses, for musculoskeletal and skin characteristics, and for cardiac or aortic disease assessed by history and/or by auscultation. This family has been followed for more than 15 years with no clinical evidence of valvular cardiac or aortic disease. A second unrelated individual, designated WMS2, was seen who had been previously diagnosed with WMS. At 18 years old, she had a history of early high myopia and presented with headaches secondary to glaucoma. She had short stature, mild brachydactyly, microspherophakia, and apparently normal joints and skin. A punch biopsy was obtained with informed consent.
Genomic DNA was extracted from cultured WMS 5016 or normal skin fibroblasts (NSF) or EDTA whole blood using standard procedures. Individual FBN1 exons were amplified by PCR of genomic DNA using intronic primers. Overlapping FBN1 cDNAs were amplified by PCR using exonic primers (for sequences see Table S3). PCR products were sequenced.
Genomic DNA was digested (using HindIII, Bsu36I, NcoI, SpeI, SspI), separated electrophoretically and transferred to a nylon membrane. The membrane was probed with a 727 bp cDNA fragment of FBN1 encompassing exons 8–12, radiolabeled with 32P-αdCTP in the presence of random and specific primers, and exposed after stringent washing to film for autoradiography.
Primers flanking FBN1 exons 9, 11, 21, and 37 were used to generate fragments of genomic NSF DNA, and the products were cloned into the pGEM-T-Easy vector (Promega). Radiolabeled probes were hybridized to total RNA of normal or WMS patient fibroblasts, followed by ribonuclease digestion to degrade unhybridized regions (RPA III kit, Ambion). Protected fragments were separated by acrylamide gel electrophoresis, and quantitated by phosphorimager (STORM, Molecular Diagnostics).
All materials used for the generation of the WMΔ mouse line originated from C57BL/6 mice (see Figure 4 for design of targeting vector). The floxed WMΔ mouse line was generated by Ozgene Pty. Ltd. (Bentley, Australia). The neomycin selection cassette was removed by breeding targeted mice to FLPe mice. Cre-mediated removal of Fbn1 exons 10–12 in all cells was accomplished by breeding floxed WMΔ mice to mice containing Cre-recombinase knocked into the Rosa26 locus (on a C57Bl/6 background). For this study, heterozygous WMΔ mice were bred to yield wildtype, heterozygous, and homozygous littermates for analyses. Genotyping was by PCR using primer pairs annealing within and outside of the deleted genomic region (for sequences see Table S4). All procedures performed on mice were approved by OHSU IACUC.
Fibrillin-1 polyclonal antibody (pAb 9543) and monoclonal antibodies (mAb) 15, 78, 201, and 69 have been previously described [9]–[12], [33]. Polyclonal antibody specific for ADAMTSL-6 was generated as described [24]. Antibody to α-smooth muscle actin was purchased from Sigma.
CRL2418, a normal dermal fibroblast cell line, was purchased from American Type Culture Collection. WMS fibroblasts were established from a punch biopsy of skin from family member 5010. Explant cultures of P4 mouse skin were established from WMΔ wildtype, heterozygous and homozygous littermates. 1 ml chamber slides were seeded at a density of 200,000 cells/ml and incubated in DMEM, including 10% fetal bovine serum, for 3 to 10 days, as indicated in the figures. Media from the 3-day incubation was collected and stored at −20°C for sandwich ELISAs. Cell layers were analyzed by immunofluorescence.
96-well ELISA plates (Corning) were coated with 100 µl of 5 µg/ml streptavidin (Pierce) and incubated overnight at 4°C. Excess streptavidin was removed by extensive washing, and 100 µl/well of 0.25 µg/ml biotinylated monoclonal antibodies (B15 or B201) were incubated at 25°C for 1 hour. After washing, wells were incubated overnight at 4°C with culture medium samples (from 3-day chamber slide cultures of WMS or control fibroblasts). In addition, serially diluted protein standards (rF11) were applied separately to wells coated with biotinylated antibodies and incubated overnight at 4°C. Unbound proteins were removed by washing, and alkaline phosphatase-conjugated monoclonal antibodies (AP201 0.05 µg/ml or AP78 0.5 µg/ml) were incubated in the wells for 1 hour at 25°C. Invitrogen's ELISA amplification system was used for colorimetric detection, according to the manufacturer's protocol. Absorbance was recorded using a Molecular Devices Emax plate spectrophotometer and was then converted to µg/ml, according to standard curve values. Calculations to determine concentration were performed on Excel software.
Skin was obtained by punch biopsy from a son of family member 5010, following informed consent. Skin was also obtained from WMΔ wildtype and mutant mouse littermates in accordance with OHSU approved IACUC procedures. Immunofluorescence of skin as well as cultured fibroblasts was performed as previously described [10], [11].
Histology was performed by the OHSU Histology Core, using standard procedures for Hematoxylin and Eosin and Masson Trichrome stains (Sigma, St. Louis, MO).
For μCT analyses, mice were sacrificed at specified time points. μCT measurements and analyses were performed with a Scanco μCT 35 (Scanco Medical, Basserdorf, Switzerland) scanner, according to the manufacturer's instructions.
qPCR using RNA from WMΔ and wildtype control mouse skin was performed as previously described [11]. Primers for mouse Col1A1, Col1A2, and Col3A1 were purchased from SABiosciences (Frederick, MD). The primers for mouse Periostin (Postn; forward: 5′-catcttcctcagcctccttg-3′; reverse: 5′-tcagaagctccctttcttcg-3′), Plasminogen activator inhibitor-1 (Pai1; forward: 5′-ctttacccctccgagaatcc-3′; reverse: 5′-gacacgccatagggagagaa-3′), and Connective tissue growth factor (Ctgf; forward: 5′-ctgcctaccgactggaagac-3′; reverse: 5′-ttggtaactcgggtggagat-3′) were individually designed and tested for amplification efficiency.
Immunoelectron microscopy of tissues from WMΔ mouse littermates was performed as described [11]. Tissues were labeled en bloc with anti-fibrillin-1 (pAb 9543) followed by 5 nm secondary gold conjugated antibodies (Amersham Biosciences). Aligned tilt series were acquired from 500 nm thick sections as described [11].
Expression vectors carrying full length human LTBP-4S and LTBP-1S were kindly provided by Dr. Jorma Keski-Oja and Dr. Daniel Rifkin. The rF90WMΔ and rF84WMΔ expression constructs were cloned from WMS fibroblast cDNA. Full-length ADAMTSL1 was obtained from human fibroblast cDNA. For cloning of ADAMTSL2, a mouse full length cDNA clone (ID RIKEN cDNA F83011122) was obtained. Constructs for ADAMTSL3 were made using a clone (RIKEN) and mouse lung cDNA. Constructs for Papilin were cloned from mouse fibroblast cDNA. A full length human ADAMTS10 clone (SC309981) was purchased from Origene, and mutations in this clone were corrected.
Generation of recombinant polypeptides representing fragments of LTBP-1 was previously described [13]. The generation of rF90 was described before [16]. All fibrillin and ADAMTSL expression constructs were transfected into 293/EBNA cells for protein expression. All proteins were purified using metal ion affinity chromatography. Protein domain boundaries for the constructs are depicted in Figure 1 and Figure S3.
Binding analyses were performed using a BIAcoreX (BIAcore AB, Uppsala, Sweden). Recombinant full length ADAMTSL-1, -2, LTBP-1, -4, and polypeptides ADAMTS-10 C-term, rL1K, rLM, rLN, rF6, rF90, and rF90WMΔ were covalently coupled to CM5 sensor chips (research grade) using the amine coupling kit following the manufacturer's instructions (BIAcore AB). Binding assays were performed at 25°C in 10 mM Hepes buffer, pH 7.4, containing 0.15 M NaCl, 3 mM EDTA, and 0.005% (v/v) P20 surfactant (HBS-EP buffer, BIAcore AB). Kinetic constants were calculated by nonlinear fitting of association and dissociation curves (BIAevaluation 3.0 software). Equilibrium dissociation constants (KD) were then calculated as the ratio of kd/ka.
Cell culture media (1 ml) from stably transfected 293/EBNA cells expressing ADAMTS-10 with a C-terminal His6-tag, and media from untransfected EBNA cells as a control were adjusted to 20 mM Tris pH 7.8, 5 mM imidazole and incubated for 1 h with 5–20 µg of rF11 (N-terminal half of fibrillin-1, without a His6-tag). Subsequently, 50 µl of a 50% Ni-NTA slurry in water was added and incubated for 2 hs. The resin was washed and boiled in 50 µl 1× SDS loading buffer. Eluted proteins were subjected to SDS-PAGE followed by immunoblotting with polyclonal anti fibrillin-1 antibody 9543.
The quantity of TGF-β1 in 100 µl culture medium from confluent fibroblasts (200,000 cells/ml grown for 72 h in 1 ml chamber slides) was determined using the TGF-β1 EMax Immnunoassay kit (Promega). WMS and control fibroblasts were utilized.
Prism 5.02 for Windows (GraphPad, San Diego, CA) was used to perform One-way Analysis of Variance (1-way ANOVA) followed by post-test analysis with Tukey's multiple comparison test. p-values<0.05 were considered significant.
|
10.1371/journal.ppat.1000106 | NK Cell–Like Behavior of Vα14i NK T Cells during MCMV Infection | Immunity to the murine cytomegalovirus (MCMV) is critically dependent on the innate response for initial containment of viral replication, resolution of active infection, and proper induction of the adaptive phase of the anti-viral response. In contrast to NK cells, the Vα14 invariant natural killer T cell response to MCMV has not been examined. We found that Vα14i NK T cells become activated and produce significant levels of IFN-γ, but do not proliferate or produce IL-4 following MCMV infection. In vivo treatment with an anti-CD1d mAb and adoptive transfer of Vα14i NK T cells into MCMV-infected CD1d−/− mice demonstrate that CD1d is dispensable for Vα14i NK T cell activation. In contrast, both IFN-α/β and IL-12 are required for optimal activation. Vα14i NK T cell–derived IFN-γ is partially dependent on IFN-α/β but highly dependent on IL-12. Vα14i NK T cells contribute to the immune response to MCMV and amplify NK cell–derived IFN-γ. Importantly, mortality is increased in CD1d−/− mice in response to high dose MCMV infection when compared to heterozygote littermate controls. Collectively, these findings illustrate the plasticity of Vα14i NK T cells that act as effector T cells during bacterial infection, but have NK cell–like behavior during the innate immune response to MCMV infection.
| An efficient immune response to viral infection requires both innate and adaptive immune cells. Natural killer (NK) cells are a critical innate cellular component of the immune response to murine cytomegalovirus (MCMV). Natural killer T (NK T) cells are non-classical T cells that have the potential to bridge the two arms of the immune system. However, the contribution of NK T cells to the anti-viral immune response has not been extensively studied. In the absence of additional stimuli, NK T cells actively participate in the immune response to MCMV infection. Interestingly, in contrast to their response to bacteria, we demonstrate that only the innate NK T cell arm is activated during viral infection while the adaptive branch, TCR engagement by CD1d, is dispensable. NK T cells display signs of activation in response to viral infection, increased expression of CD25, a rapid decrease in cell number, and production of the anti-viral cytokine IFN-γ. The NK T cell response to MCMV also influences the NK cell activity and the inflammatory cytokine profiles. Understanding the physiological function of these unique T cells in the context of infection will aid in the development of novel therapeutic and preventive treatments for viral infections.
| The β-herpes murine cytomegalovirus (MCMV) is a well-characterized model of viral infection that results in a non-replicative, chronic infection of immune-competent animals [1]. MCMV is a cytopathic virus that is known to readily infect peritoneal macrophages, dendritic cells (DC) and hepatocytes, inducing significant pathology in both the spleen and the liver [2]–[5]. The acute response to this virus is dependent on natural killer (NK) cell cytotoxicity and IFN-γ production, as animals deficient in perforin or IFN-γ signaling rapidly succumb to infection [4], [6]–[10].
The hepatic immune environment is greatly influenced by the resident cellular subsets and has been shown to be primarily tolerogenic [11],[12]. The major hepatic lymphocyte population in mice is a distinct family of T cells, Vα14 invariant NK T (Vα14i NK T) cells [13],[14]. Vα14i NK T cells are innate lymphocytes that display an effector memory phenotype, expressing CD69 and CD44 constitutively [15]. They are uniquely capable of rapidly producing TH1 and TH2 cytokines in response to antigenic stimulation [16]. The Vα14i NK T cell repertoire is highly restricted, characterized by a Vα14-Jα18 rearrangement with an invariant junction preferentially associated with Vβ8.2, Vβ7, or Vβ2 [17],[18]. In response to the ligand α-galactosylceramide (α-GalCer), Vα14i NK T cells interact with and activate other immune cells including NK cells, CD8+ T cells, DCs, and macrophages [16]. This immune cell cross-talk is facilitated by direct cell-cell contact and via cytokine release [19]–[22].
Much of the functional significance of Vα14i NK T cell activation in the context of viral infection has been provided by activating the compartment prior to or concomitantly with viral introduction in animal models [23]–[25]. Although this method examines the potential contribution of activated Vα14i NK T cells, it does not examine the physiological function of these T cells in response to viral infection without exogenous stimuli. In the context of other microbial infections, the evidence for direct Vα14i NK T cell involvement is mixed, often being dependent on the type of pathogen [26]–[32].
However, there is indirect evidence that Vα14i NK T cells play a role in anti-viral immune responses. A number of groups have clearly shown that the expression of the antigen-presenting molecule CD1d is often down-regulated by viruses in a myriad of ways, including protein degradation, alterations in transcription, or endosomal sequestration [33]–[35]. Vα14i NK T cells have also been shown to be preferential targets of infection and virus-induced cell death [36],[37]. This indicates that Vα14i NK T cells may have a potential role in the anti-viral response and it is advantageous for the pathogen to prevent their activation.
To directly assess the role of Vα14i NK T cells in the innate anti-viral response, their activation status was examined following MCMV infection in vivo. We found that Vα14i NK T cells up-regulate the high affinity IL-2 receptor-α, CD25, produce IFN-γ, but do not undergo proliferation. Importantly, we demonstrate that CD1d is dispensable for Vα14i NK T cell activation and cytokine release in the context of MCMV. However, IFN-α/β and IL-12 are both partially required for optimal activation of the Vα14i NK T cells in response to infection. We also show that in the absence of α-GalCer treatment, Vα14i NK T cells contribute significantly to the overall cytokine response and amplify NK cell-derived IFN-γ production. Collectively, our findings demonstrate a role for the NK T cells in innate sensing of viral pathogens in an unanticipated NK cell-like manner.
Inbred C57BL/6 and B6.SJL-Ptprca/BoAiTac mice were purchased from Taconic Laboratory (Hudson, NY). B6.IL-12p40−/− mice were purchased from the Jackson Laboratory (Bar Harbor, ME). B6.CD1d−/− mice (a generous gift from Dr. L. Van Kaer, Vanderbilt University, Nashville, TN) and B6.Jα18−/− mice (kindly provided by Dr. M. Taniguchi, Riken Research Center for Allergy and Immunology, Yokohoma, Japan) were bred, crossed to the B6 (>10 generations) to generate wild-type, heterozygous, and knock-out littermates. Female IFN-α/βR1−/− mice originally generated on the 129.SvEv background and backcrossed on to the C57BL/6 background were kindly provided by Dr. M. Aguet [38] and bred in our facility. All mice, except B6 mice, were bred in pathogen-free breeding facilities at Brown University (Providence, RI). All experiments were conducted in accordance with institutional guidelines for animal care.
Stocks of Smith strain MCMV salivary gland extracts were prepared as previously described [39]. Infections were initiated on day 0 with 5×104 plaque-forming units (PFU), administered via i.p. injection. For survival studies, 3×105 PFU were administered via i.p. injection. For antibody-blocking experiments, mice received blocking CD1d mAb (0.3 mg; clone 1B1; BD Pharmingen) or rat IgG control Abs in PBS at the time of the infection.
To obtain splenic lymphocytes, spleens were minced, passed through nylon mesh (Tetko, Kansas City, MO), washed once in 2% PBS-serum and cell suspensions were layered on Lympholyte-M (Cedarlane Laboratories Ltd., Canada). Hepatic lymphocytes were prepared by mincing and passage through a 70 mm nylon cell strainer (Falcon, Franklin Lakes, NJ). After washing 3 times in 2% PBS-serum, cell suspensions were layered on a two-step discontinuous Percoll gradient (Pharmacia Fine Chemicals, Piscataway, NJ). Splenocytes and hepatic lymphocytes were collected after centrifugation for 20 min at 900×g.
CD19-FITC, TCRβ-FITC, CD11b-FITC, CD11c-FITC, NK1.1-PE, CD1d-PE, B220-PerCP-Cy5, KLRG1-allophycocyanin, CD25-APC, and TCRβ-allophycocyanin were all purchased from eBioScience (San Diego, CA). NK1.1-PerCp-Cy5.5, CD11b-PerCp, CD4-PerCp, CD8-PerCp, CD11c-allophycocyanin, B220-allophycocyanin, IFN-γ-allophycocyanin and isotype control were purchased from BD Pharmingen (San Diego, CA). For NK T cell identification, CD1d tetramers were obtained from the National Institute of Allergy and Infectious Disease MHC Tetramer Core Facility at Emory University (Atlanta, GA). Additionally, the following mAbs were purchased from BD Pharmingen and used for ELISA: IFN-γ mAbs (clone R4-6A2, and clone XMG1.2), IL-4 mAbs (clone 4B11 and BVD6-24G2), IL-2 mAbs (purified JES6-N37-1A12 and biotinylated JES6-5H4) and streptavidin-peroxidase.
Hepatic lymphocytes were isolated as described above from congenic C57BL/6.SJL mice. For enrichment of hepatic NK T cells, cells were first depleted of CD8+, CD11c+, CD11b+, and CD19+ cells using the AutoMACS (Miltenyi Biotec) as instructed by the manufacturer. 5–8×106 cells were transferred via tail vein injection into Jα18−/− or CD1d−/− mice. For NK T cell positive selection hepatic lymphocytes were stained with anti-NK1.1 and anti-CD5 mAbs or with anti-NK1.1 and CD1d tetramer and sorted using a FACSAria (BD Biosciences). At the time of transfer (1×106 cells per mouse), mice were infected with 5×104 pfu MCMV. At 1.5 days post-infection, animals were sacrificed and the donor population analyzed for IFN-γ production.
For all serum-based measurements, blood was collected via cardiac puncture. Serum was separated from the cellular fraction by centrifugation at 14,000 rpm at 4°C for 30 minutes. Serum levels of cytokines were measured by ELISA or using the cytometric bead array (CBA) mouse inflammation kit (BD Pharmingen).
Following lymphocyte isolation, cells were suspended in PBS containing 2% FCS. Cells were then incubated with 2.4G2 anti-Fc receptor mAb and stained with indicated antibodies. Cells were then fixed in 2% paraformaldehyde in PBS. Intracellular staining for IFN-γ protein was performed using the Cytofix/Cytoperm kit (BD PharMingen). Depending on the experiment and the tissue, 2.5×105–1×106, events were collected on a FACSCalibur or FACSAria. The data were analyzed using CellQuest software or Diva software (Becton Dickinson, Franklin Lakes, NJ).
Statistical significance, designated as a p-value ≤0.05, was determined by paired, 2-tailed Student's T-test.
It is well documented that NK cells are necessary for the innate anti-viral immune response to MCMV infection [6],[40]. However, it is unclear if naïve Vα14i NK T cells also participate in this innate immune response. To address this issue, wild type B6 mice were infected for 20–40 hrs and the activation status of Vα14i NK T cells was examined in the liver and spleen. In both tissues, Vα14i NK T cells display signs of activation, a decrease in NK T cell numbers (Fig. 1A), and CD25 up-regulation by 20 hours post-infection (Fig. 1B and data not shown).
CD1d dependent Ag recognition by Vα14i NK T cells induces their expansion [41]. Additionally, Vα14i NK T cells have been shown to proliferate in response to infection with LPS negative bacteria [31],[42]. However, in the context of MCMV infection, the Vα14i NK T cell compartment does not expand in either number or frequency (Fig. 1D), even at the peak of activation as assessed by CD25 expression (Fig. 1C). We also performed an intra-cellular staining for TCR at different days post MCMV infection. We found that most of the cells were double positive for intracellular and cell surface TCR, ruling out a possible lack of detection of the Vα14i NK T cells due to TCR internalization (data not shown). In contrast to Vα14i NK T cells, NK cells expand during MCMV infection (Fig. 1D). Furthermore, the percentage of CD25+ Vα14i NK T cells rapidly declines in comparison to the protracted decrease in the percent of NK cells positive for the terminal maturation marker, KLRG1 (data not shown).
Vα14i NK T cells produce IFN-γ as early as 30 hours post-infection (data not shown), peaking at day 1.5 post-infection (Fig. 2A). At this time point, the frequency of IFN-γ+ Vα14i NK T cells is comparable to the frequency of IFN-γ+ NK cells in both spleen and liver (Fig. 2B). In the spleen, despite a similar frequency, the number of IFN-γ+ Vα14i NK T cells is lower than the number of IFN-γ+ NK cells. However, the number of IFN-γ+ Vα14i NK T cells is similar to the number of IFN-γ+ NK cell in the liver at day 1.5 post-infection (Fig. 2C). This indicates that these two subsets of cells contribute equally to the overall amount of IFN-γ in the liver. Notably, Vα14i NK T cells do not produce detectable amounts of IL-4 during MCMV infection in either tissue (data not shown).
In order to investigate whether MCMV induced activation of Vα14i NK T cells requires CD1d, B6 mice were treated with a blocking CD1d mAb or control antibody and infected with MCMV. On day 1.5 post-infection, the percentage of hepatic IFN-γ+ Vα14i NK T cells in mice that received the anti-CD1d mAb or control IgG was comparable (Fig. 3A). Similar results were observed in the spleen (data not shown). To directly assess the contribution of CD1d-mediated Ag presentation to MCMV-induced activation and cytokine production from Vα14i NK T cells, adoptive transfer experiments were performed. Negatively selected (purity >70%) or positively selected hepatic Vα14i NK T cells (purity >95%) from congenic wild-type B6.SJL mice were adoptively transferred into CD1d−/− or Jα18−/− deficient hosts. The recipient mice were simultaneously infected with MCMV for 1.5 days and the percentage of IFN-γ+ Vα14i NK T cells was determined. Regardless of the host expression of CD1d, donor Vα14i NK T cells produced similar amounts of IFN-γ following MCMV infection in vivo (Fig. 3B & 3C). Taken together, the results indicate that CD1d is dispensable during MCMV induced activation of Vα14i NK T cells.
High levels of IFN-α/β and bioactive IL-12 characterize the innate immune response to MCMV infection in vivo [43]. In the absence of either cytokine, the innate anti-viral response is fatally impaired [39],[44]. Here, MCMV infection of IFN-α/βR1−/− and IL-12p40−/− mice further reveals that the activation of Vα14i NK T cells at both 20 and 40 hours post-infection is independent of IL-12 and IFN-α/β, as assessed by the percentage of NK T cells (Fig. 4A & 4B) and CD25 expression (Fig. 4C). However, Vα14i NK T cell-derived IFN-γ is highly dependent on IL-12 in the liver (Fig. 5) and spleen (data not shown), similar to NK cells (Fig. 5). Notably, the Vα14i NK T cell-derived IFN-γ response is reduced by ∼50% in MCMV infected IFN-α/βR1−/− mice, further mimicking NK cell dynamics (Fig. 5).
Activated Vα14i NK T cells interact with and activate other immune cells such as NK cells, which subsequently produce cytokines [19],[20],[45]. To investigate the downstream consequences of Vα14i NK T cell absence, we measured both NK cell-derived IFN-γ and serum inflammatory cytokines in infected NK T cell deficient mice. The activation of NK cells in the spleen, as determined by IFN-γ production, was significantly reduced in CD1d−/− mice when compared to heterozygous littermates (Fig. 6A). Similarly, although not significant, a reproducible reduction of NK cell IFN-γ was also observed in MCMV infected Jα18−/− mice compared to Jα18−/+ littermates in five independent experiments (Fig. 6A). Interestingly, an overall reduction of the inflammatory cytokine profile (IL-12, IFN-γ and TNF-α) was seen in the blood of Jα18−/− animals in comparison to heterozygous littermate controls at 1.5 days post-MCMV infection in vivo (Fig. 6B). Likewise, inflammatory cytokines were diminished in CD1d−/− mice compared to CD1d+/− littermates. However, in the latter case, while IL-12 and TNF-α were reproducibly decreased, only IFN-γ was reduced significantly. Notably, IL-4 and IL-10 were not detectable in the serum of infected animals.
To address whether Vα14i NK T cells and/or CD1d participate in the early control of MCMV infection, CD1d−/− and Jα18−/− animals, as well as littermate controls were used in survival studies with high dose MCMV infection. CD1d−/− mice were more susceptible than CD1d+/− mice, as only 50% of the CD1d−/− mice lived beyond day 15. While Jα18−/− animals were not more susceptible than their littermate, heterozygous controls (Fig. 6), they were significantly more susceptible than wild-type B6 mice from outside vendors (not shown). Taken together, these results demonstrate that Vα14i NK T cells influence NK cell activity, the inflammatory cytokine profiles, and that both Vα14i NK T cells and other CD1d restricted T cells are necessary for an optimal immune response to MCMV.
The function of NK cells in anti-viral immunity has been documented; however, evidence for a direct Vα14i NK T cell role has not been examined extensively. The results presented in this report show that Vα14i NK T cells sense MCMV infection in vivo, without exogenous stimuli, such as α-GalCer. However, in contrast to bacterial infection, we provide evidence that MCMV induced Vα14i NK T cell activation is TCR independent.
Vα14i NK T cells can be activated directly by agonist glycolipids presented by CD1d. For instance, α-GalCer immunization of B6 mice leads to IL-12 independent activation of Vα14i NK T cells [16]. In this case, Vα14i NK T cells release copious amount of IL-4 and IFN-γ and subsequently proliferate. Gram-negative LPS-negative α-proteobacteria, such as Sphingomonas, Ehrlichia, Rickettsia, and Borrelia, express such agonist lipids and can directly activate Vα14i NK T cells [31],[32],[42],[46]. Bacteria that do not express agonist glycolipids have been reported to activate Vα14i NK T cells through up-regulation of self glycolipids and/or IL-12 production through recognition of endogenous lysosomal glycosphingolipids, such as iGb3, presented by LPS-activated dendritic cells [31],[47]. Therefore, there are two major mechanisms for the activation of Vα14i NK T cells against bacteria either via cognate Ag or via self-Ag with APC derived cytokines [48]. Using MCMV infection in vivo, we now demonstrate a novel activation pathway for Vα14i NK T cells, mediated principally by inflammatory cytokines.
MCMV induced activation of Vα14i NK T cells clearly differs from the two mechanisms described in response to bacterial infection. First, as opposed to α-GalCer administration [41],[49] and α-proteobacteria infection [31], Vα14i NK T cells do not proliferate nor produce IL-4 following MCMV-induced activation. Second, while IL-12 is not required for optimal stimulation of Vα14i NK T cells in response to α-GalCer or sphingomonas-derived glycolipids [20],[42], here we show that Vα14i NK T cell cytokine production is impaired in IL-12 deficient animals in response to MCMV infection. Finally, in vivo CD1d blocking experiments and adoptive transfer of Vα14i NK T cells into CD1d−/− mice demonstrates that CD1d is dispensable for MCMV induced activation of these lymphocytes. It should be noted that LPS-induced Vα14i NK T cell-derived IFN-γ in vitro does not require CD1d-mediated Ag presentation, instead exposure to IL-12 and IL-18 is sufficient to activate these cells [50].
These data raise the question of why do bacteria such as Salmonella typhimurium and Staphylococcus aureus activate Vα14i NK T cells in both an IL-12 and CD1d dependent manner, while MCMV induced activation of Vα14i NK T cells is CD1d independent? There are several non-mutually exclusive possibilities that could explain this apparent discrepancy. First, viruses unlike bacteria, do not encode enzymatic machinery for lipid synthesis. Second, the peak of the cytokine response to MCMV occurs relatively early when compared to bacteria, allowing for possible Vα14i NK T cell activation to occur prior to cytokine-driven self-Ag activation. Third, while Gram negative bacteria cytokine-driven self-Ag activation of Vα14i NK T cells was demonstrated in vitro using bone marrow derived DCs [31],[47], it has been recently demonstrated that plasmacytoid dendritic cells (pDCs) are the quasi-exclusive source of IFN-α/β, IL-12 and TNF-α early during MCMV infection [51]. It is therefore possible that depending on the pathogen and the source and/or phenotype of recruited CD1d+ DCs may lead to differential activation of Vα14i NK T cells.
pDCs and dendritic cells recognize MCMV through TLR9, an essential component of the innate immune defense against MCMV. Tabeta et al have shown that the Vα14i NK T cell response to MCMV is impaired in TLR9−/− mice [52]. Interestingly, the serum level of both IFN-α and IL-12 is reduced in TLR9−/− mice following MCMV infection [52],[53], supporting our findings that the absence of these cytokines impairs the Vα14i NK T cell response to MCMV. However, activation of Vα14i NK T cells by TLR9-stimulated dendritic cells was recently shown to be CD1d dependent [54]. The latter study was performed in vitro using BMDCs grown for 14 days prior to being pulsed with CpG for 16 hours. This procedure clearly differs from MCMV infection in vivo where the peak of the Vα14i NK T cell response is at 1.5 days post-infection. It is possible that some pathogen-derived products such as CpG may increase endogenous glycolipid presentation during the anti-viral response but that this process is not yet initiated at the peak of the innate response to MCMV.
In the context of MCMV infection, we failed to detect an expansion of the Vα14i NK T cells. Instead, there is a gradual loss of these cells in the liver and spleen following infection. Presumably, the lack of TCR engagement by CD1d during MCMV infection promotes the activation of Vα14i NK T cells that release cytokines but do not proliferate and subsequently die. It is also possible that Vα14i NK T cells preferentially undergo virus-induced apoptosis similarly to what has been reported during the anti-viral response against lymphocytic choriomeningitis virus infection in vivo [36].
The early immune response to MCMV is characterized by the production of high levels of inflammatory cytokines [51],[55]. Type I IFNs, which are critical for anti-viral immunity, can be detected very early following MCMV infection and mediate the proliferation and survival of activated lymphocytes [56]. Additionally, the classical TH1-promoting cytokine, IL-12, is also produced early and is necessary for NK cell-derived IFN-γ [57]. Infection of mice deficient in either IFN-α/β signaling or bioactive IL-12 clearly demonstrates that neither cytokine alone is sufficient to mediate an optimal IFN-γ response from Vα14i NK T cells in response to MCMV. IFN-α/β is thought to negatively regulate the production of IFN-γ via inhibition of IL-12 thus ensuring that IFN-γ does not prematurely inhibit proliferation [58]. The timing of IL-12 production may be critical for Vα14i NK T cells as MCMV infected IFN-α/βR1−/− mice produce high levels of IL-12 [56], yet we show that Vα14i NK T cell-derived IFN-γ is impaired in IFN-α/βR1−/− mice. It is also currently unclear if IFN-α/β acts directly on this innate T cell population in the context of MCMV infection. These issues warrant further inquiry.
Vα14i NK T cells are widely appreciated for their rapid cytokine production and ability to interact with and activate both innate and adaptive immune cells [15]. The cross-talk between Vα14i NK T cells and NK cells in the context of α-GalCer-mediated stimulation requires IFN-γ and IL-12 production to promote optimal NK cell activation [19],[20]. Vα14i NK T cell-mediated activation of NK cells has been shown to be required for anti-tumor immunity [59],[60] but not for viral infection. We show that in the absence of the Vα14i NK T cell population, the NK cell response to MCMV in vivo is impaired. Notably, Vα14i NK T cells also mediate activation and maturation of DCs and macrophages [61],[62], cells critical for the induction of the anti-viral immune response via production of high levels of key inflammatory cytokines such as IFN-α/β and IL-12 [62]. We are currently investigating the possibility that the function of these subsets in response to MCMV infection may also be altered in the absence of Vα14i NK T cells.
Our data suggest that Vα14i NK T cells contribute to the overall immune response to MCMV. However, it has been shown that at a low infecting dose, Jα18−/− mice and B6 wild-type control animals have equivalent viral titers [25]. Using Vα14i NK T cell deficient mice and littermate controls, we examined the pathological impact of Vα14i NK T cell absence during high dose MCMV challenge. While both Jα18−/− and CD1d−/− animals are more susceptible than wild-type B6 mice from outside vendors, only CD1d−/− mice are less resistant than heterozygous littermate control mice to high dose MCMV infection. This suggests that Vα14i NK T cells as well as other CD1d restricted T cells are required for an optimal immune response to MCMV.
Collectively, the findings presented in this report indicate that Vα14i NK T cells actively participate in the innate immune response to MCMV in vivo and directly impact the quality of the immune response. However, the mechanism of their activation differs from bacterial induced activation and in this case, NK T cell functions mirror NK cell functions. These data define a previously unappreciated role for CD1d restricted T cells in anti-viral immunity and provide additional insight into the affect of innate immune manipulation on the overall outcome of an immune response. |
10.1371/journal.ppat.1002641 | Matrix Metalloprotease 9 Mediates Neutrophil Migration into the Airways in Response to Influenza Virus-Induced Toll-Like Receptor Signaling | The early inflammatory response to influenza virus infection contributes to severe lung disease and continues to pose a serious threat to human health. The mechanisms by which neutrophils gain entry to the respiratory tract and their role during pathogenesis remain unclear. Here, we report that neutrophils significantly contributed to morbidity in a pathological mouse model of influenza virus infection. Using extensive immunohistochemistry, bone marrow transfers, and depletion studies, we identified neutrophils as the predominant pulmonary cellular source of the gelatinase matrix metalloprotease (MMP) 9, which is capable of digesting the extracellular matrix. Furthermore, infection of MMP9-deficient mice showed that MMP9 was functionally required for neutrophil migration and control of viral replication in the respiratory tract. Although MMP9 release was toll-like receptor (TLR) signaling-dependent, MyD88-mediated signals in non-hematopoietic cells, rather than neutrophil TLRs themselves, were important for neutrophil migration. These results were extended using multiplex analyses of inflammatory mediators to show that neutrophil chemotactic factor, CCL3, and TNFα were reduced in the Myd88−/− airways. Furthermore, TNFα induced MMP9 secretion by neutrophils and blocking TNFα in vivo reduced neutrophil recruitment after infection. Innate recognition of influenza virus therefore provides the mechanisms to induce recruitment of neutrophils through chemokines and to enable their motility within the tissue via MMP9-mediated cleavage of the basement membrane. Our results demonstrate a previously unknown contribution of MMP9 to influenza virus pathogenesis by mediating excessive neutrophil migration into the respiratory tract in response to viral replication that could be exploited for therapeutic purposes.
| Influenza-associated morbidity and mortality due to yearly epidemics and sporadic, devastating pandemics are a significant health and economic burden. Severe complications arising from highly virulent viruses are associated with rapid, massive inflammatory cell infiltration. Although neutrophils are the predominant cell population recruited to the lung in response to pandemic influenza viruses, the mechanisms by which they gain entry to the respiratory tract remain unclear. In this study, we show a previously unknown contribution of MMP9 to influenza pathogenesis by mediating excessive neutrophil migration into the lung, which not only controls viral replication, but also contributes to morbidity. The in vivo relevance of MMP9-derived enzymatic activity in neutrophils is controversial and understudied, but our data provide new evidence that innate recognition of influenza virus attracts neutrophils that secrete MMP9, which enables them to traverse the basement membrane of the lung by digesting the extracellular matrix. The dichotomy of MMP9 function in immunity versus pathology provides real challenges for targeting MMP9 for therapeutic purposes. Nevertheless, finding the balance to modulate neutrophil numbers following influenza virus infection will allow for innate immunity to be boosted whilst preventing pathology associated with pandemic strains.
| Influenza viruses are highly contagious and cause an average of 226,000 hospitalizations and 36,000 deaths in yearly epidemics [1]. The emergence of immunologically distinct influenza viruses to which a whole population is susceptible has resulted in sporadic, devastating pandemics, such as the recent outbreak of ‘swine flu’ that infected an estimated 61 million people and killed 12,470 [2]. Furthermore, there is growing concern that avian influenza viruses, which have a significant mortality rate, could mutate to allow easy human-to-human transmission [3]. Severe complications arising from pandemic influenza or highly pathogenic avian viruses are often associated with rapid, massive inflammatory cell infiltration [4]. Inflammatory cytokines and chemokines have been associated as their dysregulation correlates strongly with high viral load and pathology [5]. Following infection, influenza virus replication occurs in the epithelial cells of the respiratory tract where innate immune responses are initiated by recognition of the virus through the inflammasome, which consists of toll-like receptors (TLRs), retinoic acid-inducible gene-I-like receptors, and NOD-like receptors [6]. For example, viral double-stranded RNA and single-stranded RNA are recognized by TLR3 and TLR7, respectively [7], [8]. TLR3/7 signaling induces an inflammatory response that promotes a cascade of immune processes that regulate cellular recruitment and function, including the induction of cytokines and chemokines. Whereas this process results in full maturation of dendritic cells (DCs), activation and recruitment of antigen-specific T cells, and adequate anti-influenza humoral responses [9], [10], aberrant signaling has been suggested to play a key role in mediating lung pathology that is characterized by excess inflammation and pulmonary destruction [4].
MMPs are a family of proteolytic enzymes that are involved in remodeling the extracellular matrix (ECM) under both physiological and pathological conditions [11]. They can be produced by a range of cells in the respiratory tract, where they mediate wound healing, airway remodeling, and cell trafficking [12]. As such, MMPs play an important role in immunity, and their proteolytic activity can also directly dampen the inflammatory potential by downregulating cytokine and chemokine function [13]. However, excessive responses may contribute to pathology and MMP activity has been implicated in a variety of pulmonary diseases. Since MMPs have the potential to cause significant host damage, their production and function are tightly regulated. For example, MMPs are rarely stored inside cells and require gene transcription before secretion, the exceptions being neutrophil MMP-8 and -9, which are constitutively expressed in granules. Active MMPs can degrade all components of the ECM and are divided into subclasses based on their substrate specificity. Collectively known as the gelatinases, MMP2 and MMP9 cleave denatured collagens (gelatins) and type IV collagen present in basement membranes [13]. Although implicated in a range of pulmonary diseases, a detailed understanding of the involvement of gelatinases in disease pathology after viral infection is lacking. MMP9 is not expressed in healthy lungs, but may be released under inflammatory conditions (such as infections and neoplastic diseases) by macrophages, mast cells, lymphocytes, and neutrophils. Although neutrophils are the predominant leukocyte population to be recruited to the lung and appear early in the immune response to pandemic influenza viruses [14], the mechanisms by which they gain entry to the respiratory tract and their role during pathogenesis remain unclear.
In this study, we sought to identify the role of MMP9 in the recruitment of neutrophils into the respiratory tract after influenza virus infection. By increasing the infective dose, we were able to mimic many of the hallmarks of more virulent “pandemic” infections by virtue of inducing increased morbidity, inflammation, and lung pathology. High dose infection led to increased neutrophil recruitment to the lungs and airways, which contributed significantly to morbidity. We identified neutrophils as the predominant source of pulmonary MMP9. Importantly, MMP9 was functionally required for neutrophil migration into the lung in response to infection and for control of viral replication. Although MMP9 release from neutrophils was TLR-dependent, MyD88-mediated signals in non-hematopoietic cells, rather than intrinsic induction, were important for neutrophil recruitment by inducing the neutrophil chemoattractant CCL3. In addition, MyD88 signaling-induced tumor necrosis factor (TNF)α contributed to MMP9 release from neutrophils. Our data demonstrate a previously unknown role of MMP9 to influenza virus pathogenesis by contributing to excessive neutrophil migration into the respiratory tract in response to TLR-induced chemotactic factors.
Infection of humans with virulent viruses is associated with high viral titers and increased inflammation [5]. To determine the effect of infective dose on cytokine/chemokine levels and cellular recruitment, mice were infected with different doses of Influenza virus A PR/8/34 (PR8): 125, 1250, or 12500 50% egg infectious doses (EID50). Morbidity, measured by weight loss, correlated with the infective dose (Figure 1A). Mice began losing weight as early as 3 days after infection with 12500 EID50 and progressively lost more until they had to be sacrificed in accordance with animal welfare guidelines. Although delayed, mice receiving the intermediate dose (1250 EID50) lost weight, but recovered with the appearance of adaptive immunity. Depending on the infective dose, mice had increased cytokine/chemokine levels in the airways 6 days after infection as measured by multiplex analysis of the broncheo-alveolar lavage (BAL) (Figure 1B). Inflammatory cytokines, such as IFNγ, interleukin (IL)-6 and TNFα, and the chemokines, CCL2 and CCL5, were highly upregulated in the respiratory tract after infection with 12500 EID50. The increased secretion of inflammatory mediators was associated with increased cellular infiltrates in the lung 6 days after infection (Figure 1C). Whereas uninfected lungs had clear airspaces, cellular infiltrates and bronchoconstriction were observed after infection with 125 EID50. Increasing the dose tenfold resulted in complete obliteration of the normal lung architecture in many foci of infiltration (Figure 1C, far right). The increased cellularity of the lung was in part made up by a substantial number of infiltrating neutrophils (Figure 1D). Analysis of lung and BAL by flow cytometry demonstrated a significant increase in the number of Ly6G+ cells in both sites after infection with 12500 EID50. Other cell populations that increased after infection included macrophages, CD4 and CD8 T cells (data not shown). To address further neutrophil recruitment, we utilized LysM-GFP mice, which express enhanced green fluorescent protein (GFP) from the lysozyme M gene locus, as a reporter for neutrophil numbers in the blood [15]. An increased percentage of circulating neutrophils (GFPhi side-scatterhi) could be detected as early as two days after infection, but only with the higher dose of virus (Figure 1E). Their percentage kept increasing until mice had to be sacrificed 6 days after infection. In contrast, neutrophil numbers in LysM-GFP mice infected with 125 EID50 did not increase and were similar to uninfected control mice or those given allantoic fluid. Thus, by using a high dose of infectious virus, we established a model to induce severe inflammation and increased neutrophil accumulation in the respiratory tract.
Our model allowed us to investigate the contribution of MMPs to pathogenesis. Gelatinases MMP2 and MMP9 are capable of cleaving type IV collagen present in the basement membrane and have been implicated in a variety of pulmonary diseases [13]. We assessed total gelatinase activity in lung tissue by in situ zymography 6 days after influenza virus infection (Figure 2A). Enzymatic cleavage of the probe was significantly increased in the mice infected with 12500 EID50 and seemed to be restricted to defined areas, rather than displaying generalized activity throughout the lung. MMP9 expression correlated with the original infective dose (data not shown) and appeared to be expressed in the vicinity of infected cells, which was visualized using double-staining for MMP9 and viral hemaglutinin (HA) (Figure S1). Of note was the expression of MMP9 in between the airways and foci of viral replication (Figure S1B).
MMP9 can be produced in many cell types and to identify a potential candidate cell population, we transferred C57BL/6 or Mmp9−/− bone marrow into wild-type recipients and allowed the mice to reconstitute for at least four weeks (tested by flow cytometry, data not shown). Mice were infected with 12500 EID50 virus and MMP9 secretion was assessed in the airways (BAL) and lung parenchyma by ELISPOT 6 days later (Figure 2B). Mice reconstituted with C57BL/6 bone marrow had significantly more MMP9-secreting cells in both the airways and lung. However, virtually no MMP9 secretion could be observed when the hematopoietic system was deficient in mmp9, suggesting that bone marrow-derived cells produced the majority of MMP9 in the lung after infection.
To identify the cells that produce MMP9, we did extensive dual immunofluorescence analysis of lung tissues and found Ly6G+ colocalization with MMP9 (Figure 2C). Unlike the Gr-1 marker, Ly6G is specifically expressed on neutrophils, but not on plasmacytoid DC, CCR2+ inflammatory monocytes, or myeloid suppressor monocytes [16]. Thus, we identified neutrophils as the predominant cellular source of MMP9 produced after influenza virus infection.
Since neutrophil numbers were significantly increased following high dose infection and this population appeared to be the predominant source of MMP9 in the lung after infection, we sought to identify the direct function of MMP9 in this cell population. Ly6G is exclusively expressed on neutrophils and antibody treatment has been shown before to specifically deplete neutrophils only [17]. Treating C57BL/6 mice one day before infection and every two days thereafter with an antibody to Ly6G completely depleted Ly6G+ neutrophils as compared to the isotype control IgG (Figure 3A). Importantly, in the absence of neutrophils, there was a significant reduction in weight loss at 5 and 6 days after pathogenic influenza virus infection (Figure 3B). Our data suggest that neutrophils were pathogenic after high dose infection and contribute to morbidity. Furthermore, we demonstrated that MMP9 secretion correlated with morbidity, since neutrophil depletion also significantly reduced MMP9 exocytosis (Figure 3C), confirming our results that MMP9 was predominantly produced by neutrophils after influenza virus infection (Figure 2C).
In addition to containing MMP9 in preloaded granules, neutrophils are capable of secreting a variety of inflammatory cytokines or chemokines [18], which could be modulating the MMP9 response after infection indirectly. Therefore, the levels of innate mediators in the BAL were tested 6 days after infection by bead array analysis. The concentrations of the tested inflammatory cytokines and chemokines were similar between the anti-Ly6G and isotype-treated mice (Figure 3D), excluding a direct effect of these neutrophil-derived soluble mediators in influenza virus pathogenesis. Although a recent study demonstrated that neutrophils enhanced the response of virus-specific CD8 T cells [19], we did not observe differences in IFNγ production by CD8 T cells between the antibody-treated groups in our model (Figure S2). Thus, our data demonstrate that Ly6G+ neutrophils are capable of producing MMP9 and do not affect influenza virus pathogenesis through production of inflammatory mediators or by indirectly affecting CD8 T cell function.
Unlike the chemokines, our results implicate MMP9 in neutrophil-mediated morbidity. Although controversial, proteolytic digestion of the basement membrane lining the lung endothelium has been proposed to be able to mediate migration of cells [20]. To address the functional role of MMP9 during influenza virus pathogenesis, we infected C57BL/6 and Mmp9−/− mice and quantified the number of neutrophils in the respiratory tract. Ly6G+ cells were significantly decreased in both BAL and lung from the Mmp9−/− mice 6 days after infection (Figure 4A). As a consequence, the viral load in the lungs of the Mmp9−/− mice was significantly increased, demonstrating that MMP9, and possibly neutrophils, are required for viral clearance (Figure 4B). Airway levels of CXCL1, CCL2, CCL3, CCL5, IFNγ, IL-6, and TNFα were measured 3 and 6 days after infection. We could only identify significant increases in CXCL1 both days (Figure 4C and Figure S4). Recovery of alveolar macrophages (F4/80hiCD11chi) and exudate macrophages (F4/80intCD11cint), which could contribute to the inflammatory response, were not significantly different 3 days after infection of Mmp9−/− mice compared to C57BL/6 controls (Figure S3A). Likewise, lung pathology was not impacted in Mmp9−/− mice at this time point (data not shown). These results support the notion that MMP9 is crucial for neutrophil, but not macrophage, migration to the infected lung and thereby not only contributes to influenza virus pathogenesis, but is also required for control of viral replication.
We next investigated the mechanism(s) responsible for MMP9-mediated neutrophil migration. Since influenza-associated inflammation is initiated and dependent on TLRs [7], [8], we inquired if TLR signaling was also involved in MMP9-mediated neutrophil recruitment. To this end, mice deficient in the TLR adaptor molecule MyD88 (mediates all TLR signaling, except for that by TLR3) and Tlr3−/− mice were infected with 12500 EID50 and the effect on viral replication, MMP9 secretion, and neutrophil numbers was assessed. Morbidity in both strains was not significantly different (data not shown). Furthermore, overall lung pathology was only marginally ameliorated in Myd88−/− mice 3 days after infection (data not shown). However, viral titers were significantly increased in the lungs of Myd88−/− mice (p<0.01), but not Tlr3−/− mice 3 days after infection (Figure 5A). When cellularity was examined by flow cytometry, the percentage of neutrophils in the lungs was significantly reduced in Myd88−/− (p<0.001) and Tlr3−/− (p<0.05) mice (Figure 5B–C). In accordance, MMP9 secretion was also significantly reduced in the airways in both strains (Figure 5D–E). In contrast, recovery of exudate macrophages from the airways of Myd88−/− or Tlr3−/− mice compared to C57BL/6 controls was similar 3 days after infection (Figure S3B–C). However, whereas alveolar macrophage recovery was unaffected in Myd88−/− mice, there was a significant decrease in their percentage in the Tlr3−/− mice. These results highlight a role for MyD88-derived signals in MMP9 production from neutrophils.
TLRs are expressed on neutrophils [18], [21] and following ligation could induce activation. An inability to recognize the virus may therefore result in decreased MMP9 granule exocytosis. However, it was unclear from our approach whether MMP9 secretion is impaired due to the absence of TLRs in/on the neutrophils themselves or if the absence of MyD88 signaling in other cells affected neutrophil migration, thereby indirectly reducing MMP9 levels in the Myd88−/− airways. In order to distinguish the contribution of MyD88 signaling by immune cells versus non-hematopoietic cells, two bone marrow transfer approaches were used. In the first, bone marrow cells from C57BL/6 or Myd88−/− mice were transferred into Ly5.1 recipients (WT>WT and knock-out (KO)>WT, respectively). Upon reconstitution, all groups were infected with 12500 EID50 and 3 days later the number of neutrophils was enumerated by flow cytometry. No differences could be observed in the lungs (Figure 5F) or BALs (data not shown) of mice receiving wildtype or Myd88−/− bone marrow, suggesting that MyD88 signaling in neutrophils themselves is not involved. In the second set of experiments, the inverse approach was performed by transferring congenically marked wildtype cells (Ly5.1) into either C57BL/6 or Myd88−/− recipients (WT>WT and WT>KO, respectively). In these experiments, a reduction in the number of neutrophils gaining access to the respiratory tract of the Myd88−/− recipients was observed (Figure 5F). Interestingly, the data suggest that MyD88 signaling in non-hematopoietic cells is responsible for neutrophil migration into the lung, rather than TLR recognition affecting neutrophil MMP9 secretion directly.
The indirect nature of MyD88-mediated neutrophil recruitment suggests that chemotaxis to the respiratory tract could be affected. To assess whether reduced cytokine or chemokine expression contributed to the decreased neutrophil accumulation, levels of soluble mediators in the BALs of Myd88−/− mice were analyzed 3 dpi (Figure 6). CCL2, CCL5, IFNγ, IL-1β, IL-6, and IL-17 were not different between wildtype and genetically deficient mice (data not shown). However, in keeping with the reduced neutrophil levels, CCL3 and TNFα levels were reduced in the airways of Myd88−/− mice. The role of chemoattractant CCL3 has been described in neutrophil recruitment and suggested that influenza virus recognition via TLRs in the lung induces chemotactic factors that facilitate neutrophil migration into the infected lung.
We further wanted to investigate the role of TNFα in MMP9-mediated neutrophil recruitment. In vivo anti-TNFα antibody administration of C57BL/6 mice one day before infection and every day thereafter resulted in a significant reduction in neutrophil recruitment 3 days after administration of 12500 EID50 PR8 (Figure 6C). Furthermore, MMP9 secretion tended to be lower in the lung following anti-TNFα antibody treatment (p = 0.07) (Figure 6D). Since the effect of TNFα on neutrophil recruitment in vivo may be multifactorial (e.g. differential adhesion receptor expression on the endothelium), the direct consequence of TNFα on enriched neutrophils was tested in vitro. Negatively enriched neutrophils were stimulated with different doses of recombinant TNFα and MMP9 secretion tested by ELISPOT (Figure 6E). A dose-dependent increase in MMP9 secretion was observed, highlighting the direct role of TNFα in regulating MMP9 release from neutrophils. Altogether these data demonstrate a mechanistic link between influenza virus-induced MyD88 signaling in non-hematopoietic pulmonary cells to facilitate neutrophil recruitment by inducing chemotactic and proteolytic mediators needed for their motility in situ.
Although implicated in a variety of pulmonary diseases, most notably asthma and tuberculosis, a detailed understanding of the involvement of gelatinases in lung disease pathology is lacking. In this study, we addressed the contribution of MMP9 during influenza virus pathogenesis, which had been implicated recently by virtue of increased concentrations in the serum of patients with influenza and elevated enzymatic activity in mouse lung homogenates [22], [23]. However, the cellular origin and direct role of MMPs in the pathogenesis of influenza virus have not been previously examined. Thus, we focused on elucidating the effects of MMP9 in a pathological mouse model and identified neutrophils as the predominant pulmonary source of MMP9. We demonstrated that MMP9 mediated neutrophil migration into the infected respiratory tract and that it was required for viral clearance. Although MMP9 release was TLR signaling-dependent, MyD88-mediated signals in non-hematopoietic cells, rather than neutrophil TLRs themselves, were important for neutrophil recruitment. This is most likely due to MyD88-controlled induction of the known neutrophil chemoattractant, CCL3, as well as TNFα, which we found induced MMP9 secretion in neutrophils. Our data demonstrate a previously unknown role of MMP9 to influenza virus pathogenesis by contribution to excessive neutrophil migration into the respiratory tract.
Using a model wherein a lethal dose of virus mimics many of the characteristics observed after pandemic infections, with greater cytokine/chemokine levels, lung pathology, and inflammatory cells [5], [14], we showed that MMP9 activity was upregulated in infected lungs, confirming data published recently [23], [24]. We extended these studies by demonstrating that MMP9 was exclusively produced by hematopoietic cells. It is surprising that non-hematopoietic cells did not contribute to MMP9 secretion, as it can be induced under inflammatory conditions in lung epithelial and endothelial cells [25]. Nevertheless, our data further indicated that MMP9 was predominantly secreted in the lung by neutrophils. Neutrophil depletion caused a significant reduction in the number of cells being able to produce MMP9, but did not abrogate all MMP9 production, suggesting that other immune cells can also contribute. In a recent study, neutrophil depletion did not affect MMP9 levels in the BAL 5 days after sublethal influenza virus infection, which is at odds with our results [24]. However, contributing factors to this discrepancy most likely involve the timing, which was later, and the infective dose, which was lower. Nevertheless, excessive neutrophil infiltration into the lung correlated with increased MMP9 production and viral replication when macrophages were depleted [24], which supports our findings. However, the source of MMP9 was not identified. We are currently investigating the potential role of other cell populations that are capable of producing MMP9 under inflammatory conditions, such as lymphocytes [13]. However, unlike these cells, neutrophils do not require de novo generation of MMP9 and instead are preloaded in the bone marrow and exocytose their granules following stimulation. In addition, they can synthesize and secrete cytokines and chemokines [18], although we were unable to demonstrate a significant contribution of this mechanism to airway inflammation, as depletion of neutrophils did not cause a significant difference in an extensive panel of potential candidates.
While neutrophils are a major cell population recruited to the lung early in the immune response to pandemic influenza viruses [14], [26], their role during pathogenesis remains unclear. Depletion studies demonstrated enhanced susceptibility of mice to infection, exacerbated inflammation, and increased viral titers and mortality rates [24], [26], [27], [28], which suggested a critical contribution of neutrophils to innate immunity. In contrast, other studies parallel ours and show that neutrophils contribute to disease during severe influenza virus infection [29], [30], [31]. The disparate results may be a consequence of differential infective doses and genetic background. Nevertheless, we show for the first time that MMP9 may contribute to pathogenesis by mediating excessive neutrophil migration into the respiratory tract in response to viral infection. While depletion of neutrophils reduced morbidity, MMP9 was still required for viral clearance. It is likely that the same mechanism(s) that contribute to immunity cause pathology in a more virulent setting [4]. Neutrophils are able to inhibit virus replication in vitro and the release of anti-viral molecules by neutrophils could contribute to viral clearance from the lung [24], [27], [32]. However, the mechanisms that allow for neutrophil influx in response to uncontrolled or high viral load, such as MMP9 proteolytic activity, may exacerbate pathology due to collateral destruction of the lung ECM. This could explain the apparent contradicting reports on the role of neutrophils in viral infections and suggests that, while they contribute to viral clearance, excessive numbers contribute to pathology.
We showed that MMP9 functions by regulating neutrophil localization to the respiratory tract. Cellular influx into the airways is a complex process whereby cells cross the endothelial cell layer, the basement membrane, and the epithelial cell layer. Although rolling and tethering on the endothelium has been described in detail, the means by which cells further progress through the basement membrane to reach the airways remain unclear [33]. Whereas gelatinase activity has been implicated in neutrophil migration through artificial basement membranes in vitro, its relevance in vivo remains controversial [20]. Chemokine-induced migration of neutrophils in an induced lung inflammation model was demonstrated to be independent of MMP9 [34], whereas that in a dermal model did require MMP9 [35]. In part, this discrepancy may be due to the artificial nature of the experiments: while chemokines may be able to mimic inflammation and attract cells, crucial factors necessary for the release of MMP9 may not be available as they are not induced by the chemokine itself (see below). The presence of these required accessory molecules is assured in the influenza model, but at the same time can be difficult to define. Nevertheless, we were able to identify at least one, TNFα, which was able to induce MMP9. It is still of great interest to identify other factors during influenza virus infection as they have obvious clinical relevance. Thus, our data not only demonstrate a role for MMP9 in influenza pathogenesis, but also provide further in vivo evidence for MMP9 activity as a mechanism by which neutrophils digest the extracellular matrix to traverse the basement membrane in order to gain access to the infected lung epithelium. The colocalization of MMP9 and viral antigen and its expression between the endothelium and areas of viral replication support this hypothesis. In addition to its role in mediating proteolytic cleavage of the basement membrane to allow for motility within the lung, MMP9 may also modulate chemokine expression itself by digesting CXCL-1 [13]. Our finding that CXCL-1, a potent neutrophil chemoattractant, is significantly upregulated in the Mmp9−/− mice following infection hints at a negative feedback loop in which MMP9 release by neutrophils disrupts the CXCL-1 gradient and dampens further neutrophil recruitment in the absence of influenza virus.
We hypothesized that TLRs are important to induce MMP9 secretion from neutrophils. These innate pattern recognition receptors are expressed by neutrophils [18]. Furthermore, viral RNA has been detected in neutrophils [36], which could facilitate access to TLRs inside endosomes, such as TLR3 and TLR7, which have both been shown to be able to recognize influenza virus. Indeed, TLR7 signaling was demonstrated to be able to activate neutrophils in response to influenza virus, although this was dependent on GM-CSF [21]. Therefore, we examined the role of TLR signaling in neutrophil activation in Tlr3−/− and Myd88−/− mice. MyD88 is the adapter molecule for all known TLRs with the exception of TLR3, and the inclusion of both strains therefore encompassed all potential TLR-mediated effects. Although both neutrophil accumulation and MMP9 secretion were reduced following high dose influenza virus infection, our bone marrow chimera approach demonstrated that this was not a direct effect as Myd88−/− neutrophils still migrated into the respiratory tract. Furthermore, in vitro stimulation with influenza virus of neutrophils enriched from Myd88−/− mice did not affect MMP9 secretion (data not shown). Rather, Myd88−/− signaling in non-hematopoietic cells was important for neutrophil accumulation. When we examined chemokine levels in the airways of Myd88−/− mice, CCL3, a known neutrophil chemoattractant, was downregulated. Thus, it appears that rather than a direct effect on MMP9 activation, ablation of MyD88 signaling may have affected the chemokine gradient that would normally attract the neutrophils to the sites of viral replication. In addition, TNFα induced the release of MMP9 from neutrophils and antibody blocking of TNFα prevented their recruitment into the lungs in response to influenza virus infection in vivo. TNFα has been demonstrated previously to induce MMP9 exocytosis in a protein kinase C-dependent fashion [37] and we hypothesize that the reduced levels in the Myd88−/− airways prevented MMP9 release and resulted in reduced neutrophil numbers.
The role of neutrophils in viral infections has often been overlooked, on the pretext that they are mostly involved in the clearance of bacterial and fungal pathogens. Their contribution to influenza virus-induced pathology is unclear, but we showed in our pathological model that they contributed to morbidity. Furthermore, we demonstrated that the mechanism for neutrophil-mediated migration to the respiratory tract in response to influenza required MMP9 and depended on extrinsic TLR-signaling. Collectively, our results suggest that innate sensing of viral infection results in a MyD88-dependent induction of chemotactic factors that induce the recruitment of neutrophils. Following diapedesis into the lung parenchyma, MyD88-induced TNFα in non-hematopoietic cells induces MMP9 release in neutrophils and MMP9 enzymatic activity is utilized by these cells to degrade the basement membrane and facilitate their motility within the lung. These novel findings are probably not restricted to influenza pathogenesis, but may also have implications for other viral infections. The dichotomy of the necessity for MMP9-mediated immune cell migration and its role in immunopathology provides real challenges for targeting MMP9 for therapeutic purposes after influenza virus infection. Nevertheless, finding the balance to modulate neutrophils allows for innate immunity to pandemic strains to be boosted whilst preventing potential pathology.
All experiments in this study were approved by the Institutional Animal Care and Use Committee (IACUC) at the Sanford-Burnham Medical Research Institute and were 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.
Male C57BL/6J (C57BL/6J ) mice and B6.SJL-Ptprca Pep3/BoyJ (Ly5.1) were purchased from Jackson Laboratories, USA. Myd88−/− [38], Tlr3−/− [39], Mmp9−/− [40], or LysM-GFP [15] mice (all on C57BL/6 background) were obtained with permission as gifts from Drs S. Swain (University of Massachusetts, MA), M. Corr (University of California, San Diego, CA), F. Kherradmand (Baylor College of Medicine, TX), and G. Srikrishna (Sanford-Burnham Medical Research Institute, CA), respectively. All mice were used at 6–16 weeks of age and held under specific-pathogen-free conditions in the vivarium at the Sanford-Burnham Medical Research Institute.
Influenza virus A PR/8/34 (PR8, H1N1) was grown in the allantoic fluid of 10-day-old embryonated eggs (McIntyre Poultry, San Diego) and stored at −80°C until use. Before infection, mice were anesthetized by intraperitoneal injection of 100 µL ketamine/xylazine (14.3 mg/mL Ketaset (Fort Dodge, IA)/2.86 mg/mL Anased (Lloyd Laboratories, IA)). Mice were infected with 1.25×104 EID50 influenza virus (unless otherwise stated) in 50 µL by the intranasal route. Mice were terminated at the indicated timepoints or when clinical signs reached pre-defined endpoints (body weight loss of 30%) by intraperitoneal injection of avertin (2% (w/v) 2,2,2-tribromoethanol, 2% (v/v) tert-amyl-alcohol) or CO2 inhalation.
Cells from the airways and lung were obtained from infected mice at the indicated timepoints. Cells from airways were obtained via BAL by piercing the diaphragm, transecting the trachea, and infusing the lungs with 1 mL PBS through an 18-gauge catheter (Terumo, USA). The resulting cell suspension was centrifuged (5 min, 300×g), the supernatants aspirated and stored at −80°C for cytokine bead array (see below), and the cells resuspended in wash medium (HBSS (CellGro, USA) supplemented with 1% (v/v) heat-inactivated fetal calf serum, 5 mM HEPES (CellGro), and 100 U/mL Pencillin, 100 µg/mL Streptomycin, 292 µg/mL L-glutamine (Mediatech)). Perfused lungs were cut into small pieces, digested in 2 mL collagenase (1 mg/mL collagenase D (Roche, USA), 0.5% bovine serum albumin (BSA, Sigma), 100 µg/mL Dnase) for 60 min at 37°C, and filtered and homogenized over a 70 µm cell sieve. Cell suspensions were washed and resuspended in cell medium and the cell numbers enumerated.
Cytokine production in the airways after infection was measured using a Legend-plex custom cytokine/chemokine array according to manufacturer's instructions (Biolegend, USA). Supernatants aspirated from BAL were assayed for indicated combinations of the following analytes: IL-1β, IL-6, IL-12p40, IL-17a, IFNγ, CXCL1, CCL2, CCL3, CCL5, and TNFα. Samples were analyzed on a Luminex IS200.
The percentage of neutrophils present in the airways or lung after infection was analyzed by flow cytometry. Recovered cells were stained with monoclonal anti-mouse antibodies Ly6G-APC (1A8) and Gr1-PE (RB6–8C5) (Biolegend). Cells were stained for 15 min on ice before being washed and analyzed on a FACSCalibur or LSR2 Fortessa flow cytometer (Becton Dickinson, CA) using FlowJo 8.8.6 (Tree Star, Ashland, USA). The number of neutrophils in each tissue was calculated from the cell recovery and the percentage positive by flow cytometry.
For microscopic analysis of pathology, lung tissues were fixed in buffered formaldehyde-saline (pH 7.5), embedded in paraffin, and sectioned. Each lung was stained with H&E prior to examination. Tissues sections for immunofluorescence staining were cut (8 µm) using a cryostat (Leica Jung CM3000), air-dried overnight, fixed in acetone for 10 min, and stored at −35°C. Endogenous biotin in sections was blocked (Invitrogen, USA) before addition of antibodies: purified anti-MMP9 (1∶250, Millipore), goat anti-rabbit IgG Alexa Fluor 488 (1∶100, Invitrogen), anti-Ly6G-Biotin (1∶500, Biolegend), and streptavidin-AlexaFluor 568 (1∶1000, Invitrogen). All antibodies were diluted in 6% (w/v) BSA in PBS and sections washed 3× in PBS between antibodies. Sections were incubated at room temperature for 60 min with primary antibody and 30 min with the conjugated antibody. Sections were mounted using Vectashield hard set mounting media (Vector Labs) and immunofluorescent staining analyzed with a fluorescence microscope (BX50, Olympus) using UPlanFl 20× or 40× objectives (Olympus, both at ∞/0.17). Images were recorded using a SpotFlex camera and SpotSoftware 4.6 (Diagnostic Instruments).
In situ gelatinase assay was performed on frozen sections as described before [41]. Briefly, 1% (w/v) low-gelling agarose (Seaplaque) was heated, supplemented with DAPI (1 µg/mL) and allowed to cool. The agarose was mixed 9∶1 with DQ-Gelatin (1 mg/mL, Invitrogen) and layered atop fresh lung sections, which were cut as described above. Gelatinase (MMP2/MMP9) activity was assessed after an 18 hr incubation as described.
Recipient mice were irradiated 2× with 450 RAD 8 hr apart and given tetracycline (125 mg/500 mL) or sulfamethoxazole and trimethoprim (40 mg/500 mL and 200 mg/500 mL) in the drinking water. Donor cells were prepared from bone marrow collected from femurs and tibias of donor Mmp9−/−, Myd88−/−, Ly5.1, or C57BL/6 mice and single cell suspensions prepared by passing the BM through a 25G needle and lysing the red blood cells with 0.15 M NH4Cl prior to transferring 4–5×106 cells i.v. to irradiated recipients as indicated. Reconstitution of the recipients by donor BM cells was allowed for at least 4 weeks and was tested by flow cytometric analysis of the recipient's blood using the appropriate congenic markers. Briefly, blood samples (∼40 µL) were drawn by retro-orbital bleed, red blood cells lysed, and cells stained using Ly5.1 FITC (eBioscience, San Diego, CA) and Ly5.2 PerCP (Biolegend, San Diego, CA). The percentage of reconstitution was calculated after flow cytometry and was always higher than 95%.
Lungs were removed from mice, snap-frozen in liquid nitrogen, and stored at −80°C prior to viral titer. Lungs were homogenized as described before [42] and the presence of influenza virus was assessed in lungs 5 days after infection by injection of 100 µl homogenized lung supernatant into the allantoic fluids of 10-day-old embryonated eggs (McIntyre Poultry, San Diego), followed by hemagluttination assay using chicken red blood cells [43].
Alternatively, viral titers were assessed by PA copy number per lung [44]. Briefly, RNA was prepared from homogenates of the left lung lobe using Trizol (Invitrogen) and RNeasy (Qiagen), and the polymerase (PA) gene of PR8 amplified using one-step RT qPCR (Eurogentech) by ABI 7900HT (Applied Biosystems). Data were analyzed using SDS 2.3 (Applied Biosystems) and the PA copy number per lung lobe calculated using a PA-containing plasmid (a gift from Dr S. Swain (University of Massachusetts, MA) of known concentration as a standard.
Neutrophils were depleted by i.p. injection of 400 µg anti-Ly6G (1A8 clone, Biolegend) per mouse 1 day before infection and every other day following. Control groups received 400 µg rat isotype control (RatIg, Jackson Immunoresearch) at the same time points.
Secretion of MMP9 in the respiratory tract was analyzed by ELISPOT assay. Cell suspensions of BAL and lung were prepared as described above and 5×103 or 1×104 cells, respectively, resuspended in culture medium (RPMI (CellGro) supplemented with 10% (v/v) heat-inactivated fetal calf serum, 20 mM HEPES (CellGro), 50 µM β-mercaptoethanol, 100 U/mL Pencillin, 100 µg/mL Streptomycin, 292 µg/mL L-glutamine (Mediatech)). The cell suspensions were then added in triplicate to wells of a 96-well plate (Immobilin-P, Millipore), which had been coated with MMP9 capture antibody (R&D, USA) overnight at 4°C. Cells were incubated overnight at 37°C, 5%CO2, washed, and detection MMP9 antibody added overnight at 4°C. MMP9 spots were visualized with ELISPOT development module (R&D Systems) and enumerated on an Immunospot reader (Cellular Technology Ltd, Seattle).
TNFα was neutralized by i.p. injection of 200 µg anti-TNFα (XT3.11 clone, BioXcell) per mouse 1 day before infection and every day following. Control groups received 200 µg RatIg (Jackson Immunoresearch) at the same time points.
Bone marrow cells were collected from femurs and tibias of C57BL/6 mice and neutrophils were enriched using magnetic negative enrichment according to manufacturer's instructions (Stemcell Technologies). Neutrophils (purity >80% by flow cytometry) were then incubated with 0, 1, 10, 100, or 1000 pg/mL recombinant TNFα (R&D Systems) for 6 hours in wells of a 96-well plate (Immobilin-P, Millipore), which had been coated with MMP9 capture antibody (R&D Systems). The MMP9 ELISPOT was finished as described above.
All data were analyzed with Prism software (Graphpad software, USA). Unless otherwise noted, a two-tailed Mann-Whitney test was used to compare two treatment groups. Larger groups were analyzed with Kruskal-Wallis analysis of variance. Where possible, results are expressed as means ± SEM. Values of p<0.05 were considered statistically significant.
Murine proteins (Swiss-prot): MMP9: P41245, TLR3: Q99MB1, MyD88: P22366, TNF: P06804, CCL3: P10855, CXCL-1: P12850.
|
10.1371/journal.ppat.1000806 | N-Acetylglucosamine Induces White to Opaque Switching, a Mating Prerequisite in Candida albicans | To mate, the fungal pathogen Candida albicans must undergo homozygosis at the mating-type locus and then switch from the white to opaque phenotype. Paradoxically, opaque cells were found to be unstable at physiological temperature, suggesting that mating had little chance of occurring in the host, the main niche of C. albicans. Recently, however, it was demonstrated that high levels of CO2, equivalent to those found in the host gastrointestinal tract and select tissues, induced the white to opaque switch at physiological temperature, providing a possible resolution to the paradox. Here, we demonstrate that a second signal, N-acetylglucosamine (GlcNAc), a monosaccharide produced primarily by gastrointestinal tract bacteria, also serves as a potent inducer of white to opaque switching and functions primarily through the Ras1/cAMP pathway and phosphorylated Wor1, the gene product of the master switch locus. Our results therefore suggest that signals produced by bacterial co-members of the gastrointestinal tract microbiota regulate switching and therefore mating of C. albicans.
| To mate, the human fungal pathogen Candida albicans must undergo a complex phenotypic change from a round “white” to large, elongated “opaque” cell. This involves the regulation of approximately 5% of the organism's genes. Surprisingly, this complex transition is not required for mating in other related yeast. Even more surprisingly, it was found that in vitro the mating-competent opaque phenotype was unstable at 37°C, the temperature of the host body. This observation led to a paradox. If C. albicans lives primarily in an animal host, physiological temperature would thwart mating, so where does C. albicans mate? This led to the suggestion that some physiological condition in the host niche stabilizes the opaque phenotype or even induces switching from white to opaque, so cells can mate. Recently, we demonstrated that the high concentrations of CO2 found in tissue and the gastrointestinal tract induced switching from white to opaque and then stabilized the opaque phenotype. Here, we demonstrate that a second factor, N-acetylglucosamine (GlcNAc), a sugar released primarily by bacteria in the gastrointestinal tract, also induces the switch from white to opaque and stabilizes the opaque phenotype. We demonstrate by mutational analysis that GlcNAc induction is regulated primarily by the Ras1/cAMP pathway, which also regulates filamentation of C. albicans. This is perhaps not surprising given that white-opaque switching shares with filamentation several phenotypic characteristics. Finally, we show that induction by GlcNAc requires the phosphorylated master switch gene that regulates spontaneous switching, suggesting that it induces the switch from white to opaque by activating the gene product of the master switch gene. Together, our results suggest that multiple signals from bacterial co-members of the gastrointestinal tract microbiota regulate switching and, therefore, mating of C. albicans in the colonized host.
| The white-opaque transition in MTL-homozygous strains of Candida albicans affects cellular physiology, cell morphology, gene expression, virulence and biofilm formation [1]–[3]. It is repressed by the a1-α2 co-repressor in a/α cells and derepressed in cells that have undergone MTL-homozygosis to become either a/a or α/α [4]. White-opaque switching, which occurs spontaneously and reversibly, is controlled through expression of a master switch locus, WOR1, which also has been referred to as TOS9 [5]–[7]. The frequency of switching is regulated in part at the level of WOR1 transcription by a number of genes through a network of positive and negative regulatory loops [8],[9] and through changes in chromatin state [10]–[12].
After the discovery of a mating system in C. albicans [13], it was demonstrated that MTL-homozygous cells had to switch from white to opaque in order to mate [4],[14]. Paradoxically, it was demonstrated that in vitro this switch was sensitive to physiological temperature [15],[16]. When the temperature of opaque cell cultures grown at 25°C was raised to 37°C, cells switched en masse and semi-synchronously to white [17], suggesting that the opaque phenotype was unstable at physiological temperatures and that mating would, therefore, be compromised in a host, the major niche of C. albicans. Recently, we demonstrated that high levels of CO2 comparable to those found in the host gastrointestinal tract and some host tissues induced switching from white to opaque, maintained cells in the opaque phenotype, and blocked switching from opaque to white [18]. CO2 had been demonstrated previously to be a potent inducer of filamentation as well [19],[20]. Because N-acetylglucosamine (GlcNAc), which is produced by bacteria in the gastrointestinal tract [21], is also a potent inducer of filamentation [22], we therefore considered the possibility that it, like CO2, was an inducer of the white to opaque transition.
We found that G1cNAc represents a second strong inducer of the white to opaque transition and stabilizes the opaque phenotype. GlcNAc induction occurs at 25°C and is enhanced at 37°C. In addition, because there were indications that the induction of filamentation by GlcNAc was mediated by the Ras1/cAMP pathway [23]–[30], we tested whether G1cNAc induction of switching was regulated by this pathway. Our results demonstrate that GlcNAc induction is transduced primarily by the same Ras1/cAMP pathway that has been implicated in the regulation of filamentation and requires phosphorylated Wor1, the product of the master switch locus. We therefore suggest that two different signals in the host gastrointestinal tract, both produced by bacterial co-members of the gastrointestinal tract microbiota, can regulate the white to opaque transition, an essential step in C. albicans mating.
To test whether GlcNAc induces the white to opaque transition and does so as a function of culture age, as is the case for the induction of filamentation [22],[24], white cells of a/a and α/α derivatives of strain SC5314, 5314a and 5314α, respectively, were first grown at 25°C in suspension in liquid modified Lee's medium in which glucose was the sole carbon source (“liquid glucose medium”) [31] (Figure 1A). To assess GlcNAc induction as a function of culture growth [23], cells were removed at time intervals from the liquid culture, plated on nutrient agar containing either 1.25% (w/v) glucose (“glucose agar”) or 1.25% (w/v) GlcNAc (“GlcNAc agar”) as the sole carbon source (Figure 1A), and incubated at 25°C. This temperature was selected to assess induction initially, because physiological temperature (37°C) induces the reverse switch from opaque to white [15],[17], and we wanted the initial assessment to be performed in the absence of reverse induction. After five days on agar, the proportion of opaque colonies plus white colonies with opaque sectors was measured in glucose or GlcNAc agar. This proportion will be referred to as the “switching frequency” for convenience, but should not be confused with the rate of switching [1],[16],[32]. Although a/a and α/α cultures reached different final cell densities, they entered the saturation phase in liquid glucose medium at approximately the same time (Figure 1B).
For a/a cells plated on glucose agar at 25°C, the switching frequency increased from 0.4±0.4% for cells taken from exponential phase cultures after one day, to 3.5±1.2% for cells taken from late saturation phase cultures after 10 days (Figure 1C, D). For α/α cells, the proportion increased similarly from 0.5±0.5% to 2.9±0.2% (Figure 1C, D). Hence, the switching frequency of a/a and α/α cells grown in liquid glucose medium increased 9.5- and 6.0-fold, respectively, over the course of exponential growth and entrance into the saturation phase. For a/a cells plated on GlcNAc agar, the frequency of switching increased from 5.1±1.3% after one day to 88.7±4.9% after 10 days, and for α/α cells, the frequency increased from 7.3±3.4% to 92.5±3.3% (Figure 1C, D). Plating on GlcNAc agar, therefore, caused an increase in the frequency of switching of a/a and α/α cells after one day that was approximately 13- and 15-fold higher, respectively, than the frequencies on glucose agar after one day (Figure 1C, D). After 10 days, the frequency was 25- and 32-fold higher, respectively, than the frequency on glucose agar (Figure 1C, D). In Figure 1E, examples are presented of cultures from three day liquid glucose cultures of a/a cells plated on glucose agar or GlcNAc agar. Note that on GlcNAc agar, the majority of colonies were completely opaque rather than sectored, indicating that in these cases GlcNAc induction occurred very early in the life history of the colonies. Similar results were obtained for cells grown in liquid glucose medium for five days and plated on agar containing either glucose or GlcNAc ranging in concentration from 0.2% to 5% (w/v), indicating that the concentration employed (1.25%, w/v) resulted in maximum G1cNAc induction (data not shown).
The preceding experiments were performed at 25°C. To test whether GlcNAc also induced white to opaque switching at physiological temperature (37°C), we then performed experiments in which white cells of a/a and α/α derivatives of strain SC5314 were grown to mid-log phase on liquid glucose medium for 48 hr at 25°C, then plated on glucose or GlcNAc agar at either 25 or 37°C. Increasing the temperature from 25 to 37°C on glucose agar resulted in a lower frequency of switching for white cells of both the a/a and α/α strains (Table 1, data in air). In direct contrast, when white cells of both strains were grown in liquid glucose medium at 25°C and then plated on GlcNAc agar at 37°C, there was a dramatic increase in the switching frequency (Table 1, data in air). These data demonstrate that physiological temperature enhances GlcNAc induction.
The cAMP pathway, which plays a role in filamentation, includes Ras1, Cdc35, Pde2 and the protein kinase isoforms Tpk1 and Tpk2 [25]–[30]. To test whether GlcNAc induction of white to opaque switching was mediated by the cAMP pathway, we first analyzed the RAS1 deletion mutant, ras1/ras1. These experiments were performed at 25°C because a temperature of 37°C induced a portion of the cells of mutants of the Ras1/cAMP pathway to undergo filamentation (data not shown), making it difficult to assess switching from white to opaque at the level of cell phenotype. White cells of ras1/ras1 and the control strain (WT) were grown at 25°C in liquid glucose medium to saturation phase (seven days), plated on either glucose or GlcNAc agar, and analyzed for switching frequencies after five days at 25°C. The switching frequency on GlcNAc agar was 90.5±3.8% for WT cells, and 11.2±1.5% for ras1/ras1 cells (Figure 2A), indicating that Ras1 played a major, but not exclusive, role in GlcNAc induction. The frequency of switching of ras1/ras1 cells on GlcNAc agar was 9-fold lower than that of WT cells, and 16-fold higher than that on glucose agar (Figure 2A). Complementation of ras1/ras1 with RAS1 under the control of the MET3 promoter partially rescued the mutant phenotype in the activated state (Figure 2A). Rescue was incomplete due to the fact that RAS1 was controlled in the complemented strain by the MET3 rather than the natural promoter [7],[33]. It should also be noted that on glucose agar, the frequency of switching of WT cells was two-fold higher than that of ras1/ras1 cells (Figure 2A), indicating that a RAS1-dependent pathway also played a role in spontaneous switching on glucose agar.
To explore further the role of RAS1 in G1cNAc induction, we transformed strain WUM5A, a derivative of α/α strain WO-1, with RAS1V13, which encodes a constitutively activated form of Ras1 (Ras1V13 [25]) under the control of the MET3 promoter MET3p [33], to generate strain WT+MET3p-RAS1V13. The control strain was also transformed with the vector lacking the RAS1V13 to generate the control strain WT+vector. The addition of 2.5 mM methionine plus 2.5 mM cysteine (+Met, +Cys) represses MET3 promoter activity (the repressed state) and the absence (-Met, -Cys) activates it (the activated state) [33]. White cells of WT+METp-RAS1V13 and WT+vector were plated directly onto glucose agar in the presence or absence of methionine and cysteine at 25°C. In the repressed state (+Met, +Cys), the majority of colonies were white, with few sectors, but in the activated state (-Met, -Cys), nearly every colony was highly sectored (Figure 2B), indicating that overexpression of RAS1V13 in the absence of GlcNAc induced switching. Next, white cells of strains WT+vector and WT+METp-RAS1V13 were grown in liquid glucose medium in the repressed state for one day to the mid-exponential phase, then plated on either glucose or GlcNAc agar in the induced state. On glucose agar, the switching frequency of white WT+vector cells was 2.1±1.0%, and for the overexpression strain, 100% (Figure 2C). The majority of colonies of the overexpression mutant on glucose agar were highly sectored white colonies (Figure 2D). Only 2.3±1.2% were homogeneous opaque colonies. On GlcNAc agar, the switching frequency of control cells was 3.2±0.2%, while that of the overexpression mutant was 100% (Figure 2C). All of the latter colonies were homogeneously opaque (Figure 2D). The near uniformity of the opaque phenotype in the latter colonies was evident at the cellular level (Figure 2E). These results reinforce the conclusion that induction of white to opaque switching by GlcNAc is mediated primarily by Ras1.
In the cAMP pathway that is involved in filamentation, Ras1 activates adenylate cyclase, which is encoded by CDC35 [26]. The resulting increase in cAMP is kept in check by a cAMP-phosphodiesterase, which is encoded by PDE2 [34],[35]. If GlcNAc induction of white-opaque switching is mediated by the same cAMP pathway, then deletion of CDC35 should reduce the effect, while deletion of PDE2 should enhance it. When white cdc35/cdc35 cells were grown at 25°C to saturation phase in liquid glucose medium (seven days) and then plated on GlcNAc agar, the frequency of switching was 8.0±3.5%, whereas that of the WT parental control was 86.9±4.3% (Figure 3A). Complementation of the mutant cdc35/cdc35 with CDC35 under control of the MET3 promoter partially rescued the mutant phenotype (Figure 3A). These results indicate that CDC35 is necessary for the major response to GlcNAc, as was the case for RAS1. GlcNAc did, however, induce low level switching in white cdc35/cdc35 cells, indicating that although CDC35 is necessary for the major response to GlcNAc, there is a minor response that is CDC35-independent, just as we observed there is a minor response that is RAS1-independent.
When white pde2/pde2 cells were grown at 25°C to saturation phase in liquid glucose medium (five days) and then plated on GlcNAc agar, the switching frequency was 100%, compared to 78.6±6.5% for control cells (Figure 3A). When plated on glucose agar, the frequency of switching was 96.0±1.5%, compared to 0.6±0.6 for WT cells (Figure 3A). Complementation of the pde2/pde2 mutant with PDE2 under the control of the MET3 promoter partially rescued the mutant phenotype (Figure 3A).
To explore further the role of Pde2 in switching, the deletion mutant pde2/pde2 was also transformed with a vector containing PDE2 under the control of the MET3 promoter to generate the strain pde2/pde2+MET3p-PDE2. The deletion mutant pde2/pde2 was transformed with the vector lacking PDE2 to generate the control strain pde2/pde2+vector. When white cells of the parental control (WT) were grown at 25°C as a streak on glucose agar lacking methionine and cysteine (activating conditions), only rare opaque sectors formed at the periphery (Figure 3B). When the mutant pde2/pde2+vector was streaked at 25°C, opaque sectors rimmed the entire streak (Figure 3B). In contrast, opaque sectors were absent at the periphery of the streak of the overexpression mutant pde2/pde2+MET3p-PDE under activating conditions (Figure 3B). When the overexpression mutant pde2/pde2+MET3p-PDE was grown at 25°C in liquid glucose medium under activating conditions to early exponential phase (two days), then plated on glucose agar at 25°C under activating conditions, the frequency of sectoring was 2.8±1.2%, compared to 98.9±1.9% for the deletion mutant (Figure 3C). When plated at 25°C on GlcNAc agar under activating conditions, the frequency of switching of the rescued strain was 26.8±5.8% compared to 100% for the mutant (Figure 3C). Examples of colonies of strain pde2/pde2+vector and pde2/pde2+MET3p-PDE grown at 25°C on glucose or GlcNAc agar under activating conditions are presented in Figure 3D, and examples of cells from these colonies are presented in Figure 3E. These results support the conclusion that cAMP is involved in the regulation of the GlcNAc response and that Pde2 plays the role of a negative regular.
In the Ras1/cAMP pathway, cAMP activates protein kinase A (PKA) [27],[36]. In S. cerevisiae there are three PKA catalytic subunits, Tpk1, Tpk2 and Tpk3, which play roles in the cAMP pathway regulating pseudohypha formation [36],[37]. C. albicans possesses two isoforms, Tpk1 and Tpk2, which have been demonstrated to play functionally different roles in filamentation, depending upon environmental conditions [27],[38]. Consistent with previous reports [27], our lack of success in generating a double mutant of TPK1 and TPK2 suggested that this mutant may not be viable. We therefore analyzed the individual deletion mutants tpk1/tpk1 and tpk2/tpk2. White cells of the individual deletion mutants were grown at 25°C in liquid glucose medium to saturation phase (seven days) and then plated on glucose or GlcNAc agar and examined for switching after five days. Deletion of either TPK1 or TPK2 had no detectable effect on the frequency of switching on glucose or on GlcNAc agar (Figure 3F). There was, however, one noticeable difference in the GlcNAc-induced opaque colonies of tpk1/tpk1. They possessed a mixture of opaque cells and cells that had formed hyphae, suggesting that Tpk1 plays a role in the regulation of the bud-hyphae transition (data not shown).
Given that each of the alternative PKA isoforms may perform redundant function in the two mutants tpk1/tpk1 or tpk2/tpk2, we generated overexpression mutants in the wild type background WUM5A (WT) in which TPK1 or TPK2 was placed under the regulation of the strong constitutive ACT1 promoter [5]. White cells of the overexpression strains WT+ACTp-TPK1 and WT+ACTp-TPK2, as well as white cells of the control strain, were grown for five days at 25°C to saturation phase in liquid glucose medium and then plated on glucose or GlcNAc agar and examined after five days at 25°C for switching (Figure 1A). Overexpression of TPK1 had no effect on the switching frequency on glucose agar and actually suppressed switching on GlcNAc agar (Figure 3F). Overexpression of TPK2, however, caused a tenfold increase in the switching frequency on glucose agar over that of wild type cells and enhanced the frequency of switching by approximately 20% on GlcNAc agar (Figure 3F). These results suggested that in wild type cells, Tpk2 may function as the major downstream kinase in the GlcNAc induction pathway.
At 37°C, GlcNAc induction was impaired in the mutants ras1/ras1 and cdc35/cdc35 (supplemental Table S1) but to a lesser extent than at 25°C. High temperature induction, however, reinforced our conclusion that it is Tpk2 that plays the crucial role in wild type cells in transducing GlcNAc induction. Whereas at 25°C GlcNAc induction was unaffected in tpk1/tpk1 and tpk2/tpk2, at 37°C it was reduced in the tpk2/tpk2, but not tpk1/tpk1 (supplemental Table S1). These results supported our conclusion based on overexpression data (Figure 3B-E) that Tpk2 plays a role in transducing GlcNAc induction in wild type cells.
The WOR1 (TOS9) locus has been demonstrated to regulate spontaneous white-opaque switching [5]–[7]. It has been proposed that a stochastic increase in WOR1 expression above a threshold causes cells to switch from white to opaque, and that continued expression above that threshold maintains the opaque phenotype [5]–[7]. Wor1 has been shown to auto-induce at the level of transcription [5]–[7]. When activated, the cAMP pathway, which traditionally functions by cAMP-activation of protein kinase A, might increase the frequency of switching by phosphorylating either Wor1 or one of the several proteins that modulate WOR1 function through transcriptional regulatory loops [8],[9] or chromatin modification [10]–[12]. Interestingly, Wor1 possesses a single consensus PKA phosphorylation motif, between amino acids 64 and 69 with a phosphorylatable threonine at amino acid 67 [5].
To test whether Wor1 was essential for GlcNAc-activated switching, white cells of the parental strain (WT) and the WOR1 deletion mutant wor1/wor1 were grown at 25°C to saturation phase (seven days) in liquid glucose medium and then plated on nutrient agar containing glucose or GlcNAc at 25°C. The wor1/wor1 mutant did not switch on either glucose or GlcNAc agar (Figure 4A, B). Neither a single opaque colony or opaque sector was observed among more than 1,000 colonies. This was also true at 37°C (data not shown).
We then tested whether overexpression of WOR1 drove the phenotype to opaque in the ras1/ras1, pde2/pde2, cdc35/cdc35, tpk1/tpk1 and tpk2/tpk2 mutants by transforming these mutants with a construct in which WOR1 was under the regulation of the inducible MET3 promoter [33]. In the activated state, 100% of white cells of all five strains, when plated on either glucose or GlcNAc agar at 25°C, switched to opaque (Figure 4C). These results demonstrated that WOR1 is essential for the induction of switching by GlcNAc and that it plays a role downstream of the Ras1/cAMP pathway.
To test whether threonine phosphorylation is necessary for Wor1 function, the homozygous deletion mutant wor1/wor1 was transformed with the WOR1TA construct, in which the phosphorylatable threonine 67 residue was replaced with the nonphosphorylatable amino acid alanine and the construct placed under the control of the inducible tetracycline promoter (TETp) to generate strain wor1/wor1+TETp-WOR1TA. A control strain wor1/wor1+TETp-WOR1, was generated in which the mutant wor1/wor1 was transformed with a construct containing the native WOR1 ORF under the regulation of the tetracycline promoter. A second control strain, wor1/wor1+vector, was also generated in which wor1/wor1 was transformed with the vector lacking a WOR1 derivative. White cells of the three test strains were grown in liquid glucose medium at 25°C to saturation phase (five days) and then plated on glucose or GlcNAc agar at 25°C. Both the liquid and agar media contained either 50 or 200 µg per ml of the tetracycline analog doxycycline, which had been shown to induce submaximal and maximal levels of expression, respectively [39]. When native WOR1 was overexpressed both in glucose liquid medium and on glucose agar, the switching frequency was 100% at both 50 and 200 µg per ml of doxycycline (Figure 4D). When WOR1TA was overexpressed in both liquid glucose medium and then after plating on glucose agar at either 50 or 200 µg per ml of doxycycline, the frequency of switching was zero percent (Figure 4D). When native WOR1 was overexpressed in both GlcNAc liquid medium and then after plating on GlcNAc agar at 50 and 200 µg per ml of doxycycline, 100% of the colonies underwent switching (Figure 4D). At 200 µg per ml of doxycycline, over 70% of the cells in the opaque colonies exhibited the elongate opaque cell phenotype (Figure 4E). When WOR1TA was overexpressed in both GlcNAc liquid medium and then after plating on GlcNAc agar at 50 µg per ml of doxycycline, 0% of the colonies exhibited switching (Figure 4D). However, when WOR1TA was overexpressed in GlcNAc media at 200 µg per ml of doxycycline, 100% of the colonies were light pink (Figure 4D). Microscopic analysis revealed that 10% of the cells in these pink colonies exhibited the elongate opaque phenotype (Figure 4E). These results suggested that expression of the unphosphorylatable derivative of Wor1, Wor1TA, was capable of inducing switching, but with a 10-fold reduction in efficiency.
Because WOR1 and WOR1TA were fused in frame with GFP in the overexpression mutants, we used confocal microscopy to test whether Wor1TA localized normally to the nucleus and was expressed at the same level as Wor1. Both Wor1 and Wor1TA localized to the nucleus of a majority of cells of the overexpression mutants treated with doxycycline, as demonstrated by overlapping GFP fluorescence and staining with DAPI, a DNA indicator (Figure 5A). Moreover, GFP fluorescence of nuclei was qualitatively comparable for Wor1-GFP and Wor1TA-GFP (Figure 5A). These results demonstrated that although the replacement of threonine with alanine caused a dramatic decrease in its capacity to support switching, it did not affect nuclear localization or cause a decrease in the transcript level. The levels of the Wor1 and Wor1TA protein were then compared by western blot analysis using anti-GFP antibody. The levels of Wor1 and Wor1TA expressed in white cells of strains wor1/wor1+TETp-WOR1 and wor1/wor1+TETp-WOR1TA, respectively, treated with 200 µg per ml of doxycycline were similar (Figure 5B). These results indicate that the decrease in Wor1 function resulting from the replacement of threonine with alanine in the PKA consensus motif of Wor1 was due to a decrease in function, rather than to a decrease in the level of the Wor1 protein or mis-localization.
We had demonstrated that 1% CO2 induced switching submaximally and that at this concentration induction was dependent primarily upon the Ras1/cAMP signal transduction pathway [18]. We have shown here that GlcNAc induction was also submaximal when cells were grown for only two days in glucose liquid medium to mid-log phase (Figure 1C, D). We therefore tested whether cells growing at 25 or 37°C in a suboptimal concentration of CO2 (1%) and for a suboptimal period of time in glucose liquid medium enhanced GlcNAc induction. White cells of an a/a and an α/α strain were first grown at 25°C in glucose liquid medium in air for two days and then plated on either glucose or GlcNAc agar either in air or in air containing 1% CO2 at 25 or 37°C. On glucose agar in 1% CO2 at both temperatures, cells of the a/a and α/α strains exhibited switching frequencies that were significantly higher than in air alone (Table 1). When plated on GlcNAc agar in air at 25°C, the respective frequencies of the two strains were 26.7±1.2% and 22.4±3.3%, but at 37°C, they were 98.9±2% and 99.7±0.6%. However, when plated on GlcNAc agar in 1% CO2, the switching frequency at the two temperatures were 99 to 100% in both strains (Table 1). These results indicate synergy for CO2 and GlcNAc induction, and enhancement by physiological temperature.
Recently, high CO2 was demonstrated to be a potent inducer of the white-opaque transition at 37°C [18]. The high concentrations of CO2 that induce and maintain the opaque phenotype at 37°C are found in select host tissues and in the gastrointestinal tract [40],[41]. The main source of CO2 in the gastrointestinal tract is the result of metabolism by colonic bacteria [40]. Here, we demonstrate GlcNAc, also a product primarily of bacteria that cohabit the host gastrointestinal tract with C. albicans [21]–[23],[42] represents a second potent inducer of the white to opaque transition.
By mutational analyses, we have found that the components of the Ras1/cAMP pathway, Ras1, Cdc35, Pde2 and PKA (Tpk1, Tpk2), mediate the major portion of GlcNAc induction. Our results also indicate that switching in glucose medium, which is at a far lower frequency than that in GlcNAc medium, is mediated in part by the Ras1/cAMP pathway. The low level of induction in deletion mutants of the Ras1/cAMP pathway is still approximately ten-fold higher than the level caused by glucose in wild type cells. Because the pathway that transduces high CO2 induction is also Ras1/cAMP-independent and unidentified, the possibility must be entertained that it may be the same Ras1/cAMP-independent pathway that mediates the minor portion of GlcNAc induction.
Our results, especially those at 37°C, suggest that the PKA isoform Tpk2 functions as the major downstream target of cAMP in the GlcNAc response pathway. Tpk1 appears to be capable of substituting for Tpk2 in the mutant tpk2/tpk2, but when overexpressed, inhibits GlcNAc-induced switching. One possible explanation is that the two Tpk isoforms play inhibiting and stimulating roles, respectively, in the white to opaque transition, especially in light of the fact that Efg1, a negative regulator of Wor1 [8], contains a PKA phosphorylation site. Tpk isoforms have been found to play different roles in the same regulatory networks in other systems. In S. cerevisiae, while Tpk2 activates filamentation downstream of cAMP in the glucose induced pathway, Tpk1 and Tpk3 inhibit filamentation by a feedback loop [36]. In addition, in the pheromone response pathway of C. albicans, the downstream MAP kinases Cek1 and Cek2 also play both distinct and overlapping roles [43], suggesting a general pattern of functional complexity of downstream protein kinase isoforms in signal transduction pathways.
Mutant analysis revealed that both the major and minor GlcNAc induction pathways required the master switch locus WOR1. Because cAMP activates PKA, we considered the possibility that Wor1, which contains one conserved PKA phosphorylation motif between amino acids 64 and 69, might have to be phosphorylated to function in the switch event. By converting the single threonine residue at that site to alanine, GlcNAc induction was impaired dramatically, indicating that phosphorylation of threonine 67 of Wor1 is necessary for maximum induction by GlcNAc. The observation that GlcNAc induction was completely blocked in the wor1/wor1 mutant, but only impaired in the wor1/wor1+TETp-WOR1TA mutant suggested that the constitutively unphosphorylated form of Wor1 was still functional, but at far lower efficiency than the phosphorylated form. In Schizosaccharomyces pombe, the gluconate transport inducer 1 (Gti1), an ortholog of Wor1, also harbors one conserved PKA phosphorylation motif between amino acids 65 and 70, and conversion of the single threonine residue at that site to alanine causes severe impairment of function [44]. As is the case in C. albicans, Pka1, which is the only PKA in S. pombe, is involved in the regulation of Gtil. Given that in C. albicans, Wor1 has one conserved PKA phosphorylation site that must be phosphorylated to attain the major portion of GlcNAc induction, and that Tpk2 appears to be the downstream PKA involved in GlcNAc induction, it seems reasonable to suggest that GlcNAc induction may involve the direct phosphorylation of Wor1 by Tpk2, but that remains to be demonstrated.
The Ras1/cAMP-dependent pathway has been found to be the predominant one for GlcNAc induction and the minor one for low level CO2 induction [18] (Figure 6). An unidentified pathway has been found to be the predominant one for CO2 induction [18] and an unidentified pathway has the minor one for GlcNAc induction (Figure 6). Our data further suggest that glucose represents a weak but significant inducer of switching that also functions through both a Ras1/cAMP-dependent pathway and a Ras1/cAMP-independent pathway, the latter again unidentified (Figure 6). The fact that each inducer functions not only through the Ras1/cAMP pathway, but also through an unidentified pathway, leaves open the possibility that the Ras1/cAMP-independent pathway may also be common to all three inducers. We demonstrated previously that both the Ras1/cAMP-dependent and -independent pathways for CO2 induction are dependent on Wor1 [18], and we have demonstrated here that the dependent and independent pathways for G1cNAc and glucose induction are also dependent on Wor1. We have demonstrated that only the major portion of G1cNAc induction, which is transduced by the Ras1/cAMP pathway, requires the phosphorylated form of Wor1.
The induction of switching by environmental cues shares several characteristics with that of filamentation. First, both CO2 and GlcNAc induce filamentation [19],[20],[24] as they do switching. Second, the Ras1/cAMP pathway has been demonstrated to play a role in the induction of filamentation by CO2 [20], as it does in switching. The Ras1/cAMP pathway has also been demonstrated to play a role in the induction of filamentation in S. cerevisiae [36],[45], suggesting that the pathways regulating of filamentation represent an ancestral process conserved in the evolution of both the Candida and Saccharomyces groups of the hemiascomycetes. Several characteristics of the opaque phenotype are shared with hyphae, including an elongate shape, a prominent vacuole and cell surface antigens [46],[47]. Because white-opaque switching is a specific and unique characteristic of C. albicans and the closely related species Candida dubliniensis [48], it represents a newly evolved developmental process, in contrast to filamentation.
Switching of C. albicans from white to opaque at physiological temperature can therefore be influenced by two factors in the gastrointestinal tract that result primarily from gastrointestinal bacteria: high CO2 [18] and free GlcNAc. In host tissue, high CO2 is the result of metabolism by the host, but in the gastrointestinal tract, it is the product of bacterial metabolism [40],[41]. GlcNAc in the gastrointestinal tract is also a product primarily of gastrointestinal tract bacteria, but also to a lesser extent of host goblet cells [41]. Hence, bacteria of the gastrointestinal tract produce two potent inducers of the white to opaque transition, a prerequisite for mating between a/a and α/α cells [4]. These results suggest that two developmental programs of C. albicans, filamentation and switching, have evolved to respond to signals originating from bacterial co-members of the gastrointestinal microbiota.
The strains of C. albicans used in this study are listed in supplemental Table S2. For routine growth, modified Lee's medium without methionine was used [31], unless stated otherwise. For repression of MET3 promoter-controlled gene expression, 2.5 mM methionine and 2.5 mM cysteine were added to the medium. For GlcNAc induction, the carbon source glucose was replaced with GlcNAc (1.25% w/v) in nutrient medium. Here, agar containing Lee's medium, in which glucose was the carbon source, was referred to as glucose agar and Lee's medium containing GlcNAc as a carbon source was referred to as GlcNAc agar. Agar cultures were grown at a density of 80–120 colonies per 85 mm plate. Phloxine B was added to nutrient agar for opaque colony staining [46].
The PDE2 gene was disrupted using a modified Ura-blaster method [49]. Two long primers (PDE2-5DR, PDE2-3DR), each containing a different 60 nucleotide sequence homologous to the gene PDE2, were used for PCR amplification (supplemental Table S3). pDDB57, which contains the recyclable URA3-dpl200 marker, was used as template. The PCR product was transformed into WUM5A, a WO-1 derivative [50]. Transformants were grown on selective synthetic defined (SD) medium SD-Ura agar plates. To delete the second allele of PDE2, the PCR product was transformed into a spontaneous Ura- derivative of PDE2/pde2 obtained from SD agar containing 5-fluoro-orotic acid. pde2/pde2 null mutants were selected from SD-Ura agar plates and confirmed by PCR.
TPK1 and TPK2 were deleted by a PCR product-directed disruption protocol, as described in [18]. Briefly, the HIS1 and ARG4 markers were amplified by PCR from pGEM-HIS1 and pRS-ARG4-SpeI, respectively. The oligonucleotide pairs TPK1-5DR, TPK1-3DR; TPK2-5DR, TPK2-3DR (supplemental Table S3) were used for PCR amplification. The HIS1 and ARG4 markers were sequentially transformed into the host strains GH1013 [51], and heterozygous mutants. The null mutants were selected on SD-His-Arg plates and confirmed by PCR.
The primers used for plasmid constructions are listed in supplemental Table S3. To generate pMET-RAS1V13, the RAS1 ORF containing a mutation at the thirteenth amino acid (glycine to valine mutation) was amplified from pQF145.2 [25] by using primers including PstI and SphI sites, and then cloned into pCaEXP [33]. pMET-CDC35 was constructed by inserting the BamHI-SphI digested CDC35 ORF fragment into pCaEXP. To generate pMET-PDE2, the PDE2 ORF was amplified from CAI4 genomic DNA by using the primers PDE2F and PDE2R that contained BamHI and SphI sites (supplemental Table S3) and the PDE2 ORF was cloned into pCaEXP. To generate pACT-TPK1, the TPK1 ORF was amplified from CAI4 genomic DNA by using the primers TPK1F and TPK1R that included EcoRV and HindIII sites (supplemental Table S3), and the TPK1 ORF was cloned into pACT1 [18]. The TPK2 ORF was amplified from CAI4 genomic DNA by using primers TPK2F and TPK2R (supplemental Table S3). To generate pACT-TPK2, the PCR product was digested by HindIII and cloned into EcoRV/HindIII-digested pACT1. pNIM-WOR1 was constructed by inserting a SalI digested PCR fragment of WOR1 into the SalI site of pNIM1 [52]. The WOR1 ORF was amplified from CAI4 genomic DNA. The primers WOR1salF and WOR1salR were used for PCR amplification (supplemental Table S3). To generate a mutation in which threonine 67 is replaced with alanine in the WOR1 gene, a two-step PCR method [53] was used with slight modification. The primers WOR1TAF and WOR1TAR (supplemental Table S3) were used to generate site-directed mutation. The second-round PCR product was digested with SalI and subcloned into the SalI site of pNIM1. The resulting plasmid was referred to as pNIM-WOR1TA. The correct direction of WOR1 ORF and WOR1TA fragment in pNIM1 was confirmed by sequencing.
White-opaque switching on agar was analyzed as described previously [15]. Briefly, strains were first grown on agar containing supplemented Lee's medium for 6 days at 25°C. Colonies were then replated onto plates containing supplemented Lee's medium [31]. These plates were then incubated at 25°C for five days, and the proportion of colonies exhibiting different phenotypes counted.
White colonies were inoculated into a test tube containing 1 ml of supplemented Lee's medium [31] with glucose as carbon source and incubated at 25°C. The overnight culture was diluted (to 2×105 cells/ml) in 20 ml of fresh medium with glucose as the carbon source and incubated at 22°C in a shaker. Aliquots were taken out at different time points, diluted and plated onto both glucose agar and GlcNAc agar plates (Figure 1A). The plates were then incubated at 25°C for five days, and the colonies exhibiting different colony phenotypes counted. For the experiments performed at 37°C, plates were cultured at both 25 and 37°C.
Cells from liquid cultures were spun down following doxycycline treatment for 12 hours. Total protein extract was obtained using a bead beater in lysis buffer that contained 50 mM Tris-HCl, 100 mM NaCl, 5 mM MgCl2, 1 mM DTT, 1 mM EDTA, 1 mM EGTA, 0.1% Tween-20, and 5% glycerol, supplemented with a protease inhibitor cocktail (Sigma-Aldrich, St Louis, MO) and 1 mM phenyl-methylsulphonyl fluoride. An equal amount of total protein from each sample was then subjected to protein G beads (Active Motif, Carlsbad, California) for pre-clearing, followed by immuno-precipitation (IP) using rabbit GFP antibody-conjugated agarose beads (Santa Cruz Biotechnology, Santa Cruz, California). IP protein samples were subjected to SDS-PAGE (8% polyacrylamide) electrophoresis. After electrophoresis, the SDS-PAGE protein gel was transferred to a PVDF membrane (Immobilon-P, Millipore Corporation, Bedford, MA), blocked for 1 h in 3% non-fat dry milk in TBS-T (20 mM Tris-HCl, pH 7.5, 150 mM NaCl, 0.05% Tween-20), and then incubated with rabbit polyclonal GFP antibody (Santa Cruz Biotechnology, Santa Cruz, CA) overnight at 4°C [54]. After washing six times in TBS-T, the proteins on the membrane were detected with horseradish peroxidase-labelled goat anti-rabbit IgG (Promega, Madison, WI) and SuperSignal West Pico Chemiluminescent Substrate (Pierce, Rockford, IL).
Cells expressing tetracycline (doxycycline)-inducible GFP-labeled Wor1p were grown to midlog phase in the presence of 50 µg/ml doxycycline (Sigma-Aldrich, St Louis, MO, USA), harvested and simultaneously permeabilized and the nuclei labeled with 4′,6′-Diamidino-2-phenylindole (DAPI, Invitrogen, Inc.) by incubating them for 10 min at room temperature in the dark in a solution containing 5 µg/ml DAPI in 1 M Sorbitol, 0.1% Saponin, 150 mM NaCl and 20 mM Tris buffer, pH 7.4, followed by a 15–20 min incubation period on ice. Without washing, the cells were imaged using a Bio-Rad Radiance 2100 MP multi-photon microscope (Bio-Rad, Hermel, Hamstead, UK). Cells were excited at 780 nm by a Mai-Tai laser (Spectra- Physics, Newport Corp., Mountain View, CA) and three channel emission images (GFP, DAPI and transmitted) were gathered using a sequential 2.0 µm Z-series, gathered at 0.2 µm intervals to include the entire cell nucleus. GFP and DAPI images were visualized as Z-series projections. Transmitted images were a single scan at the focal plane selected from the Z-series.
|
10.1371/journal.pgen.1001249 | Functional Comparison of Innate Immune Signaling Pathways in Primates | Humans respond differently than other primates to a large number of infections. Differences in susceptibility to infectious agents between humans and other primates are probably due to inter-species differences in immune response to infection. Consistent with that notion, genes involved in immunity-related processes are strongly enriched among recent targets of positive selection in primates, suggesting that immune responses evolve rapidly, yet providing only indirect evidence for possible inter-species functional differences. To directly compare immune responses among primates, we stimulated primary monocytes from humans, chimpanzees, and rhesus macaques with lipopolysaccharide (LPS) and studied the ensuing time-course regulatory responses. We find that, while the universal Toll-like receptor response is mostly conserved across primates, the regulatory response associated with viral infections is often lineage-specific, probably reflecting rapid host–virus mutual adaptation cycles. Additionally, human-specific immune responses are enriched for genes involved in apoptosis, as well as for genes associated with cancer and with susceptibility to infectious diseases or immune-related disorders. Finally, we find that chimpanzee-specific immune signaling pathways are enriched for HIV–interacting genes. Put together, our observations lend strong support to the notion that lineage-specific immune responses may help explain known inter-species differences in susceptibility to infectious diseases.
| We know of a large number of diseases or medical conditions that affect humans more severely than non-human primates, such as AIDS, malaria, hepatitis B, and cancer. These differences likely arise from different immune responses to infection among species. However, due to the lack of comparative functional data across species, it remains unclear how the immune system of humans and other primates differ. In this work, we present the first genome-wide characterization of functional differences in innate immune responses between humans and our closest evolutionary relatives. Our results indicate that “core” immune responses, those that are critical to fight any invading pathogen, are the most conserved across primates and that much of the divergence in immune responses is observed in genes that are involved in response to specific microbial and viral agents. In addition, we show that human-specific immune responses are enriched for genes involved in apoptosis and cancer biology, as well as with genes previously associated with susceptibility to infectious diseases or immune-related disorders. Finally, we find that chimpanzee-specific immune signaling pathways are enriched for HIV–interacting genes. Our observations may therefore help explain known inter-species differences in susceptibility to infectious diseases.
| Due to our natural focus on humans, we know of a large number of diseases or medical conditions that affect humans more severely than non-human primates. Examples include progression to AIDS following infection with HIV, progression to malaria following infection with Plasmodium falciparum, Alzheimer's disease, cancer, and adverse complications following infection with hepatitis B and C (reviewed in [1], [2]). Differences in susceptibility to infectious agents between humans and other primates might be explained, at least in part, by inter-species differences in immune response to infection. Indeed, a large body of work indicates that immune systems are rapidly evolving. In particular, while very little comparative functional data in primates has been collected, recent genomic scans for signatures of natural selection have reported that genes involved in immunity processes are strongly enriched among targets of positive section in human and chimpanzee [3]–[11].
Immune responses are typically classified as either ‘innate’ or ‘adaptive.’ Historically, the focus of most immunological studies has been on the adaptive response and its hallmarks, namely the generation of a large repertoire of antigen-recognizing receptors and immunological memory. Recently, however, more effort has been expended on understanding the innate immune system, as it became clear that innate immunity is an evolutionarily ancient defense mechanism, which governs the initial detection of pathogens and stimulates the first line of host defense [12]–[15]. Moreover, innate immune responses were shown to play a pivotal role in the development of pathogen-specific humoral and cellular adaptive immune responses, which are mediated by B and T cells [16]–[18].
The recognition of pathogens by the innate immune system is primarily mediated by phagocytic cells (e.g., monocytes, macrophages, and dendritic cells) through germline-encoded receptors, known as pattern recognition receptors (PRRs) [17], [19], [20]. The PRRs recognize conserved molecular features characteristic of the microbial world, commonly referred to as pathogen-associated molecular patterns (PAMPs) [17], [19], [20]. Among the different PRRs, the Toll-like receptor (TLR) family, which comprises 10 functional members in humans, has been the most extensively studied. For example, by stimulating primary cell cultures with different TLR agonists in vitro (e.g., references [21]–[23]) and by studying mouse models that lack one or several TLRs (e.g., references [24]–[26]), it has been shown that TLRs can be activated in response to virtually any microbe that invades the host.
Once activated, TLRs play a crucial role in orchestrating the response to pathogenic microbial infections through the induction of two major regulatory programs. First, a universal regulatory response, which can be activated by all TLRs and is triggered by infection with a diverse range of microbes or TLR agonists [21], [22], [27]–[29]. This response has been interpreted as a generic ‘alarm signal’ for infection [22], [28], [29]. Second, individual TLRs can activate regulatory programs that are specific to individual microbial agents [19]. Comparative functional studies of TLR-mediated immune response in primates might therefore shed light on inter-species differences in susceptibility to certain infectious agents. However, at present, there is very little functional data with which one can study the evolution of the immune system in primates.
In order to study functional differences between the innate immune response of humans and two close evolutionary relatives, chimpanzees (Pan troglodytes) and rhesus macaques (Macaca mulatta), we stimulated primary monocytes from six individuals from each of the three species with LPS for 4, 12, and 24 hours (see Figure S1 for an illustration of the study design). LPS activates the TLR pathway (specifically, TLR4) and mimics an infection with Gram negative bacteria [16]–[18]. We chose this treatment because LPS, via TLR4, activates multiple immune signaling pathways, leading to the induction of both inflammatory and ‘viral-like’ responses [19]. Additionally, the stimulation of immune cells with LPS was shown to result in a very similar regulatory response (87% overlap) to the response to infection with a live bacteria such as E. coli [22].
To confirm that the LPS treatment activated TLR4-mediated immune responses, we used quantitative PCR to estimate the induction levels of three inflammatory cytokines (IL6, IL1β, and TNF). In all samples (from all individuals at all time points), levels of the three inflammatory cytokines were significantly higher following stimulation with LPS (Figure S2). However, we noticed that the quantitative responses to the treatment in the chimpanzee samples were lower than those of the human and rhesus macaque samples. This observation probably reflects a technical difficulty in culturing chimpanzee primary monocytes without inducing a general stress response, which results in the attenuation of the quantitative response to further stimuli (see Materials and Methods for more details). In what follows, we therefore focus primarily on qualitative rather than quantitative differences between individuals and species in the regulatory response to stimulation with LPS.
To estimate and compare gene expression levels in samples from multiple species, we used a multispecies microarray, which includes orthologous probes from human, chimpanzee, and rhesus macaque for 18,109 genes [30]. Following processing and normalization of the array data, we used a gene-specific linear mixed-effect model (see Materials and Methods) to identify inter-species differences in the regulatory response to stimulation with LPS (the ‘treatment’). To minimize the number of falsely identified differences across species, we applied two statistical cutoffs for classifying genes as responding to the treatment. Specifically, conditional on observing a treatment effect with high statistical confidence in one species, we assumed that a treatment effect likely occurred in other species as well, and relaxed the statistical cutoff for the classification of such secondary observations (see Materials and Methods for more details and the specific statistical cutoffs used). This procedure minimizes the number of falsely identified inter-species differences that might ultimately arise from incomplete power to identify differences in gene expression levels following the treatment.
Using this approach, we identified 3,170 genes whose expression levels changed following the treatment in at least one species, at any time point, of which 793 genes responded in all three species (Figure 1A, Table S1, Figure S3). As expected, genes that responded to stimulation with LPS in all three species are enriched with genes involved in immune-related biological processes such as “inflammatory-response” and “cytokine-signaling”, as well as in specific immune-related pathways including the Toll-like receptor pathway, cytokine-cytokine receptor interactions, and the Jak-STAT signaling pathway (FDR for all reported results is <0.01; Figure 1B, 1C, Table S2). Consistent with previous observations in functional studies of the immune system in mice [21], [27], we found that the conserved regulatory response to stimulation with LPS in primates included an enrichment of genes that are likely regulated by the transcription factor NF-kB (P<10−5) and several interferon regulatory factors (e.g., IRF7 and IRF1; P<10−3; Figure 1D, Table S3). Put together, these observations clearly demonstrate that the monocytes from all three species responded to the treatment with LPS by engaging TLR4-mediated regulatory pathways [19], leading to the induction of pro-inflammatory and anti-viral immune responses via the activation of NF-kB and IRF mediated pathways.
To gain further insight into the evolution of LPS-induced immune responses in primates, we classified genes as participating in either the universal regulatory response to infection (which can be triggered by a diverse range of microbes or TLR stimuli [21], [22], [27]), or in the microbial-specific response (which we then further classified as responses to either bacterial or viral infections [21], [22], [27]). Based on these classifications (Table S4), we examined how genes falling into each of the categories responded to infection in humans, chimpanzees, and rhesus macaques. We found that the majority (58%) of genes involved in universal response to infection showed a conserved regulatory response to stimulation with LPS in all three species, compared to only 31% of genes known to respond primarily to either viral or bacterial infection (χ2 test, P<0.001; Figure 2A). Viewed from a different perspective, we observed that the proportion of genes involved in immune response to viral infections is significantly higher (1.5-fold) among genes that responded to stimulation with LPS in only one of the species, compared with genes that responded to the treatment in all three species (χ2 test, P = 0.002; Figure 2B). Taken together, our data strongly support the notion that the universal TLR response is mostly conserved across primates and that much of the divergence in immune response is observed in genes that are involved in response to specific microbial and viral agents.
We proceeded by focusing on species-specific immune responses. Using the conservative approach described above, we identified 335, 273, and 393 genes as responding to stimulation with LPS exclusively in the human, chimpanzee, and rhesus macaque monocytes, respectively (see Figure 3A–3C for examples). To characterize these gene sets, we considered functional annotations based on the GO and KEGG databases (Table S5, S6, S7). Somewhat surprisingly, the only significant enrichments (after correction for multiple tests) were observed among the 335 genes that responded to the treatment exclusively in humans. We found that human-specific immune response was enriched for genes in pathways previously associated with cancer (e.g., Chronic myeloid leukemia or prostate cancer; P≤3.0×10−3, FDR<0.06), the B cell receptor signaling pathway (P = 3.2×10−3, FDR = 0.06), and pathways related to apoptosis (P = 5.0×10−3, FDR = 0.07; see Table S5 for a complete list of significant results). Further, by using the STRING database [31] to visualize all known functional interactions between these 335 genes, we found that 151 of the genes in this set (45%) are known to interact with each other – using the default cutoff suggested by STRING to define a functional interaction (Table S8). Applying a more stringent cutoff (a STRING confidence-score higher than 0.7), we identified 78 genes (23%) that interact with each other, in a functional module that is enriched with genes involved in cancer biology and apoptosis pathways (Figure 4, Table S9). In order to obtain further support for interactions across these 78 genes, we used GRAIL, a tool that uses text mining of PubMed abstracts to identify published functional interactions between genes. We found that 43 out of the 78 human-specific immune response genes (55%) had a GRAIL score of P-text<0.05, statistically supporting the notion that they have a functional interaction with at least one other gene in the list (only ∼7% of genes are expected to have GRAIL score of P-text<0.05 in randomly chosen sets of 78 genes).
We then considered networks of co-expressed genes (namely genes with coordinated patterns of expression) for each species, to find additional putative modules of interacting genes (see Materials and Methods). We found 33, 17 and 32 regulatory modules in humans, chimpanzees and rhesus macaques, respectively, with an average connectivity (|r|) higher than 0.5 (Figure 5, Table S10, S11, S12). Based on 100 random permutations of the gene expression values, we estimated that the number of clusters with |r|>0.5 expected by chance alone is 1.28±1.04, 1.16±1.08, or 1.37±1.08, using data from humans, chimpanzees and rhesus macaques, respectively, suggesting that the observed excess of regulatory modules likely describe meaningful biological relationships.
In humans, the largest regulatory module contains 82 genes, which are significantly enriched for several biological processes involved in extracellular matrix remodeling (P<0.002; Table S10). The second largest regulatory module is significantly enriched for genes involved in apoptotic pathways (P<0.04; Table S10). Interestingly, “apoptosis-related” processes appeared to also be enriched among genes in one of the largest regulatory modules identified in chimpanzees, with 23 genes (Table S11), as well as among three large regulatory modules (>30 genes each) identified in rhesus macaques (Table S12). As the regulatory modules comprise of mutually exclusive sets of genes across the three species (by the nature of the analysis), these observations support the notion that immunological-associated apoptosis mechanisms evolve rapidly in primates.
Finally, we asked whether the observed inter-species differences in immune response might provide insight into the mechanisms underlying differences in susceptibility to infectious diseases between humans and non-human primates. To do so, we considered the subsets of genes that responded to stimulation with LPS exclusively in humans, chimpanzees or rhesus macaques, and examined whether they were enriched for genes previously reported to be associated with immune disorders in humans and/or susceptibility to infectious diseases (see Materials and Methods).
We found an enrichment of “immune-related-disease-genes” among genes that responded to the treatment with LPS exclusively in humans (χ2 test, P = 0.03; Figure 6A). Interestingly, we also found that the set of genes that responded to stimulation with LPS exclusively in chimpanzees was enriched with genes that code for host cell proteins known to interact with HIV-1 (Figure 6B; χ2 test, P = 0.0002). No significant enrichment of HIV-1 interacting genes was observed among genes that responded to stimulation with LPS exclusively in either humans or rhesus macaques. This observation is robust with respect to the specific cutoffs used to classify genes that responded to stimulation in LPS in only one species (Figure S5).
We have performed a genome-wide study of LPS-mediated immune responses in primary monocytes from humans, chimpanzees, and rhesus macaques. Our study design allowed us to characterize conserved innate immune response mechanisms in primates as well as to identify species-specific regulatory responses to stimulation with LPS.
An important difficulty of all studies of gene regulation in primary tissues from primates, apes in particular, is the inability to stage the environment for each of the donor individuals across species. In our study, biological replication within species partially addresses this difficulty, but the possibility that a subset of the observed inter-species differences in gene regulation are due to differences in environments (e.g., diet) across species still exists. An additional difficulty is that most available tools for manipulating cell cultures and performing immune-related assays have not been optimized to work with non-human primate cells. We addressed this issue by performing a large number of quality controls, including the validation of the response to stimulation with LPS in each cell culture by using qPCR. Nevertheless, our observation of systematic differences in the quantitative response to infection across cultures from different species (in particular, from chimpanzees) probably has a technical rather than a biological explanation. For that reason, we chose to draw conclusions primarily based on qualitative differences between species. Thus, the inter-species regulatory differences reported in our study likely provide a lower boundary for the actual number of differences in immune response between humans, chimpanzees and rhesus macaques.
We identified 793 genes that responded to stimulation with LPS in all three species. As expected, this set of genes was significantly enriched for genes involved in immune responses, and specifically for genes involved in TLR-mediated pathways. Some examples of conserved TLR4-induced immune responses include the strong up-regulation of several pro-inflammatory cytokines, such as IL-6, IL1-β and tumor necrosis factor (TNF), and chemokines, such as CCL2, CCL3 and CCL4, whose roles are to recruit other effector cells to the inflammatory site [29], [32]. We also observed a conserved up-regulation of the anti-inflammatory cytokine IL-10, probably to control the levels of inflammatory response and avoid tissue damage [29], as well as the up-regulation of several interferon-α inducible genes (e.g., IFIH1, IFIT1, and IFIT3).
Overall, conserved immune responses were enriched for genes whose expression levels are regulated by the transcription factor NF-kB, or by several interferon regulatory factors, which are the master regulators of TLR4-dependent pathways. Interestingly, before infection, the expression levels of many of these master regulators (e.g., REL, NFKB1, RELB, IRF2, IRF9) were different across the three species, while post-infection, their expression converged to practically the same level, regardless of species (Figure S6). This observation suggests that the regulatory response of these key transcription factors likely evolve under strong evolutionary constraints, probably to ensure efficient downstream immune responses.
A known property of the regulatory programs mediated by different TLRs is the activation of both a universal response (shared by all TLRs) as well as a response that is specific to each microbial agent (or TLR ligand) [21], [22], [27]–[29]. We found that the universal TLR response is remarkably more conserved across primates compared to microbial-specific responses. From an evolutionary perspective this observation makes intuitive sense. Indeed, ‘core’ immune responses, which are critical to fight any invading pathogen, are expected to be under stronger evolutionary constraint compared to immune programs that are only important in the presence of specific microbial infections. Consistent with this expectation, our data also support the notion that adaptation of innate immune responses in primates primarily took place at the level of ‘peripheral’ responses, namely, pathogen-specific immune responses.
Among genes whose regulation was affected by stimulation with LPS in only one species, we found an enrichment of genes associated with response to viral infections. This observation might reflect the need of the host immune system to frequently devise new defense mechanisms to fight viral infection, as viruses tend to evolve faster than other microbes [33]. We also found that species-specific immune responses are enriched with genes annotated to have a role in apoptotic pathways. Apoptosis is a critical component of successful immune response as infected cells have to be efficiently removed without inciting an inflammatory reaction [34]. Moreover, controlled cell death is used to restore normal cell numbers following clonal expansion of antigen-specific lymphocytes [34]. Consistent with our observation of rapid evolution of the regulation of apoptotic pathways, coding regions of apoptosis-related genes have previously been shown to be rapidly evolving during primate evolution [7], [35]. Put together, these observations suggest that inter-species differences in apoptosis-related immune responses may be adaptive. Although the selective pressures underlying these adaptations are unclear, these observations might help elucidate the basis for important phenotypic differences between humans and non-human primates, such as differences in susceptibility to cancer.
Indeed, cancer incidence in non-human primates is low compared to that observed in humans, even when age is taken into account [1], [2], [36]–[39]. The deregulation of apoptosis has been extensively described as a hallmark of cancers [40]. Thus, while our observation that the human-specific immune response to stimulation with LPS is characterized by a significant enrichment of cancer-related genes is not surprising (because the ‘cancer-related’ and ‘apoptosis-related’ gene sets are not mutually exclusive), it may provide a first step towards understanding the mechanisms underlying the differences in cancer incidence between humans and other primates. For example, we observed that the pro-apoptotic gene CASP10 was strongly down regulated early after stimulation with LPS, exclusively in humans (Figure S7). Somatic mutations in CASP10, as well as reduced expression levels of this gene, were found to be associated with a number of different human cancers [41]–[44]. The observed inter-species differences in the regulation of CASP10 following infection may therefore be related to differences in the rates of cancer across species. Detailed comparative studies of apoptosis-related regulatory mechanisms in model organisms will be necessary to fully explore the possible connection between the predisposition to cancer and inter-species differences in immune responses.
Genes whose regulation was altered following stimulation with LPS exclusively in humans were enriched with genes known to be associated with susceptibility to infectious diseases or to immune-related diseases. We did not observe such enrichment when we considered the immune responses specific to chimpanzees or rhesus macaques. Our observations make intuitive sense, as we know more about the genes associated with diseases that affect humans than those that affect the two non-human primate species. In other words, it is reasonable to assume that we would have found similar enrichments in chimpanzees and rhesus macaques if we knew more about the genetic basis of infectious and immune-related diseases that primarily affect these two species. Our observations thus underscore the link between species-specific immune responses and susceptibility to infectious disease.
One interesting example is the enrichment of genes known to interact with HIV among genes whose regulation was affected by the stimulation with LPS exclusively in chimpanzee. This observation is intriguing because, unlike humans and rhesus macaques, a large number of studies propose that chimpanzees only rarely develop AIDS following infection with HIV [45]–[47]. This notion has recently been challenged by Keele et al. [48], who reported that wild chimpanzees naturally infected with SIVcpz do develop hallmarks of AIDS. The apparently contradictory observations in the literature might be explained by the fact that Keele et al. used data collected from an eastern subspecies of chimpanzees (Pan troglodytes schweinfurthii), whereas previous observations of increased protection from AIDS, were based on studies with the western chimpanzee subspecies (Pan troglodytes verus), the one used in our study. Differences in susceptibility to HIV between sub-species of chimpanzees might be explained by the fact that Pan troglodytes verus were physically separated from the two other sub-species prior to systemic infection with the two recombinant monkey viruses of SIVcpz.
Some examples of HIV-interacting gene that responded to stimulation with LPS exclusively in chimpanzees include ITGB2(CD18) and ITGAM(CD11b), which are the two members of the complement receptor 3 (CR3) that have been shown to play a key role in the infection of dendritic cells by C3-opsonized HIV [49], [50] and the viral transfer to CD4 T cells [50]. Interestingly, these two genes were down-regulated after LPS stimulation, exclusively in chimpanzee monocytes. Another example is the APOBEC3F gene, which is one of the most potent inhibitors of HIV replication [51] and was significantly up-regulated in response to stimulation with LPS, only in chimpanzee monocytes. The direction of regulatory change in these cases (namely, the down regulation of a receptor that may be used by the HIV virus, and the up-regulation of a known inhibitor of HIV replication), is consistent with a theoretical mechanism of increased resistance of chimpanzees (or at least Pan troglodytes verus) to progression of AIDS. That said, future studies are now required to evaluate if the down-regulation of CR3 and up-regulation of APOBEC3F are also observed after infection with HIV (or SIVcpz).
Our observations may reflect an adaptation of the chimpanzee immune system to infection with HIV/SIV or perhaps to other retroviral infection(s). Previous studies of variation at the nucleotide level have reported that genes associated with HIV infection (such as CD45, APOBEC3G and APOBEC3H) evolved under positive selection in primates [52], particularly after the divergence of humans and chimpanzees [11]. Taken together, our data suggest that regulatory changes occurring specifically in the chimpanzee lineage might explain, at least in part, why chimpanzees tend not to progress to AIDS following infection with HIV/SIV.
More generally, our observations may help to explain other inter-species differences in susceptibility to infectious agents, such as the increased resistance of chimpanzees to certain other viral infections, including hepatitis B and C, and influenza A. Our study, however, is only the first step in characterizing inter-species differences in immune response, in particular because LPS is a general stimulant. We expect future comparative studies in primates to focus on the immune response to different individual infectious agents.
This study was conducted according to the principles expressed in the Declaration of Helsinki. An IRB approved consent form was obtained from each human donor. Collection of the non-human primate samples was perform at the Yerkes Primate Center, in a manner that conformed to the animal subject regulatory standards enforced by the Emory University Institutional Animal Care and Use Committee (IACUC approved protocol #028-2009Y).
We measured gene expression levels in blood monocytes from six humans, six chimpanzees and six rhesus macaques (three males and three females from each species; see Table S13 for details on all samples). Blood samples were collected in BD Vacutainer CPT Cell Preparation Tube (BD, Franklin Lakes, NJ) and peripheral blood mononuclear cells (PBMCs) were purified according to the manufacturer's instructions. Non-human primate blood samples were collected at the Yerkes primate center and human samples were obtained from Research Blood Components.
Blood monocytes from the three species were purified from PBMCs using magnetic cell sorting technology (MACS technology from Miltenyi Biotech). Specifically, monocytes from humans and rhesus macaques were purified by positive selection with magnetic CD14 MicroBeads (Miltenyi Biotech). This method did not work well with the chimpanzee samples (less than 3% of chimpanzee PBMCs were isolated using cell sorting with a CD14 antibody). Instead, monocytes from chimpanzees were purified by depletion of non-monocyte cell types using the “Monocyte Isolation Kit II” (Miltenyi Biotech). Regardless of the method used, the purity of the isolated monocyte population was evaluated by flow cytometry. For the human and rhesus macaque samples we used a fluorochrome-conjugated antibody against monocytes (CD14-FITC; Beckmancoulter). For chimpanzees, we further confirmed the purity of the monocyte population by also using two additional fluorochrome-conjugated antibodies - against B-cells (CD20-PE; BD Bioscience) and T-Cells (CD3-APC BD Bioscience). Regardless of species, only samples with monocyte purity higher than 80% were used in subsequent experiments (Figures S8 and S9). Finally, to further increase the purity of the monocyte-enriched fraction (virtually to 100%), we performed an additional selection step based on the unique capacity of monocytes/macrophages to strongly adhere to the plastic of cell culture dishes. To do so, we cultured the cells overnight (see below for details) and washed the cell culture wells in the morning to retain only adherent cells (i.e., monocytes).
Since we used different methods to purify monocytes in samples from chimpanzees (negative selection) than in samples from humans or rhesus macaques (positive selection), we also performed a control experiment to empirically evaluate to what extent the purification method used affects the ability of the purified cells to respond to LPS stimulation. To do so, we collected whole blood samples from three additional humans, from which we purified monocytes using both a positive and a negative selection method. We then performed the same LPS treatment experiment we applied to the main samples from all three species (described below), and compared differences in response to LPS treatment between monocytes purified by positive selection and monocytes purified by negative selection. As discussed in Text S1 and Figures S15 and S16, we found that the method used to purify the monocytes (negative or positive selection) has only a minimal effect on the measured regulatory response to stimulation with LPS. That said, to be conservative, we excluded from all analyses presented in the manuscript the 192 genes identified as responding to LPS treatment only in monocytes purified by either positive or negative selection.
Monocytes were cultured in 24-well cell culture plates (Corning) in serum free media (CTL's test media) at a density of 1 million cells per ml. We used a serum free media to minimize the probability that the monocytes were non-specifically activated as a result of the undefined nature of serum products (e.g., as a result of a mitogenic serum batch). The cells were then stimulated with 1 ug/ml of LPS (Invivogen, Ultrapure LPS, E. coli 0111:B4) for 4, 12, and 24 hours. These time points were chosen based on previous observations that the transcription kinetics of immune response to infection can generally be characterized by early, middle, and late phases of response, which can be effectively captured at 4, 12 and 24 hours post infection [22]. All time course experiments were started at the same time of day (∼8am) to prevent the introduction of variation due to differences in circadian rhythm. Untreated cell cultures were kept alongside the stimulated cultures and were harvested at the same time intervals (4, 12, and 24 hours post stimulation).
Total RNA from each cell culture was extracted using RNeasy columns (Qiagen, Valencia, CA). For all samples, RNA quantity was evaluated spectrophotometrically, and the quality was assessed with the Agilent 2100 bioanalyzer (Agilent Technologies Inc, Palo Alto, CA). Only samples with no evidence for RNA degradation (RNA integrity number >8.5) were retained for further experiments. To evaluate the activation of monocytes after stimulation, we used quantitative PCR to test for an over-expression of Tumor-necrosis Factor (TNF-a), Interleukine-6 (IL-6) and IL1-β. inflammatory cytokines that are known to be induced following the activation of TLRs (primer sequences and PCR conditions can be found in Table S14). Only samples for which we observed a significant induction of these cytokines were used in downstream experiments. Once we confirmed that monocytes from all three species responded to the treatment, we performed linear amplifications of the total RNA samples by using in-vitro transcription. Specifically, 400 ng of high-quality total RNA were amplified using the MessageAmp II kit (Ambion). Unlike exponential RNA amplification methods, aRNA amplification has been shown to maintain the relative representation of the starting mRNA population [53], [54] (Figure S10).
To compare genome-wide gene expression levels between humans, chimpanzees, and rhesus macaques, we hybridized the RNA samples to the multi-species microarray described by Blekhman et al. [30]. This array contains orthologous probes from the three species, thus allowing a comparison of gene expression levels between species without the confounding effects of sequence mismatches on hybridization intensities [30]. The microarray contains probes for 18,109 genes (see Blekhman et al. [30] for a detailed description of the multi-species array). The labeling of the amplified RNA samples and subsequent hybridization to the microarray were performed by Nimblegen. For each individual we hybridized one non-stimulated and one stimulated sample at each of the three time points (4 hours, 12 hours and 24 hours). The total number of arrays analyzed was therefore 108 ( = 3 species ×6 individuals ×6 arrays per individual). Quality control, background correction and normalization of the expression data were performed as previously described [30] (Figure S11 and S12).
All the statistical analyses detailed in this and the following sections were performed using the R statistical environment (http://www.r-project.org).
We used GeneTrail (http://genetrail.bioinf.uni-sb.de) [56] to test for enrichment of functional annotations among different classes of genes (as detailed in the results). In all tests, we used a background set of 13,244 genes, which were classified as expressed (using an absolute log2 intensity cutoff of 7.5; Figure S14) in at least one species at one condition; that is, any of the time points for the non-stimulated or stimulated samples. The tests were performed using all GO categories and KEGG pathways. We calculated p-values using a Hyper-geometric distribution, and used the approach of Benjamini and Hochberg [55] to control the false discovery rate.
We applied the promoter analysis algorithm PRIMA implemented in the EXPANDER package [57]. Given a target set and a background set of genes, PRIMA identifies transcription factor binding motifs that are significantly more prevalent in the promoter of the target set than in the background set. As background we used all the genes in the array that had at least 3 homologous probes across species and that were expressed in at least one species in one of the six conditions (i.e., any of the three time-point in either treated or non-treated samples).
We looked for known functional associations between the 335 human-specific LPS response genes using the STRING database (http://string.embl.de/). STRING is a database of both known and predicted protein-protein interactions, which includes direct (physical) and indirect (functional) associations derived from numerous sources, including experimental repositories, computational prediction methods and public text collections [31]. We selected all interactions/associations available for a given node with a combined score greater than 0.7. The scores given in the STRING database define the confidence limit for each described interaction/association. A combined score of 0.7 is recommended as the high stringency criterion by the database authors.
In turn, to identify modules of co-expressed genes, we used the Modulated Modularity Clustering (MMC) algorithm, which seeks community structure in graphical data; that is, a graph of genes connected by edges whose weights reflect the degree to which their transcriptional profiles are correlated (see reference [58] for a detailed description of the method). Co-expression modules were defined using the probe-corrected expression estimates for the set of 335, 273 and 393 genes whose expression levels were altered following the treatment exclusively in humans, chimpanzees and rhesus macaques, respectively.
Genes previously reported to be associated with immune disorders and/or susceptibility to any infectious disease were identified using the Genetic Association Database (http://geneticassociationdb.nih.gov/). We downloaded the full GAD dataset on Nov 9, 2009, and parsed the all.xls table, which contains gene-disease associations. We labeled genes as immune-related if they had been associated with diseases for which the ‘disease class’ field in the GAD database was defined as ‘Immune’ or “Infectious diseases”. The list of host genes known to interact with HIV-1 proteins was retrieved from the HIV-1 Human Protein Interaction Database, which catalogues over 1,400 human proteins reported in the scientific literature to participate in HIV-1 to human protein interactions [59].
|
10.1371/journal.pcbi.1004785 | Motor Demands Constrain Cognitive Rule Structures | Study of human executive function focuses on our ability to represent cognitive rules independently of stimulus or response modality. However, recent findings suggest that executive functions cannot be modularized separately from perceptual and motor systems, and that they instead scaffold on top of motor action selection. Here we investigate whether patterns of motor demands influence how participants choose to implement abstract rule structures. In a learning task that requires integrating two stimulus dimensions for determining appropriate responses, subjects typically structure the problem hierarchically, using one dimension to cue the task-set and the other to cue the response given the task-set. However, the choice of which dimension to use at each level can be arbitrary. We hypothesized that the specific structure subjects adopt would be constrained by the motor patterns afforded within each rule. Across four independent data-sets, we show that subjects create rule structures that afford motor clustering, preferring structures in which adjacent motor actions are valid within each task-set. In a fifth data-set using instructed rules, this bias was strong enough to counteract the well-known task switch-cost when instructions were incongruent with motor clustering. Computational simulations confirm that observed biases can be explained by leveraging overlap in cortical motor representations to improve outcome prediction and hence infer the structure to be learned. These results highlight the importance of sensorimotor constraints in abstract rule formation and shed light on why humans have strong biases to invent structure even when it does not exist.
| Humans’ ability to create abstract rule structures contributes greatly to intelligence and higher cognitive functions as it affords flexible re-use across various sensory-motor transformations. However, how such rule structures develop through learning is poorly understood. Models of this process imply that cognitive rule learning scaffolds on top of mechanisms that support motor action selection learning. Here, we show that the form of motor demands across multiple stimulus-action-outcome influences the form of the abstract rule structures that are created. Our research highlights the strong reciprocal influences of low and high-level cognition, offers further insight into how we learn hierarchical structures, and reminds us that how we act constrains how we think.
| Making decisions in the complex environment of daily life often requires cognitive control to flexibly adjust our behavior: the appropriate reaction to different sensory events often depends on the current context, our goals, etc. For example, while driving, if you see a red light on your street, you will stop; if it turns green, you will go–but the opposite actions apply when the red/green light is in the intersecting street. The cognitive control literature has shown that humans use contextual cues to select abstract sets of rules (or task-sets) that are behaviorally relevant [1,2]. Specifically, cognitive control relies on a cascading hierarchical structure, where at a more abstract level, subjects choose task-sets appropriate to the context, which then constrain our action choices in response to lower-level, less abstract stimuli [3,4]. Thus, we use the term “context” to refer to those features that cue abstract task-set, and the term “stimulus” to refer to features that cue the appropriate response conditioned on the selected task-set.
In the study of executive functions, researchers tend to focus on discretized aspects of decision making–often assuming that perception systems have transformed complex, multi-dimensional sensory signals into reduced, discrete stimuli (e.g., a red circle), and that given this percept, the executive system selects among a few discrete options (e.g., left vs. right), which are then implemented by the motor system. However, the field of embodied or grounded cognition [5–7] offers strong hints that this model is over-simplified, emphasizing instead that executive functions evolved for the control of action in continuous time [5], and are thus scaffolded on existing sensorimotor processing systems. One hint that executive functions are grounded in sensorimotor processing [8,9], is the inability to isolate functionally distinct neural systems for executive functions from motor execution, such that for example, cerebellum, mostly thought of as a motor system, is strongly involved in cognitive control [10,11]. Another hint comes from the organization of cortico-basal ganglia loops, where prefrontal-striatal connectivity parallels that of premotor-cortex for action selection, with hierarchical influence of more anterior loops representing cognitive rules over more posterior loops involved in selecting motor actions [12–15]. Through learning, the consequences of motor actions are leveraged not only to improve future motor choices, but also to improve selection of higher order rules [12–16] again implying a scaffolding of more abstract “cognitive” action selection onto motor action selection. Similarly, striatal dopamine manipulations have analogous effects on reinforcement learning and cognitive action selection [17,18]. However, to date, we know of no direct evidence to indicate that constraints within the motor system influence the way our brain represents the more abstract variables needed for cognitive control, such as task-sets. Here, we propose that an intrinsic constraint of motor action representations strongly influences how we create representations of abstract hierarchical rule structures, both during learning and while applying instructed rules.
Recent research has shown that humans can acquire abstract rule representations through reinforcement learning, using only reward feedback. Structured rules are discovered when they are present in the task and helpful for learning and generalization [19–21], and there is evidence that subjects look for and create such structure, even when it is not beneficial [15,16,22]. However, creating structure is potentially effortful and costly: a learner needs not only to discover the appropriate set of rules given a defined set of “higher-level” contexts that cue these rules and “lower-level” stimuli that cue motor actions, but also how to organize the complex multidimensional world into structured components (contexts, stimuli) in the first place. Indeed, more than one hierarchical structure representation of the environment is possible and it is not always evident which features should be contexts cuing task-sets and which should be considered lower level stimuli. Often, one hierarchical structure is more useful than others because it can afford greater generalization to novel contexts (e.g., if multiple contexts cue the same task-set, they can be clustered together [15], Collins & Frank, submitted). For example, the rule to apply for red and green traffic lights is the same whether the intersecting street comes from the left or the right, such that using the position for context allows generalization, and thus simplifies the problem.
Here, we investigated whether motor patterns provide an additional constraint on imposing hierarchical cognitive rule structures. To this effect, we take advantage of motor representational structure in motor cortex. Finger movement representations in M1, while globally somatotopic, are widely overlapping for neighboring fingers [23,24], reflecting the natural statistics of hand movement structure [24]. This representational constraint may lead to pre-activation of neural networks representing frequently co-activated fingers (motor synergies [25,26]), thus facilitating further choices with those fingers. For example, cues that allow for motor preparation of adjacent, rather than segregated, finger presses facilitate action [27–29].
Therefore, we investigated whether this motor constraint could influence the nature of hierarchical rule structures when actions are selected via finger presses on a computer keyboard. Specifically, we hypothesized that subjects might naturally cluster together sets of stimulus-action associations that were more similar in motor space, for example, that involved adjacent fingers, within each task-set. This would then lead to a natural constraint on the creation of abstract rule structure: subjects should treat those feature dimensions that afford motor clustering as lower order stimulus (allowing clustering within rules), and those that afford less motor clustering as higher order context.
To test our hypothesis, we tested subjects in a structure-learning task, in which we previously showed that subjects create one of two possible hierarchical structures, even when they could learn the task as well without imposing structure at all. We show in two new data sets that the potential for motor clustering within a task-set influences the nature of the structure created. We further confirm and replicate this finding in a novel re-analysis of two published data-sets [15,22]. We then use computational modeling to investigate how these biases may emerge. Simulations show that assuming that low level motor representational biases simply influence action selection is not sufficient to account for structure learning biases observed. Rather, we implement a simple computational mechanism based on the motor scaffolding hypothesis that assumes that motor representational constraints affect outcome predictions, and thus inference of structure. We show that this mechanism can account for subject’s biases in structure learning. Finally, we show that this bias may be strong enough to induce subjects to restructure rules, even if they do not need to learn them, but know them through practice and instructions, when such rules are set up such that they conflict with motor clustering constraints.
In a first experiment, 22 subjects performed 6 independent blocks of a reinforcement learning task shown previously to induce creation of hierarchical rule structures [15,22]. Specifically, they used truthful correct/incorrect feedback to learn to select the correct one out of four possible actions (button presses with four fingers of the dominant hand) for each of four distinct visual input patterns (colored shapes, see Figs 1A and 2), presented in pseudo-randomized order.
Although such problems could be learned by standard learning algorithms assuming that each input pattern is a distinct state that simply combines color and shape, we showed in previous research [15,22,30] that subjects impose structure onto such learning problems using hierarchical rules, such that one dimension (e.g., color in Fig 2B) serves as a context that cues a task-set, and the other (e.g., shape in Fig 2B) as a stimulus that directs action selection under the constraint of the current task-set. Imposing such structure facilitates transfer of learned task-sets to novel contexts [15,22,30] and speeds learning when multiple contexts cue the same task-set. Fig 3 shows that two structures are possible for a single learning problem (e.g. with colors as the contexts cueing task-sets, or shapes as the contexts cueing task-sets). Depending on the specific mapping of actions onto four fingers of the right hand, a given structure may elicit more or less efficient motor clustering patterns; two contrasting examples are given in Fig 2, each when color is used as context.
We define a measure based on the motor representation biases introduced above: the adjacency bonus indicated whether a task-set cue clustered motor action (if actions for both stimuli require key presses from adjacent fingers within each task-set; see Fig 2 and Methods).
For a fixed assignment of actions to fingers, contrasting the two potential hierarchical structures (color vs. shape structure) leads to different motor patterns: P1 vs. P2, or P3 vs. P4 (Fig 3). We test here the theory that subjects tend to create the structure associated with greatest affordance for motor clustering, and specifically hypothesized that subjects would select structures P1 over P2 (a = 2 vs. 0), and P3 over P4 (a = 1 vs. 0).
To identify which structure was created, we relied on reaction-time task-switch costs, one of the most reliable findings in cognitive psychology [1]. Due to the additional cognitive demands associated with updating task-sets (thought to relate to processing within prefrontal cortex), response times are slower when the task-set changes from the previous trial, compared to when it repeats, and this slowing is greater than the corresponding cost associated with changes in stimuli within a task-set [1,15]; reaction time switch-cost measures this relative slowing. This measure allows us to differentiate the structures: those subjects adopting color-on-top structures should exhibit greater RT switch costs for switches in color than shape, and vice versa. We thus use the difference between reaction time switch-cost of each dimension as evidence in favor of either structure. We have previously validated this measure in that it was independently predictive of subjects’ ability to transfer the inferred structure to novel contexts that involve the same structure, and was also predictive of neural signals reflecting the same hierarchical structure [15,22].
To evaluate whether the selected structure was affected by potential for motor clustering, we define the motor clustering switch-cost as the difference between the switch costs defined by the structure affording greater motor clustering and that affording less clustering, e.g., switch cost for P1 or P3 minus the switch-cost for P2 or P4. Thus, our prediction that subjects would be more likely to create the hierarchical rule structure that affords motor clustering translated into a prediction that the motor clustering switch-cost should be positive.
Results confirmed our predictions: we found that Motor clustering RT switch cost was significantly greater than 0 (Fig 4 left, p = 0.0007, t(21) = 4.14). This measure did not differ between configuration types P1/2 and P3/4 (t = 0.36, ns). Furthermore, individual learning problems were significantly more likely to be identified as subjects having created the structure affording more motor clustering (P1-3) than the one affording less (P2-4, binomial test p = 0.0012; 66 out of 99 problems; Fig 4 right).
We sought to replicate this result in three further data-sets, using nearly identical experimental designs (two of which were previously published but where motor patterns were randomized and not assessed [15,22]), with only one learning block per subject. Identical analysis performed over all three datasets confirmed the previous results: reaction time-switch cost was significantly higher for the structure with higher adjacency bonus, measured by a positive motor-clustering switch-cost (t(91) = 3.96; p = 0.00015; S4 Fig). There was no effect of experiment on motor-clustering switch-cost (F = 0.34, p = 0.71), nor was there an effect of configuration type (P1/3 vs. P2/4; p = 0.13; S4 Fig). Furthermore, 61 (vs. 31) out of 92 subjects had a positive motor-clustering switch-cost, labeling them as having created the higher adjacency bonus structure (S4 Fig, right). This is significantly more than expected by a balanced random distribution (p = 0.0023, binomial test), supporting our previous finding.
We thus observed across four independent data sets that subjects were more likely to create hierarchical task structure that lead to grouped task sets, highlighting an influence of motor choices on abstract representations known to be crucial for executive functions [1,31]. Note that this bias has no bearing on the subjects’ actual motor choices: independently of how they implement the task, the sequence of valid motor actions to take is the same. Thus, we believe that this phenomenon emerges from strong biases highlighting constraints in how our brain represents rules. We next investigate the mechanisms that could lead to those biases.
We previously proposed a computational model of hierarchical structure learning [15]. The model simulates creation of task-set structure, allowing for clustering of multiple contexts that link to task-sets and which guide lower order stimulus-action-outcome associations. Moreover, it can infer the type of structure, e.g. which features are indicative of contexts that cue latent task sets and which should be considered low order stimuli. It does so by considering which structure is best able to predict the observed outcomes (here, reinforcement feedback) and assumes that structure describes the current environment most efficiently, and allowing for transfer of structures to novel contexts.
Recall that by design, the experimental paradigm used here can be equally well described by either structure, with either color or shape acting as context cueing task-sets, and as such our previous model did not distinguish between them during learning [15]. Here we augmented the model to investigate the potential influence of motor clustering mechanisms (e.g., the overlap in encoding of motor actions by distributed neuronal populations [23,24]) that could explain the structure patterns observed empirically. Specifically, the model assumes that the hierarchical selection of a task-set constrains pre-activation of action representations available within this task-set, allowing the model to account for specific error patterns and hierarchical neural signals that are explained by such processing [15,22]. We now further assume that this activation spreads to adjacent motor actions, given their representational overlap (Fig 5A), and can bias outcome prediction within the selected task-set. Note that this proposed mechanism is local, in that it relies on the overlap of neuronal populations encoding adjacent fingers; but it requires that this overlap is used such that it can influence selection of the higher level task-set: implying a motor constraint on cognitive rule structure.
We simulated two versions of models including this bias: in one case, the bias was used only to influence action selection (comprising a pure motor bias during choice); in the other case the motor overlap was also used to predict outcomes of selected actions (see supplement for details). Simulation results of the former showed that it could not account for any structure preference: inferred weights were in average equal for both structures (wC = wS = 0.5). Indeed, a pure motor bias only affects which keys are pressed and does not lead to inference that one structure is better than the other, and hence predicts no asymmetry in switch costs, unlike the empirical data. In contrast, when we allowed the overlap in motor representations to also influence predictions (Fig 5 –right) this simple local mechanism allowed our model to infer that the best structure was that which afforded greater motor clustering (P1>P2, P3>P4). Indeed, it allowed those task-sets involving adjacent motor actions to better predict outcomes and hence the model identified that structure as better fitting the environment. Overall, these simulations confirm that scaffolding of higher-level abstract rule learning on low level motor representation can lead to biases in abstract structure representation, as observed empirically.
The above findings showed that subjects favor hierarchical rule structures that lead to heuristically simple motor patterns within each rule, in accordance with patterns of motor cortical representations. We next asked whether this representational bias is strong enough that it would influence subjects’ task representation even when learning was unnecessary and task structures are instructed.
To this end, a separate set of subjects performed an instructed task-switching experiment in which they were told which dimension indicates the task-set and which constitutes the stimulus. Subjects were instructed the specific rules corresponding to one of each motor configuration pattern structure (groups for P1, P2, P3 and P4), and practiced those rules in a way that shaped the instructed structure (see Methods, Fig 6A). We were interested whether subjects that were instructed a rule that did not afford motor clustering (groups P2 and P4) would show signs of restructuring those rules to represent them instead within the corresponding motor clustering structure. If subjects followed the instructed representations, we expected that they should exhibit the standard reaction-time switch-cost corresponding to that instructed structure. In contrast, if the representational bias was strong enough, we expected that they would no longer show the classical instructed task-set switch cost because it would be offset (or even reversed) by the motor biases. Results showed that subjects in the P1/P3 group (allowing for motor clustering) had a significant instructed switch-cost (p = 0.002, t(11) = 3.99). In contrast, subjects in the P2/P4 group which did not afford motor clustering, did not show any instructed switch-cost (t(12) = -1.14), and this was significantly different from the other group (p = 0.004, t = -3.19, Fig 6B). Thus the classical task-switch cost was abolished when motor clustering favored the alternative structure. Indeed, over the whole group, instructed switch-cost was not significant (t(24) = 1.05). Instead, the whole group showed a motor clustering switch-cost (t(24) = 3.08, p = 0.005) with no significant difference between the instructed groups (Fig 6C t = -1.36, p = 0.19), lending support to the possibility that rather than performing the task with the instructed representations, subjects might have restructured their representation of the task according to the same motor biases observed in the learning experiment.
When learning to make choices, subjects search for structure and create representations that rely on hierarchical decision trees: at a higher level, they select abstract task-set rules based on contextual cues, which constrain how they select low-level motor actions in response to stimuli. When such a structure matches the statistical regularities of the world, subjects can discover it and leverage that structure to simplify the problem, speeding learning and facilitating transfer [15,19,32]. However, it is not always evident which structure is best. The results we presented here highlight a reversed hierarchical role in rule structure, where low-level features of the motor choices influence the nature of high-level task-set creation. We showed that a representational constraint of motor actions–adjacency overlap–influences the creation of rule structures: this bias manifests itself as a tendency to create rule structures that afford selection of adjacent motor actions for different stimuli within a same task-set. Indeed, we found across four independent data-sets that subjects were more likely to create a structure that afforded rules that primed adjacent motor actions. We showed in the last experiment that this bias was strong enough to offset classic reaction-time switch costs even in instructed task-switching experiments when the opposite structure afforded greater motor clustering.
These findings shed light on internal processes of how abstract rules are created and facilitate choice, even when such structure has no bearing on which motor responses are actually executed. Indeed, no matter what structure is used, the sequence of visual inputs and required motor actions is identical. The only difference elicited by distinct sorts of structure is observable in terms of how switching from one perceived ‘task’ to another elicits a cost, which in turn is related to the ability to transfer this task to novel contexts [15,22].
Our predominant hypothesis, indicating that executive functions are rooted in action selection, and that that implementations of decision making are scaffolded on motor circuits such that motor constraints affect cognitive processing, is inspired from a popular subfield of the embodied cognition literature [5–7,33] (though see [34] for other perspectives on what is required to be considered embodied cognition). Our neural network model of structure learning [15] hypothesized that task-set selection depended on hierarchical cortico-basal ganglia loops, with a prefrontal loop exerting control over a more posterior motor loop, but feedback from motor choices reciprocally reinforcing task-set selection in higher order loops. Here, we augmented an algorithmic version of this model to represent known features of motor cortex. Our model simulation showed that solely assuming facilitated action preparation could not account for preference in created rule-structures. Rather, our model accounts for the findings by assuming that overlap in motor patterns influences outcome predictions, such that structures involving motor clustering were better at predicting the observed outcomes within task-sets and hence inferred to be valid.
Another theory that could potentially account for the observed biases relates to spatial representation–rather than motor selection–biases. Indeed, it is possible that subjects represent the task-sets in a mental one-dimensional space (much as we represent it in the figures with a row of four squares), where a bias for adjacency in motor actions for a task-set would correspond to a spatial grouping bias (both actions on the left, in the middle, or on the right). Previous research in simpler decision-making tasks have shown that both motor and sensory biases could be important [28,29,35], particularly in conjunction, as proposed by Adam’s grouping model [28]. Disentangling both contributions here would require either systematically decorrelating the association between motor action selection and spatial action representation, or using neuroimaging to highlight the role motor cortex and overlap in motor representation play here. Both are beyond the scope of this study, but important targets for future research. However, it is important to note that such an interpretation could correspond to a visuo-motor process, involving for example eye movements, as suggested by research in other domains of high-level cognition (e.g. mental number line [36]), indicating that spatial biases do not rule out an embodied cognition interpretation.
We have focused here on motor adjacency as a low-level, sensory motor factor that influences abstract high level rule creation. Adjacency is well supported by potential mechanisms and neural data, and thus serves as a first intuitive example to establish this influence, inspired by the embodied cognition literature. However, we do not claim that it is the only factor; indeed other low level factors may also influence rule representation. As a preliminary example, we show in supplementary analysis that beyond similarity in motor space (summarized here by adjacency), similarity in the task-set space, as measured by a parallel or symmetric left-right association between stimuli and fingers, may also provide a bias on rule creation and can be captured by the same model (where symmetry influences prediction and inference). Furthermore, beyond such biases, statistics of the task itself can sometimes constrain which structures are more useful than others, e.g. when they facilitate better generalization (Collins & Frank submitted,[14]). But it is tempting to ask why, when the environment does not constrain abstract rule structure learning, low-level sensorimotor biases may do so. It is possible that this is a pure incidental byproduct of the architecture in which prefrontal-cortex learning and choice scaffolds on motor-based cortico-basal ganglia loops. However, it is also possible that this reflects the result of adaptive pressure: one might imagine that sets of rules used together in the same context tend to require more similar actions, and thus that it would be a priori useful to assume this kind of prior when learning novel structure. While this hypothesis resonates with embodied cognition literature, it remains speculative within the frame of this study.
Research in the field of cognitive control tends to think in abstract terms–of stimuli, choices, and values–disembodied from what they are–pictures or sounds, binary key-presses or joystick movements, points or food, assuming that this is processed independently upstream for inputs, and downstream for motor selection. The results presented here show that a very “low-level” feature–exactly with which finger presses choices were made–systematically influences how a very high-level, abstract representation is created, to the point of overwriting trained instructions. This highlights the need to pay careful attention to how sensory-motor factors may bias observed results, and supports the motor scaffolding theories of embodied cognition. It emphasizes how abstract representations that we build for high level-cognition and reasoning likely emerge from the constraints of interacting with connectivity to its input and output regions.
In all tasks, for a given trial, subjects were presented with a single two-dimensional visual pattern on a black screen. There were two possible features on each dimension (e.g. Green and blue Color, triangle and circle Shapes), combining to form four distinct visual input patterns (Figs 1–3). Subjects could make one of four choices, but only one lead to correct feedback, while the others lead to incorrect feedback. The correct choice was different for each input pattern (Fig 2A). In learning experiments, input-pattern order presentation across trials was pseudo-randomized to ensure equal presentation, and equal frequency of first-order transitions. Subjects were instructed to use the four fingers of the right hand to select an action, except in the R-BL experiment (see below), where they used middle and index fingers of left and right hand to press four contiguous keys.
This experiment included twelve independent learning blocks, with non-overlapping sets of stimuli for each block. 6 of those blocks, which are analyzed here, corresponded to a different structure learning problem as defined in “Learning task rules” section, each with different non-overlapping stimuli; they included 80 trials each. 22 subjects performed this experiment, with 40 blocks of configuration P1/2, 59 of configuration P3/4 and 33 blocks of configuration P5/6 (not analyzed here). 18 subjects had at least one P1/2 block, and all subjects had at least one P3/4. For switch-cost difference analysis, results were first averaged across blocks within subjects per configuration, before group analysis. For binary structure assignment analysis, we included all 99 blocks of P1/2 and P3/4 configurations independently.
To replicate findings from this learning task, we analyzed three further data-sets with similar structure, but only a single block iteration of rule learning:
Our model includes two separate “experts”, each of which considers one of the two possible structures (e.g. color as context expert, vs. shape as context expert, see Fig 5A). It then infers the most likely expert that best describes the data as a function of how well each expert predicts observed outcomes over time. Each expert rapidly learns to associate a given context to an abstract latent variable that represents the associated task-set and learns to predict outcomes (here reinforcement feedback), contingent on the current stimulus, chosen action and inferred task-set (rather than context). We first reiterate the model and then describe how we expand it to accommodate motor clustering effects.
Specifically, at each trial t, we label Ct and St the observed color and shape, with the chosen action at, and rt the obtained reward outcome. The color-structure expert learns to associate a color context C to an abstract, latent task-set variable Z, by keeping track of the probability of a given task-set given the color, P(Z|C). For a new context Cnew, the prior probability of a given task set is initialized following a Chinese Restaurant Process with concentration parameter α such that:
Thus novel contexts are assumed to be most likely linked to existing task-sets that have been most popular across variable contexts, and with some possibility of creating a novel task-set.
Following observation of an outcome, the posterior probability P(Z|C) is updated according to Bayes rule, using learned likelihood p(rt|St, at, Zt). At each trial, Zt is inferred using maximum a priori. The inferred Zt constrains both the policy that is used for choice during the current trial, and learning of S-a-r contingencies:
The shape-structure expert is identical to the color-structure expert, with the roles of color and shape reversed.
When both experts are included in the model, they both learn and select actions independently. However, the final action choice proceeds from the mixture of each expert’s policies: π(a) = wC(t)πC(a) + (1 − wC(t))πS(a), where the mixture weight wC(t) is the inferred reliability of each expert. It is initialized at wC(0) = 0.5, and updated via Bayes rule:
wC(t+1)=wC(t)p(rt|St,at,ColorStructure)wC(t)p(rt|St,at,ColorStructure)+(1−wC(t))p(rt|St,at,ShapeStructure)
In words, the expert that best predicts observed outcomes is assumed to be best for representing the current environment.
To model effects of spatial motor patterns on structure learning, we introduce biases in the outcome prediction of each expert. Specifically, we implement an adjacency bias by assuming that task-set selection may more broadly pre-activate action representations relevant to this task-set, independent of the specific current stimulus, including spreading to adjacent motor actions (see Fig 5, left). In brief, given that a task-set has been selected, the action-outcome contingencies learned are generalized to neighboring actions such that they become predictive of corresponding outcomes, even for stimuli that have not yet been encountered. This translates into a bias that modifies expected outcomes following:
bias(a|St,Zt)=∑iπ(ai|others(St),Zt)neighbor(a|ai)
where neighbor(a|ai)=1#neighbors(ai) if a is a neighbor of ai, 0 otherwise; and others(St) indicates the set of stimuli different from St. This bias may be used in a mixture for policy selection and outcome prediction. We show in main text that a model using bias for only policy selection does not account for the empirical results, while a model using it for both does.
Biases are mixed to the normal prediction with mixture weight fi (fi = 0.1) in simulations. Other model parameters are:
We simulate the model 1000 times and report asymptotic preference for either expert.
Although we assumed that motor overlap influences predictions, it is also possible that a similar effect results via imperfect credit assignment during learning. That is, when a given action is reinforced, the same overlap in motor representations could elicit partial reinforcement of adjacent motor representations. As such, the network would not only be more likely to select adjacent actions for the same task-set, but also would be better able to learn to predict accurate outcomes for a clustered task-set, leading to both better performance and greater reliability of an adjacent structure. Indeed, model simulations show that this mechanism produces similar results to the one exposed here. Our current findings cannot separate out these hypotheses, though we suspect that both could occur simultaneously (and indeed both theories rely on motor overlap).
|
10.1371/journal.pbio.2006100 | The memory for time and space differentially engages the proximal and distal parts of the hippocampal subfields CA1 and CA3 | A well-accepted model of episodic memory involves the processing of spatial and non-spatial information by segregated pathways and their association within the hippocampus. However, these pathways project to distinct proximodistal levels of the hippocampus. Moreover, spatial and non-spatial subnetworks segregated along this axis have been recently described using memory tasks with either a spatial or a non-spatial salient dimension. Here, we tested whether the concept of segregated subnetworks and the traditional model are reconcilable by studying whether activity within CA1 and CA3 remains segregated when both dimensions are salient, as is the case for episodes. Simultaneously, we investigated whether temporal or spatial information bound to objects recruits similar subnetworks as items or locations per se, respectively. To do so, we studied the correlations between brain activity and spatial and/or temporal discrimination ratios in proximal and distal CA1 and CA3 by detecting Arc RNA in mice. We report a robust proximodistal segregation in CA1 for temporal information processing and in both CA1 and CA3 for spatial information processing. Our results suggest that the traditional model of episodic memory and the concept of segregated networks are reconcilable, to a large extent and put forward distal CA1 as a possible “home” location for time cells.
| Departing from the most influential model of episodic memory (the two-streams hypothesis), we have recently proposed a new concept of information processing in the hippocampus according to which “what” one remembers and “where” it happens might be processed by distinct subnetworks segregated along the proximodistal axis of the hippocampus, a brain region tied to memory function, instead of being systematically integrated at this level. Here, we focused on the processing of temporal and/or spatial information in the proximal and distal parts of CA1 and CA3 in mice to test whether the two concepts are reconcilable. To do so, we used an imaging method with cellular resolution based on the detection of the RNA of the Immediate Early Gene (IEG) Arc, which is tied to synaptic plasticity and memory demands, and correlated imaging results with memory performance. Our data confirm the existence of subnetworks segregated along the proximodistal axis of CA1 and CA3 that preferentially process spatial and non-spatial information and suggest a key involvement of distal CA1 in temporal information processing. In addition, they show that the two models are complementary to a large extent and posit the “segregated” model as a viable alternative for the two-streams hypothesis.
| In the early 1980s, Mishkin and colleagues proposed a very influential model of episodic memory according to which spatial and non-spatial information emerging from the dorsal and the ventral visual pathways would be integrated into episodes at the level of the hippocampus [1,2]. This model considers information related to the features of objects and their location as non-spatial and spatial information, respectively, while the temporal information bound to these objects is not considered even though this dimension constitutes a key feature for the memory of episodes [3]. Clear empirical evidence for the integration of this spatial and non-spatial information at the level of the hippocampus and possible mechanisms underlying such an integration are still missing. In addition, it is not known whether such an integration would still take place if only one of the dimensions of the memory is salient, i.e., when the integration of both dimensions is not “necessary.” Moreover, the cortical areas constituting the last relay of the “extended” ventral and dorsal pathways, namely the lateral entorhinal cortex (LEC) and the medial entorhinal cortex (MEC), respectively, preferentially project at distinct proximodistal levels of the hippocampal subfield CA1. Indeed, the LEC that essentially processes non-spatial information preferentially projects to the distal part of CA1 (away from the dentate gyrus [DG], i.e., close to the subiculum; Fig 1A). In contrast, the MEC, more sensitive to spatial content, preferentially projects to the proximal part of CA1 (close to the DG and CA2) [4–13]. In addition, the proximal and distal parts of CA3 send segregated projections to CA1. Distal CA1 primarily receives projections from the proximal part of CA3 (close to the DG) and proximal CA1 from distal CA3 (away from the DG, i.e., close to CA2) [14–18]. Furthermore, the distal part of CA3 receives projections from the enclosed blade of the DG, which is tuned to spatial information, as well as from the crest and the exposed blade. Moreover, entorhinal cortex (EC) cells send most of their inputs at this level because EC cells synapse at the level of the lacunosum moleculare, which is quasi nonexistent at the proximal levels of CA3. In comparison, the proximal part of CA3 receives fewer projections from the enclosed blade of the DG and fewer entorhinal inputs, among which LEC inputs which preferentially deal with non-spatial content [19–23].
Altogether, these findings led us to recently suggest the existence of distinct “spatial” and “non-spatial” hippocampal subnetworks segregated along the proximodistal axis of the hippocampus that would preferentially be engaged either when only the spatial dimension or only the non-spatial dimension of a memory is salient, i.e., when the integration of both dimensions is not “necessary” (Fig 1B and 1C) [24,25]. These networks were termed “spatial” and “non-spatial” subnetworks with regards to their relative ability (i.e., not absolute) to process non-spatial and spatial information, i.e., the “non-spatial” subnetwork processes non-spatial information over spatial information, whereas the “spatial” network favors the processing of spatial information over non-spatial information.
Evidence for a functional segregation along the proximodistal axis of CA1 and CA3 is sparse, but its existence is supported by recent electrophysiological and Arc imaging studies. An electrophysiological study from the Moser laboratory reported a stronger engagement of proximal CA1 over distal CA1 for the processing of spatial stimuli [10]. Conversely, we showed a stronger recruitment of distal CA1 over proximal CA1 for the processing of non-spatial (odor-based) information in a previous Arc imaging study [24]. In addition, activity differences along the proximodistal axis of CA1 were reported to be attenuated with aging and proximodistal theta activity coherence to be reduced in the dorsal hippocampus of a rodent animal model of epilepsy [26–27]. Furthermore, some of these reports and others also showed a preferential involvement of proximal CA3 for the retrieval of non-spatial memory and that of distal CA3 for the processing of spatial locations [24–25]. Finally, a proximodistal segregation of CA3 was also reported in terms of pattern completion and pattern separation [28–29].
Still, very little is known about these hippocampal subnetworks. Specifically, it is unclear whether non-spatial information other than objects or odors, such as temporal information, would also preferentially recruit the proximal CA3–distal CA1 “non-spatial” network. This hypothesis is substantiated by the fact that studies focusing on temporal bridging have indeed targeted distal CA1/the distal half of CA1 [30–38]. However, their focus was not the investigation of proximodistal differences in CA1; therefore, whether temporal information also preferentially engages the “non-spatial” subnetwork remains to be thoroughly tested. Also, it is not known whether spatial information related to items such as objects (i.e., object-in-place information) would primarily engage the distal CA3–proximal CA1 “spatial” network, as is the case for locations. Finally, it is not clear whether the concept of segregated information processing in the hippocampus and the traditional model of episodic memory are supported by distinct neural substrates or whether they are variations of one and the same principle supported by the same neural networks that are recruited depending on the nature of the salient dimensions of the memory (Fig 1B–1D).
To address these questions, we investigated which areas—among the distal and proximal parts of CA1 and CA3—are tuned to temporal information, spatial information, or both types of information. To do so, we used a spontaneous object-recognition memory task that allows for the evaluation of distinct discrimination ratios for the retrieval of the temporal aspect of a memory (“when” the objects were presented) and that of its spatial aspect (“where” the objects were located) [39–42], Fig 2A).
Since performing electrophysiological recording simultaneously in 4 brain areas remains a major challenge and since the coordinates of the proximal and distal parts of CA1 and CA3 vary greatly along the transverse axis of the hippocampus because of its folding, we favored a high-resolution imaging approach (i.e., to the cellular level) over lesion/inactivation/optogenetic approaches because the latter approaches would be unlikely to yield the spatial resolution necessary to tease apart the specific function of the proximal and distal parts of CA1 and CA3 in mice. This molecular imaging technique is based on the detection of the RNA of the Immediate Early Gene (IEG) Arc that is commonly used to map brain activity in the medial temporal lobe [45–47] and is tightly linked to plasticity processes [48]. In addition, Arc is more sensitive to memory demands than other IEGs [24,49,50] and allows for each cell activated at test to be detected. Arc RNA is visualized with the help of fluorescent tags, which allow for the percentage of cells engaged at test to be evaluated (Fig 3A–3D).
Procedures were approved by the Ruhr Universität Bochum Institutional Animal Use Committee and the LANUV (8.87–51.04.20.09.323).
Adult male C57BL/6 mice (n = 26)—single-caged and kept under a reversed light/dark cycle—were tested during their active phase.
The apparatus was a 32 × 32 × 41 cm open field, placed in a dimly lit room. Extra-maze cues, as well as cues on the outside wall of the open field, served as spatial references. A video camera (Sony, HDR/CX500E) recorded the animals’ behavior for off-line analysis. Six copies of 2 different metallic objects were used for testing so that objects used during the test phase were duplicates of those during the study phases. Pilot studies showed that animals could distinguish between the 2 objects and had no aversion or preference for either object. The location of the recent, old, stationary, and displaced objects was counterbalanced between animals.
The habituation procedure occurred on 4 consecutive days, following a procedure described in Dere and colleagues [39]. In short, animals were habituated to the empty open field for 20 min during days 1 and 2. To encourage mice to explore all areas of the box (divided in 9 quadrants) and minimize the development of a spatial bias, 1 chocolate sprinkle was placed in the center of each quadrant. The absence or displacement of sprinkles and/or the presence of droppings in each quadrant at the end of each 20-min session were assessed and revealed that mice had explored each quadrant during each session. On days 3 and 4, animals were habituated to the presence of objects, which were not used on the testing day in conditions that mirrored those of the testing day (three 10-min trials, two 50-min delay periods). Animals were tested in groups of 9, plus 1 home-caged control that was placed in the same room but did not perform the task. On day 5, the testing procedure followed that of days 3 and 4, but the test phase was adapted for an optimal detection of Arc pre-mRNA (6-min test phase; Fig 2A). During study phase 1 and 2, animals were exposed to a set of 4 identical objects. During the first study phase, objects formed a triangle. During the second study phase, a different set of 4 identical objects formed a square. At test, duplicates of the objects previously studied were used. Two of the objects that had been the most recently explored (i.e., explored during study phase 2) were placed at the location they occupied then (the “recent stationary” objects). In addition, one of the objects that had been experienced earlier (i.e., during study phase 1) was also placed at the location it occupied then (the “old stationary” object), while another object of the same set was placed in a novel location (the “old displaced” object). After each trial, the open field and stimuli were cleaned with water and a solution containing 10% ethanol.
Based on animals’ natural preference for novelty [51], a successful memory for a given spatial location (e.g., a successful spatial discrimination) is observed when the “displaced old” object is explored more than the “stationary old” object, and a successful memory for the temporal context in which the object was experienced (e.g., a successful temporal discrimination) is reflected by a longer exploration of the “stationary” old object compared with that of the average of the 2 “recent” objects [39–42]. A response pattern according to which mice explore the old displaced object more than the old stationary object and concomitantly the old stationary object more than recent objects suggests that mice have the ability to establish an integrated memory for events comprising information about “what,” “where,” and “when” [39]. The exploration time for an object was defined as the time spent in exploring an object, i.e., directing the nose at a distance <2 cm to the object and/or touching it with the nose, as originally described in Ennaceur and Delacour (1988) [51]. The animals’ exploratory behavior was recorded during each phase for off-line analysis and was used to calculate spatial and temporal discrimination indices at test (spatial and temporal D2s, respectively).
Performance was scored manually by 2 independent experimenters blind to experimental conditions and averaged. The scores of both experimenters were highly correlated (r = 0.879). Specifically, D2 scores were calculated for each animal with the following equations: spatial D2 = (exploration time displaced old object − exploration time stationary old object) ÷ total exploration time for both old objects; temporal D2 = (exploration time stationary old object − average exploration time both recent objects) ÷ (exploration time stationary old object + average exploration time both recent objects).
Following the standard protocol for detection of Arc, animals were euthanized immediately after the test phase, and as Arc pre-mRNA was detected, only Arc intranuclear signal was observable [49,52–54]. In short, brains were removed, flash frozen in isopentane, and stored at −80°C until sectioning. Brains were sectioned with a cryostat (Leica CM 3050 S; 8-μm–thick coronal sections), mounted on Polylysine slides (Thermo Scientific), and stored at −80°C until in situ hybridization. Arc pre-mRNA probes were synthesized using the digoxigenin-labeled UTP kit (Roche Diagnostics). Following a similar fluorescent in situ hybridization (FISH) protocol as Nakamura and colleagues (2013) [24], slides were fixed with 4% buffered paraformaldehyde and rinsed with 0.1 M PBS. Slides were treated with an acetic anhydride/triethanolamine/hydrochloric acid mix, rinsed, and briefly soaked with a prehybridization buffer. The tissue was hybridized with the digoxigenin-labeled Arc probe overnight at +65°C. Following hybridization, slides were rinsed with buffer solutions and treated with an antidigoxigenin-horseradish peroxidase (HRP) conjugate (Roche Molecular Biochemicals) and a cyanin-5 substrate kit (CY5, TSA-Plus system, Perkin Elmer). Nuclei were counterstained with 4’, 6’-diamidino-2-phenylindole (DAPI; Vector Laboratories).
To detect Arc, 1 slide per animal was processed. Slides contained 8 nonconsecutive brain sections (approximately −3.00 mm AP; Fig 2B) [44], and images from 3 nonadjacent sections distant approximately 200 microns (i.e., covering approximately 400 microns) at this AP level were acquired. The number of activated neurons was evaluated on approximately 90 neurons per image on 3 nonadjacent sections (i.e., on approximately 270 neurons per area of interest). Of note, the distal CA3 window is located in the central portion of CA3, and not more ventrally, because the very ventral portion of CA3 belongs to proximal CA3 (close to the DG). Images were captured with a Keyence Fluorescence microscope (BZ-9000E; Japan). Images were taken with a 40× objective (z-stacks of 0.7-μm–thick pictures; see example Fig 3A–3D). Exposure time and light intensity were kept similar for image acquisition. As first described in the seminal work of Guzowski and colleagues [49], contrasts were set to optimize the appearance of intranuclear foci [43, 44, 51, 52]. To account for stereological considerations, neurons were counted on 8-μm–thick sections that contained 1 layer of cells, and only cells containing whole nuclei were included in the analysis [55]. The quantification of Arc expression was performed in the median 60% of the stack in our analysis because this method minimizes the likelihood of taking into consideration partial nuclei and decreases the occurrence of false negative. This method is comparable to an optical dissector technique that reduces sampling errors linked to the inclusion of partial cells into the counts and stereological concerns because variations in cell volumes no longer affect sampling frequencies [56]. Also, as performed in a standard manner in Arc imaging studies, counting was performed on cells (>5 μm) thought to be pyramidal neurons or interneurons because small non-neuronal cells such as astrocytes or inhibitory neurons do not express Arc following behavioral stimulation [57]. The designation “intranuclear-foci–positive neurons” (Arc-positive neurons) was given when the DAPI-labeled nucleus of the presumptive neurons showed 1 or 2 characteristic intense intranuclear areas of fluorescence. DAPI-labeled nuclei that did not contain fluorescent intranuclear foci were counted as “negative” (Arc-negative neurons) [49]. Percentage of Arc-positive neurons was calculated as follows: Arc-positive neurons ÷ (Arc-positive neurons + Arc-negative neurons) × 100. The home-caged group was generated to control for Arc baseline expression, which is known to be low (S1 Fig and S4 Data).
All statistical analyses were implemented in the R statistical package (version 3.4.2). To assess the relationship between Arc expression and the spatial and temporal discrimination indices D2space and D2time, 2 complementary analyses were conducted that led to comparable results: (i) fitting a linear mixed model to the Arc expression using the continuous discrimination indices and (ii) standard correlation analyses for each region separately.
For the first analysis, we estimated a linear mixed model that is conceptually comparable to a linear regression or partial correlations but explicitly models the repeated measurement of Arc expression from the same animals in the 4 brain regions and thus allows a comparison of the estimated effects across brain regions, as well as a comparison of the effects of both discrimination indices directly [58]. We used the “mixed” function in the “afex” package (version 0.18) [59], which in turn uses the “lme4” package (version 1.1) [60] for the estimation; p-values were computed using the Satterthwaite approximation of the degrees of freedom when assessing the significance of the fixed effects as well as using parametric bootstrapping, as implemented by the “mixed” function. The linear mixed model consisted of fixed effects for categorical variables “region” (CA1 and CA3) and “proximodistal” (proximal and distal) and the two discrimination indices (D2time and D2space) and their interactions. We specified a random intercept per animal as a random factor to explicitly model the repeated measurement of Arc expression. To conduct post hoc comparisons, we computed area-specific mean activity and the slopes of temporal and spatial discrimination indices of the fixed effects using the “lsmeans,” “lstrends,” and “cld functions” in the “lsmeans” package (version 2.27) [61]. Whether these area-specific effects were significantly different from 0 was assessed by inspecting the 95% CI—if the interval does not include 0, the effect is considered significantly different from 0.
By modeling the influence of the increase in both discrimination indices concurrently, we also estimated the increase in Arc expression when both the spatial and temporal discriminations were successful. Specifically, the slope for “space+time”—D2space+time—which quantifies this increase, was estimated by calculating the sum of the slopes of spatial and temporal discrimination indices.
For visualization purposes, contour lines were used to represent the relative relationship between the discrimination indices and the Arc activity as predicted by the fitted model. Contour lines are extrapolated beyond experimental data points by visualizing the model prediction at those hypothetical discrimination ratios. Each line represents the set of spatial and temporal discrimination indices for which the mixed model predicts the same level of Arc activity. Note that these lines are parallel because the underlying model assumes a linear combination of the discrimination indices D2time and D2space. A non-linear model including quadratic terms and the interaction of the 2 discrimination indices gave comparable results, therefore only the simpler linear model is reported here.
For the second analysis (ii), we calculated standard correlation coefficients between the Arc expression and the spatial and temporal discrimination indices, respectively. This straightforward analysis minimizes potential problems related to misspecifying the above mixed model that could lead to higher rates of false positives [62] but does not explicitly model the influence of both discrimination indices at the same time (i.e., correlations between Arc expression and the spatial D2 ignore the potential influence of temporal D2, and vice versa).
Patterns of object exploration varied between animals, leading to a substantial spread of the temporal and spatial discrimination indices (Fig 4). Discrimination indices did not correlate with the total objects exploration time during study phases 1 and 2 (66.48 s ± 6.19 s [D2time: r = 0.21; p = 0.34; and D2space: r = 0.056; p = 0.806] and 66.26 s ± 7.49 s [D2time: r = −0.07; p = 0.75; and D2space: r = −0.015; p = 0.95], respectively]) nor at testing (48.86 s ± 4.05 s [D2time: r = 0.17; p = 0.46; and D2space: r = 0.11; p = 0.64]; and see also S1 Table and S1 Data), indicating that differences in total object exploration time per se could not account for the differences in discrimination indices reported in the present study.
To assess the relationship between Arc expression and the spatial and temporal discrimination indices, D2space and D2time, the following 2 complementary analyses were conducted: (1) fitting a linear mixed model to the Arc expression using the continuous discrimination indices and (2) performing standard correlation analyses between Arc expression and the discrimination indices.
A contour plot of predicted Arc expression as a function of the spatial and temporal D2s (Fig 4) showed that, while the recruitment of all 4 areas varied with the spatial D2 (albeit to different degrees), distal CA1’s engagement varied to a larger extent with the temporal D2, as indicated by the contour lines for distal CA1 being more vertical than horizontal. In contrast, the contour lines for proximal CA1 and CA3 and distal CA3 (being more horizontal than vertical) reflected a stronger sensitivity for spatial discrimination.
In addition, statistical comparisons of the slope of D2time showed that Arc expression varied with the temporal D2—especially in distal CA1 (b = 11.32) and to a lesser extent in distal CA3 (b = 6.95)—but failed to do so in other areas (proximodistal by D2time interaction: χ2(1) = 11.13; p = 0.002; all other effects: all p > 0.12) (see also S1A and S1B Table). Moreover, further post hoc comparisons revealed that activity levels increased more as the temporal discrimination became higher in distal CA1 than in proximal CA1 (b = 0.30) or CA3 (b = 3.67) (both p < 0.005) and that distal CA3 activity varied more with D2time than proximal CA3 activity (p = 0.048). Notably, investigation of the CIs of the slopes underlined the robustness of the findings for distal CA1, as the standard 95% CI of the slope of D2time excluded 0 only in distal CA1 (i.e., the slope differed from 0), while a more relaxed 90% CI was necessary to get similar results for distal CA3. In other words, the slope of D2time differed from 0 for distal CA1 but not for distal CA3, suggesting a more robust tuning of distal CA1 than distal CA3 to temporal information. In summary, under standard statistical criteria, the present results suggest that especially distal CA1 is sensitive to the retrieval of temporal information.
In contrast to the D2time slopes, comparisons of the spatial D2 slopes showed that Arc expression increased in all areas as a function of spatial discrimination, although the extent to which this was the case differed by area (D2space effect: χ2(1) = 17.98; p = 0.001; region by D2space interaction: χ2(1) = 5.20; p = 0.030; region by proximodistal by D2space interaction: χ2(1) = 10.16; p = 0.002; no other interaction effect: p > 0.050; see also S1 Table). Indeed, post hoc comparisons showed that Arc expression in distal CA1 (b = 7.56) increased the least with increasing spatial discrimination and that this increase was not significantly different from 0 (i.e., the standard 95% CI for the slope of D2 space included 0, but a less strict 90% CI did not). In addition, the D2space slopes for the areas part of the “spatial” subnetwork—distal CA3 (b = 24.52) and proximal CA1 (b = 16.62—were larger than those part of the “non-spatial” subnetwork—distal CA1 (b = 7.56) and proximal CA3 (b = 13.49) (distal CA3 versus proximal CA3: p = 0.013; distal CA3 versus distal CA1: p = 0.0002; proximal CA1 versus distal CA1: p = 0.037). Thus, these results indicate that proximal CA1 and CA3, and distal CA3, are most tuned to spatial information, whereas distal CA1 is least tuned to spatial information.
Finally, estimating the influence of both discriminations simultaneously (i.e., predicting how Arc activity changes when both dimensions are recalled as captured by the slope of “D2space+time”) revealed that all areas were recruited at a comparable level (main effect of D2space+time: F(1,19) = 15.41; p = 0.0009; no other significant effects, all p > 0.05; see S1C Table).
Altogether, the linear mixed-model approach indicates that Arc activity in the distal CA1 most strongly relates to retrieving temporal information and that proximal CA1 and CA3 and distal CA3 are especially tuned to spatial information.
As observed with the linear mixed-model approach, distal CA1 correlated with the temporal discrimination index (r = 0.475; p = 0.026; Fig 5) but not with the spatial discrimination index (p = 0.65; S2 Fig). In contrast, all other areas correlated with the spatial discrimination index (proximal CA1: r = 0.723; p < 0.0001; proximal CA3: r = 0.675; p = 0.0006; distal CA3: r = 0.720; p = 0.0002; Fig 5) but not with the temporal discrimination index (proximal CA1: p = 0.17; proximal CA3: p = 0.14; distal CA3: p = 0.66; S2 Fig).
In summary, these analyses show, in line with the mixed-modeling results, that distal CA1 is especially tuned to temporal information and proximal CA1 and CA3 as well as distal CA3 are tuned to spatial information.
In the present study, we showed that processing temporal information bound to an object engages distal CA1 over proximal CA1, while processing spatial information about this object recruits proximal CA1 over distal CA1. In a striking contrast, distal CA3 was only slightly activated for the recall of temporal information, and both parts of CA3 were tuned to spatial information processing, albeit distal CA3 to a larger extent than proximal CA3. In addition, retrieving both types of information leads to a strong and comparable recruitment of all areas. Thus, our results suggest that retrieving one or more dimensions of a memory might rely on the same mechanism(s), supporting both the “segregated” and the “integrative” views of information processing in the hippocampus.
One question we addressed in this study is whether the temporal content of an event is preferentially processed by the proximal CA3–distal CA1 “non-spatial” subnetwork as it was the case for other non-spatial information, such as odors [24]. This appears to be the case at the level of distal CA1 because Arc expression increased in this area as the temporal discrimination ratio did (Fig 4). In addition, the slope of temporal D2 was higher in distal CA1 than in proximal CA1 (or any other areas), showing for the first time that the temporal context of an event is topographically organized along the proximodistal axis of CA1. These results also indicate that processing temporal information recruits the same part of the “non-spatial” subnetwork as other non-spatial information, such as odors [24]. Moreover, Arc expression in distal CA1 correlates only with temporal discrimination ratios (and not with spatial ratios; S2 Fig), further supporting the idea of a selective role of distal CA1 in the processing of temporal information. The recruitment of distal CA1 is unlikely to solely reflect the processing of object information as activity patterns are strikingly different between areas despite the fact that the same objects’ information (in time or space) is processed.
In contrast to CA1, CA3 was not engaged to a critical extent in processing temporal information, indicating that the temporal content is differentially computed than object or odor information at this level. These findings provide further support to the lesion, in vivo electrophysiology, and optogenetic studies that have indicated a preferential role of CA1 in temporal information processing over that of CA3 and suggest that distal CA1 is likely to be the home location of the “time cells” recently identified in CA1 [31–38,63,64]. As a support for the latter hypothesis, a thorough review of these studies showed that distal CA1/the distal half of CA1 was indeed targeted in these reports. The hypothesis of a preferential involvement of distal CA1 in the processing of time is also indirectly supported by evidence from a trace eye-blinking conditioning study showing that reversible inactivation of the LEC (which preferentially projects to distal CA1) impairs the retrieval of a memory for an association between temporally discontiguous stimuli [65]. Likewise, in vivo electrophysiology and IEG studies showed that the perirhinal cortex, which provides major inputs to the LEC, plays an important role for temporal-order memory and for object memory across large delays [66–68]. This might indicate that the LEC is the source of temporal information provided to distal CA1. Preliminary data from the Moser laboratory using population-level analyses of electrohysiological recordings partly support this hypothesis by reporting that LEC’s involvement within this frame depends on tasks’ demands, with free foraging tasks eliciting a stronger temporal representation in the LEC than continuous alternation/back-and-forth running tasks [69]; of note, such tasks’ demand dependency in the LEC were also reported in recent lesion and Arc imaging studies, albeit for the processing object and space information [70,71]. Our findings of a preferential involvement of distal CA1 in time processing depart from the standard model of episodic memory which, by extrapolation, predicts that temporal information would rather be processed by proximal CA1 because it mainly receives projections from the MEC, a part of the “where–when” pathway [2,4]. Even though the effect of the reversible inactivation of the MEC on temporal encoding in CA1 might be controversial [65,72], the latter hypothesis is supported by some electrophysiology studies that brought evidence of an involvement of the MEC in the integration of elapsed time and distance and in the temporal organization of CA1 activity [73,74]. Thus, further studies comparing directly LEC and MEC function within this frame will be necessary to clarify the nature and the extent of the contribution of the LEC and the MEC to temporal information processing and, by extension, the role of the distal and proximal parts of CA1 within this frame.
The second question of the present study was to assess whether the processing of spatial information bound to objects was topographically organized along the proximodistal axis of the hippocampus as it was shown for locations [25], i.e., whether it would also preferentially recruit the “spatial” hippocampal subnetwork. Our results show that, in addition to increasing with the temporal discrimination ratio, Arc expression in distal CA1 also increased as a function of the spatial discrimination index, indicating a relative tuning of distal CA1 to spatial information (Fig 4). However, and possibly as a further token of a preferential involvement of distal CA1 in the processing of temporal information, distal CA1 was the least tuned to the spatial discrimination when compared to all other areas. Indeed, the slopes of the spatial discrimination index for distal CA3 and proximal CA1 and CA3 were all larger than that of distal CA1. A key involvement of these regions was further supported by standard correlations and mixed-model analyses showing that Arc expression in proximal CA1 and both parts of CA3 correlated with spatial but not with temporal discrimination indices. This result confirms the central role of CA1 and CA3 in spatial memory as well as the existence of a functional segregation between the CA1 and CA3 subfields [30,75,76]. Moreover, because distal CA1 receives preferential projections from the LEC and proximal CA1 from the MEC, our findings are, by extension, in agreement with in vivo electrophysiology and Fos imaging studies that have shown that the LEC and the MEC are involved in the processing of object-in-place information [12,77–79]. These studies, however, did not directly assess the contribution of the proximal and distal parts of CA1 to the memory for object in place, but see Ito and Schuman [13]. In addition, our results show that processing spatial information bound to objects recruits the same part of the “spatial” subnetwork as processing locations because proximal CA1 was more tuned to spatial discrimination than distal CA1, and they indicate that the processing of object-in-place information is also topographically organized along the proximodistal axis of CA1. This finding, together with recent studies reporting a stronger engagement of proximal CA1 in the case of contextual changes and a weaker recruitment for non-spatial memory, shows that CA1’s functional segregation holds in various experimental settings and that the mechanism sustaining spatial information processing in CA1 could be the same when the information processed is related to a context or to an object (the object’s location) [24,10].
In CA3, a robust proximodistal segregation was also observed in terms of processing spatial information because distal CA3 was more tuned to spatial information than proximal CA3. This engagement of CA3 for spatial information processing is in line with previous lesion and electrophysiology studies—which, however, did not dissociate the contribution of proximal and distal parts of CA3 [80,81]. For example, lesions of CA3 impair object–place or odor–place–paired associations [82], and in vivo electrophysiological studies showed that spatial firing patterns in CA3 distinguish different environments in a foraging task [38]. Conversely, lesions of CA3 did not affect performance on an object–trace–odor task [63]. This stronger involvement of distal CA3 over proximal CA3 in dealing with object-in-place information matches results of a previous finding reporting a similar pattern for the processing of locations with a high-demand memory task [24], indicating that the proximodistal functional segregation in CA3 also holds independently of the type of spatial information processed (i.e., locations or object-in-place).
Thus, altogether, these data show that, at the exception of the temporal information in CA3, the retrieval of spatial (locations) or non-spatial (temporal) information bound to objects engages the same parts of CA1 and CA3 as retrieving information related to objects/odors or locations alone, respectively, indicative of a robust segregation of the spatial and non-spatial information along the proximodistal axis of CA1 and CA3.
Finally, in the present study, we also asked whether the concept of a segregated processing of spatial and non-spatial information in the hippocampus and the standard concept of an integration of this information at this level are “compatible” and based on the same networks. To be “compatible,” we hypothesized that, during memory retrieval, one of the subnetworks (“spatial” or “non-spatial”) would be recruited over the other when only one dimension of the memory is salient (spatial or non-spatial). In contrast, all areas would be activated to comparable levels when both dimensions are salient, i.e., no proximodistal differences would be observable in this case. Here, we report that, at the exception of the temporal information in CA3, proximodistal differences fitting the description of the “spatial” or “non-spatial” subnetworks were detected when animals discriminated on the basis of only spatial or temporal information as captured by the comparisons of the slopes of spatial or temporal discrimination indices, respectively. In addition, all areas were engaged to a similar extent when animals successfully discriminated based on the concurrent retrieval of both dimensions as substantiated by the comparisons of the slopes of the discrimination indices for space + time. Thus, these results indicate that the “segregated” and the “integrated” views of information processing in the hippocampus might be, to a large extent, based on the same principle(s) and networks but might differ in the nature and the number of dimensions of the memory to be retrieved.
The present study focused on assessing the tuning of the proximal and distal parts of CA1 and CA3 to spatial and temporal information. Assessing whether the spatial and the non-spatial dimension of the memory are combined or kept segregated when both dimensions are retrieved, identifying the specific processes underlying the patterns of activity reported, or the basis of interindividual differences in behavioral performance (i.e., whether mice preferentially processed temporal and/or spatial information or failed to do so) are beyond the scope of the study and will require further investigations. Moreover, despite the fact that Arc expression was reported to better reflect behavioral task demands than other IEGs, such as c-fos and zif268, and not simply stress levels or motor activity [24,49,50], the latter processes and others might still partially contribute to the levels of Arc expression observed at test. For this reason, it was crucial to keep experimental conditions (handling, number of stimuli, locomotor activity, etc.) identical across animals. Since under this condition it could be ruled out that between-area differences could not stem from differences in total objects exploration times (comparable for phase 1, 2, and 3) or from neophobia (all objects were experienced prior to the testing phase), between-area comparisons of Arc expression are expected to reflect the processing of spatial and/or temporal information. Furthermore, the proximodistal differences in patterns of activity reported here are unlikely to be a by-product of the anatomical levels at which CA1 and CA3 are imaged because proximodistal differences at these levels were also reported independently of whether the septal level, the temporal level, or the transverse axis of the hippocampus (at which the proximal and the distal parts of CA1 and CA3 are located at different dorsoventral levels) were imaged in a previous study [24].
In summary, these findings complement our recent studies that revealed that spatial information (location) and non-spatial information (odors) can be processed in a segregated manner within the hippocampus [24,25] by showing that temporal and spatial information bound to objects engage, at least part of, the same subnetworks. In addition, we identified the distal part of CA1 as a potential “hub” for time cells and showed that the new concept of segregated processing of spatial and non-spatial information within the hippocampus is, to a large extent, reconcilable with the traditional view of an integration of this information at the level of the hippocampus.
|
10.1371/journal.pgen.1008085 | Mitochondrial fusion is required for regulation of mitochondrial DNA replication | Mitochondrial dynamics is an essential physiological process controlling mitochondrial content mixing and mobility to ensure proper function and localization of mitochondria at intracellular sites of high-energy demand. Intriguingly, for yet unknown reasons, severe impairment of mitochondrial fusion drastically affects mtDNA copy number. To decipher the link between mitochondrial dynamics and mtDNA maintenance, we studied mouse embryonic fibroblasts (MEFs) and mouse cardiomyocytes with disruption of mitochondrial fusion. Super-resolution microscopy revealed that loss of outer mitochondrial membrane (OMM) fusion, but not inner mitochondrial membrane (IMM) fusion, leads to nucleoid clustering. Remarkably, fluorescence in situ hybridization (FISH), bromouridine labeling in MEFs and assessment of mitochondrial transcription in tissue homogenates revealed that abolished OMM fusion does not affect transcription. Furthermore, the profound mtDNA depletion in mouse hearts lacking OMM fusion is not caused by defective integrity or increased mutagenesis of mtDNA, but instead we show that mitochondrial fusion is necessary to maintain the stoichiometry of the protein components of the mtDNA replisome. OMM fusion is necessary for proliferating MEFs to recover from mtDNA depletion and for the marked increase of mtDNA copy number during postnatal heart development. Our findings thus link OMM fusion to replication and distribution of mtDNA.
| Mammalian mitochondria contain multiple copies of the mitochondrial genome (mtDNA), which encodes genes that are essential for the oxidative phosphorylation system. An important feature of mtDNA is that it is evenly distributed throughout the mitochondrial network. Dynamin-related GTPase proteins help control the size and shape of mitochondria by fusion and fission events and are intimately linked to maintenance and distribution of mtDNA. Certain human mutations in mitofusin 2 (MFN2) and optic atrophy protein 1 (OPA1) cause disease phenotypes, such as peripheral neuropathy and optic atrophy, which are often also associated with mtDNA depletion. However, the mechanism whereby MFNs and OPA1 are involved in maintenance of mtDNA is unclear. In this study, we demonstrate that rapid mtDNA synthesis in proliferating tissue-culture cells or cardiomyocytes during post-natal heart development requires mitochondrial fusion. However, the absence of mitochondrial fusion in mouse heart is not associated with mtDNA integrity defects but instead affects the replication of mtDNA. These findings provide direct evidence for the importance of mitochondrial fusion in maintaining mtDNA replication.
| Mammalian mitochondria are dynamic organelles present as long interconnected tubules or individual units that may undergo intracellular transport [1,2]. A family of dynamin-related GTPases regulate mitochondrial morphology through fission and fusion of the mitochondrial membranes [3–6]. The dynamin-related protein 1 (DRP1) mediates division of mitochondria, mitofusin 1 and 2 (MFN1 and MFN2) control outer mitochondrial membrane fusion, whereas optic atrophy 1 (OPA1) control inner mitochondrial membrane fusion. The above mentioned proteins are all essential for embryonic development [2,7,8], and mutated forms are known to cause human disease with phenotypes such as encephalopathy [9], peripheral neuropathy [10,11], and optic atrophy [12,13].
Mitochondrial dynamics is important for distribution and maintenance of mtDNA. Single mtDNA molecules are packaged by the DNA-binding mitochondrial transcription factor A (TFAM) into mitochondrial nucleoids that are evenly distributed throughout the mitochondrial network [14–16]. Previous research has reported that absence of mitochondrial fission results in an elongated mitochondrial network with bulbs harboring aggregates of mitochondrial nucleoids [17,18]. Additionally, inter-organellar contacts between the endoplasmic reticulum (ER) and the OMM have been proposed to facilitate the initial steps of mitochondrial division [19], and to be critical for mtDNA segregation after replication [20]. Thus, proteins acting on the OMM to promote mitochondrial division, possibly in coordination with ER contact sites, help distribute mitochondrial nucleoids. However, little is known about the role of mitochondrial fusion proteins in the distribution of nucleoids, in particular concerning mitofusins, which are also implicated in OMM-ER tethering.
Maintenance of mtDNA is absolutely dependent on mitochondrial fusion in budding yeast, whereas loss of fusion in mammals leads to depletion but not total loss of mtDNA [21–23]. The role for mitochondrial fusion in mtDNA maintenance is nicely illustrated by the double Mfn1 and Mfn2 heart conditional knockout animals (dMfn KO), where a strong reduction in mtDNA copy number impairs oxidative phosphorylation (OXPHOS) and cardiac function [23–25]. In skeletal muscle, the mtDNA depletion caused by absence of OMM fusion has been attributed to instability and increased mutagenesis of mtDNA [23]. Moreover, loss of IMM fusion through inducible OPA1 ablation in adult hearts leads to severe mtDNA depletion [26]. However, loss of mitochondrial fission induced by conditional knockout of DRP1 or reduced mitochondrial fusion due to conditional knockout of either MFN1 or MFN2 did not affect mtDNA copy number in cardiac tissue [23,27–29]. Fusion of the OMM and IMM, which serves to facilitate membrane and/or matrix content mixing, is thus critical for mtDNA copy number maintenance.
In this study, we generated dMfn KO conditional heart knockout mice and also characterized fusion-deficient mouse embryonic fibroblasts to decipher how nucleoid distribution and mtDNA maintenance are linked to mitochondrial fusion. Using sequencing of DNA, we found no increase in mtDNA rearrangements or point mutations in dMfn heart KO animals. Therefore, the low levels of mtDNA observed in the absence of OMM fusion are unlikely to be explained by increased mtDNA mutagenesis leading to instability. Additionally, we reveal that loss of OMM fusion leads to clustering of mitochondrial nucleoids, but this does not impair mitochondrial transcription. Notably, loss of OMM or IMM fusion alters the replisome protein composition, which impairs mtDNA replication activity. Altogether, we found that mitochondrial fusion is necessary to sustain high rates of mtDNA replication, but it is dispensable for maintaining integrity and transcription of mtDNA.
To study the link between mitochondrial fusion and mtDNA maintenance, we generated heart dMfn KO mice, which were born at the expected Mendelian ratios (S1A Fig) and showed loss of transcripts encoding MFN1 and MFN2 in heart tissue (S1B Fig). Most heart dMfn KO animals survived until the postnatal age of 5 weeks and displayed a significant increase in the heart weight to body weight ratio (Fig 1A). Transmission electron microscopy analysis of heart tissue revealed that dMfn KO mice exhibited a significant increase in ratio of mitochondrial/cytoplasmic area associated with aberrant mitochondrial ultrastructure and disruption of the myofibril organization (Fig 1B and 1C). In line with previous fusion-deficient models, qPCR and Southern blot analysis revealed a marked reduction of the mtDNA copy number in heart tissue from dMfn KO animals (Fig 1D and 1E) and in fusion-deficient dMfn KO and Opa1 KO mouse embryonic fibroblasts (Fig 1F). High-resolution respirometry showed a significant decrease of respiration under phosphorylating and uncoupled conditions in dMfn KO heart mitochondria (Fig 1G) and in OMM and IMM fusion-deficient MEFs (S1C Fig), consistent with the observed mtDNA depletion (Fig 1D–1F). This respiratory defect correlated with a reduction in OXPHOS proteins in dMfn KO heart mitochondria (Fig 1H), which rely on mtDNA expression for their stability [30], and caused growth impairment of fusion-deficient MEFs (S1D Fig).
It has been shown that the absence of MFN1 and MFN2 in skeletal muscle severely reduces mtDNA copy number and leads to mtDNA integrity defects [23]. This observation has led to the prevailing theory stating that mitochondrial fusion safeguards mtDNA integrity and prevents mtDNA mutations [23,31]. To examine the integrity of mtDNA in heart tissue in the absence of OMM fusion, we performed paired-end Illumina sequencing to detect unique breakpoints consistent with rearrangements of mtDNA. Similar frequencies of mtDNA rearrangements were found in both negative control samples and dMfn KO hearts (Fig 2A and 2B). The positive control sample, which is derived from muscle tissue of a mouse that harbors a mutation in the DNA helicase Twinkle (A360T), which is known to cause mtDNA rearrangements in patients [32], contained abundant breakpoints consistent with mtDNA rearrangements (Fig 2C). Although this model has been described as a deletor mouse [32], the pattern of rearrangements we observed (Fig 2C) is consistent with abundant mtDNA duplications. Moreover, we found no difference in levels of mtDNA point mutations in heart tissue between control and dMfn KO animals by using post-PCR cloning and sequencing (Fig 2D). These results demonstrate that loss of OMM fusion in the heart reduces mtDNA copy number without influencing mtDNA integrity or increasing point mutations.
To investigate if the distribution of nucleoids throughout the mitochondrial network depends on mitochondrial fusion, we performed immunofluorescence staining against DNA and mitochondria in various fusion-deficient (Mfn1 KO, Mfn2 KO, dMfn KO, and Opa1 KO) MEFs. Control MEFs immunostained with anti-DNA and anti-TOM20 antibodies, and imaged by confocal microscopy, showed abundant evenly distributed mitochondrial nucleoids throughout the mitochondrial network (Fig 3A). In contrast, dMfn KO MEFs displayed reduced abundance of nucleoids that often were present as enlarged foci in a subset of the enlarged-fragmented mitochondria (Fig 3A, white arrowheads; S2A and S2B Fig). The Gaussian distribution of confocal-determined nucleoid diameters revealed a high proportion of enlarged nucleoids in dMfn KO MEFs (Fig 3B), with ~40% of nucleoids having diameters ≥300 nm (Fig 3C). The abnormal size of nucleoids prompted us to further examine nucleoid morphology and diameter with stimulated emission depletion (STED) super-resolution microscopy [15]. In dMfn KO MEFs, the enlarged nucleoids observed by confocal microscopy were resolved into multiple nucleoids in very close proximity by STED microscopy (Fig 3D, white arrowheads). This finding was obtained by using either PicoGreen or anti-DNA antibody staining to determine nucleoid morphology (Fig 3D and S2C Fig). Control MEFs and fusion knockout MEFs of different genotypes all exhibited nucleoid diameters of about 100 nm when analyzed by STED microscopy (Fig 3E and 3F, and S2D and S2E Fig), in agreement with published reports [33,34]. These observations show that loss of OMM fusion in MEFs does not alter nucleoid size but instead causes nucleoids to cluster. We also visualized nucleoids by confocal microscopy using antibodies against TFAM, which is the core protein packaging mtDNA into nucleoids, together with antibodies against DNA. As expected, we observed that TFAM is present in every nucleoid in both control and dMfn KO MEFs (Fig 3G). The morphology of nucleoids was similar when visualized with TFAM or DNA antibodies. Notably, the stoichiometry between TFAM and mtDNA was maintained among the nucleoid population, as clustered nucleoids exhibit more TFAM reactivity than single nucleoids. This is in line with our finding that the enlarged nucleoids are composed of multiple nucleoids, and thus exhibit greater TFAM reactivity. This observation suggests that insufficient compaction by TFAM cannot account for the enlarged nucleoids observed by confocal microscopy.
To rule out the possibility that the observed nucleoid clustering phenotype could be restricted to cultured MEFs, we investigated nucleoid appearance in vivo. The mitochondrial network and nucleoids were immunostained in heart tissue sections from control and conditional dMfn KO animals. In line with the results from MEFs, confocal imaging of dMfn KO heart sections confirmed the presence of apparently large nucleoids that were resolved into multiple nucleoids by STED microscopy (Fig 4A). Furthermore, the different topological conformations of mtDNA in heart tissue were examined and we found no difference in the amount of mtDNA catenanes, which are physically interlocked molecules (Fig 4B). Therefore, changes in mtDNA topology could not explain the clustering of nucleoids in dMfn KOs. Altogether, our analyses reveal that OMM fusion does not influence nucleoid size, but it is required for proper nucleoid distribution in cultured cells and in vivo.
To determine how mitochondrial fusion impacts mtDNA copy number, we first evaluated mitochondrial transcription as the mitochondrial RNA polymerase (POLRMT) provides the RNA primers required for initiation of mtDNA replication [35]. There was a general reduction of steady-state levels of mitochondrial transcripts in hearts of conditional dMfn KO animals, whereas the positive (Mterf4 KO) and negative (Polrmt KO) control samples showed the expected increase and decrease in transcripts, respectively (S3A Fig). Also, levels of the promoter-proximal 7S RNA, previously shown to correlate with transcription initiation [35], were reduced in heart dMfn KO animals (S3B Fig). The reduction of mitochondrial transcript levels including 7S RNA was proportional to the reduction of mtDNA levels (Fig 1D). Furthermore, the levels of the key mitochondrial transcription factors TFAM and mitochondrial transcription factor B2 (TFB2M) were normal, whereas POLRMT levels were increased (Fig 5A). These findings suggest that the reduced abundance of mtDNA templates for transcription explains the reduced levels of mitochondrial transcripts. To investigate this possibility further, we performed in organello transcription assays on isolated heart mitochondria and observed an impaired mitochondrial transcription in dMfn KO heart mitochondria (Fig 5B and 5C). However, there was no difference between control and dMfn KO heart mitochondria when levels of de novo transcripts were normalized to the levels of mtDNA templates (Fig 5D). We proceeded to assess transcription in individual nucleoids by applying FISH to detect the mtDNA-encoded CoxI mRNA in MEFs (Fig 5E). The levels of CoxI mRNA as determined by FISH were reduced in dMfn KO MEFs (Fig 5F), consistent with the northern blot results (S3A Fig). Importantly, the CoxI mRNA was detected in close proximity to almost all nucleoids in both control and dMfn KO MEFs (Fig 5E and 5G) showing that nucleoid clustering does not alter transcription activity. As the FISH analyses of CoxI mRNA does not directly reflect ongoing mtDNA transcription, we also performed bromouridine (BrU) labeling to assess de novo transcription in MEFs. In agreement with the results from FISH analysis of the CoxI mRNA, we found that mitochondrial BrU incorporation was decreased by more than 50% in dMfn and Opa1 KO MEFs (Fig 5H and 5I). Notably, the vast majority of nucleoids in control as well as in OMM or IMM fusion-deficient MEFs incorporated BrU into newly synthesized transcripts (Fig 5H and 5J). These results show that mitochondrial transcription is proportional to the number of nucleoids, i.e. the abundance of mtDNA templates, and that clustering of nucleoids does not impair mtDNA transcription.
The finding of preserved transcription in dMfn KO animals (Fig 5) makes it unlikely that RNA primer formation limits mtDNA replication. Normally, the vast majority of all initiated mtDNA replication events are prematurely terminated after about 650 nucleotides to generate a short single-stranded species denoted 7S DNA [36]. Southern blot (S3C Fig) and qPCR analyses (S3D Fig) consistently showed that the levels of 7S DNA normalized to mtDNA levels were unchanged in dMfn KO hearts. We also included control samples from Mgme1 whole body KO and Polrmt heart KO animals that showed the expected increase and decrease of 7S DNA levels, respectively [35,37]. Our results thus suggest that initiation of mtDNA replication is unaffected in OMM fusion-deficient hearts and the mtDNA depletion must therefore be caused by events downstream of initiation.
The basal mitochondrial replisome is composed of the mitochondrial DNA polymerase γ (POLγ), which consists of the catatytic (POLγA) and the accessory (POLγB) subunits. In addition, the DNA helicase TWINKLE and the single-stranded DNA binding protein 1 (SSPB1) are components of the basal replisome. Western-blot analyses of POLγA, TWINKLE, and SSBP1 levels revealed a striking imbalance between the steady-state levels of replisome components in dMfn KO hearts and MEFs (Fig 6A). Low levels of SSBP1 were consistently observed in both dMfn KO hearts (Fig 6A and 6D) and MEFs (Fig 6C). TWINKLE levels were normal in mutant MEFs but upregulated in dMfn KO hearts. In contrast, the PolγA levels were drastically reduced in IMM and OMM fusion incompetent MEFs and near normal in dMfn KO hearts (Fig 6A–6D). Interestingly, the severity of replisome imbalance observed between fusion incompetent mitochondria in heart tissue and MEFs (Fig 6A–6D) nicely correlate with the extent of mtDNA depletion observed between these two models (Fig 1D–1F). To further define the nucleoid protein composition, we subjected Triton-X-lysed MEF mitochondria to density gradient centrifugation to enrich for nucleoid-associated proteins under native conditions. We found a lower ratio of POLγA per mtDNA in the nucleoid-enriched fraction in both dMfn and Opa1 KO MEFs (Fig 6E–1G). Notably, all the fusion-deficient mitochondria analyzed by density gradients consistently showed an altered stoichiometry of mitochondrial replisome factors associated with the mtDNA template (Fig 6F and 6G). In support of this finding, image analysis showed that SSBP1 co-localized to a greater extent with nucleoids in IMM and OMM incompetent MEFs compared to nucleoids in control MEFs (Fig 6H–6J). Mitochondrial DNA replication not only requires a functional replisome but also relies on stable supply of deoxyribonucleotides (dNTPs). To rule out changes in the cellular dNTP pools as a causative reason for mtDNA depletion in fusion-deficient MEFs, we determined the abundance of dNTPs by UPLC-MS and found no difference between control, dMfn KO, and Opa1 KO MEFs (S4A Fig).
The postnatal development of cardiomyocytes involves substantial mtDNA replication leading to an almost 13-fold increase of mtDNA copy number during the first four weeks of postnatal life (Fig 7A). Interestingly, the dMfn KO hearts cannot sustain this increase and develop severe mtDNA depletion at age 4 weeks (Fig 7A). The low steady-state levels of mtDNA show that loss of OMM fusion in heart tissue prevents the mtDNA increase normally occurring during postnatal heart development. To investigate whether the replisome protein imbalance directly impacts mtDNA replication in cells lacking mitochondrial fusion, we treated MEFs with ethidium bromide (EtBr) and studied mtDNA depletion and repletion dynamics. We observed that the rate of mtDNA depletion was faster in control MEFs than in dMfn or Opa1 KO MEFs (Fig 7B and 7C), suggesting that the rate of mtDNA loss induced by EtBr may depend on mtDNA replication rates. Notably, after six days of EtBr treatment, clustered nucleoids were no longer detected in dMfn KO MEFs (Fig 7D). However, during the repopulation phase, clustered nucleoids reemerged extensively in dMfn KO MEFs (Fig 7D). Moreover, during the mtDNA repopulation phase, MEFs with loss of IMM or OMM fusion showed a severe decrease in the mtDNA synthesis rate (Fig 7B and 7E). To further study the mtDNA replication process, we performed in organello assays with dMfn KO heart mitochondria and followed the incorporation of a radioactively labeled nucleotide into newly synthesized mtDNA. A marked increase in incomplete mtDNA replication products, smaller than full-length mtDNA, was visible as smears surrounding the 7S DNA in dMfn KO heart mitochondria (Fig 7F and 7G). We have previously observed that the formation of similar incomplete mtDNA replication products are linked to mtDNA replication stalling and mtDNA depletion in Mgme1 KO animals [37]. It is thus plausible that dMfn KOs exhibit processivity defects during mtDNA replication in the absence of mitochondrial fusion. Altogether, our analyses reveal that mitochondrial fusion is necessary to maintain high replicative activity of mtDNA.
The mechanism whereby loss of OMM fusion leads to mtDNA depletion is unclear and a popular theory has been that fusion stabilizes mtDNA by preventing mutagenesis and instability. However, this hypothesis, based on the observation that skeletal muscle dMfn KO mice accumulate point mutations and deletions of mtDNA [23], is unsatisfactory because the absolute levels of point mutations and deletions were extremely low (2-3x10-6 mutations/bp and 1x10-6 deletions/mtDNA). If one assumes ~103 mtDNA copies per nucleus in skeletal muscle cells, then only 1:103 of the nuclear domains in skeletal muscle will contain a deletion or a point mutation of mtDNA. Contrary to this previous model, we show here that heart dMfn KO mice have no increase of point mutations or deletions of mtDNA, which excludes this mechanism as a cause for mtDNA depletion. Importantly, our results are based on loss-of-function mutations and it should be pointed out that missense mutations in MFN2 or OPA1 in humans have been reported to cause mtDNA deletions, whereas, in support of our findings, frameshift or premature stop mutations that lead to loss of the fusion protein, do not seem to cause mtDNA deletions in humans [38,39].
Characterization of heart dMfn KO animals revealed that loss of OMM fusion causes a severe imbalance of replisome factors with mtDNA depletion as a direct consequence. However, in contrast to MEFs, TWINKLE protein levels in dMfn KO hearts are strongly upregulated, consistent with the findings in other heart conditional KO mice with mtDNA replication deficiencies. The increase of TWINKLE in KO hearts may represent a compensatory response to counteract mtDNA depletion [35]. The lack of a similar response in KO MEFs could explain why the replisome factor imbalance and mtDNA depletion are more severe in fibroblasts than in heart tissue. Altogether, our results show that both OMM and IMM fusion ensure rapid rates of mtDNA replication. This may be achieved by equilibrating the stoichiometry of replisome factors in order to allow the formation of functional replisomes throughout the mitochondrial network. This is well in line with previous in vitro studies that have shown that a balanced replisome composition is essential for efficient mtDNA replication [40–43]. Consistent with these data, mtDNA depletion phenotypes have also been reported in vivo when replisome proteins have been overexpressed or knocked-out in animals [44–49]. The results we present here thus suggest that mitochondrial content mixing induced by mitochondrial dynamics is necessary to maintain the delicate protein composition balance of the mitochondrial replisome.
Our study also shows that OMM fusion is an important player that determines segregation of mitochondrial nucleoids. The lack of nucleoid distribution was not strictly related to a general loss of fusion, because Opa1 KO mitochondria did not display nucleoid clustering. Instead, this phenomenon appears to be specific to cells with a combined loss of both MFN1 and MFN2. It has previously been reported that distribution of mitochondrial nucleoids depends on fission induced by DRP1 and that mtDNA replication preferentially may occur at ER-OMM contact sites [17,20,50]. It is thus possible that the OMM serves as an important platform for mtDNA distribution and that mitofusins are involved in this process. In line with this hypothesis, clustered nucleoids were no longer visible in dMfn KO MEFs during EtBr treatment but reappeared during the mtDNA repopulation phase. This suggests that nucleoid clusters are not caused by altered mtDNA stability, but rather result from impaired mtDNA distribution when OMM fusion is absent. Future studies are warranted to further understand the molecular link between nucleoid distribution and OMM dynamics. In summary, we report here an unexpected link between mitochondrial fusion and mammalian mtDNA replication, whereby mitochondrial fusion coordinates high rates of mtDNA replication and promotes mitochondrial nucleoid distribution.
Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Nils-Göran Larsson ([email protected])
All animal procedures were conducted in accordance with European, national and institutional guidelines and protocols were approved by the Landesamt für Natur, Umwelt und Verbraucherschutz, Nordrhein-Westfalen, Germany. Animal work also followed the guidelines of the Federation of European Laboratory Animal Science Associations (FELASA).
C57BL/6N mice with loxP-flanked Mfn1 and Mfn2 genes were previously described in Lee et al., 2012 [51]. To generate Mfn1 and 2 heart conditional double knockout mice, male mice homozygous for mitofusin 1, heterozygous for mitofusin 2, and heterozygous for expression of cre-recombinase in heart and skeletal muscle (Mfn1loxP/loxP, Mfn2+/loxP; Ckmm-cre-/+), were crossed with female mice homozygous for loxP-flanked mitofusin 1 and 2 genes (Mfn1loxP/loxP, Mfn2loxP/loxP). Littermates lacking the transgenic Cre-recombinase allele were used as controls. In some crosses, an allele allowing for the ubiquitous expression of a mitochondria-targeted YFP from the ROSA26 locus was also introduced.
Immortalized MEFs with homozygous knockout for Mfn1, Mfn2 or both Mfn1 and Mfn2, were originally generated in the lab of Dr. David Chan [2,52]. Mitofusin and Opa1 KO MEFs were obtained from Dr. Thomas Langer. MEFs were maintained in DMEM GlutaMax containing 25 mM glucose (Thermo Fisher Scientific, 31966–021) supplemented with 10% fetal bovine serum (Thermo Fisher Scientific, 10270–106), 1% penicillin/streptomycin (Thermo Fisher Scientific, 15070–063), 1% non-essential amino acids (Thermo Fisher Scientific, 11140–050), and 50 μg/ml uridine. Cells were passaged every 3 to 4 days.
Electron micrographs of heart mitochondria from control and dMfn heart double knockout mice were obtained as previously described [53]. Small pieces from the left myocardium were fixed in a mix of 2% glutaraldehyde and 1% paraformaldehyde in 0.1 M phosphate buffer pH 7.4, at room temperature for 30 min, followed by 24 hours at 4°C. Specimens were rinsed in 0.1 M phosphate buffer and post-fixed in 2% OsO4 for 2 hours, dehydrated and embedded in LX-112 resin. Ultra-thin sections (approximately 50–60 nm) were cut and sections were examined in a transmission electron microscope (Tecnai 12, FEI Company, Netherlands) at 80 kV. Digital images at a final magnification of 8,200x were randomly taken on myofibrils from sections of the myocardium. The volume density (Vv) of mitochondria was calculated on printed digital images by point counting using a 2 cm square lattice [54]. To determine the number of images needed for an appropriate sampling, we used a cumulative mean plot [54]. In total, 15 randomly taken images were used from each animal.
Mice were sacrificed by cervical dislocation and hearts were quickly washed in ice-cold PBS. Hearts were then minced and gently homogenized using a Potter S homogenizer (Sartorius) in 5 ml of mitochondria isolation buffer (MIB, 310 mM sucrose, 20 mM Tris-HCl, and 1 mM EGTA). Differential centrifugation of homogenates was performed to isolate intact mitochondria. First, homogenates were centrifuged for 10 min at 1,000xg at 4°C. Then, supernatants were collected and centrifuged again for 10 min at 4,500xg at 4°C. Crude mitochondria were resuspended in MIB and the protein concentration was determined by using the DC Protein assay (Bio-Rad).
The flux of mitochondrial oxygen consumption in isolated heart mitochondria was measured as previously described [55]. The assay was performed on 100–125 μg of crude heart mitochondria diluted in 2 ml of mitochondria respiration buffer (120 mM sucrose, 50 mM KCl, 20 mM Tris-HCl, 4 mM KH2PO4, 2 mM MgCl2, and 1 mM EGTA, pH 7.2) in an Oxygraph-2K (Oroboros Instruments) at 37°C. The mitochondria oxygen consumption rate was evaluated using either 10 mM pyruvate, 5 mM glutamate, and 5 mM malate. The oxygen consumption flux was assessed in the phosphorylating state with 1 mM ADP or in the non-phosphorylating state by addition of 2.5 μg/ml oligomycin. Lastly, mitochondrial respiration was uncoupled by successive addition of CCCP up to 3 μM to reach maximal oxygen consumption. The respiratory control ratio values were >10 with pyruvate/glutamate/malate and >5 with succinate/rotenone based on control heart mitochondria.
Heart tissue and whole cells were lyzed in RIPA buffer (150 mM sodium chloride, 1.0% NP-40 or Triton X-100, 0.5% sodium deoxycholate, 0.1% SDS (sodium dodecyl sulfate), 50 mM Tris, pH 8.0). Isolated heart mitochondria were resuspended in NuPAGE LDS Sample Buffer (novex, NP0007) with 0,1M dithiothreitol, and proteins were separated by SDS-PAGE using 4–12% precast gels (Invitrogen), followed by blotting onto polyvinylidene difluoride (PVDF) membranes or in particular cases onto nitrocellulose membranes (GE Healthcare). Membranes were blocked in 3% milk/TBS. The following primary antibodies were used: mouse anti-OXPHOS cocktail (1:1000, ab110413), mouse ATP5A (1:1000, ab14748), rabbit anti-POLγA (1:250, ab12899), rabbit anti-TFAM (1:1000, ab131607), mouse anti-VDAC (1:1000, ab14734), all from Abcam, and rabbit anti-SSBP1 (1:500, HPA002866) from Sigma. Rabbit polyclonal antisera were generated by Agrisera and recognize the mouse TWINKLE, TFB2M, LRPPRC, and POLRMT proteins [48,56,57]. All TWINKLE, TFB2M, and POLRMT rabbit polyclonal antisera were affinity purified using the corresponding recombinant protein. The following secondary antibodies were applied at a 1:10,000 dilution: donkey anti-rabbit IgG (NA9340V) and sheep anti-mouse (NxA931) both from GE Healthcare. Detection of HRP-conjugated secondary antibodies was achieved by enhanced chemiluminescence (Immun-Star HRP Luminol/Enhancer from Bio-Rad).
Mitochondrial isolation from cells and glycerol density gradients were performed as described [58]. In brief, crude mitochondria were isolated from cells by differential centrifugation, treated with DNase and RNase and purified in sucrose gradients. A 15–40% glycerol gradient containing a 20% glycerol / 30% iodixanol pad at the bottom was cast using a gradient master (model 107, Biocomp) with the following settings: 2 min and 31 sec run time, 81.65 angle, and speed set to 14. Mitochondria (250–600 μg) were lyzed in 2% Triton X-100 for 15 min on ice and afterwards the total lysate was carefully layered on top of the gradient. Samples were centrifuged at 151,000 x g at 4°C for 4 hours. The gradient was manually fractionated in 750 μl increments from top to bottom. Fractions were then divided in half, whereby one portion was used to precipitate proteins overnight with 12% trichloroacetic acid and 0.02% sodium deoxycholate, followed by centrifugation at 20,000 x g for 20 min at 4°C and acetone washes, while the other half of fraction 1 was used to isolated mtDNA by proteinase K treatment followed by phenol/chloroform extraction. Thereafter, protein samples were processed for western blot analysis and mtDNA samples for Southern blot analysis.
Cells were grown on 100 mm dishes in DMEM GlutaMax containing 25 mM glucose and supplemented with 10% FBS, 1% penicillin and streptomycin, 1% non-essential amino acids, and 50 μg/ml uridine. Positive control cells were treated with hydroxyurea (Sigma, H8627) at 2 mM for 30 hours to mildly deplete purines. Cells were quickly washed with ice-cold PBS, detached by trypsinization with 0.05% trypsin and the total number of cells was determined using a Vi-cell XR cell analyzer (Beckman Coulter). Approximately 2 million cells were centrifuged at 800xg for 5 min, washed with ice-cold PBS, and resuspended in 1 ml of ice-cold 60% methanol. Samples were then vortexed rigorously, frozen for 30 min at -20°C, and sonicated for 15 min on ice. After sonication, samples were centrifuged at 1000xg for 5 min at 4°C and supernatants collected. Extraction solution was evaporated using an Eppendorf Concentrator plus at room temperature.
Dried pellets were dissolved in 100 μl Milli-Q water, vortexed and sonicated for 2 min. After sonication samples were vortexed again and filtrated through a 0.2 μm modified nylon centrifugal filter (VWR) with an Eppendorf centrifuge 5424R set to 8°C and 12000 rpm. The external standard calibration curve was prepared in concentrations from 5 to 600 ng/ml of the dNTPs. All solutions were daily fresh prepared from stock solutions of 100 μg/ml and dissolved in Milli-Q water. dNTP analysis was conducted using a Dionex ICS-5000 (Thermo Fisher) Anion exchange chromatography using a Dionex Ionpac AS11-HC column (2 mm x 250 mm, 4 μm particle size, Thermo Fisher) at 30°C. A guard column, Dionex Ionpac AG11-HC b (2 mm x 50 mm, 4 μm particle size, Thermo Fisher), was placed before the separation column. The eluent (KOH) was generated in-situ by a KOH cartridge and deionized water. At a flow rate of 0.380 mL/min a gradient was used for the separation: 0–3 min 10 mM KOH, 3–12 min 10–50 mM, 12–19 min 50–100 mM, 19–21 min 100 mM 21–25 min 10 mM. A Dionex suppressor AERS 500, 2 mm was used for the exchange of the KOH and operated with 95 mA at 15°C. The suppressor pump flow was set at 0.6 mL/min. 10 μL of sample in a full loop mode (overfill factor 3) was injected. Autosampler was set to 6°C. The Dionex ICS-5000 was connected to a XevoTM TQ (Waters) and operated in negative ESI MRM (multi reaction monitoring) mode. The source temperature was set to 150°C, desolvation temperature was 350°C and desolvation gas was set to 50 L/h, cone gas to 50 L/h. The following MRM transitions were used for quantification: dGTP, transition 505.98 → 408.00, cone 30, collision 20; dATP, transition 490.02 → 158.89, cone 30, collision 26; dTTP, transition 480.83 → 158.88, cone 28, collision 46; dCTP, transition 465.98 → 158.81, cone 28, collision 30. The software MassLynx and TargetLynx (Waters) were used for data management and data evaluation & quantification. The calibration curve presented a correlation coefficient: r2 > 0.990 for all the compounds (response type: area; curve type linear). Quality control standards were tested during the sample analysis. The deviation along the run was between 0.5% and 40% respectively. Blanks after the standards, quality control and samples did not present significant peaks.
RNA was isolated from heart tissue either by using the ToTALLY RNA isolation kit (Ambion) or by using TRIzol Reagent (Invitrogen), and subjected to DNase treatment (TURBO DNA-free, Ambion). 1–2 μg of total RNA was denatured in RNA Sample Loading buffer (Sigma), electrophoresed in 1 or 1.8% formaldehyde-MOPS agarose gels prior transfer onto Hybond-N+ membranes (GE Healthcare). After UV crosslinking the membranes were successively hybridized with various probes at 65°C in RapidHyb buffer (Amersham) and then washed in 2x SSC/0.1% SDS and 0.2x SSC/0.1% SDS before exposure. Mitochondrial probes used for visualization of mRNA and rRNA levels were restriction fragments labeled with α-32P-dCTP and a random priming kit (Agilent). Different tRNAs and 7S rRNA were detected using specific oligonucleotides labeled with γ-32P-dATP. Radioactive signals were detected by autoradiography.
Total RNA from frozen heart tissue was isolated using RNeasy Mini Kit (Qiagen) following the manufacturer’s instructions, DNase treated using TURBO DNA-free Kit (Thermo Fisher Scientific), and RNA subjected to reverse transcription PCR (RT-PCR) for cDNA synthesis using High Capacity cDNA reverse transcription kit (Applied Biosystems). Real-time quantitative reverse transcription PCR (qRT-PCR) was performed using the following Taqman probes from Thermo Fisher Scientific: Mfn1 (Mm01289372_m1), Mfn2 (Mm00500120_m1), and β-2-microglobulin (B2M, Mm00437762_m1). The quantity of transcripts was normalized to B2M as a reference gene transcript.
Total DNA was extracted from heart and the somatic mtDNA mutation load was determined by post-PCR, cloning, and sequencing as described previously [59]. We used primers that amplify a region of the mtDNA spanning the 3’ end of mtND2 through approximately a third of mtCO1 (nucleotide pair 4950–5923 of the mouse reference mtDNA sequence GenBank NC_005089). The resulting clones were filtered to remove sequences derived from the known nuclear mitochondrial sequences (NUMTs).
DNA from isolated heart mitochondria was used to generate libraries for sequencing to detect mtDNA deletions. Total genomic DNA from the Deletor mouse was provided by Dr. Anu Suomalainen-Wartiovaara and used as a positive control for the detection of mtDNA rearrangements [32]. A standard Illumina TrueSeq paired-end library was prepared with ~500 base pair fragment inserts. Paired-end 100 base pair sequencing was conducted using an Illumina HiSeq 2500. The reads were mapped to the genomic sequence without the mitochondria using bowtie (VN: 2.1.0) to remove nuclear-genomic sequences [60]. The unmapped reads were then mapped to the mitochondrial sequence (GenBank JF286601.1) using bwa (n = 0.04, VN: 0.6.2-r126) [61] with unmapped reads undergoing an additional mapping round after trimming fastx_trimmer, VN: 0.0.13.2) to ensure higher mapping results. Using samtools [62] (VN: 1.0: samtools -f1 –F14) reads, where the two paired sequences were observed to be greater than 600 base pair apart were identified as those containing deletion breakpoints. These reads were extracted for additional analysis. The data presented in this publication have been deposited in NCBI's Gene Expression Omnibus [63] and are accessible through GEO Series accession number GSE124420 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE124420).
Total DNA was extracted from 20 mg of mouse heart tissue. Briefly, the minced tissue was lyzed in 600 μl of lysis buffer (100 mM Tris-HCl pH 7.5, 100 mM EDTA, 100 mM NaCl, 0.5% SDS, 0.5 mg/ml Proteinase K) at 55°C for 3 hours followed by 2 hours incubation on ice with premixed LiCl and K-acetate (final concentration 270 mM K-acetate, 820 mM LiCl) to precipitate contaminants. To remove the precipitate, the sample was centrifuged for 10 min at 10,000 rpm at room temperature and thereafter the DNA was precipitated with isopropanol overnight. The pelleted DNA was washed with 75% ethanol and resuspended to 10 mM Tris-HCl, 1 mM EDTA pH 8.0 followed by quantification with Qubit. To analyze the major topological isomers of the mtDNA, 200 ng of total DNA was resolved in a 0.4% agarose gel (15 x 15 cm) without ethidium bromide at 35 V for 20 hrs, followed by transfer onto Hybond-N+ membrane (GE Healthcare). The mtDNA was detected using [α-32P]-dCTP labeled probe (pAM1). To identify different topological isomers of the mtDNA, 200 ng of control total DNA was incubated at 37°C for 30 min with only buffer, SacI (New England Biolabs; 20 U), Nt.BbvCI (New England Biolabs; 10 U), Topo I (New England Biolabs; 5 U), Topo II (USB Affymetrix; 20 U), DNA gyrase (New England Biolabs; 5 U). The method was modified from [64].
Isolated heart mitochondria were purified by differential centrifugation in 320 mM sucrose, 10 mM Tris-HCl and 2 mM EDTA. From the same mitochondrial preparation: i) 300 μg of mitochondria were labeled with 50 μCi of [α-32P]-UTP (Perkin-Elmer) and processed as previously described to assess de novo mitochondrial transcription [57]; ii) as a loading control, 30 μg of mitochondria were subjected to SDS-PAGE and immunoblotted using mouse anti-VDAC antibody (Calbiochem); and iii) 300 μg of mitochondria were used to measure the ratio between 7S DNA and mitochondrial DNA. In the latter case, the mitochondrial pellet was resuspended in 50 mM Tris-HCl pH 7.5, 75 mM NaCl, 6.25 mM EDTA, 1% SDS and 1,2 mg/ml of Proteinase K and incubated for 1 hr at 37°C. After boiling for 5 min at 95°C, samples were electrophoresed in 0.8% agarose gels and transferred onto nylon membranes by Southern blotting. Both, mitochondrial DNA and 7S DNA were detected using [α-32P]-dCTP-labeled DNA probes specific for 7S DNA.
Isolated heart mitochondria were purified by differential centrifugation and 15 μg of mitochondria were used to assess loading. For de novo mtDNA replicaton, 300 μg of mitochondria were washed with 500 μl ice-cold incubation buffer (25 mM sucrose, 75 mM sorbitol, 10 mM K2HPO4, 100 mM KCl, 0.05 mM EDTA, 5 mM MgCl2, 10 mM Tris-HCl pH 7.4, 10 mM glutamate, 2.5 mM malate, 1 mg/ml BSA, and fresh 1 mM ADP, final pH 7.2), centrifuged at 9,000 rpm for 4 min at 4°C, resuspended in incubation buffer containing 50 μM of dCTP, dGTP, dTTP and 20 μCi of radioactive α-32P-dATP, and incubated for 2 hours at 37°C. Thereafter, mitochondria were centrifuged at 9,000 rpm for 4 min at 4°C and washed 3x with 500 μl ice-cold washing buffer (10% glycerol, 0.15 mM MgCl2, 10 mM Tris-HCl pH 6.8). Mitochondria were treated with 1,2 mg/ml proteinase K and DNA extracted with phenol/chloroform. The mtDNA was resuspended in 30 μl TE buffer and one half boiled at 95°C for 5 min. Next, DNA was separated by electrophoresis on a 0.8% agarose gel for 15 hours at 20 V and transferred onto nylon membranes (Hybond-N+, GE Healthcare), followed by membrane cross-linking. Both Phosphor Imager (Fuji Film FLA-7000) and autoradiography films (GE Healthcare) were used to detect the radioactive signal. After the radioactive signal had decayed, steady-state mitochondrial DNA and 7S DNA levels were assessed by using an α-32P-dATP-labeled DNA probe for 7S DNA.
The growth rate of log phase cells was assessed by equilibrating cells to galactose media (DMEM supplemented with 15 mM galactose, 1mM sodium pyruvate (Thermo Fisher Scientific, 11360–039), 10% FBS, 1% Penicillin/Streptomycin, 1% non-essential amino acids, and 50 μg/ml uridine) for 3 days, followed by plating 27,000 cells in a six-well plate containing 3 ml of galactose media. Cells were collected by trypsinization with 1 ml of 0.05% Trypsin-EDTA (Thermo Fisher Scientific, 25300–054) every 24 hours and viable cells counted using a Vi-Cell XR analyzer (Beckman Coulter).
MEFs in log phase were grown in DMEM GlutaMax containing 25 mM glucose and supplemented with 1% dialyzed fetal bovine serum, 1% penicillin/streptomycin, 1% non-essential amino acids, and 50 μg/ml uridine for 5 to 6 days. MEFs were passaged once during this timeframe. Cells were then collected by trypsinization and counted. The flux of mitochondrial oxygen consumption was determined using 1 million viable cells as described for isolated heart mitochondria, see Mitochondrial respiration, expect for the following modifications. Cells were permeabilized with 0.02 mg/ml digitonin and the oxygen consumption rate in state 3 was assessed using 10 mM succinate and 5 mM glycerol-3-phosphate.
MEFs were cultivated in 18 ml of DMEM GlutaMax containing 25 mM glucose and supplemented with 10% FBS, 1% penicillin and streptomycin, 1% non-essential amino acids, and fresh 100 ng/ml ethidium bromide. After 6 days of cultivation in the presence of ethidium bromide, cells were switched to medium containing no ethidium bromide for 6 days. Cells were passaged every 2–3 days during both conditions. Total DNA was extracted using the DNeasy Tissue and Blood Kit (QIAGEN) and mtDNA levels determined, see mtDNA quantification by qPCR.
For experiments requiring dual visualization of mRNA and proteins, FISH was performed prior to immunocytochemistry, by using the QuantiGene ViewRNA ISH Cell Assay (Affymetrix) following the manufacturer’s instructions. To visualize mitochondrial RNA, cells were hybridized 3 hours at 40°C using the Cox1mRNA probe (1:500, cat# VB4-17017, Affymetrix). To confirm probe specificity towards RNA, treatments with RNAse T1 and DNAse1 (both at 250U/mL, 37°C) were performed prior to hybridization.
Cells grown on coverslips for 24 hours in DMEM medium, without uridine supplementation were incubated with 5 mM BrU for 1 hour at 37°C and 5% CO2. A short chase using DMEM medium containing 50 μg/ml uridine was performed followed by a PBS wash prior to fixation. BrU was prepared fresh for each experiment. Standard immunocytochemistry procedures were carried out to visualize the mitochondrial network, mtDNA, and newly synthesized BrU-labeled RNA.
Imaging of heart sections was performed using a Leica TCS SP8 gated STED (gSTED) microscope, equipped with a white light laser and a 93x objective lens (HC PL APO CS2 93x GLYC, NA 1.30). For confocal images of mitochondria and DNA, Z-stacks in accordance with the Nyquist sampling criteria were taken by exciting the fluorophores at 488 nm and 594 nm, respectively, and Hybrid detectors collected fluorescent signals. Stimulated emission depletion of DNA channel was performed with a 775 nm depletion laser. 2D confocal and gSTED images were acquired sequentially with the optical zoom set to obtain a voxel size of 17 x 17 nm. Excitation was provided at 594 nm and Hybrid detectors collected signal. Gating between 0.3–6 ns was applied. Images were deconvolved with the Huygens software. Performance of the microscope and optimal depletion laser power were tested as previously described [65].
Data are presented as mean ± SEM unless otherwise indicated in figure legends. Sample number (n) indicates the number of independent biological samples (individual mice, number of cells, or wells of cells) in each experiment. Sample numbers and experimental repeats are indicated in the figures. Data were analyzed in Graphpad Prism using the unpaired Student’s t-test, one-way ANOVA using Turkey’s multiple comparison test, two-way ANOVA using Bonferroni multiple comparison test between group comparison, as appropriate. A p-value ≤ 0.05 was considered statistically significant.
|
10.1371/journal.pgen.1004396 | Early Embryogenesis-Specific Expression of the Rice Transposon Ping Enhances Amplification of the MITE mPing | Miniature inverted-repeat transposable elements (MITEs) are numerically predominant transposable elements in the rice genome, and their activities have influenced the evolution of genes. Very little is known about how MITEs can rapidly amplify to thousands in the genome. The rice MITE mPing is quiescent in most cultivars under natural growth conditions, although it is activated by various stresses, such as tissue culture, gamma-ray irradiation, and high hydrostatic pressure. Exceptionally in the temperate japonica rice strain EG4 (cultivar Gimbozu), mPing has reached over 1000 copies in the genome, and is amplifying owing to its active transposition even under natural growth conditions. Being the only active MITE, mPing in EG4 is an appropriate material to study how MITEs amplify in the genome. Here, we provide important findings regarding the transposition and amplification of mPing in EG4. Transposon display of mPing using various tissues of a single EG4 plant revealed that most de novo mPing insertions arise in embryogenesis during the period from 3 to 5 days after pollination (DAP), and a large majority of these insertions are transmissible to the next generation. Locus-specific PCR showed that mPing excisions and insertions arose at the same time (3 to 5 DAP). Moreover, expression analysis and in situ hybridization analysis revealed that Ping, an autonomous partner for mPing, was markedly up-regulated in the 3 DAP embryo of EG4, whereas such up-regulation of Ping was not observed in the mPing-inactive cultivar Nipponbare. These results demonstrate that the early embryogenesis-specific expression of Ping is responsible for the successful amplification of mPing in EG4. This study helps not only to elucidate the whole mechanism of mPing amplification but also to further understand the contribution of MITEs to genome evolution.
| Transposable elements are major components of eukaryotic genomes, comprising a large portion of the genome in some species. Miniature inverted-repeat transposable elements (MITEs), which belong to the class II DNA transposable elements, are abundant in gene-rich regions, and their copy numbers are very high; therefore, they have been considered to contribute to genome evolution. Because MITEs are short and have no coding capacity, they cannot transpose their positions without the aid of transposase, provided in trans by their autonomous element(s). It has been unknown how MITEs amplify themselves to high copy numbers in the genome. Our results demonstrate that the rice active MITE mPing is mobilized in the embryo by the developmental stage-specific up-regulation of an autonomous element, Ping, and thereby successfully amplifies itself to a high copy number in the genome. The short-term expression of Ping is thought to be a strategy of the mPing family for amplifying mPing by escaping the silencing mechanism of the host genome.
| Transposable elements (TEs) are DNA sequences that are capable of jumping from one genomic locus to another and make up a large fraction of eukaryotic genomes. More than 80% of the maize (Zea mays) and barley (Hordeum vulgare) genomes are composed of TEs [1], [2], and they constitute 35% and 14% of the genomes of rice (Oryza sativa) and Arabidopsis (Arabidopsis thaliana), respectively [3], [4]. TEs are harmful to the host because their mobilities perturb genome stability, whereas they play greatly generative roles in genome evolution such as alternation of gene structure, change of expression pattern, and rearrangement of chromosome structure [5], [6].
TEs are classified into two groups according to their transposition mechanisms: class I elements (retrotransposons) that transpose through a copy-and-paste mechanism via an RNA intermediate, and class II elements (transposons) that transpose through a cut-and-paste mechanism without undergoing an RNA intermediate. Class I elements easily attain tens of thousands of copies, whereas the majority of class II elements cannot amplify themselves to 50 copies at most. Unlike other class II elements, miniature inverted-repeat transposable elements (MITEs) have the capacity to amplify themselves to high copy numbers (hundreds or thousands) [7]–[9]. In the rice genome, MITEs are numerically predominant TEs [10], constituting 8.6% of the genome [11]. Because MITEs are too short (<600 bp) to encode any protein, their transpositions must depend on the proteins encoded by the autonomous elements. Well-studied MITEs are classified into the Stowaway and Tourist families, which belong to the Tc1/mariner and PIF/Harbinger superfamilies, respectively. Because MITEs are mainly deployed in gene-rich regions [10], [12] and affect adjacent gene expression [13], they are considered to play an important role in genome evolution. However, little is known about how MITEs attain high copy numbers.
Miniature Ping (mPing) is the first active MITE identified in the rice genome [14]–[16]. Although MITEs are deployed in the genome at a high copy number, the copy number of mPing exceptionally remains at a low level in most rice cultivars: indica and tropical japonica cultivars have fewer than 10 copies, and temperate japonica cultivars including Nipponbare have approximately 50 copies [14]. The transposition of mPing is suppressed in most rice cultivars, but, like other TEs, mPing is activated by exposure to various stress conditions such as gamma-ray irradiation [16], hydrostatic pressurization [17], cell culture [14], anther culture [15], and inhibition of topoisomerase II [18]. Introgression of distantly related genomes also causes mPing transposition [19], [20]. However, mPing is actively transposing without such stresses in the temperate japonica rice strain EG4 (cultivar Gimbozu) under natural growth conditions, and its copy number is approximately 1000 copies [21]. This indicates that mPing has overcome the silencing mechanism or established a novel strategy for its amplification in the EG4 genome. In this sense, mPing in EG4 is an appropriate material to study the amplification of MITEs in plant genomes.
The autonomous element Ping and its distantly related element Pong, which both belong to the PIF/Harbinger superfamily, provide two proteins required for mPing transposition. Both Ping and Pong have two open reading frames (ORFs), ORF1 and ORF2 [22], [23]. The former encodes a Myb-like DNA-binding protein, and the latter encodes a transposase lacking DNA binding domain. Transposase of most class II elements contains a conserved catalytic domain (DDE motif) and a DNA-binding domain [23], [24], whereas these domains are encoded separately by two ORFs in both Ping and Pong [22], [23]. The study of other members of the PIF/Harbinger superfamily suggested that the Myb-like DNA-binding protein directly binds to the subterminal regions of the transposon in order to recruit the transposase [25]. Both Myb-like protein and transposase of either Ping or Pong or both elements are necessary for mPing transposition [22], [23].
In this study, we demonstrate that mPing is actively transposing in the embryo of EG4 during the period from the regionalization of shoot apical meristem (SAM) and radicle to the formation of the first leaf primordium (3 to 5 days after pollination, DAP) with the aid of developmental stage-specific expression of Ping. Our results provide important evidence for the amplification mechanism not only of mPing but also of other MITEs.
Plants have acquired the silencing mechanism of TEs in germ cells. In Arabidopsis, for example, TEs are activated specifically in the vegetative nucleus of the pollen, and siRNAs from the activated TEs accumulate in the sperm cells [26]. On the basis of these results, Slotkin and colleagues proposed that siRNAs derived from TEs activated in the vegetative nucleus silence TEs in the sperm cells [26]. We conceived that mPing might overcome such a silencing mechanism in EG4. To confirm this hypothesis, we developed two F1 populations from reciprocal crosses between the mPing-active strain EG4 and the mPing-inactive cultivar Nipponbare, and investigated the transposition activity of mPing by transposon display (TD) analysis. Success of reciprocal crosses was confirmed by PCR analysis using locus-specific primers (Figure S1A). One of the results of TD analysis using two selective bases is shown in Figure 1A; all 16 possible primer combinations were analyzed. The banding patterns of F1 plants were almost the same as those of EG4. The bands that appeared in all F1 plants but not in the parental EG4 plant were derived from another parental Nipponbare plant (Figure S1B). Furthermore, the bands that appeared in only one of eight F1 plants but not in the parental EG4 plant are herein referred as de novo insertions. These bands were confirmed not to be PCR artifacts by sequence and locus-specific PCR analysis (Table S1 and Figure S2). We detected 15.5 de novo insertions per plant in the selfed progenies of EG4, whereas Nipponbare yielded no de novo insertions in the selfed progenies (Figure 1B). This confirmed that mPing is active in EG4 under natural growth conditions but inactive in Nipponbare. If mPing was specifically activated in the pollen of EG4, we could obtain de novo insertions only in the F1 plants from the Nipponbare/EG4 cross. However, we obtained de novo insertions in both Nipponbare/EG4 and EG4/Nipponbare populations (Figure 1B). Moreover, there was no significant difference in the number of de novo insertions per plant between the two F1 populations. This indicates that the activating factor(s) for the mPing transposition is present in both male and female gametes of EG4.
We performed TD analysis of mPing using genomic DNA samples extracted from endosperm, radicle, and leaf blades of eight progenies (S1) derived from a single parental EG4 plant (S0), and investigated the mPing transposition during ontogeny of rice plants (Figure 2A). One of the results of TD analysis using two selective bases is shown in Figure S3; all 16 possible primer combinations were analyzed. We examined de novo insertions in the same way as described above. Consequently, a total of 228 de novo insertions were detected. These insertions were divided into five groups (Figure 2B): (1) endosperm-specific insertions that appeared only in the endosperm sample, (2) radicle-specific insertions that appeared only in the radicle sample, (3) leaf-specific insertions that appeared only in one sample from the 1st to 3rd leaf blades, but not in the 4th and 5th leaf blades, (4) shoot-specific insertions that appeared in at least one sample of 1st, 2nd, and 3rd leaf blades, and in at least one sample of 4th and 5th leaf blades, and (5) radicle/shoot-specific insertions that appeared in both radicle and leaf blade samples. These de novo insertions were confirmed by sequence and locus-specific PCR analysis (Table S2 and Figure S4). Numbers of each insertion obtained in this study are summarized in Figure 2C. Plant development is divided roughly into three successive phases: embryogenesis, vegetative phase, and reproductive phase. If mPing transposed in the SAM of the S0 plant during vegetative and/or reproductive phases, the de novo insertions would segregate according to Mendel's law among the S1 progenies. We obtained no band that appeared in at least two S1 progenies and was not seen in the S0 plant. This indicates that the transmissible insertion of mPing to the next generation seldom (or never) arises during the vegetative and reproductive phases.
Flowering plants have evolved a unique reproductive process called double fertilization. In this process, either of two sperm cells in pollen fuses with either an egg cell or a central cell in the ovule, and then the egg cell fertilized with the sperm cell initiates embryogenesis [27]. In rice, the SAM and radicle are regionalized in the embryo 3 DAP, and three leaves and the radicle are already present in the mature embryo [28]. We detected only three radicle/shoot-specific insertions (Figure 2C), indicating that mPing scarcely transposes during the period from the onset of gametogenesis to the early stage (until 3 DAP) of embryogenesis. Among the 228 de novo insertions, 116 and 17 were shoot-specific and leaf-specific insertions, respectively (Figure 2C). This indicates that mPing actively transposes in the embryo during the period from the regionalization of SAM and radicle (at 3 DAP) to the formation of the 3rd leaf primordia (at 8 DAP). Of the 133 shoot- and leaf-specific insertions, 108 were of the 1st leaf blade (Figure 2D). Since the 1st leaf primordium is formed at 5 DAP, the most active phase of the mPing transposition was considered to be from 3 to 5 DAP. We detected a large number of radicle-specific insertions as well as shoot-specific insertions, and the sum of these insertions accounted for 90% of all insertions detected in this study (Figure 2C). Taken together, we concluded that mPing in EG4 most actively transposes in the 3 to 5 DAP embryo.
Endosperm is a triploid tissue that is produced by fusing a central cell containing two polar nuclei with one of two sperm cells in no particular order. The endosperm formation occurs in parallel with embryogenesis. The endosperm-specific insertions result from the mPing transposition occurred in either gametogenesis or endosperm formation. We observed only two endosperm-specific insertions (Figure 2C), supporting that mPing scarcely transposes during the period from the onset of gametogenesis to the early stage of embryogenesis. The relationship between the banding patterns obtained in TD analysis and the timing of mPing transposition is summarized in Figure S5.
In order for mPing to amplify, the de novo insertions must be transmitted to the next generation. We performed TD analysis using 12 progenies (S2) derived from the main culm and the primary tiller of a single selfed parent (S1) to investigate whether the de novo insertions detected in ontogenical analysis are inheritable (Figure S6). Both radicle-specific and leaf-specific insertions in the S1 plants were not detected in the S2 progenies (0 of 15, 0 of 2, respectively). In contrast, 85% (11 of 13) of the shoot-specific insertions that were detected in the S1 plants also appeared in the S2 progenies. This value (85%) is consistent with the estimated number of inheritable de novo insertions in our previous report [21]. Thus most of the de novo insertions that arose in the 3 to 5 DAP embryo were successfully inherited to the next generation.
We have already determined the sites of all mPing insertions (1163 in total) in the EG4 genome [13], and have investigated mPing excisions in a small EG4 population using locus-specific primer pairs [29], [30]. Here we examined the timing of the mPing excision with locus-specific PCR using the genomic DNA samples that were used for the ontogenical analysis of the de novo insertion. We randomly chose 48 markers for this study (Table S3). We divided the mPing excisions into five types with the same criteria as those used for the de novo insertions: endosperm-, radicle-, leaf-, shoot-, and radicle/shoot-specific excisions (Figure S7). There were no endosperm-specific and radicle/shoot-specific excisions, indicating that no mPing transposition occurs during the period from the onset of gametogenesis to the early stage of embryogenesis. We detected seven radicle-specific, six leaf-specific, and three shoot-specific excisions. All shoot-specific excisions were detected from the 1st leaf blade sample. These results indicate that, like the de novo insertion, the mPing excision also occurs during the period from the regionalization of the SAM and radicle to the formation of the first leaf primordium, although we cannot completely rule out the possibility that these excisions occur also in somatic cells of mature tissues. Thus, in addition to the experimental results of the de novo insertion, we concluded that mPing of EG4 was most actively transposing in the 3 to 5 DAP embryo.
Both Ping and Pong provide a Myb-like protein and a transposase, which are encoded by their ORF1 and ORF2, respectively (Figure 3A), and have been considered as autonomous elements responsible for the mPing transposition. We investigated the expression of Ping-ORF1, Ping-ORF2, Pong-ORF1, and Pong-ORF2 during embryogenesis to evaluate which autonomous element plays a predominant role in driving the mPing transposition in EG4. Reverse transcription-PCR analysis revealed that Ping-ORF1 and Ping-ORF2 constitutively expressed in the ovary during embryogenesis (Figure 3B). On the other hand, no transcriptions of Pong-ORF1 and Pong-ORF2 (Figure 3B) were observed. This strongly suggests that Ping predominantly controls the mPing transposition in EG4.
We performed real-time quantitative PCR (qPCR) analysis to compare the expression level of Ping-ORF1 and -ORF2 between EG4 and Nipponbare during embryogenesis. In all developmental stages from 1 to 6 DAP, the expression levels of both Ping-ORF1 and -ORF2 were higher in EG4 than in Nipponbare (Figure 3C, D). Since EG4 harbors seven copies of Ping, whereas Nipponbare has only one copy (Table S4), the difference in the expression levels between EG4 and Nipponbare is considered to be attributable to the different copy number of Ping. However, we found that Ping of EG4 showed different expression patterns from that of Nipponbare. In Nipponbare, the expression level of Ping-ORF1 and -ORF2 gradually declined until 3 DAP, and restored to the basal level at 6 DAP. In contrast, in EG4, the expression levels of both Ping-ORF1 and -ORF2 rapidly increased, with a peak at 3 DAP (Figure 3C, D). The ratio of relative expression level (EG4/Nipponbare) clearly demonstrated that Ping might be up-regulated in a developmental stage-specific manner in the ovary of EG4 (Figure 3E). Since mPing transposed during the period from 3 to 5 DAP, the rapid increase in Ping expression most likely drive the mPing transposition.
We investigated the spatial expression pattern of Ping by in situ hybridization using Ping-specific probes. The probe positions were indicated in Figure 3A. The Ping transcripts were detected in all tissues, viz. embryo, endosperm, and ovary wall, in both EG4 and Nipponbare (Figure 4A–C, S8). Among the tissues, the 3 DAP embryo of EG4 yielded an exceptionally strong signal, indicating a high accumulation of Ping transcripts (Figure 4A), whereas the 5 DAP embryo showed a much lower accumulation of Ping transcripts in EG4 (Figure 4D–F). Such a drastic change in accumulation quantity of Ping transcripts with the advance of embryogenesis was consistent with the change in the expression quantity of Ping with the advance of embryogenesis, which was investigated by real-time qPCR (Figure 3C–E). These results suggest that the tissue- and developmental stage-specific accumulation of the Ping transcripts triggers mPing transposition at this stage in EG4. To confirm this hypothesis, we evaluated the spatial expression pattern of Ping in the SAM during the vegetative phase. As described above, mPing hardly transposes in the SAM during this phase. The Ping transcripts were detected in all tissues including the SAM, and, as expected, there was no obvious difference in the signal intensity between EG4 and Nipponbare (Figure 4G–I). Thus the Ping transcripts proved to accumulate developmental stage-specifically only in the tissue where mPing actively transposes. We therefore concluded that the high accumulation of Ping transcripts triggers the transposition of mPing in the 3 DAP embryo of EG4.
EG4 has seven Ping elements (Ping-1 to -7), whereas Nipponbare has only one (Ping-N) (Table S4). When we sequenced and compared all Ping elements, a single nucleotide polymorphism (SNP) in the first intronic region of Ping-ORF1 was detected between EG4 and Nipponbare (Figure 5A). Ping-N has a ‘T’ nucleotide on the SNP region, whereas all Ping elements in EG4 have a ‘C’ nucleotide. We named the former ‘T-type Ping’ and the latter ‘C-type Ping’.
In addition to EG4, several Aikoku and Gimbozu landraces (hereafter AG strains) are known to exhibit high mPing activity [21]. We investigated the SNP-type of Ping and the copy number of Ping and mPing in 93 AG strains, and evaluated the effect of C-type Ping on the mPing activity. These 93 AG strains were divided into three groups according to the SNP-type of the Ping allele (Table S4): strains harboring C-type Ping; strains harboring T-type Ping; and strains harboring no Ping. The strains with C-type Ping had more mPing copies than those with T-type Ping or no Ping (Figure 5B, Steel–Dwass test, p<0.01). This implies that the C-type Ping could drive the mPing transposition. We further investigated the expression patterns of Ping-ORF1 and -ORF2 in two mPing-active strains (A119 and A123) and two mPing-inactive strains (A105 and G190) during embryogenesis (from 1 to 6 DAP). A119 and A123 have six and ten copies of C-type Ping, respectively, and both A105 and G190 have one copy of T-type Ping (Table S4). Expression analysis revealed that A105 and G190 kept low expression levels of Ping-ORF1 and -ORF2, whereas A119 and A123 showed high expression levels with a peak around 3 DAP (Figure 5C–F). This indicates that, in EG4, A119, and A123, the developmental stage-specific expression of Ping is controlled by the same factor(s) described in the Discussion.
Our final goal was to elucidate how MITEs attain their high copy numbers in the genome. To this end, we chose mPing, which is the only active MITE identified in rice, as a material and analyzed the timing of mPing transposition in the mPing-active strain EG4. Consequently, we successfully found one mechanism of the mPing amplification; mPing most actively transposes during the period from the regionalization of the SAM and radicle to the formation of the first leaf primordium (3 to 5 DAP) by the developmental stage-specific up-regulation of the autonomous element Ping.
The transpositions of TEs are categorized into germinal and somatic types according to the type of cells where the transposition takes place. LORE1a in Lotus japonicus is activated in plants regenerated from de-differentiated cells and transposes in male germ cells by the pollen grain-specific LORE1a transcription, resulting in the asymmetric transposition of LORE1a in the reciprocal crosses between the active and non-active lines [31]. Tag1 in Arabidopsis shows germinal transposition activity in both male and female germ cells. Consequently, the reciprocal crosses show symmetric transposition of Tag1 [32]. These results demonstrate that the transposition activity in reciprocal crosses reflects the tissue specificity of germinal transposition. In this study, reciprocal crosses between EG4 and Nipponbare showed the same mPing transposition activity, which may suggest that mPing in EG4 transposes in both male and female germ cells. However, we obtained only a few de novo endosperm-specific and radicle/shoot-specific insertions in the ontogenical analysis, although we detected a number of de novo shoot-specific and radicle-specific insertions. We therefore concluded that most mPing transposes not in germ cells but in somatic cells after pollination. Somatic transposition that occurs at the late stage of plant development often produces spotted and striped segments in tissues, such as maize seed coat variegation caused by Mutator excision from the bz2 gene [33], [34] and rice leaf color variegation by nDart1excision from the OsClpP5 gene [35]. In animals, somatic transposition is seldom transmitted to the next generation because germ cells are set aside from somatic cells at the early stage of embryogenesis. On the other hand, in plants, germ cells are generated from somatic cells at the reproductive stage. In rice, gametes are generated in the SAM; therefore, somatic transposition that occurred in the SAM can be transmitted to the next generation via gametes. In this study, we revealed that most mPing elements transposed in somatic cells of the embryo during the period from 3 to 5 DAP. Being a class II TE that transposes by a cut-and-paste mechanism, mPing is expected to be eliminated from genomic DNA with a certain frequency. However, a previous report demonstrated that the mPing excision sites would be repaired by utilizing a copy of mPing from either the sister chromatid or from the homologous chromosome [29]. The mPing excision site cannot be repaired if mPing transposes in germ cells, which are haploid. It is therefore considered that the somatic transposition of mPing is an important factor for mPing amplification in the genome.
The autonomous elements Ping and Pong mediate mPing transposition in the rice genome. Many japonica cultivars, including EG4 and Nipponbare, have both Ping and Pong. This study demonstrated that Ping plays a predominant role in mPing transposition in EG4. However, a heterologous expression assay using Arabidopsis and yeast showed that Pong had a higher catalytic capacity for mPing transposition than Ping [22], [23]. Furthermore, mPing transposition was observed under stress conditions in several rice cultivars harboring only Pong [14], [17], [19]. In this study, however, we detected very low expression of Pong through the development of rice plants, indicating that Pong would be epigenetically silenced at the transcriptional level in EG4. In contrast, Ping constitutively expressed in all organs including the SAM and embryo. Nevertheless, mPing could be transposing most actively in the embryo during the period from 3 to 5 DAP. Since the stage-specific up-regulation of Ping was observed during the period of mPing transposition, we hypothesized that the expression level of Ping needed to exceed a certain threshold of mPing transposition.
All mPing-active strains (EG4, A119, and A123) showed higher expression of Ping with a peak around 3 DAP than the mPing-inactive strains (Nipponbare, A105, and G190). Although further experiments are needed to elucidate the mechanism of developmental stage-specific up-regulation of Ping expression, we propose two hypotheses: (1) position- and dosage-effect, and (2) effect of SNP. The details of the hypotheses are as follows.
Chromosomal position and copy number of TE often affect the transposition activity. The former is known as ‘position effect’ and the latter as ‘dosage effect’. Eight independent Tam3 copies residing in the Antirrhinum majus genome show different transposition activities from each other [36]. In Arabidopsis, germinal reversion frequency of Tag1 increases in proportion to its copy number [32]. The mPing-inactive strains Nipponbare, A105, and G190 have only one Ping at the same locus, whereas the mPing-active strains EG4, A119, and A123 have respectively seven, six, and ten copies of Ping at different loci except for the Ping-1 locus. Furthermore, the expression pattern of Ping showed slight variation among the mPing-active strains harboring only C-type Ping. These results suggest that the developmental stage-specific up-regulation of Ping expression is probably regulated by the position-effect and/or the dosage-effect.
Intronic SNPs are known to cause drastic effects on gene expression. In humans, an intronic SNP in SLC22A4 affects transcriptional efficiency in vitro, owing to an allelic difference in affinity to the transcriptional factor RUNX1 [37]. Furthermore, a SNP located in the intronic enhancer region of the thyroid hormone receptor β gene enhances pituitary cell-specific transcriptional activity [38]. In this study, we demonstrated that a SNP is present in the intronic region of Ping-ORF1, and Ping elements in the AG strains were categorized into either T-type or C-type Ping according to the SNP-type. Since all strains that showed a peak in the expression analysis had only C-type Ping, the intronic SNP might influence the developmental stage-specific up-regulation of Ping expression. T-type Ping was present in 14 AG strains as one copy, and its chromosomal location did not differ among strains. In contrast, the copy number of C-type Ping varied from one to ten, and their chromosomal locations, except for Ping-1, differed from each other. These results indicate that T-type Ping has lost its activity, whereas C-type Ping may be still active in the rice genome. Furthermore, we found that the copy number of mPing was significantly larger in strains harboring C-type Ping than in strains harboring T-type Ping. This strongly supports that C-type SNPs in the intronic region of Ping contribute to the amplification of mPing, presumably by the developmental stage-specific up-regulation of Ping expression.
Since the transposition of TEs often damages the host genome, TEs with high transposition activity are targeted by the silencing mechanisms. Nevertheless, MITEs amplify to very high copy numbers not only in plant genomes but also in animal genomes. Very little is known about how MITEs attain their high copy numbers by escaping the silencing mechanism. The transposition of mPing is transiently induced by various stresses [14]–[18], indicating that the activity of mPing is suppressed by the silencing mechanisms in many cultivars. Thus, mPing must overcome the silencing mechanism in order to maintain the transposition activity under natural growth conditions. Our results revealed that mPing in EG4 was mobilized by the sufficient supply of Ping transcripts produced only during the period of mPing transposition. This stage-specific activation is thought to be a strategy of the mPing family to amplify mPing by escaping from the silencing mechanism of the host genome. Since no active MITEs other than mPing so far have been identified, it is very difficult to elucidate if the other MITEs also attain their high copy numbers in the same way as mPing amplifies. Given that the other active MITEs are identified, however, our study will help to understand their amplification mechanisms. Our previous study documented the generation of new regulatory networks by a subset of mPing insertions that render adjacent genes stress inducible [13]. In addition to mPing, other MITEs also contribute to gene and genome evolution via providing new promoter regulatory sequences, transcriptional termination elements, and new alternative exons [39], suggesting that the amplification of MITEs causes gene and genome evolution. Our results provide clues to further understand not only the amplification mechanism of MITEs but also the co-evolution of MITEs and the host genome.
EG4 (cultivar Gimbozu), Nipponbare, and 94 Aikoku/Gimbozu landraces were used in this study (Table S4). Aikoku/Gimbozu landraces were provided from the GenBank project of the National Institute of Agrobiological Science, Ibaraki, Japan. Reciprocal crosses between EG4 and Nipponbare were made in a green house. Before pollination, all anthers were removed from the flowers of maternal plants. The pollinated flowers were covered with protective bags to prevent outcrossing until harvest. After harvesting, success of crosses was checked with the molecular markers. For ontogenical analysis, eight progenies of EG4 (S1) derived from a single parental plant (S0) were grown in a greenhouse, and all S2 seeds were harvested. For S1 plants, each seed was cut into two halves, and the half including the embryo was germinated and the other was sampled. After germination, the radicle and the 1st, 2nd, 3rd, 4th, and 5th leaf blades were sampled. The second leaf was collected from S0 and S2 plants. For estimation of Ping and mPing copy numbers, eight bulked plants were sampled. For RNA extraction, ovaries before pollination and ovaries at 1, 2, 3, 4, 5, and 6 DAP were collected. All samples were immediately frozen in liquid nitrogen and stored at −80°C until use.
DNA extraction and transposon display was performed following a published protocol [30]. For DNA extraction from endosperm, we used GM quicker 2 (Nippon Gene).
Sequencing of mPing-flanking fragments excised from transposon display gels and primer design were performed following a published protocol [30]. The genomic locations of the mPing insertion sites were forecasted by a BLAST search in the Rice Annotation Project Database (RAP-DB; http://rapdb.dna.affrc.go.jp/) [40], [41] using mPing flanking sequences as queries. To prepare enough templates for PCR, whole genome amplification was performed using an illustra GenomiPhi V2 Kit (GE Healthcare). mPing excision was detected by PCR with mPing-sequence characterized amplified region (SCAR) markers [29]. PCR was performed in 10-µl reaction volumes containing 10 ng of the template DNA, 5 µl of GoTaq Green Master mix (Promega), 5% DMSO, and 0.25 µM of each primer (Table S3). PCR conditions were as follows: 94°C for 3 min; 40 cycles of 98°C for 10 s, 57°C for 30 s, and 72°C for 45 s; and 72°C for 5 min. To detect the presence of Ping-N, -1, -2, -3, -4, -5, -6, and -7, eight Ping-SCAR markers were used. The genomic locations of the Ping insertion sites were referred from a previous report [42]. For detection of the Ping-1 allele, PCR was performed in 10-µl reaction volumes containing 10 ng of template DNA, 0.2 U of KOD FX Neo (Toyobo), 1×PCR buffer for KOD FX Neo (Toyobo), and 0.2 µM of each primer (Table S5). PCR conditions were as follows: 94°C for 3 min; 35 cycles of 98°C for 10 s, 60°C for 30 s, and 68°C for 90 s; and 72°C for 1 min. For detection of Ping-N, -2, -3, -4, -5, -6, and -7 alleles, PCR was performed in 10-µl reaction volumes containing 10 ng of template DNA, 5 µl of GoTaq Green Master mix (Promega), 5% DMSO, and 0.25 µM of each primer (Table S5). PCR conditions were as follows: 94°C for 3 min; 35 cycles of 98°C for 10 s, 60°C for 30 s, and 72°C for 45 s; and 72°C for 1 min.
Total RNA was isolated using TriPure isolation reagent (Roche) and digested using RNase-free DNase (TaKaRa). First strand cDNA was synthesized using a Transcriptor first strand cDNA synthesis kit (Roche). For reverse transcription PCR, PCR was performed in 10 µl reaction volumes containing cDNA generated from 4 ng total RNA, 0.2 U of KOD FX Neo (Toyobo), 1×PCR buffer for KOD FX Neo (Toyobo), and 0.5 µM of each primer. PCR conditions were as follows: 94°C for 3 min; 35 or 45 cycles of 98°C for 10 s, 60°C for 10 s, and 68°C for 10 s. Relative quantification of Ping-ORF1 and Ping-ORF2 were calculated by the 2−ΔΔCT method [43] using Light cycler 1.5 (Roche). The UBQ5 gene was used as the calibrator gene. The thermal profile consisted of 10 min at 95°C; and 45 cycles of 4 s at 95°C, 10 s at 60°C, and 1 s at 72°C. Amplification data were collected at the end of each extension step. The primer pairs used in this study are listed in Table S6.
Plant samples were fixed with 4% (w/v) paraformaldehyde and 1% Triton X in 0.1M sodium phosphate buffer for 48 h at 4°C. They were then dehydrated in a graded ethanol series, substituted with 1-butanol, and embedded in Paraplast Plus. The samples were sectioned at 8-µm thickness using a rotary microtome. Fragments of Ping-ORF1 (1091 bp) and Ping-ORF2 (1368 bp) were cloned into pBlueScript SK+ (Stratagene) and sequenced. For digoxigenin-labeled antisense/sense RNA probe synthesis, in vitro transcription was performed using T7 RNA polymerase and T3 RNA polymerase. In situ hybridization and immunological detection with alkaline phosphatase were performed according to Kouchi and Hata [44].
PCR was performed in 10-µl reaction volumes containing 10 ng of template of DNA, 5 µl of GoTaq Green Master mix (Promega), 5% DMSO, and 0.25 µM of each primer. PCR conditions were as follows: 94°C for 3 min; 35 cycles of 98°C for 10 s, 60°C for 30 s, and 72°C for 30 s; and 72°C for 1 min. PCR primers used in this study are listed in Table S6. Because the original sequence contained an Afa I restriction site, one mutation was introduced into the reverse primer. The 5-µl PCR products were mixed with 5 µl restriction mixture containing 1 U Afa I (TaKaRa), 33 mM Tris-acetate (pH 7.9), 10 mM Mg-acetate, 0.5 mM Dithiothreitol, 66 mM K-acetate, and 0.01% (w/v) bovine serum albumin. After 16 h incubation at 37°C, DNA gel electrophoresis was performed. PCR products (502 bp) including +1261T SNP were not digested, whereas PCR products including +1261C SNP were digested into two fragments (352 bp and 150 bp).
To determine the copy number of Ping by Southern blot analysis, genomic DNA samples were digested with Eco RI restriction enzyme. These samples were loaded onto an agarose gel, separated by electrophoresis, blotted onto a nylon membrane, and probed with the Ping fragment. The mPing copy number was determined by real-time quantitative PCR as described previously [45] with little modification. Quantitative PCR was performed using the LightCycler 480 system (Roche). PCR was performed in 20 µl reaction volumes containing 5 µl genomic DNA (0.4 ng/µl), 1×LightCycler 480 SYBR Green I Master mix (Roche), and 0.5 µM of each primer. Specificity of the amplified PCR product was assessed by performing a melting curve analysis on the LightCycler 480 system.
|
10.1371/journal.pbio.1000402 | Innate-Like Control of Human iNKT Cell Autoreactivity via the Hypervariable CDR3β Loop | Invariant Natural Killer T cells (iNKT) are a versatile lymphocyte subset with important roles in both host defense and immunological tolerance. They express a highly conserved TCR which mediates recognition of the non-polymorphic, lipid-binding molecule CD1d. The structure of human iNKT TCRs is unique in that only one of the six complementarity determining region (CDR) loops, CDR3β, is hypervariable. The role of this loop for iNKT biology has been controversial, and it is unresolved whether it contributes to iNKT TCR:CD1d binding or antigen selectivity. On the one hand, the CDR3β loop is dispensable for iNKT TCR binding to CD1d molecules presenting the xenobiotic alpha-galactosylceramide ligand KRN7000, which elicits a strong functional response from mouse and human iNKT cells. However, a role for CDR3β in the recognition of CD1d molecules presenting less potent ligands, such as self-lipids, is suggested by the clonal distribution of iNKT autoreactivity. We demonstrate that the human iNKT repertoire comprises subsets of greatly differing TCR affinity to CD1d, and that these differences relate to their autoreactive functions. These functionally different iNKT subsets segregate in their ability to bind CD1d-tetramers loaded with the partial agonist α-linked glycolipid antigen OCH and structurally different endogenous β-glycosylceramides. Using surface plasmon resonance with recombinant iNKT TCRs and different ligand-CD1d complexes, we demonstrate that the CDR3β sequence strongly impacts on the iNKT TCR affinity to CD1d, independent of the loaded CD1d ligand. Collectively our data reveal a crucial role for CDR3β for the function of human iNKT cells by tuning the overall affinity of the iNKT TCR to CD1d. This mechanism is relatively independent of the bound CD1d ligand and thus forms the basis of an inherent, CDR3β dependent functional hierarchy of human iNKT cells.
| Our immune system uses randomly modified T-cell receptors (TCRs) to adapt its discriminative capacity to rapidly changing pathogens. The T-cell receptor (TCR) has six flexible, variable peptide loops that make contact with antigens presented to them on the surface of other cells. Invariant Natural Killer T-cells (iNKT) are regulatory T-cells with a unique type of TCR (iNKT-TCR) that recognizes lipid antigens presented by specific MHC-like molecules known as CD1d. In human iNKT-TCRs, only one of the six loops, CDR3beta, is variable. By comparing how different human iNKT clones bind and react to different CD1d-lipid complexes we uncover the existence of a hierarchical order of the human iNKT cell repertoire in which strongly CD1d-binding clones are autoreactive while weak CD1d-binding clones are non-autoreactive. Direct measurements of iNKT-TCR binding to CD1d using surface plasmon resonance recapitulated this hierarchy at the protein level. The data show that variation in the CDR3beta loop conveys dramatic differences in human iNKT TCR affinity that are independent of the CD1d bound ligand. Thus the CDR3beta loop provides the structural basis for the functional hierarchy of the human iNKT repertoire. We postulate that during the life-course, CDR3beta-dependent asymmetrical activation of different human iNKT clones leads to a bias in the iNKT repertoire, and this could result in age-dependent defects of iNKT-mediated immune regulation in later life.
| Invariant Natural Killer T (iNKT) cells are a conserved subset of highly potent and versatile T-cells which specifically recognize the non-polymorphic lipid-presenting molecule CD1d [UniprotKB P15813] [1]. iNKT cells co-express a unique T-Cell Receptor (iNKT TCR), which mediates recognition of CD1d, and the pan-NK receptor NKR-P1A (CD161). Human and mouse iNKT TCRs feature a homologous invariant TCRα chain, i.e. Vα24-Jα18 in humans and Vα14-Jα18 in mice. In addition, all human iNKT TCRs make use of a single TCR Vβ family, Vβ11, whereas mouse iNKT TCRs utilize several different TCR Vβ families.
The current paradox of iNKT biology lies in the fact that, despite their apparent innate-like simplicity, they can exert directly conflicting functions. On the one hand, several in vivo studies have demonstrated an essential role for iNKT cells in the induction and maintenance of immunological tolerance [2],[3]. Consistent with this, iNKT cells exert a protective role in animal models of spontaneous autoimmunity [4],[5], and numerical and functional defects of iNKT cells are observed in different human autoimmune diseases [6].
In contrast to these tolerogenic functions, iNKT cells can exert potent cytotoxic functions and contribute to host defense against tumors and various infectious pathogens [7],[8],[9]. Whether different subsets of iNKTs are involved in these opposed roles or whether individual iNKT clones fulfill both of these functions under different conditions is unknown. Several mechanisms underpin iNKT activation during host defense, such as TLR [10],[11],[12] and PPAR-γ activation [13], co-stimulatory molecule signaling [14], and inflammatory cytokines [15],[16]. However, it is unknown how iNKT cells are induced to mediate their tolerogenic functions under non-inflammatory conditions.
Some iNKT clones exhibit substantial activation in response to CD1d-expressing antigen-presenting cells in the absence of exogenous antigens. This autoreactive function is essential for both iNKT selection [17] and tolerogenic activity [18]. While iNKT TCR binding to CD1d is absolutely required [19], the mechanistic basis of iNKT cell autoreactivity is largely unresolved. In particular, the importance of specific CD1d-presented endogenous lipid antigens for the autoreactive interaction of the iNKT TCR with CD1d is contentious.
Studies in mice have suggested that the iNKT repertoire displays clonal heterogeneity with regard to recognition of weaker stimulatory lipid antigens, such as the α-galactosylceramide analogue OCH. These differences can be explained by the differential Vβ family usage in mouse iNKT TCRs [20],[21],[22]. However, human iNKT TCRs use a single Vβ family and so the short hypervariable complementarity determining region (CDR3β) loop in human iNKT TCRs is their only truly adaptive element. It is not known whether this is sufficient to endow the human iNKT TCR with meaningful ability to discriminate a diverse range of human CD1d-presented antigens.
Here we examined a large panel of human iNKT cell lines and clones for their binding to different CD1d-ligand tetramers and related this both to the affinity of their TCRs to different CD1d-ligand complexes and to their functional recognition of diverse antigens. The results presented here demonstrate that variations in the CDR3β loop have a profound, antigen-independent, impact on the iNKT TCR's affinity to CD1d and on iNKT cell autoreactive function.
Previous studies have shown that the CDR3β loop is dispensable for the ability of human iNKT cells to strongly react to the α-galactosylceramide antigen KRN7000 (K7), a xenobiotic glycolipid which can be presented to iNKT cells by CD1d. In fact, K7-CD1d tetramer staining does not allow discrimination of different human iNKT cell subsets by flow cytometry. We hypothesized that CD1d-tetramers loaded with weaker antigens might be better able to reveal the existence of CDR3β-dependent variation among human iNKT cells.
Therefore, we first examined whether different human iNKT subsets could be segregated by their binding to CD1d tetramers that were loaded with the synthetic iNKT partial agonist antigen OCH. For this purpose, polyclonal iNKT lines, generated from healthy donors by in vitro stimulation with K7, were tested for their binding to both K7- and OCH-CD1d tetramers. In all of these lines, K7-CD1d tetramers stained a single, clearly distinct, homogeneous, and strongly fluorescent population of iNKT lymphocytes (Figure 1A). In contrast, staining of the same lines with OCH-CD1d tetramers revealed a considerable degree of variation in fluorescence, suggesting the presence of distinct iNKT subpopulations (Figure 1A). Importantly, similar qualitative differences between K7- and OCH-CD1d tetramer staining of iNKT cells could also be observed ex vivo (Figure 1B), indicating that these differences were not due to an artifact of previous in vitro stimulation with K7. In order to examine whether the broadly heterogeneous OCH-CD1d tetramer staining of human iNKT cells resulted from stable clonal variation or from transient changes in TCR expression levels, we generated a large panel of “K7/OCH-naïve” human iNKT cell clones and lines. For this purpose, Vα24+/Vβ11+ T cells were directly sorted ex vivo from healthy human donors and expanded using the non-specific T cell mitogen phytohaemagglutinin. Ninety-seven different human Vα24+/Vβ11+ T cell lines and 256 Vα24+/Vβ11+ T cell clones from 13 different healthy donors were established and analyzed by flow cytometry with K7- and OCH-CD1d tetramers.
All Vα24+/Vβ11+ T-cell clones and lines showed bright, homogeneous staining with K7-tetramers (Figure 2), thereby confirming them as iNKT cells. Individual iNKT clones showed modest variation, up to 6-fold, in K7-CD1d tetramer mean fluorescence intensity (MFI). In contrast, multiple iNKT cell subpopulations with differing fluorescence intensities were revealed by OCH-CD1d tetramer staining in 31 of the 97 iNKT lines (Figure 2A), thereby mirroring the above described findings in K7 stimulated iNKT lines. As expected, all 256 iNKT clones stained homogeneously with OCH-CD1d tetramers. However, substantial differences, up to 200-fold, in OCH-CD1d tetramer MFI were observed between individual clones (Figure 2B). Based on the observed large differences in OCH-CD1d tetramer MFI, the 256 human iNKT clones were categorized as OCHHIGH (MFI>300; n = 41), OCHINT (MFI>50 and <300; n = 164), or OCHLOW (MFI<50; n = 51).
Importantly, the differences in OCH-CD1d tetramer staining could not be explained by differences in either TCR or CD4 co-receptor expression. Whereas K7-CD1d tetramer binding significantly correlated with surface expression levels of the Vα24 and Vβ11 TCR chains, no such association was observed for OCH-CD1d tetramer staining (Figure 2C). Furthermore, CD4 co-receptor usage was not related to the intensity of the iNKT clones' OCH or K7-CD1d tetramer staining (unpublished results).
The results of these experiments revealed that the human iNKT repertoire is broadly heterogeneous with regard to the ability of individual clones to bind OCH-CD1d tetramers, independent of either CD4 co-receptor or TCR expression levels.
The above results indicated that clonally distributed qualitative differences in iNKT TCRs were responsible for the considerable variation in OCH-CD1d tetramer binding. However, differences in iNKT TCR mediated recognition of an unnatural compound like OCH would be physiologically irrelevant if they simply reflected random differences in OCH-specific antigen selectivity. To explore this possibility, 18 iNKT clones of broadly varying OCH-CD1d MFI were tested for their ability to bind CD1d tetramers loaded with the common mammalian glycolipid β-glycosylceramide (βGC). These 18 iNKT clones displayed significant variation, up to 50-fold, in βGC-CD1d tetramer staining (Figure 3A). Importantly, a strong association was evident between OCH-CD1d tetramer staining and βGC-CD1d tetramer staining, while no correlation was seen between βGC-CD1d tetramer staining and Vα24 TCR chain surface expression (Figure 3B). These results demonstrated that the observed broad variation in OCH-CD1d tetramer binding between individual human iNKT clones was not simply due to their antigen selectivity but was a reflection of a general variability in human iNKT TCR binding to CD1d loaded with weak antigenic lipids. Furthermore, they indicated that OCH-CD1d tetramer binding can act as a surrogate marker for human iNKT cell recognition of endogenous CD1d antigens.
Based on the above results we hypothesized that the observed substantial differences in tetramer staining between OCHHIGH and OCHLOW iNKT clones resulted from significant variations in TCR:CD1d binding affinity. As expected, sequencing of the TCR Vα24 and Vβ11 chains demonstrated the usage of the known invariant Vα24-Jα18 rearrangement in all clones, while Vβ11 in these clones was rearranged with several different Jβ families, resulting in highly variable CDR3β sequences. This indicated that, in human iNKT TCRs, structural differences of the CDR3β loop have a substantial impact on iNKT TCR binding to CD1d. To test this in a cell-free system we cloned the extracellular domains of the TCR Vβ11 chains from a panel of seven OCHHIGH and OCHLOW iNKT cell clones (Table 1), as well as the invariant TCR Vα24 chain from one iNKT clone, and used them to generate soluble Vα24/Vβ11 iNKT TCRs. Binding of these recombinant iNKT TCRs to K7-, OCH-, as well as βGC- and lactosylceramide (LacCer-) loaded recombinant human CD1d complexes was measured using surface plasmon resonance (Figure 4A; Table 2).
The results of these experiments showed a striking variation, up to 40-fold, between the different iNKT TCRs in their binding affinity (KD) to a given ligand-CD1d complex (for K7-CD1d, KD: 0.24–3.67 µM; for OCH-CD1d, KD: 2.17–38.3 µM; for βGC-CD1d, KD: 2.17–85 µM; for LacCer-CD1d, KD: 2.1–54 µM; see Table 2). These findings clearly showed that the CDR3β loop of human iNKT TCRs can strongly impact on their binding to ligand-CD1d complexes.
Importantly, the binding affinities of all seven recombinant iNKT TCRs to OCH-CD1d strongly correlated with the OCH-CD1d tetramer staining (MFI) of their corresponding original iNKT clones (Figure 4B). Moreover, the binding affinity of a given iNKT TCR to OCH-CD1d also correlated closely with its affinity to either βGC- or K7-CD1d (Figure 4C). Therefore, the wide variation in affinity between our seven human iNKT TCRs contrasted to the lack of variation in antigen selectivity. In other words, the CDR3β loop of human iNKT TCRs modulated the overall binding affinity to different human ligand-CD1d complexes irrespective of the bound ligand.
Based on these findings we hypothesized that the TCRs of OCHHIGH iNKT clones could also mediate enhanced functional recognition of endogenous ligand-CD1d complexes. We tested this hypothesis by comparing autoreactive responses of OCHHIGH and OCHLOW iNKT clones to CD1d-expressing antigen-presenting cells.
We directly compared the extent of proliferation, cytokine secretion, and cytotoxicity of human OCHHIGH and OCHLOW iNKT cells in response to CD1d expressing human cell lines presenting either endogenous or specific exogenous (“pulsed”) glycolipids. Because functional responses of iNKT cells might change during long term in vitro culture, we compared different donor-matched pairs of OCHHIGH and OCHLOW iNKT cell clones with identical in vitro history, i.e. each pair was sorted from a given donor 3 wk prior to the experiment and kept under identical cell culture conditions until the day of the experiment. The selected clones were all CD4+ and were additionally matched for TCR expression levels. For all pairs, OCHHIGH iNKT clones exhibited significantly greater proliferation than OCHLOW iNKT clones in response to either unpulsed or OCH-pulsed T2-CD1d lymphoblasts. In contrast, when T2-CD1d were pulsed with the strong agonist ligand K7, both OCHHIGH and OCHLOW iNKT clones proliferated vigorously, and to similar extent (Figure 5A).
Next, we measured CD1d-dependent secretion of a panel of cytokines by OCHHIGH and OCHLOW iNKT clones. The OCHHIGH iNKT clones secreted considerably greater quantities of cytokines than their OCHLOW counterparts in response to either unpulsed or OCH-pulsed T2-CD1d cells (Figure 5B, C), while no significant differences in cytokine secretion were observed between OCHHIGH and OCHLOW iNKT clones upon stimulation with K7-pulsed T2-CD1d cells. A general Th0-type cytokine secretion pattern was observed in response to stimulation with either K7 or OCH, while a Th1 pattern was often produced by autoreactive stimulation of OCHHIGH iNKT (Figure 5C). Although most OCHLOW iNKT clones did not exhibit autoreactive cytokine release, two OCHLOW iNKT clones reproducibly secreted significant amounts of IL-13 and either IL-4 or IL-5, but no IFNγ or TNF-α, while one OCHLOW iNKT clone secreted measurable amounts of IFNγ and TNF-α, but no Th2 cytokines.
None of the tested iNKT clones secreted detectable amounts of cytokines in response to CD1d-deficient T2-lymphoblasts, and blocking of surface CD1d molecules on T2-CD1d by the monoclonal antibody CD1d42 effectively prevented autoreactive secretion of cytokines by OCHHIGH or OCHLOW iNKT cells (unpublished data). Therefore, autoreactive cytokine secretion by these iNKT clones was wholly dependent on their recognition of surface CD1d.
Finally, in Cr51 release assays, OCH-pulsed T2-CD1d were much more efficiently killed by OCHHIGH iNKT clones than their corresponding OCHLOW iNKT clones (Figure 6D). In contrast, K7-pulsed T2-CD1d were efficiently lysed by both OCHHIGH and OCHLOW iNKT clones, whereas neither OCHHIGH nor OCHLOW iNKT clones showed relevant cytotoxicity towards unpulsed T2-CD1d lymphoblasts.
Together, these results demonstrated that OCH-CD1d tetramer staining allows for identification of distinct human OCHHIGH and OCHLOW iNKT clones, which exhibit differential functional ability to respond to endogenous ligand-CD1d complexes. The above results indicated that the autoreactive potential of human iNKT clones is governed by the affinity of their iNKT TCR to CD1d, and therefore the structure of their CDR3β loop.
In order to test our hypothesis that OCHHIGH and OCHLOW iNKT TCRs differed in their binding to endogenous ligand-CD1d complexes, we generated soluble fluorescent iNKT TCR-tetramers derived from an autoreactive OCHHIGH iNKT clone and a non-autoreactive OCHLOW iNKT clone. As shown in Figure 6, both iNKT TCR tetramers bound well to K7-pulsed T2-CD1d. In contrast, only the OCHHIGH-derived iNKT TCR tetramer was able to effectively stain unpulsed T2-CD1d. These results further substantiated our hypothesis that autoreactive recognition of CD1d by human iNKT cells is primarily determined by the structure of their iNKT TCRs' CDR3β loop.
All together, these studies demonstrated that the human iNKT cell repertoire exhibits considerable clonally distributed CDR3β-dependent differences in overall TCR affinity to CD1d, irrespective of the bound ligand, and that these inherent structural differences control iNKT autoreactive activation.
iNKT cells are a conserved subset of highly potent regulatory T cells at the innate-adaptive interface. The hallmark of human iNKT cells is their unique TCR, which is composed of an invariant TCR Vα24-Jα18 alpha chain and a semi-invariant TCR Vβ11 chain. The only variable, and therefore potentially adaptive, element in human iNKT TCRs is their hypervariable CDR3β loop. The results of the present study demonstrate for the first time, to our knowledge, that the structure of the hypervariable CDR3β loop in human iNKT TCRs exerts a strong impact on CD1d binding and is a key determinant of iNKT cell autoreactivity. The magnitude of the effect of CDR3β variations on human iNKT TCR:CD1d binding observed here was unexpected as previous studies with mouse iNKT TCRs have reported only minor effects of CDR3β mutations on CD1d binding. Furthermore, they strongly suggest that CDR3β loops in autoreactive iNKT TCRs make functionally important direct protein-protein contacts with human CD1d, rather than contacts with CD1d-bound ligands, thereby affecting overall affinity rather than antigen specificity.
The role of the hypervariable CDR3β loop in human iNKT TCRs is currently unresolved. It is dispensable for binding to CD1d molecules that are loaded with the strong agonist ligand K7, and hence K7-CD1d tetramers do not support subset differentiation of human iNKT cells. Consistent with this, the recently solved structures of one human and two mouse iNKT TCR:K7-CD1d co-crystals have found no relevant contacts between CDR3β and the K7-CD1d complex [20],[23]. In contrast, recent mutagenesis studies have indicated that the CDR3β loop of mouse iNKT TCRs may exert some impact on the affinity to CD1d, particularly when CD1d was loaded with weaker antigens [24],[25],[26].
We found that human iNKT cells were surprisingly heterogeneous in their binding to CD1d tetramers loaded with the partial agonist ligand OCH, which is a synthetic analogue of K7. Up to 200-fold differences in OCH-CD1d tetramer staining were observed between individual iNKT clones, independent of variations in TCR expression. The same clones exhibited only modest differences in K7-CD1d tetramer staining, which could largely be explained simply by variations in TCR expression. Importantly, we found that the clonal variation in OCH-CD1d tetramer binding was directly related to OCH-CD1d dependent functional responses, while no such linkage was observed between K7-CD1d tetramer staining and K7-dependent functional iNKT activation. These data underpinned the notion that the five germline encoded CDR loops in human iNKT TCRs, i.e. CDR1α-3α and CDR1β-2β, are sufficient for effective iNKT cell interaction with K7-CD1d [26]. Importantly, they strongly indicated that productive iNKT TCR interactions with OCH-CD1d require additional binding energy provided by certain CDR3β loop structures. We tested this hypothesis by directly measuring the binding of K7- and OCH-CD1d complexes to a panel of seven recombinant human iNKT TCRs, which were derived from selected OCHHIGH and OCHLOW iNKT clones. These recombinant iNKT TCRs differed only in their CDR3β structure. The results of these experiments demonstrated that the broad clonal heterogeneity in OCH-CD1d tetramer staining is indeed directly determined by the iNKT clones' TCRs binding affinities to OCH-CD1d, and hence the structure of the CDR3β loop. Conversely, while all tested recombinant iNKT TCRs bound approximately 10-fold better to K7-CD1d than to OCH-CD1d, the fold-differences in affinity between the strongest and the weakest binding iNKT TCRs were similar for binding to either OCH- or K7-CD1d. Together with the above discussed tetramer-based and functional studies, this indicates that the synthetic CD1d ligand K7 pushes the interaction between human CD1d and iNKT TCRs beyond a physiological range. This is consistent with numerous in vivo and in vitro studies which showed that K7 induces concurrent massive iNKT cell secretion of TH1-, TH2-, and TH17-type cytokines, whereas OCH causes a clearly TH2-biased cytokine secretion pattern [27]. Also, addition of K7 to mouse fetal thymic organ cultures leads to effective deletion of iNKT cells [28], and K7 stimulation induces a prolonged anergy in iNKT cells [29], which supports the view that K7 is not a physiological ligand for iNKT cells. Hence, a full understanding of the biological role of CDR3β loop polymorphism will require more studies with weaker agonistic antigens, and the results of this study suggest that OCH is a good surrogate for endogenous weak agonist antigens.
There are two competing models to explain how differences in CDR3β loop structure could translate into variations of weak antigen recognition. In an “antigen-dependent” or “adaptive” model, the CDR3β loop bestows upon iNKT cells a degree of lipid selectivity by controlling iNKT TCR affinity to CD1d in a lipid antigen-specific manner. Alternatively, in an “antigen-independent” or “innate-like” model, the CDR3β loop structure modulates iNKT TCR binding affinity to CD1d via protein-protein interactions. This model would allow higher, but not lower, affinity TCR structures to recognize CD1d molecules presenting weaker lipid antigens but, crucially, without differential patterns of lipid antigen selectivity between iNKT TCRs of similar CD1d affinity. In other words, this model predicts that the inherent CDR3β sequence in a given human iNKT clone would determine its iNKT TCR's general ability to bind to diverse ligand-CD1d complexes. An important corollary of this would be a fixed hierarchy of high and low affinity iNKT clones. A prediction arising from this model would be that iNKT cells lack the ability to develop immunological memory to specific pathogens, which is a hallmark of adaptive immunity. Although iNKT TCRs clearly belong to the broader family of rearranged, and therefore “adaptive,” TCRs and BCRs, their limited structural diversity and lack of antigen-selectivity, as proposed by this model, are strongly reminiscent of innate immune receptors.
In order to test which of the two above models best explains the observed CDR3β-dependent variation in iNKT TCR binding to OCH-CD1d, we examined recognition of two β-linked glucosylceramides, βGC and LacCer, by a panel of iNKT TCRs. K7 and OCH are α-linked monosaccharide glycosylceramides and are not expressed in mammals, whereas βGC and LacCer are natural β-linked glycosylceramides of mammalian cell membranes. The different configurations of α- and β-anomeric glycolipids enforce substantial differences in the orientation of their glycosyl headgroups when presented by CD1d [30],[31]. Therefore, if the substantial variation in iNKT TCR affinity to OCH-CD1d observed in our study was mainly a function of clonal variation in lipid antigen specificity, as predicted by the “adaptive” model, there should be no association between an individual iNKT TCR's affinity to OCH-CD1d and its affinity to either βGC-CD1d or LacCer-CD1d. However, the results of the present study strongly support the “innate” model: βGC-CD1d tetramer binding to human iNKT clones correlated in a linear fashion with OCH-CD1d tetramer binding, and our binding studies with several different soluble iNKT TCRs demonstrated that the CDR3β loop of human iNKT TCRs strongly modulated the overall binding affinity to different human ligand-CD1d complexes, independent of the bound ligand.
CDR3β loop hypervariability of human iNKT TCRs therefore strongly impacts on overall affinity to CD1d but does not exert a relevant effect on antigen selectivity. The powerful effect of natural CDR3β variations on human iNKT TCR:CD1d affinity observed in our study was unexpected as previous iNKT TCR mutagenesis studies in mice have suggested only a weak impact of CDR3β structure on iNKT TCR binding affinity [24],[25],[26]. Indeed, hybridomata expressing mouse iNKT TCRs with randomized CDR3β regions only displayed moderate variability in binding to K7-CD1d tetramers, and only very few TCRs were capable of interacting with CD1d presenting endogenous lipids [25].Furthermore, previously published iNKT TCR:CD1d co-crystal structures showed a negligible contribution of the CDR3β to the interaction [20],[23]. The apparent discrepancies between these studies and the current findings could indicate relevant species differences, as the mutagenesis studies have concentrated on mouse iNKT binding or else might reflect differences in study design: the only crystal structure study of human iNKT TCR:CD1d binding was limited to a single iNKT TCR of unknown weak antigen-CD1d affinity while the current study systematically screened a large panel of naturally occurring human iNKT clones. Interestingly, while the iNKT TCR used for the human co-crystal structure study displayed very limited contacts between its CDR3β loop and CD1d, a modeling exercise of TCR Vβ11 docking onto CD1d in the same study [23] pointed to a significant degree of plasticity of the CDR3β conformation. In particular, the CDR3β loop of one of our previously published CD1d-restricted Vα24− Vβ11+ TCRs, TCR 5E [32], could make significant contacts with the alpha-2 helix of human CD1d [23]. Consistent with this, a refolded hybrid TCR of the 5E Vβ11 chain and the invariant Vα24-Jα18 chain binds with high affinity to both CD1d/OCH and CD1d/βGC (unpublished data). Therefore, certain CDR3β loop structures can potentially facilitate the recognition of human CD1d loaded with weak ligands by providing additional binding energy to the TCR-CD1d interaction.
Sequence analysis of the CDR3β loops studied did not reveal any obvious correlations between CD1d binding affinity and either physicochemical properties of the loop as a whole or the position of specific residues within the sequence. This is not surprising, given the high degree of conformational flexibility of CDR loops.
The above described considerable binding affinities of some human iNKT TCRs to naturally occurring beta-anomeric glycolipids, i.e. βGC and LacCer, have important implications for the clonal distribution of iNKT autoreactivity. CD1d-dependent autoreactivity of iNKT cells, i.e. their CD1d-mediated activation in the absence of exogenous antigens, is likely to play important biological roles, but the molecular mechanisms determining iNKT autoreactivity have been unresolved. CD1d-dependent autoreactivity is observed in approximately 30% of mouse iNKT hybridomas[19], and studies in iNKT deficient and autoimmune prone mice have shown that autoreactive CD1d-recognition is required for iNKT selection and also iNKT-mediated immunological tolerance [15],[18],[33],[34]. However, much less is known about the role of CD1d-dependent iNKT autoreactivity in humans. Neonatal human iNKT cells exhibit an activated memory phenotype, indicating their in vivo recognition of CD1d molecules in the absence of exogenous ligands [35].
An “adaptive” model has been proposed to explain autoreactive activation of iNKT cells in mouse models of bacterial infection, and it was postulated that autoreactive murine iNKT cells specifically recognize de novo synthesized antigens, such as isogloboside 3 [36]. Consistent with this model, mouse CD1d requires endosomal trafficking to elicit autoreactive activation of murine iNKT cells, which suggests that processing of the ligand-CD1d complex is essential [37]. However, in contrast to mouse iNKT cells, human iNKT cell autoreactivity is not dependent on CD1d trafficking or endosomal acidification [38], again suggesting important species differences between mouse and human iNKT cell activation.
The antigen-independent “innate-like” model discussed above offers a simpler explanation for the clonally distributed iNKT autoreactivity. iNKT clones with higher overall iNKT TCR:CD1d affinity would have an intrinsically greater autoreactive potential than low affinity clones, and these differences in autoreactive potential would be independent of de novo synthesized CD1d-bound ligands. Autoreactive activation of iNKT clones in this model would still be controlled by local conditions, such as TLR signaling [12], CD1d expression [16], or cytokine expression [39]. High affinity iNKT clones would be capable of exerting autoreactive functions under physiological conditions, while low affinity iNKT clones would only be recruited under more pro-inflammatory conditions, e.g. during bacterial infections.
Our functional analyses of autoreactive activation of OCHHIGH and OCHLOW iNKT clones support the “innate-like” model. Firstly, autoreactive activation of several matched pairs of human iNKT clones was closely associated with their OCH-CD1d tetramer binding characteristics. Secondly, only iNKT TCR-tetramers generated from OCHHIGH iNKT clones were able to bind to CD1d-expressing antigen-presenting cells in the absence of exogenous lipid. The above data therefore underpin the “innate-like” model, whereby the hypervariable CDR3β loop balances TCR binding affinity to CD1d protein, and hence the autoreactive potential of an iNKT clone, independent of the bound ligand.
The different activation thresholds of ex vivo sorted human OCHHIGH and OCHLOW iNKT clones shown herein suggest different in vivo functions of these subsets. For example, OCHHIGH and OCHLOW iNKT cells might differ in their ability to drive the formation of immature DCs and consequently in their capability to constitutively promote peripheral tolerance. Finally, it is intriguing to speculate that CDR3β-dependent asymmetrical activation of the human iNKT repertoire could, over time, skew the balance between OCHHIGH and OCHLOW iNKT clones, with ensuing consequences for iNKT-dependent functions in both host defense and immunological tolerance.
Peripheral blood mononuclear cells (PBMC) were isolated from human peripheral venous blood by density gradient centrifugation (Ficoll-Hypaque; Amersham Pharmacia and Upjohn). The study was approved by the local ethics committee (KEK, Bern, Switzerland). All donors gave informed consent. Human iNKT clones and lines were generated by FACSVantage sorting of Vα24+/Vβ11+ T cells into round-bottomed 96-well plates. Sorted cells were stimulated with 1 µg/ml phytohaemagglutinin (Remel, USA) in the presence of autologous γ-irradiated (35Gy) PBMCs. Cells were grown in T cell growth medium (RPMI 1640, 2% human AB serum (SRK, CH), 10% fetal bovine serum (FBS), 0.1 mg/ml kanamycin, 1 mM sodium pyruvate, 1% non-essential amino acids, 1% L-glutamax, and 50 µM 2-mercaptoethanol (all from Gibco Invitrogen) and IL-2 (Proleukin, Chiron) 200 IU/ml). IL-2 concentration in the medium was gradually reduced to 20 IU/ml 3 wk after sorting.
The following fluorescent reagents were used to analyze human iNKT cells: PE-conjugated human CD1d tetramers loaded with either K7, OCH, βGC [40]; FITC-conjugated anti-human TCR Vβ11, PE-anti-human TCR Vα24, (Serotec, UK); PerCP-anti-CD3, FITC-anti-CD3, APC-anti-CD4, APC-anti-CD8, (BD Pharmingen). After addition of staining reagents, cells were incubated at 4°C for 45 min, washed twice in ice-cold PBS/1% FBS, and acquired on a four-color FACSCalibur flow cytometer (Becton Dickinson). Propidium iodide was used to exclude dead cells. Data were processed using CellQuest Pro software (BD Biosciences, USA). Staining with PE-streptavidin conjugated iNKT-TCR tetramers (4C12 and 4C1369) were carried out in the same way as CD1d-tetramer stainings.
Soluble TCR heterodimers were generated as previously described [41]. Briefly, the extracellular region of each TCR chain was individually cloned in the bacterial expression vector pGMT7 and expressed in Escherichia coli BL21-DE3 (pLysS). Residues Thr48 and Ser57, respectively, of the α- and β-chain TCR constant region domains were both mutated to cysteine. Expression, refolding, and purification of the resultant disulfide-linked iNKT TCR αβ heterodimers was carried out as previously described [32].
Streptavidin (∼5,000 RU) was linked to a Biacore CM-5 chip (BIAcore AB, UK) using the amino-coupling kit according to manufacturer's instructions, and lipid-CD1d complexes or control proteins (βGC-CD1b and HLA-A2*01-NY-Eso-1(157-165) complex) were flowed over individual flow cells at ∼50 µg/ml until the response measured ∼1,000 RU. Serial dilutions of recombinant iNKT TCRs were then flowed over the relevant flow cells at a rate of 5 µl/min (for equilibrium binding measurements) or 50 µl/min (for kinetic measurements). Responses were recorded in real time on a Biacore 3000 machine at 25°C, and data were analyzed using BIAevaluation software (Biacore, Sweden). Equilibrium dissociation constants (KD values) were determined assuming a 1∶1 interaction (A+B ↔ AB) by plotting specific equilibrium binding responses against protein concentrations followed by non-linear least squares fitting of the Langmuir binding equation, AB = B×ABmax/(KD+B), and were confirmed by linear Scatchard plot analysis using Origin 6.0 software (Microcal, USA). Kinetic binding parameters (kon and koff) were determined using BIAevaluation software.
Stable human CD1d-expressing T2-lymphoblast lines and clones (T2-CD1d) were generated by spin infection of T2 lymphoblasts with lentiviral particles encoding the human CD1d gene. VSV–G pseudotyped lentiviral particles were generated as previously described [42]. The following primers were used to clone full-length human CD1d into the lentiviral vector pHR'SIN18: 5′-AGCGGGATCCGCCGCCACCATGGGGTGCCTGCTGTTTCTGCTG-3′ (forward), and 5′-GCGTCTCGAGTCACAGGACGCCCTGATAGGAAGTTTG-3′ (reverse). In brief, HEK293T cells were co-transfected with 5 µg of pVSV-G [43], 10 µg of the packaging plasmid pCMV δ8.91 [44], and 15 µg of the human CD1d-encoding transfer vector pHR'SIN18-hCD1d by calcium phosphate method. Viral supernatants were harvested 48–60 h post-transfection, filtered, and concentrated by centrifugation at 25,000 rpm, 4°C for 90 min. Viral pellets were resuspended in 1 ml fresh RPMI 1640 for transduction. Transduced cells were maintained in growth medium for 10 d before sorting of human CD1d-expressing T2 single cells and lines on a FACSVantage SE apparatus (Becton Dickinson, USA), using PE-conjugated anti-human-CD1d antibody CD1d42 (Pharmingen, Switzerland).
T2 lymphoblast cells (T2-) and CD1d-expressing T2 lymphoblast cells (T2-CD1d) were used as antigen presenting cells (APC). 5×104 iNKT cells were plated in a 96-well round-bottom plate in triplicates with either medium alone, with 2.5×104 T2-CD1d, or with T2 lymphoblasts. Before use, T2-CD1d and T2 lymphoblasts were treated with 0.1 mg/ml mitomycin C for 1 h at 37°C and extensively washed with PBS. Lipid antigens (K7, OCH, and βGC) were added at a final concentration of 100 ng/ml. Lipids were solubilized at 200 µg/ml by sonication in vehicle (0.5% Tween-20), which was also used as a negative control. IL-2 was added to the culture medium at a final concentration of 10 IU/ml. Proliferation was measured during the last 18 h of a 96 h incubation by addition of 1 µCi [3H]-methyl-thymidine (1 Ci = 37 GBq, Amersham Pharmacia), followed by harvesting and scintillation counting (Perkin Elmer beta counter).
Levels of IL-4, IL-5, IL-10, IL-13, GM-CSF, IFN-γ, and TNF-α were measured in the cell supernatants, collected after 48 h of incubation, by Bio-Plex suspension array system (Bio-Rad, USA), according to manufacturer's recommendations.
T2 lymphoblasts and T2-CD1d were cultured for 16 h either in the presence of lipid antigens at 100 ng/ml concentration or an equivalent quantity of vehicle. They were then labeled with 100 µCi of 51Cr (GE Healthcare, UK) for 1 h at 37°C and washed 3 times with warm RPMI 1640 supplemented with 1% FBS.
iNKT cells were added in duplicates at different effector-to-target cell ratios and cultured for 4 h. Maximal 51Cr release was determined from target cells lysed by hydrochloric acid. The percentage of specific lysis was calculated by the following formula: [(experimental cpm − spontaneous release cpm)/(maximum release cpm − spontaneous release cpm)] ×100%. Percentage of unspecific lysis was always <20%.
Soluble iNKT-TCR heterodimers were biotinylated via an engineered BirA motif on the C-terminus of their TCR β-chain and then conjugated to PE-streptavidin (Molecular Probes, USA). Multimeric complexes were purified by FPLC (Pharmacia, Sweden) on an SD200 column (Pharmacia, Sweden) and concentrated to 1 mg/ml using Vivaspin20 concentrators (Vivascience, UK).
|
10.1371/journal.pgen.1006929 | Comprehensive analysis of nucleocytoplasmic dynamics of mRNA in Drosophila cells | Eukaryotic mRNAs undergo a cycle of transcription, nuclear export, and degradation. A major challenge is to obtain a global, quantitative view of these processes. Here we measured the genome-wide nucleocytoplasmic dynamics of mRNA in Drosophila cells by metabolic labeling in combination with cellular fractionation. By mathematical modeling of these data we determined rates of transcription, export and cytoplasmic decay for 5420 genes. We characterized these kinetic rates and investigated links with mRNA features, RNA-binding proteins (RBPs) and chromatin states. We found prominent correlations between mRNA decay rate and transcript size, while nuclear export rates are linked to the size of the 3'UTR. Transcription, export and decay rates are each associated with distinct spectra of RBPs. Specific classes of genes, such as those encoding cytoplasmic ribosomal proteins, exhibit characteristic combinations of rate constants, suggesting modular control. Binding of splicing factors is associated with faster rates of export, and our data suggest coordinated regulation of nuclear export of specific functional classes of genes. Finally, correlations between rate constants suggest global coordination between the three processes. Our approach provides insights into the genome-wide nucleocytoplasmic kinetics of mRNA and should be generally applicable to other cell systems.
| All mRNAs start from production in the nucleus, undergo exportation through nuclear pores and finally are degraded in the cytoplasm. A comprehensive characterization of the kinetic rates of all mRNAs is an important prerequisite for a global understanding of the regulation of the transcriptome and the cell. By conducting a time-series experiment and building a mathematical model, we trace the dynamics of mRNAs from the nucleus to the cytoplasm and determine the rates at each kinetic step at transcriptome-wide level. This information allows us to associate mRNA kinetic rates with a wealth of biological features and made some intriguing discoveries. We show mRNA decay is positively linked to transcript length while mRNA export is negatively linked to the length of the 3' UTR. We show binding of splicing factors is associated with faster rates of mRNA export. We provide evidence for global coordination between nuclear export an decay of mRNA. We show genes sharing specific functions tend to have similar nucleoplasmic kinetics, in which ribosomal proteins possessing special kinetic features exclusively stand out. Altogether, our integrated approach to quantitatively determine the rates of kinetic steps on a gene-by-gene basis provides a blueprint to obtain the global understanding of RNA regulation.
| The production, nuclear export and degradation of mRNA are key steps in the control of cytoplasmic mRNA levels. Steady-state levels of transcripts in the cytoplasm are determined by the rates of these three processes. Hence, our understanding of gene regulatory systems requires quantitative knowledge of the relative contributions of each of these steps. Comparison of the kinetic rate constants for these steps across genes may provide insights into mechanisms of differential gene regulation.
Recent technological advances have enabled genome-wide measurements of mRNA dynamics [1, 2] and subcellular distribution [3]. In particular, the utilization of 4-thiouridine (4sU) as a reagent to metabolically label newly synthesized RNA has provided the means to monitor RNA dynamics with a minimal perturbation [1]. Using this approach, various fundamental kinetic rates of mRNA, such as synthesis, splicing and decay have been quantified at genome-wide level in a number of cell types from different species [4–9]. Global quantification of RNA kinetic rates has revealed at least four major biological insights: 1) different classes of genes utilize distinct kinetic strategies to sustain/alter their expression levels; 2) transcription is the primary determinant of the steady-state level of RNA/protein, with contributions much higher than degradation rates; 3) motif analyses and experimental approaches have identified a range of RNA binding proteins that regulate RNA stability [10, 11]; 4) the average rates of RNA decay differ dramatically between species.
These studies of mRNA kinetics have taken the total cellular mRNA as a single entity to calculate the overall turnover rate, overlooking nucleocytoplasmic transportation, which is thought to be a key aspect of mRNA dynamics and has been shown to be regulated by a variety of evolutionarily conserved mechanisms [12, 13]. Here we combined metabolic labeling of mRNA with cellular fractionation to systematically determine mRNA transcription, nuclear export and decay rates for thousands of genes. We developed a mathematical framework that infers nucleocytoplasmic kinetic rate constants from such labeling and fractionation time series data.
We chose Drosophila Kc167 cells as a representative model for metazoan cells, because of their ease to perform experiment and the availability of a wealth of genome-wide information. We report kinetic rate constants for 5420 genes and determine the relative contributions of each of transcription, nuclear export and decay to overall cytoplasmic abundance. Moreover, we uncover links between the three kinetic steps and transcript features, interactions of specific RNA-binding proteins and specific gene classes.
To obtain genome-wide measurements of the nucleocytoplasmic dynamics of mRNA we followed a strategy as outlined in S1 Fig. Briefly, we performed a time series of metabolic labeling of RNA in Drosophila Kc167 cells using 4-thiouridine (4sU). We then isolated nuclear and cytoplasmic fractions from cells at each time point, and determined the relative abundance of “old” (unlabeled) mRNA for thousands of genes by high-throughput sequencing, as a function of time in both fractions. We then fed these measurements into a computational model that describes the process of sequential mRNA transcription, export and decay as a set of differential equations. Parameter fitting of the model to the measurements yielded kinetic rate constants for each of these three steps. Below we describe each step of the approach in more detail.
To label newly synthesized RNA we used 4sU, which is known to have no major effects on gene expression in Drosophila [14]. We further tested the impact of 4sU on gene expression of Kc167 cells by genome-wide comparison of mRNA expression levels between cells treated for 480 minutes with 4sU. The overall gene expression profile was not much affected by 4sU labeling (Spearman’s ρ = 0.98; P = 0) (S2A Fig). A small set of 59 genes that were influenced by 4sU labeling were excluded from subsequent analysis (S2A Fig). We then exposed cells to 4sU for 0, 30, 90, 180, 300, and 450 minutes and subsequently fractionated the cells into nuclear and cytoplasmic portions by hypotonic lysis and centrifugation. In each sample we spiked in a fixed amount of total RNA from the yeast Saccharomyces cerevisiae for normalization purposes, analogous to a previously reported approach [15]. The sum of nuclear and cytoplasmic portions showed very good genome-wide consistency with the unfractionated total transcriptome that was independently measured, indicating that the loss of RNA during fractionation was generally low (ρ = 0.89; P = 0; S2B Fig). We removed 107 genes for which this consistency did not hold up (S2B Fig).
We then purified pre-existing (i.e., unlabeled) RNA by removal of newly synthesized RNA through sulfhydryl conjugation and biotin-streptavidin pull-down [1]. Finally, we isolated poly-adenylated mRNA from the unlabeled fractions and subjected it to high-throughput sequencing. Because it is known that 4sU labeling shows a bias for long genes, we corrected for such bias as described previously [8]. From the changes in mRNA abundance in the two fractions over time we then inferred kinetic rate constants (see below).
We conducted these experiments as two biological replicates, and the reproducibility of the detected mRNA levels was high for all time points (S3A–S3D Fig). In bulk, the reads of both nuclear mRNA and cytoplasmic mRNA showed a continuous decrease over time relative to the yeast spike-in, reflecting the expected replacement of unlabeled mRNA by labeled mRNA (S3E Fig). The amount of unlabeled mRNA eventually asymptotes to a plateau of 7.3±1.2% (S3F Fig; Methods), which may reflect a pool of highly stable transcripts or incomplete removal of labeled mRNA. In order to check the purity of the nuclear and cytoplasmic fractions we monitored intron:exon ratios for a number of transcripts by quantitative reverse transcription polymerase chain reaction (qRT-PCR). This revealed predominant presence of introns in the nuclear fraction, as expected (S3G and S3H Fig). Furthermore, analysis of the high-throughput sequencing reads indicated a substantial enrichment (9.3±2.0 fold) of rRNA in the cytoplasmic fraction (S3I Fig). These results indicate that our measurements of pre-existing mRNA abundance over time in both the nuclear and the cytoplasmic compartments were generally robust.
Subsequently, we applied mathematical modeling to the time series measurements to estimate rates of transcription, nuclear export and cytoplasmic decay for each transcript. We designed a set of first-order ordinary differential equations to describe the nucleocytoplasmic dynamics (Fig 1A; see Methods). We assumed a steady-state model of a pool of non-synchronously dividing cells in which mature transcripts are produced in the nucleus, transported to the cytoplasm, and degraded in the cytoplasm, with each step described by a first-order reaction rate constant. In this model we assumed that transport of mRNA across the nuclear pore complex is unidirectional, which is generally supported by previous studies [12, 13, 16–19]. However, because transcripts redistribute between the nuclear and cytoplasmic compartments in the period between nuclear envelope breakdown and reformation during mitosis, we included kinetic terms to model this process (Fig 1A). Furthermore, we followed the prevailing model that degradation of polyadenylated mRNA occurs predominantly in the cytoplasm [20, 21]. Lastly, because it is not feasible to accurately quantify the abundance of every alternative transcript, we combined sequence reads from alternative transcripts, yielding a single kinetic model for each gene. An analogous mathematical framework based on similar assumptions was reported recently [22].
For each gene, we fitted this model to the experimental data from each of the two biological replicates (S1 Table). As examples, we show the fitting results of the genes Arc1 and Bacc (Fig 1B and 1C). Genome-wide, the goodness of fit was high, as indicated by the coefficients of determination (r2) globally being close to 1 for both nuclear and cytoplasmic compartments (Fig 1D and 1E). For subsequent analyses, out of 5,730 genes that were detected in all the time points of measurements after correction for 4sU labelling bias and exclusion for 4sU labeling and fractionation influence, we retained 5420 genes that have r2 > 0.8. To test whether the data are compromised by undersampling, we down-sampled the sequencing reads to only 25 percent and repeated the modeling. The resulting estimated kinetic rates are generally consistent with the results based on the full dataset, demonstrating the robustness of the modeling (S4A–S4C Fig).
Analysis of the genes that did not fit the first order kinetic model well (r2 < 0.8) indicated that they have substantially lower transcription rates (S5A Fig). This could mean that the modeling is less accurate at low expression levels. However, genes with relatively poor fits also tend to have long introns (S5B Fig); we speculate that processing of these mRNAs is more complex and cannot be captured accurately by our computational model. Importantly, for the set of 5,420 genes that have r2 > 0.8, all three modeled parameters have very good reproducibility between the two replicates (Fig 1F–1H), ruling out overfitting as the cause for the good agreement between the modeled values and experimental data. We therefore used these 5,420 genes for subsequent biological analysis.
Due to normalization to the spiked-in yeast mRNA, transcription is expressed in arbitrary units per minute. We emphasize that these transcription rates refer to the speed of production of polyadenylated mature mRNA, not the distance travelled by RNA polymerase as a function of time. The rates of export and cytoplasmic decay are expressed as fraction per minute, with a median value of 0.83% per minute (corresponding to a half-life of 1.4 hours), and 1.40% per minute (a half-life of 0.8 hours), respectively.
For the majority of genes, the rate constants of transcription, export and cytoplasmic decay span about 5, 0.5 and 0.9 orders of magnitude, respectively (Fig 1F–1H). This prompted us to calculate the relative contributions of the three processes to the genome-wide variance in steady-state mRNA abundance. The results (Fig 1I) indicate that transcription explains most of the variance (89.4%), while the contribution of cytoplasmic decay is lower but still substantial (10.1%). In contrast, the contribution of nuclear export to the genome-wide variance is negligible (0.5%). Our estimation of the relative contribution of mRNA decay to the steady state mRNA abundance is lower than previously estimated for yeast (~30%) [8], but higher than estimated for mouse embryonic stem cells (~1.4%) [23].
Next, we sought to identify potential determinants of individual kinetic steps. First, we investigated a possible relationship between transcript length and kinetic rates. This revealed that transcription rate has a considerable negative correlation with mRNA length (ρ = -0.36; P = 1.1E-164; Fig 2A), suggesting that mature transcripts are generally less efficiently produced from long genes than from short genes. In part, this may be explained by a more extensive (co-transcriptional) splicing of long transcripts. Indeed, we find that transcription rate is negatively correlated with intron content (ρ = -0.24; P = 1.1E-69; Fig 2B) and with the number of exons (ρ = -0.22; P = 1.8E-60; Fig 2C). This is in agreement with observations that elongation tends to slow down at exons [24] and that transcribed length has negative relationship with the rapidity of RNA polymerase II (Pol II) recruitment [25].
Export rates show no correlation with total mRNA length (ρ = 0.01; P = 0.6), suggesting that mRNA length is not a major limiting factor for transport through the NPC. However, export rate does show a notable negative correlation with the length of the 3’ UTR (ρ = -0.26; P = 4.6E-85; Fig 2D) and to a much lesser extent with the length of the coding region (ρ = 0.13; P = 3.2E-22) or 5' UTR (ρ = -0.12; P = 4.3E-18). We speculate that the binding of regulatory proteins to the 3’ UTR may slow down export or actively retain transcripts in the nucleus.
Interestingly, cytoplasmic decay rate shows a considerable positive correlation with total mRNA length (ρ = 0.41; P = 3.8E-215; Fig 2E). We propose two possible explanations for this surprising link. First, cytoplasmic mRNA degradation may be initiated by stochastic attack by an endonuclease. In this model, long mRNAs simply have a higher probability to be cleaved than short mRNAs. The decay rate generally scales sub-linearly with mRNA length, as indicated by a linear regression slope of 0.25 in log-log space (Fig 2E). This may reflect that mRNA is generally folded in a three-dimensional ribonucleoprotein particle, and the proportion of the mRNA that is buried inside this particle may increase with the linear length. A second, not mutually exclusive explanation may be that long mRNA molecules are more likely to contain motifs that have affinity to proteins that target the mRNA to the cytoplasmic decay machinery. Interestingly, decay rate has virtually no correlation with the length of the 3’UTR (ρ = -0.05; P = 8.2 E-4; Fig 2F) which is the primary site where miRNAs act [26, 27]. miRNA-directed degradation may therefore not be the chief mechanism for cytoplasmic decay in Drosophila Kc167 cells.
RNA-binding proteins (RBPs) are well known for their regulatory roles in specific steps of RNA metabolism [28–30], but genome-wide assessment has not been yet carried out. We took advantage of recently published transcriptome-wide RNA interaction profiles of 20 RBPs [31] to uncover putative links with kinetic properties of mRNA. The interaction profiles were generated from Drosophila S2 cells, which are similar to the Kc167 cells that we used in our study. We compared the median kinetic rates of mRNAs that are bound and not bound by each RBP (Fig 3A, 3C and 3D). This revealed that about two-thirds of the RBPs are significantly correlated with each step of kinetic regulation. The differences range from ~8.4 fold for transcription, to ~1.2 fold for export and ~1.9 fold for cytoplasmic decay. Overall, the correlations between RBP binding and the rate constants were similar for long and short transcripts (S6A–S6F Fig), indicating that transcript length is not a substantial confounding factor in this RBP analysis.
The RBP that is associated with the highest transcription rate is Cbp20, a key component of the nuclear cap-binding complex (Fig 3A). Its human homolog has previously been reported to facilitate Pol II release from promoters through interaction with transcription elongation through P-TEFb [32], which should lead to reduction of the fraction of paused Pol II. We investigated this by computing the ‘pausing index’ of Pol II for each gene [33]. This pausing index is generally inversely correlated with transcription rates, and we found Cbp20-bound transcripts to have the lowest pausing index (Fig 3B). It thus is likely that Cbp20 also has a role in releasing paused Pol II in Kc167 cells. The Exon Junction Complex protein Upf1 that links to nonsense-mediated decay also correlates with high transcription (Fig 3A) and a relatively low pausing index (Fig 3B). A previous study in S. cerevisiae indicated that RNA decay factors can boost transcription independent of their function in RNA degradation [34].
The RBPs that are associated with the slowest transcription rate are msi and elav, which have little overlap in their mRNA binding specificities (S6J Fig). However, these two proteins are expressed at extremely low levels in Kc167 cells [35]. It is therefore unlikely that they act as repressors of transcription. We speculate that their target mRNAs may in fact require binding of elav/msi for efficient transcription, and thus may be transcribed at higher levels in the cell types where these proteins are present (e.g., nervous system cells). We tested this hypothesis by comparing the expression level of transcripts in Kc167 that were shown to be bound by the ectopically expressed msi and elav [36], to that of ML-DmBG2-c2, a cell line representing cells of the central nervous system where these two proteins are expressed. Indeed, those genes that can be bound by msi and elav show statistically higher expression in ML-DmBG2-c2 cells (msi: P = 1.9e-7; elav: P = 2.2e-4; two-sided Wilcoxon test). This suggest that msi and elav can promote transcription in the corresponding tissues, but further experimental evidence is needed.
Remarkably, we found that all the 12 RBPs known to be involved in splicing [31] are associated with somewhat higher export rates. These RBPs, which tend to have overlapping mRNA binding specificities (S6J Fig) occupy the top 12 positions when ranked by mean export rate (Fig 3C). This is in agreement with previous reports that have linked splicing factors to nuclear export [37, 38]. In particular, among the splicing factors, SF2 has been shown as the adapter protein for TAP-dependent mRNA export [39]. Paradoxically, we find that mRNAs from intron-containing genes are generally not more rapidly exported than mRNAs from intron-less genes; there is even a slight opposite trend (S6G Fig), which may be due to the fact that intronless genes generally have shorter 3’UTR (S6H Fig), which in turn is associated with higher export rate (Fig 2D). Analysis of the published RBP binding data [31] indicates that transcripts from genes with introns are not enriched for the binding of splicing-related RBPs, compared to transcripts from genes lacking introns (S6I Fig). Together, these data suggest that splicing factors may promote export of mRNA at least in part independently of their role in splicing.
The RBP that is associated with the highest cytoplasmic decay rate is CG6227, a putative DEAD-box containing RNA helicase with so far unknown function (Fig 3D). Lastly, the factor that is associated with the slowest cytoplasmic decay rate is Cbp20. Transcripts with a 7-methylguanosine cap are thought to be bound by Cbp20. This protein is generally restricted to the nucleus, while other proteins take over the cap-binding function in the cytoplasm [40]. Although it is not confirmed that this is also the case in Drosophila, it is therefore unlikely that Cbp20 directly affects cytoplasmic decay. Rather, Cbp20 binding to mRNA as detected in a total cell lysate [31] probably reflects the presence of 7-methylguanosine on the transcripts, and this capping is regulated in Drosophila cells and known to inhibit cytoplasmic decay [41] [42]. In conclusion, this analysis identifies candidate proteins that may control specific steps of nucleocytoplasmic mRNA kinetics.
Chromatin is well-known for its role in regulating transcription, but there is also evidence that it may control the downstream fate of RNA. For example, specific histone modifications can influence alternative splicing [43, 44] and promoter-bound proteins can direct cytoplasmic mRNA stability [45] [46]. We therefore asked whether the kinetic parameters derived from our measurements are correlated with the chromatin environments of the genes. To this end we stratified these parameters by the previously characterized five principal chromatin states [47] at both the transcription start sites (TSSs) and transcription termination sites (TTSs).
As expected, the modeled transcription rates differ widely among chromatin states, both at TSSs and TTSs (Fig 4A–4D). BLUE chromatin, which is marked by H3K27me3 and Polycomb proteins, is associated with very low transcription rates. This is consistent with a wide body of literature indicating that Polycomb complexes directly repress transcription [48]. BLACK chromatin, a hitherto poorly characterized repressive chromatin type that carries H3K27me2 but not Polycomb proteins [14, 47, 49], shows similarly low transcription rates, suggesting that BLACK chromatin also acts at the level of transcription. GREEN chromatin, marked by H3K9me2 and HP1 shows intermediate transcription rates, while the euchromatic YELLOW and RED states show high transcription activity, as expected.
We also observed modest correlations between chromatin states and post-transcriptional kinetic parameters. For both TSS and TTS, transcripts arising from BLACK and BLUE chromatin showed lower export rates than those from YELLOW and GREEN chromatin (Fig 4E–4H), but the difference is only ~1.1-fold. For cytoplasmic decay, we also observed significant chromatin-state-associated differences (Fig 4I–4L), but only at TTSs and of minor magnitude (up to ~1.2 fold). Notably, HP1-containing GREEN chromatin is associated with highest decay rates, which is consistent with a previous study that demonstrated that HP1 mediates heterochromatic transcript decay in S. pombe [50].
We obtained similar results with a 9-state chromatin state model, which is mostly based on histone modification maps [51] (S7 Fig). In particular state 6, roughly equivalent to BLUE chromatin, shows associations with low transcription and low export, especially when present at TSSs. Other states correlate with transcription levels but show only minor differences in export and decay rates. Together, these results indicate that chromatin states primarily affect transcription, and may have intriguing but subtle links to mRNA export and decay.
Cross-talk has previously been observed between mRNA transcription and decay in yeast [3, 9, 34, 52]. This prompted us to analyze possible relationships between the kinetic rate parameters in our data. Pairwise scatterplots revealed a number of interesting patterns. First, there is virtually no correlation between the rates of transcription and export (ρ = 0.06, P = 3.6E-05, Fig 5A). Second, a moderate negative relationship exists between the rates of transcription and cytoplasmic decay (ρ = -0.23, P = 5.3E-66, Fig 5B), which is in part due to a separate group of genes that we will discuss below. Third, we observe a positive relationship between rates of export and cytoplasmic decay (ρ = 0.54, P ≈ 0, Fig 5C). These data suggest global coordination between mRNA decay and both transcription and nuclear export. The underlying mechanism is unclear; in yeast two subunits of RNA polymerase II have been implicated in such coordination [53–55].
One group of genes stands out in the scatterplots (blue dots, Fig 5A–5C). Virtually all genes in this group encode for cytoplasmic ribosomal proteins (cRPs). These genes are characterized by very high transcription, fairly low export, and very low cytoplasmic decay (Fig 5A–5C). It is noteworthy that the cRP genes are completely separated from nuclear genes encoding mitochondrial ribosomal proteins (mRPs), which also cluster in the scatterplots but show less extreme values (green dots in Fig 5A–5C). The basis for this unique regulation is unclear, but we speculate that the 5'-terminal oligopyrimidine tract (TOP) motif, which is found in most cRP mRNAs [56, 57], plays a role in this.
We searched for other functional categories of genes with distinct kinetic parameters by gene ontology (GO) analysis (Fig 5D, S2 Table). Translational machinery genes, primarily comprising of cRP genes, exhibit relatively high transcription, low export and low cytoplasmic decay, as expected. Genes that are responsible for primary metabolic processes are highly transcribed, highly exported and slowly degraded, representing prominent efficiency of expression. The other house-keeping genes, including various synthetic and transport activities of macromolecules, are highly transcribed and lowly decayed but do not possess characteristic export rates. Lastly, it is intriguing that genes linked to neural differentiation and cellular response to stress are enriched for high export, and the latter process is also enriched for high decay. Presumably this provides routes to activate or inactivate gene regulatory cascades in a rapid and flexible manner when extrinsic stimuli are received in developmental processes or stress response. We note that genes with low transcription or high decay may be under-represented in this GO analysis, because they are less likely to have passed our stringent filters for model fitting. Overall, these results reveal curious links between specific functional gene modules and kinetic properties of their transcripts.
By combination of metabolic labeling, cell fractionation and mathematical modeling we determined key parameters of the nucleocytoplasmic dynamics of mRNA for 5,420 genes in Drosophila cells. Our subsequent analyses revealed that these kinetic rates are linked to various molecular components, have relationships with each other and are linked to specific biological processes.
The export and cytoplasmic decay rates deduced from from our measurements and modeling are both on average in the range of ~1% per minute. These values are generally similar to rates estimated by previous studies. Genome-wide studies have determined median mRNA decay rates to range from 1.4% (H. sapiens, [58]), 0.8% (M.musculus, [6]), to 6.3% (S. cerevesiae, [8]), while focused analyses of individual transcripts have yielded values ranging from ~1% per minute in zebrafish [59] to 1.67% per minute in mouse tissues [22], which is very similar to our estimates. For nuclear export, recent microscopy studies estimated the retention time of mRNAs in human cells to be about 40–60 minutes, which corresponds to a rate of 1.2%-1.7% per minute [60], and 8.6 minutes (~8% per minute) in mouse tissue [22]. The latter export rate is somewhat higher than we typically observed, which may be explained by differences in cell type, species, or techniques used. Our estimates of transcription rates are in arbitrary units and can therefore not be compared to other studies.
Our results indicate that nuclear export generally has a relatively minor impact on steady-state mRNA levels. Nevertheless, links with 3'UTR length and the binding patterns of RBPs point to mechanisms that regulate mRNA export. This is in line with previous gene-specific studies indicating that sequences in the 3'UTR of mRNA can affect nuclear export [61, 62]. Our observation that several functional classes of genes show higher or lower export rates points to a certain degree of coordination of the export of mRNAs belonging to the same pathway. Some of the processes that we identified involve responses to DNA damage, stress and nutrients, as well as differentiation. This extends previous observations that the export of individual transcripts can be under control of such signaling events [62–67]. It will be interesting to study the changes in global nucleoplasmic kinetics of mRNAs when these pathways are activated by the appropriate stimuli.
It is also noteworthy that transcripts derived from genes bound by Polycomb complexes (BLUE chromatin) show slightly slower export rates. We speculate that this may be caused by the broad affinity of Polycomb complexes for RNA [68, 69], which may lead to some sequestration of transcripts in the nucleus. This may reduce the availability of the transcripts in the cytoplasm to some degree. Another interesting possibility is that temporary nuclear retention of transcripts may buffer bursts of transcription [22].
We observed that the rate of cytoplasmic decay is positively correlated with transcript length. This is somewhat surprising, because it is generally thought that degradation of mRNA is primarily mediated by the 3' exonuclease activity of the exosome [8, 70–72], which is unlikely to lead to a faster decay for a longer RNA. Rather, the positive correlation with transcript length points to decay mediated by stochastic activity of an endonuclease [73]. A candidate for such endonuclease activity is the Drosophila exosome subunit Dis3, which harbors ribo-endonuclease activity and is expressed in Kc167 cells [74].
In summary, we outlined a generally applicable experimental strategy and a mathematical framework to determine important parameters of nucleoplasmic dynamics for thousands of genes. One possible extension of our strategy is to quantify both the unlabeled and the 4sU-labeled mRNA fractions over time, rather than the unlabeled fractions alone. This may provide an even more precise view of the nucleoplasmic kinetics, particularly of transcripts with high transcription and export rates. The dataset reported here–as well as the uncovered links with mRNA characteristics, RBPs, and chromatin states–provides a foundation to begin to untangle the underlying mechanisms. Application of this approach to other cell types and species will help to understand the global principles of mRNA regulation in the context of differentiation and evolution.
Drosophila Kc 167 were cultured as previously described [75].
Around 1 million Drosophila kc167 cells were separately labelled in 5 ml medium in 10 cm culture dishes with 300 μM 4-thiouridine (Sigma-Aldrich, Cat No. T4509) for 0, 30, 90, 180, 300, 450 minutes. Cells were spun down, washed with serum-free medium and suspended with 120 μl hypotonic buffer consisting of 10mM NaCl, 2mM MgCl, 10mM Tris-HCL (pH = 7.8), 5mM dithiothreitol (DTT), 0.5% nonylphenoxypolyethoxylethanol (NP-40). Suspensions were put on ice for 5 min and spun down at 2000 g at 4 degree for 5 minutes. Supernatants were taken out as cytoplasmic fraction and the pellets were suspended in 120μl hypotonic buffer as nuclear fraction. 700 μl TRIsure (BIOLINE, Cat No. BIO-38032) containing 1 ng/μl total RNA from S. cerevisiae as spike-in was added to both fractions and RNA extractions were performed following the protocol of a published study [1].
1 μg of nuclear and cytoplasmic RNA samples were reverse-transcribed (BIOLINE, Tetro reverse transcriptase, Cat No. BIO-65050) with random-hexamers (BIOLINE, Cat No. BIO-38028). The reaction was subsequently diluted 20 times with water, 4 μl of which was used for qPCR. The primers used are:
Lam exon forward, GAAGACCTGAATGAGGCGCT; Lam exon reverse, TGGTGTTCTCCAGGTCAACG; Lam intron forward, AAGTGCGTGGAAACTGAATCG; Lam intron reverse, CTTGCTTGAAACCACGCCTT; Fmo-2 exon forward, TGATGCAGTGCTTCCACAGT; Fmo-2 exon reverse, ATGTTCTGCACCGGCTACAA; Fmo-2 intron forward, GGCCCCGTGAGATCGATTAG; Fmo-2 intron reverse, TGGTAGCGACGTCACGTATT.
Nuclear and cytoplasmic RNA were labeled with EZ-Link™ HPDP-Biotin (ThermoFisher Cat No. 21341) and pre-existing RNA were purified by removal of biotinylated newly-synthesized RNA as previously described [1]. Total RNA was directly extracted from unfractionated cells following the protocol of a published study [1].
Polyadenylated RNA was purified by oligo-dT beads for both nuclear and cytoplasmic fractions, reverse transcribed (SuperScript II Reverse Transcriptase, Invitrogen, # 18064–014) and constructed into strand-specific libraries using the TruSeq Stranded mRNA sample preparation kit (Illumina Inc., San Diego, RS-122-2101/2) according to the manufacturer's instructions (Illumina, # 15031047 Rev. E). The generated cDNA fragments were 3' end adenylated and ligated to Illumina Paired-end sequencing adapters and subsequently amplified by 12 cycles of PCR. The libraries were analyzed on a 2100 Bioanalyzer using a 7500 chip (Agilent, Santa Clara, CA), diluted and pooled equimolar into a 12-plex for each replicate and subjected to sequencing with 50 base single reads on a HiSeq2500 using V4 chemistry (Illumina Inc., San Diego). The two replicates were sequenced in two separate lanes. The total reads for the two replicates are 182,747,266 and 177,771,796, with even reads distribution for each time point.
Reads were mapped first to the transcriptome of S. cerevisiae (Saccharomyces_cerevisiae.R64-1-1.78) and then to the transcriptome of D. melanogaster (Drosophila_melanogaster.BDGP5.77) by Tophat [76].
For each time point, the number of reads that were mapped to Drosophila transcriptome was divided to the number of reads that were mapped to Saccharomyces transcriptome to obtain the factor for normalization. The number of Fragments Per Kilobase of transcript per Million mapped reads (FPKM) for each gene was calculated using the default setting of Cufflinks [77] and multiplied by the factor for normalization to obtain the relative abundance of transcripts.
We intended to exclude genes that show strong anomalous behaviors which may be due to 4sU labeling or the fractionation procedure. Instead of linear exclusion, we also took the non-uniform distribution of dispersion into account. Considering two transcriptomes x and y of comparison, we used a simple non-linear function to depict the dispersion:
f(x)=sd(y)=m+ne−px
where x is the RNA abundance and f(x) is deviation from the perfect diagonal which corresponds to the standard deviation of y. The abundance of transcripts in the data were divided into 1000 intervals and the corresponding values were calculated. The non-negative parameters m,n,p were then fitted by the Levenberg—Marquardt algorithm using the ModFit function in the package of FME in R. We defined outliers as genes that exceed twice the amount of technical dispersion:
outliers∣{y>x+2f(x) or x>y+2f(y)}
The labelling bias of 4sU as function of the length of genes was corrected previously described [8]. The correction factor is calculated as
F=1−(1−pr)Nu
where pr is the labeling probability that is determined to be around 0.01 [8] and Nu is the number of uridine in the transcripts of individual genes.
To calculate the magnitude of remaining newly-synthesized RNA that contains 4sU after streptavidin removal, we assume first order turnover of total transcripts (detailed in the next section of Quantitative modeling), and add a term for the remaining fraction of 4sU (C) that has not been removed by streptavidin pulldown, similar to a previous approach [6].
Using the notations from the next section of Quantitative modeling, the abundance of newly synthesized RNA over time is
W4sU(t)=W0(egt−e− kTt)
Considering potential contamination factor (U) of W4sU into the unlabeled fraction, and assuming the pre-exsiting RNA follows first order turnover, the pre-existing fraction P is
P(t)= (W0−U⋅W4sU(t))e− kTt=W4sU(t)=W0(1−Uegt+Ue− kTt)e− kTt
And the contamination fraction R(t) is
R(t)= U⋅W4sU(t)=UW0(egt−e− kTt)
Therefore, the expected unlabeled fraction with contamination is
W(t)=P(t)+R(t)
The contamination factor U is estimated by minimizing
d=1n∑i=1n(L(Wm(ti)logWm(ti)W(ti))2)
Where L is the loess function described in the Eq (18) of the next section and Wm(ti) is the measured abundance at a time point i. We used the Levenberg—Marquardt algorithm to fit experimental data using the ModFit function in the package of FME in R. U is estimated to be (7.3 ±1.2) %.
The non-compartmentalized overall mRNA dynamics of non-synchronized Kc167 cells can be described by a simple ordinary differential equation
dW(t)dt=kS−kTW(t),
(1)
whereby for a given gene, W stands for the total amount of bulk mRNA with the unit of Fragments Per Kilobase Of Exon Per Million Fragments Mapped (FPKM), t for time in minutes (min), ks for transcription rate in FPKM·min-1, kT for overall turnover rate in min-1. The equation satisfies the quasi-steady-state assumption, i.e.,
dW(t)dt=gW(t),
(2)
since in standard culture medium, cells in the dish do nothing but merely doubling. Let g be the proliferation rate, which can be calculated from the measured doubling time of 24 hours of kc167 cells,
g=100 ln224*60%=0.048%min−1,
(3)
For pre-existing RNA, total amount Wp satisfies,
dWp(t)dt=− kTWp(t),
(4)
Denoting Wp(0) by W0,
Wp(t)=W0e− kTt.
(5)
Similarly, the simplest model for the nucleocytoplasmic dynamics of mRNA can be written as
(dN(t)dtdC(t)dt)=[kS− (kE+k′f) (kE+k′f)kf−(kD+kf)](N(t)C(t))=g (N(t)C(t)),
(6)
whereby for a given gene, N and C stand for the total amount of mRNA in the, correspondingly, nuclear and cytoplasmic compartments in FPKM. For post-transcriptional kinetic rates, kE stands for exportation rate from the nucleus to the cytoplasm, kD for cytoplasmic decay rate, and kf stands for cytoplasm-to-nucleus inward transfer rate while kf′ for nucleus-to-cytoplasm outward transfer rate during mitosis, which will be discussed later. All post-transcriptional kinetic rates are in the unit of min-1. The equation also satisfies the quasi-steady-state assumption, at t = 0, denoting N(t) by N0 and C(t) by C0,
[kS− (kE+k′f) (kE+k′f)kf−(kD+kf)](N0C0)=g (N0C0).
(7)
For pre-existing Np and Cp,
(dNp(t)dtdCp(t)dt)=[− (kE+k′f) (kE+k′f)kf−(kD+kf)](Np(t)Cp(t)).
(8)
Because
Wp=Np+Cp,
(9)
rewrite Eq (8) in terms of Cp and Wp,
(dCp(t)dtdWp(t)dt)= [− (kE+k′f+kD+kf)0kE+k′f−kT](Cp(t)Wp(t)).
(10)
Therefore,
dCp(t)dWp(t)= (kE+k′f+kD+kf)Cp(t)kTWp(t)−kE+ k′fkT,
(11)
from which we can get the analytical solution of Cp in terms of Wp
Cp(Wp(t))=C0(Wp(t)W0) kE+k′f+kD+kfkT+Wp(t) kE+k′f kE+k′f+kD+kf−kT(1−(Wp(t)W0) kE+k′f+kD+kf−kTkT),
(12)
Because of Eq (5),
Cp(t)=C0e−( kE+k′f+kD+kf)t+W0 kE+k′fkE+k′f+kD+kf−kT(e−kTt−e−(kE+k′f+kD+kf)t).
(13)
Similarly,
Np(t)=N0e−( kE+k′f+kf)t+W0kfkE+k′f+kf−kT(e−kTt−e−(kE+k′f+kf)t).
(14)
To determine the transfer rates of k′f, kf in mitosis, we considered the process of cell cycle. Because of the nature of the quasi-steady state in which stable proportionality of each phase of the cell cycle exists, we can determine during the doubling time of D = 24 hours, the duration of G1/S (FG1/S) takes about 20% of the time and G2/M (FG2/M), in which cells have roughly two fold of cellular content compared to G1/S, takes about 80% of the time, based on published the FACS profile[78] [79] [80]. The duration of mitosis (FM) of drosophila cells takes around 1 hour. Kc167 cells have relatively large nuclei with the ratio of the diameters between the nucleus and the cell equals to rnc = 4:5. Therefore, for every hour the cytoplasmic RNA that diffuses into the nucleus at the end of telophase is
(rnc3×2)D×(FG1/S ×1+FG2/M ×2)(N+C)=2rnc3(N+C)D(FG1/S +2FG2/M ).
(15)
And for every hour the nuclear RNA that diffuses into the cytoplasm at the beginning of M phase is
2ND×(FG1/S ×1+(FG2/M −FM )×2)=2ND(FG1/S +2FG2/M −2FM ).
(16)
To obtain the kinetic rates of nucleocytoplasmic dynamics, we considered four attributes that ought to be satisfied,
RNA-seq experiments render an over-dispersed non-Gaussian distribution for technical noise [81]. To adjust for this effect, we performed local polynomial regression fitting with the coefficient of variation (CV) with the mean value (m) for all the data points from the two biological replicates using the function loess in the package of stats in R, by which we generated function L that represents the numerical correspondence of loess.
Thus, taking differential dispersions at individual time points into account we compute the difference on logarithmic scale, and minimized the corresponding four-component fitting gradient by least square.
We used the Levenberg—Marquardt algorithm to fit experimental data using the ModFit function in the package of FME in R.
To investigate the robustness of modeling in relation to the depth of sequencing, we randomly down-sampled the sequencing reads to only 25 percent of the original number. In this case, the number of genes that were detected in all samples after length correction was reduced to 3519, of which 3403 pass the threshold of r2 > 0.8. The consistency of the modeled rates was very high between the original and the down-sampled data, indicating that the performance of modeling is quite resilient to the reduction of sequencing depth (S4A–S4C Fig).
Annotations from BioMart (http://www.biomart.org/) for Drosophila melanogaster genome BDGP5 were used for these analyses.
For every gene, the length of transcript, intron, 3’UTR and 5’UTR and the number of exons were defined by the maximal values in each category from BioMart annotations. Spearman’s and Pearson’s correlations were calculated to associate length with kinetic rates.
RBP binding data were retrieved from Stoiber et al [31], in both binary form (bound and unbound) and quantitative form (binding scores of the bound genes). To compare the kinetic difference between transcripts bound and unbound by a specific RBP, binary information was used to stratify the genes into two groups and two-sided Wilcoxon rank sum test were performed to calculate the statistical significance that was adjusted with the Holm—Bonferroni correction for multiple comparisons. Because RBP binding may have been underestimated for the 30% lowest expressed transcripts (Figure S2 of [31]), we restricted this analysis to the 70% highest expressed genes (4457 genes total). Overlapping binding of RBP X and RBP Y were calculated as
Overlap(X,Y)=No. of genes bound by X and YNo. of genes bound by X
Overlap(Y,X)=No. of genes bound by X and YNo. of genes bound by Y
To compute the pausing index of Pol II, we used Pol II Chip-seq data of Drosophila Kc167 cells from the modENCODE project (DCCid: modENCODE_5569), and calculated the ratio of Pol II signal within 200 bp around TSS and Pol II signal from 201bp to the end of the gene, similar to previous studies [33]. To investigate the relationship between binding strength of a specific RBP and kinetic rates, Spearman’s correlations were calculated.
Chromatin states data were from Filion et al [47] and Kharchenko et al [51]. For every gene, the coordinates of the most 5’ TSS and the most 3’ TTS from BioMart annotation were defined as the TSS and TTS of the gene, for which corresponding chromatin states were assigned. kinetic rates associated with each chromatin state were compared by ANOVA and the statistical significance was calculated with Tukey’s range test.
GO analysis was performed using the single ranked list method on the Gorilla server ([82], http://cbl-gorilla.cs.technion.ac.il/). Corresponding p values were retrieved and gene ontology processes with p < 10−11 for at least one kinetic processes were displayed in a heatmap generated by the ‘pheatmap’ R package.
|
10.1371/journal.pgen.1002822 | A Positive Feedback Loop Links Opposing Functions of P-TEFb/Cdk9 and Histone H2B Ubiquitylation to Regulate Transcript Elongation in Fission Yeast | Transcript elongation by RNA polymerase II (RNAPII) is accompanied by conserved patterns of histone modification. Whereas histone modifications have established roles in transcription initiation, their functions during elongation are not understood. Mono-ubiquitylation of histone H2B (H2Bub1) plays a key role in coordinating co-transcriptional histone modification by promoting site-specific methylation of histone H3. H2Bub1 also regulates gene expression through an unidentified, methylation-independent mechanism. Here we reveal bidirectional communication between H2Bub1 and Cdk9, the ortholog of metazoan positive transcription elongation factor b (P-TEFb), in the fission yeast Schizosaccharomyces pombe. Chemical and classical genetic analyses indicate that lowering Cdk9 activity or preventing phosphorylation of its substrate, the transcription processivity factor Spt5, reduces H2Bub1 in vivo. Conversely, mutations in the H2Bub1 pathway impair Cdk9 recruitment to chromatin and decrease Spt5 phosphorylation. Moreover, an Spt5 phosphorylation-site mutation, combined with deletion of the histone H3 Lys4 methyltransferase Set1, phenocopies morphologic and growth defects due to H2Bub1 loss, suggesting independent, partially redundant roles for Cdk9 and Set1 downstream of H2Bub1. Surprisingly, mutation of the histone H2B ubiquitin-acceptor residue relaxes the Cdk9 activity requirement in vivo, and cdk9 mutations suppress cell-morphology defects in H2Bub1-deficient strains. Genome-wide analyses by chromatin immunoprecipitation also demonstrate opposing effects of Cdk9 and H2Bub1 on distribution of transcribing RNAPII. Therefore, whereas mutual dependence of H2Bub1 and Spt5 phosphorylation indicates positive feedback, mutual suppression by cdk9 and H2Bub1-pathway mutations suggests antagonistic functions that must be kept in balance to regulate elongation. Loss of H2Bub1 disrupts that balance and leads to deranged gene expression and aberrant cell morphologies, revealing a novel function of a conserved, co-transcriptional histone modification.
| Modification of histone proteins is an important transcriptional regulatory mechanism in eukaryotic cells. Although various histone modifications are found primarily within the coding regions of transcribed genes, how they influence transcription elongation remains unclear. Among these modifications is mono-ubiquitylation of histone H2B (H2Bub1), which is needed for co-transcriptional methylation of histone H3 at specific sites. Here we show that H2Bub1 and Cdk9, the kinase component of positive transcription elongation factor b (P-TEFb), are jointly regulated by a positive feedback loop: Cdk9 activity is needed for co-transcriptional H2Bub1, and H2Bub1 in turn stimulates Cdk9 activity toward one of its major substrates, the conserved elongation factor Spt5. We provide genetic evidence that the combined action of H2Bub1 on Spt5 phosphorylation and histone methylation accounts for the gene-regulatory effects of this modification. Surprisingly, our genetic and genome-wide studies indicate that P-TEFb and H2Bub1 act in opposition on elongating RNA polymerase. We suggest that the positive feedback linking P-TEFb and H2Bub1 helps to maintain a balance between their opposing actions. These results highlight a novel regulatory role for a conserved histone modification during transcription elongation.
| The elongation phase of transcription is a point of regulation for many genes transcribed by RNAPII in eukaryotes, and control of elongation is critical for coupling of transcription to downstream steps in gene expression [1], [2]. Whereas the regulation of transcription at the initiation step has been studied extensively, many of the mechanisms governing elongation in vivo remain to be elucidated. Transcription is accompanied by post-translational modification of nucleosomal histones in a highly conserved pattern, stereotypical features of which include methylation of histone H3 Lys4 (H3K4me) and acetylation of histones H3 and H4 at 5′ ends, methylation of histone H3 Lys36 (H3K36me) towards 3′ ends, and mono-ubiquitylation of histone H2B at a conserved site in the carboxyl-terminus (H2Bub1) throughout coding regions of genes [3]. Conservation of this pattern suggests an important role in coordinating gene expression, but the precise functions of individual modifications in elongation control are poorly understood.
H2Bub1, which appears to play a central role in the interplay between chromatin and the RNAPII elongation complex, is catalyzed in budding yeast by the ubiquitin-conjugating enzyme Rad6 and the E3 ubiquitin ligase Bre1 [4]. Formation of H2Bub1 on transcribed chromatin also requires PAF, a conserved complex with multiple functions during elongation [5]. H2Bub1 is required for co-transcriptional generation of H3K4me by the methyltransferase Set1 in yeast, and contributes to global H3K4me levels in metazoans [6]–[9]. In vitro, H2Bub1 directly stimulates activity of Set1 towards a reconstituted chromatin substrate [10], [11].
H2Bub1 also acts independently of histone methylation; our work in the fission yeast S. pombe revealed that the Set1-independent pathway is important for normal cell growth and morphology, and for elongation by RNAPII at select target genes [12]. Similar findings have now been reported in S. cerevisiae and mammalian cells [13]–[15], but the mechanism of methylation-independent effects of H2Bub1 is unknown.
Co-transcriptional histone modifications are regulated by a conserved subset of cyclin-dependent kinases (CDKs) associated with the transcription machinery [16]. In metazoans these include Cdk7, a component of the initiation factor TFIIH, and Cdk9, catalytic subunit of positive transcription elongation factor b (P-TEFb). Cdk7, Cdk9 and their yeast orthologs phosphorylate multiple proteins important for elongation, including the Rpb1 subunit of RNAPII, at Ser2, Ser5 and Ser7 positions within the repeated YSPTSPS motif of its carboxyl-terminal domain (CTD) [17]–[19]; and the Spt5 subunit of a conserved elongation factor, known in metazoans as DRB-sensitivity inducing factor (DSIF) [20], [21]. Those phosphorylations control recruitment of pre-mRNA-processing and chromatin-modifying enzymes [17], [22], [23].In metazoans, moreover, Cdk9 activity overcomes promoter-proximal pausing imposed by unphosphorylated DSIF and a negative elongation factor (NELF) [20], [24], [25]. Pausing is thought to act both as a quality control over gene expression, by facilitating recruitment of mRNA-processing factors [26], [27]; and as a rate-limiting determinant of expression for subsets of stringently regulated genes [28], [29].
P-TEFb also regulates histone modification. In mammalian cells, levels of H2Bub1, H3K4me, and H3K36me decrease after depletion of Cdk9 by RNA interference (RNAi), or treatment with flavopiridol, an inhibitor of Cdk9 and related kinases [30], [31]. Bur1, the essential ortholog of Cdk9 in budding yeast, is required for co-transcriptional H3K36me and consequent action of the Rpd3S histone deacetylase complex to prevent transcription initiation within coding regions [32]–[34]. Similarly, co-transcriptional H2Bub1 depends on Bur1 [35], [36] and the carboxyl-terminal region of Spt5, which contains sites phosphorylated by Bur1 [37], [38]. Budding yeast contains another, non-essential CDK, Ctk1, which is the major Rpb1-Ser2 kinase in vivo and is required for H3K36me [39], [40]. Although Ctk1 was proposed to be a yeast-specific P-TEFb paralog [41], putative orthologs of Ctk1 have recently been identified in human and Drosophila cells [42], [43], suggesting faithful conservation of the entire CTD kinase network.
CDKs are themselves regulated by histone modification. In mammals, P-TEFb binds the bromodomain protein Brd4, whose recruitment to promoter-proximal sites is stimulated by acetylation of histone H4 [44], [45]. Phosphorylation of histone H3 also favors recruitment of P-TEFb [46], [47]. On the other hand, certain histone marks work in opposition to specific CDKs during elongation. For example, in budding yeast, lethality of a BUR1 deletion is suppressed by deletion of the H3K36 methyltransferase Set2 [34], [48]; and H2B de-ubiquitylation by Ubp8—a component of the SAGA complex—is required for Ctk1 recruitment and Ser2 phosphorylation at some genes [49].
In fission yeast, the P-TEFb ortholog is the essential Cdk9/Pch1 complex. In vitro, Cdk9 phosphorylates the Rpb1 CTD at both Ser5 and Ser2 [50]; inhibition of Cdk9 in vivo reduced the apparent stoichiometry of Rpb1 phosphorylation but did not cause selective loss of either Ser5 or Ser2 signals [51]. Cdk9 also phosphorylates the CTD of S. pombe Spt5 [52]; spt5 mutations that prevent this phosphorylation cause slow growth, defects in transcript elongation and, when combined with partial truncations of the Rpb1 CTD, synthetic lethality [53]. Genetic epistasis suggests, moreover, that the Spt5 CTD contains an exclusive, albeit nonessential, target of Cdk9 in vivo [51]. The potential roles of CDKs and their substrates in regulating histone modifications in fission yeast have not been explored.
Here we uncover a mutual dependence of H2Bub1 and P-TEFb function in S. pombe: mutations that impair phosphorylation of Spt5 by Cdk9 diminish levels of H2Bub1 and, reciprocally, mutants unable to generate H2Bub1 have reductions in chromatin-associated Cdk9 and phosphorylated Spt5. Ablation of the preferred phosphoacceptor site in Spt5 combined with deletion of set1+ phenocopies morphological defects of htb1-K119R mutants, suggesting that Cdk9 and Set1 govern separate, partially redundant pathways downstream of H2Bub1. Despite the dependencies between H2Bub1 and Spt5 phosphorylation, mutations that impair Cdk9 function suppress abnormal cell morphologies and reverse an RNAPII distribution defect, both due to loss of H2Bub1. Conversely, growth of htb1-K119R cells is resistant to selective Cdk9 inhibition, relative to that of htb1+ cells. Taken together, biochemical and genetic results suggest that P-TEFb and H2Bub1 oppose one another to regulate RNAPII elongation, but that proper balance between the two is ensured via a positive feedback loop, in which phosphorylation by Cdk9 stimulates H2Bub1 and vice versa.
Our previous results suggested that H2Bub1 promotes expression of a subset of genes [12]. To ask whether H2Bub1 might act more generally in transcription, we combined ChIP with hybridization of recovered DNA sequences to microarrays that cover the entire S. pombe genome at 200 base-pair intervals (ChIP-chip). We first mapped genome-wide distribution of H2Bub1 in wild-type cells with an antibody that specifically recognizes the ubiquitylated form of histone H2B [[54], Figure S1A]. A FLAG-epitope tag fused to the carboxyl-terminus of histone H2B allowed us to normalize for total H2B occupancy by parallel ChIP-chip with anti-FLAG antibody. The FLAG-tagged H2B was ubiquitylated to a similar extent as the native protein ([12], Figure S1B). To correlate the H2Bub1 pattern with transcription, we also performed ChIP-chip with an RNAPII-specific antibody. We grouped genes into quartiles according to their total levels of RNAPII enrichment and plotted average distributions of H2Bub1 within each group (Figure 1A). This analysis confirmed the presence of H2Bub1 throughout coding regions of transcribed genes in S. pombe, similar to its distribution in S. cerevisiae and metazoans [13], [55]–[57]. Furthermore, H2Bub1 enrichment was correlated with that of RNAPII throughout the genome (r = 0.58).
Because of this genome-wide association, we sought to determine the impact of H2Bub1 loss on RNAPII distribution within genes, by ChIP-chip in wild-type and htb1-K119R mutant cells. This analysis revealed alterations in RNAPII occupancy within gene coding regions in the htb1-K119R mutant: a global, downstream shift in average RNAPII density, reflecting both a decrease in occupancy within the 5′ halves of genes and a more pronounced increase near the 3′ ends (Figure 1B and 1C). These changes occurred in all classes of genes, but varied quantitatively depending on their overall levels of RNAPII cross-linking (Figure 1D; Dataset S1). Genes highly enriched for RNAPII (Figure 1D; red line) had a significant (P<10exp-5) decrease in RNAPII 5′-end occupancy in the absence of H2Bub1, whereas the increase at 3′ ends failed to reach significance (P>0.01). In the other three classes of genes the increases in RNAPII occupancy at the 3′ end were the most significant changes (Figure 1D). Therefore, H2Bub1 globally affects RNAPII distribution in S. pombe genes.
Previous reports have shown that H2Bub1 affects nucleosome stability [9], [14], [58], [59]. We performed ChIP-chip with an antibody that recognizes histone H3 to determine whether the changes in RNAPII distribution caused by H2Bub1 loss were correlated with changes in nucleosome occupancy. The average histone H3 occupancy in gene coding regions did not differ significantly between wild-type and htb1-K119R cells (Figure S2; Datasets S2 and S3). These data suggest that the observed changes in RNAPII occupancy are unlikely to be caused solely by altered nucleosome distribution.
To ascertain whether genes affected by the htb1-K119R mutation were associated with particular cellular functions, we searched for Gene Ontology (GO) terms that were significantly enriched among the 500 genes with the largest increases or decreases in RNAPII occupancy in the htb1-K119R mutant. The genes exhibiting loss of RNAPII occupancy were enriched for ribosomal protein genes (P = 3.58exp-26, 17.7%), consistent with a role for H2Bub1 in promoting RNAPII occupancy at heavily transcribed genes.
The ChIP-chip analysis revealed that RNAPII density was reduced throughout coding regions of 25 of 40 genes whose expression decreased by ≥2-fold in our previous microarray analysis of gene expression [12](Table S1). Furthermore, this group of genes is significantly enriched among the highest RNAPII-occupancy class that we defined by ChIP-chip (P = 7.96exp-5). Thus, the ChIP-chip results and our previous gene-expression data argue that loss of H2Bub1 has a particularly strong negative impact on steady-state levels of poly(A+) mRNA produced from highly transcribed genes. However, the global redistribution of RNAPII in htb1-K119R cells (Figure 1D) suggests that H2Bub1 might affect gene expression more broadly through additional, post-transcriptional mechanisms.
Loss of H2Bub1 caused changes in cell morphology and RNAPII distribution that resemble those produced by inactivation of Rpb1 CTD kinases or loss of Rpb1-CTD Ser2 phosphorylation [12], [51], [60], [61]. We therefore investigated possible interactions between H2Bub1 and three CDKs implicated in transcription: the TFIIH-associated Cdk7 ortholog Mcs6 [61], Cdk9, and the Ctk1 ortholog Lsk1 [62]. First, we determined the effect on H2Bub1 of selective inhibition of Mcs6 or Cdk9, made possible by mutating the “gatekeeper” residue in each kinase to Gly, to render the enzyme sensitive to bulky adenine analogs that do not affect wild-type kinases [63]. We previously replaced mcs6+, cdk9+ and lsk1+ with alleles encoding analog-sensitive (AS) mutant versions, each of which was sensitive to the inhibitory analog 3-MB-PP1 [51]. Selective inhibition of Cdk9, by addition of 20 µM 3-MB-PP1 to cdk9as cells, caused diminution of H2Bub1 signals in whole-cell extracts (Figure 2A). Cdk9 inhibition also decreased levels of H3K4 di- and trimethylation (H3K4me2 and H3K4me3, respectively) and H3K36 trimethylation (H3K36me3) (Figure 2A and data not shown)—modifications likewise associated with actively transcribed genes [3]. Inhibition of the TFIIH-associated kinase in mcs6as cells had small or no effects on H2Bub1, H3K4me2, H3K4me3 or H3K36me3. These results suggest a predominant role for Cdk9 in co-transcriptional histone modification, and a possibly exclusive requirement in generating H2Bub1.
Bur1 in budding yeast and Cdk9 in mammalian cells are proposed to stimulate H2Bub1 by favoring recruitment of PAF, which can in turn associate directly with the H2B ubiquitylation machinery [10], [37], [38], [64]. Inhibition of Cdk9as decreased association of both the PAF component Rtf1 and the E2 ubiquitin-conjugating enzyme Rhp6 with transcribed chromatin (Figure S3A and S3B), suggesting faithful conservation of the P-TEFb-H2Bub1 pathway in S. pombe.
To further probe the dependence of H2Bub1 on P-TEFb, we tested effects of other mutations that perturb Cdk9 function (Figure 2B). Two mutant strains in which Cdk9 activity is reduced by lack of phosphorylation at Thr212 of the activation segment (T loop)—cdk9-T212A and csk1Δ, which is missing the CDK-activating kinase (CAK) responsible for that phosphorylation [50]—had nearly undetectable levels of H2Bub1. In contrast, a cdk9-T212E mutation, which substitutes a Glu residue for Thr212 to mimic constitutive phosphorylation [65], had only a small negative effect on H2Bub1. Loss of H2Bub1 occurred uniquely in mutants with impaired Cdk9 function; neither mcs6-S165A, a T-loop mutation that renders Mcs6 refractory to CAK [66], [67], nor deletions of lsk1+ or lsc1+ [which encodes the cyclin partner of Lsk1 [68]], affected bulk H2Bub1 levels (Figure 2B). Therefore, among S. pombe CDKs implicated in transcript elongation, Cdk9 is uniquely required for H2Bub1.
Impaired Cdk9 function could lead to H2Bub1 loss by decreasing phosphorylation of the Spt5 CTD nonapeptide repeat TPAWNSGSK at Thr1, a site modified by Cdk9 in vitro [52]. To test this we measured H2Bub1 levels in a series of spt5 truncation mutants: spt5ΔC, in which the entire CTD is deleted; or variants containing seven repeats, either of wild-type sequence (spt5-WT), or with every Thr1 position mutated to Ala (spt5-T1A) or Glu (spt5-T1E) [53]. H2Bub1 decreased in spt5ΔC and spt5-T1A mutants, but not in spt5-WT or an spt5-T1E mutant that mimics constitutive phosphorylation (Figure 2C). We conclude that Cdk9 activity and Spt5-CTD phosphorylation are required for H2Bub1 in vivo.
To investigate this connection, and to assess relative contributions of different CDKs to Spt5 phosphorylation in vivo, we generated an antibody specific for the Spt5 CTD phosphorylated on the Thr1 residue (anti-Spt5-T1P), which recognized Spt5 only after it had been phosphorylated by Cdk9 in vitro (Figure S4A). In whole-cell extracts, the phosphorylated Spt5 signal was undetectable after treatment of cdk9as cells with 20 µM 3-MB-PP1, which had little or no effect on Spt5 phosphorylation in wild-type, mcs6as or lsk1as cells, and did not alter Spt5 expression levels in any of the strains (Figure 3A). Spt5-Thr1 phosphorylation was nearly abolished in cdk9as cells at 3-MB-PP1 doses as low as 300 nM (Figure S4B). In contrast, the IC50 for inhibition of cdk9as cell growth was ∼10–15 µM, and Rpb1-CTD Ser2 phosphorylation was relatively resistant to inhibition of Cdk9 alone [51]. There was no additive sensitivity of Spt5-P to the analog in mcs6as cdk9as cells, nor any loss of Spt5-P signal in mcs6as cells treated with 3-MB-PP1 doses up to 40 µM (Figure S4B, S4C). We conclude that Cdk9 is the major, and possibly sole, kinase responsible for phosphorylation of Spt5-Thr1 in vivo, and that levels of Spt5-P are sensitive even to sublethal reductions in Cdk9 activity.
There is a precedent for H2Bub1 regulating the function of a CTD kinase [49]. We wondered whether Cdk9 and the H2Bub1 machinery could be reciprocally regulated, such that the H2Bub1-defective mutants htb1-K119R and brl2Δ (which lacks the Bre1 ortholog required for H2Bub1 [12]) would have global alterations in Spt5-P. Indeed, Spt5-P was decreased relative to total Spt5 (detected with antibodies specific for wild-type Spt5-CTD irrespective of phosphorylation state; Figure S4D) in both htb1-K119R and brl2Δ strains. Conversely, Spt5-P was slightly increased in cells lacking the ubiquitin-specific protease encoded by ubp8+ (Figure 3B). (Cells lacking a second putative H2B de-ubiquitylating enzyme, encoded by the UBP10 homolog ubp16+ [69], had no change in levels of H2Bub1 or Spt5-P [data not shown]). These data suggest a positive feedback loop connecting Cdk9 and the H2Bub1 machinery in vivo.
Previous reports have documented cooperative action of H2Bub1 and the FACT complex, a conserved elongation factor that controls nucleosome dynamics during RNAPII elongation [14], [70], [71]. This prompted us to ask whether FACT might also influence Spt5-P levels in vivo. There was no reduction in levels of Spt5-P, relative to total Spt5, in a strain deleted for pob3+, which encodes a subunit of FACT (Figure S5A). This argues that promotion of Spt5-P reflects a FACT-independent function of H2Bub1.
Another consequence of H2Bub1 loss is reduction in H3K4me—a chromatin modification strongly implicated in gene activation. We showed previously, however, that loss of H2Bub1 caused more severe phenotypes than did ablation of H3K4me due to deletion of set1+. Those phenotypes include increased frequency of cells with division septa in asynchronous populations, and the appearance of cells with abnormal septation patterns (multiple septa separating two nuclei and unseparated chains of cells in which septa had formed) [12]. To test whether impaired Spt5-Thr1 phosphorylation might account for the more severe phenotypes caused by H2Bub1 loss, relative to those due to absence of H3K4me, we constructed an spt5-T1A set1Δ double mutant; the combination, which abolished two protein modifications shown to depend on H2Bub1, phenocopied the increased and aberrant septation produced by H2Bub1 loss (Figure 3C, 3D). Unlike the htb1-K119R and brl2Δ mutants, a set1Δ mutant showed no reduction in Spt5 phosphorylation (Figure 3E). Conversely, the spt5-T1A mutant retained detectable levels of H3K4me3 despite reduced H2Bub1 (Figure 3F). Therefore, H2Bub1 promotes Spt5-P and H3K4me by independent pathways. Importantly, the spt5-T1A mutation by itself did not cause morphological defects, even though it lowered H2Bub1 levels. We attribute this to the residual Set1 function present in the spt5-T1A strain (Figure 3F). A similar double mutant carrying spt5-T1E was morphologically normal, as was an spt5-T1A set2Δ strain, confirming the specificity of the genetic interaction (Figure 3C, 3D). (The pob3Δ mutation had no effect on cell morphology by itself, and did not modify phenotypes caused by htb1-K119R, suggesting that these effects are also FACT-independent [Figure S5B].) Together, these results suggest that Cdk9 and Set1 operate downstream of H2Bub1 to modify components of the transcription machinery and chromatin, respectively, and thereby regulate transcript elongation.
Next, to investigate the basis for reduced Spt5-P in H2Bub1-deficient cells, we measured Spt5-P occupancy at specific genes in the htb1-K119R mutant by ChIP. Control experiments confirmed that the Spt5-P antibody was suitable for immunoprecipitation and ChIP (Figure S6A, S6B). As predicted by immunoblot results, treatment of cdk9as spt5-myc cells with 3-MB-PP1 abolished the Spt5-P ChIP signal without affecting total Spt5 recruitment.
We then compared cross-linking of Spt5-P to chromatin in htb1+ and htb1-K119R cells, at the Cdk9-sensitive eng1+ and Cdk9-insensitive aro1+ genes [51] (Figure 4A and 4B); and at two loci (nup189+ and SPBC354.10+) that, according to our RNAPII ChIP-chip analysis, displayed a 3′ shift in RNAPII distribution in the absence of H2Bub1 (Figure S7A, S7B). A non-transcribed sequence served as a negative control (Figure S8). Although Spt5-P was broadly enriched within gene coding regions, as we observed for H2Bub1, the patterns of Spt5-P and H2Bub1 crosslinking across eng1+ and aro1+ differed, suggesting that the interdependence between these modifications is influenced by other, potentially locus-specific factors (Figure 4A, 4B, 4I, and 4J). Nevertheless, levels of Spt5-P were reduced to varying degrees on all four genes in the htb1-K119R mutant (Figure 4A and 4B, Figure S7A and S7B). Total Spt5 crosslinking was unchanged in htb1-K119R relative to wild-type cells, and the effect on Spt5-P was most evident near the 5′ ends of genes we tested, suggesting that H2Bub1 promotes co-transcriptional phosphorylation (but not recruitment) of Spt5 during the transition from initiation to elongation (Figure 4C–4F; Figure S7C–S7F).
We hypothesized that the reduction in Spt5-P could be due to a defect in recruitment of Cdk9 to chromatin in the htb1-K119R mutant. To test this possibility we compared Cdk9 crosslinking to eng1+, aro1+, nup189+ and SPBC354.10+ by ChIP in htb1+ cdk9-myc and htb1-K119R cdk9-myc strains (Figure 4G and 4H, Figure S7G and S7H). This analysis revealed that Cdk9 recruitment was impaired at all four loci in htb1-K119R cells, consistent with the reduction in Spt5-P signals. Diminished Spt5-P and Cdk9 crosslinking to these loci was not an indirect consequence of reduced RNAPII recruitment, because RNAPII crosslinking and mRNA levels were not reduced at three of the four genes examined (Figure S9). We conclude that lack of H2Bub1 specifically impedes Cdk9 recruitment to transcribed chromatin, leading to impaired phosphorylation of Spt5 within gene coding regions.
Biochemical analyses revealed a reciprocal relationship between Spt5-P and H2Bub1, suggestive of a positive feedback loop. The synthetic phenotype of spt5-T1A set1Δ double mutants, moreover, indicated that the Cdk9-Spt5 axis was one of two partially redundant gene-regulatory pathways downstream of H2Bub1. To detect and characterize direct genetic interactions between P-TEFb and H2Bub1, we tested sensitivity of a cdk9as htb1-K119R double mutant strain to 3-MB-PP1. Unexpectedly, cells bearing a cdk9as allele were less sensitive to growth inhibition by 3-MB-PP1 in an htb1-K119R, compared to an htb1+, background (Figure 5A). In contrast, the htb1-K119R mutation exacerbated 3-MB-PP1-sensitivity of mcs6as and lsk1as cells (data not shown). Therefore, lack of H2Bub1 specifically reduced dependency on active Cdk9, consistent with the two pathways acting antagonistically. Absence of H2Bub1 did not fully bypass the requirement for Cdk9 activity, however, because higher doses of 3-MB-PP1 were still capable of arresting proliferation of cdk9as htb1-K119R cells.
In the absence of inhibitory analogs, cdk9as htb1-K119R cells displayed hyperseptated and branched morphologies characteristic of H2Bub1-defective mutants. After treatment with 10 µM 3-MB-PP1 for 7 hr, however, cell morphology reverted to normal (Figure 5B). This effect was reversible; after washout of 3-MB-PP1 and return to drug-free medium, the “reverted” cdk9as htb1-K119R cells re-acquired an aberrant, hyperseptated morphology.
To confirm that impairment of Cdk9 function could suppress phenotypes caused by the absence of H2Bub1, we combined a cdk9-T212A mutation, which prevents activation of Cdk9 by a CAK but allows cell viability [50], with htb1-K119R or brl2Δ. As was the case with chemical inhibition of Cdk9, cdk9-T212A rescued morphologic phenotypes of H2Bub1-defective mutants (Figure 5C, 5D). Importantly, there was no suppression of htb1-K119R or brl2Δ by lsk1 loss-of-function alleles (Figure 5D and data not shown), establishing the specificity of the interaction between Cdk9 and H2Bub1. Similarly, neither cdk9-T212E nor mcs6-S165A could correct the aberrant morphologies of htb1-K119R cells (Figure 5C). Reversal of septation phenotypes by cdk9-T212A was accompanied by loss of the flocculation observed in htb1-K119R single mutants; ∼15% of wild-type or cdk9-T212A htb1-K119R double mutant cells settled out of liquid cultures after 1 hr, compared to ∼60% of cdk9+ htb1-K119R cells (Figure 5E), consistent with suppression of htb1-K119R by reduction of Cdk9 activity.
We next asked if aberrant morphologies due to htb1-K119R could be suppressed by mutations in the known Cdk9 substrates: Spt5 and Rpb1 [51]. By itself, spt5-T1A did not modify htb1-K119R phenotypes (Figure 6A), implicating another Cdk9 target (or targets) in the observed suppression. We attempted to verify this by combining cdk9as with spt5-T1A in an htb1-K119R background, to allow selective inhibition of Cdk9 in the absence of Spt5-Thr1 phosphorylation. Unexpectedly, the triple-mutant cells grown in the absence of inhibitory analogs had nearly normal morphology—a septation index close to that of the wild-type strain, <5% multiseptated cells and no chained cells (Figure 6B). In vitro, the activity of Cdk9as was reduced ∼3-fold relative to that of wild-type Cdk9, even in the absence of drugs (Figure S10). This is likely to be due to decreased affinity for ATP—a known consequence of the gatekeeper mutation in other AS kinases [72], [73]. The data suggest that loss of Spt5-Thr1 phosphorylation, combined with the partial reduction of Cdk9 activity, suppressed morphological abnormalities caused by loss of H2Bub1.
These results reveal multiple, distinct roles for Spt5-Thr1 phosphorylation in mediating regulatory interactions between H2Bub1 and Cdk9. When Cdk9 activity levels are normal, Spt5-P serves, in combination with Set1, to promote H2Bub1 functions in maintaining normal cell morphology (Figure 3C, 3D). When H2Bub1 is absent, however, Spt5-P, together with at least one other Cdk9-dependent pathway, contributes to aberrant and excessive septation (Figure 6B). A similar distinction between specific elimination of Spt5-P and a general reduction in Cdk9 activity is apparent in a set1Δ background, in which spt5-T1A resulted in cell-morphology defects but cdk9-T212A had no effect (Figure 3D). Thus, the positive feedback between H2Bub1 and Cdk9, involving a single Cdk9 target (Spt5), is genetically distinguishable from the antagonism, which involves multiple targets (Spt5 and one or more others).
Cdk9 also contributes to phosphorylation of Ser2 and Ser5 of RNAPII CTD repeats [50], [51], [74]. An rpb1-S2A mutation that replaced Ser2 in all repeats with Ala in the context of a truncated but still functional CTD [53] failed to suppress htb1-K119R in either a cdk9+ or cdk9as background (Figure 6C). Combined loss of both Spt5-Thr1 and Rpb1-Ser2 in a cdk9+ background also did not suppress htb1-K119R septation phenotypes. Similarly, in a cdk9as background, spt5-T1A but not rpb1-S2A suppressed flocculation due to an htb1-K119R mutation (Figure 6D). Therefore, diminished Cdk9 activity must rescue phenotypes caused by H2Bub1 loss through reduced phosphorylation of Spt5-Thr1 and another target (or targets) besides Rpb1-Ser2.
The suppression of hyperseptation and flocculation was not simply a consequence of reduced growth rate, because hyperseptated, flocculating cdk9+ spt5-T1A htb1-K119R cells and suppressed cdk9as spt5-T1A htb1-K119R cells grew at similar rates in liquid culture. Moreover, the non-suppressed, cdk9as rpb1-S2A htb1-K119R strain grew more slowly than did the suppressed strain (Figure 6E). These data indicate that suppression of htb1-K119R is a specific consequence of reduced Cdk9 activity.
Our results thus far indicated that loss of H2Bub1 partially alleviated the cellular requirement for Cdk9 activity and that, conversely, selective impairment of Cdk9 function suppressed phenotypes of H2Bub1-defective mutants. This mutual suppression led us to hypothesize that P-TEFb and H2Bub1 might act antagonistically to control transcript elongation. We tested such an interaction by asking whether a reduction in Cdk9 activity could reverse the altered RNAPII distribution observed in an htb1-K119R mutant. We analyzed genome-wide RNAPII crosslinking by ChIP-chip in cdk9-T212A and cdk9-T212A htb1-K119R strains (Figure 7; Datasets S4 and S5). The cdk9-T212A mutation caused a global redistribution of RNAPII occupancy that was essentially opposite to that observed in the htb1-K119R strain: an increase in RNAPII density in the 5′ halves of genes, and a decrease toward the 3′ ends of genes that was particularly significant at highly transcribed loci (compare Figure 7C and Figure 1D). Impaired Cdk9 activity also led to reversal of the RNAPII 3′ shift caused by htb1-K119R (Figure 7F). Global RNAPII occupancy profiles in cdk9-T212A htb1+ and cdk9-T212A htb1-K119R mutants were nearly identical, suggesting that, when Cdk9 activity was reduced, H2Bub1 had little or no effect on RNAPII dynamics at the majority of genes (Figure 7A–7E). Loss of H2Bub1 reduced RNAPII occupancy within the 5′ halves of genes even in the context of reduced Cdk9 activity, however, suggesting that this effect is partially Cdk9-independent (Figure 7D, 7E). Nonetheless, full Cdk9 activity is required for the increase in RNAPII density at the 3′ ends of genes in the absence of H2Bub1 (Figure 7E). These data argue that P-TEFb and H2Bub1 antagonize one another to govern the distribution of elongating RNAPII within genes (Figure 8).
Here we combined chemical genetics and genomics to reveal reciprocal relationships between a CDK and a co-transcriptional histone modification in fission yeast. First, we showed interdependence of Cdk9-mediated Spt5 phosphorylation and H2Bub1—evidence for co-regulation governed by positive feedback. Importantly, we found that this feedback loop and the histone methyltransferase Set1 support independent functions of H2Bub1 in vivo. Second, we uncovered genetic interactions between mutations that compromise either Cdk9 activity or H2Bub1, which indicate opposing functions. To explain this paradox we propose a homeostatic mechanism that ensures optimal balance of P-TEFb and H2Bub1 functions during elongation.
A key Cdk9 substrate in the H2Bub1 pathway is Spt5, an ancient component of the transcription machinery that promotes RNAP processivity [75]–[77]. Eukaryotic Spt5 orthologs contain carboxyl-terminal sites phosphorylated by P-TEFb; S. pombe Spt5 has a contiguous array of nonapeptide repeats, which are phosphorylated at Thr1 by Cdk9 in vitro [52]. We showed that Cdk9 activity is necessary for phosphorylation of Spt5-Thr1 in vivo, consistent with genetic epistasis between Spt5-CTD truncation and selective Cdk9 inhibition [51]. Furthermore, reduction of Cdk9 activity or mutation of Spt5-Thr1 to Ala reduced H2Bub1. In budding yeast, Spt5 phosphorylation promotes H2Bub1 through recruitment of the PAF complex [37], [38], and Spt5 also influences H2Bub1 in mammalian cells [31], [78], suggesting a conserved signaling pathway.
Cdk9-dependent phosphorylation of the Rpb1 CTD has been implicated in formation of H2Bub1 in mammalian cells [30], [79]. The sites phosphorylated by Cdk9 within the CTD repeat—Ser2 and Ser5—are essential for cell viability, however, precluding a definitive proof of their involvement in this pathway. Because Ser2 is not essential in S. pombe [53], [68], we have been able to test its contribution to H2Bub1 more directly; the rpb1-S2A mutation had no effect on H2Bub1 levels (data not shown). Understanding the relative contributions of the Rpb1 CTD and other Cdk9 substrates in directing H2Bub1 will probably require better tools for study of metazoan P-TEFb and its targets in vivo.
The identification of Spt5-Thr1 phosphorylation as a specific and sensitive indicator of Cdk9 activity allowed us to uncover another, unexpected relationship between P-TEFb and H2Bub1 in vivo: Cdk9-mediated phosphorylation of Spt5 depends on H2Bub1. ChIP analysis suggested a possible explanation—decreased recruitment of Cdk9 to H2Bub1-deficient chromatin. In mammalian cells, basal association of P-TEFb with the HIV-1 promoter decreased upon depletion of the Bre1 ortholog RNF20, consistent with a conserved role for H2Bub1 in P-TEFb recruitment [80]. Based on our data and results in budding yeast, we propose that a positive feedback loop connects Cdk9 activity and H2Bub1 through the sequence: 1) phosphorylation of Spt5 by Cdk9; 2) PAF recruitment by Spt5-P; 3) recruitment and/or activation of the Rhp6 (S. pombe ortholog of Rad6)-Brl1/2 complex by PAF, to generate H2Bub1; and 4) Cdk9 recruitment, leading to reiteration of the cycle (Figure 8). Although the mechanism by which H2Bub1 facilitates Cdk9 recruitment remains to be determined, there is precedent in metazoans for P-TEFb association with modified histones [44], [45].
Our results indicate that H3K4me and Spt5-P are independently stimulated by H2Bub1 in vivo. Importantly, elimination of either modification alone did not cause septation phenotypes, but their combined ablation phenocopied the aberrant cell morphologies produced by H2Bub1 loss. Therefore, we have identified a molecular intermediate—Spt5-P—in the H3K4me-independent pathway downstream of H2Bub1 (Figure 8). A challenge for the future will be to ascertain how Set1 cooperates with Cdk9, acting through Spt5-Thr1, to regulate gene expression.
The mutually reinforcing relationship between Cdk9 activity and H2Bub1 seemed to predict the two would act in concert during transcription. Instead, genetic analysis indicated they work in opposition: loss of H2Bub1 partially relieved the requirement for Cdk9 activity in vivo, and reduction in Cdk9 activity suppressed cell-morphology phenotypes of H2Bub1-deficient mutants. Simultaneous mutation of two known sites of Cdk9-dependent phosphorylation (Spt5-Thr1 and Rpb1-Ser2) did not suppress htb1-K119R, implicating at least one other Cdk9 target in opposing H2Bub1 function. Consistent with the genetic interaction, Cdk9 and H2Bub1 exerted inverse effects on global RNAPII distribution: whereas RNAPII density shifted toward the 3′ ends of genes in an htb1-K119R mutant, lowering Cdk9 activity produced the opposite effect in an htb1+ background, and prevented the 3′ shift in the double-mutant strain.
The antagonism we observe between Cdk9 activity and H2Bub1 is incomplete—loss of H2Bub1 did not entirely bypass the essential function of Cdk9, and led to altered RNAPII occupancy in the 5′ portions of genes even when Cdk9 activity was reduced. Conversely, although the presence of cdk9-T212A eliminated aberrant cell morphologies associated with htb1-K119R, the growth rate of double mutant cells was slower than that of either single mutant (data not shown). These results imply that other pathways can modulate the functions of Cdk9 and H2Bub1 in transcript elongation.
Our results strengthen the notion that P-TEFb activity is required to overcome repressive effects of histone modifications, first suggested by work in budding yeast. In S. cerevisiae, the instructive genetic interactions occurred between BUR1 and the H3K36 methyltransferase Set2; lethality of a bur1 null mutation was suppressed by deletion of SET2 [34], [48]. In fission yeast, however, set2 deletion did not suppress growth arrest due to inhibition of Cdk9as (L.V. and R.P.F., unpublished observations). Interestingly, bur1Δ was also weakly suppressed by deletion of RAD6 [48], implying that a regulatory interaction between P-TEFb and H2Bub1 might be conserved in budding yeast.
In mammalian cells, the Bre1/Brl2 ortholog RNF20 selectively represses transcription of genes implicated in tumorigenesis; its knockdown reduces H2Bub1, enhances the transcriptional response to growth factors and promotes cellular transformation [13]. These effects were recently shown to depend on the transcription elongation factor TFIIS [81]. This extends the parallel between mammals and fission yeast, by suggesting that H2Bub1 selectively modulates gene expression in both settings by restraining transcript elongation, albeit through different factors—TFIIS or P-TEFb, respectively. Given the established dependence of H2Bub1 on CDK9 activity in mammalian cells [30], [31], it will now be important to ask if metazoan P-TEFb function is also reciprocally influenced by H2Bub1.
How can we reconcile biochemical evidence for mutual dependence of Cdk9 activity and H2Bub1 with genetic evidence for their antagonism? We propose a homeostatic mechanism whereby Cdk9-H2Bub1 interdependence ensures a balance between their opposing functions. In this scenario, Cdk9 promotes H2Bub1 to regulate its own effects on elongation (Figure 8). In an htb1-K119R mutant, the balance is disrupted, resulting in Cdk9-dependent accumulation of RNAPII in 3′ regions of genes. We suggest this pattern reflects unchecked elongation that is poorly responsive to mRNA processing signals, and may imply enhancement of the RNAPII pausing normally associated with mRNA 3′ end processing [82], [83]. H2Bub1 has been linked to various aspects of RNA processing in budding yeast and in mammalian cells, including splicing, 3′ end formation, and nuclear export [30], [57], [84]–[86]. In fission yeast, Cdk9 has been implicated in mRNA 5′-end formation, through its role in recruiting a capping enzyme to transcribed chromatin [87]. Still to be determined are the full range of H2Bub1-dependent functions that depend on P-TEFb activity and the mechanism(s) by which H2Bub1 and Cdk9 influence each other.
In cell division, faithful genome duplication and segregation are ensured by checkpoints—extrinsic signaling pathways that enforce dependency between intrinsically independent events [88]. Transcription and mRNA-processing could be intrinsically coupled through chromatin structure, because removal of nucleosomal barriers is likely to be a limiting step for elongation [89]. Here we have uncovered a role for a covalent chromatin modification, H2Bub1, in regulating elongation through an extrinsic pathway involving Cdk9 and Spt5. We propose a checkpoint-like function of H2Bub1, to set thresholds of Cdk9 activity and Spt5-P required for elongation and mRNA maturation, with the balance between opposing functions of Cdk9 and H2Bub1 maintained by virtue of their interdependence. The absence of H2Bub1 (and consequent decreases in Spt5-P and H3K4me) might hinder mRNA processing events normally coupled to Cdk9-mediated phosphorylations, without fully alleviating the Cdk9-dependence of elongation; “rescue” of these defects by reducing Cdk9 activity would be analogous to rescue of cell-cycle checkpoint mutants by drugs or mutations that slow another process.
Strains used in this study are listed in Table S2. Cells were grown in YE medium containing 250 mg/L each of adenine, leucine, histidine, and uracil (YES). Strains were constructed by standard genetic techniques [90].
To detect Spt5 phosphorylation by immunoblotting, S. pombe whole-cell extracts were prepared as previously described [37], [91]. To detect chromatin modifications, cells were lysed in trichloroacetic acid, as described [92]. The Spt5-P antibody was raised against the peptide acetyl-NSGNK[pT]PAWNVGNK[pT]PAWNSC-amide injected into rabbits, and purified from whole serum after depletion with the same peptide in unphosphorylated form immobilized on resin, by adsorption to and elution from resin-bound acetyl-AWNSGSK[pT]PAWNSGSC-amide by 21st Century Biochemicals (Marlboro, MA). Other antibodies are listed in Text S1.
ChIP was carried out as described previously [12], [93]. Amplification, labeling, and hybridization of control and immunoprecipitated DNA samples for ChIP-chip were carried out as described previously and normalized data are included as Datasets S1, S2, S3, S4, S5 [94]. Further details are included in Text S1. Sequences of primers used for qPCR analysis are available upon request.
S. pombe cells were fixed and stained with diamino-phenylindole (DAPI) and calcofluor as described previously [51]. Cells were viewed using a Leica DM5000b microscope and photographed with a CCD camera. Images were processed using Volocity software.
Cultures were treated with 3-MB-PP1 at indicated doses, or with vehicle (DMSO), for indicated times at 30°C prior to extract preparation. Dose-response to 3-MB-PP1 was determined as previously described [51]. To analyze effects of 3-MB-PP1 on cell morphologies, cells were grown at 30°C to OD600∼0.1 in YES and treated with 10 µM 3-MB-PP1 or DMSO for 7 hr, then fixed and stained for microscopy immediately or after return to growth in drug-free YES for 7 hr.
Assays were carried out as described previously [95].
|
10.1371/journal.pgen.1006259 | The PU.1-Modulated MicroRNA-22 Is a Regulator of Monocyte/Macrophage Differentiation and Acute Myeloid Leukemia | MicroRNA-22 (miR-22) is emerging as a critical regulator in organ development and various cancers. However, its role in normal hematopoiesis and leukaemogenesis remains unclear. Here, we detected its increased expression during monocyte/macrophage differentiation of HL-60, THP1 cells and CD34+ hematopoietic stem/progenitor cells, and confirmed that PU.1, a key transcriptional factor for monocyte/macrophage differentiation, is responsible for transcriptional activation of miR-22 during the differentiation. By gain- and loss-of-function experiments, we demonstrated that miR-22 promoted monocyte/macrophage differentiation, and MECOM (EVI1) mRNA is a direct target of miR-22 and MECOM (EVI1) functions as a negative regulator in the differentiation. The miR-22-mediated MECOM degradation increased c-Jun but decreased GATA2 expression, which results in increased interaction between c-Jun and PU.1 via increasing c-Jun levels and relief of MECOM- and GATA2-mediated interference in the interaction, and thus promoting monocyte/macrophage differentiation. We also observed significantly down-regulation of PU.1 and miR-22 as well as significantly up-regulation of MECOM in acute myeloid leukemia (AML) patients. Reintroduction of miR-22 relieved the differentiation blockage and inhibited the growth of bone marrow blasts of AML patients. Our results revealed new function and mechanism of miR-22 in normal hematopoiesis and AML development and demonstrated its potential value in AML diagnosis and therapy.
| We found that miR-22 is transcriptionally activated by PU.1 during monocyte/macrophage differentiation and miR-22 promotes the differentiation via targeting MECOM (EVI1) mRNA and further increasing interaction between c-Jun and PU.1. We also show that miR-22 is a tumor repressor and that PU.1-miR-22-MECOM regulation is involved in AML development; moreover, we demonstrate that reintroduction of miR-22 relieves the differentiation blockage and inhibits the growth of AML bone marrow blasts.
| Hematopoiesis is a highly ordered multistep process that is orchestrated by various regulators including transcriptional factors [1], cytokines [2], and noncoding RNAs [3]. Deregulation of important regulators in hematopoiesis could induce hematopoietic cancers including acute myeloid leukemia (AML) [4]. In recent years, microRNAs (miRNAs) are emerging as novel regulators in myelopoiesis and AML. The aberrant expression of the miR-17-92 cluster which is epigenetically regulated by PU.1 contributes to leukaemogenesis [5]. MiR-223 acts as a fine-tuner of granulocyte production and the inflammatory response in mice [6,7], while miR-142-3p and miR-29a promote myeloid differentiation [8]. Abnormally expressed miRNAs are associated with AML, serving as powerful prognostic indicators [9–13]. Several miRNAs, such as the miR-29 family [14] and the miR-181 family [15] have proven promising in AML therapy.
To reveal more miRNAs participating in myelopoiesis and AML development, we performed miRNA chip analysis and identified several miRNAs with expression change during monocyte/macrophage differentiation and in AML patients. Among them, miR-22 exhibited evident up-regulation during the differentiation and abnormal down-regulation in AML patients.
MiR-22 is reported to play an important role in several physiology processes and cancers. In mice, miR-22 targets Irf8 mRNA and controls the differentiation of dendritic cells [16]. Moreover, miR-22 is a critical regulator in cardiomyocyte hypertrophy and cardiac remodeling [17,18]. Interestingly, miR-22 can act as either a tumor suppressor or an oncogene in different cancers. miR-22 inhibits cell growth and induces cell-cycle arrest, apoptosis and senescence in breast cancer, colon cancer and lung cancer [19–22]. miR-22 was also reported to promote chronic lymphocytic leukemia B cell proliferation via activation of the PI3K/AKT pathway [23].
MECOM (MDS1 and EVI1 complex locus), also termed EVI1 (Ecotropic viral integration site 1), was first identified as a murine common locus of retroviral integration in myeloid leukemia [24]. Several studies have demonstrated MECOM as a regulator in the maintenance [25] and differentiation [26] of mouse hematopoietic stem cells. However, the function of MECOM in human hematopoiesis is poorly understood. The inappropriate high expression of MECOM is an adverse prognostic marker in AML [27]. MECOM can act as a transcriptional factor [25], epigenetic regulator [28], or repressor of key transcriptional factors in hematopoiesis such as PU.1 and GATA1 via protein—protein interaction [26,29]. MECOM mRNA was previously identified as a miR-22 target in metastatic breast cancer cells [30].
Here, we showed that miR-22 is transcriptionally activated by PU.1 during monocyte/macrophage differentiation, and that miR-22 promotes the differentiation by targeting MECOM mRNA and further increasing interaction between c-Jun and PU.1. We also showed miR-22 to be a repressor miRNA in AML development and examined whether it could be a therapeutic target for AML therapy.
We performed quantitative real-time PCR (qRT-PCR) to detect miR-22 expression in peripheral blood (PB) mononuclear cells (MNCs) derived from 79 primarily diagnosed AML patients (S1 Table) and 114 healthy donors, as well as in bone marrow (BM) MNCs and in BM CD34+ hematopoietic stem cells and progenitors (HSPCs) derived from limitary healthy donors and AML patients. Significantly decreased miR-22 levels were observed in the AML patients as compared with the healthy donors for each kind of the materials (Fig 1A). Receiver-operating characteristic curve analysis of miR-22 suggested that the miR-22 level in each kind of the materials could be as a reference marker with high sensitivity and specificity for AML diagnosis (S1 Fig).
We performed qRT-PCR and Northern blot analyses to detect changes in miR-22 expression during monocyte/macrophage differentiation. The results revealed that miR-22 was gradually elevated during phorbol myristate acetate (PMA)-induced monocyte/macrophage differentiation of HL60 and THP1 cells, as well as during monocyte/macrophage induction of CD34+ HSPCs derived from human umbilical cord blood (UCB) (Fig 1B).
As miR-22 increased substantially during monocyte/macrophage differentiation, we examined if it was regulated by key transcriptional factors. We first performed rapid amplification of cDNA 5’ end (5’ RACE) and identified the transcriptional start site (TSS) of miR-22 at 60 base-pair downstream of the predicted 5’ end of C17orf91 where miR-22 located (Fig 1B). Two potential PU.1 binding sites PB1 and PB2 (Fig 1C) were identified within the region of -2 to +0.2 kb from the TSS using the Transcription Element Search System. Moreover, PU.1 inhibition by siPU1 transfection resulted in obviously reduced miR-22 expression in HL60 and THP1 cells (Fig 1D). Furthermore, PU.1 knockdown by Lenti-shPU.1 infection impaired miR-22 primary transcript (Pri-miR-22) expression, while ectopic expression of PU.1 by Lenti-PU.1 infection induced Pri-miR-22 expression (Fig 1E).
To examine whether PU.1 physically interacts with miR-22 promoter in vivo, we performed chromatin immunoprecipitation (ChIP) assays in HL60 and THP1 cells. The DNA fragments immunoprecipitated by the PU.1 antibody were amplified with two pairs of PCR primers surrounding PB1 and PB2. Only the fragments containing PB2 were detected (Fig 1F). Moreover, ChIP-qPCR showed that the interaction between PU.1 and miR-22 promoter became stronger with PMA induction. The specificity of PU.1 occupancy in PB2 was demonstrated by the absence of immunoprecipitated chromatin fragments corresponding to an unrelated genomic region (UR) and the presence of immunoprecipitated chromatin fragments corresponding to a known PU.1-binding site within the CSF1R promoter (pro-CSF1R) [31] (Fig 1G). We then cloned the genomic fragment surrounding PB2 (PB2_WT) or a mutant version (PB2_Mut), into a promoterless luciferase reporter plasmid pGL3-basic. The recombinant plasmid expressing PU.1 pcDNA3.1-PU.1 (or pcDNA3.1 control), together with the luciferase reporter plasmid pGL3-PB2_WT (or pGL3-PB2_Mut), were co-transfected into 293T cells. Notably, the PB2_WT fragment yielded robust PU.1-dependent transcriptional activity, while the mutation eliminated the transcriptional activity (Fig 1H).
These findings provide compelling evidence that miR-22 is transcriptionally regulated by PU.1.
To examine the effect of miR-22 on monocyte/macrophage differentiation, we firstly transfected THP1 and HL60 cells with miR-22 mimics or anti-miR-22, and then induced monocyte/macrophage differentiation. The overexpression or inhibition of miR-22 was confirmed by qRT-PCR (S2 Fig). Remarkably, the flow cytometry data revealed a higher percentage of CD14-positive cells in the miR-22 mimics-transfected cells, while a lower percentage was observed in the anti-miR-22-transfected cells (Fig 2A). qRT-PCR analysis also demonstrated that the ectopic expression of miR-22 increased while the inhibition of miR-22 expression impaired CD14 and CSF1R mRNA levels in the cells (Fig 2B). In addition, May-Grünwald Giemsa staining demonstrated that the overpresence of miR-22 promoted monocyte/macrophage cell development in the PMA-induced cells (Fig 2C).
Next, we infected HSPCs with a recombinant lentivirus harbouring a miR-22 precursor (Lenti-miR-22) or expressing hairpins that exhibited anti-miR-22 activity (Lenti-ZIP-miR-22) and a relative control (Lenti-Con or Lenti-ZIP-Con), and induced monocyte/macrophage differentiation. As shown in Fig 3A, miR-22 overexpression increased while miR-22 knockdown decreased CD14 mRNA levels. Flow cytometry analysis revealed that miR-22 overexpression increased (Fig 3B, left) while knockdown of miR-22 decreased the percentage of CD14-positive cells (Fig 3B, right). Furthermore, Lenti-miR-22 infection significantly promoted the colony-forming activity of CFU-M and CFU-GM, both in clone size and number (Fig 3C). May-Grünwald Giemsa staining confirmed that Lenti-miR-22 infection led to higher percentage of more mature monocyte/macrophage cells (Fig 3D).
Altogether, these data indicated that miR-22 is a positive regulator in monocyte/macrophage differentiation.
To identify miR-22 targets contributing to the phenotypes observed, we looked over the potential targets predicated by TargetScan and noticed MECOM, a transcription factor and oncoprotein, which was documented to be involved in hematopoietic stem cell function [32] and blood generation [33] (Fig 4A). To determine whether MECOM is directly regulated by miR-22, the 3’UTR of MECOM was inserted downstream of the luciferase ORF of pMIR-REPORT. Significant repression of luciferase activity caused by miR-22 transfection was observed and the repression effect was abrogated by mutagenesis of the core miR-22 binding site in the 3’UTR of MECOM (Fig 4B). Furthermore, enforced miR-22 expression reduced the MECOM protein level, while inhibition of miR-22 expression elevated the MECOM protein level in THP1 and HL60 cells (Fig 4C). These data demonstrated that MECOM is a target of miR-22 in the AML cell lines.
We then examined the function of MECOM in monocyte/macrophage differentiation of THP1 cells. MECOM knockdown by Lenti-shMECOM infection in THP1 cells significantly increased the percentage of CD14-positive cells (Fig 4D), and also mRNA levels of the differentiation markers CD11b, CD14 and CSF1R (Fig 4E). In addition, May-Grünwald Giemsa staining showed that MECOM knockdown promoted monocyte/macrophage development (Fig 4F).
To confirm whether the miR-22 regulation of monocyte/macrophage differentiation occurred via its regulation on MECOM, we performed rescue assays. As shown in Fig 4G, the increase in MECOM expression (left) was accompanied by decreased CD14-positive cells (middle and right) after anti-miR-22 transfection (b vs. a). As expected, retransfection with siMECOM reduced the increase of MECOM expression resulting from anti-miR-22 treatment (left, c vs. b), which was accompanied with restoration of the percentage of CD14-positive cells (middle and right, c vs. b).
Collectively, these results demonstrated that the enhancement of monocyte/macrophage differentiation induced by miR-22 occurred at least partially via its negative regulation on MECOM.
MECOM was reported to impair myeloid differentiation via blocking the association of PU.1 with c-Jun [26], a critical coactivator of PU.1 transactivation. A recent report demonstrated antagonism between PU.1 and GATA2 in the transcriptional regulation of some genes [34]. GATA2 was also reported to inhibit the binding of PU.1 to c-Jun [35]. Interestingly, MECOM was reported to directly target the GATA2 promoter to promote its transcription [25]. In addition, MECOM can interfere with interaction between JNK (c-Jun N-terminal kinases) and c-Jun, thus reducing the level of phosphorylated c-Jun (p-c-Jun) [36], which is capable of inducing c-Jun expression via the formation of the heterodimer AP-1 with c-Fos [37]. Based on this evidence, we examined whether miR-22 affects the interaction between c-Jun and PU.1 through regulating MECOM in HL60 and THP1 cells. As shown in Fig 5A, the levels of MECOM and GATA2 decreased, whereas the levels of p-c-Jun, c-Jun and PU.1 increased following the PMA-induced monocyte/macrophage differentiation of THP1 and HL60 cells. We assessed the effects of miR-22 on the expression of these factors. In THP1 cells, ectopic expression of miR-22 reduced MECOM and GATA2 levels and increased p-c-Jun and c-Jun levels; but it barely affected PU.1 expression. In contrast, anti-miR-22 transfection led to an increase of MECOM and GATA2 levels, and a concomitant reduction of p-c-Jun and c-Jun levels, but little change was observed in PU.1 expression (Fig 5B). Similar results were obtained from the transfected HL-60 cells (S3 Fig). Furthermore, MECOM knockdown by Lenti-shMECOM infection showed similar results as miR-22 overexpression (Fig 5C) in THP1 cells. We next performed co-immunoprecipitation analysis on the infected THP1 cells. As shown in Fig 5D, miR-22 overexpression increased c-Jun levels but barely affected PU.1 expression (lane 1 and 2). Normal mouse IgG was not able to immunoprecipitate PU.1 and c-Jun (lane 3 and 4). However, anti-PU.1 antibody immunoprecipitated more endogenous c-Jun in the cells infected with Lenti-miR-22 than those infected with Lenti-Con (IP-PU.1, up, lane 5 and 6), while the immunoprecipitated endogenous PU.1 level was almost the same in the two groups (IP-PU.1, down, lane 5 and 6). Moreover, the anti-c-Jun antibody immunoprecipitated more endogenous c-Jun and PU.1 in the cells infected with Lenti-miR-22 than in the Lenti-Con-infected cells (IP-c-Jun, lane 5 and 6). These results demonstrate that miR-22 increases interaction between PU.1 and c-Jun.
To further determine if the mechanism by which miR-22 regulates monocyte/macrophage differentiation revealed in THP1 and HL-60 cells also exists in normal hematopoiesis, we analyzed the expression of miR-22 and its target protein MECOM. A gradual increase in miR-22 levels whereas a decrease in MECOM mRNA and protein levels were detected during the monocyte/macrophage induction culture of CD34+ HSPCs derived from human HCB (Fig 6A). Western blotting revealed a decrease in MECOM and GATA2, and an increase in c-Jun levels in the induction culture of the Lenti-miR-22-infected HSPCs (Fig 6B, left). Conversely, Lenti-ZIP-miR-22 infection caused increased MECOM and GATA2 levels and decreased c-Jun levels (Fig 6B, right). We also examined the effect of MECOM on monocyte/macrophage differentiation of HSPCs. Flow cytometry demonstrated that knockdown of MECOM by Lenti-shMECOM in the HSPCs increased percentages of CD14-positive cells (Fig 6C). Western blot analysis revealed a reduced GATA2 levels but increased c-Jun levels (Fig 6D) in the induction culture of the Lenti-shMECOM-infected HSPCs. These results confirmed that the mechanism by which miR-22 promotes monocyte/macrophage differentiation of HSPCs is identical to that in the cell lines.
We performed Taqman real-time PCR to detect MECOM mRNA expression [40] in PBMNCs derived from 40 AML patients and 43 healthy donors. Significantly higher MECOM mRNA levels were detected in the AML patients compared to the healthy donors (Fig 6E, left), while miR-22 levels were much lower in the same AML samples compared to the healthy donors (Fig 6E, middle). Moreover, miR-22 expression was conversely associated with MECOM mRNA expression in the tested projects (Fig 6E, right).
As PU.1 has been proved to regulate miR-22 expression, we questioned whether the down-regulation of miR-22 was related to PU.1 expression in AML. We examined their expression in PBMNCs derived from 42 AML patients and 39 healthy donors and found decreased PU.1 levels in AML patients (Fig 6F, left). Moreover, the PU.1 levels were positively associated with miR-22 levels in the tested projects (Fig 6F, right).
Collectively, these results at least partially confirm PU.1-miR-22-MECOM regulation in AML development.
Since a remarkable decrease of miR-22 was observed in AML patients, and since myeloid differentiation blockage is one of the key characterizations in AML, we examined whether reintroduction of miR-22 could relieve the differentiation blockage. The BM CD34+ HSPC samples derived from seven AML patients were infected with Lenti-miR-22 or Lenti-Con and subjected to monocyte/macrophage induction. Flow cytometry demonstrated that Lenti-miR-22 infection significantly improved the differentiation of HSPCs from all seven patients (Fig 7A and S4A Fig). May-Grünwald Giemsa staining also showed that Lenti-miR-22 infection improved monocyte/macrophage development of the AML HSPCs (Fig 7B and S4B Fig). Analysis with qRT-PCR confirmed miR-22 overpresence in the Lenti-miR-22-infected cells (Fig 7C and S4C Fig). Western blot analysis displayed significantly decreased MECOM and GATA2 levels and increased c-Jun levels in the induction cultures of Lenti-miR-22-infected-AML HSPCs as compared with the control infection samples (Fig 7D). These results demonstrated that the reintroduction of miR-22 could partially relieve differentiation blockage in AML BM blasts.
We also examined the effects of miR-22 on the growth of HL60 and THP1 cells, and found that miR-22 significantly inhibited cell growth (S5 Fig). Following this observation, we further demonstrated that lentivirus-mediated miR-22 reintroduction inhibited cell growth during the monocyte/macrophage induction culture of AML BM CD34+ HSPCs (Fig 7E).
PU.1 has been shown to play a decisive role in lympho-myeloid development and its stage-specific expression is critical to prevent leukemic transformation [39,40]. Other studies have revealed that monocyte/macrophage development from hematopoietic stem cells requires PU.1-coordinated miRNA expression [41,42]. It was also reported that miR-22 was transcriptionally regulated by P53 and c-Myc [43–45]. However, the TSS of miR-22 has not been confirmed. In this paper, we identified the TSS and showed that PU.1 activates miR-22 transcription by directly binding to the miR-22 promoter.
MiR-22 has been reported to play an important role in several physiologic processes and cancers, and several target genes of miR-22 have been identified in different cell types [16–23,30,46,50,51]. Here, we demonstrated that miR-22 is a positive regulator and MECOM a negative regulator in monocyte/macrophage development. We also showed that miR-22 promotes the differentiation via targeting and downregulating MECOM mRNA, at least partially.
MECOM was reported to impair the function of PU.1 by competing with c-Jun, a critical coactivator of PU.1 [47] Similarly, GATA2, which can be transcriptionally activated by MECOM [27], is able to interfere with the interaction between c-Jun and PU.1 [37]. In addition, other studies have revealed that MECOM blocks JNK-dependent phosphorylation of c-Jun [38], thus reducing p-c-Jun levels, which can form heterogeneous or homogeneous AP-1 to activate c-Jun transcription [39]. In this study, we found that a decrease in miR-22-mediated MECOM resulted in increased c-Jun-PU.1 protein complexes via increasing c-Jun levels and by relieving MECOM- and GATA2-mediated interference in the interaction between c-Jun and PU.1, which promotes monocyte/macrophage differentiation.
In the present study, we also detected abnormally decreased expression of miR-22 in de novo AML patients, suggesting that it acts as a tumor suppressor in AML development. Additionally, we detected a negative association between MECOM mRNA and miR-22 expression and a positive association between miR-22 and PU.1 expression in AML patients, which suggests that PU.1-miR-22-MECOM regulation is involved in AML development.
According to the above results, we summarized molecular models underlying miR-22’s involvement in monocyte/macrophage differentiation regulation (Fig 8A) and AML development (Fig 8B).
Until now, there have been two published leukemia/miR-22-related reports with opposite conclusions. Song et al. reported that miR-22 is an oncogenic miRNA and is abnormally upregulated in myelodysplastic syndrome (MDS) and MDS—derived leukemia [50]; they also showed that miR-22 transgenic mice developed MDS and hematological malignancies [50]. Jiang et al. reported that miR-22 plays an anti-tumor role and is abnormally downregulated in de novo AML [51], which is consistent with our results. Mechanistically, Song et al. reported that miR-22 regulated methylation status via targeting TET2 mRNA [50], while Jiang et al. reported that TET1 could repress miR-22 transcription, and that miR-22 targets multiple oncogenes, including CRTC1, FLT3 and MYCBP, and thus repressing the CREB and MYC pathways [51]. Our present paper demonstrates that miR-22 is transcriptionally activated by PU.1, and can enhance PU.1–c-Jun interaction by targeting MECOM and thus affecting GATA2 and c-Jun levels. These findings illustrate how miR-22 and the transcription factors MECOM, GATA2, c-Jun, and PU.1 are orchestrated in normal monocyte/macrophage differentiation regulation and AML development.
Using oncogenes-transformed mouse models, Jiang et al. demonstrated miR-22’s therapeutic potential in AML. Using BM CD34+ cells obtained from AML patients, our present paper shows that the reintroduction of miR-22 could relieve the differentiation blockage and inhibit the growth of AML BM blasts, which also suggests its potential in AML therapy.
High MECOM expression defines a subgroup of AML with a poor prognosis [38,48,49]. We found that the reintroduction of miR-22 significantly improved monocyte/macrophage differentiation in the patients with either high or low MECOM expression. Interestingly, it seems that Lenti-miR-22 infection improved differentiation better in MECOMhigh patients than in MECOMlow patients (18.63 ± 5.41% vs. 9.13 ± 2.39%, p = 0.002, see S6 Fig); however this finding needs further demonstration.
In conclusion, our data revealed new function and mechanism of miR-22 in human monocyte/macrophage differentiation and AML development, and demonstrated its potential value in AML diagnosis and therapy.
Human UCB was obtained from normal, full-term deliveries from Beijing Hospital. The PB and BM samples of AML patients and normal volunteers were obtained from the 303 hospital and the 307 Hospital according to the protocols approved by the Ethics Committees of the Institutional review Board of Institute of Basic Medical Sciences, Chinese Academy of Medical Sciences. The informed consent was obtained from all of the examined subjects.
The human promyelocytic cell line HL60 was maintained in IMDM (Gibco-BRL, Paisley, UK) containing 2 mM glutamine, 25 mM HEPES, 1.5g/L sodium bicarbonate, 50 U/mL penicillin and 50 μg/mL streptomycin (Sigma, St. Louis, MO, USA), supplemented with 10% FCS (PAA, Pashing, Austria), at 37°C in 5% CO2. Acute monocytic leukemia cell line THP-1 was maintained in RPMI-1640 medium (Gibco-BRL) containing 2 mM glutamine, 25 mM HEPES, 1.5 g/L sodium bicarbonate, 50 U/mL penicillin, and 50 μg/mL streptomycin (Sigma), supplemented with 10% fetal bovine serum (FBS), at 37°C in 5% CO2. Lentivirus packaging cell line 293TN was cultured in DMEM medium (Gibco-BRL), supplemented with 10% FBS. For monocytic/macrophagic induction, PMA (Sigma) was added to a final concentration of 16 nM.
Total RNA was isolated from the cell harvest using Trizol (Invitrogen, CA, USA) according to the manufacturer’s instructions. One μg of total RNA was used to generate cDNA by M-MLV reverse transcriptase (Invitrogen). Stem-poop RT primers were used for the reverse transcription of miRNAs, and Oligo(dT)18 was used for the reverse transcription of mRNAs. qRT-PCR was carried out in the Bio-Rad IQ5 real-time PCR system (Bio-rad, CA, USA) or in the ABI PRISM 7900HT Sequence Detection System (Applied Biosystem, CA, USA) according to the manufacturer’s instructions. Each qRT-PCR assay was performed in triplicate. The data were normalized using the endogenous GAPDH mRNA or U6 snRNA. The primers for reverse transcription of miRNAs and qRT-PCR as well as the Taqman probes are described in S2 Table.
Human CD34+ cells that contain HSPCs were collected using a human CD34 MicroBead Kit (Miltenyi Biotec, Cologne, Germany) from MNCs isolated from UCB, PB or BM by percoll density gradient (d = 1.077) (Amersham Biotech, Little Chalfont, UK). The CD34+ cells were cultured in IMDM (Gibco-BRL, Paisley, UK) with 30% FBS, 1% bovine serum albumin, 2 mM L-glutamine, 0.05 mM 2-mercaptoethanol, 50 U/ml penicillin, 50 μg/ml streptomycin, 50 ng/ml stem cell factor and 20 ng/ml IL-3. To induce monocyte/macrophage differentiation, a cytokine cocktail of 50 ng/ml M-SCF, 1 ng/ml IL-6 and 100 ng/ml Flt-3 L was used. All of these cytokines were purchased from Peprotech (Rocky Hill, NJ, USA).
The total RNA isolated from THP1 cells treated with PMA for 72 hours was used and RACE was performed using a 5’-Full RACE kit with TAP (Takara, Dalian, China). Primer sequences are listed in S2 Table.
Twenty μg of denatured total RNA was loaded onto a 15% polyacrylamide TBE gel and separated in a 1 X TBE running buffer, followed by transfer onto a N+ membrane (Amersham, London, UK) at 200 mA for two hours in an electro-transferring system and crosslinking under ultraviolet radiation for 150 seconds. The miRNA-specific oligo was 5’ end labelled with γ-32P-ATP through T4 polynucleotide kinase (Takara), according to the manufacturer’s protocol. The oligo probes were designed based on individual miRNA sequence information deposited in miRBase (http://microrna.sanger.ac.uk). An antisense oligo of U6 snRNA was used to detect U6 snRNA from each sample as a loading control. After prehybridisation using hybridizing buffer (BioDev, BJ, China), blots were hybridized with 32P-labelled DNA probes (2 μmol/ml) overnight at 37°C. After washing, the hybridized membranes were exposed to Kodak X-omat BT film.
miR-22 mimics, anti-miR-22 (miR-22 inhibitor), small interference RNAs (siPU.1) and negative controls (scramble control, inhibitor control and siRNA control) were purchased from Dharmacon (IL, USA). Small interference RNAs (siMECOM) were purchased from Origene (MD, USA). These oligonucleotides were transfected into HL-60 and THP1 cells using a DharmaFECT1 reagent (Dharmacon) at a final concentration of 100 nM.
Total proteins were extracted from cells or tissues using a RIPA buffer (50 mM Tris-HCl, pH 7.4, 150 mM NaCl, 1 mM EDTA, 1% Triton X-100, 1% sodium deoxycholate, 0.1% SDS) supplemented with 1 mM PMSF, 5 μg/ml aprotinin and 5 μg/ml leupeptin. The protein concentration was determined with a BCA Protein Assay Kit (Vigorous, China). The total protein (15–30 μg) was loaded onto a 10% SDS-PAGE gel, probed with mouse or rabbit mAb against MECOM (Epitomics, MA, USA), GATA2 (Proteintech, IL, USA), p-c-Jun (Bioworld Technology, St. Louis, USA), c-Jun (Bioworld Technology), PU.1 (Cell Signaling Technology) and GAPDH (Proteintech) followed by horseradish peroxidase-conjugated sheep anti-mouse or rabbit Ig (ZSGB-BIO). GAPDH was detected as a loading control.
The harvested cells were washed twice with PBS, resuspended in 100 μl cold PBS and stained with PE- or APC-conjugated anti-CD14 (eBioscience, San Diego, CA, USA) for 30 minutes at 4°C in the dark. Finally, Stained cells were washed using cold PBS and analyzed on an Accuri C6 flow cytometer (Becton Dickinson Biosciences, San Jose, CA, USA). Usually, in an identical experiment group, the fluorescence gate was set based on untreated cells which endured neither DNA delivery treatment nor induction, and at the place where the percentage of differentiated cells in untreated cells was less than 1% and fixed. The unstained cells, which endured the same treatment except for antibody staining, were used for controls to exclude the effect of the background fluorescence of the cells caused by treatment.
Cells were re-seeded into 96-well plates at a density of 10000 cells per well after transfection or infection for 12 hours. Cell viability was measured every 24 hours for HL60 and THP1 cells or every three days for AML BM CD34+ cells by adding 10% CCK-8 (DOJINDO, Japan) and then incubating at 37°C for three hours. The optical density was read at 450 nm with a microplate spectrophotometer. Each experiment was carried out three times.
The 3′UTR of MECOM containing the predicted binding site of miR-22 was amplified from human cDNA by PCR using the primers described in S2 Table, and then inserted into the pMIR-REPORT luciferase reporter vector (MECOM_WT). MECOM_Mut contained the sequences with mutations in the binding site. Fragments containing wild-type PB2 (TTCCTC) were amplified from genomic DNA by PCR using the primers described in S2 Table and then cloned into pGL3-basic to get pGL3-PB2_WT, while those containing mutant PB2 (CATGAG) were also cloned into pGL3-basic to get pGL3-PB2_Mut. The recombinant construct (pGL3-PB2_WT or pGL3-PB2_Mut), pRL-TK and miR-22 mimic or pcDNA3.1-PU.1 and related controls were co-transfected into HEK293T cells using Lipofectamine 2000 (Invitrogen, CA, USA). The plasmid pRL-TK containing Renilla luciferase was used as an internal control. The cells were harvested after transfection for 48 hours and the luciferase activity was measured using a Dual Luciferase Assay System (Promega, WI, USA) according to the manufacturer’s instructions. Data were obtained by normalization of firefly luciferase activity to renilla luciferase activity. All transfection assays were performed three times.
CD34+ HSPCs infected with Lenti-miR-22 or Lenti-Con were cultured in a 6-well plate with human methylcellulose medium without EPO (R&D, SystemsGmbH, MN, USA) according to the manufacturer’s instructions. After 12 days of incubation at 37°C in a 5% CO2 incubator, colony-forming unit granulocyte/macrophages (CFU-GM) and colony-forming unit-macrophages (CFU-M) were analyzed and quantified using Eclipse TS100 phase-contrast microscopy (Nikon, Tokyo, Japan).
The HL60, THP1 or CD34+ HSPCs induced monocyte/macrophage differentiation were harvested at the indicated time and stained with May-Grünwald for 5 min and Giemsa for 30 minutes. Then, the cell smears were washed with water, air-dried, and observed under Olympus BX51 optical microscopy (Olympus, Tokyo, Japan).
Cells (2 x 107) were treated with 1% formaldehyde in a medium for 10 minutes at 37°C, followed by addition of glycine (final concentration, 0.125 M). After washing with PBS, cells were lysed on ice and sonicated to obtain 500–1000-bp sheared chromatin fragments. Subsequent ChIP steps were performed according to the protocols from Upstate Biotechnology (Charlottesville, VA, USA). Each reaction included 2 μg anti-PU.1 (Cell Signaling Technology, Beverly, MA, USA); anti-IgG (Santa Cruz Biotechnology, Santa Cruz, CA, USA) served as the unspecific control. The presence of target DNA sequences was detected by PCR and qRT-PCR. PCR products were resolved by 2% agarose gel electrophoresis. qRT-PCR analysis of fragments containing validated PU.1-binding site, the positive control (pro-CSF1R) and the negative control (UR) were carried out three times with the primers listed in S2 Table. The relative occupancy of the immunoprecipitated factor at a locus is examined via the comparative threshold method [52]. For every promoter studied, a ΔCt value was calculated for each sample by subtracting the Ct value for the input (to account for differences in amplification efficiencies and DNA quantities before immunoprecipitation) from the Ct value obtained for the immunoprecipitated sample. A ΔΔCt value was then calculated by subtracting the ΔCt value for the sample immunoprecipitated with PU.1 antiserum from the ΔCt value for the corresponding control sample immunoprecipitated with normal rabbit serum. Fold differences (PU.1 ChIP relative to control ChIP) were then determined by raising 2 to the ΔΔCt power. The equation used in these calculations is summarized as fold difference (PU.1 ChIP relative to control ChIP) = 2[Ct(control)—Ct(PU.1)], where Ct = Ct (immunoprecipitated sample)–Ct (input).
THP1 cells were infected with a lentivirus overexpressing miR-22 or GFP control, then treated with PMA for 48 hours for co-immunoprecipitation. The Dynabeads Protein G (Invitrogen) was incubated with anti-PU.1 antibody (Santa Cruz Biotechnology) or anti-c-Jun antibody (Santa Cruz Biotechnology) or IgG (Santa Cruz Biotechnology) in antibody binding and washing buffer at room temperature with a 20-minutes rotation. The Dynabeads-antibody complexes were washed one time using antibody binding and washing buffer then incubated with the whole cell lysates at 4°C overnight. For Western blot analysis, the Dynabeads-antibody-antigen complexes were washed four times with washing buffer, and the proteins were separated by SDS-PAGE.
For construction of the recombinant lentiviruses that expresses specific shRNAs against MECOM or PU.1, the targeted sequences (see S2 Table) were synthesized and inserted into the pLentiLox 3.7-RNAi plasmid (Invitrogen) following the manufacturer’s protocols. For construction of the recombinant lentivirus that expresses miR-22, a 300-bp DNA fragment containing the miR-22 precursor was amplified and inserted into pMiRNA1 vector. The miRZip lentivector construct expressing miRZip shRNAs targeting miR-22 (Lenti-ZIP-miR-22) was purchased from SBI (Mountain View, CA, USA). The virus packaging was performed using a packaging kit from SBI (Mountain View) according to the manufacturer’s instructions. The lentivirus particles (Lenti-miR-22, Lenti-Con, Lenti-ZIP-miR-22, Lenti-ZIP-Con, shMECOM, shPU.1, shCon) were harvested and concentrated using PEG-it Virus Precipitation Solution (SBI). The lentiviral particles were added into the THP1 cells or CD34+ cells in the presence of Polybrene (5 μg/mL; Sigma, St. Louis, MO, USA). The cells were washed with PBS 24 hours after infection and exposed to lineage-specific differentiation cultures or plated for colony-forming assay.
A Student’s t-test (two-tailed) was performed to analyze the data. The correlation between miR-22 and MECOM mRNA as well as between miR-22 and PU.1 mRNA was examined by Pearson correlation analysis. P-values <0.05 were considered to be significant.
|
10.1371/journal.ppat.1000843 | The RING-CH Ligase K5 Antagonizes Restriction of KSHV and HIV-1 Particle Release by Mediating Ubiquitin-Dependent Endosomal Degradation of Tetherin | Tetherin (CD317/BST2) is an interferon-induced membrane protein that inhibits the release of diverse enveloped viral particles. Several mammalian viruses have evolved countermeasures that inactivate tetherin, with the prototype being the HIV-1 Vpu protein. Here we show that the human herpesvirus Kaposi's sarcoma-associated herpesvirus (KSHV) is sensitive to tetherin restriction and its activity is counteracted by the KSHV encoded RING-CH E3 ubiquitin ligase K5. Tetherin expression in KSHV-infected cells inhibits viral particle release, as does depletion of K5 protein using RNA interference. K5 induces a species-specific downregulation of human tetherin from the cell surface followed by its endosomal degradation. We show that K5 targets a single lysine (K18) in the cytoplasmic tail of tetherin for ubiquitination, leading to relocalization of tetherin to CD63-positive endosomal compartments. Tetherin degradation is dependent on ESCRT-mediated endosomal sorting, but does not require a tyrosine-based sorting signal in the tetherin cytoplasmic tail. Importantly, we also show that the ability of K5 to substitute for Vpu in HIV-1 release is entirely dependent on K18 and the RING-CH domain of K5. By contrast, while Vpu induces ubiquitination of tetherin cytoplasmic tail lysine residues, mutation of these positions has no effect on its antagonism of tetherin function, and residual tetherin is associated with the trans-Golgi network (TGN) in Vpu-expressing cells. Taken together our results demonstrate that K5 is a mechanistically distinct viral countermeasure to tetherin-mediated restriction, and that herpesvirus particle release is sensitive to this mode of antiviral inhibition.
| To replicate efficiently in their hosts, viruses must avoid antiviral cellular defenses that comprise part of the innate immune system. Tetherin, an antiviral membrane protein that inhibits the release of several enveloped viruses from infected cells, is antagonized by the HIV-1 Vpu protein. The K5 protein of the human pathogen Kaposi's sarcoma-associated herpesvirus (KSHV) modulates the cell surface levels of several host proteins including tetherin. We show that KSHV release is sensitive to tetherin, and that K5 expression is required for efficient virus production in tetherin-expressing cells. K5 is also capable of rescuing Vpu-defective HIV-1 virus release from tetherin. K5 expression induces a down-regulation of cell-surface tetherin levels and degradation in late endosomes, which depends on a single lysine residue in the tetherin cytoplasmic tail. Finally, we show that the ESCRT pathway, which promotes the trafficking of cell surface receptors for degradation, is required for K5-mediated tetherin removal from the plasma membrane. Thus, we demonstrate that herpesviruses are sensitive to the antiviral effects of tetherin and that KSHV has evolved a mechanism for its destruction. These findings extend the list of viruses sensitive to tetherin, suggesting that tetherin counter-measures are widespread defense mechanisms amongst enveloped viruses.
| The inhibitory effect of type 1 interferons (type 1 IFN) on the replication of mammalian viruses has been documented for over 50 years. However the effecter mechanisms that interfere with virus replication have not been well characterized. While many IFN response genes are known, few definitive antiviral functions have been ascribed to them. Amongst the best characterized are PKR/2′5′oligoadenylate synthetase, MxA and ISG15, all of which have broad activity against a variety of mammalian RNA viruses [1]. Recently the identification of retroviral restriction factors including members of the APOBEC3 family of cytidine deaminases, as well as TRIM5 and other members of the tripartite motif protein family, has highlighted innate intracellular defense mechanisms as key determinants of tropism for human and primate immunodeficiency viruses [2], [3]. Moreover, these antiviral activities have driven the acquisition of viral countermeasures [2], [4] and thus interferon-inducible restriction factors are now thought to represent an important arm of the antiviral innate immune system [3].
Tetherin, (BST2/CD317) has recently been shown to inhibit the release of HIV-1 particles that are defective for the accessory protein Vpu [5], [6]. In the absence of Vpu expression, nascent HIV-1 particles assemble at the plasma membrane but remain tethered to the surface of tetherin expressing cells via a protease-sensitive linkage. Tethered virions are then endocytosed leading to their accumulation in late endosomes [5], [7], [8]. Tetherin colocalization with restricted viral particles on cell surfaces and in endosomes is well documented [5], [6], [9]. Strikingly, it is tetherin's unusual topology that is thought to be directly responsible for its mode of action [10]. Tetherin is a dimeric type-II membrane protein consisting of an N-terminal cytoplasmic tail, an extracellular domain with a putative coiled coil, and a C-terminal GPI anchor which is required for its antiviral function [5], [11]. It forms dimers which are thought to cross-link viral and cellular membranes during viral budding [10]. Tetherin appears to have no direct association with any viral structural proteins and is therefore able to restrict a range of unrelated viruses including retroviruses, filoviruses and arenaviruses [9], [12], [13]. It is expressed on mature B cells and plasmacytoid dendritic cells, but can be induced in many cell types by type-I interferon (IFN) [8], [14], [15], [16], [17]. Sequence analysis of orthologues of tetherin from primates indicates high levels of positive selection during their evolution suggesting selective pressure from pathogenic viral infections [18], [19]. Together, these observations suggest that tetherin may be an important antiviral defense against enveloped virus replication, necessitating the acquisition of viral countermeasures to antagonize its activity.
Interestingly, other potential viral countermeasures to tetherin may exist. Kaposi's sarcoma-associated herpesvirus (KSHV), also known as Human herpesvirus 8 (HHV8), encodes two immuno-modulatory membrane associated RING-CH (MARCH) E3 ubiquitin ligases named K3 and K5 [20]. K5 has been shown to mediate the down-regulation of a variety of cell surface markers including MHC class I, PE-CAM-1, CD80/CD86, ICAM-1, IFNγ receptor and NKG2D [21], [22], [23], [24], [25], [26], [27]. In a recent proteomics screen Bartee and colleagues used a methodology called stable isotope labeling of amino acids in cultured cells (SILAC) to identify proteins that are removed from the plasma membrane on K5 expression. One of the novel K5 targets found was tetherin (BST2) [28]. Here we have tested the hypothesis that tetherin restricts assembly/release of KSHV particles and that an important function of K5 is to overcome this process.
Tetherin has the capacity to restrict the release of diverse enveloped viruses including filoviruses and arenaviruses [9], [12], [13]. Given the capacity of the KSHV protein K5 to reduce cell surface expression of tetherin [28] we tested whether KSHV could be restricted by the expression of increasing amounts of exogenous human tetherin. We reasoned that the complex envelopment strategy of herpesviruses [29] and the tropism of KSHV for tetherin-positive mature B cells made this virus potentially sensitive to restriction by tetherin. To test this we generated a HeLa cell line harboring latent recombinant rKSHV.219 episomes under puromycin selection which encode eGFP driven by the human EF-1α promoter and DsRed driven by the KSHV PAN promoter that responds to the KSHV immediate early protein RTA [30]. Transfection of these cells with an RTA encoding plasmid (pCMV RTA) induces KSHV transcription, lytic replication and the release of KSHV particles, the efficiency of which can be assessed by measurement of RFP expression. The amount of virus released 48h post-RTA induction can be measured by quantitative PCR (Q-PCR) detecting DNAse-I protected genomes in the supernatant or by titration of infectious virus onto 293T cells and enumeration of GFP expressing cells by flow cytometry.
We first determined whether transfecting a plasmid expressing tetherin together with pCMV-RTA would impact on the amount of infectious virions released into the supernatant. Figure 1A clearly demonstrates a linear decrease in release of KSHV infectious virus when increasing amounts of the tetherin-expressing plasmid are transfected with a constant amount of pCMV-RTA. Similarly, Q-PCR performed on DNAse-I resistant genomes showed a greater than 10-fold decrease in total virions released for the highest dose of tetherin-expressing plasmid used (Figure 1B). To control for transfection efficiency, cells were recovered after virus collection and cell lysates subjected to western blotting (Figure 1C). As expected, increasing amounts of tetherin were detected as increasing amounts of tetherin-expressing plasmid was transfected. Importantly, equal RTA expression was found in all samples confirming that RTA levels were not impacted by tetherin expression. To further rule out an effect of tetherin expression on KSHV reactivation we also measured lytic viral RNA production by quantitative RT-PCR. Ct values for ORF37 were normalized to those obtained for cellular GAPDH. Figure 1D shows very similar ORF37/GAPDH ratios for each condition. Whilst unlikely, it is also possible that exogenous expression of increasing amounts of tetherin might lower genomic replication, thus leading to a reduced amount of virions and/or DNase-I resistant genomes in the supernatant. To rule this out we measured intracellular episomes in DNA extracts by Q-PCR. Figure 1E clearly demonstrates that episome copy number was similar across all conditions and could not account for reduced viral production.
We then tested whether K5 expression was required for efficient KSHV release from r219-HeLa cells. We reasoned that if K5 is a KSHV antagonist for tetherin then reducing its expression by RNA interference (RNAi) should inhibit KSHV release. Three shRNA hairpins were designed and expressed in 293T cells together with a HA-tagged K5 protein. Reduction of K5 expression by the hairpins was assessed by recovering the cells 48 hours after transfection and subjecting the lysates to western blot (Figure 2A). Blots probed with the anti-HA antibody were stripped and re-probed for alpha tubulin to control for equal loading (Figure 2A). All three hairpins reduced K5 expression with hairpin 3 (sh-K5iii) being the most potent. We then expressed the hairpins in r219-HeLa cells using lentiviral vectors [31]. 72 hours later the cells were re-seeded and transfected with RTA to induce KSHV lytic replication. Release of infectious KSHV was measured by titration of supernatants on 293T as before (Figure 2B). We found that expression of specific anti-K5 shRNA reduced KSHV titer in the supernatant whereas expression of an empty shRNA vector did not (Figure 2B). The number of DNAse-I resistant KSHV genomes in the supernatant was also reduced by all three shRNA vectors, as shown by taqman Q-PCR (Figure 2C). Messenger RNA for the late gene ORF37 was measured by quantitative RT-PCR in each sample and values were corrected for cellular GAPDH mRNA levels. This demonstrated that hairpin expression did not significantly inhibit KSHV reactivation (Figure 2D). In fact hairpin 2 appeared to stimulate ORF37 expression in one experiment leading to the large error bar. However, this experiment shows that inhibition of reactivation cannot account for the loss of KSHV in the supernatant. To further rule out an effect of hairpin expression on genomic replication we also measured intracellular KSHV episomes by DNA Q-PCR for ORF37 (Figure 2E) This control confirmed that the number of KSHV episomes in the cells did not account for the K5 hairpin induced defect in KSHV release and supports the notion that K5 antagonizes tetherin.
These results suggest that K5 is required for efficient KSHV particle release in a cell line that constitutively expresses tetherin and that expressing increasing amounts of exogenous tetherin further inhibits KSHV release. Together our data suggest that tetherin has antiviral activity against KSHV and that K5 has an important role in overcoming tetherin-mediated restriction.
Having established a role for K5 in counteracting tetherin during KSHV release we then tested whether K5 expression could functionally substitute for the HIV-1 encoded tetherin antagonist Vpu and rescue the release of tetherin restricted HIV-1(delVpu) particles [5]. We transfected HeLa cells with the HIV-1 molecular clone NL4.3, or a Vpu-defective counterpart, together with expression vectors for HA-tagged K5 or Vpu. As predicted, K5 expression potently rescued HIV-1(delVpu) release to levels achieved by Vpu expression (Figure 3A). This was evidenced both by measurement of HIV-1 p24 capsid protein released into the supernatant by western blots as well as by titration of infectious HIV-1 virus released into the supernatant on sensitive indicator cells (Figure 3A, B and C). Importantly K5 expression had no effect on wild-type HIV-1 particle release or on HIV-1 structural protein expression. These data demonstrate that K5 is a functional homolog of HIV-1 Vpu and that KSHV encodes a tetherin countermeasure.
The fact that K5 has been demonstrated to be a RING dependent E3 ligase for ubiquitin [20] suggested that the RING-CH domain is important for its function. We therefore made a RING deletion mutant of K5 (K5delRING) and tested its ability to counteract tetherin restriction in an HIV-1 release assay as above. Despite similar expression levels of the mutant and wild type K5 proteins (Figure 3B), the RING-defective K5 was unable to rescue HIV-1(del Vpu) release from HeLa cells (Figure 3A and C). To confirm that K5 can antagonize tetherin function we repeated the experiment but in a tetherin-deficient cell line (HT1080) [5] stably expressing a tetherin protein in which an HA-tag had been inserted into the ectodomain at amino acid position 154 [9]. HT/THN-HA cells were then transduced with retroviral vectors co-expressing DsRed and Vpu, or K5 or K5delRING at doses sufficient to give >90% DsRED-positive cells as demonstrated by flow cytometry 48 hours later. The cells were then re-seeded and infected with VSV-G pseudotyped HIV-1(wt) or HIV-1(del Vpu) at a multiplicity of infection of 0.2. Since HT1080 cells are devoid of CD4, VSV-G pseudotyping allows the measurement of a single round of viral replication. Similar to the experiment in HeLa cells, HT/THN-HA released HIV-1(del Vpu) particles approximately 20 fold less efficiently than they released wild type virus. Furthermore, expression of either K5 or Vpu, but not K5delRING, could rescue the tetherin mediated defect in virus release (Figure 3D and E). Measurement of gag levels in supernatants and extracts of infected cells by western blot indicated that the effects of tetherin, Vpu and K5 were on HIV-1 release and not gag protein expression. Intriguingly, K5 could not antagonize tetherin function in transiently-transfected 293T cells, even when tetherin expression was titrated to vary its expression level (Figure S1). We propose that 293T cells lack a co-factor essential for K5's antagonism of tetherin, but not for Vpu function. Interestingly, 293T cells are also unable to support HIV-2 Env's anti-tetherin activity [32], suggesting that this particular cell line might be poorly suited for characterization of some tetherin antagonists.
Vpu has been shown to remove tetherin from the cell surface after transfection [6] and in HIV-1-infected cells [32]. We therefore tested whether K5 expression had the same effect on surface tetherin levels in HT/THN-HA cells by expressing K5 via retroviral transduction. Both K5 (Figure 4A) and HIV-1 Vpu (Figure S2) expression led to a marked reduction of tetherin from the cell surface. Importantly, the K5delRING protein was inactive in this assay as predicted by data in Figure 3. HIV-1 Vpu-mediated tetherin antagonism displays distinct species specificity for primate tetherin genes with non-human primate orthologues being largely insensitive to HIV-1 Vpu [18], [19], [33]. Sensitivity to HIV-1 Vpu maps to the tetherin transmembrane domain and mutation of the T residue at position 45 to the I present in Rhesus macaques (T45I) coupled with an in frame deletion of a GI pair at the N-terminus of the human tetherin protein's TM (delGI-T45I) results in a human tetherin that is completely resistant to HIV-1 Vpu [18] (Figure S2). We therefore examined whether K5 also displays similar species–specific effects (Figure 4B). Unlike HIV-1 Vpu, K5 was able to down-regulate the THN(delGI-T45I) mutant human tetherin. However, K5 was unable to down-regulate the rhesus macaque tetherin protein. Thus K5 also displays species-specificity in its antagonism of primate tetherins but the determinants of this specificity are distinct from those of HIV-1 Vpu.
We then examined whether K5 could also down regulate human tetherin from the surface of a B cell line carrying KSHV (Figures 4C and D). Body-cavity-based lymphoma (BCBL) 1 cells were transduced with a lentiviral vector encoding K5 or with an empty vector and surface tetherin expression assessed by flow cytometry analysis, using a specific antibody against human tetherin. K5 expression led to a reduction of cell surface tetherin as compared to the levels on cells treated with the empty vector or levels detected on un-transduced cells (Figure 4C). The mean of mean fluorescence intensity (MFI) values for 3 representative experiments are also shown (Figure 4D).
We next examined the fate of tetherin in K5-expressing cells. HT/THN-HA cells were transduced to stably express K5 or Vpu and intracellular levels of tetherin were compared to levels in unmanipulated cells. Tetherin appears in western blots as a heterogeneous smear of glycosylated species that varies between cell types [10], [17]. After K5 and Vpu expression, tetherin levels in the modified cell lysates were decreased (to 4% of wildtype in the case of K5 and to 31% of wildtype for Vpu), suggesting that, like Vpu, K5 induces tetherin degradation (Figure 5A). At present it is unclear whether Vpu induces tetherin degradation via a proteasomal-dependent [19], [34], [35] or lysozomal pathway [36], [37]. This is further complicated by the fact that endolysozomal degradative pathways are often also dependent on ubiquitin, and thus proteasomal inhibition can inhibit them through depletion of free cytoplasmic ubiquitin levels. We addressed the K5 mechanism of action by examining the role of lysozomal degradation using the vacuolar ATPase inhibitor bafilomycin A1 or by inhibiting proteasomal degradation using MG132. We then measured steady state levels of tetherin in K5 or Vpu expressing cells. A 16h treatment with BafA1 or MG132 substantially rescued tetherin levels from both Vpu and K5 expression. BafA1 treatment rescued not only mature tetherin species, but also lower molecular weight fragments that are likely to be partially degraded tetherin molecules. This suggested that K5 degrades tetherin via an endolysozomal process, similar to that by which it degrades Class I MHC [38]. The rescue of tetherin degradation products in HT/THN-HA cells that do not express HIV-1 Vpu or K5 suggested that endosomal processing of tetherin contributes to its natural turnover. In contrast, whilst MG132 treatment also rescued partially tetherin levels in HT/THN-HA-K5 cells, mature species were predominant. In Vpu-expressing cells MG132 appeared more potent for tetherin rescue than BafA1, and again differential tetherin species were rescued by each inhibitor. Together, these data suggest that while tetherin degradation by K5 and Vpu are sensitive to both classes of inhibitor, the stages of degradation that are affected are different.
Next we determined whether we could observe differences in the cellular localization of tetherin induced by K5 or Vpu. In HT/THN-HA cells, tetherin localized predominantly to the plasma membrane with a small amount of the intracellular localizations coincident with the late endosomal marker CD63 (Figure 5B) This is consistent with the notion of natural tetherin turnover in endolysozomal compartments. In K5 expressing cells, while tetherin levels are markedly reduced, remaining tetherin was much more often found associated with CD63+ late endosomes than in controls (Figure 5B), implicating K5-induced endosomal degradation. In contrast, in Vpu-expressing cells, tetherin was never observed in CD63+ endosomes, but instead localized predominantly to compartments that stain positive for the Trans-Golgi marker TGN46 (Figure 5B and C). This localization is similar to that observed after expression of the HIV-2 and SIVtan envelope glycoproteins [32], [39]. Thus, the subcellular localization of tetherin after K5 expression suggests that K5 causes it to be delivered to late endosomes for degradation. Importantly, this distinguishes K5 and Vpu-induced tetherin antagonism and implies distinct mechanisms.
Since the RING domain of K5 is required for tetherin down-regulation from the cell surface, and degradation is sensitive to the drug MG132, which also causes ubiquitin depletion, we hypothesized that K5 mediated ubiquitination might drive tetherin delivery to late endosomes. We therefore examined whether proteasomal or lysozomal inhibition could rescue cell-surface tetherin levels (Figure 5D). We found that whilst endosomal inhibition with BafA1 treatment rescued tetherin protein in the cell extracts of K5 expressing cells (Figure 5A) neither BafA1 or concanamycin A could rescue tetherin levels at the cell surface (Figure 5D). However, consistent with a role for ubiquitination in the delivery of tetherin for endosomal degradation, MG132 treatment of tetherin-expressing cells did substantially rescue tetherin levels at the surface of K5-expressing cells (Figure 5D and E). In contrast, none of the inhibitors significantly rescued tetherin surface expression in Vpu-expressing cells. These observations demonstrate that K5-induces an endosomal degradation of tetherin and suggest that an ubiquitin dependent process is required for its delivery into this pathway. Conversely, HIV-1 Vpu causes tetherin degradation by a distinct mechanism that is associated with its localization to the TGN (Figure 5B–C).
Next we asked whether the double-tyrosine based endocytic motif in the tetherin cytoplasmic tail was required for K5-mediated degradation. This motif binds the clathrin adaptors AP1 and AP2 and has been reported to be important for tetherin endocytosis and recycling [40]. HT1080 cells expressing THN-HA Y6,8A were generated and tetherin localization determined by immunofluorescence microscopy. As expected this mutant tetherin was found almost exclusively at the plasma membrane (Figure 5F). Interestingly, expression of K5 in these cells still led to significant tetherin down-regulation and degradation (Figure 5F) as demonstrated by flow cytometry and western blot. Thus tetherin trafficking to late endosomes induced by K5 is independent of its endocytic motif, suggesting that K5 targets tetherin for degradation via a pathway that is independent from its normal subcellular trafficking.
K5 targeting of class I MHC molecules depends on membrane proximal lysine residues in their cytoplasmic tails [20]. Tetherin also has two membrane proximal lysines, K18 and K21. To seek a role for these residues we mutated them to arginine, singly or in combination, in THN/HA and stably expressed these mutant tetherins in HT1080 cells. We then stained the cell surface tetherin via the HA tag and measured cell surface tetherin levels by flow cytometry (Figure 6A). All tetherins were expressed at the cell surface, but the lysine mutants were expressed at enhanced levels as compared to the wild-type protein. This suggested that these two lysines might be involved in natural tetherin turnover. When K5 was expressed in the mutant tetherin cell lines, tetherin down-regulation was almost completely prevented for proteins bearing a K18R substitution. In contrast, K21R mutation had no effect on K5-induced cell surface down-regulation. Furthermore, K21R but not K18R or K18,21R mutants were consistently redistributed to CD63+ compartments upon K5 expression (Figure 6B). We then further examined whether K5 retained the ability to disrupt tetherin function in the absence of cell-surface tetherin down-regulation, as has been suggested for Vpu [17]. HT/THN-HA and HT/THN-HA K18R cells expressing either K5 or Vpu were infected with HIV-1wt and HIV-1(delVpu) VSV-G pseudotyped viruses at an MOI of 0.2. Supernatants were collected 48h later and infectious virus output was measured on HeLa-TZM cells. THN-HA K18R expressing cells restricted Vpu-defective HIV-1 release similarly to the wild type protein, however expression of K5 failed to rescue the release of HIV-1(delVpu) from HT/THN-HA-K18R cells (Figure 7A). By contrast Vpu-mediated antagonism of tetherin was unaffected by mutation of K18 in HT1080 cells (Figure 7A) or indeed either of the lysine residues in 293T cells (Figure 7B). Thus, removal of tetherin from the cell surface and delivery to late endosomes is required for K5-mediated antagonism of its antiviral action and this is dependent on the lysine at position 18. Measurement of gag levels in cell extracts and supernatants by western blot demonstrated that tetherin expression did not impact on gag expression (Figure 7C). Measurement of tetherin levels ensured that tetherin was expressed as expected (Figure 7C).
Since lysine residues serve as targets for ubiquitination we next sought evidence for tetherin ubiquitination in the presence of K5. HeLa cells were transfected with THN-HA or THN-HA K18, 21R in the presence of an ubiquitin bearing a 6-histidine tag. The cells were treated for 8h with BafA1 to block tetherin degradation and ubiquitinated proteins were isolated by incubating whole cell lysates with nickel-agarose beads. In the presence of either co-transfected K5 or Vpu, THN-HA molecules could be isolated from the transfected cells (Figure 8). The tetherin predominantly formed a single species at a size suggestive of mono-ubiquitination. Interestingly, we failed to pull down ubiquitinated tetherin molecules when their cytoplasmic lysine residues had been mutated, after either K5 or Vpu transfection. This implies that the action of both K5 and Vpu leads to ubiquitination of the tetherin cytoplasmic tail. In the case of K5, this suggests that ubiquitination of K18 antagonizes tetherin-mediated restriction and directs it to endosomal compartments for degradation. Intriguingly, mutating the lysine residues does not make tetherin insensitive to Vpu suggesting that tetherin becomes ubiquitinated as a consequence of tetherin antagonism by HIV-1 Vpu but that this ubiquitination is not required for the antagonistic process.
There are several ubiquitin-dependent mechanisms by which K5 might achieve tetherin degradation. For example, ubiquitination of tetherin's cytoplasmic tail could stimulate its internalization and mediate classical recruitment of the ESCRT pathway through engagement of TSG101, as has been shown for K3 targeting of MHC I [41]. This would lead to the budding of tetherin containing vesicles into the lumen of multivesicular bodies for degradation in lysozomes [42]. To examine the role of the ESCRT pathway in K5-mediated tetherin degradation, we tested whether K5-mediated loss of tetherin from the cell surface was sensitive to expression of a dominant negative form of VPS4. VPS4 is the essential AAA-ATPase that provides the energy for ESCRT disassembly and recycling during the final membrane scission event in the sorting of cell surface receptors for endosomal degradation [43]. Co-transfection of HeLa cells with GFP-dnVPS4(E228Q) [44] substantially rescued cell surface tetherin levels from K5 as assessed by flow-cytometry (Figure 9A and B). Expression of K5delRING has no effect on THN cell surface expression, concordant with Figure 4, and neither does the dominant negative VPS4 protein when expressed with the tetherin RING mutant. Together with the demonstration that K5 leads to tetherin ubiquitination and degradation, these observations strongly suggest that K5 induces a VPS4 and ubiquitination-dependent trafficking of tetherin from the cell surface to late endosomes for destruction. Importantly, this mechanism is similar to that used by K3 to down-regulate Class I MHC molecules [38], [41].
Tetherin has emerged as a potent inhibitor of enveloped virus release [5], [6], [9], [12], [13]. Recent evidence has demonstrated that tetherin dimers act as a physical linkage between the membranes of the infected cell and nascent virions [10]. This mechanism lends itself well to a general non-specific antiviral inhibition that restricts virus release and thereby interferes with viral spread to new target cells. It also suggests that mammalian viruses may not be able to easily mutate to avoid tetherin because tetherin does not directly interact with any viral structural proteins. Sensitive viruses must therefore evolve specific ways of counteracting it. In the case of primate lentiviruses, several tetherin antagonists have now been identified. The Vpu accessory protein antagonizes tetherin in HIV-1 infected cells [5], [6], whereas in a variety of SIVs that do not encode a Vpu gene, the Nef protein can overcome the tetherin orthologues from their host species [33], [45]. Interestingly, HIV-2 [32] and at least one strain of SIV [39] have acquired the ability to antagonize tetherin with their envelope glycoproteins. Outside the Retroviridae, the ebolavirus glycoprotein has anti-tetherin activity [13] and here we propose that human herpesvirus KSHV antagonizes tetherin with K5.
While this study was in revision, a study from Mansouri and colleagues showed that tetherin could be degraded by K5 in an ubiquitin-dependent manner [46]. Our data demonstrate that K5 can fully substitute for Vpu in mediating the efficient release of HIV-1 particles from tetherin-expressing cells. This requires the K5 RING domain and leads to a cell-surface down-regulation of tetherin followed by its degradation in endosomal compartments. Down-regulation of tetherin by K5 and its targeting to endosomes requires the membrane proximal lysine residue K18 in the cytoplasmic tail, and this process is sensitive to inhibition of the proteasome with MG132. Since K5-induced ubiquitination of tetherin is dependent on its cytoplasmic tail lysines, the effects of proteasomal inhibition are likely to be due to depletion of cytoplasmic ubiquitin levels, rather than blocking proteasomal degradation of tetherin. Furthermore, K5-mediated tetherin degradation requires a functional ESCRT pathway as shown by the rescue of surface tetherin levels after expression of a dominant negative VPS4 protein. These observations suggest that K5 targets tetherin by inducing an ubiquitin-dependent sorting of tetherin to multivesicular bodies where it is destroyed in a lysozomal compartment, a similar mechanism to that of K3-mediated degradation of Class I MHC molecules [41]. K3 targets MHCI for ESCRT-dependent sorting and destruction via addition of a single ubiquitin moiety to the molecule's cytoplasmic tail through recruitment of the E2 enzymes UbcH5B and/or C [38]. Lysine-63 poly-ubiquitination is then induced through the subsequent recruitment of Ubc13 [38]. While we cannot rule out that K5 induces poly-ubiquitination of tetherin, we were only able to observe species consistent with mono-ubiquitination. Thus, while K5 induces ESCRT-dependent tetherin degradation, the precise molecular details of the endosomal targeting may differ between K3 and K5 targets. Whether endocytosis of tetherin is stimulated by K5, or whether it is routed to endosomes from the Golgi, bypassing the cell surface remains an interesting question. Our observation that surface down-regulation of tetherin is independent of its tyrosine-based endocytic-sorting signal [40] suggests the latter. Since K5 localizes mainly to the ER [47], tetherin ubiquitination could happen very early after synthesis leading to endosomal routing independently of the cell surface. Similarly, Mansouri and colleagues suggested that in K5-expressing cells, little tetherin reaches the PM, based on surface biotinylation experiments [46]. However, neither of these results is unambiguous, especially if cell surface turnover is fast. Further studies of the molecular details of K5 mechanism are required to fully dissect its effects on tetherin trafficking.
An important aspect of our study is the comparative analysis of the mechanisms by which Vpu and K5 achieve tetherin antagonism. Both proteins lead to cell surface down-regulation and degradation of tetherin but the mechanism of Vpu remains unclear. Several studies have shown that tetherin degradation is blocked by proteasomal inhibition [19], [34], [35], whereas others suggest endosomal degradation [36], [37]. It is clear that whilst Vpu-mutants that cannot interact with βTRCP2 cannot mediate tetherin degradation [36], [37], they can nonetheless antagonize tetherin and rescue viral release [17]. Thus, the SCF-Skp1-cullin 1 ubiquitin ligase complex and perhaps an ER-associated degradative pathway are implicated in tetherin degradation and this process presumably follows tetherin antagonism at the cell surface [35], [36]. Tetherin down-regulation and degradation might therefore not be as strictly linked during Vpu-mediated antagonism as it is during K5 mediated antagonism of tetherin. Furthermore, proteasomal inhibition appeared to be more potent than endosomal inhibition in rescuing cellular levels of tetherin in Vpu-expressing cells. Unlike K5, we saw no evidence of tetherin redistribution to endosomes in response to Vpu. Rather, residual tetherin can be seen in the TGN, consistent with a recent study suggesting that Vpu localization to the TGN is important for tetherin antagonism [48]. Similarly, proteasomal inhibition does not restore tetherin to the surface of Vpu-expressing cells, and neither does dominant negative VPS4 (not shown). Also consistent with these observations, is our observation that Vpu can induce ubiquitination of tetherin cytoplasmic tail lysine residues, but these are dispensable for Vpu sensitivity. Thus their ubiquitination appears to be a consequence of tetherin antagonism rather than the absolute requirement for K18 demonstrated for K5 sensitivity. In this respect, we suggest tetherin antagonism by Vpu precedes, and may not be dependent on, degradation, but rather results in the sequestration of tetherin away from budding virions, preventing incorporation. Thus there are more parallels with the mechanism by which HIV-2 and SIVtan envelopes antagonize tetherin through sequestration in TGN-associated compartments [32], [39]. K5's apparent inability to antagonize tetherin in 293T cells, cells that support Vpu's antagonism of tetherin, suggests that Vpu and K5 may require different cellular cofactors. Clearly, further comparative mechanistic studies will allow us to dissect the differences and similarities in the mode of action of these two very different proteins.
We, and others [46], have also shown that productive KSHV release is restricted by tetherin expression and knockdown of K5 expression imparted a block to virus release. Furthermore, tetherin expression is reduced on B cells after K5 expression. Importantly, this indicates that herpesvirus particle assembly is sensitive to the antiviral effects of tetherin. The mechanism of herpesvirus assembly and envelopment is complex and controversial, most studies have focused on herpes simplex virus type 1 (HSV-1). Immature HSV-1 capsids may bud through nuclear membrane, re-entering the cytoplasm, and then bud again into secretory vesicles [29] via an ESCRT dependent process [49]. Our data suggest that in the case of KSHV, at least one budding stage is through a membrane accessible to tetherin. Tetherin is highly expressed in terminally differentiated B cells and plasma cells, important cellular targets for KSHV [50]. B cell to plasma cell differentiation activates KSHV lytic replication through the activation of XBP-1 and the unfolded protein response [51]. Therefore the virus undergoes productive replication in cells that express high levels of tetherin. This cellular tropism may have provided the selective pressure for K5 evolution to target tetherin. K5 has the ability to modulate the expression of a variety of cell surface molecules involved in immuno-recognition (MHC and NK receptors) and cell adhesion, suggesting targeting of tetherin is part of a wider immuno-evasion strategy by KSHV [52]. Is tetherin antagonism found in other mammalian herpesviruses? B cell expression of tetherin would certainly suggest that this might be the case for Epstein Barr Virus (EBV). Removal of MHC and related molecules is certainly a common feature of several human herpesvirus immune evasion strategies [52], [53], [54], and it is likely that if other herpesviruses are sensitive to tetherin-mediated restriction, proteins with analogous function to K5 might also target tetherin. MARCH ligase homologues are also found in a variety of poxviruses [20], which also have highly complex envelopment strategies [55].
Aside from the cytoplasmic tail lysine residue K18, we do not yet fully understand the determinants of sensitivity for tetherin's antagonism by K5. Data presented herein show that while K5, like Vpu [18], [19], cannot target tetherin from Rhesus macaques, the TM domain positions that determine the difference in Vpu-sensitivity do not confer resistance to K5. Tetherin has been under high levels of positive selection during mammalian evolution, particularly in the cytoplasmic tail and TM domain, areas of the protein likely to be topologically accessible to viral antagonists [18], [19]. Several residues in the TM domain are responsible for the species-specific activity of Vpu. However, positive selection in other parts of tetherin may be due to selective pressure exerted by distinct viral countermeasures. We speculate that some mammalian herpesviruses, which show strong co-evolution with their hosts, may have contributed to this evolutionary pressure. Tetherin may therefore represent an extremely interesting case study in the evolution of an antiviral gene under regular assault by unrelated viral pathogens, as has been suggested for PKR [56].
In conclusion, here we have demonstrated that tetherin is capable of restricting a human herpesvirus and show that in the case of KSHV, the virus has co-opted an immuno-modulatory ubiquitin ligase to target this antiviral effector.
All adherent cells were maintained in Dulbecco's Modified Eagle Medium (DMEM) supplemented with 10% fetal calf serum and antibiotics. BCBL-1 cells were maintained in Roswell Park Memorial Institute (RPMI) medium supplemented with 10% FCS and antibiotics. HeLa, 293T and HT1080 cells were obtained from the ATCC. The HIV-1 indicator cell line, HeLa-TZM, that expresses HIV-1 receptors and an integrated HIV-1-LTR controlling expression of a beta-galactosidase reporter gene were kindly provided by John Kappes and Xiaoyun Wu via the NIH AIDS reagents program. HT1080/THN-HA is a clonal cell line expressing human tetherin with an internal HA tag inserted at nucleotide position 463 [18] expressed from an integrated MLV proviral vector, pLHCX (Clontech). Mutants of tetherin were constructed by standard methods and expressed from pCR3.1 and pLHCX. Vpu, K5 and a K5 mutant lacking amino acids 1–62 encompassing the RING domain (K5delRING) were amplified by PCR and inserted into pCR3.1 with C-terminal HA and mCherry fusion tags and the retroviral vectors pCxCR, which also encodes dsRED express as a marker gene, and pCMS28, a pMigR1 derivative encoding puromycin under control of an Internal Ribosome Entry Site (IRES). The molecular clones of HIV-1 NL4.3 and the Vpu-defective counterpart have been described previously [7]. Anti-HA monoclonal antibody HA1.11 was obtained from Covance, rabbit anti-HA polyclonal was obtained from Rockland and anti-BST2 monoclonal antibody was obtained from Abnova. Secondary Alexa Fluor 488, 594 and 633 antibodies for flow cytometry and microscopy were obtained from Molecular Probes. For quantitative western blotting, Licor 680 and 800nm secondary abs were used and blots scanned using a Licor scanner.
HeLa cells infected with KSHV were made by infecting HeLa with rKSHV.219 a recombinant KSHV virus encoding RFP, GFP and puromycin resistance, a gift from Jeff Vieira [30]. Cells were selected and kept under puromycin selection. To induce rKSHV.219 into the lytic cycle, r219-Hela cells were seeded at 3.105 cells/well onto 6-well plates and transfected with 1.5 µg of an expression factor for RTA (pCMV-RTA) (a gift from Adrian Whitehouse) using Fugene-6. RFP expression was visible a day after transfection. The transfection mix was removed and replaced with 2 ml of fresh medium. Virus was recovered another 24 hours later, i.e. 48 hours after transfection, and filtered through 0.45 µM device (Millipore) to remove any cellular debris. Infectious virus was measured by titration of supernatants onto 293T cells and total virus by QPCR on DNAse-I resistant genomes. To assay for KSHV release in the presence of tetherin, r219-Hela cells were transfected with pCMV-RTA and increasing doses of tetherin expression vector pCR3.1-THN [5]. Plasmid dose was kept constant using empty vector (pCDNA3.1). To assay for KSHV release under conditions where K5 expression is reduced, r219-Hela cells were seeded at 105 cells per well in 6-well plates 24 hours prior to transduction with the lentiviral vectors encoding the K5 specific hairpins, or the empty vector, at a multiplicity of infection (MOI) of 5. 72 hours post-transduction, cells were counted, re-seeded into 6-well plates and transfected with RTA encoding plasmid as above.
293T cells were seeded at 105 cells/well in 12-well plates 24 hours prior to titration. For each viral collection 250 µl (a sixth) of the final volume of filtered virus was added to fresh medium to a final volume of 1 ml and used to infect 293T cells. Titrations were performed in duplicate using polybrene (4µg/ml). 12-well plates were subjected to spinoculation at 500 g for 1 hour at RT before being returned to the incubator. Cells were analyzed by flow cytometry 48 hours post-inoculation.
KSHV genomes were quantified by extracting total DNA (QIAamp, QIAGEN) from DNAse-I treated supernatant (70 units/ml for 2 hours, RQ1 Promega, UK) with 40µg of salmon sperm DNA as a carrier (Sigma, Poole UK). 5µg of purified DNA was subjected to quantitative Taqman PCR for KSHV early gene ORF37 as described [57]. Absolute copy number was determined with reference to a standard curve derived by QPCR against serial dilutions of an ORF37 amplicon encoding plasmid, a gift from David Bibby and Duncan Clark, as described [58].
Total mRNA was extracted from r219-HeLa cells after virus collection, 48 hours post-RTA transfection (RNeasy, QIAGEN). cDNA syntheses were performed on 4 µl of the RNA (SuperScript II Reverse Transcriptase, Invitrogen) according to manufacturer's instructions. cDNAs were treated with 2 units of RNAse H (Invitrogen) for 20 minutes at 37°C before being used in taqman Q-PCR reactions for ORF37 and GAPDH. GAPDH primers were GAPDH forward primer, 5′-GGCTGAGAACGGGAAGCTT-3′; GAPDH reverse primer, 5′-AGGGATCTCGCTCCTGGAA-3′; GAPDH probe, 5′-FAM-TCATCAATGGAAATCCCATCACCA-TAMRA-3′. Absolute copy number was determined with reference to a standard curve derived by QPCR against serial dilutions of a GAPDH amplicon encoding plasmid. QPCR for ORF37 was performed as described above.
Cellular DNA was extracted from r219-HeLa cells after virus collection, 48 hours post-RTA transfection (QIAamp, QIAGEN) and QPCR for ORF37 performed, as described above, on cellular associated episomes. Copy numbers were normalized to quantities of DNA used for each reaction.
Semi-confluent 293T cells on 6-well dishes were transfected with 1µg of vector plasmid, 1µg of pMLVgag-pol or p8.91 (HIV-1 gag-pol, tat and rev expression vector) and 0.2µg pCMV-VSVG. For full length replication competent HIV-1 (VSV-G) pseudotypes, 293T cells were transfected with 2µg of pNL4.3 or pNL4.3(del Vpu) and 0.2µg of VSV-G. This leads to production of replication competent HIV-1 that has both VSV-G and HIV-1 gp160 envelope proteins on its surface. Virus and vector stocks were harvested 48h post-transfection. Those encoding florescent markers were titrated on HT1080 cells and analyzed by flow cytometry; lentiviral vectors encoding shRNA hairpins were titered by Q-PCR on 293T cells with primers specific for the HIV-1 LTR; and the endpoint titer of full length HIV-1 pseudotypes was determined on HeLa-TZM cells by staining infected foci with X-Gal 48h post-infection.
Subconfluent HeLa cells were transfected with 500ng of HIV-1 proviral plasmid and 100ng of pCR3.1-Vpu or pCR3.1-K5-HA/K5delRING-HA using lipofectamine 2000 (Invitrogen). Viral supernatants and cell lysates were harvested 48h later. Supernatant virions were filtered and pelleted at 30000g through 20% sucrose/PBS cushion for 90 minutes. Lysates and virions were then separated by SDS-PAGE and HIV-1 Gag proteins detected using an anti-p24 monoclonal antibody CA183 (provided by B Chesboro through the NIH AIDS Reagents repository). In parallel, harvested supernatants were titered onto HeLa-TZM indicator cells and 48 hours later beta-gal activity was determined in cell lysates using a Tropix Beta-galactosidase activity kit (Molecular Probes). For assays involving stable HT1080/THN-HA cell lines, 105 cells were plated per well in a 12 well plate and infected with 2×104 infectious units (MOI 0.2) of HIV-1 (VSV-G) or HIV-1(del Vpu)(VSV-G) viral stocks. 48h later cell lysates and viral supernatants were treated as above.
Cell surface staining for tetherin and THN-HA was performed using the appropriate antibodies by standard methods and analyzed on a FACS-Calibur (Becton-Dickinson). Cells for microscopy were grown on glass coverslips and transfected or treated with BafA1 (100nM), concanamycin A (100nM) or MG132 (1µg/ml) (Sigma, UK) for 10h. The cells were then fixed in 4% paraformaldehyde, permeabilized with 0.1% Triton-×100 and stained with the required primary and secondary antibody. Cells were then examined using a Zeiss confocal microscope.
HeLa cells seeded on 10cm plates were transiently transfected with the indicated THN-HA expression vector (5ug) in combination with K5, Vpu or GFP (2µg) expression vectors and in the presence or absence of a 6-His-tagged ubiquitin encoding plasmid (2µg). 48h after transfection, cells were treated for 8h with BafA1 (Sigma) to prevent further tetherin degradation. Cells were lysed in 5M guanidinium hydrochloride, sonicated and ubiquitinated proteins were isolated by binding to 50µl of Ni2+ Nti-agarose (Invitrogen) for 3h at room temperature. The beads were eluted with 100mM imidazole. Lysates and pull-downs were then analyzed by western blot for THN-HA.
|
10.1371/journal.ppat.1004868 | Discovery of a Small Non-AUG-Initiated ORF in Poleroviruses and Luteoviruses That Is Required for Long-Distance Movement | Viruses in the family Luteoviridae have positive-sense RNA genomes of around 5.2 to 6.3 kb, and they are limited to the phloem in infected plants. The Luteovirus and Polerovirus genera include all but one virus in the Luteoviridae. They share a common gene block, which encodes the coat protein (ORF3), a movement protein (ORF4), and a carboxy-terminal extension to the coat protein (ORF5). These three proteins all have been reported to participate in the phloem-specific movement of the virus in plants. All three are translated from one subgenomic RNA, sgRNA1. Here, we report the discovery of a novel short ORF, termed ORF3a, encoded near the 5’ end of sgRNA1. Initially, this ORF was predicted by statistical analysis of sequence variation in large sets of aligned viral sequences. ORF3a is positioned upstream of ORF3 and its translation initiates at a non-AUG codon. Functional analysis of the ORF3a protein, P3a, was conducted with Turnip yellows virus (TuYV), a polerovirus, for which translation of ORF3a begins at an ACG codon. ORF3a was translated from a transcript corresponding to sgRNA1 in vitro, and immunodetection assays confirmed expression of P3a in infected protoplasts and in agroinoculated plants. Mutations that prevent expression of P3a, or which overexpress P3a, did not affect TuYV replication in protoplasts or inoculated Arabidopsis thaliana leaves, but prevented virus systemic infection (long-distance movement) in plants. Expression of P3a from a separate viral or plasmid vector complemented movement of a TuYV mutant lacking ORF3a. Subcellular localization studies with fluorescent protein fusions revealed that P3a is targeted to the Golgi apparatus and plasmodesmata, supporting an essential role for P3a in viral movement.
| In order to maximize coding capacity, RNA viruses often encode overlapping genes and use unusual translational control mechanisms. Plant viruses express proteins required for movement of the virus through the plant, often from non-canonically translated open reading frames (ORFs). Viruses in the economically important Luteoviridae family are confined to the phloem (vascular) tissue, probably due to their specialized phloem-specific movement proteins. These proteins are translated from one viral mRNA, sgRNA1, via initiation at more than one AUG codon to express overlapping genes, and by ribosomal read-through of a stop codon. Here, we describe yet another gene translated from sgRNA1, ORF3a. Translation of ORF3a initiates at a non-standard (not AUG) start codon. We found that ORF3a is not required for viral genome replication, but is required for long-distance movement of the virus in the plant. The movement function could be restored in trans by providing the ORF3a product, P3a, from another viral or plasmid vector. P3a localizes in the Golgi apparatus and adjacent to the plasmodesmata, supporting a role in intercellular movement. In summary, we used a powerful bioinformatic tool to discover a cryptic gene whose product is required for movement of a phloem-specific plant virus, revealing multiple levels of translational control that regulate expression of four proteins from a single mRNA.
| RNA viruses are models of efficiency in compressing maximum information, such as coding and regulatory signals, into minimum sequence space. To do this, RNA viruses often employ noncanonical translation mechanisms [1, 2]. For example, many viruses encode genes in overlapping open reading frames (ORFs), some of which can be very short. Also, the arrangement of the ORFs themselves can regulate their expression. To decipher a virus life cycle, it is imperative to identify all the coding regions and to understand their function and how they are regulated. However, small functional ORFs, often lacking conventional initiation sites, can be very difficult to detect. Thus, specialized bioinformatic tools are often required to detect key viral genes. Here we use such tools to identify an essential ORF conserved in the two main genera in the Luteoviridae family, and provide evidence of its role in infection.
Viruses in the economically important Luteoviridae family are paragons of translational control. They employ leaky scanning to initiate translation at separate start codons, ribosomal frameshifting, and stop codon readthrough to express various genes, some of which overlap (Fig 1). The Luteoviridae family comprises over 33 viruses distributed among three genera, including the wide-spread Barley yellow dwarf virus (BYDV) in genus Luteovirus, and Potato leafroll virus (PLRV), Turnip yellows virus (TuYV) and Cereal yellow dwarf virus (CYDV) in genus Polerovirus, and Pea enation mosaic virus 1 (PEMV1), the sole member of genus Enamovirus [3]. All Luteoviridae species are transmitted in a persistent and circulative manner by aphids, but they do not replicate in the aphid, and all but PEMV are confined to the phloem in the plant [4]. Genus Luteovirus differs from the others in that the RNA-dependent RNA polymerase (RdRp) and the translational and replication control signals throughout the genome resemble those of the Tombusviridae family [5]. In addition, like the Tombusviridae, the genome of the Luteovirus genus members has no 5’ modification [6], whereas the genomes of poleroviruses and the enamovirus have a genome-linked protein (VPg) covalently attached to the 5’ end [7, 8].
Viruses in all three genera have a 5.2 to 6.3 kb positive-sense RNA genome from which a subgenomic mRNA (sgRNA1) is generated in infected cells [9]. sgRNA1 consists of the 3’-terminal half of the genome in genus Luteovirus, or the 3’-terminal third of the genome in the Polerovirus and Enamovirus genera (Fig 1A and 1C). sgRNA1 serves as the mRNA for translation of ORFs 3, 4, and 5. ORF3 encodes the coat protein (CP), ORF4 codes for P4 and is embedded within the CP ORF but in a different reading frame. ORF5 is translated by in-frame readthrough of the CP ORF stop codon, a process stimulated by sequences adjacent to, and far downstream of, the leaky stop codon [10, 11]. Thus, the translation product of ORF5 (RTD; readthrough domain) does not exist on its own, but is present only as a C-terminal extension of the CP in the CP-RTD protein. The icosahedral (T = 3) virion contains 180 copies of the CP, of which, in the case of BYDV, an estimated 10 to 25% are present in the form of CP-RTD [12]. The RTD is required for aphid transmission [13, 14] and in some viruses also for virus movement in certain plant hosts [13–15]. P4 is likely a cell-to-cell movement protein [16–18]. The sole official enamovirus, PEMV1, lacks ORF4, has a truncated RTD and depends on a helper virus, PEMV2 (genus Umbravirus), for efficient movement in the plant [19].
Translation of ORF4 depends on a leaky scanning mechanism whereby some scanning 40S ribosomal subunits fail to initiate at the ORF3 AUG initiation codon and instead continue scanning to the downstream ORF4 AUG initiation codon [20]. This mechanism is facilitated by the generally poor context of the ORF3 initiation codon. In mammals, potential initiation codons with an A at -3, or a G at -3 and a G at +4, (where the A of the AUG is nucleotide +1) may be regarded as being in a ‘strong’ context for initiation [21], while other contexts facilitate leaky scanning. Similar context rules appear to apply in plants [22–24]. Leaky scanning may also be facilitated by the use of non-AUG initiation codons. The near-cognate codons CUG, GUG, ACG, AUU, AUA, UUG and AUC are, under certain circumstances, able to support a significant level of initiation (typically 2–15% of the level of initiation at an AUG codon in a similar context), with CUG being the most efficient non-AUG initiation codon in many systems [1, 25]. Initiation at non-AUG codons normally requires a strong initiation context, but may also be enhanced in less predictable ways by RNA structure within the message [26]. In several plant viruses, a combination of non-AUG and poor-context AUG initiation codons allows production of three or even four functional proteins from a single transcript [27–29].
The 5’ end of sgRNA1 of luteoviruses and poleroviruses, where known, varies between 188 and 302 nt upstream of the ORF3 AUG [30–32], usually including sequence that encodes the 3’ end of ORF2. This is an unusually long leader sequence for a viral sgRNA because, as mentioned above, RNA viruses tend to minimize sequence length wherever possible. A long 5’ untranslated region (UTR) often implies the presence of a translational enhancer such as an internal ribosome entry site [33], but the cap-independent translation element for genus Luteovirus, called a BYDV-like translation enhancer (BTE), is located in the 3’ UTR [34, 35], and only a small stem-loop at the 5’ end of sgRNA1 is needed for cap-independent translation [36]. Instead of being a long 5’ UTR, we report here that the 5’ end of sgRNA1 of poleroviruses and luteoviruses (but not of the enamovirus) encodes a small ORF, termed ORF3a, that initiates at a non-AUG codon. The encoded protein P3a of TuYV is not required for replication in protoplasts but is required for systemic infection in plants.
The coding potential of ORF3a was detected initially by applying the gene-finding program MLOGD to luteovirus and polerovirus sequence alignments. MLOGD uses nucleotide and amino acid substitution matrices to model sequence evolution in dual-coding, single-coding and non-coding regions [37]. It can be used to predict novel coding sequences via an approximate likelihood-ratio test. Although originally developed to analyze overlapping genes, MLOGD can also be used to analyze the coding potential in each of the three reading frames relative to a 'null' model in which the sequence is presumed to be non-coding in that reading frame.
Fig 1B illustrates the application of MLOGD to an alignment of 76 Barley yellow dwarf virus (BYDV) sequences (including serotypes PAV, PAS, MAV, GAV and the highly divergent Ker-II) with full or near-full genome coverage, using a 40-codon sliding window separately in each reading frame. A positive coding signature was observed in the correct reading frame throughout the ORF1, ORF2 and ORF5 regions. As expected, due to the lower number of substitutions in dual-coding regions, the coding signature was weaker in the ORF3/ORF4 overlap region, but still mainly positive. Furthermore, a positive coding signature was observed in the ORF6 region, supporting earlier evidence that it may encode a functional product [38–40]. Unexpectedly, a short region of positive coding potential was observed upstream of ORF3, in a region hitherto presumed to be part of the sgRNA1 non-coding leader (pink band, Fig 1B). Moreover, the region of positive coding potential coincided with a conserved absence of stop codons in the corresponding reading frame (Fig 1B). We named this open reading frame ORF3a. We observed that ORF3a is conserved in the three clades identified in the genus Luteovirus, (i) the BYDVs, (ii) Bean leafroll virus, Soybean dwarf virus and relatives, and (iii) Rose spring dwarf-associated virus (S1 Datafile).
We next analyzed 97 full- or nearly full-length genome sequences of viruses in genus Polerovirus. Recombination, particularly between the 5' replication and 3' capsid/movement gene blocks, is a common feature of polerovirus and luteovirus evolution [5, 41]. In view of this, the 5' ORF0-ORF1-ORF2 and 3' ORF3-ORF4-ORF5 gene blocks were extracted from polerovirus sequences and aligned separately. 178 nucleotides of 5' flanking sequence were included in the ORF3-ORF4-ORF5 alignment in order to include the potential ORF3a region and some upstream flanking sequence in the analysis. MLOGD revealed a positive coding signature in the correct reading frame throughout most of the ORF0, ORF1, ORF2, ORF3, ORF4 and ORF5 regions (Fig 1D). Again, a short but clear region of positive coding potential was observed just upstream of ORF3 and, once again, this coincided with a conserved absence of stop codons in the corresponding reading frame (pink band, Fig 1D). Note that this analysis does not provide information on any additional ORFs that may be restricted to just one or a few polerovirus species, such as the Rap1 ORF reported only in the PLRV genome [42].
Although the presence of ORF3a is conserved throughout the Luteovirus and Polerovirus genera, in nearly all sequences it lacks a suitable AUG initiation codon. Thus we searched for potential non-AUG initiation codons. With few exceptions (see below), all available sequences with coverage of the ORF3a region contain a near-cognate non-AUG potential initiation codon, in a strong initiation context, near the 5' end of the maximal open reading frame (which is determined by the next upstream in-frame stop codon). Representative sequences (NCBI species RefSeqs) are shown in Fig 2; additional sequences are shown in S1 Datafile. The great majority of sequences contain an AUU, ACG, AUA or CUG ORF3a-frame codon (green shading, Fig 2), flanked by a favorable translation initiation context, i.e. an A at position -3 and frequently also a G at +4. Initiation at one of these codons would give rise to a 45–48 amino acid P3a product. The presence of ORF3a-frame stop codons (amber shading, Fig 2) in many sequences shortly upstream of this site further suggests that this is the site of initiation. As only a fraction of scanning 40S ribosomal subunits initiate translation at any given non-AUG initiation codon, it is possible that in some species multiple non-AUG initiation sites are utilized (e.g. closely spaced ACG and AUU codons, both with an A at -3, in BYDV sequences NC_002160, NC_003680, NC_004750 and NC_004666; Fig 2). ORF3a normally terminates shortly downstream of the ORF3 initiation codon, overlapping ORF3 in the +2 reading frame (Fig 2). ORF4 overlaps ORF3 in the +1 reading frame and, in almost all viruses, the ORF4 initiation codon is shortly downstream of the ORF3a termination codon (Fig 2).
Four of the 27 NCBI RefSeqs differ from this general pattern. Uniquely, in one luteovirus (NC_006265, Carrot red leaf virus) the ORF3a stop codon is upstream of the ORF3 AUG codon. This is due to replacement of the canonical ORF3 AUG codon with ACG (Fig 2). The single nucleotide deletion that disrupts the ORF3a reading frame in NC_004756 (Beet western yellows virus; pink '-' in Fig 2) is not present (i.e. is replaced with a nucleotide) in all other available Beet western yellows virus sequences in NCBI (>30 sequences) suggesting that the RefSeq may have a sequencing error or represent a defective genome. In NC_003491 (Beet mild yellowing virus, BMYV) and NC_002766 (Beet chlorosis virus), ORF3a is shorter and initiates with an AUG codon in a weak context instead of a non-AUG initiation codon in a strong context (Fig 2). However, in other sequences of these two species the AUG is replaced with AUUG giving rise to the full-length canonical ORF3a (see S1 Datafile). Whether these RefSeqs represent defective sequences or functional variants remains to be determined. However, an infectious clone of BMYV contains an intact ORF3a with AUUG rather than the AUG sequence present in the RefSeq [43].
We suggest that these exceptions are due to sequencing errors or sequences of nonviable viral RNAs. It should be noted that NCBI RefSeqs are often derived from older sequences (typically the first full-length sequence obtained for a species) and are sometimes prone to sequencing errors that are not supported by later sequencing (e.g. see S1 Datafile). While insertion/deletion errors that occur in known coding ORFs are generally corrected, errors elsewhere often escape notice. The long standing confusion regarding the genome organization of sobemoviruses, which arose as a result of insertion/deletion errors in a number of early sequences, is a case in point [44]. When we analyzed all 459 sequences available in GenBank with coverage of the ORF3/3a region, only 10 were found to be defective or potentially defective with respect to ORF3a as a result of insertions, deletions or premature termination codons, while only one sequence lacked a strong initiation context at the canonical ORF3a initiation site (S1 Datafile).
The protein product of ORF3a, P3a, has a predicted molecular mass normally in the range 4.8 to 5.3 kDa. Its amino acid sequence is generally highly conserved between divergent virus species (S1 Fig). Moreover, all sequences, except the two which are N-terminally truncated (AUG initiation; see above), contain a predicted transmembrane region towards the N-terminus of P3a (S1 Fig).
In order to experimentally evaluate the expression of ORF3a, we selected the polerovirus Turnip yellows virus (TuYV), and performed in vitro translation experiments in wheat germ extracts using T7-derived transcripts starting at nt 3259 which corresponds to the 5’ end of TuYV sgRNA1. This places ORF3a in its natural context in the TuYV sgRNA1, beginning upstream of ORFs 3 and 4. Based on alignments, translation of ORF3a is suspected to start with an ACG (Fig 2, NC_003743.1, nt 3365) and to stop with a UAG (nt 3502), producing a theoretical protein of 45 amino acids (MW 5.1 kDa). This ACG displays a favorable translation initiation context (A at -3 and G at +4) at a position that is conserved among most poleroviruses and luteoviruses (Fig 2; S1 Datafile).
To monitor ORF3a expression and function, several mutants were constructed (Figs 3A and S2). As a positive control, the putative ORF3a start codon (ACG) was mutated to AUG (mutant “AUG”). As a negative control, the ACG was mutated to AGC, a codon that should not function as a translation initiation site [1, 45]. Capped in vitro transcripts of the corresponding subgenomic RNAs (WT, AUG and AGC) were translated in wheat germ extracts and a band migrating at about 6.8 kDa was generated from the WT and AUG constructs (Fig 3B). Although migrating more slowly than the predicted size of 5.1 kDa, evidence below supports the notion that this is the product of ORF3a, initiating at the predicted ACG codon. For example, as predicted, translation of this protein increased when the ACG was changed to AUG (Fig 3B, lanes 2 and 3). Moreover, no band of similar size was observed from the construct in which ACG was mutated to AGC (Fig 3B, lane 4), but two minor bands (migrating at 4.6 and 7.3 kDa), also present with the WT construct, were observed. These bands could result from alternative translation initiation events, with the 7.3 kDa-migrating product potentially arising via initiation at an in-frame AUU codon four triplets upstream of the ACG (S2 Fig). The lower mass product of 4.6 kDa could result from initiation at one of several in-frame downstream alternative initiation codons (AUA, GUG and AUC), the first one being located 16 codons downstream of the ACG (Figs 2 and S2). To verify that the higher mass proteins arose from ORF3a, tandem stop codons (UAA UAG) were introduced by site-directed mutagenesis of two internal codons (UCA UCG, 14 codons downstream of the proposed ACG initiation codon) in order to prematurely interrupt translation of ORF3a (Figs 3A, 2stop construct, and S2). Translation of the corresponding sgRNA1 yielded neither the main 6.8 kDa protein nor the minor 7.3 kDa band (Fig 3B, lane 5). This observation confirms that translation of both products initiated at codons upstream of, and in-frame with, the introduced stop codons, which is in agreement with a major translation initiation at the aforesaid ACG codon (nt 3365) generating the P3a protein.
In a first attempt to detect the P3a protein in vivo (see below), a tagged version was constructed with a FLAG tag (DYKDDDDK) positioned directly after the P3a ACG initiation codon, by adding five codons (DDDDK) to the DYK encoded in the WT TuYV ORF3a (Figs 3A and S2, FLAG mutant). Translation of the FLAG construct generated a band of a slightly larger size (8 kDa) than the P3a protein expressed from the WT construct, with an additional faint band above it, supporting the ACG as the major initiation codon of ORF3a, and an upstream codon (presumably the aforementioned AUU) being used as an alternative initiation site (Figs 3B, lane 6, and S2). All together, these experiments showed that ORF3a can be translated in vitro in wheat germ extracts from a synthetic subgenomic RNA to produce a protein of 6.8 kDa apparent MW.
Translation of the wild type and mutant subgenomic RNAs also produced major bands corresponding to the P4 and the CP proteins (19.5 kDa and 22.5 kDa, respectively) (Fig 3B). The ORF3a AUG mutation reduced accumulation of the 22 kDa product when compared with the WT and the other constructs (Figs 3B and S3). No significant variation in the accumulation of either protein was noticed with any of the other mutants.
To determine whether ORF3a is expressed during infection and whether it plays a role in viral replication, the previously described mutations were introduced into the T7-based TuYV full-length clone (pTuYV-WT, formerly named pBW0, [46]). Capped in vitro transcripts derived from the mutated viral clones pTuYV-3aAUG, -3aAGC, -3a2stop and -3aFLAG mutants were inoculated to Chenopodium quinoa protoplasts. At 44 hours post-inoculation (hpi), northern blot hybridization revealed that all four mutants produced genomic (gRNA) and the major subgenomic RNA (sgRNA1) at levels similar to TuYV-WT (Fig 4A). Thus viral RNA replication is independent of the presence of P3a.
Expression of ORF3a in vivo was analyzed by western blot using specific antibodies generated against the fifteen C-terminal amino acids of P3a. The P3a protein was detected with an apparent MW of 6.5 kDa in protoplasts infected with TuYV-WT or the TuYV-3aAUG mutant and 8.5 kDa in TuYV-3aFLAG-infected protoplasts (Fig 4B). TuYV-3aAUG yielded much more P3a protein than the WT virus, as already observed in the in vitro translation experiments (see above). Conversely, and as expected, no P3a protein was detected in protoplasts infected with the null-3a mutants TuYV-3aAGC or TuYV-3a2stop (Fig 4B). No other major P3a-related products were detected in protoplast extracts suggesting that the 7.3 kDa and 4.6 kDa products generated in cell-free translation (Fig 3B) were due to aberrant translation initiation that does not occur in vivo. The FLAG-tagged P3a was also detected using commercial antibodies against the FLAG epitope (S4 Fig). These results indicate unambiguously that the P3a protein is expressed in vivo during viral infection.
To investigate the effect of ORF3a on expression of the other proteins produced from the sgRNA1 in infected cells, CP-RTD, CP and P4 protein accumulation was analyzed by western blotting of proteins extracted from infected protoplasts. All mutants, except TuYV-3aAUG, produced amounts of CP, P4 and CP-RTD proteins similar to those of TuYV-WT (Fig 4C). In contrast, accumulation of CP, P4 and CP-RTD from the TuYV-3aAUG mutant was drastically impaired and not visible on the blot in Fig 4C. Loading increasing amounts of TuYV-3aAUG-infected protoplasts revealed a band corresponding to less than one-tenth of the amount of CP-RTD present in TuYV-WT-infected cells (S5 Fig). No CP could be detected in 30,000 protoplasts, while as few as 3,000 cells allowed CP detection in TuYV-WT-inoculated protoplasts (S5 Fig). Thus, the CP-RTD protein appears to accumulate in higher levels relative to CP in TuYV-3aAUG-infected protoplasts than in TuYV- WT-inoculated cells. Alternatively, the CP antibodies may have a nonlinear response to dilution and are unable to detect CP below a certain threshold at which the CP-RTD-specific antibodies can still detect CP-RTD. To conclude, while absence of P3a expression had little effect on expression of the other sgRNA1-encoded proteins (CP, P4, CP-RTD), overexpression of P3a from a strong initiation codon inhibited expression from downstream AUG codons much more strongly than in wheat germ extract.
To explore the role of the P3a protein in the viral infection process in planta, we first analyzed the outcome of infection with the TuYV-3a mutants in inoculated leaves of Arabidopsis thaliana. Full-length viral cDNAs containing the different mutations, and driven by the Cauliflower mosaic virus (CaMV) 35S promoter, were agroinfiltrated into A. thaliana leaves. Total RNA was analyzed by northern blot hybridization. To minimize sample variation, agroinfiltrated leaves from three different plants were collected for each time point. Viral RNAs (gRNA and sgRNA1) were detected in TuYV-WT-inoculated leaves at 54 hpi and accumulation reached a maximum at 72 hpi and remained at that level throughout the 138 hour experiment (Fig 5A). Both the P3a-overexpressing TuYV-3aAUG mutant and the null TuYV-3aAGC and TuYV-3a2stop mutants displayed similar replication kinetics to wild type (Fig 5A). Western blot analysis of TuYV-WT-infected samples detected P3a at 114 hpi (Fig 5B). P3a was detected as early as 54 hpi in TuYV-3aAUG-infected leaves and accumulated to much higher levels than in TuYV-WT-infected leaves. The tagged P3a protein from the TuYV-3aFLAG mutant was also detected earlier and accumulated to slightly higher levels (72 hpi) compared to TuYV-WT (Fig 5B). However this observation might be related to a higher epitope accessibility of P3aFLAG compared with the wild type protein. As expected, no P3a was detected in the TuYV-3aAGC-infected leaves (Fig 5B).
We also examined the accumulation levels of the CP, CP-RTD and P4 proteins in the samples inoculated with the different ORF3a mutants. All three proteins were detected by 90 hpi in TuYV-WT-infiltrated leaves (Fig 5C). Leaves infiltrated with the TuYV-3aAUG mutant reproducibly contained no detectable CP, CP-RTD or P4 at any time, even with longer blot exposure (Fig 5B). Conversely, in TuYV-3aAGC-infiltrated leaves, CP and CP-RTD accumulated in higher amounts at early time points compared to the levels in TuYV-WT-infiltrated leaves (Fig 5C). Expression of CP, CP-RTD and P4 was reduced in TuYV-3aFLAG-infiltrated leaves (Fig 5C). Thus, the ORF3a mutations not only affect P3a synthesis, but they also significantly alter accumulation of the other proteins encoded by sgRNA1.
Systemic infection was investigated by analyzing the presence of viral RNA in upper non-inoculated leaves of A. thaliana plants that had been agroinfiltrated with TuYV-WT or with one of the TuYV-3a mutants. While the efficiency of TuYV-WT systemic infection was 92%, all the TuYV-3a mutants (whether over-expressor TuYV-3aAUG, knock-out TuYV-3aAGC and TuYV-3a2stop, or tagged TuYV-3aFLAG) were poorly infectious as shown by the low percentage of infected plants (Table 1). Moreover, accumulation of the viral RNAs in these plants was very low especially for the mutants TuYV-3aAGC, TuYV-3a2stop and TuYV-3aFLAG as shown by northern blot hybridization with one positive sample for each mutant (Fig 6, lanes 12, 15, 17). The stability of the engineered mutations in the viral progeny was investigated by RT-PCR and sequencing. In the two plants out of 29 that became infected with TuYV-3aAUG (e.g. Fig 6, lane 7), the initial AUG mutation had reverted to the WT ACG sequence. No other modifications were found in the entire subgenomic sequence. In the progeny of the three other mutants, TuYV-3aAGC, TuYV-3a2stop and TuYV-3aFLAG, no base changes were observed (3 plants analyzed for TuYV-3aAGC, 2 plants for TuYV-3a2stop and 5 for TuYV-3aFLAG) (Table 1), showing that the modifications introduced in the viral sequence of these three mutants were maintained through viral replication.
We hypothesize that lack of systemic movement of TuYV-3aAGC is due to absence of P3a protein, and that lack of systemic movement of TuYV-3aAUG is due to insufficient translation of the downstream ORFs encoding the proteins CP, CP-RTD and P4. If this is the case, then the two mutant viruses may be able to complement each other to facilitate systemic infection. As we intended to investigate in parallel direct complementation by a P3a-expressing vector that had to be carried out in N. benthamiana, we chose this plant as a common host for both complementation experiments. Ten plants inoculated with TuYV-3aAGC alone triggered local infection but did not develop systemic infection with one exception, which had an extremely low level of viral RNA (Fig 7B, plant #8). This shows that P3a is required for long distance movement of TuYV in N. benthamiana, as well as A. thaliana. Similarly TuYV-3aAUG was able to accumulate only in infiltrated leaves however none of the 10 plants inoculated with TuYV-3aAUG showed viral spread in upper leaves (Fig 7, plants 11–20), as was also observed in A. thaliana. When N. benthamiana plants were co-infiltrated with TuYV-3aAGC and TuYV-3aAUG, eight out of ten co-infiltrated plants showed wild type levels of viral RNA accumulation in the non-inoculated leaves at 21 d.p.i. as shown by northern blot hybridization and real-time RT-PCR (Fig 7, plant numbers #28–37).
In order to identify the nature of the progeny viral genomes that moved systemically and multiplied in the upper leaves, we performed sequence-specific qRT-PCR, using primers designed to detect only WT, only TuYV-3aAUG, or only TuYV-3aAGC mutants (S6 Fig). The WT-specific primer indeed detected only WT RNA in the WT-inoculated plants (four plants tested #22–25) and gave no amplification of RNA in the eight systemically infected plants co-inoculated with TuYV-3aAGC and TuYV-3aAUG (S6 Fig, plants #28–37). Moreover, the TuYV-3aAGC and TuYV-3aAUG primers detected viral RNA only in plants inoculated with those mutants and not in the WT-infected plants. These quantifications were first normalized with a reference gene (GAPDH) whose expression was shown to remain stable upon various viral infections [47] before being normalized with the sample’s value obtained with the common set of primers and finally normalized with a positive sample (taken arbitrarily). Therefore the relative values shown in S6 Fig reflect only presence or absence of the corresponding virus and do not allow comparison of levels of one mutant versus another. The ratio of the two mutant viruses present in the systemic leaves of the co-inoculated plants could be estimated by comparing the cycle threshold (Ct) values of the samples to a calibration curve generated with the corresponding plasmid. A mean value of 2.4 was obtained for the ratio TuYV-3aAGC/TuYV-3aAUG (extreme values: 1.7–3.1). The RT-PCR results revealed that the mutations were maintained, and that the mutants did not revert to wild type. Recombination was not an issue as the mutations changed the same nucleotides (S6A Fig). Thus, both mutant viruses, each individually incapable of systemic infection, can complement each other to move systemically as efficiently as WT virus.
To determine whether the above complementation of TuYV-3aAGC by TuYV-3aAUG is due to the provision of P3a by the latter virus, we tested whether expression of P3a alone is capable of complementing TuYV-3aAGC. Indeed, co-infiltration of leaves with agrobacteria expressing TuYV-3aAGC and agrobacteria containing a plasmid that expresses only P3a driven by the CaMV 35S promoter, yielded significant levels of TuYV-3aAGC RNA in systemic leaves of 10 out of 10 plants at 21 d.p.i. (Fig 7). The RNA levels were generally less than those in the plants complemented with TuYV-3aAUG, but consistently and significantly above levels in systemic leaves of plants inoculated with TuYV-3aAGC alone (Fig 7, plants #38–44). Plasmid expressing P3a with a C-terminally fused green fluorescent protein (P3a-GFP) did not efficiently complement TuYV-3aAGC, except for one plant out of 10 (Fig 7, plant #62). Nevertheless most plants were positive by qRT-PCR, albeit at very low levels (Fig 7B see insert; compare plants #58–67 with mock-inoculated plant #96), suggesting that the GFP fusion reduced function or expression of P3a to levels that did not permit efficient complementation. Transient expression of P3a and P3a-GFP in presence or absence of the viral mutant TuYV-3aAGC was confirmed in infiltrated leaves by western blotting using specific antibodies (S7 Fig). Curiously, as observed by northern blot analysis in A. thaliana plants inoculated with TuYV-3aAGC (Table 1), one plant infiltrated with TuYV-3aAGC alone and one infiltrated with TuYV-3aAGC plus GFP gave positive but weak signals by qRT-PCR analysis in N. benthamiana (Fig 7B, plants # 8 and 81), suggesting the potential for rare escape events. Overall, the data provided in this work strongly support the notion that P3a is necessary for viral systemic infection and that it can facilitate long distance movement when provided in trans.
Because virion formation is a prerequisite to TuYV long-distance movement [48], we investigated the ability of the TuYV-3a mutants to form particles. Immunosorbent electron microscopy (ISEM) performed on purified viral preparations from leaves agroinfiltrated with TuYV-3aAGC, TuYV-3a2stop and TuYV-3aFLAG mutants revealed typical virus particles which did not differ in conformation from WT virions (Fig 8). A few particles were detected on grids from the TuYV-3aAUG mutant. Interestingly, in protoplast infections where only one replication cycle occurs, no particles were observed for this mutant while particles were easily detected for the other mutants (S8 Fig). This suggests that the particles found in TuYV-3aAUG-inoculated leaves may be due to rare reversion events as described earlier (Table 1 and Fig 6); Therefore, the inability of both TuYV-3a knockout mutants to move efficiently to non-inoculated leaves of agroinfected plants cannot be attributed to the absence of capsid formation but rather to the inhibition of another step required for viral long-distance spread. This conclusion is reinforced by the ability of P3a to complement movement of the TuYV-3aAGC mutant. Taken together, these results show that the P3a protein plays a crucial role in systemic infection.
To further address the role of the P3a protein in viral infection, its subcellular localization was observed in epidermal cells of Nicotiana benthamiana. Because P3a contains a putative trans-membrane domain near its N-terminus, whose function might be affected by the fusion with a bulky marker, ORF3a was fused at its 3’ end to a GFP or RFP ORF and expressed under the CaMV 35S promoter. When agroinfiltrated into N. benthamiana leaves, both constructs expressed fusion proteins of the expected size (S9 Fig). Both P3a-GFP and P3a-RFP proteins visualized by confocal laser scanning microscopy showed cytoplasmic punctuate structures (Fig 9A–9D). Co-expression of P3a-GFP and a cis-Golgi marker, α-1,2 mannosidase-1 fused to RFP (Man1-RFP) [49] showed a perfect co-localization of P3a-GFP with Man1-RFP (Fig 9E–9G), suggesting that P3a is associated with individual Golgi bodies. Fluorescent spots were also observed at discrete areas near the cell wall of epidermal cells. To pinpoint the locations of these spots, we co-expressed the P3a-RFP protein with a plasmodesmata marker (plasmodesmata-localized protein-1; PDLP-1) fused to GFP [50] (Fig 9H–9N). Whereas P3a-RFP localized near plasmodesmata, precise co-localization with the PDLP-1-GFP marker was not observed. Higher magnification views of some spots showed that P3a-RFP was adjacent to plasmodesmata, and appeared to remain essentially outside of the cell wall (Fig 9K–9N). To confirm this specific position of P3a, the leaf discs infiltrated with the construct P3a-RFP were stained with aniline blue, a callose marker. Callose is known to be deposited at the neck region of plasmodesmata [51]. Blue staining of callose was observed at potential positions of plasmodesmata in the cell wall while the P3a protein was consistently observed close to the labeled callose but not merged with it (S10 Fig).
By applying bioinformatics tools to genome sequences of luteoviruses and poleroviruses, we have discovered a previously overlooked essential gene, ORF3a, that is conserved throughout the Luteovirus and Polerovirus genera. Translation of ORF3a depends on non-AUG initiation on the sgRNA. Thus, via additional leaky scanning and stop codon readthrough, four distinct proteins (P3a, CP, P4 and CP-RTD) are expressed from a single sgRNA species. This work adds to the increasing known prevalence of non-AUG codons as start codons. While the use of ACG, AUA, AUU and CUG as weak start codons has been known for some time, including in plants [25] and viruses [52–54], this was thought to be a rarity, until many more were revealed by ribosome profiling and bioinformatics approaches [55–57]. This opens up a vast increase of potential coding capacity in viruses and host mRNAs for proteins needed only in small quantities [58].
In addition to containing ORF3a, the sequence of the sgRNA1 5’ end plays a key role in cap-independent translation of BYDV and other viruses in genus Luteovirus. These viruses contain a BYDV-like cap-independent translation element (BTE) in the 3’ UTR which must base pair to a stem-loop at the 5’ end of the mRNA [35] (upstream of ORF3a in sgRNA1) to facilitate translation initiation. In competitive conditions, sgRNA1 of BYDV translates more efficiently than the full-length genomic RNA, and this efficiency is conferred by what was thought to be the 5’ UTR, including ORF3a [59]. This differential translation efficiency was proposed to be due to the relative lack of secondary structure in the sgRNA1 5’ end, but the role of ORF3a in this preferential translation is unknown. This role for the 5’-terminal sequence of sgRNA1 in translation may apply only to genus Luteovirus, because poleroviruses are not known to harbor a 3’ cap-independent translation element. Truncation of the 5’UTR of the PLRV sgRNA1 was reported to increase translation efficiency of CP and P4 [60], an effect that is likely due to the absence of ORF3a and not solely to the shorter 5’ end per se.
Knock-out of ORF3a did not evidently alter the expression level of CP, P4 and CP-RTD in protoplast infections. In leaves infected with the AGC knock-out mutant, the CP and CP-RTD proteins appeared to accumulate slightly earlier relative to WT (Fig 5C) suggesting an influence of ORF3a on translation of the other sgRNA1 ORFs. Conversely, increasing translation of ORF3a dramatically inhibited translation of the other sgRNA1-encoded proteins. These drastic effects were not seen in vitro, most likely due to the more efficient and less competitive translation conditions of the wheat germs extract, where ribosomes are not limiting.
The dispensability of P3a for replication in protoplasts was expected, because previous deletion analysis of infectious clones showed that large deletions that included ORF3a and ORFs 3, 4, and 5 [61], or mutations that prevented sgRNA1 synthesis [32] did not significantly reduce replication of BYDV RNA in protoplasts. Similarly, none of the products of ORFs 3, 4 or 5 of TuYV are needed for RNA replication in protoplasts [46].
Agroinoculation of the two hosts tested, A. thaliana and N. benthamiana, triggered local infection by all TuYV-3a mutants. However systemic infection was very inefficient or nonexistent in these hosts. TuYV particle formation is a prerequisite for long distance trafficking [48]. However viral particles were easily observed in leaves inoculated with these mutants (TuYV-3aAGC and TuYV-3a2stop, Fig 8). This indicates that P3a is not required for the encapsidation process and that the lack of systemic movement is not due to packaging deficiency, thus reinforcing a direct role for P3a in viral systemic spread. The P3a-overexpressing mutant TuYV-3aAUG showed a drastically reduced expression of CP, CP-RTD and P4 that correlates with its incapacity to form virions as shown in Fig 8 and explains its deficiency in systemic movement.
Importantly, the P3a-defective mutant TuYV-3aAGC was capable of long distance trafficking when P3a was supplied in trans by either replicating virus (TuYV-3aAUG) or from a non-replicating plasmid (Fig 7). These results demonstrate unequivocally that the P3a protein is a key factor in long distance movement that functions in trans. This raises the issue of the mode of action of P3a. One hypothesis could be that P3a functions only in the cell where it is expressed and assists the virus in its exit from the infected phloem cells and loading into the phloem. This could be achieved by increasing the size exclusion limit of the specialized plasmodesmata that connect phloem companion cells and the sieve tube, the so-called pore plasmodesmal unit (PPU) [62]. Polerovirus particles have indeed been detected in these PD [15, 63]. In this case, virus accumulation in systemic leaves would result exclusively from replication at primary infection sites. A second hypothesis could be that P3a either facilitates its own movement far beyond the agroinfiltrated cells, and/or that the complemented virus (TuYV-3aAGC) brings the plasmid-expressed P3a protein along with it to the phloem to permit unloading from the sieve element into phloem cells of neighboring leaves.
Bioinformatic predictions highlighted a putative trans-membrane domain which seems in contradiction with this hypothesis, except if during the infection cycle P3a could associate with another factor (i.e. CP or CP-RTD) to move long distance in the phloem. The structural proteins CP and CP-RT* (the encapsidated truncated form of the CP-RTD), and also CP-RTD possibly in its free form, were detected in phloem exudates of CABYV-infected cucumber plants [64]. These proteins might interact with P3a and move in the phloem until they reach, with or without virions, new sites in upper leaves.
Wild-type TuYV generates only minute amounts of P3a from a non-AUG initiation codon (Figs 3, 4, and 5) suggesting that only low amounts of P3a are required. The P3a-defective AGC mutant should therefore not be limited in P3a supply when complemented with the AUG mutant which overproduces P3a (Figs 4 and 5). In contrast, the AUG mutant—defective in CP, CP-RTD and P4 production—requires much larger amounts of these proteins from the complementing AGC mutant. This may explain the skewed ratio TuYV-3aAGC/TuYV-3aAUG in favor of the AGC mutant, with a mean value of 2.4 in double-infected plants.
Sub-cellular localization studies with fluorescent protein fusions in infiltrated N. benthamiana leaves showed that P3a is targeted to the Golgi apparatus, and also close to plasmodesmata (Fig 9). Both subcellular locations are in accordance with a role in viral movement. Thus, we speculate that the inability of P3a-GFP to significantly complement the P3a-deficient mutant TuYV-3aAGC was due to the GFP domain interfering with interactions of P3a with viral components (RNA or protein) necessary for movement, rather than being due to impaired subcellular localization. The specific targeting of P3a-GFP indicates a close association with the host endomembrane network likely through the transmembrane domain predicted in P3a. The ER and the Golgi apparatus constitute the core components of the secretory pathway, suggesting movement processes similar to thoses of other viruses. Small movement proteins of carmoviruses [65, 66] and potyviruses [67–69] also usurp the secretory pathway, and the TGB3 protein of potexviruses drives TGB2 protein-induced vesicles via the ER to form punctate caps on the cytoplasmic orifices of PD, similarly to the P3a protein [70].
Although we have shown function and localization only for TuYV P3a, it is highly likely that P3a of the other poleroviruses and luteoviruses has the same function, given (i) that P3a is required for TuYV movement in both Arabidopsis and N. benthamiana, (ii) the amino acid sequence conservation of P3a among diverse luteovirids (S1 Fig), and (iii) the functional conservation of all the neighboring ORFs on sgRNA1. It is noteworthy that, in addition to P3a, the P4 movement protein of PLRV also localizes to PD [16], facilitated by actin- and ER-Golgi-dependent transport [71]. Ectopically expressed TuYV P4 similarly targets PD but the trafficking pathway has not been studied yet (Julia De Cillia and V.Z-G. personal communication). Like conventional movement proteins MPs, P4 binds single-stranded RNA, dimerizes, is subject to phosphorylation, and increases the PD size exclusion limit [16, 18, 72, 73]. Remarkably P4 was found to be a host-specific movement protein. PLRV and TuYV P4-deficient mutants were reported to spread systemically in some, but not all, hosts [17, 74]. This raises questions such as whether P3a and P4 of TuYV act cooperatively on the same viral entity (virions or ribonucleoprotein complexes, RNP), or whether one protein promotes movement of RNP and the other virions. Another alternative mode of action of both P3a and P4 proteins could be that they function on specific PD of certain phloem cells, at certain development stages or even in specific hosts [17, 71]. Are P3a and P4 proteins specialized for a specific virus transport through “conventional” PD or through PPU? Since we have shown that in the absence of P3a TuYV long-distance transport is impaired, it seems more likely that P3a could play a role in the viral movement across PPUs.
In addition to P4 and P3a, CP and CP-RTD also participate in polerovirus and luteovirus movement. The CP is essential for TuYV long-distance movement through its ability to form particles [48]. CP-RTD occurs in planta in two forms, the full-length protein (the non-structural form) and CP-RT*, which is a C-terminally truncated form incorporated into virions [75, 76]. The N-terminal part of the CP-RTD is required for TuYV to move between nucleated vascular cells [15]. Both CP-RTD and CP-RT* were shown to be required for efficient long distance movement of CABYV [64]. The discovery that the P3a protein is involved in systemic trafficking adds more complexity to this process.
Interestingly, assigned (PEMV1) and putative (Citrus vein enation virus) members of the third Luteoviridae genus, Enamovirus, lack ORF3a. They also lack P4, and the carboxy-terminal half of the readthrough domain, both of which have been implicated in cell-to-cell and systemic movement [17, 75]. Instead, PEMV1 relies on a protein or proteins encoded by an associated umbravirus (PEMV2) for systemic movement in the plant beyond the phloem cells [19] which renders both viruses mechanically transmissible. Apparently, because PEMV has a movement mechanism different from the other Luteoviridae, it does not require P4, the C-terminus of RTD, or P3a, which suggests that P3a may act in concert with P4 and the C-terminus of RTD for virus movement. Further understanding of luteo/poleroviral movement will require us to decipher the precise function and interplay of these multiple viral proteins involved in movement.
Sequences were processed and analyzed using EMBOSS [77], BLAST [78], CLUSTALW [79] and MLOGD [37]. Transmembrane regions were predicted with TMHMM v2.0 [80].
All Luteoviridae sequences available in GenBank as of 16 Nov 2013 were downloaded. Patent sequences were removed. The remaining nucleotide sequences were used to generate a BLAST database [78]. Sequences with coverage of the ORF3/3a region were identified by applying TBLASTN to the NC_004750 (BYDV) P3 amino acid sequence and retaining sequences with ≥75% coverage and ≥30% identity (parameters sufficient to retrieve all luteovirus and polerovirus sequences with ORF3 coverage, as well as enamovirus sequences which were subsequently excluded), and then retaining sequences with sufficient flanking sequence 5' of the ORF3 AUG in order to cover the ORF3a region. In total, 459 luteovirid ORF3a-region sequences were retrieved (see S1 Datafile).
Sequences with complete or near-complete genome coverage were initially selected by taking a length cut-off limit of 5000 nt, followed by semi-automated inspection of ORF lengths and alignments. Sequences which were defective due to obvious disruptions (e.g. premature termination codons or insertions/deletions that disrupted the reading frame) in ORFs 0, 1, 2, 3, 4 or 5 were removed. Sequences were easily clustered into poleroviruses, luteoviruses and enamoviruses based on genome organization and ORF lengths. BYDV sequences (serotypes PAV, MAV, GAV, PAS and KerII) were separated from other luteovirus clades using a P1 phylogenetic tree (CLUSTALW amino acid alignment; CLUSTALX tree).
The full-length BYDV nucleotide sequences were aligned initially with CLUSTALW. For the MLOGD analysis and stop codon plots, to produce meaningful results it is important that sequences are aligned in-frame within coding ORFs. To ensure this, the ORF1-ORF2 and ORF3-ORF5 coding blocks were extracted from the BYDV nucleotide alignment; ORFs 1 and 2 were fused in-frame through the artificial insertion of 'N' at the frameshift site; then each of the two regions was translated, re-aligned as amino acid sequences, back-translated to a nucleotide alignment, the previously inserted 'N's in the ORF1-ORF2 alignment were removed, and the re-aligned regions were reinserted into the full-genome alignment. For the pan-polerovirus alignment, nucleotide sequences were too divergent for an initial full-genome nucleotide alignment to provide a suitable scaffold. Thus, for the polerovirus alignment, untranslated regions were not included in the alignment, except for 178 nt of sequence 5' of ORF3 in order to encompass the ORF3a region. For each genome sequence, the ORF0-ORF1-ORF2 and ORF3a-ORF3-ORF5 regions were extracted and the ORFs in each region were fused in-frame through the artificial insertion of 'NN' before the start of ORF1, 'N' at the ORF1/ORF2 frameshift site, and 'NN' before the start of ORF3 (except for NC_006265 where ORFs 3a and 3 do not overlap). Then each of the two regions was translated, aligned as amino acid sequences, back-translated to a nucleotide alignment, and the previously inserted 'N's were removed.
For the MLOGD analysis and stop codon plots (Fig 1B and 1D), reading frames were defined by mapping sequences onto a specific reference sequence (NC_004750.1 for Fig 1B and NC_003743.1 for Fig 1D). This is important since reading frames are often not preserved in intergenic regions. This 'correction' is relevant to four of the 97 full-genome polerovirus sequences (three detailed in Fig 2 and discussed in text, plus HM439608) where the reading frame of ORF3a is disrupted, besides NC_006265 (Carrot red leaf virus) where the reading frame of ORF3a with respect to that of ORF3 differs from normal due to a 3-nt intergenic region between ORFs 3a and 3.
The BYDV analysis (Fig 1B) is based on GenBank accessions AF218798, AF235167, AJ810418, AY220739, AY610953, AY610954, AY855920, D11028, D11032, D85783, EF043235, EF521828, EF521829, EF521831, EF521832, EF521833, EF521834, EF521835, EF521836, EF521837, EF521838, EF521840, EF521841, EF521842, EF521843, EF521844, EF521845, EF521846, EF521847, EF521849, EF521850, EU332307, EU332308, EU332309, EU332310, EU332311, EU332312, EU332313, EU332314, EU332315, EU332316, EU332317, EU332318, EU332319, EU332320, EU332321, EU332322, EU332323, EU332324, EU332325, EU332326, EU332327, EU332328, EU332329, EU332330, EU332331, EU332332, EU332333, EU332334, EU332335, EU332336, EU402386, EU402387, EU402388, EU402389, EU402390, EU402391, HE985229, KC571999, KC572000, KF523378, KF523379, KF523380, KF523381, KF523382 and X07653. The polerovirus analysis (Fig 1D) is based on GenBank accessions AB594828, AF157029, AF235168, AF352024, AF352025, AF453388, AF453389, AF453390, AF453391, AF453392, AF453393, AF453394, AF473561, AJ249447, AM072750, AM072751, AM072752, AM072753, AM072754, AM072755, AM072756, AY138970, AY695933, AY956384, D00530, D10206, D13953, D13954, DQ132996, EF521827, EF521830, EF521839, EF521848, EF529624, EU000534, EU000535, EU313202, EU636990, EU636991, EU636992, EU717545, EU717546, FM865413, GQ221223, GQ221224, GU167940, GU190159, GU327735, GU570006, GU570007, GU570008, HM439608, HM804471, HM804472, HQ245316, HQ245317, HQ245318, HQ245319, HQ245320, HQ245321, HQ245322, HQ342888, HQ388348, HQ388349, HQ388350, HQ388351, HQ439023, HQ827780, JF507725, JF925152, JF925153, JF925154, JF925155, JF939812, JF939813, JF939814, JQ346189, JQ346190, JQ346191, JQ420901, JQ420902, JQ420903, JQ420904, JQ420905, JQ700305, JQ700306, JQ700307, JQ700308, JQ862472, JX855134, KC121026, KC921392, L25299, X13063, X74789, X76931, X83110 and Y07496.
pUC19 expression vector containing sgRNA cDNA sequence was obtained by cloning the region corresponding to the sgRNA (3259–5641 bases) with the upstream primer containing an XbaI restriction site and T7-promoter sequence (in italics), GAGGTCTAGATAATACGACTCACTATAGGGACACCCGATACCAGGAGAG, and the downstream primer containing an NcoI restriction site, GAGGCCATGGAGTGCCCAACTCTCTTTGG. The mutations were introduced via the QuikChange site-directed mutagenesis procedure (Agilent Technologies) using mutagenic primers for PCR and subsequent DpnI treatment of PCR mixture. Oligonucleotides used for mutagenesis (mutations in bold and italics): for the 3aAUG mutant: CTTAAGCAAACCCAATTAAAGATACAATGGATTACAAATTCCTAGCAGGCTTCGCC and the reverse complement, for the 3aAGC mutant: CTTAAGCAAACCCAATTAAAGATACAAGCGATTACAAATTCCTAGCAGGCTTCGCC and the reverse complement, for 3a2stop mutant: CAGGCTTCGCCGCAGGCTTCGTTTAATAGATACCAATATCCGTGATCAGTATC and the reverse complement, for FLAG tag mutant: CCCAATTAAAGATACAACGGATTACAAAGACGACGACGATAAGTTCCTAGCAGGCTTCGCCGCAGGC and the reverse complement.
In order to obtain the same mutations in the agroinfection vector pBinBW0 [13, 81] containing full-length TuYV cDNA sequence, the SpeI/SalI fragment of pBinBW0 was replaced with the corresponding mutated sequences. All constructs were sequenced to confirm the presence of the mutations. pBin-derived constructs were introduced by electroporation into Agrobacterium tumefaciens strain GV3101 [82].
Capped TuYV sgRNA transcripts were obtained by in vitro transcription using the bacteriophage T7 RNA polymerase and BglII-linearized pUC19 vectors containing WT and mutant sgRNA sequences [83]. Transcripts were translated in wheat germ extracts (Promega) according to the manufacturer's instructions using 80 mM potassium acetate and 0.77 μg of corresponding transcript in 12.5 μl reactions containing the amino acid mix without methionine and 0.6 μCi [35S]-methionine. Reactions were performed for 90 minutes and terminated by addition of an equal volume of 2×SDS-PAGE buffer [84] and incubated at 95°C for 5 min. Samples were run on a Tricine-SDS gel [85] (6% and 16% acrylamide gels for stacking and resolving gels respectively) for 2.5 hours. The gel was washed 3 times for 1 hour and fixed for 16h in 20% ethanol/10% acetic acid solution and then successively washed for 30 min with solutions containing 15%/7,5%/5%, 10%/5%/10%, 5%/2,5%/15% ethanol/acetic acid/PEG550 and finally with 20% PEG 550 for 1 hour to prevent gel cracking (http://sciphu.com/2008/03/use-of-polyethylene-glycol-for-drying-polyacrylamide-gel). The gel was dried for 2 hours at 70°C and exposed either with an X-ray film or with a PhosphoImager screen.
The CP and P4 relative quantities were calculated as areas under the corresponding peaks and normalized to the WT CP or P4 intensities. The quantification was performed using ImageJ software according to the standard procedure of the peak surface measurement (e. g. http://openwetware.org/wiki/Protein_Quantification_Using_ImageJ).
Full-length TuYV RNA transcripts were obtained by in vitro transcription using the T7 RNA polymerase and SalI-linearized pBS vectors containing WT or mutant TuYV cDNA sequences [83]. Capped transcripts were then used to inoculate Chenopodium quinoa protoplasts by electroporation as described previously [83], using 5 μg transcripts for 250,000 protoplasts and a pulse of 180 V. Protoplasts were harvested 44 hours post-inoculation (p.i.), and total proteins or RNAs were extracted as described previously [46, 83].
Agrobacterium tumefaciens GV3101 [86] containing empty pBin19 vector, pBinBW0, derived mutant vectors or protein-expressing vectors were grown for 24 hours, pelleted and incubated in buffer containing 10 mM MES (pH 5,6), 10 mM MgCl2 and 0.15 mM acetosyringone for 2 hours. Agro-infiltration was performed at an OD600 of 0.5 (when mixed infiltrations, OD600 was 0.5 for each culture) to 5-week old A. thaliana plants (ecotype Col0) or to 6-week old N. benthamiana plants. New upper leaves were harvested 3 weeks pi for RNA or protein analysis (100 mg). For infiltrated leaves analysis the samples were collected at indicated time points.
Protoplasts were disrupted by addition of hot 2×SDS-PAGE buffer with subsequent heating at 95°C for 5 min. Plant total proteins were extracted by grinding 100 mg A. thaliana leaves in 250 μl of hot 2×SDS-PAGE buffer [84]. Proteins were separated in 10% or 12% SDS-PAGE gels and transferred onto PVDF Immobilon-P membrane (Millipore). Membranes were then blocked in PBS-Tw 1% buffer (150 mM NaCl, 2.7 mM KCl, 10 mM Na2HPO4, 1.5 mM KH2PO4, 1% Tween-20) with 5% fat-free milk for 2 hours and incubated with specific primary antibodies raised against CP, RTD or P4 proteins [10, 46]. The protein/antibody complex was detected by chemiluminescence (Lumi-LightPLUS kit, Roche).
To immunodetect the 3a protein, protein samples were run on a 16% acrylamide Tricine-SDS gel [85] for 2.5 hours as described above and transferred onto Immobilon-PSQ membrane (Millipore). Membranes were then blocked in PBS-Tw 0.1% buffer with 1% BSA and incubated with primary antibodies raised against the FLAG epitope (Sigma) or a peptide corresponding to the P3a C-terminal 15 amino acids. The protein/antibody complex was detected by chemiluminescence (Lumi-LightPLUS kit, Roche).
RNAs from protoplasts were extracted as described by Veidt et al. [83]. Samples from infiltrated or upper A. thaliana leaves were ground in liquid nitrogen and RNAs were extracted using TriReagent (Sigma-Aldrich) according to the manufacturer instructions. 7.5 μg of RNA extracted from leaves or from 100,000 protoplasts were denatured and fractionated on a 1% formaldehyde-agarose gel [83] and transferred to nitrocellulose (Amersham Hybond-NX, GE Healthcare). Prehybridization was performed at 60°C for 2 h in PerfectHyb Plus buffer (Sigma). The radioactive probe was generated using the Prime-a-Gene labeling system (Promega) and a PCR product corresponding to the 3’-terminal 600 bases of TuYV genome as template. After hybridization and washing, the membrane was exposed onto an X-ray film or a Phosphoimager screen.
2 μg of RNA isolated from upper non-inoculated leaves were used as a template for the reverse transcription reaction using the SuperScript III system (Life Technologies) and a reverse complement oligonucleotide to the last 19 bases of TuYV genomic RNA as a primer. PCR was performed using Qiagen Taq polymerase and the oligonucleotides corresponding to the first and the last 19 bases of TuYV sgRNA1. Purified fragments were thereafter sequenced.
Real-time PCR was performed on cDNA corresponding to 20 ng of total RNA extracted from upper leaves of N. benthamiana plants infiltrated with the various recombinant agrobacteria using a LightCycler 480 II instrument (Roche). The reactions were carried out using the SYBR Green I Master (Roche). In order to distinguish the viral mutants present in the upper leaves infected with the mixture of AUG and AGC mutants, or to verify the progeny in the singly infected plants, four sets of primers were designed: one set of common primers to detect any TuYV RNA (named co-Tu-LP (CCAGGAGAGTAAAGAAGAAGAAAG) and co-Tu-RP (AAGCCTGCTAGGAATTTGTAATC)) and three sets of primers able to recognize specifically the TuYV WT, AUG or AGC mutated sequence (see S6 Fig). The forward oligonucleotide (co-Tu-LP, S6 Fig) located 74 nucleotides upstream of the mutation site was common for all viral RNA and the reverse primers ended precisely at the mutation site so that the last 1 or 2 nucleotides were different in WT, AUG and AGC primers. The specificity of the primers was confirmed with plasmids used for T7-transcription of WT and mutated viral sgRNAs. The N. benthamiana GAPDH gene (JQ256517.1) was used as reference gene. The corresponding forward and reverse primers used are GTGCCAAGAAGGTTGTGATC and CAAGGCAGTTGGTAGTGCAA respectively. We then normalized the values with those obtained with the common TuYV primers and finally for each specific primer set by one of the RNA samples (extracted from plants #22 for TuYV-WT, #28 for TuYV-3aAUG and TuYV-3aAGC). Therefore the values presented in S6 Fig can only be considered as relative and not quantitative values.
Virus particles were purified from 2.5 g of A. thaliana agroinfiltrated leaves using the classical protocol adapted to small volumes [87]. Virions were visualized by ISEM as described by Hipper et al. [48] using a TuYV polyclonal antiserum to capture viral particles on the grids before observation by transmission electron microscopy.
ORF3a was mobilised into pB7FWG2 or pB7RWG2 vectors [88] to obtain GFP- or RFP- fusions, respectively. Transient expression was performed by agroinfiltration on six week-old N. benthamiana using a bacterial OD600 of 0.3. For co-expressions, a 1:1 mixture of the two Agrobacteria transformants was infiltrated. For mRNA stabilization Agrobacteria containing the silencing suppressor P19-encoding vector were used at the final OD600 of 0.1. Confocal observations were performed between 24 and 30 hpi with leaf discs mounted with water and vacuum infiltrated. Confocal microscopy images were obtained with a Zeiss LSM700 or LSM780 inverted confocal laser microscope using a 40×oil immersion objective. The excitation wavelength for GFP and RFP detection was 488 and 561 nm, respectively.
To visualize PD-localized callose, leaf disks were vacuum-infiltrated with aniline blue solution (0.1% aniline blue in 67 mM phosphate buffer pH 8). Leaf disks were incubated in dark at room temperature for 15 minutes before imaging using a Zeiss LSM700/780 laser scanning confocal microscope. The excitation wavelength for aniline blue was 405 nm.
|
10.1371/journal.pgen.1006975 | Isocitrate protects DJ-1 null dopaminergic cells from oxidative stress through NADP+-dependent isocitrate dehydrogenase (IDH) | DJ-1 is one of the causative genes for early onset familiar Parkinson’s disease (PD) and is also considered to influence the pathogenesis of sporadic PD. DJ-1 has various physiological functions which converge on controlling intracellular reactive oxygen species (ROS) levels. In RNA-sequencing analyses searching for novel anti-oxidant genes downstream of DJ-1, a gene encoding NADP+-dependent isocitrate dehydrogenase (IDH), which converts isocitrate into α-ketoglutarate, was detected. Loss of IDH induced hyper-sensitivity to oxidative stress accompanying age-dependent mitochondrial defects and dopaminergic (DA) neuron degeneration in Drosophila, indicating its critical roles in maintaining mitochondrial integrity and DA neuron survival. Further genetic analysis suggested that DJ-1 controls IDH gene expression through nuclear factor-E2-related factor2 (Nrf2). Using Drosophila and mammalian DA models, we found that IDH suppresses intracellular and mitochondrial ROS level and subsequent DA neuron loss downstream of DJ-1. Consistently, trimethyl isocitrate (TIC), a cell permeable isocitrate, protected mammalian DJ-1 null DA cells from oxidative stress in an IDH-dependent manner. These results suggest that isocitrate and its derivatives are novel treatments for PD associated with DJ-1 dysfunction.
| The molecular pathogenesis of Parkinson’s disease (PD) is still elusive even though many causative genes for the disease have been identified. In this study, we demonstrated that isocitrate dehydrogenase (IDH), the enzyme responsible for converting isocitrate into α-ketoglutarate, is critical for the pathogenesis of PD by providing NADPH as a reducing power in the cell. IDH mutant animals showed increased reactive oxygen species (ROS) levels and phenotypes related to PD including dopaminergic (DA) neuron degeneration and locomotor defects. Conversely, elevating IDH function either by overexpression or treating a cell-permeable derivative of isocitrate, trimethyl isocitrate (TIC), made DA cells resist oxidative stress and reduce ROS level, thereby suppressing PD phenotypes induced by DJ-1 mutations. These results demonstrate that IDH protects DA neurons from ROS at the downstream of DJ-1 and cell-permeable isocitrates can be novel treatments for PD.
| Parkinson’s disease (PD) is the second most common neurodegenerative disease and is characterized by typical movement disorders and selective loss of dopaminergic (DA) neurons in the substantia nigra pars compacta (SNpc) [1]. Accumulated evidence has firmly linked the death of these neurons to oxidative stress, the state of imbalance between generation and elimination of reactive oxygen species (ROS) [2]. Postmortem brain analysis showed that markers of oxidative damage to lipids, proteins, and nucleic acids are substantially elevated in the SNpc of PD patients [2]. High levels of somatic mitochondrial DNA (mtDNA) deletion are also found in the SNpc neurons from PD patients [3], suggesting a vicious cycle of oxidative damage to mtDNA and other mitochondrial components, thus increasing ROS production in the course of DA neuron degeneration. The link between oxidative stress and DA neuronal loss is further supported by modeling parkinsonism in various animals using oxidative stress-inducing agents, such as 1-methyl-4-phenyl-1,2,3,6-tetrahydropyridine (MPTP), rotenone, paraquat, and 6-hydroxydopamine (6-OHDA) [4–8]. In addition to PD, other neurodegenerative diseases including Alzheimer’s disease (AD), Huntington’s disease (HD), and amyotrophic lateral sclerosis (ALS) are also associated with oxidative stress, further strengthening the correlation between oxidative stress and neurodegeneration [9]. However, the molecular mechanisms resulting in DA neuron degeneration under oxidative stress have not been fully elucidated.
Although PD mainly occurs in a sporadic manner, it could also occur by monogenic mutations [10]. Because familial forms of PD are often clinically and pathologically indistinguishable from sporadic ones, they are likely to have common pathogenic mechanisms [11]. Moreover, recent genome-wide association studies (GWAS) have revealed variations in several familial PD genes as significant risk factors for the development of sporadic PD [12]. Therefore, investigating how PD gene mutations cause familial PD could potentially reveal the molecular pathogenesis of sporadic PD.
Among PD-linked genes, DJ-1 is most closely associated with oxidative stress [2]. DJ-1 was first identified as an oncogene that transforms mouse NIH3T3 cells in cooperation with activated Ras [13]. Later, Bonifati et al. found that DJ-1 is associated with an autosomal recessive early onset type of familial PD [14]. DJ-1-deficient animal models showed hypersensitivity to oxidative stress [15–18], and further cell biological studies revealed that DJ-1 is a multifunctional protein that participates in transcription regulation, anti-apoptotic signaling, protein stabilization and degradation, and mitochondrial regulation to respond to oxidative stress [19]. DJ-1 is sequentially oxidized on its cysteine residues, and its activity and subcellular localization are regulated by its oxidative status [20–23]. Excessive oxidation of DJ-1 inactivates it, and this oxidized form is observed in the brains of patients with sporadic PD and AD [24, 25], suggesting that DJ-1 participates in the pathogenesis of sporadic PD as well as familial PD. In Drosophila, there are two homologues of mammalian DJ-1; DJ-1α and β. DJ-1α is predominantly expressed in the testes, whereas DJ-1β is expressed throughout the whole body [16, 18, 26], similar to the expression pattern of mammalian DJ-1 [13]. DJ-1β loss-of-function mutants show locomotive dysfunction and loss of DA neurons, resembling the phenotypes seen in PD patients [18, 27].
In this study, we found that DJ-1 is critical for maintaining transcription of NADP+-dependent isocitrate dehydrogenase (IDH) under oxidative stress induced by pesticides like rotenone that have been associated with onset of PD in recent epidemiologic studies [28]. IDH catalyzes decarboxylation of isocitrate into α-ketoglutarate and CO2, and also produces NADPH, which provides a reducing power to antioxidant processes scavenging ROS [29]. Indeed, our Drosophila IDH mutants showed decreased NADPH levels with increased ROS production and hyper-sensitivity to oxidative stress. Moreover, loss of IDH induced age-dependent mitochondrial defects and DA neuron degeneration, very similar to the phenotypes of Drosophila PD models [30]. Consistently, overexpression of IDH in DJ-1 mutants successfully enhanced their survival rates and ameliorated DA neuron loss under oxidative stress. Further genetic analysis revealed that DJ-1 maintains IDH expression by regulating the Kelch-like ECH-associating protein 1 (Keap1)-nuclear factor-E2-related factor2 (Nrf2) pathway. In addition, trimethyl isocitrate (TIC), a cell permeable form of isocitrate, markedly restored oxidative stress-induced decrease of NADPH level and inhibited subsequent cell death in mammalian DA cells with DJ-1 deficiency. These results consistently support that the activity of NADP+-dependent IDH is critical in protecting neurons from oxidative stress and DJ-1 mutation.
To find out a new molecular mechanism in which DJ-1 protects cells from oxidative stress, we treated rotenone, a well-known ROS inducer associated with PD [28], to wild type and DJ-1β-deficient flies, and investigated gene expression in both of them through RNA-sequencing (RNA-seq) analysis. Based on the role of mitochondria as a center for generating and controlling ROS [9], we hypothesized that a ROS controlling protein located in mitochondria would act downstream of DJ-1. We looked over the RNA-seq result and found that oxidation-reduction process gene ontology was the most changed in biological process terms (S1 Table) and oxidoreductase activity gene ontology was the most changed in molecular function terms (S2 Table) between wild type and DJ-1β-deficient flies, consistent with the role of DJ-1 in oxidative stress responses. We further looked into the gene list falling into two groups: oxidation-reduction process (GO: 0016491) and mitochondrion (GO: 0005739). As a result, 173 genes in oxidation-reduction process ontology and 237 genes in mitochondrion ontology were found to be different between wild type and DJ-1ß null mutants in RNA-seq analysis. Interestingly, 34 genes were included in both ontologies (S1A Fig), and most of the genes diminished their mRNA expression in DJ-1β null mutants (S1A and S1B Fig). Among them, 3 genes showed statistically significant difference in false discovery rate (FDR) < 0.05 and satisfied fold change > 1.5 at the same time. As we expected, IDH, which encodes a protein regulating intracellular ROS level by producing NADPH, was one of the 3 genes (Fig 1A and Table 1). In quantitative RT-PCR, IDH gene expression was decreased in DJ-1β null mutants compared to wild type controls under oxidative stress, confirming the RNA-seq data (Fig 1B). The reduction in IDH expression in DJ-1β null mutants was observed in heads, thoraces and abdomens, indicating that it may be not tissue-specific (S1C–S1E Fig).
In mammalian organisms, IDH1 and IDH2 are located in cytosol and mitochondria, respectively, and they are expressed from independent genes [29]. However, in Drosophila, a cytosolic isoform (IDHc) and mitochondrial isoforms (IDHm1 and IDHm2) are expressed from the single gene IDH, although Drosophila IDHs are highly homologous to human counterparts (S2A and S2B Fig). IDHP is a Drosophila mutant line containing a P-element in an exon that is shared by all isoforms of IDH (S2C Fig). IDHP flies successfully developed into adults (Fig 1C), but they showed decreased IDH mRNA expression, IDH enzyme activity, and NADPH/NADP+ ratio compared to wild type controls (Fig 1D–1F). All these phenotypes were rescued in the revertant (RV) generated by precise P-element excision (Fig 1D–1F). These results demonstrated that the P-element insertion successfully inhibits gene expression, activity, and function of IDH in the IDH mutants. In addition, the inserted P-element slightly increased expression of CG17352, a gene located next to IDH (S2C and S2D Fig). However, compared to control flies, CG17352 transgenic flies showed no difference in survival rates under oxidative stress (S2E Fig), ruling out the possibility that this small increase of CG17352 expression affects the oxidative stress-related phenotypes of IDHP mutants that we examined in following experiments.
To understand physiological functions of IDH, we checked the lifespan of IDHP flies that exhibited notably shorter life spans than WT and RV controls (S2F Fig). Furthermore, ROS markers such as dihydroethidium (DHE) and 5-(and-6)-chloromethyl-2′,7′-dichlorodihydrofluorescein diacetate (CM-H2DCFDA) demonstrated that in vivo ROS accumulated highly in IDH mutants (Fig 1G and 1H). This finding was further confirmed by measuring the level of hsp22 mRNA expression which increases according to mitochondrial oxidative stress (S2G Fig) [31]. Consistently, the survival rate of IDHP flies radically declined under rotenone treatments showing that IDH is indispensable for sustaining resistance to oxidative stress (Fig 1I). To further investigate the effect of the in vivo ROS accumulation, the phenotypes of 3- and 30-day-old IDHP adult flies were compared to those of RV control flies without oxidative insults. While ATP level in the muscles, which indicates the function of mitochondria, and mtDNA content, which shows the amount of mitochondria, did not present any significant changes in 3-day-old flies, those of 30-day-old IDHP mutants were markedly dropped compared to RV control flies (S3A–S3D Fig). Based on the fact that oxidative stress and mitochondrial defects are vital factors in onset and deterioration of a variety of human diseases, particularly neurodegenerative diseases including PD [2], the phenotypes related to PD in IDH mutants were analyzed. The results showed that climbing ability and the number of DA neurons of 30-day-old IDH mutants substantially decreased (~30% for climbing ability, ~10% for DA neuron number) compared to those of RV control flies (Fig 1J and 1K, and S3E and S3F Fig). Specifically, among the four major DA neuron clusters [dorsolateral clusters 1 (DL1), dorsomedial clusters (DM), posteriomedial clusters (PM), and dorsolateral clusters 1 (DL2)] [32], only the DA neurons in DL1 and DM clusters were degenerated (Fig 1J and 1K), as previously shown in PINK1 and Parkin mutant flies [32]. These results showed that DA neuronal degeneration induced by mutating IDH gene can be clearly distinguished from non-specific neuronal degeneration caused by simply increasing ROS.
Drosophila phenotypes related to PD symptoms, such as loss of climbing ability and DA neuronal degeneration, were also observed in IDHP flies. Consistent with these results, when IDHm1, IDHm2, and IDHc isoforms were overexpressed in DJ-1β mutant flies, survival rates of DJ-1β mutant flies increased dramatically in a rotenone-induced oxidative stress condition (Fig 2A). IDH overexpression rescued the reduced climbing ability caused by DJ-1β mutation under rotenone treatment (Fig 2B), and suppressed the DA neuronal degeneration caused by rotenone or H2O2 treatment in DJ-1β mutants (Fig 2C, 2D and S4 Fig). Moreover, IDHP mutation decreased survival rates and the number of DA neurons under oxidative stress, but the genetic combining of IDHP mutation did not have an effect on DJ-1β mutants (S5 Fig). From these results, we concluded that IDH is an important antioxidant enzyme that protects cells downstream of DJ-1 under oxidative stress. Especially, expressing IDHm1 and IDHm2 more effectively restored the phenotypes of DJ-1β mutants induced by oxidative stress (Fig 2A–2D and S4 Fig), indicating that mitochondrial IDH plays an important role in DJ-1-regulated cell protection against oxidative stress. Consistent with this idea, in DJ-1β deficient flies stressed with rotenone, the mRNA expression of mitochondrial IDH isoforms (IDHm1 and IDHm2) was severely decreased, but that of cytosolic IDH isoform (IDHc) showed no significant change (Fig 2E–2G). In addition, when we overexpressed IDH genes in PINK1 mutants which showed reduced levels of ATP and mtDNA in muscles, loss of climbing ability, and DA neuronal degeneration in the absence of oxidative stress [32], none of these phenotypes were rescued (S6 Fig). These results suggested that IDH is specifically related to DJ-1 in PD pathology.
The transcription factor Nrf2 induces expression of various antioxidant proteins in response to oxidative stress [33]. Under normal conditions, Nrf2 is degraded by Keap1, an E3 ligase. Upon exposure to oxidative stress, Nrf2 is stabilized and translocated to the nucleus [34–37]. Interestingly, DJ-1 stabilizes Nrf2 by preventing its association with Keap1 and thereby preventing the subsequent degradation [38]. Previous data showed that Nrf2 also controls IDH mRNA expression [39]. Thus, we suspected Nrf2 to be a possible transcription factor that mediates IDH mRNA expression regulated by DJ-1. When we checked mRNA levels of each IDH isoform, the expression levels of IDHm1 and IDHm2 were rescued by Keap1 mutation in DJ-1β null flies under oxidative stress (Fig 3A and 3B). In contrast, Keap1 mutation failed to change IDHc mRNA expression in the stressed DJ-1β mutants, suggesting that DJ-1 specifically induces mitochondrial IDH isoforms via Keap1-Nrf2 pathways (Fig 3C). In addition, Keap1 mutation also increased survival rates of DJ-1β mutants on rotenone- or H2O2-containing media (Fig 3D and 3E). Consistently, Cap’n’Collar C (CncC), a fly orthologue of mammalian Nrf2 [40], could induce IDHm1 mRNA expression (Fig 3F), but failed to change IDHc mRNA expression (Fig 3G). When Keap1 was co-overexpressed, IDHm1 mRNA expression decreased, and increased again with the addition of DJ-1β overexpression (Fig 3F). Moreover, when CncC was overexpressed in DJ-1β mutants, DA neuronal death induced by rotenone or H2O2 was suppressed (Fig 3H–3K). These results confirmed that DJ-1 accelerates mRNA expression of mitochondrial IDH isoforms through the Keap1-Nrf2 pathway to protect the cells from oxidative stress.
In the process of investigating the mechanism of Nrf2-induced IDH expression, we found that a putative Nrf2-binding element, also known as Antioxidant Response Element (ARE), (TGACGGGGC) [33] is located on the promoter region of the IDH gene (S2C Fig), and cloned this IDH promoter region in a luciferase reporter plasmid. We co-transfected CncC cDNA and this ARE reporter plasmid in Drosophila S2 cell line, and found that the promoter activity increased with CncC expression and decreased with ARE mutation (Fig 3L). Therefore, we concluded that DJ-1 controls IDH expression through Nrf2 and ARE in an evolutionarily conserved manner.
In the above experiments, we showed that IDH complements DJ-1β mutation in Drosophila. Using mouse DA cell line SN4741 [41] and CRISPR/Cas9-mediated genome editing, we generated DJ-1 null DA cells to further investigate the relationship between DJ-1β and IDH in mammalian systems (S7A and S7B Fig). When we examined IDH1 and IDH2 mRNA expression in SN4741 cells under oxidative stress, there was no significant difference in IDH1 mRNA expression between wild type and DJ-1 null SN4741 cells with or without H2O2 treatment (Fig 4A). In contrast, IDH2 mRNA expression in wild type SN4741 was elevated with H2O2 treatment, but it was almost nullified in DJ-1 null SN4741 cells (Fig 4B). Moreover, Keap1 knockdown completely rescued IDH2 mRNA expression in H2O2-treated DJ-1 null cells, showing that Nrf2 also mediates between mitochondrial IDH and DJ-1 in mammalian cells (Fig 4B and S7C Fig).
In MTT assays, cell viability increased when IDH1 and IDH2 were expressed in DJ-1 null SN4741 cells under oxidative stress (Fig 4C). We performed both annexin V and propidium iodide (PI) staining to observe cell death, but only PI staining was positive in DJ-1 null SN4741 cells in H2O2-containing media, indicating necrotic cell death (Fig 4D). Overexpression of IDH1 or IDH2 rescued this necrotic cell death upon H2O2 treatment in DJ-1 null SN4741 cells (Fig 4D and 4E). These results in mammalian DA cells were highly similar to the DJ-1 mutant phenotype recovery by IDH overexpression in Drosophila except that IDHc overexpression failed to protect DA neurons in some DA neuron clusters of DJ-1β mutants (Fig 2 and S4 Fig). Moreover, CM-H2DCFDA showed that intracellular ROS level dramatically increased in DJ-1 null SN4741 cells when H2O2 was treated, but IDH1 and IDH2 overexpression significantly decreased the ROS level (Fig 4F and 4G). Furthermore, when mitochondrial ROS level was measured by MitoSOX, mitochondrial ROS drastically increased by oxidative stress in DJ-1 null SN4741. However, the mitochondrial ROS level also decreased with overexpression of IDH1 and IDH2 in DJ-1 null SN4741 (Fig 4H and 4I). These results suggested that the relationship between DJ-1 and IDH genes is evolutionarily well conserved and is important in controlling intracellular ROS level, especially mitochondrial ROS level. In addition, a mitochondrial-specific ROS scavenger, MitoTEMPO, successfully inhibited oxidative stress-induced cell death in DJ-1 null SN4741 cells (Fig 4J), further supporting the roles of DJ-1 in controlling mitochondrial ROS.
To further investigate the antioxidant effect of IDH, we searched for a compound that increases IDH activity, but found no such existing activator. To increase NADPH production by IDH, we tried to deliver isocitrate, a substrate of IDH, into the cells. We decided to attach 3 methyl groups to isocitrate to make it permeable to the cells. When treated with H2O2, NADPH/NADP+ ratio in DJ-1 null SN4741 cells decreased dramatically compared to wild type cells (Fig 5A). This result showed that DJ-1 is indispensable for maintaining NADPH/NADP+ ratio under oxidative stress. Surprisingly, trimethyl isocitrate (TIC) treatment substantially restored NADPH/NADP+ ratio (~40–50%) and cell viability (~70–80%) dose-dependently in DJ-1 null SN4741 cells under oxidative stress (Fig 5A and 5B), suggesting that TIC successfully crossed through the plasma membrane and increased NADPH production under oxidative stress. In an additional experiment using annexin V and PI, PI uptake was decreased when cells were pre-treated with TIC (Fig 5C and 5D). These results indicate that TIC inhibited necrotic cell death caused by H2O2 treatment. When SN4741 cells were stained with ROS indicators, CM-H2DCFDA and MitoSOX, intracellular and mitochondrial ROS levels were increased in H2O2-treated DJ-1 null SN4741 cells. However, when cells were pre-treated with TIC, ROS levels immensely decreased in H2O2-treated DJ-1 null SN4741 cells (Fig 5E–5H). These results implied that TIC raises NADPH/NADP+ ratio which detoxifies ROS and ultimately protects DJ-1 null DA cells from oxidative stress. To see whether IDH is indeed responsible for the cell protecting effect of TIC, we suppressed IDH expression in DJ-1 null SN4741 cells with siRNA technology. We observed that when gene expression of IDH1 and IDH2 was blocked simultaneously, TIC had no effect in protecting cells against oxidative stress, confirming that TIC functions through IDH (Fig 5I).
Recently, it has been reported that oxidative stress and mitochondrial maintenance are highly correlated with PD [2]. In this study, we performed RNA-seq to find a novel gene which provides resistance against oxidative stress at the downstream of DJ-1. We assumed that there are important target genes related to the mitochondrion, which is the main organelle of ROS generation. Through RNA-seq analysis (Fig 1A and S1A and S1B Fig), we found that IDH was a gene included in mitochondrion ontology downstream of DJ-1. Loss-of-function flies for IDH were generated, whose in vivo ROS levels were increased (Fig 1G–1H). These IDH mutants displayed reduced survival rates against oxidative stress and decreased DA neurons and climbing ability (Fig 1I–1K, and S3E and S3F Fig). IDH mutant flies showed decreased ATP generation and mtDNA contents (S3A–S3D Fig) that were also observed in IDH2 knockout mouse [42]. These phenotypes are very similar with the PD-related phenotypes of PINK1, Parkin and DJ-1β mutant flies [18, 27, 32, 43, 44], supporting that IDH is highly implicated in PD pathology. Consistently, under MPTP treatments, IDH2 KO mouse showed more severe PD-related phenotypes than those of wild type controls [45]. More interestingly, IDH overexpression in DJ-1β null mutants rescued the PD-related phenotypes of DJ-1β mutants (Fig 2), but not those of PINK1 null flies (S6 Fig). This specific rescue of the phenotypes in DJ-1β mutants implies that IDH is a unique downstream regulator of DJ-1-dependent and/or oxidative stress-induced PD pathologies.
In searching for the molecular link between DJ-1 and IDH, overexpression of Nrf2, a well-known transcription factor that responds to oxidative stress [33], suppressed the loss of DA neurons in DJ-1β mutants under oxidative stress (Fig 3H–3K). Further genetic analysis demonstrated that Keap1 loss-of-function mutation restores decreased IDH expression and survival rates of DJ-1β mutants under oxidative stresses (Fig 3A–3E), supporting Nrf2 as the molecular mediator that links DJ-1 and IDH under oxidative stress. Interestingly, only the expression of mitochondrial IDH isoforms was suppressed by DJ-1 mutation in Drosophila and mouse SN4741 DA cells under oxidative stress (Figs 2E–2G, 4A and 4B). This specific regulation of mitochondrial IDHs is mediated by the Keap1-Nrf2 pathway in Drosophila (Fig 3A–3C, 3F and 3G) and mammalian system (Fig 4B). When these mitochondrial IDHs were overexpressed, oxidative stress-induced DJ-1 null defects, including DA neuron loss in Drosophila brain, decreased Drosophila survival rates, and increased intracellular and mitochondrial ROS levels in SN4741 DA cells, were successfully rescued (Figs 2 and 4). These data established the role of the DJ-1-Nrf2-IDH pathway in DA neuronal protection against oxidative stress and implied that DJ-1 effectively eliminates intracellular ROS by reducing ROS generation in mitochondria through mitochondrial IDHs.
Although most data converge on this conclusion, there are some results that may raise further questions. In Drosophila and SN4741 cells, cytosolic IDHs can also ameliorate the DJ-1 null defects mentioned above, including increased mitochondrial ROS levels. Is regulating mitochondrial ROS levels indeed important to protect cells? Shin et al. reported that reducing cytosolic ROS level decreases mitochondrial ROS levels [46]. Consistently, in our experiment, overexpression of IDH1, a cytosolic mammalian IDH, substantially reduced mitochondrial ROS accompanying decreased cellular ROS in SN4741 cells (Fig 4I). However, IDH2, the mitochondrial counterpart, reduced more mitochondrial ROS than IDH1 in SN4741 cells (Fig 4I). In Drosophila brains, IDHc failed to protect DA neurons in some DA neuron clusters (Fig 2C, 2D and S4 Fig). These results indicated that indirect elimination of mitochondrial ROS by cytosolic IDHs is not sufficient to protect cells compared to direct elimination via mitochondrial IDHs, especially in DA neurons that may have more complex and varied stresses compared to cultured cells. Furthermore, mitochondria-specific antioxidant MitoTEMPO strongly inhibited cell death induced by oxidative stress in DJ-1 null SN4741 cells, further confirming the importance of mitochondrial ROS reduction in DJ-1-mediated anti-oxidative stress responses (Fig 4J). In addition, although all Drosophila IDHs are encoded from the same gene locus, only the expression of mitochondrial isoforms is regulated by DJ-1 and CncC (Fig 3). In DNA sequence analyses, we found a putative CpG island between the transcription start sites of IDHc and IDHm1 and 2, raising the possibility that DNA methylation inhibits CncC to induce IDHc expression (S2C Fig). However, we could not find the experimental evidence that DNA methylation is involved in the mitochondrial IDH-specific expression, so the molecular mechanism of this expression will be a future topic.
After confirming IDH as a downstream target of DJ-1, we investigated recent issues on IDH research based on our findings. It has been reported that IDH1 and IDH2 neomorphic mutations are prevalent in various cancers [29]. These mutant IDH enzymes convert α-ketoglutarate into D-2-hydroxyglutarate, which promotes tumorigenesis [47]. Therefore, we tested whether these neoenzymes can affect DJ-1 mutant phenotypes. Overexpression of Drosophila IDHm1 R166K and R134Q mutants, which correspond to the cancer-associated human IDH2 R172K and R140Q mutants, respectively [29], failed to increase IDH activity in flies and restore the decreased survival rates and DA neuron numbers of DJ-1β mutants (S8 Fig). This is consistent with the reports that the cancer-associated IDH mutants do not produce, but consume NADPH [47]. In contrast, wild type IDHm1 consistently increased in vivo IDH activity and rescued the PD-related phenotypes (S8 Fig). Thus, these results implicate that the activity of wild type IDH, not the cancer-related one, is linked to DJ-1-associated PD, suggesting that the drugs developed to target cancer-related IDH mutant enzymes are not appropriate to treat DJ-1-associated PD. To overcome this limitation, we hypothesized that excess concentration of cell-permeable isocitrate, the substrate of IDH, would help treat DJ-1-associated PD by raising NADPH/NADP+ ratio to increase the reducing power in the cell. As expected, TIC substantially elevated NAPDH/NADP+ ratio and strongly reduced intracellular and mitochondrial ROS levels in H2O2-treated SN4741 cells (Fig 5A and 5E–5H). TIC also increased survival rate and lowered necrotic cell death against oxidative stress (Fig 5B–5D). IDH expression knockdown inhibited this TIC-mediated cell protection, supporting the idea that TIC protects cells via IDH (Fig 5I). Since TIC treatment significantly increased cell viability in DJ-1 null SN4741 cells, we expected a similar degree of increase in NADPH/NADP+ ratio. However, the observed ratio was less than expected, indicating that NADPH is being rapidly used to resist oxidative stress (Fig 5A). Overall, these results confirmed the protective role of IDH against oxidative stress, and also suggested cell-permeable isocitrates as putative drug candidates for the treatment of DJ-1 deficiency-associated human pathology including PD (Fig 6).
da-GAL4, hs-GAL4, and elav-GAL4 strains were obtained from the Bloomington Stock Center. IDHP mutants (G9298) were obtained from KAIST-GenExel Drosophila library and backcrossed to w1118 controls for 6 generations to remove genetic background effects. The insertion site of the P-element in IDHP is located at +301 of IDHm1 ORF, +244 of IDHm2 ORF, and +205 of IDHc ORF. A revertant (IDHRV) was generated by precise excision of the P-element in IDHP after backcrossing. In DNA sequencing analysis, IDHRV showed a precise excision of the P-element with no insertion or deletion of nucleotides. IDHm1, IDHm2, IDHc, and CG17352 cDNAs were sub-cloned from GH01524, RE70927, AT04910 and GH02239 BDGP cDNA clones, respectively. IDHm1 R134Q and R166K mutant cDNAs were generated by QuikChange™ site-directed mutagenesis kit (Agilent Technologies) using following primer pairs: IDHm1 R134Q F (gcc caa cgg tac cat cca aaa cat ctt ggg agg aac), R134Q R (gtt cct ccc aag atg ttt tgg atg gta ccg ttg ggc), R166K F (gaa gcc tat tgt gat cgg taa aca tgc cca cgc cga tca gt) and R166K R (act gat cgg cgt ggg cat gtt tac cga tca caa tag gct tc). The IDH cDNAs were inserted into the pUAST vector with C-terminal HA-tag and microinjected into w1118 embryos. The CG17352 cDNA was inserted into the pACU2 vector and microinjected into y1 w1118; PBac{y+-attP-3B}VK00001 embryos. PINK1B9, DJ-1βex54 and UAS-DJ-1β flies were generated as previously described [18, 32]. The tyrosine hydroxylase (TH)-GAL4 fly was a gift from Dr. S. Birman. The Keap1EY5, UAS-Keap1 and UAS-CncC lines were provided by Dr. D. Bohmann.
For oxidative stress assay, three or four groups of 3-day-old 30 male flies (n = 90 or 120) were starved for 6 h and transferred to a vial containing a gel of phosphate-buffered saline (PBS), 5% sucrose and an oxidative stress agent (5 mM rotenone or 1% H2O2) as indicated in figure legends. Dead flies were counted at the indicated time points. For life span assay, three or four groups of 30 male flies (n = 90 or 120) were transferred to fresh fly food vials and scored for survival every 3 or 4 days.
To check climbing activity of IDHP mutants, groups of fifteen 3- or 30-day-old males grown on normal media were transferred into climbing ability test vials and incubated for 1 h at room temperature for environmental acclimatization. After tapping the flies down to the bottom, the number of climbing flies was counted for 10 seconds. For each group, ten trials were performed, and the climbing score (percentage ratio of the number of climbed flies against the total number) was obtained. To check climbing ability of IDH expressing DJ-1β mutant males under oxidative stress, groups of 3-day-old 30 males were starved for 6 h and transferred to a vial containing a gel of phosphate-buffered saline (PBS), 5% sucrose and 0.5 mM rotenone. After 4 days, they were re-grouped in size of fifteen and tested according to the procedures above. The average climbing score with standard deviation was calculated for five independent tests.
For mtDNA PCR, total DNA from five thoraces of 3- or 30-day-old male flies was extracted. Then, quantitative real-time PCR was performed as previously described [32]. Genomic DNA levels of rp49 were measured for internal controls. The results were expressed as fold changes relative to the control. For ATP assay, five thoraces from 3-day-old male flies were dissected, and ATP concentration was measured as previously described [32]. The relative ATP level was calculated by dividing the measured ATP concentration by the total protein concentration. Protein concentration was determined by a bicinchoninic acid (BCA) assay (Sigma). In the mtDNA PCR and ATP assay, the average value with standard deviation was obtained from three independent experiments.
S2 cells were cultured and transiently transfected with IDHm1, IDHm2, and IDHc plasmids used to generate UAS-IDH flies as described previously [48]. To induce IDH protein expression in pUAST vector, we co-transfected pMT-GAL4 plasmids that contained GAL4 gene with metallothionein promoter. Twenty-four hours before cell staining, CuSO4 was treated to induce expression of GAL4 and IDHs. Cells were pre-incubated with 5 μg/mL MitoTracker Red CMXRos (Molecular Probes) for 1 h at 25°C and then subjected to the standard immunocytochemistry using anti-HA antibody (Invitrogen).
To check the change of the DA neuron numbers in DJ-1β mutants under oxidative stress, 30 male flies (3-day-old) were starved for 6 h and incubated for 3 days in a vial containing a gel of phosphate-buffered saline (PBS), 5% sucrose and an oxidative stress agent (0.2 mM rotenone or 1% H2O2). To check the change of the DA neuron numbers in IDH or PINK1 mutants without oxidative insults, 30 male flies were transferred to a fresh normal media vial every 3 or 4 days for the time points indicated in figure legends. To stain DA neurons, adult brains from ten randomly chosen flies were fixed with 4% paraformaldehyde and stained with anti-TH rabbit antibody (1:50, Pel-Freez, P40101-150) as previously described [32]. Brains were observed and imaged by LSM 700 confocal microscope (Zeiss). For imaging ROS production in fly tissues, the indirect flight muscles from 3-day-old males were dissected in Schneider’s medium (Sigma) and incubated for 5 min in Schneider’s medium containing 30 μM dihydroethidium (DHE, Invitrogen). Muscles were observed and imaged by BX-50 microscope (Olympus).
IDHRV (IDHRV/IDHRV); IDHP (IDHP/IDHP); hs (hs-GAL4/+); hs>CG17352 (hs-GAL4/UAS-CG17352); hs DJ-1βex54 (hs-GAL4/+; DJ-1βex54/DJ-1βex54); hs>IDHm1 DJ-1βex54 (hs-GAL4/UAS-IDHm1; DJ-1βex54/DJ-1βex54); hs>IDHm2 DJ-1βex54 (hs-GAL4/UAS-IDHm2; DJ-1βex54/DJ-1βex54); hs>IDHc DJ-1βex54 (hs-GAL4/UAS-IDHc; DJ-1βex54/DJ-1βex54); elav (elav-GAL4/+); elav DJ-1βex54 (elav-GAL4/+; DJ-1βex54/DJ-1βex54); elav>IDHm1 DJ-1βex54 (elav-GAL4/UAS-IDHm1; DJ-1βex54/DJ-1βex54); elav>IDHm2 DJ-1βex54 (elav-GAL4/UAS-IDHm2; DJ-1βex54/DJ-1βex54); elav>IDHc DJ-1βex54 (elav-GAL4/UAS-IDHc; DJ-1βex54/DJ-1βex54); DJ-1βex54 (DJ-1βex54/DJ-1βex54); Keap1EY5/+ (Keap1EY5/+); DJ-1βex54 Keap1EY5/+ (DJ-1βex54Keap1EY5/DJ-1βex54); WT (+/Y); elav>CncC DJ-1βex54 (elav-GAL4/UAS-CncC; DJ-1βex54/DJ-1βex54); hs>CncC (hs-GAL4 UAS-CncC/+); hs>CncC Keap1 (hs-GAL4 UAS-CncC/UAS-Keap1); hs>CncC Keap1 DJ-1β (hs-GAL4 UAS-CncC/UAS-Keap1; UAS-DJ-1β/+); IDHP DJ-1βex54 (IDHP/IDHP; DJ-1βex54/DJ-1βex54); da (da-GAL4/+); B9 da (PINK1B9/Y;; da-GAL4/+); B9 da>IDHm1 (PINK1B9/Y; UAS-IDHm1/+; da-GAL4/+); B9 da>IDHc (PINK1B9/Y; UAS-IDHc/+; da-GAL4/+); TH (TH-GAL4/+); B9 TH (PINK1B9/Y;; TH-GAL4/+); B9 TH>IDHm1 (PINK1B9/Y; UAS-IDHm1/+; TH-GAL4/+); B9 TH>IDHc (PINK1B9/Y; UAS-IDHc/+; TH-GAL4/+); hs>IDHm1 (hs-GAL4/UAS-IDHm1); hs>IDHm1RQ (hs-GAL4/UAS-IDHm1R134Q); hs>IDHm1RK (hs-GAL4/UAS-IDHm1R166K); hs>IDHm1RQ DJ-1βex54 (hs-GAL4/UAS-IDHm1R134Q; DJ-1βex54/DJ-1βex54); hs>IDHm1RK DJ-1βex54 (hs-GAL4/UAS-IDHm1R166K; DJ-1βex54/DJ-1βex54); elav>IDHm1RQ DJ-1βex54 (elav-GAL4/UAS-IDHm1R134Q; DJ-1βex54/DJ-1βex54); elav>IDHm1RK DJ-1βex54 (elav-GAL4/UAS-IDHm1R166K; DJ-1βex54/DJ-1βex54).
To measure transactivation activity of CncC on the IDH gene, the promoter and 5’ untranslated region (S1C Fig) were subcloned into pGL3 reporter plasmid (Promega) using following primers: IDH promoter F (gcg ggt acc cag tta ttc gct gcg tct gat tgg) and IDH promoter R (gcg gga tcc gaa ccg acc gac gac tgg aaa cg). For generating the IDH reporter with ARE mutation, the first five bases (TGACG) of the putative ARE (TGACGGGGC) were deleted by QuikChange™ site directed mutagenesis kit (Agilent Technologies). S2 cells were transfected with wild type or ARE mutant IDH reporter, pUAST-CncC, pRL-TK Renilla reporter, and pMT-GAL4 plasmids. Two days later, CncC expression was induced by CuSO4 treatment. After 24 h, luciferase assays were performed using Dual-Luciferase™ reporter assay kit (Promega) according to the manufacturer's instructions. The average luciferase activity with standard deviation was obtained from three independent experiments.
Total RNA from heads, thoraces, abdomens, or whole bodies of 3-day-old flies or SN4741 cells was extracted and reversely transcribed as previously described [49]. To check the inhibition of IDH expression in IDHP mutants, 5 whole bodies were used (Fig 1D and S2D Fig). To check the expression change of IDH and its isoforms in DJ-1β mutants, 5 heads and, thoraces, reported to be predominantly damaged in PD-gene-defected flies were used (Figs 1D and 2E–2G) [32]. To check whether the expression change of IDH is tissue-specific, 10 heads, 10 thoraces, or 10 abdomens were used (S1C–S1E Fig). To confirm the gene expression of each isoform, 5 whole bodies were used (Fig 3A–3C, 3F and 3G). SN4741 cells were seeded in 6-well plates at a density of 1 × 106 cells per well. Then, quantitative real-time PCR was performed using SYBR Premix Ex Taq (Takara) on Prism 7000 Real-Time PCR System (ABI). rp49 levels or mouse actin levels were measured for internal control of Drosophila or SN4741 samples, respectively. The results were expressed as fold changes relative to the control. The average mRNA level with standard deviation was obtained from three independent experiments. For primer pairs, we used rp49-F (gct tca aga tga cca tcc gcc c) and rp49-R (ggt gcg ctt gtt cga tcc gta ac), IDH-F (cct tcc tgg aca ttg agc tg) and IDH-R (gta ccg ttg ggc gac ttc cac), CG17352-F (cac atc tcg ttg aga gtg gat gac) and CG17352-R (cga atg tag tag cca ttg agg atg), hsp22-F (gtc ctg acc atc agt gtg c) and hsp22-R (cca gtc tgc tcg atg gtc ac), IDHm1-F (cat cag cgc cgc gat gg) and IDH-R (gta ccg ttg ggc gac ttc cac), IDHm2-F (gtg agc gag atg gcc cag aag) and IDH-R (gta ccg ttg ggc gac ttc cac), IDHc-F (gta tgc tct ccc gaa cag atg g) and IDH-R (gta ccg ttg ggc gac ttc cac), mouse IDH1-F (cct ggg cct gga aaa gta ga) and mouse IDH1-R (tcc tgg ttg tac atg ccc at), mouse IDH2-F (cta tga cgg gcg ttt caa gg) and IDH2-R (cct tga gcc agg atg tca ga), mouse actin-F (ttc ttt gca gct cct tcg tt) and mouse actin-R (tgg atg gct acg tac atg gc), and mouse Keap1-F (tgc ccc tgt ggt caa agt g) and mouse Keap1-R (ggt tcg gtt acc gtc ctg c).
SN4741 cells were established from the substantia nigra region of wild type and DJ-1 knock out mouse embryos, and were characterized for expression of the neuronal markers including TuJ1 and NeuN, and the DA cell marker TH as previously described [41]. The SN4741 cells were grown in RF medium (DMEM supplemented with 10% fetal bovine serum, 1% glucose, and 2 mM L-glutamine) at 33°C in a humidified atmosphere with 5% CO2. pCMV14 vector, pCMV14 FLAG-IDH1, or pCMV14-IDH2 was transfected using Lipofectamine Plus Reagent (Invitrogen) according to the manufacture’s protocol. siRNAs for control (Bioneer, #SN-1003), mouse IDH1 (Bioneer, #1371568), mouse IDH2 (Bioneer, #1371576), or mouse Keap1 (Bioneer, #1367293) was transfected to SN4741 cells using the RNAiMAX reagent (Invitrogen) according to the manufacture’s protocol.
CRISPR genome editing technique was used for the deletion of DJ-1. The guide RNA sequence (gtg gat gtc atg cgg cga gc) was cloned into the px459 vector. The plasmid was transfected into SN4741 cells. 48 h after transfection, transfected cells were selected by 5 μg/mL puromycin for 3 days and then single colony was transferred onto 96-well plates with one colony in each well. The clones were screened by immunoblot with anti-DJ-1 antibody (1:1,000, Novus Biology, #NB100-483).
For detection of IDH1, IDH2, β-tubulin, and HA- or FLAG-tagged protein, S2 or SN4741 cells were lysed with Lysis Buffer [48]. The lysates were purified by centrifugation and boiled in SDS sample buffer. The samples were subjected to SDS-PAGE and proteins were transferred to nitrocellulose membrane. The membrane was incubated for 30 min in Blocking Solution and further incubated with anti-IDH1 antibody (1:1,000, Bethyl, #A304-162A-T), anti-IDH2 antibody (1:1,000, Bethyl, #A304-096A-T), anti-FLAG antibody (1:1,000, MBL, #M185-3L), anti-HA antibody (1:1,000, Invitrogen, #26183), anti-DJ-1 antibody (1:1,000, Novus Biology, #NB100-483), or anti-β-tubulin antibody (1:1,000, DSHB, Clone E7) as described previously [48]. Membrane-bound antibodies were detected with ImageQuant LAS 4000 system (GE Healthcare Life Sciences).
Cells were seeded in 12-well plates at a density of 6 × 105 cells per well. After pre-treatment of TIC (LegoChem Biosciences) or MitoTEMPO (Sigma, 10 nM) at the indicated concentrations for 1 h, cells were treated with 1.5 mM H2O2. After 6 h incubation, the culture medium was removed and replaced with a medium containing 0.5 mg/mL of MTT dissolved in PBS (pH 7.2). After 4 h, the formed formazan crystals were dissolved in 400 μL of DMSO, and the absorbance intensity was measured at a wavelength of 595 nm using Infinite 200 pro (TECAN). The relative cell viability was expressed as a percentage relative to the untreated control cells. The average viability with standard deviation was obtained from three independent experiments.
SN4741 cells were seeded in 6 well plates with cell density of 1 × 106 cells per well. Treatment of TIC (5 mM) and H2O2 (1.5 mM) was performed as described above. The cells were stained using the Annexin V-FITC Apoptosis Detection kit (BD Biosciences) according to the manufacturer's protocol. Stained cells were analyzed by flow cytometry using BD FACSCanto II (BD sciences). A total of 10,000 events was analyzed for each sample, and the necrotic cell death rates obtained from three independent experiments were presented as the mean values with standard deviations.
SN4741 cells were pre-treated with TIC (5 mM) for 1 h. Following 2 h treatment of 1.5 mM H2O2, cells were incubated with 5 μM of 5- and 6-chloromethyl-2′,7′-dichlorodihydrofluorescein diacetate (CM-H2DCFDA, Invitrogen) for 30 min at 33°C. The cells were trypsinized, washed with PBS, suspended in PBS, and analyzed with BD FACSCanto II (BD sciences). A total of more than 5,000 events was analyzed for each sample, and the results obtained from three independent experiments were presented as the mean values with standard deviations.
SN4741 cells were pre-treated with TIC (5 mM) for 1 h. Following 2 h treatment of 1.5 mM H2O2, cells were incubated with 1 μM MitoSOX (Invitrogen) for 10 min at 33°C. The cells were trypsinized, washed with PBS, suspended in PBS, and analyzed with BD FACSCanto II (BD sciences). A total of more than 5,000 events was analyzed for each sample, and the results obtained from three independent experiments were presented as the mean values with standard deviations.
SN4741 cells were pre-treated with TIC (5 mM) for 1 h. Following 6 h treatment of 1.5 mM H2O2, the cells were lysed with 0.2 N NaOH with 1% dodecyl trimethyl ammonium bromide (DTAB, Sigma). To measure NADPH/NADP+ ratio in flies, five 3-day-old male flies were homogenized in 0.2 N NaOH with 1% DTAB. Samples were centrifuged to obtain supernatants. NADP+ and NADPH levels of the lysates were individually measured by using NADP/NADPH-glo™ assay kit (Promega) according to the manufacturer's instructions, and NADPH/NADP+ ratio was calculated. The average NADPH/NADP+ ratio with standard deviation was obtained from three independent experiments.
Ten 3-day-old male flies were homogenized in 40 mM Tris buffer (pH 7.4). Supernatants from samples were each added to the Tris buffer-containing NADP+ (2 mM), MgCl2 (2 mM), and isocitrate (5 mM). IDH activity was determined by monitoring the kinetics of NADPH production at 340 nm at 25°C with SpectraMax M2 multi-mode microplate reader (Molecular Devices). The average relative IDH activity with standard deviation was obtained from three independent experiments.
For quantification of DA neurons, four major DA neuron clusters from more than 15 brains of each genotype were observed in a blind fashion to eliminate bias (n = 30~40). To compare three or more groups, we used one-way ANOVA with Sidak correction. For two-group comparison, we used Student’s two-tailed t test. The Kaplan-Meier estimator and the log-rank test were conducted on the survival data to determine whether each treatment had any effect on the longevity of individuals using Online Application Survival Analysis Lifespan Assays (http://sbi.postech.ac.kr/oasis). All n values defined in the figure legends refer to biological replicates unless otherwise indicated. The experiments were not randomized. To obtain consistent results, we incubated flies for at least three days after eclosion and excluded dead or malformed flies before any fly assay in this report.
20 male flies (3-day-old) were starved for 6 h and transferred to a vial containing a gel of PBS, 5% sucrose and 5 mM rotenone. 16 h later, total RNA from ten heads and thoraces of ten randomly chosen stressed flies was extracted. Indexed RNA-seq libraries were constructed using Illumina TruSeq RNA Sample Prep Kit version 2. Each library was sequenced in paired end using Illumina HiSeq2500 platform. Raw reads (n = 3) were aligned to the Ensembl Drosophila melanogaster reference genome (BDGP6) using Tophat2. The read alignments were assembled into transcriptome assembly. Fragments per kilobase of transcripts per million reads (FPKM) as normalized expression levels were calculated using Cufflinks. The assemblies for each replicate were merged together using Cuffmerge. Differentially expressed gene (DEG) analysis was performed using Cuffdiff workflow to screen DEGs with false discovery rate (FDR) adjusted by P-value of < 0.05 and fold change of > 1.5. Gene ontology (GO) analysis was performed for term enrichment using g:Profiler and Amigo2. We filtered GO tree hierarchy and statistical significance threshold was FDR < 0.05. A volcano plot and hierarchical clustering in a heat map were generated by statistical package R.
|
10.1371/journal.pgen.1005524 | Ernst Rüdin’s Unpublished 1922-1925 Study “Inheritance of Manic-Depressive Insanity”: Genetic Research Findings Subordinated to Eugenic Ideology | In the early 20th century, there were few therapeutic options for mental illness and asylum numbers were rising. This pessimistic outlook favoured the rise of the eugenics movement. Heredity was assumed to be the principal cause of mental illness. Politicians, scientists and clinicians in North America and Europe called for compulsory sterilisation of the mentally ill. Psychiatric genetic research aimed to prove a Mendelian mode of inheritance as a scientific justification for these measures. Ernst Rüdin’s seminal 1916 epidemiological study on inheritance of dementia praecox featured large, systematically ascertained samples and statistical analyses. Rüdin’s 1922–1925 study on the inheritance of “manic-depressive insanity” was completed in manuscript form, but never published. It failed to prove a pattern of Mendelian inheritance, counter to the tenets of eugenics of which Rüdin was a prominent proponent. It appears he withheld the study from publication, unable to reconcile this contradiction, thus subordinating his carefully derived scientific findings to his ideological preoccupations. Instead, Rüdin continued to promote prevention of assumed hereditary mental illnesses by prohibition of marriage or sterilisation and was influential in the introduction by the National Socialist regime of the 1933 “Law for the Prevention of Hereditarily Diseased Offspring” (Gesetz zur Verhütung erbkranken Nachwuchses).
| Psychiatric genetics was established as a scientific discipline in the early 20th century. The intention was to provide evidence for the inheritance of mental illnesses. This was to open up new paths in prevention, since therapeutic options at the time were meagre. Applying modern study designs and statistical methods, the German psychiatrist Ernst Rüdin (1874–1952) found lower inheritance of affective disorders than anticipated. He surmised that external factors were also important in their development. From the vantage point of the present, his results were sound. However, this major study, titled “On the Inheritance of Manic-Depressive Insanity”, though developed as a manuscript, was never published. The question arises whether Rüdin doubted his own scientific methods and hence results, or if he was influenced by his ideological views and the expectations of the scientific community and German society. Whilst withholding his results, he continued to promote prevention of assumed hereditary mental illnesses by prohibition of marriage or sterilisation.
| Formal genetic psychiatry was established in the early 20th century at a time of therapeutic nihilism, when psychiatry was strongly influenced by the theory of hereditary degeneration [1–7]. This combined with economic considerations, led to a change of emphasis from healing to prevention of the more severe types of psychiatric illness [1, 3, 8]. This was reflected in the popularity of the eugenics movement, which advocated measures such as restrictions on marriage or sterilisation to prevent inherited disease [3, 9]. A key aim of psychiatric genetic research was to provide scientific evidence that severe mental illnesses were inherited, thus strengthening the case for eugenic measures.
Ernst Rüdin (1874–1952) [11] has been credited as the originator of modern psychiatric human genetics on the basis of his research aiming to establish inheritance estimates, that is the risk of a relative developing the same disease as the index patient, which he termed the “empirische Erbprognose” (“empirical heredity prognosis”) [8, 10].
During his formative school and university years, Rüdin’s worldview was heavily influenced by his brother-in-law, the eugenicist Alfred Ploetz (1860–1940), and by Auguste Forel’s (1848–1931) abstinence movement [8] (for detailed biographical information see Table 1).
Rüdin advocated the goals of eugenics early in his career and called, in numerous publications and reviews, for prohibition of marriage and sterilisation of “inferior persons” such as mentally ill and handicapped persons, criminals, alcoholics and prostitutes [15, 19–21].
Rüdin formulated the primary goal of his “psychiatric-genealogical” research, namely to create a scientific basis for racial hygiene measures, thereby solving the “social issue” (soziale Frage, social ills of industrialised societies in 19th century Europe [22]). Rüdin, who had worked under Emil Kraepelin (1856–1926) in Heidelberg and Munich, adopted his psychiatric nosology and was influenced by Kraepelin’s belief in a strong hereditary component [8, 15, 23, 24].
Rüdin’s research methodology was based on systematic family studies with large sample sizes, using statistical analyses such as Wilhelm Weinberg’s (1862–1937) age correction, simple sib and proband method [25], to identify Mendelian rules of inheritance, and to generate estimates of genetic risk on this basis [26]. Rüdin applied this methodology for the first time at The German Research Institute of Psychiatry (Deutsche Forschungsanstalt für Psychiatrie, DFA) in Munich in his 1916 study “On the Problem of the Inheritance and Onset of Dementia Praecox”, which was the foundation of his reputation as a psychiatric geneticist [24, 27]. 701 families of patients with “dementia praecox” (a group of endogenous psychoses [4, 23]) were examined in the study. From his results, Rüdin derived the hypothesis of a two-locus recessive model, though he did not obtain conclusive proof of recessive inheritance.
At the DFA, Rüdin then worked on his second major inheritance study, on the genetic inheritance of “manic-depressive insanity” (Zur Vererbung des manisch-depressiven Irreseins) [28]. The original manuscript in the Historical Archives of the Max Planck Institute of Psychiatry in Munich is undated; on the basis of the publications cited the compilation can be dated to the period 1922–1925 [8]. It was never published despite being intended for chapter 4 of “Studies on the Inheritance and Origin of Mental Illness”, one of a series of monographs from the publisher Springer-Verlag of Berlin, covering the full domain of neurology and psychiatry [29]. The Rüdin biographer and psychiatrist Matthias Weber described the study as “the most comprehensive and probably most significant of Rüdin’s works” [8].
The manuscript consists of approximately 160 pages of unbound typescript, each chapter paginated separately, with Rüdin’s hand-written corrections, and numerous large-format, hand-written charts [28]. Some chapters are incomplete and untitled and are not considered in this paper. An additional 250 pages of Rüdin’s hand-written notes, unpaginated, unsystematised and partially in shorthand are mostly illegible. The bibliography is hand-written and only partially intelligible.
Rüdin characterised the study as a complement to his 1916 dementia praecox study and used the same methodology [24]. Whilst not excluding a role for environmental factors, he assumed inheritance of a disposition to affective disorders as generally acknowledged but unproven, citing Kraepelin [30]. Rüdin predicted Mendelian dominant inheritance of affective disorders [26]. He intended to prove this on the basis of the segregation pattern in the families of his probands, or determine another mode of Mendelian inheritance. Criteria could then be derived for selection of persons for the eugenic measures he had formulated earlier [20].
The inclusion criteria encompassed all patients admitted as inpatients to the Psychiatric University Clinic of Munich (Psychiatrische Universitätsklinik München) with a diagnosis of “manic-depressive insanity” since 1904. No preference was given to patients with a dense family history; the diagnosis alone was the criterion for inclusion. This type of systematic ascertainment was unusual in psychiatry at a time when spectacular single case studies or densely affected families were the main focus of research [31, 32]. Follow-ups were conducted to evaluate the course of the illness and verify the diagnosis.
Rüdin sought accurate clinical diagnostics for inclusion and exclusion of probands, using Kraepelinian nosology, which was widely recognised though not without controversy [33, 34]. Under the diagnostic rubric of “manic-depressive insanity”, Kraepelin included “periodic and circular insanity”, “simple mania”, “melancholy” and “amentia” [30]. This corresponds approximately to the modern-day categorisation of affective disorders [35–38].
For the diagnosis Rüdin depended above all on the clinical symptoms and less on the course of the disease. Kraepelin regarded “manic-depressive insanity” as “curable” and dementia praecox as “incurable”. Rüdin emphasised however that cases of both disorders were documented that did not follow this rule. He called upon a second clinician with no knowledge of the family history to verify the diagnoses [28]. The sample were hospitalised patients, thus biasing to more severe cases [15]. The patients and, whenever possible, family members were interviewed; further sources included medical and administrative files. However, when a medical diagnostic interview could not be arranged for family members, these diagnoses were often based on descriptions of maladjusted or otherwise conspicuous family members, how often this occurred is not recorded. The most important information (age, diagnosis, medical history and family history) was noted on standard cards, as implemented by Kraepelin, and in detailed family trees [24].
Rüdin integrated Weinberg’s statistical methods and Mendelian genetics into a research methodology which sought to predict the passing on of mental illnesses within families: the “empirical heredity prognosis”. His second goal was to prove a Mendelian mode of inheritance via the systematic ascertainment of as many patients with affective disorders as possible. Rüdin recognised the problem that in recessive modes of inheritance, only families in which the disposition was expressed could be counted, while families with heterozygote carriers were lost. In order to be able to apply Mendelian rules to a sample, the ascertained morbidities would also ideally have to correspond to those in the population. Rüdin applied Weinberg’s simple sib and proband method to that end [25]. He stayed in close contact with Weinberg during the evaluation of his results; their correspondence is cited in several places in the manuscript.
The overrepresentation of patients with affective disorders in the sample was corrected by the simple sib method, whereby index cases, referred to as “probands” (Probanden), were excluded from the calculation. Affected siblings were entered into the calculation as “secondary cases” (Sekundärfälle). The proband method was used to correct for multiple ascertainment in families with several probands.
Rüdin also applied two different methods for age correction, allowing for the possibility that a family member who was healthy at interview could become ill later [28, 39]. On the basis of age distribution, Rüdin determined the period of risk for the onset of affective disorders to be between the ages of 14 and 68. The number of “lifetimes at risk” (Bezugsziffer) was then calculated from the sum of siblings, corrected for their age (Table 2).
Rüdin’s “empirical heredity prognosis” was based on the calculation of a predictive risk of illness, the “morbid risk” (Morbiditätsrisiko), which included the above-mentioned Weinberg methods. The morbid risk was calculated from the number of affected persons in the sample, corrected by the simple sib and proband methods, divided by the number of lifetimes at risk, that were calculated by the age correction method (Table 2).
Calculating the morbid risk, Rüdin assumed full penetrance for the inherited traits in question and excluded a possible influence of external factors or interference with other genes. The thus calculated morbid risk was compared with the proportions expected from a Mendelian crossing in order to prove Mendelian inheritance and thereby the inheritance of affective disorders as such. Rüdin also used morbid risk as a predictive value for any particular person with certain preconditions (e.g. a parent with an affective disorder) to develop an affective disorder at some point in their life, and therefore serve as a kind of “genetic counseling”, or as he called it, an “empirical heredity prognosis” [28].
Rüdin’s use of statistical processes was a groundbreaking approach at the beginning of the 20th century and led to sound results that in part remain valid today [32, 40]. However, while focussing on the methods of Weinberg, who was also chairman of the “Society for Racial Hygiene” and, like Rüdin, a staff member of the “Archive of Racial and Social Biology”, Rüdin ignored some other statistical methods, such as correlational analyses, which were in common use [15].
After excluding unconfirmed diagnoses, the sample comprised 661 probands from 650 families, or “sibships”, comprised of 4351 siblings in total (Fig 1). In 566 families both parents were healthy and in 84 families one parent had an affective disorder (Table 3).
Rüdin analysed these two groups independently (Table 4). Consideration was given to other aspects such as alcoholism in the parents or mental illness in other relatives, in which case the families’ standard cards were re-sorted according to the question posed, and the proportions recalculated.
The morbid risk was then calculated for various categories, and Rüdin found an increase of affective disorders in children of parents who had affective disorders. Contrary to the widely held assumption of a high inheritance rate of affective disorders [30), Rüdin calculated a substantially lower proportion with 7.43% affected children of healthy parents and 23.82% affected children of an affected parent. For an overview of Rudin’s study results see Table 5.
Weinberg calculated the proportions expected from crossing different genetic strains, which were then compared with the morbid risks determined by Rüdin in order to find as close a match as possible. Because a simple recessive or dominant inheritance had to be excluded, Rüdin, with the aid of Weinberg’s calculations, took more and more complicated modes of Mendelian inheritance into consideration. Instead of looking more closely at other factors such as patients’ living conditions, comorbidity or other external influences, calculations with assumptions of up to 12 interacting alleles were made in order to come as close as possible to the expected proportions. With the assumption of a three-locus model with two recessive and one dominant factor, Rüdin’s proportions ultimately best matched those calculated by Weinberg, which is why he postulated this mode of inheritance as the most probable (Table 6). However, implementation of the “empirical heredity prognosis” was unsuccessful; it was still not possible to use a standardised formula to predict the probability that a particular individual would develop an illness. Hence the way one should conceive the genotype of a certain patient was still far from clear.
Rüdin emphasised the preliminarity of his results and called for further studies with larger sample size and avoidance of assortative mating. In order to better measure the effects of environmental factors, Rüdin initiated further family and twin studies into the inheritance of affective disorders at the GDA [41–43].
Aubrey Lewis (1900–1975), in 1934, considered Rüdin’s studies the starting point in the field, refuting allegations against psychiatric genetics “that it was bad psychiatry and bad genetics” [44]. Lewis distinguished between Rüdin’s research and demands he made for changes in health policy, sharply criticising the eugenic measures in Germany’s “Law for the Prevention of Hereditarily Diseased Offspring”, albeit anonymously [45, 46]. In German-speaking psychiatry, there were well-known critics of Rüdin’s eugenic demands on society; important psychiatrists like Karl Jaspers (1883–1969), Oswald Bumke (1877–1950) and Eugen Bleuler (1857–1939) argued the mode of inheritance was unproven; external and social factors influenced disease course, and the right to personal self-determination had to be respected [15, 47–51] (for further insight into resistance of medical professionals against negative eugenic measures, see [52–54]).
Rüdin’s concept of an empirical heredity prognosis served as a methodological model for many subsequent studies at the DFA, known as the Munich School [55–60]. European and American scientists, some of whom had been fellows of the Genealogic-Demographic Department (GDA) of the DFA, used Rüdin’s research methodology in psychiatric genetic studies [27, 27, 60–65, 65–69]. The results of the twin and family studies undertaken at the Munich School remained valid in their methodology and results for decades [31, 42, 70–72].
Whilst Rüdin actively published over decades [8], only one article reported the methodology he prized, namely his large-scale study on the inheritance of dementia praecox [24], with which he acquired his reputation as a psychiatric geneticist [15]. It is all the more surprising that the similarly laid out study on “manic-depressive insanity”, which was elaborately prepared and carried out over the course of years was never published [8, 28]. Contributing factors may include Rüdin, concentrating on his political career, paid less attention to the practical implementation of his research plans [15]. This is somewhat contradicted, by the fact that the study and manuscript were essentially completed. There is some evidence that Rüdin doubted his results. He concurred with eugenicist and heredity theorist Ludwig Plate (1862–1937), that diagnostic uncertainty necessitated skeptical application of Mendelian rules [73–75]. Most likely is that his demands for negative eugenic measures against patients with affective disorders and their families could not be justified on the grounds of the heredity figures he had calculated.
With the benefit of hindsight, the inheritance figures Rüdin calculated have been confirmed repeatedly [71, 76], and the search for replicable gene variants leading to the onset of affective disorders continues [77]. In a 1924 lecture Rüdin even recognised that environmental factors combined with disposition to illness, trigger onset of disease [78]. Rüdin’s actual results were therefore reasonably sound from today’s perspective, notwithstanding methodological limitations and that his studies were undertaken prior to knowledge of DNA, the double helix, and the intricacies of molecular genetics, epigenetics and endophenotypes [79–84].
Selective publication of positive results remains contentious today. The German human geneticist Peter Propping considers it the greatest danger to psychiatric genetics; “a silent coalition exists between an author and an editor: both are interested in publishing positive findings” [40]. He calls for a platform where all relevant results are accessible to the scientific public, to minimise bias in publishing [40]. In contrast to Propping, prominent scientists like Christiane Nüsslein-Volhard who received the Nobel Prize in Physiology or Medicine in 1995 recommend not publishing negative results, because science should increase knowledge, not merely produce more data. In a recent interview she points to methodological errors as a potential reason for negative results [85]. Rüdin too may have been dubious about his results and therefore refrained from publishing the study.
Rüdin had already clearly formulated his research aim before his study began. He unethically promulgated his eugenic ideology based on a selective and at times patently false reading of his results or even ignoring them [40, 86]. For further discussion on publication bias, see [87, 88].
In 1933, Rüdin chaired the committee for racial hygiene and racial policy at the ministry of the interior of the NS regime and collaborated on the “Law for the Prevention of Hereditarily Diseased Offspring” [8, 17]. For more detailed information about the role that scientists played in bolstering the racial theories of the NS regime see [3]. According to this law, all persons who had been determined to be hereditarily diseased according to medical science were to be compulsorily sterilised [17] (for the origins of the law see [8]). In his commentary on the law’s implementation, Rüdin justified sterilising psychiatric patients based on the results of his study of the inheritance of dementia praecox, extrapolating the allegedly proven inheritance to other psychiatric and neurologic diagnoses [15, 17, 89, 90]. In doing so, Rüdin reinterpreted his results as much more conclusive and reliable than he had in earlier commentaries, but without citing any later studies [24, 74].
As the story of the 1933 sterilisation-law shows, Rüdin’s research results were accepted uncritically as a scientific basis for legislation, strengthening the case for the National Socialists’ health and social policy. Between the law coming into force in 1934, until 1945, between 350,000 and 400,000 persons were sterilised [91–93]. The DFA, founded by Kraepelin and later led by Rüdin, was among the Institutes that issued registration forms required for sterilisation of patients [13]. The number of sterilisations only decreased when, in 1940 and 1941, the Nazis progressed to killing mentally ill and handicapped patients, in “Aktion T4”, under the euphemism “euthanasia” [94] (for the history of eugenics and euthanasia, see also [95–97]). Rüdin became aware of this secret operation at the end of 1939 at the latest, but was not directly involved in its preparation or execution [8]. However, his backing may have influenced important decisions at the highest political levels in favour of killing patients [98, 99]. Rüdin supported research projects which included killing designated patients for post mortem material [100, 101]. Also the meticulously recorded registration of patients in Rüdin’s studies later facilitated locating the victims for forced sterilisation and “Aktion T4” [98, 102]. Because his scientific interests were so consistent with Nazi ideology, the DFA was supported by the various centers of power in the National Socialist state [96, 103]. Rüdin did not consider his associations with these ideologies to compromise the scientific quality of his empirical studies [16]. In celebration of the tenth anniversary of the National Socialist state, Rüdin composed a laudation in praise of Adolf Hitler’s services to racial hygiene [104]. In a memorandum he referred to “euthanasia” as a component of “therapeutic reform” [10, 105].
Contemporary views of Rüdin’s work diverge widely, and are subject to ongoing controversy [13, 90, 99, 106]. Some would strip him entirely of his rank as a scientist [15]; others credit him with the establishment of modern psychiatric human genetics through the development of the “empirical heredity prognosis” [8, 31, 40]. Appeals to discontinue citing Rüdin’s scientific work have been put forward [107]. Certainly any scientific analysis of Rüdin’s works should provide information about his political role [108]. His legacy is deeply troubling but highly illustrative of the nexus between science, ideologies, ethics and humanity.
|
10.1371/journal.ppat.1003968 | Evolution of the Retroviral Restriction Gene Fv1: Inhibition of Non-MLV Retroviruses | Fv1 is the prototypic restriction factor that protects against infection by the murine leukemia virus (MLV). It was first identified in cells that were derived from laboratory mice and was found to be homologous to the gag gene of an endogenous retrovirus (ERV). To understand the evolution of the host restriction gene from its retroviral origins, Fv1s from wild mice were isolated and characterized. Most of these possess intact open reading frames but not all restricted N-, B-, NR-or NB-tropic MLVs, suggesting that other viruses could have played a role in the selection of the gene. The Fv1s from Mus spretus and Mus caroli were found to restrict equine infectious anemia virus (EIAV) and feline foamy virus (FFV) respectively, indicating that Fv1 could have a broader target range than previously thought, including activity against lentiviruses and spumaviruses. Analyses of the Fv1 sequences revealed a number of residues in the C-terminal region that had evolved under positive selection. Four of these selected residues were found to be involved in the novel restriction by mapping studies. These results strengthen the similarities between the two capsid binding restriction factors, Fv1 and TRIM5α, which support the hypothesis that Fv1 defended mice against waves of retroviral infection possibly including non-MLVs as well as MLVs.
| We have followed the evolution of the retroviral restriction gene, Fv1, by functional analysis. We show that Fv1 can recognize and restrict a wider range of retroviruses than previously thought including examples from the gammaretrovirus, lentivirus and foamy virus genera. Nearly every Fv1 tested showed a different pattern of restriction activity. We also identify several hypervariable regions in the coding sequence containing positively selected amino acids that we show to be directly involved in determining restriction specificity. Our results strengthen the analogy between Fv1 and another capsid-binding, retrovirus restriction factor, TRIM5α. Although they share no sequence identity they appear to share a similar design and appear likely to recognise different targets by a mechanism involving multiple weak interactions between a virus-binding domain containing several variable regions and the surface of the viral capsid. We also describe a pattern of constant genetic change, implying that different species of Mus have evolved in the face of ever-changing retroviral threats by viruses of different kinds.
| Viruses co-evolve with their hosts, upon which they are completely dependent for replication. As the host acquires strategies to restrict virus infection the invaders develop counter measures to evade restriction. The ensuing genetic conflict can play out over an extensive timeframe [1], [2], [3], [4]. Due to the unique replication strategy employed by retroviruses where integration of viral genetic information into the host genome occurs [5], the conflict between virus and host can take an interesting twist. When integration occurs in germ or embryonic cells, the virus can become an endogenous retrovirus (ERV) and inherited through the germ line [6], [7]. As a result, viral gene products can be conscripted to serve as defensive forces against further viral infection [8]. The murine retrovirus restriction gene, Fv1, provides perhaps the prototypic example of one such gene [9].
Fv1 restriction was first described in the early 1970s [10], [11] as an activity protecting mice against infection with murine leukemia virus (MLV). Two semi-dominant alleles were identified, Fv1n and Fv1b, that provide protection against B-tropic and N-tropic MLVs, respectively [12], [13]. The crucial difference between N-tropic and B-tropic MLV maps within the viral gag gene to a single codon encoding amino acid 110 of the mature capsid (CA) protein [14] indicating that CA represents the target for the restriction factor. MLVs insensitive to Fv1, called NB-tropic, carry further changes in CA [15], [16]. The mode of action of the Fv1 protein is not fully understood but indirect evidence suggests that it binds to CA on the cores of incoming virions shortly after virus entry into the cell without inhibiting viral reverse transcription [17] but somehow preventing entry of newly synthesized viral DNA into the nucleus [9]. Based on sequence similarity, Fv1 appears to be derived from the gag gene of an ancient ERV called MERV-L (murine endogenous retrovirus with a leucine tRNA primer binding site) though it appears only distantly related to MLV [18].
Amino acid 110 of CA also determines sensitivity of MLV to another retrovirus restriction factor, TRIM5α [19], best known for its ability to restrict HIV-1 [20]. While there is no similarity between Fv1 and TRIM5α at the primary sequence level, both molecules share a similar domain organization [9]. The N-terminal domains both contain an essential coiled coil motif involved in multimerization while the respective C-terminal domains are required for specific virus binding [21], [22]. Indeed, the C-terminal domain of Fv1 can be replaced with CypA, a molecule that binds HIV-1 CA, resulting in a factor that restricts HIV-1 [23]. TRIM5 has been isolated from a number of mammals including a variety of primates, rabbits and cows [20], [24], [25], [26], [27]. These have been shown to restrict a range of retroviruses from different genera. In particular, TRIM5 from the cotton top tamarin can restrict gammaretroviruses, lentiviruses, spumaviruses and betaretroviruses [24], [28], [29]. Comparison of the target sequences show little identity and although the gammaretroviral, betaretroviral and lentiviral CA molecules show a similar tertiary structure [30] the spumavirus target is folded very differently [31]. Residues in the C-terminal B30.2 domain of TRIM5α that determine viral recognition, and thus restriction specificity, are under strong positive selection [32], [33] and are thought to evolve under pressure imposed by retroviral infection [3], [34], [35]. However in no case have the viruses involved been identified unambiguously [36], [37], [38].
In contrast, changes in Fv1 and the acquisition of its antiviral activity are less well defined. Based on its distribution in different subgenera of Mus, it appears that the Fv1 gene was inserted around 4–7 million years ago [39], [40]. However this finding is somewhat paradoxical because the only known target for Fv1, MLV, probably arose considerably more recently as judged by the distribution of its endogenous forms [41], [42]. What then drove the spread and survival of the Fv1 open reading frame? Could it be that viruses other than MLV selected for Fv1? To address this question we have developed a panel of Fv1 genes from different mice and investigated their anti-viral activity against a variety of retroviruses. These studies reveal an extraordinary degree of plasticity in the Fv1 gene as well as two non-MLV viral targets suggesting that a number of different viruses have moulded its evolution.
To study the evolution of Fv1, we set out to clone the gene from a variety of species of Mus. Consistent with previous reports [39], [40], it proved possible to clone Fv1 from multiple species of the subgenus Mus as well as single examples of the subgenera Mus nannomys and Mus pyromys (Table 1). However, we failed to amplify Fv1 from Mus coelomys, Apodemus and Rattus despite multiple attempts [43], suggesting that the insertion leading to Fv1 arose about five million years ago, at the time when the ancestors of Pyromys, Nannomys and Coelomys diverged [44]. Sequencing revealed open reading frames in all cases except Mus mus terricolor (dunni) and M. m. cookii (Figure S1). In M. m terricolor, this was due to a single base pair deletion at position 224 that causes a frameshift and premature stop, while in M. m. cookii, a base pair transition from C to T at position 650 coupled with a 5 base pair deletion causes the formation of a premature stop codon. Interestingly, in three cases, M. m. molossinus, M. m. spretus and M. m. caroli, 2 different sequences were amplified in reproducible fashion. Pairs clustered together in phylogenetic analyses, suggesting the presence of more than one segregating allele in these subspecies of mice.
Sequence comparisons reveal that the N-terminal region of Fv1, which encodes an extended coiled coil region necessary for restriction activity [23], [45], is well conserved (Figure S1). Compared to M. m. caroli, M. m. famulus, Mus nannomys minutoides and Mus pyromus platythrix, all the other Fv1s contained a 3 amino acid insertion near the N-terminus (Figure S1). This change was not important for restriction activity. By contrast, the C-terminal domain shows significant variation in regions important for Fv1 function. Four variable regions, which we designate VA–D can be distinguished (Figure 1). The first variable region (residues 247–276) overlaps a sequence called the Major Homology Region (MHR) that is present in the CA protein of all retroviruses as well as Fv1 [18], [46], [47] and is essential for Fv1 function [48]. Variable regions B–D (amino acids 345–358, 375–401 and the extreme C-terminus) contain the residues we had previously shown to distinguish the predicted products of the n and b alleles of Fv1. These differences are found at amino acids 358, 399 and the very C-terminus of the Fv1 protein where an apparent deletion of 1.3 kb in genomic DNA resulted in a nineteen amino acid length difference [18]; together they appear responsible for the differences in restriction specificity [48]. The present analysis showed that the more divergent mice contained the residues that are found in Fv1n at positions 358 and 399 but they did not contain the 1.3 kb deletion. This suggested that Fv1n arose from the progenitor Fv1, which was similar in length to Fv1b, through an internal deletion, while Fv1b evolved through the substitution of the residues at positions 358 and 399.
Variable regions A–C appear to arise by point mutation but region D shows more significant changes in nucleotide sequence. The three most distantly-diverged mice, M. n. minutoides, M. m. famulus and M. p. platythrix each appear to have B1 repeat sequences inserted, apparently independently, near the deletion site that gave rise to the Fv1n allele (Figure 2). They contribute the last few amino acids of Fv1 resulting in C-termini that are rather different from either Fv1n or Fv1b. Other differences in this region arise from short insertions or deletions perhaps resulting from polymerase slippage during DNA replication. Thus the clones, SPR1 and SPR2, we amplified from M. m. spretus of French and Spanish origins differed by four amino acids; the same difference also seen between Fv1b and CAS2 (Figure 2).
By analogy with other restriction factors it seems likely that the variation seen in Fv1 arose by selection and this would be reflected in changes in restriction specificity. To test this we examined the restriction properties of Fv1 we had cloned from different mice (Table 1) as well as a number of genes synthesized on the basis of published sequences (Table 2) [40], [49]. A tree based on these sequences is shown in Figure 3; it shows good agreement with the accepted phylogeny of genus Mus [50]. The Fv1 genes were introduced into pLgatewayIYFP and tested for restriction of different MLVs using a two-colour FACS assay [19], [51]. The results are shown in Table 3.
As previously shown, the Fv1n gene restricted B-MLV but not N- or NB-MLV while Fv1b, when expressed at protein levels seen in transduced cells, restricted N- and NB- MLV and also had a weak activity against B-MLV [51]. The Fv1 gene from M. m. molossinus and M. m. spretus restricted both N- and B-MLV but not NB-MLV, suggesting that they could be similar to Fv1nr [16]. Hence, we also examined three N-MLV CA variants (D82N, H114R and L117H) that confer resistance to Fv1nr [16]. Variants N-MLV H114R and N L117H, which we have found to be NR tropic, were also resistant to both the Fv1MOL1 and Fv1SPR1 proteins. In contrast, N-MLV N82D was restricted by Fv1MOL1 but not by Fv1SPR1, suggesting that Fv1MOL1 is subtly different from Fv1nr. We have cloned the Fv1nr gene from 4 different strains of mice (129SvEv, NZB/B1NJ, NZW/LacJ and RF/J); all contained a single nucleotide change compared to Fv1n, causing a serine to phenylalanine substitution at residue 352. While Fv1MOL1 also encoded phenylalanine at position 352, Fv1SPR1 possesses a serine at the corresponding position. Taken together, these results suggest that other changes could also be involved in determining the nr-specificity. Perhaps surprisingly, Fv1 from two closely related species, M. m. castaneus and M. m. spicelegus lacked perceptible Fv1 activity against MLV.
Other Fv1 genes displayed a variety of restriction phenotypes. Fv1 from two asian members of the Mus subgenus, M. m. caroli and M. m. cervicolor, lacked Fv1 activity directed against MLV as did the two members of the Pyromys subgenus that we tested, M. p. platythrix and M. p. saxicolor (Table 3). By contrast two members of the Nannomys subgenus, M. n. ninutoides and M. n setulosis were active with the M. n. minutoides clones restricting all six MLVs tested. The M. m. famulus sample, whose position in Mus phylogenetic trees is relatively poorly defined, showed weak activity against N, B, and NB-tropic MLVs. Thus more than half of the Fv1 genes with intact open reading frames did not seem to have any activity against MLV, the target that defines the Fv1 gene, even though they were expressed at similar levels to restricting genes in transduced cells (Figure S2). Further, the extent and specificity of restriction of different MLVs varies significantly. Clearly, the properties of the restriction gene have changed since the gene first became part of the mouse germ line but whether MLV alone was responsible for selecting such changes remained an open question.
Prompted by the example of TRIM5α that can restrict multiple genera of retrovirus [24], [29], we decided to investigate the hypothesis that non-MLV retroviruses might play a role in shaping the evolution of Fv1, by testing a number of different retroviral vectors for restriction by Fv1 from wild mice. These included other gammaretroviruses like Gibbon Ape Leukemia Virus (GALV), Feline leukemia Virus (FeLV) and Porcine Endogenous Retrovirus-A (PERV-A), lentiviruses such as HIV-1, HIV-2, SIVmac, Equine Infectious Anemia Virus (EIAV) and Feline immunodeficiency Virus (FIV), as well as foamy viruses including Prototypic Foamy Virus (PFV), Simian Foamy Virus (SFV) and Feline Foamy Virus (FFV). Some of these results are presented in Table 4. The data show that Fv1 from M. m. caroli, that lacked activity against MLV, restricted FFV strongly and PFV weakly. Moreover, Fv1 from M. m. spretus, which restricted N- and B-MLV, and from M. m. macedonicus, which inhibited N-MLV, were also active against the lentivirus EIAV. By contrast GALV, FeLV, PERV-A, SFV, HIV-1, SIVmac and FIV were not restricted by any of the Fv1 genes in the panel (Table 4 and data not shown). Formally it remains possible that the novel specificities observed result from over expression. Unfortunately, no cell lines expressing Fv1CAR1 and Fv1SPR1 at endogenous levels are available, precluding a direct test of this idea. However we are not aware of any examples of complete restriction of novel viruses resulting from such a mechanism.
To further characterize restriction mediated by Fv1CAR1 and Fv1SPR1 stable cell lines were derived by transducing MDTF cells with retroviral vectors carrying these genes and selecting for G418-resistant single cell clones. These cell lines were used in virus titrations by measuring the percentage of transduced cells by FACS with different amounts of virus (Figure 4A,B). As expected, the titre of EIAV was dramatically reduced in the cell line expressing Fv1SPR1 compared to the untransduced MDTF control (Figure 4A). Similarly titres of FFV and PFV were greatly reduced in MDTF cells expressing Fv1CAR1 compared to untransduced while titres of SFV were unaffected by the presence of the Fv1 gene (Figure 4B). These results confirm the observations made with the 2 colour FACS assay that Fv1 from some wild mice can restrict non-MLV retroviruses.
Fv1 is thought to interfere with MLV replication by preventing nuclear import of newly synthesized viral DNA [9]. To test whether this was also true for EIAV and FFV, we examined the fate of viral DNA in restricting cell lines. Testing EIAV replication in Fv1SPR1 cells shows no inhibition of reverse transcription as measured by levels of newly synthesized late DNA products (Figure 4C). However levels of 2-LTR circles, which are thought to form only after nuclear entry [52], are substantially reduced suggesting a block in nuclear uptake. In Fv1CAR1 cells a reduction in FFV 2-LTR circles, with no change in late RT products, was also observed (Figure 4D), again consistent with a block in nuclear import. However, interpretation of these data is complicated by the fact that the majority of FFV DNA synthesis is thought to occur in the producer cells [53]. Nevertheless, it appears likely that Fv1 is acting to block lentivirus and foamy virus replication at the same stage in the viral life cycle as seen with MLV.
To identify the specificity determinants of these novel restriction activities, chimeric Fv1 genes were constructed and tested for restriction. To look at FFV restriction, we made chimeras between Fv1CAR1, which restricted only FFV, and Fv1n, which restricted B-MLV. Schematic views of the constructs made and the corresponding restriction data are shown in Figure 5A. Replacement of a C-terminal fragment of Fv1n (from residue 318) with the corresponding fragment from Fv1CAR1 generated a chimera (Fv1nC4) capable of restricting FFV. Replacement with a shorter fragment starting from residue 353 (Fv1nC5) was insufficient to confer restriction, suggesting that the determinants of FFV restriction were found between residues 316 and 352 of Fv1 from M. m. caroli. In the reciprocal chimeras, replacing the small C-terminal segment of Fv1CAR1 beginning from residue 352 with that from Fv1n did not result in any loss of activity against FFV. However, when a larger fragment beginning at residue 316 was replaced, activity was lost, confirming the presence of the determinants of FFV restriction within the region of Fv1CAR1 between residues 316 and 352. Within this region, there are 5 residues that differ between Fv1n and Fv1CAR1. These were systematically changed to identify the residues involved in specificity determination (Figure 5B). No single change could endow Fv1n with the ability to restrict FFV (Figure 5B). However two single changes at positions 349 and 352 of Fv1CAR1 resulted in loss of FFV restriction. We therefore mutated both these positions in Fv1n to the corresponding amino acids found in Fv1CAR1. This generated a construct (Fv1nE349KS352Y) capable of restricting FFV. Taken together, these results indicate that both lysine 349 and tyrosine 352 in Fv1 from M. m. caroli are crucial for FFV restriction. We had previously shown that MLV recognition maps downstream of this region [48]; it was therefore interesting to see that Fv1nE349KS352Y (and chimera Fv1Cn5) could recognize both B-MLV and FFV in an additive fashion.
To examine EIAV restriction by Fv1 from M. m. spretus, a second set of chimeras was made between Fv1SPR1, which restricts N-MLV, B-MLV and EIAV, and Fv1n, which only restricts B-MLV. Restriction of EIAV was seen with chimeras only when amino acids from positions 191 and 271 were derived from Fv1SPR1 (Fig. 6A) suggesting that the determinants of EIAV restriction lay between these residues. Interestingly, the determinants for MLV restriction were slightly different from those of EIAV. Replacing a short segment of C-terminus of Fv1n (from residue 366) with that from Fv1SPR1 in Fv1nS3 was sufficient to confer restriction of N-MLV, suggesting that this region contained determinants of N-MLV restriction. However, a reciprocal change in Fv1SPR1 (Fv1Sn3) did not abolish N-MLV restriction. It was only when a C-terminal segment beginning with residue 191 was replaced from Fv1SPR1 (Fv1Sn1) that the restriction of N-MLV was lost. This suggested that additional requirements for N-MLV restriction were found between residues 191 and 271 of Fv1 from M. m. spretus, perhaps overlapping with those that determined EIAV restriction.
There are 5 differences between Fv1n and Fv1SPR1 in the segment between residues 191 and 271 (Figure 6B). To identify the residues involved in restriction, site-directed mutagenesis was employed to change the residues in Fv1n to those present in Fv1SPR1. Reciprocal mutations were also made in Fv1SPR1. These mutants were tested for restriction of EIAV, N- and B-MLV. The substitution from arginine to cysteine at position 268 in Fv1n was sufficient to confer the ability to restrict both N-MLV and EIAV. The reciprocal change in Fv1SPR1 resulted in a partial reduction in restriction of all three viruses. These results indicated that residue 268 was the major determinant of EIAV restriction by Fv1SPR1 and had an influence on MLV restriction but that other neighbouring residues were also important. A lysine to glutamine change at residue 270 in Fv1n resulted in low but reproducible restriction of EIAV and N-MLV though the reciprocal change in Fv1SPR1 had little effect. Interestingly, substitution of the residue at position 261 in both Fv1n and Fv1SPR1 seemed to abolish the restriction of B-MLV, indicating that this residue was involved in the interaction with B-MLV. We conclude that residues 261, 268 and 270 in Fv1 from M. m. spretus are all involved in virus recognition. However, it would appear that recognition of EIAV by Fv1 from M. m. macedonicus has arisen in a different manner as it contains arginine rather than cysteine at position 268.
In this study of Fv1 evolution we have demonstrated that Fv1 shows substantial sequence variation in its C-terminal half, the region of the protein thought to contain determinants of restriction specificity. In addition we have shown that Fv1 is capable of restricting viruses other than its previously defined targets and identified the sequence variation responsible for these novel targets. We note that some Fv1 alleles do not appear to possess an associated restriction activity; it would be of considerable interest to determine whether they recognize other targets.
A previous study had identified six codons, specifying Fv1 amino acids 261, 265, 270, 362, 299 and 401, that show evidence for positive selection during the course of Mus evolution [40]. These represent potential sites of interaction between Fv1 and its target viruses. Combining these data with our previous studies of Fv1 specificity [16], [48], it seems reasonable to conclude that the four variable regions defined in Figure 1 constitute four domains collectively or individually involved in target selection and binding (Figure 7). Thus VRA (amino acids 247–276) includes the positively selected residues 261, 265 and 270 as well as three residues, 261, 268 and 270, shown to be important for EIAV restriction by Fv1SPR1 while VRB (amino acids 345–358) has positively selected amino acid 352, amino acids 349 and 352 important for FFV recognition by Fv1CAR1 as well as residues 352 and 358 important for NR- and N- versus B-tropism, respectively [16], [48]. Variable region C (amino acids 375–401) contains positively selected amino acids 399 and 401 while residue 399 was also implicated in determining N- versus B-tropism [48]. The nature of the length variation at the C-terminus precludes computational analysis for positive selection; nevertheless functional studies [48] provide compelling evidence that this region can also alter restriction specificity.
We have previously noted that CA binding restriction factors Fv1 and TRIM5α share certain design features despite lack of sequence similarity [9]; the present study strengthens this analogy. Both factors possess an N-terminal coiled-coil region allowing dimer formation. They also contain other sequences facilitating the formation of higher order multimers. Both contain a C-terminal domain responsible for virus binding that can be substituted with the cellular CA binding cyclophilin A protein to give a fusion protein capable of restricting HIV-1 and other lentiviruses [23], [54]. We now provide evidence that the CA binding domain of Fv1, like TRIM5α [24], [32], [55], [56], appears to comprise multiple variable regions, showing attributes of positive selection, implying virus driven evolution [3]. Further, Fv1 is capable of recognizing multiple genera of retrovirus. It seems possible that the ability to recognize multiple viruses by low affinity binding with avid binding provided by multimerisation [57] represents a common theme in restriction factor design. Further insights into the interaction between virus and restriction factor requires detailed structural information; unfortunately both Fv1 and TRIM5α are relatively recalcitrant to such studies.
The origin of Fv1 remains unclear. It is only present in Mus and appears related to the gag gene of the endogenous retrovirus family ERV-L [18], [47]. This suggests that Fv1 might be derived from an endogenous retrovirus following the loss of both LTRs and pol coding sequences [58]. Interestingly a significant increase in MERV-L copy number took place at around the time of the separation of Mus subgenera [59], the time when Fv1 became part of the Mus germline. However sequence alignments indicate that Fv1 and MERV-L share only 43% amino acid identity whereas the different genomic MERV-L elements are much more closely related to one another (<5% nucleic acid divergence). BLAST searches of the NCBI non-redundant genome databases reveal no sequences intermediate between Fv1 and MERV-L. This suggests that Fv1 might be derived from an exogenous virus related to ERV-L that has not made its home as an intact ERV, at least not in any species so far sequenced, and may no longer exist in infectious form. As such Fv1 might be the last remnant of an ancient extinct virus, or paleovirus [2]. Unfortunately this inability to identify the proximal precursor for Fv1 prevents us from determining whether or not the original transgene showed restriction activity and, if so, against which virus.
The selection and continuing existence of the Fv1 open reading frame implies that it provides an evolutionary advantage, presumably by providing protection against retroviral infection. The observation of multiple restriction specificities suggests that a variety of unknown viruses have contributed to this process. Taken together with frequent genetic changes to inactivate [60] or block MLV receptors [61], these data imply that multiple virus epidemics have occurred in the course of mouse evolution [62]. One might postulate that at least four significant virus exposures have occurred during Mus evolution (Figure 8). One took place after the divergence of Nannomys; a second occurred in M. m. caroli; a third in mice in countries surrounding the Mediterranean Sea and a fourth in the Mus musculus subfamily. In turn this prompts the question of how the current properties of a restriction factor reflect the properties of the viruses involved in selection, a question that is as relevant for TRIM5α as for Fv1. Specifically one might ask whether the ability to restrict one genus of retrovirus reflects prior exposure to that kind of virus. An affirmative answer might resolve the vexed question of whether foamy viruses have deleterious effects on their hosts [63], possibly as co-pathogens [64] since both Fv1 (this paper) and TRIM5α [29] have evolved to see one or more such virus. Alternatively, changes selected by, say, a gammaretrovirus like MLV, might fortuitously result in recognition of a lentivirus like EIAV or a foamy virus like FFV. In light of the shorter generation time of mice compared to primates Fv1 could provide a more useful system for studying evolution of restriction specificity than does TRIM5α. The observation of multiple alleles of Fv1 might also suggest that selection is an ongoing process offering opportunities for experimental analysis. In particular, the evolution of restriction activity against the lentivirus EIAV, which appears to have happened in two different ways in M. m. spretus and M. m. macedonicus as well as the kind and source(s) of the virus(es) involved would appear worthy of more detailed investigation.
Genomic DNA samples for Mus musculus laboratory mouse strains C57BL/6J, AKR/J, DBA/2J, 129/SvEv, and LG/J, M. m. spretus (M. spretus), M. m. caroli (Mus caroli), M. m. molissinus (MOLD/Rk) and M. m. castaneous were purchased from the Jackson Laboratory. Genomic DNA from M. p. platythrix, M. m. cookii, M. m. spicilegus, M. m. spretus, M. m. castaneous and M. m. bactrianus were gifts from Dr. F. Bonhomme (Laboratoire Genome et Populations, Universite de Montpellier II, CNRS). M. n. minutoides genomic DNA was a gift from Dr. B. Mock (National Cancer Institute, NIH), while M. m. famulus and M. m. cervicolor genomic DNA were gifts from Dr. John Coffin (Tufts University School of Medicine, Boston). M. m. terricolor (dunni) genomic DNA was prepared from a Mus dunni tail fibroblast (MDTF) line [65] using the DNeasy blood and tissue kit (Qiagen). The Fv1 ORF was PCR amplified from mouse genomic DNA using primers PL80 and GT17 (see Table S1 for primer sequences) that permit the amplification of a sequence starting from 3056 bp upstream of the start codon of Fv1 to 2684 bp downstream of the start codon. Sequence analysis of this region from in-bred mice identified 2 SacI sites downstream of the PL80 primer-binding site, while GT17 contained a SalI site. The PCR products were hence cloned initially as SacI/SalI fragments into M13 phage and sequenced. Subsequent subcloning of Fv1 ORFs was carried out following amplification with the primers GatewayFv1F and Gateway Fv1rev. The PCR product was used in a second amplification reaction with primers UniversalF and UniversalRev to attach the attB sites to the ends of the fragment. This was then inserted into pDNR221, which is an entry vector to the Gateway Cloning system, using BP clonase (Invitrogen). Finally, the entry clone was used in a LR reaction with LR clonase to insert the Fv1 ORF into either pLgatewayIRESEYFP or pLgatewaySN to generate retroviral delivery vectors carrying either the EYFP or G418 resistance marker. Details of these different clones as well as the abbreviations used for their designation are summarized in Table 1.
Fv1 open reading frames from M. m. macedonicus, M. n. minutoides, M. n. gratus, M. n. setulosis, M. n. triton and M. p. saxicolor were synthesized chemically (GENEART, Life Technologies) based on their published sequences [40] with added attB sites and introduced into pLgatewayIRESEYFP via pDNR221. These clones are also summarized in Table 2.
Fv1 chimeras were generated by overlapping PCR. Briefly, a 5′ fragment was amplified from one parental sequence while a 3′ fragment was amplified from the other. The two fragments were then combined in a third amplification reaction using forward and reverse primers that annealed to the 5′ and 3′ ends of Fv1 respectively. Internal primer pairs were designed to target regions of identity between the two parental sequences. The sequences of the primers are shown in Table S1.
To generate the Fv1nC series, the 5′ fragments were amplified from Fv1n using TopoFv1F and either C1Rev, C3Rev, C4Rev or C5Rev, while the 3′ fragments were amplified from Fv1caroli (CAR1) using either C1F, C3F, C4F or C5F and Fv1caroliRev. The 2 fragments were joined in a reaction using TopoFv1F and Fv1caroliRev to yield Fv1nC1, Fv1nC3, Fv1nC4 and Fv1nC5. Similarly, the 5′ fragments for the reciprocal series Fv1Cn were amplified from Fv1caroli (CAR1), using the same primer pairs as the Fv1nC series, while the 3′ fragments were amplified from Fv1n using either C1F, C3F, C4F or C5F and Fv1nRev. These fragments were joined using primer pair TopoFv1F and Fv1nRev, yielding Fv1Cn1, Fv1Cn3, Fv1Cn4 and Fv1Cn5.
The 5′ fragments for the Fv1nS series were amplified from Fv1n using TopoFv1F and either S1Rev, S2Rev or S3Rev, while the 3′ fragments were amplified from Fv1spretus (SPR1) using either S1F, S2F or S3F and Fv1spretusRev. The 2 fragments were joined together in a reaction using TopoFv1F and Fv1spretusRev to yield Fv1nS 1, Fv1nS2 and Fv1nS3. Similarly, the 5′ fragments for the reciprocal series Fv1Sn were amplified from Fv1spretus (SPR1), using the same primer pairs as the Fv1nS series, while the 3′ fragments were amplified from Fv1n using either S1F, S2F or S3F and Fv1nRev. These fragments were joined using primer pair TopoFv1F and Fv1nRev, yielding Fv1Sn1, Fv1Sn2 and Fv1Sn3.
The chimeric fragments were cloned into pENTR/D-TOPO (Invitrogen) and verified by sequencing before transferring into the retroviral vector pLgatewayIRESEYFP.
The point mutants were generated by site directed mutagenesis using the primer pairs listed in Table S1. Mutagenesis was carried out in 50 microlitre reactions containing 2.5 units of Pfu ultra, 10 ng of template, 0.2 mM dNTP and 125 ng each of the forward and reverse primer. The reaction was performed in a thermal cycler at 95°C for 2 minutes followed by 18 cycles of 95°C for 30 seconds, 55°C for 1 minute and 68°C for 9 minutes 30 seconds. The PCR product was then digested with10 units of DpnI (Roche) for 1 hour before transforming XL10Gold cells (Agilent technologies). Colonies were screened by restriction digest and the mutations were verified by sequencing.
MDTF and 293T cells were maintained in DMEM containing 10% foetal calf serum and 1% penicillin and streptomycin. Viruses were made by the transient transfection of 293T cells as previously described [19], [51]. Delivery viruses were produced by co-transfecting pcz-VSVG, pHIT60 and a retroviral vector containing Fv1 and either the EYFP or G418 resistance gene. N-, B- and NB-tropic MLV tester viruses were generated by co-transfection of pczVSVG, pczCFG2fEFPf and either pCIGN, pCIGB or pHIT60 respectively, while the NR-tropic viruses were made using a mutagenized form of pCIGN as previously described [16]. EIAV tester viruses were made using pczVSVG, pONY3.1 and pONY8.4ZCG or pONY4.1Z [66], while PFV, SFV and FFV were produced with pciSFV-1envwt and either pczDWP001, pcDWS001 or pcDWF003 respectively [29]. HIV-1 tester viruses were generated by co-transfecting pczVSVG with p8.91 and pCSGW. MLV and HIV-1 were frozen in aliquots at −80°C while EIAV and foamy viruses were freshly prepared for each experiment.
Restriction activity was routinely assayed using transient two colour FACS analyses as described previously [19], [51]. Briefly, Fv1 was introduced into MDTF cells together with an EYFP marker in a retroviral delivery vector. Three days post-transduction, the cells were challenged with tester viruses carrying the EGFP markers. The cells were then subjected to FACS analyses three days later and the percentages of tester virus positive cells in EYFP (i.e. Fv1) - positive and - negative cells determined and compared. Ratios of less than 0.3 were taken as restriction while those that were greater than 0.7 were taken to represent no restriction. Numbers between 0.3 and 0.7 were taken to represent partial restriction.
Alternatively, single cell clones stably expressing restricting Fv1s were derived by transducing MDTF cells in 12 well plates with limiting dilutions of retroviral vectors carrying Fv1 and a G418 resistance marker. The cells from each well were transferred to a 10 cm dish and G418 was added to a concentration of 1 mg/ml. Well-separated colonies were picked from the dishes when they appeared 7 to 10 days after antibiotic selection was started. Typically, 6 to 8 colonies were picked for each Fv1 cell line, expanded and tested for restriction before being used for virus titration. To titrate tester viruses, MDTF cells and their derivatives were seeded in 12 well plates at a density of 5×104 cells per well 24 hours prior to infection. Increasing amounts of viruses carrying the EGFP marker were then added to the wells and the percentage of infected cells was determined by FACS 3 days post infection.
MDTF cells and their derivatives stably expressing Fv1 were seeded in 6 well plates at a density of 5×105 cells per well 24 hours prior to infection. The cells were transduced at an m.o.i. of 1 with equal amounts of viral vectors that had been pre-treated with 10 units/ml of DNase (Promega) for 1 hour at room temperature. The cells were harvested 7 or 18 hours post-infection for quantification of late RT products and 2 LTR circles respectively. Total genomic DNA was extracted using the DNeasy blood and tissue kit (Qiagen) and 250 mg or 500 mg was used for quantitative PCR to detect late RT products and 2 LTR circles respectively. Primers and probes directed against EGFP [67] were used for quantifying late RT products from MLV and FFV while those directed against LacZ were used for EIAV. The retroviral vectors fEGFPf and pHIT111 were used as standards for EGFP and LacZ quantification respectively. Primers and probes for the detection of MLV 2 LTR circles have been described previously [68]. In order to detect EIAV and FFV 2 LTR circles, primers and probes that amplified and bound to a fragment spanning the 2LTRs were designed. For EIAV 2 LTR circle detection, EIAV2LTRCF (5′ACTCAGATTCTGCGGTCTGAG3′), EIAV2LTRCRev (5′ACCCCTCATAAAAACCCCAC3′) and EIAV2LTRCprobe (5′FAM-CTCAGTCCCTGTCTCTAGTTTGTCTGTTCG-Tamra3′) were used while FFV2LTRCF (5′CCAGAACTCACATGAGTGGTG3′), FFV2LTRCRev (5′CTCATCGTCACTAGATGGCAG3′) and FFV2LTRCprobe (5′FAM-GAAGGACTAACCTATCCCAGGTATAGGCCG3-Tamra') were used for the quantification of FFV 2LTR circles. The primer pairs were used to amplify fragments spanning the 2 LTRs from genomic DNA of EIAV or FFV infected cells. The fragments were cloned into pCR-BluntII-TOPO (Invitrogen) to be used as standards. Quantitative PCR was performed in 25 ml reactions using the ABsolute QPCR Rox mix from Abgene with 300 nM of each primer and 200 nM of probe. A programme of 50°C for 2 minutes, 95°C for 15 minutes followed by 40 cycles of 95°C for 15 s and 60°C for 1 minute was employed in the Applied Biosystems 7500 real time PCR system.
Trees were generated using the MegAlign programme from the DNASTAR Lasergene package. The distance values were calculated using the Kimura distance formula that takes into account the number of non-gap mismatches and silent substitutions.
|
10.1371/journal.pntd.0002356 | Inter-epidemic Transmission of Rift Valley Fever in Livestock in the Kilombero River Valley, Tanzania: A Cross-Sectional Survey | In recent years, evidence of Rift Valley fever (RVF) transmission during inter-epidemic periods in parts of Africa has increasingly been reported. The inter-epidemic transmissions generally pass undetected where there is no surveillance in the livestock or human populations. We studied the presence of and the determinants for inter-epidemic RVF transmission in an area experiencing annual flooding in southern Tanzania.
A cross-sectional sero-survey was conducted in randomly selected cattle, sheep and goats in the Kilombero river valley from May to August 2011, approximately four years after the 2006/07 RVF outbreak in Tanzania. The exposure status to RVF virus (RVFV) was determined using two commercial ELISA kits, detecting IgM and IgG antibodies in serum. Information about determinants was obtained through structured interviews with herd owners.
An overall seroprevalence of 11.3% (n = 1680) was recorded; 5.5% in animals born after the 2006/07 RVF outbreak and 22.7% in animals present during the outbreak. There was a linear increase in prevalence in the post-epidemic annual cohorts. Nine inhibition-ELISA positive samples were also positive for RVFV IgM antibodies indicating a recent infection. The spatial distribution of seroprevalence exhibited a few hotspots. The sex difference in seroprevalence in animals born after the previous epidemic was not significant (6.1% vs. 4.6% for females and males respectively, p = 0.158) whereas it was significant in animals present during the outbreak (26.0% vs. 7.8% for females and males respectively, p<0.001). Animals living >15 km from the flood plain were more likely to have antibodies than those living <5 km (OR 1.92; 95% CI 1.04–3.56). Species, breed, herd composition, grazing practices and altitude were not associated with seropositivity.
These findings indicate post-epidemic transmission of RVFV in the study area. The linear increase in seroprevalence in the post-epidemic annual cohorts implies a constant exposure and presence of active foci transmission preceding the survey.
| Rift Valley fever (RVF) is an arthropod-borne viral disease that affects people, livestock and wild animals. It occurs mostly in Africa, and epidemics have been reported in the Arabian Peninsula. RVF is transmitted to humans and animals by mosquitoes, but people can also get the infection through direct contact with blood or tissues of infected animals. The disease occurs in epidemic form in a cycle of 5–15 years, but some reports also indicate occurrences of the disease during non-epidemic periods. We report here inter-epidemic period transmission of RVF in livestock population, evidenced by demonstration of RVFV antibodies in animals that were born after the 2006/07 RVF outbreak in Tanzania and demonstration of immunoglobulin M (IgM), a short lived class of antibodies, following infection by RVF virus in 9 samples. We have also identified hotspots of transmission in the study area, with exposure being higher away from the main flood plain. There was a linear increase in percent seropositivity from 1 year olds to age 5 years, implying a possible annual challenge.
| Rift Valley fever (RVF) is known to occur in outbreaks in cycles of 5–15 years in the Eastern Africa region and the Horn of Africa, following unusual high precipitations that lead to sustained flooding [1], [2]. In recent years, evidence of RVF transmission during the inter-epidemic periods in some parts of the African continent has increasingly been reported [3]–[5]. The inter-epidemic transmissions generally pass undetected clinically, but can be revealed where active serological surveillance is regularly done in either livestock or human populations [4], [6], [7].
Rift Valley fever is a mosquito borne viral zoonosis that affects both livestock and wild ruminants [4], [5], [8]. It is caused by Rift Valley fever virus (RVFV) belonging to the genus Phlebovirus of the family Bunyaviridae and in susceptible animals is manifested clinically by high fever, and causes abortion in susceptible pregnant animals irrespective of the gestation period and high mortality in newborn animals [9]. In humans, RVF can be asymptomatic, but can also cause mild illness (associated with headache, fever, muscle and joint pains) or severe illness (associated with hemorrhagic fever, encephalitis or ocular disease) [10]–[12].The disease was first described in the early 1910s and the aetiological agent was isolated in the 1930s in Kenya [13]. The disease pattern in the Eastern Africa region and the horn of Africa is driven by climatic conditions linked to the El Niño/Southern Oscillation (ENSO) phenomenon, which leads to unusual high rainfall and floods alternated by long dry spells [2]. In other parts of Africa, RVF emerged in relation to the construction of hydroelectric power dams along Senegal river and thereafter established itself as endemic disease [14], [15]. In the Arabian Peninsula RVF was introduced through trading of live animals with countries in the horn of African [16], [17].
RVF maintenance in nature between epidemics both in the mammalian host and vector populations has not been fully explained [18], [19]. This is partly due to the factors driving its maintenance at fine geographical scales that are yet to be described in detail. Such factors interact in diverse ways in different geographical regions of Africa or beyond and may play a crucial role in vector population dynamics and the disease transmission [20].
The disease in humans is primarily an occupational hazard affecting people in close contact with infected animals through blood and body fluids, e.g. livestock keepers, abattoir workers, laboratory workers and veterinarians in the course of taking care and treating sick animals, butchering, or disposing of dead animals and aborted fetuses [21]–[25]. Consumption of animal products such as unpasteurized milk and raw meat has been identified as exposure risk factors in some observational studies [26] but attempt to show infection through oral route in experimental studies had mixed results [27], [28]. People can also acquire the disease through infectious mosquito bites [24], [29].
In the past two decades, RVF outbreaks have been reported twice in Tanzania, with the most recent outbreak occurring in 2006/07. During this outbreak, the disease occurred in 10 of the 21 regions of Tanzania mainland, including Morogoro region where the Kilombero river valley is located, with clinical cases in people and livestock [30]–[32].Given the ecology and climatic factors that support mosquito vectors in the valley we hypothesized a continued transmission of RVFV during the inter-epidemic periods. The goal of our study was therefore to investigate potential inter-epidemic period transmission of RVF in an area experiencing seasonal floods annually. Specifically, the study aimed 1) to establish baseline information on the exposure status to Rift valley fever in livestock as determined by presence of antibodies against RVFV in livestock blood samples during the inter-epidemic period and 2) to assess the relationship between RVF seropositivity, animal characteristics and environmental factors.
This study was conducted in the Kilombero river valley, which spans Kilombero and Ulanga districts in Morogoro region. The Kilombero valley is a seasonally inundated flood plain up to 52 km wide, at high water, between the densely forested escarpment of the Udzungwa mountains, which rise to 2576 m above sea level (asl) on the north-western side and the grass covered Mahenge mountains, which rise to 1516 m asl on the south-eastern side (Figure 1). The valley receives an average annual rainfall of 1200 mm–1800 mm and temperatures range between 25°C and 32°C. The valley has a diverse ecology and demography with villages consisting largely of numerous distinct groups of houses located on the margins of the flood plain where rice cultivation is the predominant economic activity. Other land use types include hunting, fishing, forestry, pastoral livestock rearing and other crops cultivation. There has been intense malaria transmission in the valley with the main vectors being Anopheles gambiae complex and Anopheles funestus [33]. Other mosquito species inhabiting the valley, some of which are vectors of RVF virus, include Culex spp, Aedes spp and Mansonia spp [34].
The sero-survey was done from May to August 2011, approximately four years after the end of the 2006/07 RVF outbreak in Tanzania. Blood samples were collected from three livestock populations, namely cattle, goats and sheep from 44 villages in both districts. These villages were selected based on their inclusion in a demographic surveillance system [35], [36] and the proximity to the Kilombero river flood plains (figure 1). In each village four households that kept at least one of the three species, were randomly selected from the list of livestock keepers in that particular village. For each participating household a maximum of 18 samples were collected (i.e. maximum 10 cattle, maximum 4 goats and maximum 4 sheep, the actual numbers sampled depending on the number of a particular species present). The sampling strategy was based on selecting animals of all age groups in order to be able to characterize epidemic and inter-epidemic transmission. The blood samples were collected in 6 ml vacutainer tubes with clot activator, labeled and stored in a cooler box with ice packs while in the field. After coagulation, serum was eluted from the whole blood into a 1.8 ml cryovial tube, labeled and stored in a car fridge until transfer to the laboratory for laboratory analysis. Characteristics of individual animals together with the herd history were obtained through a structured interview with the herd owner.
Individual animal age was estimated using history taking, review of available records on date of birth and dentition. Records were available only from households keeping exotic dairy cattle. Dentition was used in determining age of cattle between 24 to 54 months only [37]. When the above two methods yielded no useful results, age was estimated by probing the head of the household, herd boys and other members of the household for the animal's month/season and year of birth and for female animals also by taking into account number of births and average birth rate for the particular species in the valley.
To determine the individual animals' longstanding exposure status to previous Rift Valley fever virus (RVFV), a commercial, inhibition enzyme-linked immunosorbent assay (c-ELISA) for the detection of antibodies against RVFV in humans, domestic and wildlife ruminants was used (Biological Diagnostic Supplies Limited, Dreghorn, United Kingdom) [38]. Recent infection was determined using a commercial IgM ELISA kit (Biological Diagnostic Supplies Limited, Dreghorn, United Kingdom) [39].The IgM ELISA test was employed for c-ELISA positive samples only since the c-ELISA detects both IgG and IgM antibodies against RVFV [38].
The data was analysed in STATA version 12 [Stata Corp., College Station, Texas, USA]. To examine the determinants for RVF seropositivity, first a univariable analysis of individual factors was performed by fitting a logistic regression model with wards and villages as random effects to account for clustering. Variables with p-value <0.25 were selected as potential covariables in the multivariable analysis, where a p-value ≤0.05 was considered statistically significant. Forward model-building was done with subsequent models evaluated against sparser models by means of the Akaike information criterion (AIC). Two-way interactions between variables included in the model were also tested. Lastly, all factors that were dropped in the process of model building were later tested for any confounding effect. Factors were considered a confounder if they led to a change of ≥25% in the coefficient estimates of other determinants.
The spatial analysis of the seropositivity was performed using ArcGIS software version 10 (ESRI, Redlands, USA) using hot spot analysis and inverse distance weighted (IDW) tools. The Getis-Ord Gi* statistic for each feature (household) was computed with the resultant Z-score values indicating where households with either high (hot spot), median (random) or low (cold spot) values cluster spatially. The IDW is a deterministic interpolation model that assigns values to locations where no measurements have been taken to produce a surface pattern, based on how far those locations are to the sentinel locations where measurements have been taken.
The blood collection procedure from livestock was performed by a qualified veterinarian following proper physical restraint of animals that ensured both personnel and animal safety. Livestock owners were explained the study purpose and procedures and upon agreeing to participate they provided a written consent prior to study procedures and blood collection from their animals. Ethical approval for this study protocol was obtained from the Institutional Review Board of the Ifakara Health Institute and Medical Research Coordination Committee of the Tanzania's National Institute for Medical Research (permit number NIMR/HQ/R.8a/Vol. IX/1101).
A total of 1680 livestock serum samples were tested by RVF c-ELISA, of which 1234 samples were from Kilombero district and 446 samples were from Ulanga district. Out of the samples tested 970, 455 and 255 were from cattle, goats and sheep respectively. Several potential animal-level risk factors were investigated; table 1 shows the univariable logistic regression model output of the risk factors. The proportion of seropositive animals by c-ELISA was 11.3%. A seroprevalence of 5.5% was recorded among animals that were born after the 2006/07 RVF outbreak (less than 4 years of age), compared to 22.7% in those that were born before and thus present during the outbreak. There was a linear increase in proportion sero-conversion up to age 5, Figure 2. The samples included less than one year olds in which the youngest positive individual was 7 months old at which age the maternally acquired immunity has waned out. There was no significant difference in seropositivity between the different species. The sex difference in seroprevalence in animals born after the previous epidemic was not significant (6.1% vs. 4.6% for females and males respectively, p = 0.158) whereas there was a significant sex difference in seroprevalence in animals that were present during the outbreak (26.0% vs. 7.8% for females and males respectively, p<0.001) and overall (table S1).
Nine out of c-ELISA positive samples were positive for RVFV IgM antibodies indicating a recent infection. The IgM positive samples originated from 7 villages out of the 44 sampled. The spatial distribution of sero-conversion in the study area exhibited no particular pattern but rather a few hotspots (Figure 1). Each district had an almost equal share of the disease, and this was true for the wards. Very few villages had zero seroprevalence. There was a significant relationship between proportion seropositivity and being >15 km from the flood plains as compared to <5 km (Table 2).
The findings from this serosurvey indicate post-epidemic and recent transmission of RVFV in livestock populations in the Kilombero valley. The demonstration of RVFV antibodies in animals as young as one year old, and the observed linear increase in the proportion of sero-converted animals from the age of 1 year to 5 years implies a constant exposure to infectious mosquito bites. The IgM antibodies detected in some animals illustrate presence of active foci of recent transmission preceding the serosurvey as the median duration of IgM antibodies to RVFV is two months [39]–[41]. These observations add to the increasing body of serological [3], [4], [42], [43] and virological [44], [45] evidence pertaining to RVFV transmission during the inter-epidemic periods in parts of Africa. Inter-epidemic transmission can be detected where active disease surveillance is in place in the livestock and/or human populations, as most of the inter-epidemic infections either are subclinical or mistaken for other diseases in the absence of public awareness of RVF presence [6], [46], [47].
The high prevalence observed in animals that were present during the outbreak is not surprising as during epidemics there is high exposure to RVFV with resulting high herd immunity [48], [49]. Such increased prevalence with age was also reported in sero-surveys in Madagascar, Nigeria and Senegal [47], [48], [50]. One of the few other studies reporting sex differences in RVF prevalence in livestock, a study in a slaughter house in Chad also found higher prevalence in female animals [51]. In contrast to our observation, a serosurvey in Madagascar reported higher prevalence in male animals [50]. Our observation might be as a result of female animals staying longer in a herd due to their role in reproduction and consequently most of sampled female animals were relatively older than their male counterparts. The different timing of our survey and the Madagascar survey, i.e. 4 years versus 3 months post-outbreak, might explain the different result on risk associated with sex in livestock populations.
Despite the sero-conversions observed in animals born after the 2006/07 outbreak, there have been no reports of epidemic or clinical disease in the area during the study period. This might be as a result of high herd immunity following the RVF outbreak as demonstrated by high prevalence in animals that were present during the previous epidemic in the study area. In such scenario high proportion of offspring born to immune dams would acquire maternal antibodies thus protected during vulnerable young age whereas old animals of local breeds are naturally less susceptible to clinical disease [52], [53]. Another explanation could be a circulation of non-virulent strain of RVFV [54] during inter-epidemic period in the area as it has been hypothesized in other reports of sero-conversion with no previous epidemic or clinical disease reports [46]. The sporadic cases of RVF could easily be confused with other livestock diseases which present similar clinical features of fever, including sporadic abortions and thus overlooked and under reported [53].
The contact between infected vectors and naïve mammalian hosts is the main determinant of arboviral disease transmission [55]. Mosquito vector population dynamics are driven by environment and ecology, which provide essential life resources. The trend observed in this study of increased seroprevalence away from the main flood plain and in high altitude is contrary to findings in other studies [6], [46]. This could be an indication of other factors playing a role including localized floods unrelated to the main flood area, but also proximity to dense vegetation (forested environments) that could harbor a variety of mosquito vectors [6], as the main forested areas are further away from the flood plain. Another explanation could be movement of previously exposed animals from one locality to another within the area through animal trade among livestock keepers, dowry or establishment of new households. In such circumstances, naïve animals can be transferred to an infected area or infected animals can be introducing the disease into a naïve location. In this way vector populations which are abundant and of diverse species within the valley [34] are exposed and maintenance mechanisms established where conditions are favorable. The observed separate hotspots further point to fine scale factors playing major role in transmission dynamics. On the other hand the only cold spot is located around Ifakara town, a semi urban environment in which livestock keeping is characterized by small herds, sedentary and mixed feeding practices as compared to relatively large herds with extensive system in the villages. The urban environment might be unfavorable to the main vector species but also due to limited daily livestock movements there is little interaction between neighbouring herds thus possibility of infection to limit itself to isolated herds in the event of disease outbreak.
Given the possible active transmission observed in this study within Kilombero valley and the moving out of livestock from the valley as a result of environmental degradation of wetland due to overstocking and overgrazing, there are chances for incubating or sick animals to introduce the disease into new areas. This might be enhanced by the quick means of transportation employed and possibilities for animals to harbour the RVFV for up to three weeks [52], [56]. In view of that, follow up of these moved herds and their new environment through serological and vector population monitoring will help to inform various stakeholders of the currently unidentified consequences, as livestock movement have been implicated to spread RVF in previously free areas [16], [57].
The interaction of livestock keepers with their animals is intense and includes milking, taking care of sick animals, grazing, using as draft animals, slaughtering and butchering and even children playing with animals. Future work should establish to what degree this inter-epidemic zoonotic circulation of RVFV leads to human infection as well. If considerable transmission to humans exists, health care providers within the Kilombero valley should consider RVF in their differential diagnosis in all fever cases presented in their facilities as RVF in humans may present with similar clinical signs to malaria, which is thought to be the main cause of fever in the valley [58]. We think this should be of priority in particular when dealing with patients from agro-pastoralist communities, especially when the malaria test is negative.
The findings from this study indicate post-epidemic and recent transmission of RVFV in livestock populations in the Kilombero river valley. The linear increase in prevalence of RVFV antibodies in the post-epidemic annual cohorts implies a constant exposure and presence of active foci of recent transmission preceding the survey.
|
10.1371/journal.pgen.1008288 | The Drosophila ERG channel seizure plays a role in the neuronal homeostatic stress response | Neuronal physiology is particularly sensitive to acute stressors that affect excitability, many of which can trigger seizures and epilepsies. Although intrinsic neuronal homeostasis plays an important role in maintaining overall nervous system robustness and its resistance to stressors, the specific genetic and molecular mechanisms that underlie these processes are not well understood. Here we used a reverse genetic approach in Drosophila to test the hypothesis that specific voltage-gated ion channels contribute to neuronal homeostasis, robustness, and stress resistance. We found that the activity of the voltage-gated potassium channel seizure (sei), an ortholog of the mammalian ERG channel family, is essential for protecting flies from acute heat-induced seizures. Although sei is broadly expressed in the nervous system, our data indicate that its impact on the organismal robustness to acute environmental stress is primarily mediated via its action in excitatory neurons, the octopaminergic system, as well as neuropile ensheathing and perineurial glia. Furthermore, our studies suggest that human mutations in the human ERG channel (hERG), which have been primarily implicated in the cardiac Long QT Syndrome (LQTS), may also contribute to the high incidence of seizures in LQTS patients via a cardiovascular-independent neurogenic pathway.
| Neurons are extremely sensitive to diverse environmental stressors, including rapid changes in the ambient temperature. To buffer environmental stress, many animals have evolved diverse physiological mechanisms to protect neuronal activity from acute and chronic stressors. Failures of these safeguards often lead to hyperexcitability, episodic seizures, and chronic epilepsy. However, although seizures and related syndromes are common, their underlying molecular and genetic factors, and their interactions with environmental triggers, remain mostly unknown. Here, we show that in the fruit fly, mutations in the ERG voltage-gated potassium channel seizure (sei), an ortholog of the human hERG channel that has been previously implicated in the cardiac Long-QT syndrome, also increases seizure susceptibility. We demonstrate that in addition to its cardiac expression, the sei channel is broadly expressed in the nervous system. In neurons, sei channels are enriched in axonal projections, and are specifically required in excitatory and octopaminergic modulatory neurons, as well as the non-neuronal glia, for maintaining organismal resistance to heat-induced seizures. Thus, our work indicates that the previously reported increase in seizure susceptibility in individuals with mutations in hERG is possibly related to its neuronal action, independent of its cardiac functions.
| Neuronal homeostatic responses to acute and long-term environmental stressors are essential for maintaining robust behavioral outputs and overall organismal fitness [1–3]. Many environmental stressors, such as changes in temperature or oxygen availability, impact various aspects of neuronal system function [1]. Nervous systems must therefore compensate, in a homeostatic manner, in order to continue functioning in the presence of these stressors. At the neuronal level, the homeostatic response to stress depends on both synaptic and cell-intrinsic physiological processes that enable neurons to stably maintain optimal activity patterns [4–6]. The synaptic processes include both presynaptic mechanisms related to neurotransmitter release and postsynaptic mechanisms controlling neurotransmitter receptor localization, turnover, and control of downstream signaling pathways [2]. Previous theoretical and empirical studies in both invertebrate and mammalian species have suggested that neuronal intrinsic robustness depends on the expression and activity of specific combinations of ion channels and transporters, which can vary across neuronal cell types and individuals [7–10]. While some of the transcriptional and physiological processes that enable neurons to adjust their intrinsic activity levels in response to long-term stressors have been identified, primarily via the altered conductance of voltage-gated ion channels [11–13], most of the genetic and molecular mechanisms that mediate susceptibility to acute, environmentally-induced seizures, such as fever-induced febrile seizures, remain unknown [14–16].
In humans, seizures result from a diverse set of mechanisms that lead to an abnormal increase in electrical activity of the nervous system. A wide range of stressors have been associated with triggering seizures, including fevers, flickering lights, sleep deprivation, and emotional stress [17, 18]. A handful of genetic mutations have been linked to febrile [19–21] and photosensitive [22, 23] seizures, yet these only account for a small percentage of individuals experiencing seizures in response to these and other stressors.
Because of its small size, large surface-to-volume ratio, and its inability to internally regulate body temperature, the fruit fly Drosophila melanogaster, represents an excellent model for studying mechanisms underlying the neuronal response to acute heat stress [24]. To date, forward genetic screens in Drosophila have identified several mutations that lead to heat-induced seizures and paralysis [24–27]. These mutations seem to primarily affect the function of genes that encode voltage-gated sodium and potassium channels, and proteins associated with their neuronal function [28–31]. Here we tested the hypothesis that the knockdown of genes that are specifically important for the intrinsic neurophysiological homeostatic response to acute heat stress, would have little impact on fly behavior at permissive temperatures, but would lead to rapid paralysis under acute heat-stress conditions.
To test our hypothesis, we first employed a reverse genetic approach to identify candidate genes specifically involved in the neuronal homeostatic response to acute heat stress. By using a tissue-specific RNAi knockdown screen of voltage-gated potassium channels, we identified seizure (sei), the fly ortholog of the mammalian hERG channel (KCNH2) [26, 29, 32–35], as an essential element in the neuronal homeostatic response to acute heat stress. The sei gene was originally identified in a Drosophila forward mutagenesis screen for temperature-sensitive (ts) structural alleles of essential genes associated with neuronal excitability [33–35]. Although the original screen was designed to identify point-mutations that would lead to proteins that are functional at permissive temperature but would inactivate at non-permissive high temperature due to misfolding, we have recently shown that null alleles of sei and neuronal RNAi knockdown lead to high temperature-induced seizures, as in the original “ts” alleles [35]. These data indicate that the original sei alleles isolated were not true ts alleles. Instead, these data suggest that sei is not required for baseline neural excitability, but does play a role in the ability of neurons to maintain adaptive firing rates when exposed to acute heat stress. Furthermore, we have previously shown that temporal downregulation of sei expression specifically in neurons of the adult fly, or pharmacologically blocking SEI channel activity only in the adult stage, is sufficient to increase susceptibility to acute heat-induced seizures [35]. These data suggest that the impact of sei mutations on stress-induced seizures is primarily a consequence of physiological rather than developmental processes.
Previous studies have indicated that individuals who carry some of the dominant hERG mutations that cause the cardiac Long QT Syndrome (LQTS) [36, 37], often also suffer from high prevalence of generalized seizures [38, 39]. Yet, it is currently assumed that seizures in these patients represent a derived secondary outcome of the primary LQTS cardiac pathology [40–42]. However, the data presented here, as well as previous studies that showed that ERG channels are expressed in mammalian neuronal tissues [43, 44], and contribute to intrinsic spike frequency adaptation in cultured mouse neuroblastoma cells and cerebellar Purkinje neurons [45, 46], suggest that ERG channels also have a specific function within the nervous system.
By utilizing existing and novel genetic tools, here we show that the ERG channel sei is indeed essential for maintaining neuronal robustness under acute heat stress conditions in Drosophila. Specifically, we used an intersectional approach, combining the UAS-GAL4 and LexAOp-LexA binary transgene expression systems [47, 48], to show that sei is broadly expressed in the nervous system, in both neurons and glia. Yet, using RNAi to downregulate sei expression in specific cell types, we demonstrate that the contribution of sei to organismal behavioral resistance to acute heat stress is primarily mediated via its specific action in excitatory cholinergic and glutamatergic neurons, the octopaminergic system, as well as non-neuronal glia. Furthermore, by generating a CRISPR/cas9-derived GFP-tagged allele of the native sei locus, we also show that at the subcellular level, sei exerts its action primarily in axons and associated glia. Together, these studies indicate that mutations in hERG-like potassium channels may contribute directly to the etiology of stress-induced seizures in susceptible individuals by limiting the intrinsic neuronal homeostatic response to acute environmental stressors, possibly via homeostatic axonal spike frequency adaptation.
Previously published theoretical models and empirical studies have indicated that the actions of diverse voltage-gated potassium channels mediate action potential repolarization, and modulate the action potential threshold [49–52], which are important for the homeostatic regulation of synaptic activity and excitability [53–55]. Yet, which genes regulate the intrinsic capacity of neurons to buffer environmentally-induced hyperexcitability is mostly unknown. Thus, we initially hypothesized that the intrinsic ability of neurons to buffer acute heat stress is mediated, at least in part, by the action of specific voltage-gated potassium channels. Because most of these channels are expressed in both neuronal and non-neuronal tissues, we tested our hypothesis by using a neuronal-specific RNAi-dependent knockdown screen of all genes that encode voltage-gated potassium channels in the Drosophila genome. This screen revealed that the threshold to heat-induced seizures is lowered by neuronal knockdown of the genes seizure (sei) and Shab, and raised by neuronal knockdown of Shal (Fig 1A and 1B). These data suggest that different members of the Kv4-type voltage-gated potassium channels play diverse physiological roles in regulating the organismal susceptibility to acute heat stress.
Because previous studies by us and others have shown that mutations in sei promote low neuronal and organismal resistance to acute heat stress [33–35], and sei had the strongest effect on lowering the threshold to heat-induced seizures in our screen (Fig 1A and 1B), we focused our following primary working hypotheses on the contribution of the sei channel to the intrinsic neuronal homeostatic response to acute environmental stress. Thus, we next used a null allele of sei [56, 57] to demonstrate that sei activity is specifically required for the ability of adult flies to resist the impact of acute heat stress (Fig 2A and 2B). In addition to its role in the neuronal response to acute heat stress, we also show that the sei mutation impaired the ability of flies to adapt to a gradual heat stress, where vials of flies were placed in an incubator, starting at room temperature (26°C) and increased 2°C every half hour (Fig 2C and 2D). Together, these data indicate that sei activity is not required for basal neuronal excitability, but is essential for the ability of neurons to maintain stable and adaptive firing rates under fluctuating environmental conditions.
In contrast to adult flies, which often have the ability to escape non-ideal environmental conditions, such as high temperatures, by flying, larvae are much more constrained. Thus, the protective role of sei might be more ecologically relevant to the pre-adult developmental stages. Our data indicate that, as in the adult, sei activity is also necessary for normal larval locomotion under acute heat stress conditions, as tested by placing larvae on a heated agar surface (Fig 2E). Furthermore, because larvae can sense acute nociceptive stimuli, such as heat, via the activity of their cuticular multidendritic (md) sensory neurons [58–60], we next tested the hypothesis that heat induced hyperexcitability in sei mutants would lead to nociception hypersensitivity. Indeed, we found that sei mutant larvae exhibit a significantly faster response when touched with a heated probe, relative to wild type control (Fig 2F), suggesting nociceptive system hypersensitivity, due to hyperexcitability of md sensory neurons, downstream circuits, or both. Together, these studies indicate that the sei channel plays an important role in maintaining neuronal stability and robustness, and protecting Drosophila neurons from environmentally-induced hyperexcitability.
Because previous investigations of the impact of temperature changes on neuronal activity have shown that neurons will respectively increase or decrease their firing rates in response to a rise or fall in ambient temperature [61–63], we next hypothesized that sei mutant flies might be protected from the effect of acute cold stress on neuronal activity. However, we found no effect of the sei mutation on larval locomotion on a 13°C cooled agar surface relative to wild type controls (Fig 2H). Thus, although the precise biophysical role of hERG-type voltage-gated potassium channels in regulating neuronal excitability remains elusive, the in vivo data presented here, as well as previously published in vitro studies [45, 46], indicate that hERG channels play a specific role in maintaining optimal neuronal activity by protecting neurons from environmentally-induced hyperexcitability but not hypoexcitability [35].
Next, we asked whether the susceptibility of sei mutants to heat stress represents a more general sensitivity to any environmental stressor. To answer this, we studied the effect of the sei mutation on survival during exposure to hydrogen peroxide (H2O2), a reactive oxygen species that induces oxidative stress, harming many molecular compounds and processes within cells [64, 65], and shown to lead to death in Drosophila [66, 67]. Surprisingly, we found that sei mutants actually exhibit higher resistance to the toxic effects of hydrogen peroxide relative to wild type controls (Fig 2J). While we do not yet know whether this phenotype is due to nervous system SEI function, nor the mechanism by which sei mutants have a slight resilience, these data nevertheless indicate that the impact of sei mutation in increasing susceptibility to heat stress is relatively specific, and does not generalize to all environmental stressors.
Previous studies by us and others have indicated that the sei gene is expressed in diverse neuronal and non-neuronal cell types, including cardiac and muscle cells [43, 68, 69], and that mutations in sei increase the overall organismal sensitivity to acute heat stress, resulting in shorter latency to heat-induced seizures and paralysis [33–35]. Yet, whether this organismal phenotype is driven by the action of sei in all cell types that express it was unknown. Therefore, we next used tissue-specific RNAi knockdown to determine which cell types require sei activity to protect animals from heat-induced seizures. Similarly to our previous work [35], we found that neuronal-specific knockdown of sei is sufficient to phenocopy the effect of the null allele on the susceptibility of adult flies to heat-induced seizures (Figs 1A, 1B, 3A and 3B). However, we were also surprised to find that, although not as striking as in the neuronal knockdown, the organismal response to heat stress also depends on the activity of sei in non-neuronal glia (Fig 3C and 3D). In contrast, although a previous study has indicated that sei activity is important for heart physiology [69], we found that sei knockdown in the heart or body muscles has no effect on seizure susceptibility (Fig 3E–3H). Together with the strong effect of pan-neuronal sei knockdown on heat-induced seizures, these data suggest that the effects of the sei mutations on seizure susceptibility are independent of sei action in the heart. Furthermore, the organismal resistance to acute heat-stress specifically depends on sei activity in the two primary cell types of the nervous system.
Because the fly CNS is a compact mosaic of different cell types, it is hard to establish whether sei is primarily expressed in neurons or glia. To address this, we generated a transgenic Drosophila line that expresses the LexA activator under the control of the putative sei promoter sequences [48, 70]. We then used this line to express a nuclear localized EGFP reporter (GFPnls) [71, 72], which indicated that sei is broadly expressed throughout the central nervous system (Fig 4A and 4B). To specifically identify sei-expressing glia, we next combined the sei-LexA line driving GFPnls with a red fluorescent protein DsRed including a nuclear localization signal (RedStinger) driven by the glia-specific Repo-GAL4 line [73]. Confocal imaging of co-labeled brains showed that in addition to its broad neuronal expression pattern, sei is also expressed in a small fraction of brain glia (Fig 4C–4E, arrows). These data further indicate that the action of sei in the fly CNS is primarily mediated by its action in most neurons and some glia.
Neurons are comprised of diverse cell types with different physiological properties and varying contributions to systems-level neural excitability. Therefore, we next wished to determine which neuronal subtypes might require sei activity for enabling the organismal response to acute heat stress. A broad screen of sei knockdown using several neuronal type-specific GAL4 driver lines revealed that the organismal response to heat stress depends on the expression of sei in cholinergic (Fig 5A and 5B) and glutamatergic (Fig 5C and 5D) excitatory neurons, but not in GABAergic inhibitory neurons (Fig 5E and 5F). We also observed a significant effect of sei knockdown in the modulatory octopaminergic system (Fig 5G and 5H) but not in the dopaminergic, serotonergic, peptidergic, or the peripheral sensory systems (Fig 5I–5P). Thus, our data suggest that sei plays an important role in protecting the nervous system from environmental stressors that could lead to general hyperexcitability and seizures by maintaining the neuronal robustness of excitatory and some neuromodulatory neurons.
Similarly to neurons, Drosophila glia are comprised of several subtypes based on their localization, cellular morphology, and function [74, 75]. Therefore, we also wished to determine which glia require sei expression for the organismal response to acute heat stress. By screening a collection of recently published glia subtype-specific GAL4 lines [75], we found that the response to acute heat stress specifically depends on sei action in neuropile ensheathing glia (Fig 6A and 6B), and to a lesser extent in perineurial glia (Fig 6C and 6D), but not in astrocyte-like, subperineurial, cortex or tract ensheathing glia (Fig 6E–6L). Although the contribution of SEI channel activity to specific glia subtypes remains unknown, these data suggest that hERG-like potassium currents in some glia play an important role in maintaining organismal robustness to some acute environmental stressors.
Previous studies have shown that in cultured mammalian neurons, ERG-related channels contribute to the regulation of action potential firing rate via spike frequency adaptation, a cell-intrinsic process [45, 46]. These data suggest that sei plays a similar role in fly neurons, which likely explains the heat-induced hyperexcitability and rapid seizure development observed in sei mutant animals. Nonetheless, a previous study suggested that mutations in sei can also lead to a mild increase in axonal branches and boutons at the larval neuromuscular junction (NMJ) [76], which suggests that the observed mutant phenotype may be also driven, at least in part, via synaptic mechanisms. Therefore, we next examined the synaptic morphology of the larval and adult NMJs. We found that although the sei mutation does lead to an increase in the number of branches at the larval NMJ (Fig 7A), it has no effect on synaptic bouton numbers (Fig 7B). Furthermore, we observed no effects of the sei mutation on either synaptic branches or bouton numbers in the NMJs of the adult ventral abdominal muscles (Fig 7E and 7F). Together, these data indicate that the effects of sei mutations on heat-induced seizures in adult flies are primarily mediated via its intrinsic physiological action in neurons rather than via processes associated with synaptic development.
The subcellular localization of various voltage gated ion channels plays an important role in determining how they might be contributing to neuronal signaling and excitability [77–79]. For example, ion channels localized to axons generally impact action potential generation, propagation, and modulation, while those localized to dendrites influence integration of synaptic inputs, propagation of electrical activity to the soma, and action potential backpropagation [77, 80, 81]. Nevertheless, ion channels with specific subcellular enrichment in either dendrites, cell bodies, axons, or presynaptic terminals have all been implicated in human epilepsies [82]. Therefore, we next determined the subcellular localization of native SEI channels by generating a C-terminus GFP-tagged allele of the endogenous sei locus (Fig 8A). We found that the response of homozygous seiGFP flies to heat stress is not different from wild type animals, which indicates that the tagged protein forms wild-type like channels (Fig 8B and 8C). We next used an anti-GFP antibody to probe the subcellular spatial distribution of seiGFP channels in the larval and adult nervous systems. These studies revealed that SEI is primarily localized to the axonal membranes of most neurons, but not in dendrites or somas, as visualized by the lack of colocalization between GFP and the nuclear stain DAPI (Fig 8D–8M), and the enriched sei localization to sensory axonal tracks in the adult brain (Fig 8D), thoracic ganglion (Fig 8F), and motor and sensory neuron axons in larvae (Fig 8H). While the glial function of SEI channels remains unknown, the enrichment of SEI channels in axons supports a model whereby ERG channels contribute to the intrinsic homeostatic regulation of optimal neuronal activity via the modulation of action potentials. This model is further supported by the previously reported influence of mammalian ERG channels on spike frequency adaptation in cultured mammalian neurons [45, 46].
Previous theoretical and empirical studies of neural circuit adaptability, and by extension, the ability of animals to maintain robust and adaptive behavioral outputs in unstable environments, depends on both the intrinsic homeostatic capacity of neurons to maintain an optimal activity pattern, and the ability of neural circuits to maintain stable outputs via the homeostatic regulation of neuronal connectivity and synaptic activity [4–6]. Yet, despite its high incidence, the majority of genetic and molecular factors that regulate neuronal homeostasis, and increase susceptibility to seizures, remain mostly unknown [14–16]. Here, we show that the Drosophila voltage-gated potassium channels sei, Shab, and Shal impact the neuronal homeostatic response to acute heat stress. Furthermore, we show that the organismal capacity to buffer the effects of acute heat stress depends on the independent activity of sei in both neurons and glia. We also found that although sei is broadly expressed in the nervous system, its contribution to the overall organismal resistance to acute heat stress seems to be specifically driven by its action in cholinergic and glutamatergic excitatory neurons, and neuropile-ensheathing glia, as well as to a lesser extent in the modulatory octopaminergic system and perineurial glia. In addition, we developed genetic tools to show that SEI is expressed in the axons of neurons and in glia. Together, our data highlight the important role of sei in the organismal homeostatic response to acute environmental stress, by providing robustness to both the intrinsic activity of specific neuronal populations, and the neural circuits that harbor them. We expect that future work on other voltage gated potassium channel genes, and well as other ion channels and transporters, in both neurons and glia, will continue to shed light on the genetic programs that control robust and homeostatic processes within the nervous system.
How ERG-type potassium channels might contribute to neuronal intrinsic homeostasis during bouts of acute stress is not well understood. Nonetheless, in vivo and in vitro studies in Drosophila and mammalian models have suggested that ERG channels have little effect on baseline neuronal firing rate, but can prevent rapid firing in response to environmental or electrophysiological stimuli that induce hyperexcitability [35, 45, 46]. We have previously shown that in Drosophila motor neurons, basal neuronal firing patterns are unaffected by the sei mutation at optimal 25°C, but become hyperexcitable in response to a rapid temperature increase [35]. Similarly, in electrophysiological studies of mammalian brain slices, in vitro cultured neurons, and heterologously-expressed mammalian ERG channels, pharmacological blockers of hERG channels have little effect on firing rates in response to small current injections, but greatly diminish spike frequency adaptation in response to large current injections, resulting in rapid firing rates [45, 46]. The presence of SEI channels specifically in axons (Fig 8) suggests that they do not affect the propagation of dendritic potentials, but rather limit the rate of action potential generation and propagation, which is sufficient to prevent rapid firing rates. Together, these data suggest a model whereby ERG-like potassium channels play a crucial role in mediating the neuronal homeostatic response to acute stress by protecting neurons from rapid increase in firing rates, and therefore, support neuronal robustness when exposed to extreme environmental fluctuations.
At the neuronal network level, seizures are thought to result from an imbalance between excitatory and inhibitory neural signaling pathways [15, 83]. We found that knocking down sei specifically in all cholinergic neurons, the primary excitatory pathway in the fly central nervous system, is sufficient to phenocopy the effects of sei null mutations on the organismal resistance to heat-induced seizures. These results are similar to previous studies, which showed that increasing activity of the cholinergic system in flies, via genetic manipulations of voltage gated sodium channels and optogenetic neural activation, is sufficient to increase seizure-related and paralytic behavior [84–87]. The simplest interpretation of these data together is that the lack of sei in cholinergic excitatory neurons makes them hypersensitive to heat-induced hyperexcitability, which subsequently surpasses the buffering capacity of the inhibitory neurotransmission pathways, and therefore leads to the rapid development of generalized seizures and paralysis.
Additionally, we observed a large increase in seizure susceptibility when sei is knocked down in glutamatergic neurons. Although we and others have previously demonstrated that the activity of motor neurons, which in insects are primarily glutamatergic, is increased in seizure-susceptible mutant flies [84, 88], this finding suggests that the decreased intrinsic ability of motor neurons to resist acute stress is sufficient for inducing organismal seizure-like phenotype (Fig 5C and 5D). Nevertheless, we currently cannot exclude the possibility that the observed effect of knocking down sei expression in glutamatergic neurons on organismal sensitivity to heat stress is mediated via the action of a small number of modulatory glutamatergic neurons within the central nervous system [89].
We also observed an impairment in the organismal homeostatic response to acute heat stress when sei is specifically knocked-down in the modulatory octopaminergic system (Fig 5G and 5H). Previous studies of the octopaminergic system in Drosophila and other insects have indicated that octopamine and related biogenic amines have broad impact on diverse neuronal processes at the developmental and physiological timescales [90, 91]. Because sei mutant flies seem to have normal behaviors when housed under constant optimal conditions, it is likely that the effects of knocking down sei in octopaminergic neurons on heat-induced seizures are physiological, not developmental. Although we currently do not know which specific elements of the octopaminergic system play a role in the organismal response to acute heat stress, previous work has shown that exogenous application of octopamine in Drosophila increases contraction force of muscles and their response to synaptically driven contractions [92]. Therefore, one possible mechanism by which sei knockdown in octopaminergic neurons might affect observed heat-induced seizures is via the direct modulation of the neuromuscular junction. Octopamine has also been shown to play important roles in the central nervous system, including modulation of behaviors related to motivation, sleep, aggression, social behaviors and learning and memory [90, 93–96]. Therefore, sei knockdown in octopaminergic neurons may result in a broader shift in synaptic processes associated with the homeostatic maintenance of the balance between excitatory and inhibitory pathways under acute heat stress conditions.
The important role of sei activity in regulating the capacity of the nervous system to buffer acute environmental stress is further supported by our discovery that its knockdown in glia also increased susceptibility to acute heat-induced seizures (Figs 3C, 3D and 6). These data are in agreement with previously published studies, which demonstrated that the knockdown of genes associated with ionic homeostasis in glia can increase seizure susceptibility in Drosophila [97–101]. Previous studies have suggested that some glia are important for maintaining synaptic activity and homeostasis, and that disrupting glia functions could contribute to the etiology of seizures because of their role in modulating extracellular potassium concentration, adenosine levels, the size of the extracellular space, and uptake of neurotransmitters [97, 102–106]. Of all the glia, our data indicate that sei is specifically important in neuropile-ensheathing glia. A complete picture of the specific functions of neuropile ensheathing glia has yet to emerge, yet studies manipulating genes in this cell type have implicated roles in phagocytosis of injured neurons [107], organization of neural circuits [108], and glutamate metabolism [109]. However, how the action of voltage gated ion channels such as sei in glia might affect these specific processes remains mostly unknown. Nevertheless, glial expression of another voltage gated potassium channel that is associated with human epilepsy, KCNJ10, has been shown to lead to epileptic activity in a mouse model, possibly via its role in buffering extracellular potassium and glutamate [110, 111]. Whether hERG-like channels play a similar role in glia remains to be explored.
Together, the data we present here provide important insights into the possible role of hERG channels in regulating neuronal robustness and susceptibility to stress-induced seizures. From a clinical perspective, our data suggest that the high incidence of generalized seizures that has been reported in LQTS patients that carry mutations in the hERG genes [38] might not be a secondary cardiogenic comorbidity, as is currently often assumed [40–42]. Instead, because about 40% of patients that carry LQTS-related hERG mutations have reported a personal history of seizures, as compared to less than 20% in LQTS patients with similar cardiac pathologies that are due to mutations in other genes [38], we hypothesize that seizure etiology in many LQTS patients is likely due to the direct impact of mutations in hERG on nervous system functions, independent of their cardiovascular condition. Therefore, we predict that it is possible that some unidentified mutations in hERG might be causally related to epilepsies, independent of the presentation of any LQTS-related pathologies, and may represent novel genetic risk factors for seizures.
The studies we describe here provide compelling evidence that hERG channels play an essential role in protecting the nervous system from acute environmental stressors, such as heat, which could potentially lead to hyperexcitability and seizures. Furthermore, we show that in Drosophila, the activity of the ERG channel sei contributes to neuronal and behavioral robustness via its action in independent cell types in the nervous system. These important insights should help us to better understand how the nervous system responds to acute environmental stressors, and possibly provide important mechanistic insights into some of the known pathologies associated with hERG mutations in human patients.
Flies (Drosophila melanogaster) were raised on standard corn syrup-soy food (Archon Scientific) at 25°C temperature, 70% humidity, on a 12:12 light/dark cycle. Unless specifically noted, wild type control line used was w1118. All fly strains were either produced in the Ben-Shahar lab or obtained from the Bloomington Stock Center (stock numbers in parentheses). UAS-RNAi TRiP lines [112] used in the initial screen included sei (#31681), shab (#25805), eag (#31678), shaker (#53347), shaw (#28346), shal (#31879), elk (#25821) and kcnq (#27252). The TRiP UAS-Luciferase RNAi was used a control (#35789), and UAS-RNAi lines were driven by the elav-GAL4; UAS-Dicer2 line (#25750) (Fig 1A and 1B). For the cell-type-specific sei knockdown screen (Figs 3, 5 and 6), the UAS-RNAi for sei and luciferase were each recombined with UAS-Dcr2. The phenotypic assessment of these RNAi lines indicates that the observed effects were specific to some GAL4 lines but not all, suggesting that the UAS-RNAi transgene alone has no effect on sei expression. The original null seiP allele from Bloomington (#21935) was backcrossed for 6 generations into the w1118 wild type strain (Bloomington #6326) [113]. Other transgenic lines from the Bloomington stock center included: UAS-RedStinger (#8546), LexAOp-GFPnls (#29954), elav-GAL4 [114] (#458), Repo-GAL4 (#7415), hand-GAL4 (#48396), ChAT-GAL4 (#6798), VGlut-GAL4 (#60312), Gad1-GAL4 (#51630), ple-GAL4 (#8848), Tbh-GAL4 (#39939), Trh-GAL4 (#49258), and C929-GAL4 (#25373). BG57-GAL4 [115] and PO163-GAL4 were from the Dickman (USC) and Zlatic (HHMI Janelia Research Campus) labs respectively. The following GAL4 lines were used for glia subtype-specific expression [75]: neuropile ensheathing, R56F03-GAL4 (#39157); tract ensheathing, R75H03-GAL4 (#39908); perineurial, R85G01-GAL4 (#40436); subperineurial, R54C07-GAL4 (#50472); cortex, R54H02-GAL4 (#45784); astrocyte-like, alrm-GAL4 (#67032).
The C-terminus GFP-tagged allele of sei was generated via CRISPR/Cas9-dependent editing by using a modified “scarless” strategy (www.flyCRISPR.molbio.wisc.edu)[116, 117]. Specifically, four sgRNAs TGTAAGCGAATACCACGTTG, GACAGCATTCTCCCGCAACG, GAAGCAGAAGCAGGTAACTC, AGGTGAGTGAGTTACTCATC, which flank the targeted genomic sei sequence were designed using flyRNAi.org/crispr. Complementary oligos that correspond to each individual sgRNA (IDT) were cloned into the pDCC6 plasmid (a gift from Peter Duchek, Addgene plasmid # 59985), which also includes the coding sequence for Cas9 [118], by using the BbsI restriction enzyme (NEB). The donor plasmid for homologous recombination was constructed by using a Golden Gate assembly [119] to recombine four DNA elements: 1) A backbone with ampicillin resistance (pBS-GGAC-ATGC plasmid, a gift from Frank Schnorrer, Addgene #60949)[120]; 2) Eye-specific dsRed reporter driven by the 3XP3 promoter, flanked by two PiggyBac transposase recognition sites, which was PCR-amplified from the pHD-ScalessDsRed plasmid (a gift from Kate O’Connor-Giles, Drosophila Genome Resource Center #1364) using the following primers 5’-CACACCACGTCTCATTAACCCTAGAAAGATAATCATATTGTG-3’ and 5’- CACACCACGTCTCACCCTAGAAAGATAGTCTGCGT-3’ (the primers included BsmBI restriction enzyme sites and overhangs corresponding to the left and right homology arms); 3–4) The left and right homology arms, which consisted of 1kb genomic DNA fragments upstream and downstream of the sgRNA sites respectively. Single base pair mutations were introduced into the PAM sequence at each sgRNA binding site on the homology arms to prevent Cas9-dependent cutting of the donor plasmid. The right homology arm also included the GFP coding sequence in frame with the 3’ end of the last coding exon of sei, immediately upstream of the endogenous stop codon (Fig 8A). pDCC6 plasmids containing sgRNA and cas9 sequences (100 ng/μL) and the donor plasmid (500 ng/μL) were co-injected into y-w- background. Subsequently, correct genomic integration of the GFP tag was verified by screening for DsRed-positive animals, followed by sequencing of genomic PCR fragments. The final tagged sei allele was generated by removing the DsRed cassette via the introduction of the piggyBac transposase (Bloomington #8285) (Fig 8A).
The sei-LexA transgenic flies were generated by amplifying a 2612 bp genomic DNA fragment upstream of the sei start codon by using the following PCR primers: 5’-GTCGACCGCCGGCAAAGTATCAACAT-3’ and 5’-GCGGCCGCTTTTAAGTCTGCAAAGTATAGAAACG-3’, followed by cloning into the pENTR1A plasmid (ThermoFisher) with SalI and NotI restriction enzymes. The sei promoter fragment was then recombined into the pBPnlsLexA::p65Uw vector (a gift from Gerald Rubin, Addgene plasmid # 26230)[48] by using the Gateway reaction (ThermoFisher). The sei-LexA containing plasmid was integrated into the fly genome by using a line carrying a PhiC31 integrase landing site on Chromosome III (Bloomington #24483)[121].
Assays were performed as previously described [35]. In short, two-day old flies were collected and transferred into standard vials containing food (five per sex in each vial). On the following day, flies were flipped into an empty vial, and tested within the next hour. For testing, vials with 10 flies were individually submerged into a 41–42°C water bath and observed for seizure-like behavior and paralysis. The cumulative number of paralyzed flies (immobile at the bottom of the vial) was recorded every 15 seconds. The time at which 50% of the flies in a given vial were paralyzed was used as a measure of seizure susceptibility for statistical analysis.
Two-day old flies were housed in groups of 10 as above. Subsequently, vials were placed in a temperature-controlled incubator (Fisher Scientific Isotemp) with a glass door, which allowed continuous video recording of their behavior. To test for the ability of flies to adapt to gradual temperature increase, flies were first acclimated to 26°C followed by a 2°C increase every 30 minutes to a maximum of 42°C. The number of paralyzed flies was recorded every two minutes.
A 60mm plastic petri dish was filled with 3% agar, and placed on a Peltier plate surface of a PCR machine, which was set to either 37°C or 13°C for the heat or cold stress respectively. Identical tests at 25°C were used as controls. For the heat stress condition, locomotion was assayed by placing individual third instar foraging larvae on the agar surface and video recording (Logitech C920 Webcam) for one minute. Larva locomotion was analyzed by assessing the amount of time spent executing the following specific behaviors: peristaltic locomotion, whipping, head thrashing, rolling and no movement [122](S1 Video). For the cold stress condition, individual larvae were allowed to acclimate for 30s before their behavior was recorded for four min. Behaviors of all larvae were tracked and analyzed by using a custom designed motion tracker system [123], which enabled computer-assisted analyses of distance traveled.
Foraging 3rd instar larvae were removed from bottles, washed in water, and placed on a water-saturated 3% agar plate. Behavioral tests were conducted in constant 25°C and 70% humidity. Larvae were allowed to acclimate to the agar plate for 10 seconds. A custom-made heat probe (Thermal Solutions Controls and Indicators Corporation) set to 50°C was used to gently touch the side of the larva’s body, and the amount of time for the larva to roll over was recorded. For each genotype, 47–50 larvae were individually tested.
At seven days of age, male flies were transferred into empty standard fly vials (Genesee Scientific), housed ten per vial. Daily, 900 uL of 1% sucrose solution, with or without 3% hydrogen peroxide (H2O2), was added to a Kimwipe packed tightly at the bottom of each vial. Survival was assessed every 12 hours.
Analyses of NMJ morphology were done as previously described [124, 125]. In short, animals were dissected to expose larval body wall muscles or adult ventral abdominal muscles. Samples were pinned on sylgard plates, fixed in 4% PFA for 20 minutes, washed with PBST+0.1% Triton-X (PBST), and then blocked with Superblock (ThermoFisher) for one hour. Samples were then incubated with FITC-conjugated goat anti-HRP (123-095-021, Jackson ImmunoResearch) to label neurons, diluted 1:1000 in Superblock, for 1 hour at room temperature. Samples were then washed in PBST, mounted using vectashield (Vector Laboratories), and imaged using a laser scanning confocal microscope (Leica TCS SP5).
Third-instar larval brains were dissected and then fixed, washed and blocked as above. Brains were subsequently incubated overnight at 4°C with rabbit anti-GFP (A-11122, ThermoFisher) diluted in Superblock at 1:1000. After washing with PBST, larval brains were incubated for 2 hours at room temperature with secondary antibody Alexa Fluor 488-conjugated goat anti-rabbit (A-11034, ThermoFisher) and FITC-conjugated goat anti-HRP (123-095-021, Jackson ImmunoResearch) to label neurons, each diluted 1:1000 in Superblock. After secondary incubation, brains were washed in PBST, mounted using FluoroGel II with DAPI (ThermoFisher), and imaged using an laser scanning confocal microscope (Nikon A1Si).
Adult brains were dissected, fixed and blocked as above, then subsequently incubated overnight at 4°C in rabbit anti-GFP (A-11122, ThermoFisher) diluted at 1:1000 and mouse anti-Brp (NC82; Developmental Studies Hybridoma Bank) diluted 1:33 in Superblock. After washing with PBST, adult brains were incubated overnight at 4°C with secondary antibodies Alexa Fluor 488-conjugated anti-rabbit (A-11034, ThermoFisher) and Alexa Fluor rhodamine-conjugated donkey anti-mouse (sc-2300, Santa Cruz Biotechnology), each diluted 1:500 in Superblock. Adult brains were mounted and imaged as above. To image live brains expressing either RedStinger or GFPnls nuclear markers, the tissues were dissected, mounted in PBS, and imaged on a confocal microscope within one hour of the dissection.
Indicated statistical comparisons were analyzed by using Excel (Microsoft) and Prism 7 (GraphPad). Statistical significance was set at p<0.05.
|
10.1371/journal.ppat.1007538 | Urease is an essential component of the acid response network of Staphylococcus aureus and is required for a persistent murine kidney infection | Staphylococcus aureus causes acute and chronic infections resulting in significant morbidity. Urease, an enzyme that generates NH3 and CO2 from urea, is key to pH homeostasis in bacterial pathogens under acidic stress and nitrogen limitation. However, the function of urease in S. aureus niche colonization and nitrogen metabolism has not been extensively studied. We discovered that urease is essential for pH homeostasis and viability in urea-rich environments under weak acid stress. The regulation of urease transcription by CcpA, Agr, and CodY was identified in this study, implying a complex network that controls urease expression in response to changes in metabolic flux. In addition, it was determined that the endogenous urea derived from arginine is not a significant contributor to the intracellular nitrogen pool in non-acidic conditions. Furthermore, we found that during a murine chronic renal infection, urease facilitates S. aureus persistence by promoting bacterial fitness in the low-pH, urea-rich kidney. Overall, our study establishes that urease in S. aureus is not only a primary component of the acid response network but also an important factor required for persistent murine renal infections.
| Urease has been reported to be crucial to bacteria in environmental adaptation, virulence, and defense against host immunity. Although the function of urease in S. aureus is not clear, recent evidence suggests that urease is important for acid resistance in various niches. Our study deciphered a function of S. aureus urease both in laboratory conditions and during host colonization. Furthermore, we uncovered the major components of the regulatory system that fine-tunes the expression of urease. Collectively, this study established the dual function of urease which serves as a significant part of the S. aureus acid response while also serving as an enzyme required for persistent kidney infections and potential subsequent staphylococcal metastasis.
| Bacterial pathogens often encounter acidic environments within host tissues and employ several direct and indirect defense measures [1]. Direct measures include the utilization of proton pumps and generation of alkaline compounds such as ammonia to neutralize pH. Indirect methods such as damage repair, biofilm formation, and metabolic alterations, are utilized to rescue cell viability. Staphylococcus aureus is a leading cause of opportunistic infections in community and health care settings [2, 3]. S. aureus resides in multiple acidic niches during colonization and infection of the human host that include the surface of the skin and within abscesses [4–6]. It is important to note that S. aureus is sensitive to acetic acid stress when growing in the presence of excess glucose [7]. Weak acids such as acetic acid are unique in potentiating stationary phase cell death, in that unlike strong acids that fully dissociate in water, the undissociated weak acids can easily enter into the cytoplasm and reduce the intracellular pH by releasing protons. Therefore, S. aureus must overcome different kinds of acid stress to maintain viability. However, mechanisms of acid resistance in S. aureus are not well described. It has been shown that sodA, which encodes a superoxide dismutase, is induced upon acid stress and facilitates acid tolerance by alleviating the cell damage caused by reactive oxygen species [8]. In addition, a σB–dependent acid-adaptive response has been described that facilitates S. aureus survival in media with a pH of 2 after pre-exposure to a sub-lethal pH of 4 [9]. Based on global transcriptional studies, increased urease activity is thought to be a major contributor to acid resistance in S. aureus [10–13].
In humans, urea is produced in the liver via the urea cycle as a means to remove excess nitrogen. Urea enters the bloodstream, becomes concentrated in the kidneys, and is excreted during urination. The concentration of urea in the blood is normally 2.5–7.1 mM, and it is found in other body fluids such as gastric acid, sweat, and saliva. The level of urea in the saliva is 3–10 mM in healthy individuals but can reach 15 mM in patients with renal diseases [14]. Notably, the re-absorbance of urea from the collecting ducts makes the interstitium of the kidney inner medulla a urea-rich environment.
Urease (EC: 3.5.1.5) is a nickel-dependent metalloenzyme that catalyzes the hydrolysis of urea into ammonia (NH3) and carbon dioxide (CO2) [15–17]. For some bacterial species, urease is an integral part of the bacterial acid response network, as the hydrolysis product ammonia is readily protonated into ammonium (NH4+), during which process protons are consumed, resulting in an increase in pH [1]. Urease is crucial for niche adaptation of many bacterial pathogens. For example, urease is essential for the survival of Helicobacter pylori in the stomach lining, where the pH can be as low as 2.5 [18]. With a high affinity for urea, urease from H. pylori is required not only for the establishment of infections but also for the maintenance of a chronic infection [19]. Also, Streptococcus salivarius produces urease to utilize salivary urea as a nitrogen source for growth while resisting acid stress [20]. Over 90% of S. aureus strains are urease-producing [21], which is encoded by the urease gene cluster ureABCEFGD. The α, β, and γ subunits that comprise the apoenzyme are encoded by ureC, ureB, and ureA, whereas ureEFGD genes encode accessory proteins. Previous studies have shown that urease genes are highly transcribed during biofilm growth conditions [13, 22]. However, the function and utilization of urease in S. aureus has not been comprehensively studied.
In this work, we explored the in vitro and in vivo functions of urease in S. aureus. We found that S. aureus primarily utilizes urease to facilitate pH homeostasis under weak acidic stress, but it does not utilize urea as a nitrogen source under neutral pH. Lastly, our data demonstrate that urease is essential for S. aureus to persist in mouse kidneys, where a significant pH gradient exists and urea is an abundant nitrogen source.
Previous work has documented that aerobic growth of S. aureus in tryptic soy broth (TSB) containing excess glucose (35–45 mM) impairs stationary phase survival of S. aureus [7]. Under these growth conditions, the acetate derived from glucose catabolism is not consumed as a secondary carbon source, and the pH in the medium remains low, which potentiates cell death. Based on those observations, we hypothesized that the presence of urea in TSB containing 45 mM glucose would rescue cell death due to ammonia generation via urease activity. To test this hypothesis, we performed a growth assay in which JE2 wildtype (WT) and JE2 ureB::ΦΝΣ (ureB) were aerobically cultured in TSB containing 45 mM glucose with or without 10 mM urea. Over the period of 120 h, colony forming units (CFU/ml) and extracellular pH were monitored every 24 h. In addition, culture supernatant was analyzed to measure glucose, acetate, urea, and ammonia concentrations. In the absence of urea supplementation, both WT and the ureB mutant showed a drastic decrease in cell viability (~9 log10 difference) while maintaining an acidic extracellular pH (~4.8) (Fig 1A and 1B). However, we observed a urease-dependent increase in viability (~8 log10 difference) and medium pH (~4 pH unit difference) in the presence of exogenous urea (Fig 1A and 1B). Both pH and viability phenotypes of the ureB mutant were complementable by integrating ureABCEFGD into the chromosomal SaPI1 attC site (S1 Fig). Glucose in the medium was depleted by 24 h for both WT and the ureB mutant either with or without urea supplementation (Fig 1C). However, only WT grown in TSB supplemented with 10 mM urea consumed acetate (Fig 1D). Lastly, we observed a urease-dependent consumption of urea coincident with the generation of ~20 mM NH3 (Fig 1E and 1F). Thus, these results suggested that the S. aureus urease functions as part of an acid response network to facilitate pH homeostasis in the presence of urea. Weak acids and subsequent intracellular acidification has been previously shown to generate endogenous reactive oxygen species and potentiate cell death [7]. Thus, to evaluate the physiological status of WT and the ureB mutant in the above growth assay, flow cytometry was performed to assess cellular respiration and reactive oxygen species (ROS) levels. The results confirmed that urease-mediated pH homeostasis rescued cellular respiration and protects cells from endogenous ROS under weak acid stress in the presence of urea at 72 h of growth when viability differences are evident (Fig 1A and S2 Fig). It is important to note that in TSB the concentration of arginine, which can be catabolized to generate NH3 via arginase/urease, nitric oxide synthase (NOS), or two separate arginine deiminase (ADI) systems (Fig 2A), is not sufficient to rescue the survival of JE2 in this assay. Therefore, we repeated the assay with excess arginine (5 mM) in TSB containing 45 mM glucose. As a result, excess arginine was unable to rescue viability of JE2 to the same extent as urea, although we did observe an arginine deiminase-dependent increase in viability (~2 log10) (S3A and S3B Fig). Collectively, these data suggest that urea derived from arginine via RocF and subsequent urease activity is not functional in this assay. However, Staphylococcus epidermidis catabolized either arginine (S3C and S3D Fig) [23] or urea (S3E and S3F Fig) to rescue growth under weak acid stress.
To further investigate the transcriptional regulation of the ure operon, a lacZ reporter plasmid pNF315 was generated in which the promoter of ure was fused to the promoterless lacZ gene and transduced into JE2. The previous experiments (Fig 1) suggested that ure transcription or urease function is induced under weak acid stress [10]. Indeed, as the pH deceased due to the accumulation of acetate, ure transcription was induced 3.3-fold at 6 h comparing to 2 h, in TSB containing 45 mM glucose with or without 10 mM urea (Fig 3A–3C). When the media were buffered to a pH of 7.25 with 100 mM 3-(N-morpholino) propanesulfonic acid (MOPS), the transcription of ure was significantly inhibited regardless of urea supplementation (Fig 3A). These results indicate that the transcription of urease genes is induced by weak acid stress. Multiple global transcriptional studies have suggested that the accessory gene regulator (Agr) quorum sensing system, as well as global regulators CcpA and CodY, function to regulate the transcription of the urease operon [24–26]. To assess these relationships in our model, JE2 WT, ΔccpA::tetL (ΔccpA), Δagr::tetM (Δagr), and ΔcodY::ermB (ΔcodY) each containing pNF315 were grown aerobically in TSB containing 45 mM glucose and 10 mM urea. β-galactosidase activity assays were performed with cell lysate collected during early- (2 h), mid- (6 h), and post- (10 h) exponential phases of growth. The transcription of the ure operon was significantly decreased in ΔccpA/pNF315 (6 h) and Δagr/pNF315 (6 and 10 h), and significantly increased in ΔcodY/pNF315 at 6 and 10 h (Fig 3D), indicating that the transcription of ure genes is activated by CcpA and Agr and negatively regulated by CodY. However, it is unclear if ure transcriptional regulation by CcpA, Agr, or CodY is via direct or indirect regulation. To corroborate the effects of these regulators on urease activity, JE2 WT, Δure, ΔccpA, Δagr and ΔcodY were grown in TSB containing 45 mM glucose and 10 mM urea for 120 h (Fig 3E–3H). As expected based on the transcriptional analysis, ΔcodY essentially phenocopied WT with regards to pH, acetate production and viability (Fig 3E–3H). However, since Δagr displayed reduced ure transcription, it was predicted that the viability would be significantly reduced in the 120-h growth assay. Indeed, the extracellular pH of Δagr was significantly different from WT by 6 h of growth (Fig 3E) and remained at 5.1 from 24–120 h similar to Δure. Further, in comparison to JE2 WT, the viability of Δagr decreased ~6 log10 to ~102 CFU/ml at 120 h similar to Δure. Lastly, based on Fig 3D and the β-galactosidase activity assay documenting a decrease in ure transcription, we would expect the ΔccpA mutant to have decreased viability similar to Δagr and Δure. However, we found that the pH was significantly higher than WT/ ΔcodY from 6–12 h of growth (Fig 3E). In addition, it produced less extracellular acetate than WT (Fig 3F), presumably due to consumption of acetyl-CoA via the tricarboxylic acid (TCA) cycle as CcpA represses TCA cycle activity [27–29]. Further, it survived as well as WT, and the pH remained alkaline over the entire 120 h (Fig 3G and 3H).
Based on our previous work documenting the function of CcpA in repressing amino acid catabolism [30], we hypothesized that the survival of the ΔccpA mutant in the above growth assay was urease-independent due to derepression of amino acid catabolism and subsequent generation of ammonia. Previous investigations have documented that in the presence of a preferred carbon source such as glucose, CcpA represses amino acid catabolic genes including gudB (encoding glutamate dehydrogenase), rocF (encoding arginase), putA (encoding proline dehydrogenase), and arcA1/arcA2 (encoding arginine deiminases) [23, 24, 30–32], all of which produce ammonia as a byproduct. To determine whether the survival of ΔccpA was dependent upon urease or catabolism of a particular amino acid, a growth assay was performed in which JE2 WT, ΔccpA, ΔccpA/gudB::ΦΝΣ, ΔccpA/Δure, ΔccpA/putA::ΦΝΣ, and ΔccpA/arcA1::kan/arcA2::ΦΝΣ were cultured in TSB containing 45 mM glucose. In the absence of urea, JE2 WT was unable to survive as expected, presumably due to a dramatic decrease in pH observed over the 120 h experimental timeframe (Fig 4A and 4B). However, ΔccpA/gudB::ΦΝΣ, ΔccpA/Δure, ΔccpA/putA::ΦΝΣ, ΔccpA/arcA1::kan/arcA2::ΦΝΣ all phenocopied ΔccpA, suggesting that it was not the catabolism of one specific amino acid that was responsible for cell survival in this assay (Fig 4A and 4B). These data led to the hypothesis that the derepression of overall amino acid catabolism provides the ΔccpA mutant a growth advantage. Further, as shown in Fig 3F and previously it is known that ΔccpA mutants produce less extracellular acetate due to derepression of the TCA cycle [27–29]. To address this hypothesis, JE2 WT and ΔccpA were grown in TSB containing 45 mM glucose. As expected, the ΔccpA mutant generated significantly more ammonia and the extracellular acetate was eventually consumed by ΔccpA compared to JE2 WT (Fig 4C). Amino acid analysis of the same supernatant demonstrated that the ΔccpA mutant consumed histidine, aspartate, proline, glutamate, alanine, and arginine at a faster rate than WT (Fig 4D–4H). Taken together, these data suggest that the absence of CcpA in S. aureus facilitates survival in the presence of excess glucose due to decreased acetate generation and increased ammonia generation by amino acid catabolism of multiple amino acids.
In previous experiments using TSB containing 45mM glucose, catabolism of arginine in the media is repressed by CcpA (Fig 4H). Therefore, we were unable to determine the potential function, if any, of endogenously derived urea. Further, it is unclear if urease is active at neutral pH and generates NH3 for use in nitrogen metabolism. To determine if urease is functional in media where arginine is rapidly catabolized thus generating urea from arginine via arginase (RocF), we grew JE2 in buffered chemically defined medium (CDM) lacking glucose [32, 33]. CDM is a defined medium that lacks glucose but contains 18 amino acids except glutamine and asparagine [34], buffered at pH 7.5. We reasoned that if urease catalyzes the reaction generating NH3 from urea, the ammonia would be actively utilized by glutamine synthetase to synthesize glutamine from glutamate (Fig 2A). Both glutamine and glutamate are major amino donors for cellular reactions [35]. Therefore, JE2 WT and the ureABCEFGD deletion mutant (Δure) were grown aerobically in CDM containing 0.1 g/L 15N-arginine (guanidino-labeled only) and cells were harvested at 7 h, at which time both supernatant and intracellular components were assessed by nuclear magnetic resonance (NMR) to detect 15N-glutamine [36]. Since the ammonia generated from amino acid catabolism is also utilized for glutamine synthetase, the rapidly catabolized serine was labeled with 15N as a control [31]. As a result, 17 times more 15N-glutamine was detected when 15N-serine was added to CDM than when 15N-arginine was added (Fig 2B). In addition, no difference was noted when comparing the 15N-glutamine detected from WT or Δure when grown in CDM containing 15N-arginine. Therefore the small amount of detected 15N-glutamine was derived from ADI or NOS-generated 15NH3 (Fig 2B). However, in CDM containing 15N-labeled arginine, significant 15N-labeled urea was detected extracellularly in both WT and Δure (Fig 2C), indicating that the nitrogen from arginine catabolism does not enter the intracellular nitrogen pool, but is rather excreted as urea under neutral pH.
Our data demonstrated that urease facilitates pH homeostasis and cell survival in vitro under weak acid stress in the presence of urea. However, it is unclear if urease functions to facilitate staphylococcal colonization or virulence. One niche where urease may be important is the host skin, where S. aureus resides within hair follicles and sweat glands [37]. Moreover, it is known that human sweat contains 22.2 mM urea [38] and the pH of human skin is ~4 to ~6 [39]. However, animal models of S. aureus skin colonization are difficult to replicate since mice and other rodents do not sweat. Therefore, we reasoned that another host niche where urease might be required was in the colonization of the kidney, which has a low tissue pH and a relatively high concentration of urea. In addition, it is well known that S. aureus causes chronic kidney infections in mice and thus the kidney provides a nidus for subsequent staphylococcal metastasis (S4 Fig) [40]. To test this hypothesis, we used a mouse bacteremia model in which C57BL/6 mice were retro-orbitally injected with JE2 WT and Δure. On days 8, 12 and 19 post-infection, bacterial burden in the kidney was determined (Fig 5). Although no difference between WT and Δure was noted on day 8 (Fig 5A), kidneys infected with Δure had significantly lower bacterial burden on days 12 and 19, with more kidneys below the limit of detection infection compared to WT (Fig 5B and 5C), indicating that urease contributes to the persistence of S. aureus during a mouse chronic kidney infection. To determine if the host immune response differed between mice infected with either WT or Δure, leukocyte populations were assessed from infected kidneys on day 8, an interval where bacterial burdens were equivalent, to prevent bias from animals that had cleared the infection. However, no significant differences were noted between these two groups (S5 Fig) suggesting the absence of ure did not skew the immune response to facilitate enhanced clearance.
Acid stress, along with other environmental risk factors such as extreme temperatures, osmotic pressure, and nutrient depletion, is challenging for bacterial survival [41]. Accordingly, a variety of strategies are utilized to resist low pH which also contribute to bacterial virulence [1]. In Escherichia coli, four main acid resistance systems (ARs) have been described: the oxidative system AR1, the glutamate-dependent AR2, the arginine-dependent AR3, and the lysine-dependent AR4 [42]. The amino acid-dependent ARs are composed of a decarboxylase which consumes protons, and an inner membrane antiporter which imports the decarboxylase substrate while exporting the product. Bacillus cereus activates not only the general stress response genes via σB but also proton transporters and amino acid decarboxylases, as well as the ADI system that produces ammonia [43]. H. pylori is known for the ability to proliferate in extremely low pH environments such as the host gastric acid [44, 45]. In addition to amino acid catabolism, enhanced urease activity sustains favorable intracellular pH by generating ammonia [18, 45–48]. Mycobacterium tuberculosis resists acid stress through nitrogen assimilation from asparagine hydrolysis [49], as well as urea hydrolysis [50]. However, there is little known about acid resistance in S. aureus although it proliferates in multiple mildly acidic niches of the human host.
In the aerobic growth assay of S. aureus cultured in TSB containing 45 mM glucose, the acetate produced via glucose catabolism is excreted into the growth medium (Fig 1D). When the extracellular pH of the growth medium reaches the pKa of acetic acid (4.8), acetate becomes protonated and is able to traverse the cell membrane and release the proton in the near neutral pH of the cytoplasm. The drop in intracellular pH potentiates cell death in S. aureus by intracellular acidification and ROS generation [7]. We documented that in the presence of urea, urease functions to facilitate pH homeostasis in a weak acid environment through the generation of ammonia that inhibits the acetate-dependent intracellular acidification (Fig 1E and 1F) [7]. In addition, it was confirmed that the urease activity and subsequent NH3 generation rescued cellular respiration and prevented endogenous ROS generation (S2 Fig). The ammonia generated facilitated acetate consumption through acetyl-CoA synthetase and the TCA cycle for subsequent cell growth. Collectively, these data suggest that urease is a significant component of the acid response network of S. aureus in the presence of urea.
Many gram-positive species, including S. epidermidis, utilize arginine catabolism as a rapid mechanism to generate ammonia during acidic pH stress [23, 51]. The pathway most utilized is the ADI pathway yielding ammonia, ornithine, and ATP. However, in contrast to S. epidermidis [23] (S3C and S3D Fig), excess arginine was unable to remarkably rescue S. aureus during weak acid stress suggesting that ammonia generating pathways via arginine are not significantly active in our aerobic assay containing glucose (S3A and S3B Fig). These results agree with a previous report documenting that the ADI pathway genes are not significantly induced under weak acid stress in S. aureus [10]. In this work we also confirmed that in S. epidermidis 1457, additional arginine and urea provided a growth advantage under weak acid stress (S3C and S3E Fig), suggesting that both ADI and urease are active in S. epidermidis. This result is consistent with another study documenting the differential transcriptional response following sapienic acid stress in S. epidermidis and S. aureus [52]. Under these growth conditions, S. aureus upregulates urease whereas S. epidermidis upregulates ADI, the NreABC nitrogen regulation system, in addition to the nitrate and nitrite reduction pathways. The lack of NH3 generation via arginine catabolism in S. aureus is not unexpected as catabolism of arginine via RocF is under the control of carbon catabolite repression and CcpA [30, 31]. Our results suggest that when S. aureus is growing in the presence of glucose or another preferred carbon source, urease must utilize exogenous urea to facilitate pH homeostasis. Previous results from our laboratory demonstrated that when S. aureus grows in a defined medium lacking glucose, arginine is catabolized via RocF generating ornithine and urea [31]. Thus, we wanted to determine if urease was active under growth conditions where urea was generated via arginine catabolism and the medium was not acidic. These NMR experiments suggested that under neutral growth conditions, little ammonia from urea is detected (via detection of 15N-labeled glutamine). In fact, the majority of urea was detected in the culture medium as it is excreted, and is potentially used as a nitrogen storage molecule. Indeed, we found that a ccpA mutant grown in TSB lacking urea was able to survive weak acid stress via derepression of global amino acid catabolism. This observation suggests that when S. aureus is growing in acidic environments where peptide and amino acids are the major carbon source, arginine catabolism and urease activity is not required to facilitate pH homeostasis, which is due to rapid catabolism of amino acids and subsequent NH3 release.
Previous transcriptional analyses have suggested that ureABCEFGD is upregulated upon acid stress [10, 11, 53]. In the current study, we confirmed via β-galactosidase assays that the transcription of the ure genes was inhibited when the medium was buffered to a pH of 7.25 with MOPS (Fig 2A). Moreover, we found that the transcription of the urease genes was activated by CcpA and Agr, while inhibited by CodY (Fig 2B). The regulation of urease transcription by CcpA is supported by the putative catabolite-responsive element (cre) site identified 139 base pairs upstream of the ureA start codon [24]. Also, CcpA activation of ure transcription agrees with a previous finding that S. aureus urease, as a part of the CcpA regulon, has higher transcription as well as enzymatic activity in WT as compared to a ΔccpA mutant [24]. The significant decrease in ure gene transcription in the Δagr mutant that we observed is consistent with the transcriptional array data documenting that urease genes are upregulated by Agr [25]. The involvement of the urease genes in the Agr regulon strengthens the link between urease activity and virulence in S. aureus. Importantly, phagosomal acidification induces Agr activity, which is essential for S. aureus survival inside macrophages [54]. Under this circumstance, Agr may upregulate urease to counter acidic pH in coordination with enhanced virulence. CodY is also a global regulator that controls the expression of a variety of genes in gram-positive bacteria [55]. In particular, CodY senses the level of branched-chain amino acids and intracellular GTP and controls the transcription of many metabolic genes that are involved in amino acid synthesis, TCA cycle, and carbon overflow metabolism [56]. Our results agreed with previous reports that CodY represses urease gene transcription in Bacillus subtilis [57, 58] and S. salivarius [26]. Although urease genes are not the direct targets of CodY in S. aureus UAMS-1 [59], it is possible that CodY negatively regulates urease gene transcription through repressing Agr, since the agrA gene encoding the Agr response regulator is upregulated in the UAMS-1 ΔcodY mutant [59]. More in-depth future studies are required regarding the regulation of urease, as other transcriptional regulators such as Sae [60, 61], ClpP [62–64], and MgrA [65], are suggested to contribute to the regulatory network that fine-tunes urease activity. Lastly, it is interesting that CcpA activates ure transcription but represses rocF (arginase) transcription. These data suggest that the generation of urea by arginase is not linked to urease activity. Thus, ure transcription is activated when S. aureus is growing with a preferred carbon source such as glucose, which generates weak acids such as lactate or acetate. However, this also suggests that the urea utilized by urease must be exogeneous and not generated by arginase activity, which is only active when S. aureus is growing on non-preferred carbon sources such as peptides and amino acids.
To interrogate the function of urease in vivo, we hypothesized that the kidney is a favorable niche for S. aureus colonization and resisting the host immune response for the following reasons. First, renal blood flow is about 20% of the cardiac output [66]. Thus, S. aureus has ample opportunities to invade kidney tissue during blood filtration. Second, as urea becomes concentrated when transported through renal tubules during the production of urine, the collecting ducts in the inner medulla display the highest permeability to urea [67]; hence, not only the renal tubules but also the medullary interstitium is rich in urea, providing sufficient substrates for urease. Third, kidney medullary interstitium has a low pH (~5.5), comparing to the neutral cortical interstitium pH (~7.4) [68, 69]. In our mouse S. aureus bacteremia model, the temporal distribution of the organ bacterial burden followed what has been previously documented (S4 Fig) [40, 70, 71]. Among all the examined organs, the kidney was the only niche that developed a chronic infection over time. On days 12 and 19, mice inoculated with the Δure mutant had a significant decrease in CFU count in the kidneys (Fig 5), indicating a selective pressure in S. aureus to maintain urease function, similar to what has been reported in H. pylori [19]. The increased persistence of JE2 WT over Δure during chronic kidney infections demonstrated that urease functionality enhances the fitness of S. aureus within low pH and high urea environments such as the kidney. In order to determine whether differences in leukocyte recruitment are responsible for the changes in bacterial persistence between JE2 WT and Δure, individual kidneys were analyzed by flow cytometry (S5 Fig). Overall, no significant changes in the leukocyte populations were observed, indicating that urease primarily enhances bacterial persistence rather than directly altering leukocyte infiltration. The reason why we chose to evaluate the leukocyte populations on day 8 was that the kidney infections started to be cleared on approximately day 12 (Fig 5). Thus the drastic differences in the bacterial burden between JE2 WT and Δure could skew the immune responses on day 12. Further studies are required to determine the mechanism of how urease facilitates survival during infection. It is also possible that urease allows for persistence in the phagolysosomes upon phagocytosis by macrophages [72]. As macrophages are found in renal medulla [73], S. aureus needs to employ strategies to survive within or escape from the phagolysosomes during colonization in the kidney. Moreover, the phagolysosomes are acidic in pH, which may induce urease activity for acid resistance and survival of S. aureus [74]. Indeed, anti-inflammatory macrophages, which are prevalent during late stages of S. aureus infection, produce urea via arginase-1 [75]. For the above reasons, it would be appropriate to expand our future studies to examine the function of urease in phagolysosome survival or the escape of S. aureus, especially regarding kidney macrophages.
In summary, we identified that urease in S. aureus functions to facilitate pH homeostasis and survival under weak acid stress in the presence of urea; in non-acidic conditions, the endogenous urea derived from arginine is secreted extracellularly but not catabolized to fuel nitrogen metabolism. We found that urease is induced by weak acid stress and is within the regulation network that consists of CcpA, Agr, and CodY, interconnecting S. aureus stress response, metabolism, and virulence. We illustrated that urease provides a fitness advantage for S. aureus to persist during chronic kidney colonization of mice. These data all point to the conclusion that urease is not only a critical component of the acid stress response system of S. aureus, it is also an important factor in S. aureus pathogenesis.
Animal experiments were performed in ABSL2 facilities in accordance with a protocol (#11-076-08-FC) approved by the Institutional Animal Care and Use Committee (IACUC). All animals at the University of Nebraska Medical Center are maintained in compliance with the Animal Welfare Act and the Department of Health and Human Service “Guide for the Care and Use of Laboratory Animals.” Animals were anesthetized with ketamine and xylazine. Post injection of S. aureus retro-orbitally, all anesthesized mice were continuously monitored until they regained sternal recumbency and were capable of holding their heads up. The animals were monitored once/day on a daily basis following infection to ensure animal welfare. At all monitoring intervals, post-infection general appearance and body weights were recorded. Animals were euthanized by exposure to CO2 in a chamber (chamber was not pre-charged). Animals were in the CO2 filled chamber for 5 minutes after all evidence of respiration and cardiac function was absent. CO2 was chosen as a method of euthanasia because it has a rapid anesthetic effect and quickly results in loss of consciousness and respiratory arrest.
The E. coli, S. aureus, and S. epidermidis strains, plasmids, as well as primers used in this study are listed in S1 Table. E. coli cultures were grown in Luria-Bertani broth (LB; Difco; Becton, NJ). S. aureus and S. epidermidis were grown in tryptic soy broth (TSB; Difco; Becton, NJ) containing 14 or 45 mM glucose. CDM was prepared essentially as previously described [34], and no glucose was added. Overnight cultures grown in TSB were washed with phosphate-buffered saline (PBS) twice before inoculation to an optical density at 600 nm (OD600) of 0.05. Cultures were grown aerobically at 37°C in flasks with a 10:1 flask-to-volume ratio shaking at 250 rpm. When necessary, antibiotics were added to cultures as follows: ampicillin (50 μg/ml); erythromycin (10 μg/ml); tetracycline (10 μg/ml); and chloramphenicol (10 μg/ml). Bacterial growth yield was assessed by measuring the OD600. Culture pH was measured with a pH meter (Mettler Toledo, Columbus, OH). Bacterial viability was measured as CFU/ml by serial dilutions on TSB agar plates.
PCR amplifications were performed using Q5 High-Fidelity DNA polymerase (New England Biolabs, Beverly, MA), Midas Mix (Monserate Biotechnology Group, San Diego, CA), and oligonucleotides (S1 Table) synthesized by Sigma-Aldrich (St. Louis, MO). Restriction endonucleases and ligase from New England Biolabs (Beverly, MA) were used for DNA digestion and ligation. Purification of DNA fragments prior to subsequent cloning steps was achieved by recovery from agarose gels using a DNA Clean and Concentrator-5 Kit (Zymo Research, Orange, CA). Recombinant plasmids were purified using a Zyppy Plasmid Miniprep Kit (Zymo Research, Orange, CA). All plasmid inserts were sequenced at Eurofins Genomics (Louisville, KY) to ensure the absence of mutations.
The reporter plasmid pNF315 was constructed by amplifying the intergenic region upstream of ureA with primers 2833 and 2835 so that the native ribosomal binding site (RBS) was replaced with a plasmid-encoded RBS. The DNA fragment was digested and ligated into the BamHI and XhoI sites of the vector plasmid pJB185, which contains a promoterless lacZ [76]. pNF315 was electroporated into S. aureus RN4220 and subsequently transduced into JE2 strains using bacteriophage Φ11-mediated transduction [77].
To create the markerless JE2 Δure mutant, the allelic exchange plasmid pNF320 was generated by inserting the DNA sequences 1 kb upstream and 1 kb downstream of the ureABCEFGD operon into the temperature-sensitive E. coli-S. aureus shuttle vector plasmid pJB38 [78], using a NEBuilder HiFi DNA Assembly Cloning Kit (New England Biolabs, Beverly, MA), with primers 2980 and 2983, as well as primers 2986 and 2987. pNF320 was electroporated into S. aureus RN4220 and subsequently transduced into JE2 WT using bacteriophage Φ11-mediated transduction. Once the plasmid pNF320 was introduced into JE2, the allelic replacement to introduce the deletion mutation into the S. aureus chromosome was performed as previously described [79]. The deletion of the urease operon was confirmed phenotypically by plating on a Christensen's urea agar plate and by PCR using primers 2984 and 2985.
Chromosomally complementation of Δure was performed as previously described [80]. Briefly, plasmid pNF363 was constructed containing ureABCEFGD genes with their native promoter by amplifying an approximately 5.5 kb region from the JE2 genome using primers 2991 and 3306. The resulting DNA fragment was inserted into BamHI and PstI sites of the shuttle vector pJC1111 yielding pNF363. pJC1111 and pNF363 were subsequently transformed into RN9011 for chromosomal integration. Φ11 mediated transduction was performed to move the integrated pJC1111 and pNF363 into both JE2 WT and JE2 ureB::ΦΝΣ.
For all metabolite assays, 1 ml bacterial culture was collected and pelleted for 2.5 min at 15,000 rpm. The supernatant was collected and stored at -80°C until use. Glucose, acetate, urea, and ammonia concentrations were determined using commercial kits (R-Biopharm AG, Darmstadt, Germany) according to the manufacturer's instructions.
As previously described [31, 36], five independent 50 ml cultures of S. aureus JE2 WT and the Δure mutant were grown to stationary phase (OD600 = 1.9) in CDM containing 15N2-labeled arginine (Isotec, Sigma-Aldrich, Miamisburg, OH) or 15N-labeled serine (Isotec, Sigma-Aldrich, Miamisburg, OH). For each culture, a total OD600 of 40 was collected and pelleted by centrifugation at 4000 rpm for 5 min at 4°C. 2 ml culture supernatant was collected as the media sample. Pellets were washed with 10 ml of cold sterile water twice and resuspended in 1 ml cold sterile water. The cells were lysed using a bead ruptor (OMNI International, Kennesaw, GA) and centrifuged at 15,000 rpm for 15 min at 4°C. The pellet was re-extracted with 1 ml cold sterile water. The combined cell lysate supernatant from both extractions, as well as the culture supernatant, were snap frozen in liquid nitrogen and lyophilized using a FreeZone freeze dryer.
The data collection and analysis of NMR was conducted as previously described [36]. A Bruker AVANCE IIIHD 700 MHz spectrometer equipped with a 5 mm quadruple resonance QCI-P cryoprobe (1H, 13C, 15N, and 31P), an automatic tune and match system (ATM), and a SampleJet automated sample changer system with Bruker ICON-NMR software were utilized. The 2D 1H−15N HSQC spectra collected for S. aureus cell lysates and culture media were assigned using a database of 2D 1H−15N HSQC reference spectra for known metabolites [36]. A chemical shift tolerance of 0.08 ppm for 1H and 0.25 ppm for 15N were used to match metabolites to our reference database.
Beta-galactosidase assays were performed essentially as previously described [81]. Briefly, overnight cultures of JE2/pNF315 grown in TSB were inoculated in TSB containing 45 mM glucose, TSB containing 45 mM glucose buffered with 100 mM MOPS, TSB containing 45 mM glucose with 10 mM urea, and TSB containing 45 mM glucose with 10 mM urea buffered with 100 mM MOPS. At 2 h and 6 h, 2 ml and 0.5 ml of cells were collected and centrifuged (Fig 3A). Additionally, overnight cultures of JE2/pNF315, ΔccpA/pNF315, Δagr/pNF315, ΔcodY/pNF315 grown in TSB were inoculated to TSB containing 45 mM glucose and 10 mM urea. At 2 h, 6 h, and 10 h, 2 ml, 0.5 ml, and 0.5 ml of cells were collected and centrifuged respectively (Fig 3D). The cell pellets were resuspended in 1.2 ml Z-buffer (60 mM Na2HPO4, 40 mM NaH2PO4, 10 mM KCl, 1 mM MgSO4, 50 mM β-mercaptoethanol, pH 7.0) and lysed with a bead ruptor (OMNI International, Kennesaw, GA). 700 μl supernatant of the cell lysate was collected, and 140 μl of 4 mg/ml ortho-nitrophenyl-β-galactoside (ONPG) was added. The samples were incubated at 37°C until the color turned slightly yellow (under OD420 1.0). 200 μl of 1 M Na2CO3 was added to stop the reaction. Protein concentrations were determined by Bradford assays using the Protein Assay Dye Solution (Bio-Rad, Hercules, California). Absorbances at 420 nm and 550 nm were measured with an Infinite 200 plate reader (Tecan, Männedorf, Switzerland).
Overnight cultures of JE2 WT and ΔccpA were inoculated to an OD600 of 0.05 in TSB containing 45 mM glucose. At 0 h, 3 h, 6 h, 9 h, and 12 h, 0.5 ml culture was collected and pelleted for 3 min at 15,000 rpm. Supernatant was collected and filtered through the Pierce Protein Concentrators (3,000 molecular weight cutoff; Thermo Scientific, Rockford, IL) according to the manufacturer’s instructions. Amino acid analysis was performed with a Hitachi L-8800 amino acid analyzer by the Protein Structure Core Facility, University of Nebraska Medical Center.
The flow cytometry analyses following the growth assays were performed as previously described [7], using a BD LSRII flow cytometer (Becton and Dickinson, San Jose, California). Cells collected at 24 h and 72 h from the growth assay where JE2 WT and ureB::ΦΝΣ were cultured in TSB containing 45 mM glucose and 10 mM urea. Cell samples were washed with PBS to a final concentration of 107 cells/ml and stained with 5-cyano-2,3-ditolyl tetrazolium chloride (CTC, 5 mM) and 3-(p-hydroxyphenyl) fluorescein (HPF, 15 μM). The fluorescence-activated cell sorting (FACS) was performed at a flow rate of ∼1,000 cells per second with 10,000 events per sample. Samples were excited at 488 nm, with HPF emission being detected at 530±30 nm, and CTC emission being detected at 695±40 nm. The FlowJo software was used to analyze the raw data.
For the flow cytometry analyses following the animal experiments, kidneys were collected in 1.0 mL of FACS buffer, which was composed of PBS and 2% heat-inactivated fetal bovine serum (FBS). Kidneys were homogenized with the blunt end of a 3.0 mL syringe and filtered through a 70 μm filter (BD Falcon, BD Biosciences). Filtrate was washed with PBS and collected by centrifugation (300 x g, 5 min), whereupon the filtrate was digested with Collagenase A and DNase while mixing at 37°C. The reaction was stopped after 15 minutes with heat-inactivated FBS on ice, filtered, and washed with FACS buffer, whereupon red blood cells (RBC) were lysed using the RBC Lysis Buffer (BioLegend, San Diego, CA). Single cell suspensions were washed and resuspended in FACS buffer and incubated with TruStain fcX (BioLegend, San Diego, CA) to minimize non-specific antibody binding. Samples were divided in two to analyze innate immune cell (MDSCs, neutrophils, monocytes, and macrophages) populations and lymphocyte (CD3, CD4, CD8, and γδ T cells) populations separately. Both samples were stained with Live/Dead Fixable Blue Dead Cell Stain (Invitrogen, Eugene, OR). Innate immune cells were stained with CD45-APC, Ly6G-PE, Ly6C-PerCP-Cy5.5, and F4/80-PE-Cy7, CD11b-FITC (BioLegend, San Diego, CA). Lymphocytes were stained with CD45-PE-Cy7, CD3-APC, CD4-PacBlue, CD8-FITC, and γδTCR-PE. An aliquot of pooled cells was stained with isotype-matched control antibodies to assess the degree of non-specific staining per treatment group [82]. For individual samples, 10,000–100,000 events were analyzed using BD FACSDiva software with cell populations expressed as percentage of total viable CD45+ leukocytes.
Seven-week-old male and female C57BL/6 mice (Charles River Laboratories, Wilmington, MA) were used in all animal experiments. Overnight cultures of S. aureus JE2 WT and Δure in TSB were washed with PBS twice and suspended in PBS to yield an OD600 of 10. The cultures were further diluted 1:50 with PBS, prior to the retro-orbital injection of 50 μl (106 CFU) final bacterial suspension. The inocula were verified by serial dilution plating and colony enumeration on TSB agar plates. Mice were anesthetized by intraperitoneal injection of ketamine/xylazine (60 mg/kg and 3 mg/kg, respectively). Mice were euthanized for the quantification of bacterial burden (expressed as Log [(CFU/g of tissue) +1]), by serial dilution plating and colony enumeration of homogenized organs.
For all studies, statistical analysis was performed using GraphPad Prism 5.0 software (La Jolla, CA). P-values < 0.05 were considered significant. For comparisons of two groups, Mann-Whitney test was used. One-way analysis of variance (ANOVA) was performed to compare three or more groups. Two-way repeated measures ANOVA was performed to compare differences between groups with two independent variables.
|
10.1371/journal.ppat.1003535 | MAP Kinase Phosphatase-2 Plays a Key Role in the Control of Infection with Toxoplasma gondii by Modulating iNOS and Arginase-1 Activities in Mice | The dual specific phosphatase, MAP kinase phosphatase-2 (MKP-2) has recently been demonstrated to negatively regulate macrophage arginase-1 expression, while at the same time to positively regulate iNOS expression. Consequently, MKP-2 is likely to play a significant role in the host interplay with intracellular pathogens. Here we demonstrate that MKP-2−/− mice on the C57BL/6 background have enhanced susceptibility compared with wild-type counterparts following infection with type-2 strains of Toxoplasma gondii as measured by increased parasite multiplication during acute infection, increased mortality from day 12 post-infection onwards and increased parasite burdens in the brain, day 30 post-infection. MKP-2−/− mice did not, however, demonstrate defective type-1 responses compared with MKP-2+/+ mice following infection although they did display significantly reduced serum nitrite levels and enhanced tissue arginase-1 expression. Early resistance to T. gondii in MKP-2+/+, but not MKP-2−/−, mice was nitric oxide (NO) dependent as infected MKP-2+/+, but not MKP-2−/− mice succumbed within 10 days post-infection with increased parasite burdens following treatment with the iNOS inhibitor L-NAME. Conversely, treatment of infected MKP-2−/− but not MKP-2+/+ mice with nor-NOHA increased parasite burdens indicating a protective role for arginase-1 in MKP-2−/− mice. In vitro studies using tachyzoite-infected bone marrow derived macrophages and selective inhibition of arginase-1 and iNOS activities confirmed that both iNOS and arginase-1 contributed to inhibiting parasite replication. However, the effects of arginase-1 were transient and ultimately the role of iNOS was paramount in facilitating long-term inhibition of parasite multiplication within macrophages.
| Toxoplasma gondii is a protozoan (single cell) parasite that can be transmitted to humans via infection stages in cat feces or from eating under-cooked meat containing parasite cysts. The disease, which infects some 30% of the world population is a major cause of congenital infection and abortion and can be fatal in immune compromised hosts. Control of parasite growth has been shown to be largely, but not wholly, dependent on the production of nitric oxide (NO) as a result of the enzymatic activity of inducible nitric oxide synthase (iNOS) on the substrate L-arginine. It has been suggested that arginase-1 by depleting the L-arginine store can promote parasite growth. We have recently shown MAP kinase phosphatase-2 (MKP-2) to increase iNOS and decrease arginase-1 levels and mice lacking this gene were found to be more susceptible to T. gondii. Inhibition of iNOS activity in genetically intact mice demonstrated it was of paramount importance in controlling parasite growth. Surprisingly, however, inhibition of arginase-1 in MKP-2−/− mice demonstrated arginase-1 was also protective. Significantly T. gondii cannot make its own arginine and requires this from the host cell and consequently iNOS and arginase-1 can work together to starve the parasite of this essential metabolite.
| Toxoplasma gondii is an obligate intracellular protozoan parasite of significant public health importance, being a major cause of congenital infection and abortion as well as a significant and often fatal infection in immune compromised hosts. The early acute stage of infection is characterized by widespread tachyzoite dissemination and tissue damage. The rapid onset of immunity, initiated in large part by the well characterized T. gondii pathogen associated molecular patterns (PAMPS) controls parasite replication [1]–[4], and results in the life-long chronic stage of infection associated with encystment of the parasites in skeletal muscle and the central nervous system [reviewed in 5]. Protection against acute disease is mediated primarily by the interaction of neutrophils, dendritic cells, macrophages and natural killer (NK) cells that as part of the innate response not only limits parasite growth, but initiates an effective cytotoxic CD8+ T cell response that is responsible for long-term protection and prevention of encephalitis via IFN-γ production [6]–[9]. The mechanisms by which IFN-γ, the major effector cytokine mediating resistance during T. gondii infection, promotes anti-toxoplasma activity are not yet fully clear and vary between host species studied. Several IFN-γ-regulated genes including iNOS [6], [10], indoleamine 2,3 dioxygenase (IDO) [11], [12], and more recently, p47 GTPases, have been implicated in playing significant roles in mediating these protective responses [13]–[21].
Despite being essential to control parasite replication, an overactive type-1 response and overproduction of IFN-γ, TNF-α and NO can result in severe pathology and death. Consequently, protective immunity to T. gondii that needs to effectively control parasite proliferation without excessive inflammation is dependent on the regulation of type-1 responses by Th2 cells and Treg cells [22]–[29]. An overactive type-2 response could equally promote enhanced parasite proliferation and host death and indeed alternative macrophage activation has been shown to promote parasite growth [30]. Interestingly, the classic hallmark indicator of alternative macrophage activation, arginase-1, can be induced innately by T. gondii apparently via both STAT-6 dependent [30]–[32] and independent mechanisms [30], [33]. Paradoxically, arginase-1 has been associated with enhancing infection with T. gondii by competing with iNOS for their common substrate L-arginine [33] and promoting parasite replication by providing the polyamines needed for cell division [32]. Conversely arginase-1 might also limit parasite growth by starving the parasite of the L-arginine it requires for this process [31], [34].
A recent study has demonstrated that the dual specific phosphatase, MKP-2 is not only a negative regulator of macrophage arginase-1 but also a positive regulator of iNOS expression [35]. Furthermore MKP-2−/− C57BL/6 mice have been found to display enhanced susceptibility to the intracellular parasite Leishmania mexicana. Susceptibility in MKP-2−/− mice was in large part due to enhanced parasite growth that could be reversed by inhibiting arginase-1 activity. Given the importance of NO production as well as the apparently contradictory roles of arginase-1 in murine T. gondii infections it is likely that the role of MKP-2 in T. gondii infection would be one of major significance. We consequently studied the course of T. gondii infection in MKP-2−/− and MKP-2+/+ C57BL/6 mice. We identified that MKP-2 deficiency results in increased susceptibility to T. gondii infection and that this correlated strongly with impaired iNOS activity in vivo. Conversely we demonstrated an arginase-1 dependent mechanism responsible for control of parasite growth that functioned to partially protect the host independently of iNOS mediated effects.
All animal procedures conformed to guidelines from The Home Office of the UK Government. All work was covered by two Home Office licences: PPL60/3929, “mechanism of control of parasite infection” and PPL60/3439,“genetic models of cancer and inflammation” with approval by the University of Strathclyde ethical review panel.
MKP-2+/+ and MKP-2−/− mice on a C57BL/6 background were bred and maintained and all experiments carried out in house at the Strathclyde Institute of Pharmacy and Biomedical Sciences, Glasgow, UK. Six to eight week old, male mice were used for infection and aged matched within each experiment.
Beverley cysts were maintained in vivo by intraperitoneal (i.p.) passage of infective brain tissue homogenates through outbred CD1 albino mice. For experimental infections MKP-2+/+ and MKP-2−/− mice were infected i.p. with 10 tissue cysts in 200 µl sterile PBS.
Tachyzoites were routinely maintained in confluent human foreskin fibroblasts (HFFs) grown in DMEM complete medium comprising; Dulbecco's Modified Eagle Medium (DMEM) containing L-glutamine (Invitrogen, UK), 10% foetal calf serum (PAA, UK), 100 U/ml penicillin (Cambrex Bioscience, Veniers, Belgium), 100 µg/ml streptomycin (Cambrex Bioscience, Veniers, Belgium) and 50 U/ml amphotericin B (Cambrex Bioscience, Veniers, Belgium) at 37°C in 5% CO2.
In vivo parasite burden was assessed using bioluminescent imaging using type II Prugniaud T. gondii transfected with the firefly luciferase [36]. The light data was quantified using Living Image software (Caliper Life Science)
Briefly, intracellular FLUC T. gondii was harvested from in vitro culture. Culture media was removed and the HFF monolayer washed once with sterile PBS (Lonza, UK) prior to disruption with a cell scraper and harvesting in 10 ml sterile PBS. The intracellular parasites were then released from the host cells by passage through a 21 gauge needle and centrifuged at 1200 rpm for 10 minutes. The bioluminescent activity of the FLUC T. gondii was quality controlled for each experiment prior to infection using a standard curve in a black-walled 24 well plate by in vitro imaging using the IVIS Spectrum (Caliper Life Sciences) (Figure S1). Mice were then infected with 20,000 tachyzoites i.p. in a volume of 400 µl sterile PBS. FLUC T. gondii infected mice were imaged using the IVIS Spectrum (Caliper Life Sciences) to determine parasite burden. Mice to be imaged were given 150 mg/kg of D-luciferin potassium salt solution i.p. prior to anaesthesia with isoflurane. For optimal imaging, 1 minute exposures were taken with medium binning, 20 minutes post luciferin injection. For ex vivo imaging of the brain, mice were injected with 150 mg/kg of D-luciferin potassium salt solution. After 10 minutes the mice were sacrificed by CO2 inhalation. The brains were removed and soaked in 15 mg/ml D-luciferin potassium salt, dissolved in warmed RMPI 1640, for a further 10 minutes before being imaged.
RH tachyzoites where harvested from acutely infected BALB/c mice, by intra-peritoneal washout and lysate antigen (TLA) was prepared as described previously [37].
Spleen cell suspensions from infected mice were prepared as described previously [24]. For re-stimulation, 2×105 cells were incubated with 5 µg TLA in a total volume of 200 µl for 72 h at 37°C in 5% CO2. Supernatants were then used for IFN-γ, IL-5, IL-6 and IL-10 ELISA, using paired antibodies from BD Bioscience. Absorbance was the read at 405 nm using a microtitre plate reader (Spectramax 190, Molecular Devices, USA). Concentrations were determined against standard curves for each cytokine (R&D Systems).
Serum nitric oxide levels were determined by Griess assay. Blood was collected by cardiac puncture and cells removed by centrifugation at 13,000 rpm for 10 minutes. Protein was removed from the serum by adding ZnSO4 to a final concentration of 15 mg/ml, vortexing thoroughly and centrifuged at 13,000 rpm for 10 minutes. The supernatant was retained for the Griess assay. Griess reagent (equal volumes of 2% sulphanilamide in 5% H3PO4 and 0.2% Napthylene diamine HCL in ddH2O) was added to samples and standards in a 96 well plate and incubated in the dark for 10 minutes. Absorbance was read at 540 nm and serum nitrite concentrations were calculated against a standard curve.
To inhibit iNOS and Arginase-I in vivo, mice received intraperitoneal doses of N (G)-nitro-L-arginine methyl ester (200 mg/kg, L-NAME, Sigma UK) or Nω-hydroxy-nor-Arginine (100 µg nor-NOHA, Merck, UK), respectively [38], [39]. Treatment commenced one day prior to infection.
Bone marrow derived macrophages were cultured by flushing the femurs and tibiae of 6 week old MKP-2+/+ or MKP-2−/− mice with DMEM. Cells were then cultured in DMEM containing 20% heat inactivated FCS, 30% L-cell conditioned medium, 5 mM L-glutamine, 100 U/ml penicillin and 100 µg/ml streptomycin and incubated at 37°C for 10 days [35]. After this time adherent cells were harvested and then seeded into either 12 well plates (1×106 cells/ml) or black-walled 96 well plates (0.5×105).
For infection, 5×104 bone marrow derived MKP-2+/+ or MKP-2−/− macrophages were seeded into black-walled 96 well plates in complete phenol red free RPMI 1640 medium. The cells were infected with 5×104 type II Prugniaud strain T. gondii tachyzoites, expressing YFP (Donated by Marcus Meissner, University of Glasgow). Wells were made up to a final volume of 200 ml. Macrophages were treated with LPS (100 ng/ml), IFN-γ (100 U/ml), 50 µM nor-NOHA or 5 mM L-NAME as appropriate. Levels of YFP expression were assayed at 24, 48 and 72 h using the transillumination feature of the IVIS Spectrum (Caliper Lifescience). An excitation wavelength of 500 nm and emission wave of 540 nm was used. The light data was quantified using Living Image software (Caliper Lifescience) using the uninfected macrophages as background controls.
For re-stimulation, 1×106 splenocytes were incubated with 5 mg/ml TLA, ionomycin and phorbol-13-myristate-12-acetate (PMA, Sigma, UK) (0.5 µg/ml and 10 ng/ml, respectively) for 6 h at 37°C and 5% CO2. After 3 h incubation Brefeldin A (Sigma) was added for a further 3 h at a final concentration of 10 ng/ml. Cells were then harvested for staining and incubated with Fc-Block (5 µg/ml αCD16/CD32, BD Bioscience, 1% mouse serum in RPMI) for 10 minutes at room temperature, cells were stained for 1 hour at 4°C with αCD3-PerCP, αCD4-APC-H7 (BD Biosciences) and αCD8-Alexa488 (ebioscience). Cells were fixed using Fix & Perm (Life Technologies) following manufacturer's instructions. Intracellular cytokine staining was performed using αTNF-PE and αIFN-γ-APC (eBioscience). A total of 500,000 events per sample were acquired on a BD FACSCanto and data analysis was carried out using FlowJo software.
Natural killer or NKT cells from intraperitoneal exudates (1×106) from infected mice were surface stained with αPanNK-APC and αCD3-PerCP-eFluor710 before fixing and permeabilisation. Intracellular staining for IFN-γ was performed using αIFN-γ-PE or the respective isotype control, rat IgG1,κ-PE (all eBioscience). A total of 200,000 events was recorded and analysed using Kaluza software (Beckman Coulter). All values were normalized to their respective isotype control.
Cell lysates (20 µg protein/lane) were separated by 10% SDS-PAGE and transferred onto a nitrocellulose membrane and probed for Arginase-1 as previously described [35]. Arginase activity from murine BMD-macrophages was measured using an assay based on a reaction with α-isonitrosopropiophenon (ISPF) as described previously [35].
Peritoneal exudates from mice were counted and 1×106 cells from each sample taken for RNA extraction using RNeasy Mini Kit (Qiagen, UK) following manufacturer's instructions. For each reaction, 2 µg of total RNA and 300 ng of random primers (Promega) were incubated at 65°C for 5 minutes. Following a 10 minute incubation at room temperature 2 µl of AffinityScript reverse transcriptase (RT) buffer (Agilent Technologies), 10 mM dithiothreitol, 4 mM dNTP mix (Promega) and 1 µl of AffinityScript multiple RT (Agilent Technologies) were added to each sample. These were then incubated for 10 minutes at 25°C, 1 hour at 50°C and 15 minutes at 70°C.
qRT-PCR experiments were conducted on the Stratagene Mx3000p system (Stratagene, Agilent). Each reaction consisted of 5 µl Brilliant III Ultra-fast SYBR Green QPCR Master Mix (Agilent Technologies), 3.5 µl of molecular grade water (Life Technologies), 25 pmol of the forward and reverse primers (Table S1), and 1 µl of the cDNA template. PCR reactions were carried out with the following thermoprofile: one cycle at 95°C for 10 minutes, 40 cycles of 20 seconds at 95°C, 20 seconds at 64°C and 30 seconds at 70°C. Relative gene expression was calculated based on the ΔCT for each gene, normalized to that of the housekeeping gene (TBP).
Statistical analysis was performed using GraphPad Prism Program (Version 5.0, GraphPad Software, California). All results are shown as standard error of the mean (SEM). Statistically significant differences were determined using students t-test for parametric and Mann Whitney U test for non-parametric data, respectively. P values equal or below 0.05 were considered significant.
Intraperitoneal infection with 20 cysts of the Beverley (type-II) strain of T. gondii resulted in significant mortality, typically between 80–100%, 15–25 days post-infection in MKP-2−/−, but not MKP-2+/+ mice (Figure 1A). In order to determine whether this was associated with any inability of the MKP-2−/− mice to control parasite growth, MKP-2−/− and MKP-2+/+ mice were infected i.p. with 20,000 type-II strain Prugniaud-FLUC tachyzoites and parasite burdens measured at 2 day intervals by bioluminescent imaging (Figure 1B and C). At days 8 (Figure 1B) and 10 the bioluminescent intensity and consequently parasite burdens were significantly greater (p<0.031 and p<0.0079 respectively) in MKP-2−/− compared with MKP-2+/+ mice (Figure 1C). Greater parasite numbers in infected MKP-2−/− as opposed to MKP-2+/+ mice during acute infections were confirmed by qRT-PCR (Figure S2) The increased susceptibility of MKP-2−/− mice infected with Prugniaud-FLUC tachyzoites compared with their wild-type counterparts was also evident during the chronic phase of the disease: mortality was typically 40–60% in MKP-2−/− mice and by day 30 post-infection the excised brains of MKP-2−/− mice were revealed to have significantly higher parasite burdens (p<0.02) than similarly infected MKP-2+/+ animals (Figure 1D and E). Increased parasite burdens in infected MKP-2−/− compared with MKP-2+/+ mice were also evident in other tissues at this late stage (Liver: MKP-2−/− 4.76×105±4.7×103p/s, MKP-2+/+ 3.85×105±4.6×104p/s; Kidney: MKP-2−/− 1.31×106±4×104p/s, MKP-2+/+ 4.76×105±1.3×105p/s p<0.05).
In order to determine whether the increased susceptibility of MKP-2−/− mice infected with T. gondii was the result of an impaired adaptive immune response spleens were removed and splenocytes stimulated with TLA at days 10, 20 and 30 post-infection and IFN-γ, IL-4, IL-5 and IL-10 production measured in the supernatants. There were generally no differences between the ability of splenocytes from MKP-2−/− or MKP-2+/+ mice infected with T. gondii to produce cytokines when stimulated with parasite lysate antigen. This was consistent through both acute and chronic infection for the primary mediator of parasite control IFN-γ (Figure 2). There was also no difference in the absolute number of cytokine producing cells between MKP-2−/− and MKP-2+/+ mice (Figure S3). Flow cytometry analysis (Figure 3) of splenocytes from MKP-2−/− and MKP-2+/+ mice day 10 post-infection either under resting conditions, following antigen stimulation, or following treatment with ionomycin and PMA revealed no differences in the overall numbers of CD4+ or CD8+ T cells. In addition no differences in the frequencies of CD4+ (Figure 3B) or CD8+ (Figure 3C) T cells producing IFN-γ, or TNF-α, or both IFN-γ and TNF-α (Figure 3) were noted. In addition, while no IFN-γ was detected in NK cells equivalent amounts of NKT cells producing IFN-γ were detected from both infected MKP-2−/− and MKP-2+/+ mice (Figure S4). Consequently, there was little evidence that MKP-2−/− mice were limited in their ability to mount a type-1 response. Furthermore we examined the expressions of IFN-γ dependent GTPases by qRT-PCR and found these to be comparable in both infected MKP-2−/− and MKP-2+/+ mice (Figure S2).
Serum nitrite levels were significantly higher (p<0.0028) in MKP-2+/+ mice compared with MKP-2−/− at day 10 post-infection coincident with the noted inverse differences in T. gondii parasite burdens between MKP-2−/− and MKP-2+/+ mice (Figure 4A). Nitrite levels were below the level of detection in both non-infected MKP-2−/− and MKP-2+/+ mice. At the same time spleen arginase-I expression was higher in MKP-2−/− compared with MKP-2+/+ mice in both, the uninfected as well as in the T. gondii infected groups (Figure 4B). Infection with T. gondii in general increased splenic arginase-I levels in all mice independent of the genotype (Figure 4B).
In order to determine whether the enhanced NO production by T. gondii infected MKP-2+/+ mice over their MKP-2 deficient counterparts was responsible for their increased resistance to parasite growth and survival in vivo, NO production was inhibited in infected mice by intraperitoneal injection with the iNOS inhibitor L-NAME. While L-NAME treatment of infected MKP-2+/+ mice resulted in 100% mortality by day 10 post-infection, surprisingly the majority of L-NAME treated infected MKP-2−/− mice survived until day 12 post-infection (Figure 5A). No nitrite was detectable in L-NAME treated mice. All non-treated infected MKP-2−/− and MKP-2+/+ mice survived until day 12 (Figure 5A) and in agreement with previous studies at day 10 there were significantly more parasites in MKP-2−/− mice (p<0.03). Measurement of bioluminescence (Figure 5B) allowed quantification of parasite burdens and demonstrated a significant increase in parasite growth at days 6 and 8 post-infection (day 6 p = 0.017, day 8 p = 0.005) in L-NAME treated as opposed to non-treated MKP-2+/+ mice (Figure 5C). By comparison treatment of infected MKP-2−/− mice with L-NAME did not significantly alter parasite growth (Figure 5C).
Unlike their wild-type counterparts, inhibition of NO production did not increase parasite growth or the early mortality of MKP-2−/− mice infected with T.gondii. This implied an alternative mechanism of controlling early infection in the absence of MKP-2. As T. gondii is an L-arginine auxotroph [40] we suspected that over expression of arginase-1 in MKP-2−/− mice could also be playing a protective role. To determine the possibility that arginase-1 expression might arrest parasite growth through arginine depletion, infected MKP-2−/− and MKP-2+/+ mice were treated with the arginase-1 inhibitor nor-NOHA. Nor-NOHA did not significantly increase mortality in either MKP-2−/− nor MKP-2+/+ mice infected with T. gondii compared with non-drug treated controls (Figure 6A). However, measurement of bioluminescence (Figure 6B) demonstrated significantly greater (p = 0.02 parasite burden day 8 post-infection in nor-NOHA treated as opposed to non-treated MKP-2−/− mice (Figure 6C). This is despite nor-NOHA treatment up-regulating iNOS activity in MKP-2−/− mice as measured by increased nitrite production by as much as ×2. By comparison nor-NOHA treatment had little effect on parasite burden during early T. gondii infection in MKP-2+/+ mice (Figure 6C).
We next determined whether the increased susceptibility of MKP-2−/− mice to infection with T. gondii was a result of their macrophages being more permissive to parasite growth. MKP-2−/− and MKP-2+/+ bone marrow derived macrophages were infected with YFP expressing Prugniaud T. gondii tachyzoites at a 1∶1 ratio. In addition the macrophages were either treated with LPS and IFN-γ with or without L-NAME or nor-NOHA. Parasite growth was then determined by assessing YFP fluorescence. At 24 hours, through to 72 hours post-infection parasite growth was similar in non-stimulated MKP-2−/− and MKP-2+/+ macrophages (Figure 7A). In addition, following LPS/IFN–γ stimulation parasite growth was significantly, and equally, reduced over the non-treated macrophages (p = 0.0001) and this was irrespective of whether the macrophages were derived from MKP-2+/+ or MKP-2−/− mice (Figure 7A). Treatment of LPS/IFN–γ activated MKP-2−/− and MKP-2+/+ macrophages with L-NAME partially ablated their ability to control parasite replication at 24 hours and totally ablated their ability to control parasite growth through 48 hours and 72 hours post-infection (p<0.0001) (Figure 7A). Conversely, treatment of LPS/IFN-γ activated MKP-2−/− and MKP-2+/+ macrophages with nor-NOHA partially ablated their ability to control parasite growth at 24 hours (MKP-2+/+ p = 0.005 and MKP-2−/− p = 0.006 respectively) but not at 48 or 72 hours post-infection (Figure 7A). While LPS/IFN-γ activation stimulated NO production, as measured by nitrite production 72 hours post-infection in the supernatants of T. gondii infected MKP-2−/− and MKP-2+/+ macrophages, this was significantly higher in the infected MKP-2+/+ culture supernatants (p<0.0001) (Figure 7B). Treatment of LPS/IFN-γ activated macrophages with L-NAME, but not nor-NOHA, largely ablated NO production by both MKP-2−/− and MKP-2+/+ macrophages (p<0.0001) (Figure 7B). T. gondii tachyzoite infection was found to enhance macrophage arginase-1 expression (Figure 8A) although this was consistently higher in MKP-2−/− than in MKP-2+/+ macrophages up to 24 hours post-infection. Increased expression of arginase-1 following macrophage infection with T. gondii (Figure 8A) was also reflected in up-regulated enzyme activity at 24 hours post-infection in both MKP-2−/− (p<0.05) and MKP-2+/+ macrophages (Figure 8B). Arginase-1 activity was significantly greater in MKP-2−/− than in MKP-2+/+ infected macrophages (p<0.001 for non-activated macrophages and p<0.0001 for LPS/IFN–γ activated macrophages) but was effectively and equally inhibited by nor-NOHA in both MKP-2−/− and MKP-2+/+ macrophages. Interestingly while L-NAME treatment significantly increased arginase-1 activity in non-activated MKP-2−/− and MKP-2+/+ macrophages (p<0.02 for both) and LPS/IFN-γ activated MKP-2−/− cultures (p<0.01) this was not noted in activated MKP-2+/+ macrophages (Figure 8B).
The present study demonstrates an important role for MKP-2 in controlling infection with T. gondii as infected MKP-2−/− C57BL/6 mice were found to be less able to control parasite growth during both acute and chronic infection as well as displaying increased mortality compared with their wild-type counterparts. The enhanced susceptibility of MKP-2−/− mice was associated with increased tissue arginase-1 expression, generally associated with Th2 responses, and at the same time lowered serum nitrite levels, generally associated with type-1 responses. Nevertheless, increased susceptibility was not associated with any significant modifications of the adaptive immune response between MKP-2 deficient and wild-type mice and the type-1 responses generated in infected MKP-2−/− mice were at least as strong as in their MKP-2+/+ counterparts. Rather, the recently identified unique feature of MKP-2 as a negative regulator of macrophage arginase-1 expression and a positive regulator of macrophage iNOS expression [35] would appear to underlie the important role of this member of the dual specific phosphatase family in controlling infection with T. gondii. While highlighting the importance of iNOS and NO production in controlling T. gondii infection, the present study also uncovered a protective role for arginase-1 in controlling parasite multiplication that can compensate for NO deficiency during early infection.
Generally protective immunity against T. gondii is associated with a type-1 response where IFN-γ synergizes with TNF-α to activate macrophages and induce the expression of inducible nitric oxide synthase (iNOS) that catalyzes the formation of nitric oxide (NO) from L-Arginine. While NO can directly kill the parasites [6], [10] some studies have also shown that NO promotes tachyzoite conversion to the much slower dividing bradyzoite form of the parasite through inhibition of mitochondrial and nuclear enzymes essential for parasite respiration [41]. Although iNOS seems to be the predominant pathway used by classically activated macrophages to control T. gondii proliferation in tissue culture, the role of NO during in vivo infection is less clear. Studies using iNOS-deficient mice have shown that mice lacking iNOS are able to survive and control tachyzoite growth during the acute stage of infection via an IFN-γ dependent mechanism and only succumb during chronic toxoplasmosis [42]. Death was associated with uncontrolled proliferation of tachyzoites in the brain, suggesting that the protective anti-toxoplasmic effect in the brain was iNOS-dependent [42], [43]. Nevertheless, the observation that iNOS deficient mice were able to survive acute infection in an IFN-γ mediated manner suggested that there were alternative pathways other than NO production mediating anti-toxoplasma resistance in vivo [44]. In addition, a further IFN-γ and NO independent control of T. gondii has been identified in IFN regulatory factor-1 gene deficient mice [45].
In the present study we found that L-NAME treatment of wild-type mice infected with 20,000 type-II Prugniaud-FLUC tachyzoites but not MKP-2−/− mice resulted in enhanced parasite burden and mortality early in infection. This indicated that an NO independent mechanism was playing a protective role in MKP-2−/− mice and controlling parasite growth under conditions where NO generation was being inhibited. Previously induction of the IFN-γ inducible gene IDO has been implicated in mediating some of the IFN-γ dependent NO independent anti-Toxoplasma activity [11]. IDO catalyzes the degradation of the essential amino acid L-tryptophan through the kynurenine pathway, thereby depriving the parasite of this essential amino acid [12]. Interestingly the relative contributions of iNOS and IDO to parasite control appear to be tissue specific [12]. More recently immunity-related GTPases (IRGs) have emerged as potent effectors of T gondii killing in mice [13], [14]. Thus, murine astrocytes have been shown to have the ability to kill intracellular T. gondii independently of iNOS and IDO via IFN-γ inducible IRGs [18]–[20]. Different IRGs have been shown to play roles in controlling acute and chronic infection [16], [17] although the mechanisms through which p47 GTPases confer resistance to T. gondii infection have not been determined [21]. However, as no difference in IFN-γ production was noted between infected MKP-2+/+ and MKP-2−/− mice infected with T. gondii this would suggest GTPase production was not associated with the NO independent resistance demonstrated by MKP-2−/− mice. We have now examined qRT-PCR the expression of the IFN-γ inducible genes LRG47 and IGTP GTPase in infected MKP-2−/− and MKP-2+/+ mice and indeed found them comparable. Consequently, this would suggest that an IFN-γ independent mechanism was operating to protect MKP-2−/− mice in the absence of NO production.
Butcher and colleagues (2011) have recently demonstrated that the type-1 strain T. gondii parasites deficient in ROP16 have enhanced growth in macrophages, and in vivo infection results in increased parasite multiplication and dissemination in the host. This has been associated with the inability of ROP16 deficient parasites to induce STAT-6 mediated arginase-1 expression resulting in maintenance of host cell L-arginine which is exploited by the parasite. It has previously been demonstrated that T. gondii is an L-arginine auxotroph [40] and parasite multiplication in the host cell is L-arginine dependent. Previously studies have demonstrated that polymorphisms in ROP-16 from type-2 parasites result in their inability or greatly reduced ability to activate STAT3/6 [46], [47]. However, in the course of the present study these type-2 parasites were shown to increase arginase-1 expression and activity in both MKP-2+/+ and MKP-2−/− macrophages. While often thought of as a Th2 associated product of alternative macrophage activation, innate macrophage activation via TLR-4 ligation [48], [49] has also been shown to induce arginase-1 expression. As T gondii has been demonstrated to have a number of TLR-4 ligands such as GPI anchors [50] and HSP70 [51], [52] this is the likely source of the STAT-6 independent induction of arginase-1 by type-2 strain parasites as demonstrated in the present study. In vivo treatment with nor-NOHA to inhibit arginase-1 activity during the course of the present study resulted in enhanced parasite multiplication in MKP-2−/− mice. This indicated that the enhanced expression of arginase-1 expression in these mice could in some part compensate for iNOS deficiency compared with MKP-2+/+ animals.
Many studies suggest that macrophage killing of parasites via iNOS catalyzed NO production is the main mechanism by which T. gondii parasite multiplication is controlled [33], [53] and that L-arginine depletion by arginase-1 counter-regulates the effectiveness of iNOS and facilitates parasite growth. Our in vitro studies utilizing classically activated BMDM clearly demonstrate that iNOS catalyzed NO production plays the major role in controlling parasite growth and this could be reversed by treatment of cultures with L-NAME. This was irrespective of whether MKP-2−/− or MKP-2+/+ macrophages were examined. Interestingly despite producing less NO than activated MKP-2+/+ macrophages, activated MKP-2−/− macrophages were equally able to control parasite growth. Indeed inhibition of arginase-1 activity by nor-NOHA treatment did not facilitate parasite growth inhibition in activated MKP-2−/− or MKP-2+/+ macrophages, but in fact reversed inhibition of parasite growth early under conditions of activation. This clearly indicates a role for arginase-1 in protection against T. gondii. Surprisingly no differences in intracellular parasite growth in MKP-2−/− versus MKP-2+/+ macrophages in vitro were noted, suggesting the differential weighting of the alternative control mechanisms were compensating for each other over the course of the experiment. Nevertheless, the in vivo consequences of MKP-2 deficiency are significant in terms of parasite growth and long-term survival, and this would be in keeping with NO playing a role post-acute infection. It is also likely that cells other than macrophages are contributing to the dysregulated iNOS/arginase-1 expression bias in vivo in infected MKP-2−/− mice. Likely candidates would be neutrophils: these have previously been shown to play a significant role in early T. gondii infections [54] and have been shown to be major sources of arginase-1 activity in vivo and influence the disease process in humans [55]. This is currently under investigation in the murine model.
Our studies on the consequences of MKP-2 deficiency have revealed some fascinating insights into the control of T. gondii infection. Firstly NO induction is ultimately of paramount significance in controlling parasite multiplication and host survival. However, in the absence of NO production enhanced arginase-1 is able to in part compensate for this deficiency presumably by starving the parasite of L-arginine as previously demonstrated [31]. Previously it has been suggested that arginase-1 stimulates parasite growth by converting L-arginine to the polyamines needed by the parasite [32] or inhibiting iNOS activity by L-arginine depletion [33]. We could find no evidence for this. Rather we would propose that arginase-1 and iNOS work together to control parasite multiplication by a combination of L-arginine starvation (arginase-1 and iNOS) and NO killing (iNOS).
Overall our results demonstrate that MKP-2 through its ability to reciprocally modulate arginase-1 and iNOS expression is a key regulator in L-arginine metabolism and consequently this has clear consequences for the control of intracellular parasites. Furthermore as arginase-1 has also been shown to have potent T cell modulatory effects [55], [56] MKP-2 influences are likely to have significant consequences for inflammatory disease and cancers where arginase-1 and iNOS have already been identified as key players [57]–[59]. These observations identify manipulation of MKP-2 expression or activity as a significant target for future therapeutic strategies.
|
10.1371/journal.pcbi.1005799 | Multivariate pattern dependence | When we perform a cognitive task, multiple brain regions are engaged. Understanding how these regions interact is a fundamental step to uncover the neural bases of behavior. Most research on the interactions between brain regions has focused on the univariate responses in the regions. However, fine grained patterns of response encode important information, as shown by multivariate pattern analysis. In the present article, we introduce and apply multivariate pattern dependence (MVPD): a technique to study the statistical dependence between brain regions in humans in terms of the multivariate relations between their patterns of responses. MVPD characterizes the responses in each brain region as trajectories in region-specific multidimensional spaces, and models the multivariate relationship between these trajectories. We applied MVPD to the posterior superior temporal sulcus (pSTS) and to the fusiform face area (FFA), using a searchlight approach to reveal interactions between these seed regions and the rest of the brain. Across two different experiments, MVPD identified significant statistical dependence not detected by standard functional connectivity. Additionally, MVPD outperformed univariate connectivity in its ability to explain independent variance in the responses of individual voxels. In the end, MVPD uncovered different connectivity profiles associated with different representational subspaces of FFA: the first principal component of FFA shows differential connectivity with occipital and parietal regions implicated in the processing of low-level properties of faces, while the second and third components show differential connectivity with anterior temporal regions implicated in the processing of invariant representations of face identity.
| Human behavior is supported by systems of brain regions that exchange information to complete a task. This exchange of information between brain regions leads to statistical relationships between their responses over time. Most likely, these relationships do not link only the mean responses in two brain regions, but also their finer spatial patterns. Analyzing finer response patterns has been a key advance in the study of responses within individual regions, and can be leveraged to study between-region interactions. To capture the overall statistical relationship between two brain regions, we need to describe each region’s responses with respect to dimensions that best account for the variation in that region over time. These dimensions can be different from region to region. We introduce an approach in which each region’s responses are characterized in terms of region-specific dimensions that best account for its responses, and the relationships between regions are modeled with multivariate linear models. We demonstrate that this approach provides a better account of the data as compared to standard functional connectivity in two different experiments, and we use it to discover multiple dimensions within the fusiform face area that have different connectivity profiles with the rest of the brain.
| Cognitive tasks recruit multiple brain regions [1–4]. How do these regions work together to generate behavior? A variety of methods have been developed to study connectivity both in terms of the anatomical structure of the brain [5], and of the relations between timecourses of responses during rest [6] and during specific experimental tasks [7–11]. Functional Magnetic Resonance Imaging (fMRI) has proven to be a valuable instrument in this enterprise, offering noninvasive recording with good spatial resolution and whole-brain coverage.
In parallel to this literature, multivariate pattern analysis (MVPA; [12]) has drastically increased the potential of fMRI for the investigation of representational content, making it possible to detect information at a level of specificity that was unthinkable with previous univariate analyses [13–17]. Despite the success of MVPA, relatively few attempts have been made to transport the potential of multivariate analyses to the domain of dynamics and connectivity.
A recent study [18] used trial-by-trial classification accuracy of color and shape in area V4 and in the lateral occipital complex (LOC) to predict trial-by-trial accuracy of object classification in the anterior temporal lobe (ATL). Earlier work by the same group [19] used a continuous measure of classification based on correlations, offering a richer description of each brain region’s patterns. These studies are important steps towards exploiting the wealth of information encoded in patterns of BOLD response to study connectivity, but they both characterize the information encoded in a brain region using a single measure (a given classification), rather than in terms of values along multiple dimensions.
An additional property of both these methods [18, 19] is that they use classification along experimenter defined categories. This approach can be useful to probe a specific hypothesis about a given classification. However, it might disregard other information encoded by the regions studied which is orthogonal to the categories chosen by the experimenter. As a consequence, the results depend on the experimenter’s choice of the categories, and on how well the chosen categories capture the functional role of the regions studied.
Multivariate pattern dependence (MVPD) is a novel method to investigate the ‘connectivity’ between brain regions in terms of multivariate spatial patterns of responses. In keeping with the statistical literature [20], we will replace the term ‘connectivity’ with the term ‘statistical dependence’, which we consider more accurate. MVPD is composed of three main stages. In the first stage, the representational space in each brain region is modeled extracting a set of data-driven dimensions (rather than chosen by the experimenter), that correspond to spatial response patterns that ‘best’ characterize that region’s responses over time. In the second stage, the multivariate timecourses of responses in each region are reparametrized as trajectories in the representational spaces defined by these dimensions. In the third stage, the multivariate relations between the trajectories in the representational spaces of different regions are modelled. In a procedure analogous to MVPA, independent data are used to train and test the models. The dimensions and the parameters modelling the relationship between two regions are estimated with all runs but one, and then used to model the relation between those regions in the remaining run.
We demonstrate the potential of MVPD in two different experiments, analyzing the statistical dependence between the posterior superior temporal sulcus (pSTS) during the recognition of faces and voices, and of the fusiform face area (FFA) during the recognition of faces. In both experiments, MVPD identified dependencies between regions not detected by standard functional connectivity, and explained more variance in individual voxels responses than univariate methods. In the end, MVPD revealed different connectivity profiles associated with different dimensions of FFA’s responses.
The volunteers’ consent was obtained according to the Declaration of Helsinki (BMJ, 1991, pp. 302, 1194). The project was approved by the Human Subjects Committees at the University of Trento and Harvard University.
The data were collected on a Bruker BioSpin MedSpec 4T at the Center for Mind/Brain Sciences (CIMeC) of the University of Trento using a USA Instruments eight-channel phased-array head coil. Before collecting functional data, a high-resolution (1 × 1 × 1 mm3) T1-weighted MPRAGE sequence was performed (sagittal slice orientation, centric phase encoding, image matrix = 256 × 224 [Read × Phase], field of view = 256 × 224 mm2 [Read × Phase], 176 partitions with 1 mm thickness, GRAPPA acquisition with acceleration factor = 2, duration = 5.36 minutes, repetition time = 2700, echo time = 4.18, TI = 1020 msec, 7° flip angle). Functional data were collected using an echo-planar 2D imaging sequence with phase oversampling (image matrix = 70 × 64, repetition time = 2000 msec, echo time = 21 msec, flip angle = 76°, slice thickness = 2 mm, gap = 0.30 mm, with 3 × 3 mm in plane resolution). Over three runs, 1095 volumes of 43 slices were acquired in the axial plane aligned along the long axis of the temporal lobe.
Data were preprocessed with SPM12 (http://www.fil.ion.ucl.ac.uk/spm/software/spm8/) and regions of interest were generated with MARSBAR [25] running on MATLAB 2010a. Subsequent analyses were performed with custom MATLAB software. The first 4 volumes of each run were discarded and all images were corrected for head movement. Slice-acquisition delays were corrected using the middle slice as reference. Images were normalized to the standard SPM12 EPI template and resampled to a 2 mm isotropic voxel size. The BOLD signal was high pass filtered at 128s and prewhitened using an autoregressive model AR(1). Outliers were identified with the artifact removal tool (ART), using both the global signal and composite motion. Datapoints exceeding experimenter-defined thresholds were removed from the analysis. An additional noise-removal step was performed with CompCorr [26]. In each individual participant, a control region was defined combining the white matter and cerebrospinal fluid masks obtained with SPM segmentation, and five principal components were extracted. Since the control region does not contain gray matter, its responses are thought to reflect noise. For each run, the timecourses of the components extracted from the control region were regressed out from the timecourses of every voxel in gray matter. For both experiments, the global signal and six motion regressors generated by SPM during motion correction were also included as regressors of no interest. For the FFA seed, data were analyzed both with and without these additional regressors, and results are reported for both analyses.
For experiment 1, we defined a seed region of interest in the right pSTS using the independent functional localizer. Data were modeled with a standard GLM using SPM12, and the seed ROI was defined in each individual participant as a 6mm radius sphere centered in the pSTS peak for the faces vs houses contrast (mean MNI coordinates: 54,-54,13).
For experiment 2, we defined a seed region of interest in the right FFA using the independent functional localizer. Data were modeled with a standard GLM using SPM12, and the seed ROI was defined in each individual participant as a 6mm radius sphere centered in the FFA peak for the faces vs houses contrast (mean MNI coordinates: 40,-48,-20).
We defined a gray matter mask by smoothing (with a 6mm FWHM gaussian kernel) and averaging the gray matter probabilistic maps obtained during segmentation. The average maps were then thresholded obtaining approximately 130000 gray matter voxels (127821). For each voxel in the gray matter mask, we defined a 6mm radius sphere centered in that voxel, and calculated the statistical dependence between the responses in the seed region and the responses in the sphere. Spheres contained 123 voxels. Spheres at the edge of the brain were restricted to the voxels within the gray matter mask.
Functional connectivity was calculated low-pass filtering at 0.1Hz the mean response in the seed region and the mean response in the searchlight spheres, and calculating Pearson’s correlation between the low-pass filtered responses in the seed and each sphere, thus obtaining a whole-brain functional connectivity map. Statistical significance across participants was assessed with statistical nonparametric mapping [27] using the SnPM extension for SPM (http://warwick.ac.uk/snpm).
Let us consider the multivariate timecourses in the seed region: Y1, …, Ym and in a sphere: X1, …, Xm, for experimental runs from 1 to m. Each multivariate timecourse Yi is a matrix of size Ti × ny, where ny is the number of voxels in the seed region and Ti is the number of timepoints in run i. Analogously, each multivariate timecourse Xi is a matrix of size Ti × nx, where nx is the number of voxels in the sphere. Data analysis followed a leave-one-run-out procedure: for each choice of an experimental run i, data in the remaining runs were concatenated, obtaining
Y t r a i n = ( Y 1 , … , Y i - 1 , Y i + 1 , … , Y m ) ; X t r a i n = ( X 1 , … , X i - 1 , X i + 1 , … , X m ) .
Principal component analysis (PCA) was applied to Ytrain, and Xtrain:
Y t r a i n = U Y S Y V Y T X t r a i n = U X S X V X T
Dimensionality reduction was implemented projecting Ytrain and Xtrain on lower dimensional subspaces spanned by the first kY and kX principal components respectively:
Y ˜ t r a i n = Y t r a i n V Y [ 1 , … , k Y ]
X ˜ t r a i n = X t r a i n V X [ 1 , … , k X ]
where V T [ 1 , … , k T ] is the matrix formed by the first kY columns of VY and V X [ 1 , … , k X ] is the matrix formed by the first kX columns of VX. In the first analysis, the number of components kY and kX was chosen for each sphere and iteration using the Bayesian Information Criterion (BIC). In the second analysis, the incremental contribution of each component was tested by comparing the results obtained choosing 1, 2 and 3 components. We can take a moment to reflect on the interpretation of the procedure we just completed. For each region, each dimension obtained with PCA is a linear combination of the voxels in the region, whose weights define a multivariate pattern of response over voxels. Considering as an example the seed region, the loadings of a dimension j are encoded in the j-th column of Y ˜ t r a i n, and represent the intensity with which the multivariate pattern corresponding to dimension j is activated over time.
The mapping f from the dimensionality-reduced timecourses in the sphere X ˜ t r a i n to the dimensionality-reduced timecourses in the seed Y ˜ t r a i n was modeled with multiple linear regression 1:
Y ˜ t r a i n = B t r a i n X ˜ t r a i n + E t r a i n (1)
the model parameters were estimated using ordinary least squares (OLS).
After having estimated parameters Btrain, predictions for the multivariate responses in the left out run i were generated by 1) projecting the sphere data in run i on the sphere dimensions estimated with the other runs, and 2) multiplying them by the parameters estimated using data from the other runs. More formally, for each run i, we generated dimensionality reduced responses in the sphere:
X ˜ t e s t = X t e s t V X [ 1 , … , k X ],
where VX was calculated using the training data. Then, we calculated the predicted responses in the seed region in run i:
Y ^ t e s t = B t r a i n X ˜ t e s t
using the parameters Btrain independently estimated with the training runs.
In keeping with the use of correlation in standard functional connectivity, we calculated the correlation between the predicted and observed timecourses in each dimension in the seed region. First, we projected the observed voxelwise timecourses in the seed region onto the lower dimensional subspace using V Y [ 1 , … , k Y ]:
Y ˜ t e s t = Y t e s t V Y [ 1 , … , k Y ], (2)
where VY was calculated using the training data. Then, we computed
r j = corr ( Y ^ t e s t j , Y ˜ t e s t j )
for each dimension j = 1, …, kY of the seed region’s subspace. In the end, we generated a single summary measure r ‾, computing the average of the values rj weighted by the proportion of variance explained by the corresponding dimensions j:
w j = S Y ( j , j ) ∑ l = 1 k Y S Y ( l , l ),
r ¯ i = ∑ j = 1 k Y w j r j
(see the relationship between the eigenvalues along the diagonal of S and variance explain in PCA). This procedure is motivated by the observation that if a dimension explains more overall variance in the total multivariate response, then explaining variability in that dimension should be weighted more. See Fig 1 for an outline of the method. The values r ‾ i obtained for the different runs i = 1, …, m were averaged yielding r ‾. This procedure was repeated for each searchlight sphere, obtaining a whole brain map of r ‾ values for each participant. The significance of r ‾ was tested across participants with statistical nonparametric mapping [27] using the SnPM extension for SPM (http://warwick.ac.uk/snpm).
The value r ‾ is a convenient measure of statistical dependence: it reflects how well the prediction generated by MVPD correlates with the observed data. However, in this measure, the target of the prediction is the multivariate timecourse Y ˜ t e s t. Instead, ‘standard’ univariate connectivity based on the mean timecourse aims to predict a different target: mean(Ytest, 2). This is important because the proportion of variance explained (cross-validated R-squared) is given by the amount of variance explained divided by the total variance of the target of the prediction. Univariate connectivity and MVPD could explain the same amount of absolute variance, but still have different proportions of variance explained, because the total variances of the targets of the prediction differ. One way to think about this is that mean-based univariate connectivity ‘gives up’ on predicting variability orthogonal to the mean: if the mean response is predicted perfectly, then the proportion of variance explained will be 100%. In contrast, if MVPD tries to predict the mean as well as other dimensions, it could predict the mean perfectly like univariate connectivity, and still its proportion of variance explained could be less than 100%, because of residuals in the other dimensions. To compare the cross-validated R-squared of univariate connectivity and of MVPD, therefore, we need a measure of their ability to predict a common target. For this reason, for each searchlight sphere we calculated the cross-validated R-squared of mean-based univariate connectivity and of MVPD in the timecourses of individual voxels in the seed region. Predicting the timecourses of all voxels in the seed region is a common target for both univariate connectivity and MVPD, and therefore it makes the cross-validated R-squared of the two methods comparable. To calculate the cross-validated R-squared for both methods, we needed to use a variant of functional connectivity that can perform leave-one-out predictions. The variance explained in functional connectivity is r2, and it is equal to the variance explained by a linear regression estimated and tested in the same data. We used linear regression estimated in all runs minus one, and tested the variance explained in the left-out run, thus obtaining a leave-one-out variant of mean-based univariate functional connectivity (that uses the same data-split used in MVPD). The linear regression yielded a prediction of the mean response in the seed region. Each voxel’s response was then predicted with the predicted mean response in the seed region. For MVPD, we predicted each voxel’s response projecting the multivariate prediction Y ^ t e s t from its low-dimensional subspace of principal components to voxel space, using the matrix V Y [ 1 , … , k Y ]. Each voxel’s response was reconstructed as the sum of the dimensions’ loadings on the voxels weigthed by the dimensions’ loadings at each timepoint. It can be helpful here to note that this is equivalent to the product
Y ˘ t e s t = Y ^ t e s t V Y [ 1 , … , k Y ] T ,
where Y ˘ t e s t is the voxel-wise prediction (see 2 and consider that ( V Y [ 1 , … , k Y ] ) T = ( V Y [ 1 , … , k Y ] ) − 1). In the case of the mean-based univariate functional connectivity, the voxelwise prediction can be written as
Y ˘ t e s t = Y ^ t e s t 1 T ,
where Y ^ t e s t is the predicted mean response in the seed region and 1 is a nY × Ti vector of ones, making explicit the common form of the prediction for MVPD and for mean-based univariate connectivity: in the latter the mean is treated as a single dimension with equal loadings for each voxel.
For each voxel j in the seed region, variance explained was calculated as
v ( j ) = 1 - S S ( Y t e s t ( : , j ) - Y ˘ t e s t ( : , j ) ) S S ( Y t e s t ( : , j ) )
where Y ˘ are the predicted voxelwise timecourses, and the values v(j) were averaged to obtain a single measure
v ¯ = ∑ j = 1 n Y v ( j ) n Y
for each searchlight sphere.
In Experiment 1, standard functional connectivity identified statistical dependence between the right pSTS and more anterior regions of right STS (peak MNI: 54 -9 -15) and with the left STS (peak MNI: -52 -27 -6) (Fig 2, S1 Table). MVPD, but not standard functional connectivity, identified statistical dependence with the posterior cingulate (peak MNI: 0 -71 34), and with larger portions of posterior STS bilaterally (Fig 2, S2 Table).
To evaluate the separate effects of predicting independent data with a leave-one-out approach and of transitioning from univariate to multivariate statistical dependence, we additionally measured univariate statistical dependence with a leave-one-out procedure. As anticipated, predicting independent data reduced the number of significant voxels for univariate dependence (or ‘connectivity’) as compared to standard functional connectivity (Fig 3A), in line with the expectation that predicting independent data is a more stringent test. MVPD, despite predicting independent data, outperformed both variants of univariate dependence (Fig 3A). As a further comparison between univariate dependence and MVPD, we calculated the proportion of variance explained by each model in independent data. Univariate dependence did not explain more than 5% of the variance in any brain region, while MVPD explained more than 20% of the variance in several regions, including the STS bilaterally and posterior cingulate (Fig 3B).
As an additional test of the potential of MVPD, we analyzed multivariate dependence between the pSTS seed and the rest of the brain after subtracting the univariate signal (Fig 4). By doing so, we obtained an analysis procedure which is entirely complementary to standard functional connectivity, which relies entirely on the univariate signal. Even after removing the univariate signal, MVPD detected significant statistical dependence between the right pSTS and posterior cingulate (peak MNI: 0 -63 28) as well as the left STS (peak MNI: -58 -10 -13).
In Experiment 2, standard functional connectivity identified statistical dependence between the FFA seed and other regions of ventral temporal cortex, as well as with early visual cortex (peak MNI coordinates: 12,-90,-6), the right insula (peak MNI: 34,26,1), the thalamus (peak MNI: -9,-23,11), dorsal visual stream area V7 (14, -70, 43) and intraparietal sulcus (IPS, peak MNI: 30,-66,32; Fig 5, in blue, FWE-corrected p < 0.05, S3 Table). MVPD additionally identified statistical dependence between the FFA and posterior cingulate (pCing, peak MNI: 8,-46,38), the right superior temporal sulcus (rSTS, peak MNI: 51,-25,-4), the right anterior temporal lobe (rATL, peak MNI: 26 6 -33), right dorsomedial prefrontal cortex (rDMPFC, peak MNI: 8 57 30), and the dorsal visual stream area V3A (peak MNI 15,-88,31; Fig 5, in yellow, FWE-corrected p < 0.05, S4 Table). MVPD, unlike standard functional connectivity, did not detect significant statistical dependence between FFA and the amygdala (peak MNI for standard functional connectivity: 22,0,-20). Even after regressing out the global signal and six motion regressors generated by SPM during motion correction (S2 Fig), MVPD detected significant dependence in the posterior cingulate (peak MNI: -2 -39 40), the dorsal visual stream (peak MNI: -29 -76 29; 30 -75 32), occipital cortex (peak MNI: 18 -87 -10; -43 -79 -9).
Analysis of voxelwise cross-validated R-squared was performed for mean-based univariate connectivity, and for MVPD with 1, 2, and 3 principal components. Increasing the number of principal components led to a corresponding increase in the voxelwise cross-validated R-squared (Fig 6A for voxels explaining more than 5% of voxelwise variance, (Fig 6B for voxels explaining more than 10% of voxelwise variance). Cross-validated R-squared was also computed after regressing out six motion parameters and the global signal as additional nuisance regressors (S3 Fig). As expected, the greatest voxelwise cross-validated R-squared was observed in the right fusiform gyrus, in the proximity of the seed region’s location. Thanks to the additional contribution of the second and third principal components, variance explained above the 5% threshold was also observed more posteriorly extending towards the occipital face area (OFA), in the left fusiform, and anteriorly extending towards the medial portions of the anterior temporal lobes (ATL). These portions of cortex have been implicated together with FFA in the recognition of faces. [1, 14, 15]. The inclusion of dimensions beyond the first PC improved the modeling of statistical dependence between FFA and other regions implicated in face recognition. The voxelwise cross-validated R-squared with univariate dependence remained below 5% in the whole brain.
Including additional dimensions beyond the first improved our ability to characterize the statistical dependence between responses in the FFA seed and responses in other brain regions that have been implicated in face processing.
As in the case of Experiment 1, we performed an additional analysis removing the univariate signal, obtaining a fully complementary analysis to standard functional connectivity. This analysis revealed multivariate dependence between the FFA and ventral occipital regions despite the univariate signal was removed (S4 Fig).
We then averaged the MVPD-searchlight maps for the first PC and for the second and third PCs, and we studied the spatial distribution of the top 5000 voxels in the brain showing greatest statistical dependence with the first PC (Fig 7B in yellow) and the top 5000 voxels in the brain showing greatest statistical dependence with the second and third PCs (Fig 7B in blue). The first PC showed greatest statistical dependence with voxels extending posteriorly towards early stages in the visual processing hierarchy, and dorsally towards regions in the dorsal visual stream. By contrast, the second and third PCs showed a different profile: strongest statistical dependence was found with regions extending anteriorly, towards the medial ATL. MVPD revealed different connectivity profiles for different dimensions of FFA’s representational space, individuating two subspaces showing disproportionate statistical dependence with regions involved in early and late visual processing respectively.
This article introduces multivariate pattern connectivity (MVPD), a new method to investigate the multivariate statistical dependence between brain regions. MVPD is characterized by several key properties. First, the BOLD signal in each brain region is modeled as a set of responses along multiple dimensions, with each dimension corresponding to a function of the voxels in that region. Second, MVPD investigates the statistical dependence between two regions by computing the extent to which the responses in the multiple dimensions characterizing one region can predict the responses in the multiple dimensions characterizing the other region over time. Third, with an analogy to MVPA methods, MVPD uses a cross-validation procedure in which independent data are used for training and testing of the models. A subset of the runs are used as a training set to generate parameters which are then tested assessing their ability to predict responses in a left-out independent run. This leave-one-out approach mitigates the impact of noise, improving on most current methods that do not test the extent to which relationships between regions are sufficiently stable to generalize to independent data.
There are two senses in which MVPD is multivariate. First, PCA identifies weighted combinations of multiple voxels that covary over time explaining most of each region’s variance. Therefore, the dimensions that describe the representational space in each region are a combination of multiple dependent variables. Second, in standard functional connectivity, statistical dependence between two regions is measured by correlating two one-dimensional timecourses (the average responses in each of two regions). Instead, in MVPD, statistical dependence is measured by modeling a multiple linear regression that predicts a multi-dimensional timecourse (the responses along the multiple dimensions in one region) as a function of another multi-dimensional timecourse (the responses along the multiple dimensions in the other region).
In the examples described in the present article, dimensions are obtained with PCA as linear combinations of the voxels that tend to be jointly activated or deactivated over time. From a neuroscientific perspective, we can think of each region as consisting of multiple neural populations with selectivities for different properties of the stimuli that have different distributions over the course of the experiment. Each population has different spatial distributions over voxels. This leads different weighted combinations of voxels to having different timecourses of responses, whose dynamics can provide deeper insights into the interactions between regions than the investigation of average responses. Of course, while different populations with different selectivities and different spatial distributions can lead to dimensions with different time courses, it is unlikely that individual dimensions obtained with PCA correspond in a one-to-one relationship to neural populations with a specific selectivity profile. For example, more than one neural population might be collapsed in a single principal component, or populations might not be assigned to dimensions in a one-to-one mapping because of the orthogonality constraints imposed by PCA.
Like standard functional connectivity, MVPD revealed statistical dependence between the FFA and more posterior regions of ventral temporal and occipital cortex, and with regions in early visual cortex. However, MVPD additionally revealed statistical dependence between the FFA and the right ATL, previously implicated in the recognition of face identity [1, 13–15]. Furthermore, MVPD (but not standard functional connectivity) identified statistical dependence between the FFA and the right STS, implicated in the recognition of person identity [21, 28–30] and of facial expressions [31–33]. Standard functional connectivity, but not MVPD, identified statistical dependence between FFA and the amygdala. This can be due to less stable predictive relationships between responses in the amygdala and FFA dimensions beyond the first PC.
Previous studies investigating the functional connectivity of the FFA reported connectivity with the STS in resting state data specifically when the responses in regions selective for other categories were regressed out [34]. MVPD can help to disentangle different kinds of information in the study of statistical dependence: face-specific information might load differentially on different principal components, and the mapping between region can learn to rely specifically on the relevant information. Significant MVPD between FFA and STS might be observed thanks to the potential of the method to rely selectively on a relevant subset of the information encoded in FFA responses.
A recent study investigated effective connectivity between FFA and early visual cortex and STS, including participants with developmental prosopagnosia as well as neurotypical controls [35]. Feedforward connections from EVC to FFA and EVC to pSTS showed reduced strength in DP participants. A promising direction for future research consists in applying MVPD to study differences between patient populations and neurotypical controls, to investigate more closely whether neural differences affect specific subsets of the information encoded within a brain region. In the case of developmental prosopagnosia, MVPD could be used to test whether the reduced connectivity from EVC to FFA and pSTS is specific to particular response dimensions within EVC and FFA.
MVPD led to important improvements in cross-validated R-squared at the voxel level over mean-based univariate connectivity (Fig 6). MVPD using a single principal component already improved variance explained over a mean-based univariate approach. Adding a second and a third PC further improved variance explained in ventral temporal cortex as well as the anterior temporal lobes. In the end, MVPD allowed us to separately investigate the connectivity profiles of different dimensions of FFA’s representational space. In particular, different dimensions showed stronger dependence with posterior and anterior regions respectively. Previous connectivity studies found support for the view that posterior ventral stream regions are an entry node in the face recognition network [36], and previous MVPA studies found evidence of invariant representations of faces in anterior regions [1, 15]. In this context, the present evidence suggests that different FFA dimensions encode information related to FFA’s inputs and outputs respectively.
Future work can investigate the differences in MVPD between different tasks. Whether or not MVPD is sensitive to task differences remains an open question. We consider this a key research direction, in which the greater sensitivity of MVPD can reveal task-dependent changes in the interactions between regions that cannot be detected by standard functional connectivity.
In this study, we showed that MVPD can be sensitive to statistical dependence between regions that is not detected by standard functional connectivity. MVPD has the potential to study in even greater detail how statistical dependence is affected by different tasks. For example, in different tasks, the dependence between two regions could remain similar in overall magnitude, but shift from relying on a particular subset of dimensions to a different subset. MVPD could be used to detect this type of task-dependent changes by analyzing not only the overall amount of variance explained, but the matrix of parameters Btrain obtained in different tasks. If some dimensions in one region have a greater influence on responses in another region in one particular task, the parameters in Btrain corresponding to those dimensions will increase in that task.
MVPD differs in important respects from previous techniques aimed at studying the dynamic interactions between brain regions in terms of the information they encode. Unlike previous techniques [18, 37], MVPD does not rely on discrimination between categories determined by the experimenter, but on dimensions derived in a data-driven fashion. The data-driven dimensions can be related to properties of the stimuli or the task with a subsequent model (for instance regressing dimensions on conditions, or on stimulus properties using a forward model). Another difference between MVPD and the methods introduced by Coutanche and Thompson-Schill [18, 19] is that the latter characterize each region with a single measure (how well the pattern in a given timepoint can be assigned to one condition or another), while MVPD adopts multiple measures (the values along the multiple dimensions), which can provide a richer characterization of a region’s representation at any given time.
An innovative study [37] investigated the relations between brain regions measuring the correlation between representational dissimilarity matrices in different regions. This approach provides a richer characterization of each region’s representational structure by comparing similarity matrices instead of classification accuracies, but it discards trial-by-trial variability. Furthermore, correlations between dissimilarity matrices can only be computed if the same set of conditions are used to generate the dissimilarity matrices in each region. When the conditions correspond to individual stimuli as in [37] this is not problematic, but if stimulus categories are used it raises the question of whether it is appropriate to characterize the representational spaces of very different brain regions in terms of the dissimilarities between the same set of categories. Taking images of objects as an example of stimuli, categorization based on animacy could be most appropriate for some brain regions, while categorization based on color could be more appropriate for other regions.
An additional approach has used distance correlation to capture multivariate dependences between regions [38], finding more robust results than traditional correlations for inhomogeneous regions. MVPD offers as advantages over this approach the ability to test stability of the dependence between two regions in independent data, and to analyze dependence for different representational subspaces (e.g. Fig 7). This feature of MVPD makes it possible to relate the dimensions characterizing a region’s responses to stimulus properties using forward models, to then investigate what representational content drives statistical dependence between two regions.
More generally, methods to model multivariate statistical dependence can be described by the way in which they model the responses of individual regions, and by the way in which they model the dependence between the regions. Some methods (e.g. [18, 19]) use multivariate methods to generate a unidimensional quantity (e.g. classification accuracy), and measure statistical dependence relating these unidimensional quantities between regions (e.g. with correlation). Other methods (e.g. [37, 39]) map directly the responses along multiple voxels in one region onto responses along multiple voxels in another. MVPD combines the two strategy by initially mapping the multi-voxel responses in each region onto a small set of dimensions (thus reducing the number of parameters that need to be estimated), and then modeling the multivariate relationship between these dimensionality-reduced patterns (e.g. with multiple regression).
By virtue of modeling the statistical dependence between patterns of responses in different regions, which likely correspond to different processing stages, multivariate measures of dependence are related to some extent to the approach of developing computational models of information processing and using them to predict neural responses [40, 41]. Two key differences between these approaches are that at present, computational models of information processing have more sophisticated tools to relate neural responses to stimulus properties, but the model parameters are trained independently of neural responses. By contrast, while multivariate dependence does not yet have the same level of sophistication in linking neural responses to stimulus properties, it gives the neural data a more predominant role in shaping the resulting models, by estimating parameters directly using the fMRI measurements. A recent article [42] built a model of visual cortex more closely inspired to the architecture of the brain, making a step in the direction of combining these two strengths. Future work will be necessary to constrain computational models taking full advantage of the wealth of information available in neural measurements while also tying the neural responses to the stimulus content they represent.
The most important asset of MVPD is probably its flexibility. The framework of 1) modelling representational spaces in individual regions, 2) considering multivariate timecourses as trajectories in these representational spaces, and 3) fitting models predicting the trajectory in the representational space of one region as a function of the trajectory in the representational space in another offers a wealth of possibilities to build increasingly refined models, both in terms of the characterization of representational spaces and in terms of the models of their interactions. For the characterization of representational spaces, in this article we adopted PCA as a simple example, but other methods such as independent component analysis (ICA) and nonlinear dimensionality reduction techniques can also be used. For modelling interactions between regions, we limited the current application to simultaneous, non-directed interactions following an approach similar to functional connectivity, but MVPD makes it possible to model nonlinear maps between representational spaces [43], and to use models that investigate the directionality of interactions using temporal precedence, along the lines of Granger Causality [8], Dynamic Causal Modelling [7], and Dynamic Network Modelling [11].
|
10.1371/journal.pgen.1006933 | Systematic tissue-specific functional annotation of the human genome highlights immune-related DNA elements for late-onset Alzheimer’s disease | Continuing efforts from large international consortia have made genome-wide epigenomic and transcriptomic annotation data publicly available for a variety of cell and tissue types. However, synthesis of these datasets into effective summary metrics to characterize the functional non-coding genome remains a challenge. Here, we present GenoSkyline-Plus, an extension of our previous work through integration of an expanded set of epigenomic and transcriptomic annotations to produce high-resolution, single tissue annotations. After validating our annotations with a catalog of tissue-specific non-coding elements previously identified in the literature, we apply our method using data from 127 different cell and tissue types to present an atlas of heritability enrichment across 45 different GWAS traits. We show that broader organ system categories (e.g. immune system) increase statistical power in identifying biologically relevant tissue types for complex diseases while annotations of individual cell types (e.g. monocytes or B-cells) provide deeper insights into disease etiology. Additionally, we use our GenoSkyline-Plus annotations in an in-depth case study of late-onset Alzheimer’s disease (LOAD). Our analyses suggest a strong connection between LOAD heritability and genetic variants contained in regions of the genome functional in monocytes. Furthermore, we show that LOAD shares a similar localization of SNPs to monocyte-functional regions with Parkinson’s disease. Overall, we demonstrate that integrated genome annotations at the single tissue level provide a valuable tool for understanding the etiology of complex human diseases. Our GenoSkyline-Plus annotations are freely available at http://genocanyon.med.yale.edu/GenoSkyline.
| After years of community efforts, many experimental and computational approaches have been developed and applied for functional annotation of the human genome, yet proper annotation still remains challenging, especially in non-coding regions. As complex disease research rapidly advances, increasing evidence suggests that non-coding regulatory DNA elements may be the primary regions harboring risk variants in human complex diseases. In this paper, we introduce GenoSkyline-Plus, a principled annotation framework to identify tissue and cell type-specific functional regions in the human genome through integration of diverse high-throughput epigenomic and transcriptomic data. Through validation of known non-coding tissue-specific regulatory regions, enrichment analyses on 45 complex traits, and an in-depth case study of neurodegenerative diseases, we demonstrate the ability of GenoSkyline-Plus to accurately identify tissue-specific functionality in the human genome and provide unbiased, genome-wide insights into the genetic basis of human complex diseases.
| Large consortia such as ENCODE [1] and Epigenomics Roadmap Project [2] have generated a rich collection of high-throughput genomic and epigenomic data, providing unprecedented opportunities to delineate functional structures in the human genome. As complex disease research rapidly advances, evidence has emerged that disease-associated variants are enriched in regulatory DNA elements [3, 4]. Therefore, functional annotation of the non-coding genome is critical for understanding the genetic basis of human complex diseases. Unfortunately, categorizing the complex regulatory machinery of the genome requires integration of diverse types of annotation data as no single annotation captures all types of functional elements [5]. Recently, we have developed GenoSkyline [6], a principled framework to identify tissue-specific functional regions in the human genome through integrative analysis of various chromatin modifications. In this work, we introduce GenoSkyline-Plus, a comprehensive update of GenoSkyline that incorporates RNA sequencing and DNA methylation data into the framework and extends to 127 integrated annotation tracks covering a spectrum of human tissue and cell types.
To demonstrate the ability of GenoSkyline-Plus to systematically provide novel insights into complex disease etiology, we jointly analyzed summary statistics from 45 genome-wide association studies (GWAS; Ntotal≈3.8M) and identified biologically relevant tissues for a broad spectrum of complex traits. We next performed an in-depth, annotation-driven investigation of Alzheimer’s disease (AD), a neurodegenerative disease characterized by deposition of amyloid-β (Aβ) plaques and neurofibrillary tangles in the brain. Late-onset AD (LOAD) includes patients with onset after 65 years of age and has a complex mode of inheritance [7]. Around 20 risk-associated genetic loci have been identified in LOAD GWAS [8]. However, our understanding of LOAD’s genetic architecture and disease etiology is still far from complete. Through integrative analysis of GWAS summary data and GenoSkyline-Plus annotations, we identified strong enrichment for LOAD associations in immune cell-related DNA elements, consistent with other data suggesting a crucial role for the immune system in AD etiology [9–11]. Jointly analyzing GWAS summary data for LOAD and Parkinson’s disease (PD), we identified substantial enrichment for pleiotropic associations in the monocyte functional genome. Our findings provide support for the critical involvement of the immune system in the etiology of neurodegenerative diseases, and suggest a previously unsuspected role for an immune-mediated pleiotropic effect between LOAD and PD.
We use our previously established statistical framework to calculate the posterior probability of functionality for each nucleotide in the human genome [12]. Integrating tissue and cell-specific genomic functional data available through Epigenomics Roadmap Project [2], we make available GenoSkyline-Plus scores for 127 individual tissue annotation tracks (Methods; S1 Table). H3K4me3 and H3K9ac, known markers of open chromatin and active transcription [13], are shown to have the largest odds ratios of predicting functionality across the genome (Fig 1A). Identifying H3K4me3 and H3K9ac as strong indicators of genomic functionality is a finding consistent with previous studies of gene regulation through chromatin marks [14]. In contrast, H3K9me3, a well established repressive mark [13], has a reversed effect on genome functionality. The bimodal pattern of GenoSkyline scores [6] allows us to impose a score cutoff to robustly define the functional genome. Using a cutoff of 0.5, 3% of the genome is considered functional on average across all annotation tracks (Fig 1B). This functionality percentage varies from 1% in pancreatic islet cells to 8% in PMA-I stimulated T-helper cells. Our findings on functionality across all tracks are consistent with previous findings [12]; 34% of the intergenic human genome is predicted to be functional in at least one annotation track (Fig 1C). Additionally, coding regions of the genome are predicted to have much greater proportions of functionality in multiple tissues than intronic and intergenic regions.
To assess the ability of GenoSkyline-Plus to capture tissue and cell-specific, non-coding functionality in the human genome, we consider a diverse set of known non-coding regulatory elements studied across the genome. To start, we examined microRNAs (miRNA), which are known to regulate a variety of cellular processes through the translational repression and degradation signaling of transcripts [15]. Recent work by Ludwig et al. profiled miRNA expression in 61 different human tissues and identified miRNAs with functionality unique to single tissues through a tissue specific index [16, 17] (TSI; Methods). We applied GenoSkyline-Plus scores to miRNA with tissue-specific functionality by calculating the total proportion of nucleotides predicted to be functional in each tissue. We next looked for which annotation tracks are able to predict the highest proportion of functionality for these known functional regions. The best predictors of high functionality for the three tissues with the largest sample sizes (i.e. brain, liver, and muscle) are tracks for brain structures, the liver track, and the muscle track, respectively (Fig 2A).
We next examined long non-coding RNAs (lncRNA), another non-coding element known for its tissue-specific regulatory action [18]. Using a custom-designed microarray targeting GENCODE lncRNA, Derrien et al. profiled the activity of 9,747 lncRNA transcripts [19]. In order to reidentify and validate the set of lncRNA transcripts that are specific to their respective tissues, we calculated the previously described TSI and selected lncRNAs with expression specific to only a few cell types. Physiologically matching tracks show a higher proportion of predicted functionality than unmatched tracks in complex, heterogeneous tissue structures like the midbrain. More functionally uniform tissues, such as the thymus or placenta, show the highest functional proportion in matching annotation tracks (Fig 2B).
We also assessed enhancers, non-coding elements that can remotely regulate transcription of an associated promoter elsewhere on the genome with important roles in cell-type specificity [20]. We extracted tissue and cell type-specific enhancer facets identified through the FANTOM5 cap analysis of gene expression (CAGE) atlas and positive differential expression when compared against other defined facets [21]. To determine the utility of the large library of immune cells available in the Epigenomics Roadmap Project for which we developed annotation tracks, we focused on enhancer facets with differential CAGE expression in immune cells. While the method by which enhancers are defined to be differential in a facet is liberal (Methods) and does not imply facet-specific expression, GenoSkyline-Plus still showed outstanding ability to identify matching cell types. Indeed, matched annotation tracks for T-cells, natural killer cells, and monocytes show consistently higher functional proportions than other, non-matched immune cell annotation tracks (Fig 2C).
Finally, we present a case study of the IL17A-IL17F locus control region (LCR) in humans, a ~200kb regulatory region surrounding the IL17A gene locus. IL17A encodes the primary secreted cytokine effector molecule IL-17 of T helper 17 (Th17) cells [22]. The LCR has been studied in mouse models and is found to contain many potential human-conserved intergenic regulatory elements that bind transcription factors that are essential for Th17 cell differentiation and effector function [23, 24]. Experimentally, these conserved noncoding sequences (CNS) acquire functionally permissive H3 acetylation marks at much greater magnitudes under Th17-inducing conditions than naïve or combined Th1 and Th2 populations [25]. Comparing annotation tracks for naïve CD4+ T-cells, differentiated Th17 cells, and differentiated Th1/Th2 cell populations, we identified highly Th17-specific functionality in the conserved regions of the human genome corresponding to known murine CNS regions (Fig 2D and 2E). CNS sites and their flanking regions showed substantially higher functional proportion in Th17 cells than in naïve CD4+ T-cells or Th1/Th2 cell subsets.
We jointly analyzed three tiers of annotation tracks that respectively represent the overall functional genome, 7 broad tissue clusters, and 66 tissue and cell types (Methods; S2 Table), with summary statistics from 45 GWAS covering a variety of human complex traits (S3 Table). We applied LD score regression [26] to stratify trait heritability by tissue and cell type, and identified a total of 226 significantly enriched annotation tracks for 34 traits after correcting for multiple testing (S4–S7 Tables). In general, GWAS with a large number of significant SNP-level associations showed stronger heritability enrichment in the predicted functional genome (Fig 3A and 3B). Tissue and cell tracks refined the resolution of heritability stratification and provided additional insights into the genetic basis of complex traits (Fig 3C and 3D).
The immune annotation track was significantly enriched for 7 immune diseases, namely celiac disease (CEL), Crohn’s disease (CD), ulcerative colitis (UC), primary biliary cirrhosis (PBC), rheumatoid arthritis (RA), systemic lupus erythematosus (SLE), and multiple sclerosis (MS). Using tracks for cell types, we identified several significant enrichments, including monocytes for CD (p = 2.9e-11) and B cells for PBC (p = 2.3e-6), RA (p = 1.2e-5), and MS (p = 2.2e-6). Inflammatory bowel diseases showed significant enrichment in the gastrointestinal (GI) annotation track (CD: p = 1.4e-4; UC: p = 5.6e-5). Another autoimmune disease with a well-established GI component, CEL, also showed nominal enrichment in the GI annotation track (p = 3.7e-4).
Several brain annotation tracks were significantly enriched for associations of schizophrenia (SCZ), education years (EDU), and cognitive performance (IQ). Bipolar disorder (BIP), neuroticism (NEU), and chronotype (CHT) all showed nominally significant enrichment in the anterior caudate annotation track. Body mass index (BMI) and age at menarche (AAM) were significantly enriched in multiple brain annotation tracks. Compared to other brain regions, the substantia nigra annotation track showed weaker enrichment for these brain-based traits, which is consistent with its primary function of controlling movement.
Hundreds of height-associated loci have been identified in GWAS [27]. Such a highly polygenic genetic architecture is also reflected in our analysis. 59 of 66 tier-3 tissue and cell annotation tracks were significantly enriched for height associations, with breast myoepithelial cell (p = 6.2e-14) and osteoblast (p = 8.5e-14) being the most significant. Waist-hip ratio (WHR), birth weight (BW), and three blood pressure traits showed significant enrichment in the adipose annotation track. Overall, cardiovascular (CV) annotation tracks showed strong enrichment for blood pressure and coronary artery disease (CAD). Interestingly, the aorta annotation track is significantly enriched for pulse pressure (PP) but not systolic or diastolic blood pressure (SBP and DBP). CAD and 4 lipid traits, i.e. high and low density lipoprotein (HDL and LDL), total cholesterol (TC), and triglycerides (TG), shared a similar enrichment pattern in liver, adipose, and monocyte annotation tracks, which is consistent with the causal relationship among these traits [28].
Our results demonstrated that annotations with refined specificity could provide insights into disease etiology while broader annotations have greater statistical power. Age-related macular degeneration (AMD) was significantly enriched in broadly defined annotation tracks including immune, brain, CV, and GI, despite the non-significant enrichment results using tier-3 annotation tracks. Analyses based on all three tiers of annotations could systematically provide the most interpretable results for most traits. Importantly, we note that greater GWAS sample sizes will effectively increase statistical power in the enrichment analysis while leaving the overall enrichment pattern stable (S1 Fig). Therefore, many more suggestive enrichment results are likely to become significant as GWAS sample sizes grow. Finally, some traits, e.g. type-II diabetes (T2D) and age at natural menopause (AANM), showed strong enrichment in the general functional genome but not in specific tissues, suggesting that we may be able to gain a better understanding of these traits when annotation data for tissues or cell types more relevant to these traits are made available.
Next, we performed an integrative analysis of stage-I GWAS summary statistics from the International Genomics of Alzheimer’s Project [8] (IGAP; n = 54,162) with GenoSkyline-Plus annotations (Methods). SNPs located in the broadly defined immune annotation track, which account for 24.4% of the variants in the IGAP data, could explain 98.7% of the LOAD heritability estimated using LD score regression (enrichment = 4.0; p = 1.5e-4). Somewhat surprisingly, the signal enrichment in DNA elements functional in immune cells was substantially stronger than the enrichment in brain and other tissue types (Fig 4A). To investigate if immune-related DNA elements are also enriched for associations of other neurodegenerative diseases, we analyzed a publicly accessible GWAS summary dataset for PD [29] (n = 5,691; Methods). Again, the immune annotation track was the most significantly enriched annotation (enrichment = 6.3; p = 7.5e-6), followed by epithelium and CV (Fig 4A).
Analysis based on 66 tissue and cell tracks further refined the resolution of our enrichment study. Monocyte (enrichment = 10.9; p = 2.0e-5) and liver (enrichment = 16.6; p = 4.1e-4) annotation tracks were significantly enriched for LOAD associations (Fig 4B). In fact, the combined functional regions in monocyte and liver covered 8.8% of the SNPs in the IGAP data, but could account for 99.6% of the LOAD heritability currently captured in the IGAP stage-I GWAS (Fig 4C). In PD GWAS, signal enrichment in liver was absent, but monocyte-functional regions remained strongly enriched (enrichment = 16.3; p = 8.5e-7).
Our findings support the critical role of innate immunity in neurodegenerative diseases [10]. Significant enrichment for LOAD associations in liver-specific DNA elements also provides additional support for the possible involvement of cholesterol metabolism in LOAD etiology [30, 31]. LOAD signal enrichment in liver remained significant after removing the APOE region (chr19: 45,147,340–45,594,595; hg19) from the analysis (S2 Fig), suggesting a polygenic architecture in this pathway. Finally, some adaptive immune cells also showed enrichment for AD and PD associations. LOAD signal enrichment in the B cell annotation track was nominally significant, while multiple T cell annotation tracks were significantly enriched for PD associations. These results not only suggest the involvement of adaptive immunity in neurodegenerative diseases, but also hint at distinct mechanisms of such involvement between AD and PD. Finally, for comparison, we applied several other annotations including CADD [32], GWAVA [33], and EIGEN [34] to the LOAD GWAS data. GenoCanyon and GenoSkyline annotations for seven tissues were also included in the comparison. Our annotations outperformed these methods, showing stronger fold enrichment and more significant p-values (S8 Table).
Our results showed strong enrichment for both AD and PD in the monocyte functional genome. Next, we investigate if the enrichment for both diseases is through shared or distinct genetic components. Recent studies have failed to identify statistically significant genome-wide pleiotropic effects between AD and PD [35]. We instead hypothesize that the same set of immune-related genetic components are involved in both diseases. Therefore, we aim to identify enrichment for pleiotropic effects in the genome localized to regions of monocyte functionality.
We first partitioned AD and PD heritability by chromosome. Chromosome-wide heritability showed moderate correlation between the two diseases (correlation = 0.65; Fig 5A). When focusing on monocyte functional elements, chromosome-wide heritability showed high concordance between AD and PD (correlation = 0.96; Fig 5B). Interestingly, such high concordance cannot be fully explained by chromosome size. In fact, the correlation between chromosome size and per-chromosome heritability estimates is 0.56 for AD and 0.59 for PD, both lower than the correlation between AD and PD’s per-chromosome heritability estimates, especially in the monocyte functional genome. The percentage of explained LOAD heritability on chromosome 19 is lower than previous estimation [36] due to removal of SNPs with large effects in the APOE region (Methods). Next, to quantify the shared genetics between AD and PD, we identified significant enrichment for pleiotropic effects in monocyte functional regions (enrichment = 1.8; p = 9.4e-4) using a window-based approach (Methods). To account for potential bias due to the moderate sample overlap between the two GWAS as well as other confounding factors, we applied a permutation-based testing approach (Methods). Enrichment for pleiotropic effects in the monocyte functional genome remained significant (p = 4.6e-3). In addition, these results were robust with the choice of window size.
We identified 15 candidate loci for pleiotropic effects (Methods; S9 Table), among which signals at SLC9A9 and AIM1 are the clearest (Fig 5C and 5D). SLC9A9, whose encoded protein localizes to the late recycling endosomes and plays an important role in maintaining cation homeostasis (RefSeq, Mar 2012), is associated with multiple pharmacogenomic traits related to neurological diseases, including response to cholinesterase inhibitor in AD [37], response to interferon beta in MS [38], response to angiotensin II receptor blockade therapy [39], and multiple complex diseases including attention-deficit/hyperactivity disorder [40], autism [41], and non-alcoholic fatty liver [42]. Gene AIM1 is associated with stroke [43], human longevity [44], and immune diseases including RA [45] and SLE [46].
A few candidate loci pointed to clear gene candidates but showed unclear or distinct peaks of association (S3 Fig). These include an inflammatory bowel disease risk gene ANKRD33B [47]. PRUNE2 is a gene associated with response to amphetamine [48] and hippocampal atrophy which is a quantitative trait for AD [49]. HBEGF is associated with AD in APOE ε4- population [50] and involved in Aβ clearance [51]. PROK2 is a gene involved in Aβ-induced neurotoxicity [52]. Additionally, the protein product of AXIN1 negatively affects phosphorylation of tau protein [53]. Other gene candidates include CCDC158, PRSS16, and ZNF615, which are previously identified risk genes for PD, SCZ, and BIP, respectively [54–56]. Some other windows showed complex structures of linkage disequilibrium (LD) and contained large association peaks spanning a number of genes (S4 Fig), which include the region near PD risk gene PRSS8 [54] and the HLA region. Interestingly, we also identified the surrounding region of MAPT, a gene that encodes the tau protein which is a critical component of both AD and PD pathologies [50, 54, 57, 58].
Pathway enrichment analysis for genes in 15 pleiotropic candidate loci identified significant enrichment in immune-related pathways staphylococcus aureus infection (KEGG:05150; p = 1.9e-5) and systemic lupus erythematosus (KEGG:05322; p = 3.7e-04; Methods). Both pathways remained significant after removing two HLA loci from our analysis.
Finally, we reprioritize AD risk loci using monocyte and liver annotation tracks. We integrated IGAP stage-I summary statistics with GenoSkyline-Plus using genome-wide association prioritizer (GenoWAP [59]), and ranked all SNPs based on their GenoWAP posterior scores (Methods). Under a posterior cutoff of 0.95, we identified 8 loci that were not reported in the IGAP GWAS meta-analysis using monocyte annotation and 4 loci using the liver annotation track (S10 Table).
We then sought replication for SNPs with the highest posterior score at each of these loci using inferred IGAP stage-II z-scores (Methods). After removing shared SNPs between monocyte- and liver-based analyses, 10 SNPs remained in the analysis, 7 of which showed consistent effect directions between the discovery and the replication cohorts (Fig 6A). One SNP was successfully replicated in the inferred IGAP stage-II dataset, i.e. rs4456560 (p = 0.013). SNP rs4456560 is located in SCIMP (Fig 6B), a gene that encodes a lipid tetraspanin-associated transmembrane adaptor protein that is expressed in antigen-presenting cells and localized in the immunological synapse [60].
A moderate replication rate in the IGAP stage-II cohort was expected since we focused on loci that did not reach genome-wide significance in the IGAP meta-analysis and the IGAP stage-II cohort is relatively small (n = 19,884) compared to the data in the discovery stage. Furthermore, data from IGAP stage-II cohort are not publicly available and we were limited to the inverse inference approach shown here. It is possible additional loci will replicate when IGAP stage-II summary or individual-level data are made available. However, all identified loci have been linked to AD or relevant phenotypes in the literature. RPN1 was linked to AD through a network-based technique [61]. Association between ECHDC3 and AD risk was established through a joint analysis of AD and lipid traits [62]. Association between DLST and AD has also been previously reported [63]. BZRAP1 and MINK1 were shown to be associated with cognitive function and blood metabolites, respectively [64, 65]. A pleiotropic effect candidate gene HBEGF showed up again in the SNP reprioritization analysis. Multiple genes in the sorting nexin family have been found to participate in APP metabolism and Aβ generation [66]. Association between SNX1 and AD has also been previously identified using gene-based tests [67]. Finally, during the peer review process of this paper, three new genome-wide significant loci (i.e. PFDN1/HBEGF, USP6NL/ECHDC3, and BZRAP1-AS1) were reported in a trans-ethnic GWAS meta-analysis for AD [68], all of which were among our reprioritized list of risk loci. Further, the most significant SNPs at loci PFDN1/HBEGF (rs11168036, p = 7.1e-9) and BZRAP1-AS1 (rs2632516, p = 4.4e-8) matched with our top reprioritized SNPs (Fig 6A).
Increasing evidence suggests that non-coding regulatory DNA elements may be the primary regions harboring risk variants in human complex diseases. In this work, we have substantially expanded our previously established GenoSkyline annotation by incorporating RNA-seq and DNA methylation into its framework, imputing incomplete epigenomic and transcriptomic annotation tracks, and extending it to more than 100 human tissue and cell types. With the help of integrative functional annotations, we identified strong enrichment for LOAD heritability in functional DNA elements related to innate immunity and liver tissue using hypothesis-free tissue-specific enrichment analysis. This enrichment was also found in immune-related DNA elements using PD data. Our analysis also clearly indicated that monocyte functional elements in particular appear to be highly relevant in explaining AD and PD heritability. Of note, we analyzed 45 complex diseases and traits in addition to AD and PD. The substantial enrichment for multiple psychiatric and neurological traits in the brain functional genome shows that the lack of brain enrichment in neurodegeneration is not due to poor quality of brain annotations. Further, the monocytes annotation track was the most significantly enriched for Crohn’s disease among the 45 GWAS, and was not ubiquitously enriched for a large number of traits. Consistent and biologically interpretable enrichment results on a large collection of complex traits demonstrate the effectiveness of our approach and increase the validity of novel findings.
It is worth noting that multiple studies have highlighted the role of myeloid cells in the genetic susceptibility of neurodegenerative diseases [11]. Several genes expressed in myeloid cells (e.g. ABCA7, CD33, and TREM2) have been identified in GWAS and sequencing-based association studies for AD [8, 69, 70]. Further, AD risk alleles identified in GWASs have been shown to enrich for cis-eQTLs in monocytes [9]. In addition, two recent papers identified enrichment for AD heritability in active genome regions in myeloid cells [71, 72], which suggested a polygenic genetic architecture for immune-related DNA elements in AD etiology and hinted at a large number of unidentified, immune-related genes for AD. Compared to the aforementioned work, our study utilizes a better set of tissue-specific genome annotations and explicitly accounts for the similarity between different cell types through a multiple regression model. One major limitation in our analysis is lack of data for other potentially AD-relevant cell types such as microglia. Whether our findings correctly reflected the direct involvement of peripheral immune cells in neurodegenerative diseases rather than the detection of epigenomic similarities between monocytes and microglia remains to be carefully investigated in the future.
Furthermore, we successfully identified enrichment for shared genetic components between AD and PD in the monocyte functional genome, which hints at a shared neuroinflammation pathway between these two neurodegenerative diseases. We note that several candidate loci with potential pleiotropic effects showed fairly marginal associations with AD and PD, which explains why they have been missed in traditional SNP-based association analysis. Importantly, SNPs in immune-related DNA elements explain a large proportion of AD and PD heritability in total. These results suggest that weak but pervasive associations related with immunity still remain unidentified. Further evaluations of these relationships using GWAS with larger sample sizes may provide insights into the shared biology of these neurodegenerative conditions.
Through multi-tier enrichment analyses on 45 GWAS, an in-depth case study of neurodegenerative diseases, and validation of known non-coding tissue-specific regulatory machinery, we have demonstrated the ability of GenoSkyline-Plus to provide unbiased, genome-wide insights into the genetic basis of human complex diseases. The analyzed GWAS represent a variety of human complex diseases and traits, highlighting the effectiveness of our method in different contexts and genetic architecture. However, while our non-coding validation study demonstrated that GenoSkyline-Plus annotations indeed captured tissue-specific activity in a variety of intergenic machinery, there is a need to develop a more statistically robust framework to identify new non-coding elements rather than validate existing ones. Our approach of identifying the functionally active proportion of all elements in aggregate is only able to identify tissue specificity while considering large groups of highly specific non-coding elements. The availability of over 100 different annotation tracks introduces many multiple-testing issues that should be addressed in the case of a statistically sound analysis for tissue-specificity. We have also demonstrated how GenoSkyline-Plus and its explanatory power improve with the addition of more data. Currently, functionality in 28% of exonic regions still remains to be identified. As the quantity and quality of high-throughput epigenomic data continue to grow, GenoSkyline-Plus has the potential to further evolve and provide even more comprehensive annotations of tissue-specific functionality in the human genome. We will update our annotations when data for new tissue and cell types from the Roadmap consortium become available. Finally, several recent papers have introduced novel models to integrate functional annotations in tissue-specific enrichment analysis [73, 74]. Many models that do not explicitly incorporate functional annotation information have also emerged in transcriptome-wide association studies and other closely-related applications in human genetics research [75–78]. Our annotations, in conjunction with rapidly advancing statistical techniques and steadily increasing sample sizes in genetics studies, may potentially benefit a variety of human genetics applications and promise a bright future for complex disease genetics research.
Chromatin data were extracted from the Epigenomics Roadmap Project’s consolidated reference epigenomes database (http://egg2.wustl.edu/roadmap/). Specifically, ChIP-seq peak calls were collected for each epigenetic mark (H3k4me1, H3k4me3, H3k36me3, H3k27me3, H3k9me3, H3k27ac, H3k9ac, and DNase I Hypersensitivity) in each Roadmap consolidated epigenome where available. Peak calls imputed using ChromImpute [79] were used in place of missing data. Next, peak files were reduced to a per-nucleotide binary encoding of presence or absence of contiguous regions of strong ChIP-seq signal enrichment compared to input (Poisson p-value threshold of 0.01).
DNA methylation data were also collected from the Roadmap’s reference epigenomes database. CpG islands were identified in each sample using the CpG Islands Track of the UCSC Genome Browser (http://genome.ucsc.edu/), and unmethylated islands were those CpG islands with less than 0.5 fractionated methylation based on imputed methylation signal tracks in the Roadmap reference epigenomes database. Presence of an unmethylated CpG island was then encoded for each nucleotide as a binary variable. Finally, Roadmap’s RNA-seq data were dichotomized using an rpkm cutoff of 0.5 at 25-bp resolution and included in our annotations.
We adapt the existing framework established by Lu et al. to a broader set of genomic data [12]. Briefly, given a set of Annotations A and a binary indicator of genomic functionality Z, the joint distribution of A along the genome is assumed to be a mixture of annotations at functional nucleotides and non-functional nucleotides. Assuming that each of the annotations in A is conditionally independent given Z, we factorize the conditional joint density of A given Z as:
f(A|Z=c)=∏i=110fi(Ai|Z=c), c=0, 1
(1)
All annotations have been preprocessed into binary classifiers, and the marginal functional likelihood given each individual annotation can be modeled with a Bernoulli distribution
fi(Ai|Z=c)=picAi(1−pic)1−Ai, i=1,…,10; c=0, 1
(2)
With an assumed prior probability π of functionality, the parameter pic of each individual annotation can be estimated with the Expectation-Maximization (EM) algorithm. The posterior probability of functionality at a nucleotide, known as the GenoSkyline-Plus score, is then:
P(Z=1|A)=π∏i=110fi(Ai|Z=1)π∏i=110fi(Ai|Z=1)+(1−π)∏i=110fi(Ai|Z=0)
(3)
Giving us with 21 parameters for each annotation track:
Θ=(π,p1,0,p2,0, … ,p10,0,p1,1,p2,1, … ,p10,1)
(4)
These parameters were estimated using the GWAS Catalog, downloaded from the NHGRI website (http://www.genome.gov/gwastudies/). 13,070 unique SNPs found to be significant in at least one published GWAS were expanded into 1kb bp intervals and formed a sampling covering 12,801,840 bp of the genome. This sampling method has been shown to be a robust representation of functional and non-functional regions along the genome [6]. Notably, other models have been recently developed to predict functional non-coding SNPs [34].
Quantile-normalized expression values were downloaded for all mature miRNAs profiled in Ludwig et al [17]. Due to inconsistent levels of miRNA specificity in the two donors in this study and to avoid diluting miRNA specificity, we used miRNA data from only body 1, which had a higher fraction of tissue specific miRNAs. TSI values were calculated as described in the study:
TSIj= ∑i=1N(1−xj,i)N−1
(5)
Where N is the total number of tissues measured, xj,i is the expression intensity of tissue i divided by the maximum expression across all tissues for miRNA j. We extract any miRNAs with a TSI score greater than the median value of 0.75 to produce a sufficiently large collection of miRNAs with expression highly specific to only a few tissues that we can then attempt re-identify using GenoSkyline-Plus. We next download genomic positions and identify the highest expressed tissue for each TSI-filtered miRNA. miRNA coordinates were extracted from miRbase (http://mirbase.org/) and mapped to hg19 using the UCSC liftover tool (http://genome.ucsc.edu/). lncRNA data was prepared similarly to miRNA. Expression data of 9,747 lncRNA transcripts based on GENCODE v3c annotation across 31 human tissues [19] (GEO accession: GSE34894) was downloaded. As above, the TSI of each lncRNA transcript was calculated, and transcripts with a TSI greater than 0.75 were labeled for genomic position and maximally expressed tissue.
Pre-defined enhancer differentially expressed cell facets [21] were downloaded from PrESSto database (http://enhancer.binf.ku.dk/presets/). Andersson et al. define their enhancer sets via bi-directional CAGE expression collected by the FANTOM consortium [80]. Cell facets were manually constructed using hierarchical FANTOM5 cell ontology term mappings to create mutually exclusive and broadly covered histological and functional annotations. Enhancers were considered differentially expressed in a facet using Kruskal-Wallis rank sum test and subsequent pair-wise post-hoc tests to identify enhancers with significantly differential expression between pairs of facets. Based on this method, an enhancer is considered differentially expressed in a facet if it is significantly differentially expressed compared to any other facet and has overall positive standard linear statistics.
For each of the three data validation sets, functional specificity is assessed by calculating the per-nucleotide functional proportion of all non-coding elements across a tissue. Functionality is defined by a Genoskyline-Plus score greater than 0.5 at that nucleotide. For Roadmap samples with multiple donors (e.g. skeletal muscle and rectal mucosa) we took the average GenoSkyline-Plus score at each nucleotide across the samples. For each set of non-coding elements we selected the top three tissues with the largest sample size that had matching annotations in Genoskyline-Plus. For example, we did not calculate scores for enhancers with maximal expressions in human testis because there is no corresponding Roadmap sample in which we would detect tissue-specific functionality.
To examine cell-specific functionality of the IL17A LCR in T-cell subsets, we extracted GenoSkyline-Plus scores for each nucleotide along the ~200 kilobase region between the genes PKHD1 and MCM3 [23]. While scores for Th17 and Th1/Th2 subsets (i.e. ‘CD4+ CD25- IL17+ PMA-Ionomycin stimulated Th17 Primary Cells’ and ‘CD4+ CD25- IL17- PMA-Ionomycin stimulated MACS purified Th Primary Cells’; S1 Table) were extracted as-is, we took the average score of the two available CD4+ naïve T-cell subsets (i.e. ‘CD4 Naïve Primary Cells’ and ‘CD4+ CD25- CD45RA+ Naïve Primary Cells’). We identified the analogous human regions of previously identified functional murine CNS regions [25] by taking the top 20 most conserved intergenic sites between mouse and human in the LCR region using the VISTA browser (http://pipeline.lbl.gov/cgi-bin/gateway2). GenoSkyline-Plus scores in the 20 CNS sites and their 200-bp flanking regions were compared across different cell types.
Summary statistics for 45 GWAS are publicly accessible. Details for these studies are summarized in S3 Table. IGAP is a large two-stage study based upon genome-wide association studies (GWAS) on individuals of European ancestry. In stage-I, IGAP used genotyped and imputed data on 7,055,881 SNPs to meta-analyze four previously-published GWAS datasets consisting of 17,008 Alzheimer's disease cases and 37,154 controls (The European Alzheimer's disease Initiative–EADI, the Alzheimer Disease Genetics Consortium–ADGC, The Cohorts for Heart and Aging Research in Genomic Epidemiology consortium–CHARGE, and The Genetic and Environmental Risk in AD consortium–GERAD). In stage-II, 11,632 SNPs were genotyped and tested for association in an independent set of 8,572 AD cases and 11,312 controls. Finally, a meta-analysis was performed combining results from stages I and II. IGAP stage-I GWAS summary data is publicly accessible from IGAP consortium website (http://web.pasteur-lille.fr/en/recherche/u744/igap/igap_download.php). GWAS summary statistics for PD was acquired from dbGap (accession: pha002868.1). Details for AD and PD studies have been previously reported [8, 29].
Heritability stratification and enrichment analyses were performed using LD score regression implemented in the LDSC software (https://github.com/bulik/ldsc/). Annotation-stratified LD scores were estimated using dichotomized annotations, 1000 Genomes (1KG) samples with European ancestry [81], and a default 1-centiMorgan window. Enrichment was defined as the ratio between the percentage of heritability explained by variants in each annotated category and the percentage of variants covered by that category.
A resampling-based approach was used to assess standard error estimates [26]. Three tiers of annotations of different resolutions were used in enrichment analyses:
The smoothing strategy for GenoCanyon improves its ability to identify general functionality in the human genome [59]. GenoSkyline-Plus and smoothed GenoCanyon annotations were dichotomized using a cutoff of 0.5. Such dichotomization is robust to the cutoff choice due to the bimodal nature of annotation scores [6]. We selected 66 annotation tracks in the tier-3 analysis by removing all the fetal and embryonic cells, and taking the union of different Roadmap epigenomes for the same cell type (S2 Table). The 53 baseline annotations of LD score regression were always included in the model across all analyses as suggested in the LDSC user manual. Smoothed GenoCanyon annotation track was also included in tier-2 and tier-3 analyses to account for unobserved tissue and cell types. Of note, the proposed multiple regression model explicitly takes the overlapped functional regions across biologically-related cell types into account. Further, the linear mixed-effects model in LDSC does not assume linkage equilibrium, and therefore LD will most likely not introduce bias into heritability estimation and enrichment calculation. We removed the MHC region from our analysis due to its unique LD patterns.
A slightly different strategy was adopted when comparing the performance of different computation annotation tools. To make fair comparison, we dichotomized all annotation tracks using each score’s top 90% quantile calculated from SNPs with minor allele count greater than five in 1000 Genomes samples with European ancestry. We then followed the suggested protocol of LDSC and kept baseline annotations in the model while adding each annotation track one at a time.
We calculated chromosome-by-chromosome heritability percentage through summing up and normalizing per-SNP heritability estimated using LD score regression and tier-3 annotation tracks. Of note, since only GWAS summary statistics were used as the input, popular heritability estimation tools such as GCTA [82] could not be applied. The sums over complete chromosomes are compared with the sums over monocyte functional regions only. Notably, LDSC is conceptually different from some other tools (e.g. GCTA [82]) in its estimation of trait heritability. GCTA estimates the proportion of phenotypic variability that can be explained by SNPs in the GWAS dataset while LDSC aims to estimate the proportion of phenotypic variability explained by all the SNPs in samples from the 1KG Project. In practice, LDSC only uses HAPMAP SNPs to fit the LD score regression model and assumes that HAPMAP SNPs are sufficient for tagging all 1KG SNPs through LD [26]. Additionally, LDSC applies a few stringent SNP filtering steps for quality control reasons, e.g. removing SNPs with very large effect sizes (i.e. χ2 > 80), which leads to the removal of some SNPs in the APOE region in our analysis. Finally, we note that a recent method may potentially improve the heritability estimates based on LDSC [75].
To evaluate enrichment of pleiotropic sites in the monocyte functional genome, we partition the genome into windows with length of 1M bases. Sex chromosomes and windows without SNPs are removed in our datasets. For each disease (i.e. AD and PD), we label a window 1 if the following criteria are met.
Otherwise, the window is labeled 0. This labeling results in two binary vectors, one for each disease. A window marked as 1 for both AD and PD is a window of interest that suggests a possible association in monocytes-related DNA for both diseases in that region. We use a hypergeometric test to assess if such a pattern of local association appears more often than by chance. Windows marked as 1 for both diseases are subsequently curated to identify the association peaks that potentially have pleiotropic effects for AD and PD.
There is a moderate overlap of control samples between IGAP AD GWAS and the PD GWAS (KORA controls, N~480). To account for the bias introduced by sample overlap and other confounding factors, we designed a permutation-based approach. In each permutation step, we shuffle the annotation status while keeping the total proportion of annotated regions, and then pick out windows that meet condition 2. We calculate the p-value through comparing the observed number of windows that meet conditions 1 and 2 for both diseases with the empirical distribution acquired in permutations.
Of note, we also applied this approach using a window size of 500K bases. Results in all related tests remained similar.
We briefly describe the SNP reprioritization approach implemented in the GenoWAP software available on our server (http://genocanyon.med.yale.edu/GenoSkyline). First, we identify three disjoint cases for SNPs in a given GWAS dataset.
A useful metric for prioritizing SNPs is the conditional probability that the SNP is classified under case-1 given its p-value in the GWAS study, i.e. P(ZD = 1, ZT = 1 | p). We can denote this probability using Bayes formula as follows:
P(ZD=1,ZT=1 | p)= P(Case 1 | p)= f(p|Case 1)×P(Case 1)∑k=13f(p|Case k)×P(Case k)
(7)
First, P(Case 3) = 1 − P(ZT = 1) can be directly identified using GenoSkyline-Plus scores. We partition all the SNPs into two subgroups based on a mean GenoSkyline-Plus score threshold of 0.1. Notably, these probabilities are not sensitive to changing threshold [6]. In this way, we can directly estimate f(p|Case 3) = f(p|ZT = 0) by applying a histogram approach on the SNP subgroup with low GenoSkyline-Plus scores.
Next, we assume that SNPs that are functional in a tissue but not relevant to the phenotype will have the same p-value distribution to all other SNPs that are not relevant to the phenotype, which in turn behave similarly to SNPs that are not functional at all. We have previously demonstrated that this assumption is backed by empirical evidence [6]. More formally, this relationship is denoted as follows:
f(p|Case 2)=f(p|ZD=0, ZT=1)=f(p|ZD=0)=f(p|Z=0)
(8)
We estimate the distribution f(p|Z = 0) by using a similar approach to estimating f(p|ZT = 0), but partitioning SNPs using the general functionality GenoCanyon score instead of tissue-specific GenoSkyline-Plus score.
Finally, all remaining terms in Formula 6 can be estimated using the EM algorithm. The p-value distribution of the subset of SNPs located in tissue-specific functional regions (i.e. ZT = 1) is the following mixture:
f(p|ZT=1)=P(ZD=1|ZT=1)×f(p|Case 1)+P(ZD=0|ZT=1)×f(p|Case 2)
(9)
Density function f(p|Case 2) has been estimated in Formula (8) and f(p|Case 2) is assumed to follow a beta distribution, which guarantees a closed-form expression in the EM algorithm.
Notably, the APOE region was removed in the SNP reprioritization analysis for LOAD.
Summary statistics from both IGAP stage-I GWAS and stage-I+II meta-analysis are publicly available (http://web.pasteur-lille.fr/en/recherche/u744/igap/igap_download.php). We inferred z-scores from IGAP stage-II replication cohort using the following formula.
In this formula, Z1 and Z1+2 indicate z-scores from the stage-I GWAS and the combined meta-analysis, respectively. Ni indicates the sample size from the ith stage. This formula was derived from the sample size based meta-analysis model, an approach known to be asymptotically equivalent to inverse variance based meta-analysis [83].
GenoSkyline-Plus annotation tracks, tiers 1–3 LD score files, and scripts for generating GenoSkyline-Plus scores are freely available on the GenoSkyline server (http://genocanyon.med.yale.edu/GenoSkyline). All annotation tracks can be visualized using UCSC genome browser. Web server g:Profiler was used to perform pathway enrichment analysis [84]. The g:SCS threshold implemented in g:Profiler was applied to account for multiple testing. Locus plots were generated using LocusZoom [85]. Gene plots were generated using R package “Gviz”.
|
10.1371/journal.pgen.1004437 | Caspase Inhibition in Select Olfactory Neurons Restores Innate Attraction Behavior in Aged Drosophila | Sensory and cognitive performance decline with age. Neural dysfunction caused by nerve death in senile dementia and neurodegenerative disease has been intensively studied; however, functional changes in neural circuits during the normal aging process are not well understood. Caspases are key regulators of cell death, a hallmark of age-related neurodegeneration. Using a genetic probe for caspase-3-like activity (DEVDase activity), we have mapped age-dependent neuronal changes in the adult brain throughout the lifespan of Drosophila. Spatio-temporally restricted caspase activation was observed in the antennal lobe and ellipsoid body, brain structures required for olfaction and visual place memory, respectively. We also found that caspase was activated in an age-dependent manner in specific subsets of Drosophila olfactory receptor neurons (ORNs), Or42b and Or92a neurons. These neurons are essential for mediating innate attraction to food-related odors. Furthermore, age-induced impairments of neural transmission and attraction behavior could be reversed by specific inhibition of caspase in these ORNs, indicating that caspase activation in Or42b and Or92a neurons is responsible for altering animal behavior during normal aging.
| The approaching era of an “aging society” is receiving considerable attention amongst biomedical researchers in advanced nations. In order to understand the molecular mechanisms underlying age-related alterations of neural circuitry, we focused on caspase-3, a cysteine protease that induces apoptotic cell death, using the fruit fly Drosophila melanogaster, a model often used to study aging due to a short lifespan of approximately 30–60 days. Here, we describe the spatiotemporal activation of caspase-3 in aged fly brains and show that caspase-3 is specifically activated in select olfactory neurons essential for innate odor attraction behavior. Furthermore, we discuss how inhibition of caspase-3 activation in those select olfactory neurons can rejuvenate the sensitivity of innate attraction behavior in aged flies. These findings suggest that caspase-3 plays an active role in producing age-related alterations to neuronal physiology and circuit function associated with animal behavior.
| Neuronal dysfunction and cell death are hallmarks of age-related neurodegenerative disorders, such as Alzheimer's disease. Epidemiological and biomedical studies have demonstrated that both genetic and age-related factors are crucial for the development and progression of these disorders. Attempts to understand the underlying mechanism of functional alterations in neural circuits during “normal aging” are receiving considerable attention [1]–[3] and should provide new insights toward preventing and treating age-related disorders. However, our knowledge about whether and how neural circuits are remodeled and/or maintained during normal aging is still very limited.
Caspases are highly conserved cysteine proteases, which function as central regulators of apoptosis [4], [5]. Knockout mice lacking caspase-3, caspase-9, or the caspase activator, apaf-1, all exhibit reduced neuronal apoptosis and brain malformation [6]–[11], indicating that caspases are essential for normal brain development. In addition to their role in apoptosis, non-apoptotic roles for caspases, particularly in the nervous system, are being reported in vivo [12]. These roles include dendritic pruning in the developing Drosophila [13], [14], song habituation in birds [15], [16], synaptic long-term depression (LTD) in rat hippocampal neurons [17], [18], synaptic maturation of olfactory sensory neurons in mice [19], and early synaptic dysfunction in a mouse model of Alzheimer's disease [20], [21]. Although the essential role of caspases in developing and adult brains has been documented, the in vivo activation pattern of caspases has not yet been systematically investigated.
In this report, we began with mapping caspase activity throughout the entire lifespan of the fruit fly. Using a genetic probe for caspase-3-like activity (DEVDase activity) [13], we revealed spatiotemporal caspase activation in the adult brain. Moreover, we found that this caspase activation was particularly prominent in the antennal lobe (AL) and ellipsoid body, which are brain structures responsible for olfaction and visual place memory, respectively [22]–[24]. Interestingly, when we further investigated caspase activity in the antennal lobe, we determined that caspases were activated in an age-dependent manner in select ORNs, particularly in Or42b and Or92a neurons that are essential for mediating innate attraction to food odors [25], and that elevation of caspase activity caused ORN death. Furthermore, two-photon calcium imaging of projecting neural dendrites (secondary neurons receiving input from ORNs) indicated that aging reduced sensitivity of the related olfactory glomeruli, which could be suppressed by the expression of p35, a caspase inhibitor. Lastly, we found that the age-related impairment of innate attraction behavior was also significantly suppressed by the inhibition of DEVDase in Or42b and Or92a neurons. Taken together, our data suggest that caspase activation in the aging brain is spatio-temporally regulated and actively contributes to age-related alterations of neural function.
To monitor DEVDase activity in the brain of the adult Drosophila, we used a genetically encoded DEVDase probe consisting of a transmembrane mouse CD8 (mCD8) protein and a yellow fluorescent protein (Venus) linked by the caspase-3-cleavage sequence derived from human poly ADP ribose polymerase (PARP) [13] (Figure 1A). The activated form of DEVDase cleaves this probe, known as mCD8::PARP::Venus, into two fragments. Moreover, an antibody against cleaved PARP (anti-cPARP Ab) can specifically detect one of these two fragments; the immunohistochemical cPARP signal thus generated reflects levels of activated DEVDase.
We expressed mCD8::PARP::Venus in postmitotic adult neurons marked by elav-Gal4 and found that the brains of very young (0–1 day old) and very old (30–45 days old) flies exhibited higher cPARP signaling frequency than other age groups (Figure 1B). In young adult brains, cPARP signals were primarily detected in the subesophageal ganglion (SOG) and in the midline region; however, the intensity of these signals varied in the SOG of individual brains (Figure 1C and 1D). The cPARP brain pattern was similar between males and females, although cPARP appeared more frequently in males than in females (Figure 1B, 1F and 2C). These results are consistent with previous findings obtained using the terminal deoxynucleotidyl transferase dUTP nick end labeling (TUNEL) assay and an antibody aimed to detect active forms of caspase-3 [26]. In contrast, brains of aged flies tended to exhibit cPARP in the dorso-medial corner of the AL and in the ellipsoid body (Figure 2A–2C). Other neuronal processes showed cPARP signals in the aged brain, but the labeling appeared to be random (Figure 2D and 2E; some data not shown). Importantly, we determined that the cPARP pattern in the AL was highly stereotyped in aged flies of both sexes (32.2% of male brains and 8.2% of female brains, at 45 days post-eclosion); hence, we focused on the AL neural circuit.
The AL is the first olfactory center in the Drosophila brain that consists of ∼50 glomeruli, which are ball-shaped synaptic structures that receive axons of ORNs and dendrites of projection neurons (PNs) and are interconnected by local interneurons (LNs) [22]. To identify the neuronal subtypes with DEVDase activity, we expressed mCD8::PARP::Venus using pebbled-Gal4 (all ORNs), NP1227-Gal4 (GABAergic LNs), and GH146-Gal4 (two-thirds of the PNs). Only pebbled-Gal4 generated reproducible cPARP signals in aged fly brains, which were suppressed by p35 (Figure 3). Further, we used 17 Or-Gal4 drivers to express mCD8::PARP::Venus in each ORN subtype. Surprisingly, we found that DEVDase was frequently activated in the axons of the Or42b, Or92a, and Or35a neurons but rarely in the other classes of ORNs that we tested (Figure 4A–4C). Or47b neurons also showed DEVDase activation, but cPARP signal intensity in these cells was very low. These data indicate that DEVDase activation is age-dependent in specific subsets of ORNs.
Strong activation of DEVDase in a cell body typically leads to apoptosis [5]. To determine whether DEVDase activation in Or42b and Or92a neurons caused apoptosis, we examined the activation state of DEVDase in ORN cell bodies located in the third segment of the fly antenna. In aged flies, among mCD8::PARP::Venus-positive Or42b and Or92a neurons, we found that a small fraction of neurons were positive for cPARP (Figure 5A): 6.7% of Or42b (n = 60 cells) and 3.1% of Or92a (n = 65 cells). On the other hand, none of the ORN cell bodies showed cPARP signals in young flies (0% Or42b neurons, n = 71 cells; 0 of Or92a neurons, n = 81 cells). In addition, some Or42b neurons are positive for both TUNEL and cPARP signal (Figure 5B). To further examine death of ORNs of aged flies, we expressed a nuclear-localized enhanced cyan fluorescent protein (ECFP) (Histone H2B::ECFP) in each type of ORN and found that there was a significant decrease in Or42b and Or92a neurons during aging (Figure 5C and 5D), while the number of Or85a neurons remained constant (Figure 5E). The expression of several apoptosis inhibitors (p35; the dominant-negative form of Drosophila caspase-9 Dronc, Dronc-DN [27]; and microRNA for reaper, hid, and grim, miRHG [28]) effectively reversed the trend of an age-dependent decrease in Or42b and Or92a neurons (Figure 5C and 5D). In contrast, the expression of these apoptosis inhibitors did not affect neuron numbers in young flies (Figure S1). These data suggest that Or42b and Or92a neurons die, at least in part, by caspase-mediated cell death in an age-dependent manner.
Since the above-mentioned ORN subtypes die with DEVDase activation during aging, we hypothesized that odor-evoked behavior through the Or42b and Or92a neurons would be impaired with age. To test this possibility, we measured innate attraction behavior to apple cider vinegar in young and aged flies. Apple cider vinegar excites six glomeruli including DM1 and VA2, which are innervated by the axons of Or42b and Or92a, respectively [25]. We found that attraction to apple cider vinegar was significantly decreased in aged flies, and that this effect was reversed by p35 expression in Or42b and Or92a neurons (Figure 6A). These results clearly indicate that DEVDase activation in Or42b and Or92a neurons is the main cause of age-related impairments to innate attraction behavior.
Lastly, we wanted to investigate age-dependent change of glomerular sensitivity by using two-photon microscopy imaging with a genetically encoded calcium sensor, GCaMP [29]. We measure the sensitivity of a given glomerulus by monitoring its output projection neurons. Specifically, we image the dendritic calcium levels of PNs innervating each glomerulus, we applied apple cider vinegar to the flies bearing GH146-LexA and LexAop-GCaMP1.3-ires-GCaMP1.3. We found that, in young fly brains, the DM1 glomerulus was robustly activated in response to apple cider vinegar in a concentration-dependent manner (Figure 6B and 6C). In contrast, the DM1 glomerulus was only weakly activated even at high concentrations of apple cider vinegar in aged flies. Moreover, the expression of p35 in Or42b neurons increased sensitivity of DM1 to vinegar (Figure 6D and 6E). These results are consistent with our observations on attraction behavior (Figure 6A), and indicate that the olfactory response of the DM1 glomerulus is impaired during aging due to DEVDase activation in Or42b neurons.
In the current study, we demonstrate that normal aging increases caspase activity, leading to age-related cell death, reduced olfactory sensitivity, and impaired innate attraction behavior. Since caspase-3 activity appears to contribute to synaptic LTD in the rat hippocampus [17], [18] and to early synaptic dysfunction in mouse models of Alzheimer's disease [20], further studies of age-related increases in caspase activity and its role in Drosophila neuronal excitability and cell death are warranted. Our discovery that specific types of ORNs show DEVDase activation and cell death is the first example of age-related, stereotyped cell death of neurons within a specific network. The activation of the DM1 and VA2 glomeruli, which are innervated by Or42b and Or92a neurons, respectively, is essential for innate food attraction behavior [25]. Thus, our observations may help to explain age-related changes in innate animal behavior.
In addition to the olfactory system, we found stereotyped caspase activation in the ellipsoid body, the brain region involved in olfactory memory consolidation [23] and visual place memory [24]. This observation could imply the possible contribution of caspase activation to age-related memory impairment (AMI). Like other animals, aged flies exhibit AMI, which corresponds to an increase of cAMP-dependent protein kinase (PKA) in the mushroom body but not in the ellipsoid body [30]. Because caspase is required in synaptic LTD [17], it might be interesting to investigate the possible implication of caspase activation in the ellipsoid body for olfactory or visual memory and whether its role is apoptotic or non-apoptotic.
The results of our current study reveal an interesting phenomenon in that age-related caspase activation only occurred in select ORNs. One possible explanation for this is the continuous activation of Or42b and Or92a neurons by food odors. As previously discussed, Or42b and Or92a respond to odors that flies recognize as food, such as apple cider vinegar [25]. Under regular experimental conditions, flies are cultured in a food-containing vial leading to continuous activation of Or42b and Or92a for the duration of a fly's lifespan. To test whether this continuous ORN activation was responsible for the eventual age-related caspase activation in these neurons, we aged flies in a vial containing yeast paste and examined cPARP signal in Or42b neurons. Interestingly, we found that the age of onset and strength of caspase activation in Or42b neurons was similar to what we found in flies that had been cultured in normal food (data not shown), suggesting that a continuous food odor is not solely responsible for inducing age-related caspase activation. As for the involvement of neural activity in the maintenance of ORN axons [31], further studies of culturing conditions containing restricted odors and the genetic manipulation of ORN neural activity would help to clarify these issues.
In addition to neuronal activity, aging itself might produce ORNs with differential sensitivity to neuronal excitability or toxicity. This idea is supported by a report from Tonoki et al. (2011), showing that forced expression of a truncated form of human Machado-Joseph disease protein with an expanded polyglutamine domain in the adult Drosophila eye at 20–24 days after eclosion causes more severe neurodegeneration than expression at 0–4 days of age [32]. These observations suggest that neuronal identity, including sensitivity to a toxic factor generated by age-related neuronal excitability, may be continuously changing over the course of a fly's lifecycle.
Differential expression of the effector caspases, drICE and Dcp-1, may determine the spatiotemporal specificity of caspase activation. Previous studies have suggested that expression levels of these caspases reflect the apoptotic potential of cells, and that drICE is more effective than Dcp-1 to induce apoptosis [33]. Interestingly, it has been shown that activation of drICE and Dcp-1 can only be detected in degenerating dendrites but not in the cell body of Drosophila C4da neurons [34]. We also previously reported a similar phenomenon in mice where caspase-3 could be detected in the developing axons of olfactory sensory neurons but not in the cell body [19]. In the current study, we found that caspases were activated in both the axon and cell body of a subset of Or42b neurons that eventually go on to die in an apoptotic manner, while the subset that did not show elevated levels of caspases went on to survive. Therefore, we expect that both drICE and Dcp-1 were likely activated in dying Or42b neurons, while either drICE or Dcp-1 was activated in the degenerating axons of surviving Or42b neurons.
It has been reported that the Or42b and Or92a genes are the most conserved in the drosophilid olfactory subgenome and seem to be utilized to detect odors from wild lily (Solomon's lily) in other drosophilid species [35]. Thus, Or42b and Or92a seem to possess the most fundamental function among the ∼50 types of olfactory neurons. Moreover, this suggests that caspase activation in these neurons might have a greater impact on animal behavior. Therefore, we believe that this would be an ideal experimental paradigm to investigate age-dependent changes of innate behaviors.
Lastly, our current findings suggest that while it is clear that caspase activation plays a crucial apoptotic role in the adult olfactory circuit, caspase activation may also have non-apoptotic functions. This is in light of that fact that while we were only able to detect a few TUNEL-positive cells among cPARP-positive ORNs, we noted a significant reduction in odor-evoked neural activity in the DM1 glomerulus of aged flies. Richard et al. recently identified an age-dependent disruption of a specific synaptic layer in the mouse olfactory bulb without any detectable neuronal loss [36]. In addition, it has been shown that caspase-9 is activated in aged olfactory bulb neurons without affecting the number of these cells [37]. These observations, together with our study, prompt questions concerning the ecological and pathological significance of caspase activation, or synaptic dysfunction, in specific groups of neurons or synapses of the adult olfactory circuit. Investigating the relationship between age-related alterations in neural circuits may provide clues to understanding the neural basis of impaired sensory and cognitive performance during normal aging and senile dementia.
The following transgenic lines were used: elav-Gal4, Or-Gal4, (Bloomington Stock Center), NP1227-Gal4 (Kyoto Drosophila Stock Center), UAS-mCD8::PARP::Venus [13], pebbled-Gal4 [38], GH146-Gal4 [39], UAS-miRHG [28], UAS-reaper [40], UAS-p35 (a gift from Bruce Hay), UAS-Dronc-DN [27], UAS-mCD8::GFP [41], UAS-H2B::ECFP [42], LexAop-GCaMP1.3-ires-GCaMP1.3 [25], and GH146-LexA [43]. All flies were maintained in a 25°C incubator and transferred to vials with fresh food every 3 to 4 days.
Immunohistochemistry of the Drosophila adult brain was performed as previously described [44]. To stain ORN cell bodies, we dissected antennae from flies, fixed them in 4% (vol/vol) paraformaldehyde/0.3% (vol/vol) phosphate-buffered saline with Triton X-100 (PBT) at room temperature (R.T.) for 30 min, mounted them in OCT, and cut 14 µm-thick sections on a cryostat. Slides were then re-fixed with 4% (vol/vol) paraformaldehyde/0.3% (vol/vol) PBT at R.T. for 30 min, washed with 0.3% (vol/vol) PBT, and labeled using standard techniques. Antibodies used include rat anti-mouse CD8 antibody (1∶100, MCD0800, Invitrogen), rabbit anti-cleaved PARP (Asp214) antibody (1∶100, #9541, Lot.7, Cell Signaling), rabbit anti-cleaved PARP antibody [Y34] (1∶100, ab32561, Abcam), nc82 mouse monoclonal antibody (1∶40, Developmental Studies Hybridoma Bank), anti-rat Alexa488 (1∶250), anti-rabbit Cy3 (1∶1000), and anti-mouse Cy5 (1∶1000) (Jackson Laboratory). Confocal images were captured using a Leica SP5 confocal microscope.
TUNEL assay of antennal cryosections was performed using an In Situ Cell Death Detection Kit, TMR red (Roche). Tissues were mounted in SlowFade Gold Antifade Reagent with DAPI (Life Technologies).
Cantonized w1118 [w(CS10)] flies were used as behavioral controls in our experiments. The flies used for behavioral assays were out-crossed to the w(CS10) background. All fly stocks were maintained at 25°C and 70% relative humidity under a 12/12 h light-dark cycle. For behavioral studies, about 50 male flies were placed into food vials and transferred to fresh food vials every 3 or 4 days until the age for behavioral assay was reached. Behavioral assays were performed under dim red light at 25°C and 70% relative humidity. Attraction to apple cider vinegar was measured for 3 or 30-day-old flies. Briefly, flies of each type were loaded into a T maze in which they could make a choice between two arms. Flies were allowed 2 min to choose between an odor of apple cider vinegar or air. Apple cider vinegar was diluted in water to 0.3% (v/v), which elicits robust attraction behavior in 3-day-old flies. The performance index (P.I.) was defined as the ratio of the difference in number of flies that chose the air laced with or without apple cider vinegar odor to the total number of flies that chose either side.
Calcium imaging was performed as described [25], [29], [45]. For odor stimulation experiments, a constant airflow of 1 L/min was applied to the antennae via a tube of 12 mm diameter. Odor onset was controlled by mixing a defined percentage of carrier air redirected through odor bottles (presented as percent saturated vapor pressure, or %SV) as previously described [45].
Third segments of the antennae were dissected in phosphate-buffered saline (PBS) and mounted with FocusClear mounting solution (Cedarlane Laboratories). All cell images (H2B::ECFP) were taken live by a Leica SP5 confocal microscope within 15 min. Cell numbers were manually counted with ImageJ software and statistical analyses were performed using Stastical Package for the Social Sciences (SPSS 16.0) (IBM) and Prism software (GraphPad).
|
10.1371/journal.pcbi.1003703 | Inferring Clonal Composition from Multiple Sections of a Breast Cancer | Cancers arise from successive rounds of mutation and selection, generating clonal populations that vary in size, mutational content and drug responsiveness. Ascertaining the clonal composition of a tumor is therefore important both for prognosis and therapy. Mutation counts and frequencies resulting from next-generation sequencing (NGS) potentially reflect a tumor's clonal composition; however, deconvolving NGS data to infer a tumor's clonal structure presents a major challenge. We propose a generative model for NGS data derived from multiple subsections of a single tumor, and we describe an expectation-maximization procedure for estimating the clonal genotypes and relative frequencies using this model. We demonstrate, via simulation, the validity of the approach, and then use our algorithm to assess the clonal composition of a primary breast cancer and associated metastatic lymph node. After dividing the tumor into subsections, we perform exome sequencing for each subsection to assess mutational content, followed by deep sequencing to precisely count normal and variant alleles within each subsection. By quantifying the frequencies of 17 somatic variants, we demonstrate that our algorithm predicts clonal relationships that are both phylogenetically and spatially plausible. Applying this method to larger numbers of tumors should cast light on the clonal evolution of cancers in space and time.
| Cancers arise from a series of mutations that occur over time. As a result, as a tumor grows each cell inherits a distinctive genotype, defined by the set of all somatic mutations that distinguish the tumor cell from normal cells. Acertaining these genotype patterns, and identifying which ones are associated with the growth of the cancer and its ability to metastasize, can potentially give clinicians insights into how to treat the cancer. In this work, we describe a method for inferring the predominant genotypes within a single tumor. The method requires that a tumor be sectioned and that each section be subjected to a high-throughput sequencing procedure. The resulting mutations and their associated frequencies within each tumor section are then used as input to a probabilistic model that infers the underlying genotypes and their relative frequencies within the tumor. We use simulated data to demonstrate the validity of the approach, and then we apply our algorithm to data from a primary breast cancer and associated metastatic lymph node. We demonstrate that our algorithm predicts genotypes that are consistent with an evolutionary model and with the physical topology of the tumor itself. Applying this method to larger numbers of tumors should cast light on the evolution of cancers in space and time.
| Many clones exist within each cancer, and selective pressure imposed by environmental factors, most notably treatments directed at tumor eradication, favors the emergence of clones that grow increasingly resistant to successive rounds of therapy. Incorporating this intra-tumor heterogeneity into strategies for planning, monitoring, and revising cancer treatment could improve outcomes for oncologists and their patients. Therefore, methods for estimating the number, size and mutational content of clones within a patient's tumor are being explored.
New approaches are being developed to assess the clonal content of a given tumor. Methods based on the interrogation of individual cells have relied on the use of fluorescent markers [1], [2] or single cell sequencing [3]–[6]. Whereas fluorescence-based approaches are inevitably limited by the relatively small number of features they can accommodate, single cell sequencing brings the highest possible resolution to characterizing an individual patient's tumor. Nonetheless, single cell sequencing also faces obstacles to its widespread implementation. Evaluating sufficiently large numbers of single cells to obtain statistical power can be prohibitive, for technical or financial reasons. Additionally, it is often difficult to ascertain the identity of the cells being sequenced, and details regarding the spatial positioning of cells relative to each other and to other cells in the tumor are lost when the single cells are obtained. These disadvantages pose significant challenges to the widespread adoption of single cell sequencing as a means for assessing tumor heterogeneity.
Complementing single cell approaches are efforts to deconvolve clonal subpopulations based on the frequencies of mutated alleles within one or more bulk tumor specimens. Shah et al. [7], who sequenced a breast cancer at the time of diagnosis and nine years later, after metastasis, pointed out that allele frequencies of the mutations shared between the two samples could be used to segregate primary mutations into those that occur in a dominant versus subdominant clone. This insight is the basis for a variety of approaches that apply clustering algorithms to mutation allele frequencies, including kernel density estimation [8] and Dirichlet process modeling applied either to the allele frequencies [9] or to a combination of allele frequency, loss-of-heterozygosity status and copy number [10]–[13].
Clearly, statistical power to infer variants and, ultimately, clonal composition, is increased if multiple samples are available for analysis. Accordingly, various studies have examined the progression of cancer within one or more patients over time. Sets of variants that exhibit similar allele frequencies within a single sample are suggestive of a clonal population. Hence, clustering methods to identify groups of mutations associated with a single clone have been applied. For example, kernel density estimation has been applied to allele frequencies from tumor-relapse pairs from eight acute myeloid leukemia (AML) patients [14] and from seven secondary AML patients [15].
An orthogonal approach taken by Newberger et al. [16] employs triplet samples of neoplasia, matched normal and carcinoma from six patients to infer lineages of various genetic events. They characterize each locus in terms of a binary vector representing the presence of the mutation across the various samples and then group the loci into classes on the basis of these vectors. After filtering low frequency classes, the classes are used to manually construct a phylogenetic tree. The focus of the study is to identify the shared characteristics of the evolutionary process across six patients with breast cancer.
In the current study, we adopt an alternative approach to identify clonal structure. Rather than measuring allele frequencies in multiple samples from the same patient over time, we physically subdivide a single breast cancer specimen and measure allele frequencies within each subsection (Figure 1). We are aware of two previous studies that have adopted such an approach. Yachida et al. [17] analyzed seven metastatic pancreatic cancers, sequencing from multiple samples per patient. Clones are initially defined relative to sample types (peritoneal, liver and lung metastases). Subsequently, the tumors from two patients are resected and a clonal phylogeny is inferred manually. More recently, Gerlinger et al. [18] carried out exome sequencing followed by targeted deep sequencing on samples from four patients with renal carcinoma. Each primary tumor was divided into 9 regions, and a phylogeny was manually constructed by assuming that higher alternate allele frequencies correspond to earlier mutations. In neither of these studies was an algorithm proposed to automatically infer from such data both the clonal genotypes and the relative frequencies of the clones within each subsection.
The method proposed here bears some similarity to the recently proposed Tree Approach to Clonality (TrAp) method [19]. The TrAp algorithm aims to identify the number, relative frequencies and genotypes of clones within a tumor using a formalism somewhat similar to ours, based on matrix decomposition. However, rather than analyzing data from multiple sections, the authors use as input a single set of variant allele frequencies and then constrain the resulting optimization problem by introducing a series of four assumptions about cancer evolution. It is not clear whether the method can easily generalize to analysis of data from multiple sections or multiple time points.
Here we describe a generative binomial model that incorporates information from multiple sections from a single tumor at a single time point to infer the frequencies and genotypes for a specified number of clones. An implementation of our algorithm is available through Bioconductor as an R package called Clomial (http://www.bioconductor.org/packages/release/bioc/html/Clomial.html). We use Clomial version 1.1.7 to apply this approach to a breast cancer specimen and demonstrate that the results from our model predict relationships that are phylogenetically and spatially plausible.
We assume that a tumor is comprised of multiple populations of cells (“clones”), each with a unique genotype, and that these populations are heterogeneously distributed within the tumor itself. We collect, from several physical subsections of the tumor, shotgun sequencing reads. We also collect sequencing data from a non-tumor subsection from the same patient. Using the called genotypes from the normal subsection, and restricting ourselves to positions that are homozygous in the normal subsection, each read from a tumor subsection exhibits either a normal allele or a variant allele at each location. We exclude positions that exhibit homozygous normal alleles in all of the tumor subsections. Our goal is to infer, from the remaining mutated positions, the genotype of each clonal population and their relative frequencies within each physical subsection of the tumor.
Formally, the problem can be stated as follows. Note that we use bold face letters for random variables, and that and respectively denote the row and the column of matrix . We are given two primary input matrices and , where is the number of mutated loci, is the number of subsections (of which one is normal and are tumor), is the total number of reads (i.e., the coverage) at locus in subsection , and is the number of cancerous reads (those supporting the mutation) at locus in subsection . We assume, without loss of generality, that the first of the subsections corresponds to normal tissue, and that the remaining subsections are from the tumor. In addition, we consider , the number of distinct clones in the tumor, as a hyperparameter, and train a model based on a given value of . We assume that the first clone corresponds to the normal cell population and the tumor is composed of tumor clones. Later, we will discuss whether can be estimated from the data. Our task is to infer two matrices: a clone frequency matrix in which is the proportion of cells of clone in subsection , and a genotype matrix in which if clone has the variant allele at locus , and otherwise. The first column of contains all zeroes because it represents the “normal clone.” By definition, each column of sums to . Also, by construction, the first column of corresponds to the normal subsection and hence consists almost entirely of zeroes, although small non-zero counts may be possible due to contamination from tumor or due to sequencing error. If the first column of consisted entirely of zeroes, then we would expect the first column of to be of the form , but in order to allow for the possibility that the allegedly normal subsection can have slight tumor contamination, we infer the first column of (as well as the other columns).
We propose to solve this problem using a generative model whose parameters are learned via expectation-maximization (EM) [20]. Accordingly, we define a matrix of hidden variables representing the unknown genotypes of the clones; for instance, if , then the clone has a tumor allele at the locus. We assume that each follows an independent Bernoulli distribution with parameter , i.e.,(1)
We also assume that if a mutation is present in a particular clone, then at that locus the clone is heterozygous with copy number equal to 1. Therefore, for subsection , if clone has a mutation at locus (), then its contribution to the observed count of cancer alleles is by , half of its proportion in the subsection. Conversely, if a clone does not have a mutation at (), then it does not contribute to the count of variant alleles. By summing up the contributions of all clones, we obtain the total probability that an observed read corresponds to a variant allele rather than a normal allele. Therefore, the probability that a read contains the variant allele at locus in subsection is given by(2)where is the row of , and is the column of . Finally, we introduce a matrix of random variables representing the observed data, where is the number of reads exhibiting the variant allele at locus in subsection . This matrix encodes our primary assumption about the distribution of the data: for each and , we observe an independent sample of that has a binomial distribution with two parameters and , i.e.,(3)
The first parameter of this distribution is the (known) total number of reads at locus in subsection . The second parameter, , is the probability of observing a variant allele; it will be inferred by EM.
Given the joint distribution over observed variables and latent variables , governed by parameters , our goal is to maximize the likelihood function . We do so using EM, exploiting three assumptions: (1) that each subsection contains non-zero normal contamination, i.e., for all , (2) independence of the subsections from each other, and (3) independence of mutations from each other. The first assumption is based on the widely accepted difficulty associated with obtaining perfectly pure samples of tumor cells [21], [22]. The two independence assumptions essentially state that each locus and each sample is informative. These assumptions are unavoidable: in the presence of very high dependence, only very limited information about the underlying clonal composition of the tumor would be provided by the loci and samples. Furthermore, it is worth noting that these independence assumptions are made conditional on the parameters in the model: that is, the elements of are independent conditional on and . In other words, if we knew the true underlying parameters for the model (that is, the true genotypes for the clones, and the true proportion of each clone present in each sample), then the actual number of “tumor” reads that we would observe for each locus-sample pair would be independent.
While the formulation of our inference problem shows some similarity to well-studied matrix factorization problems [23]–[25], such techniques cannot be directly applied here. Unlike most matrix factorization techniques, which assume a normal distribution, our observations are binomially distributed. Moreover, the elements of the latent matrix are binary, and each column of must sum to 1. These constraints required us to develop a customized inference algorithm.
To frame the EM optimization, we consider the following complete-data log likelihood function of the model:(4)which can be computed as follows (for details see Note S4 in Text S1):(5)where .
Our goal is to find the parameters which maximize the likelihood. Because our model involves the hidden variable , we cannot directly maximize the given in Equation 5 with respect to . Instead, we use the EM algorithm to fit the model to the data [26]. EM is an iterative algorithm with two steps—E (for expectation) and M (for maximization)—in each iteration. In the E step, we use the current estimates of the parameters, , to compute the conditional expectation of . In the M step, we find the new parameters that maximize the conditional expectation.
To validate our implementation of the EM optimization procedure and to understand our model's behavior, we produced simulated deep sequencing data and measured the extent to which the model successfully recovers the true clonal structure of the data.
For each simulation, we began by randomly generating four matrices. First, we generated a simulated matrix of total read counts with respect to a fixed number () of loci and a fixed number () of subsections with a mean coverage of 1000 reads per locus. The matrix was generated by independently sampling each column (corresponding to a single subsection) from a multinomial distribution , where the parameters and correspond to the total number of trials, and the probability of success for each of the loci, respectively. Second, for any clone number , we generated a corresponding Boolean matrix , in which the entry at row and column indicates whether locus exhibits the variant allele in clone . Entries in were generated independently from a Bernoulli distribution with a probability of success , with the exception of the first (“normal”) column of , which contains all zeroes. Third, we generated a clone frequency matrix as follows: each element of is independently drawn from a Uniform distribution, and then each column of was divided by the column sum, so that the columns summed to 1. We then set so that the first column of corresponds to the normal subsection. Finally, for each locus and subsection , we generated the observed number of variant alleles by sampling from a binomial distribution with parameters (representing the total number of reads) and (representing the probability that a given read corresponds to the variant allele). This last step complies with our primary assumption about the distribution of the data (Equation 3).
We ran the EM algorithm using the simulated data and and then evaluated the extent to which the estimated clone frequency matrix and mutation probability matrix differed from the corresponding true matrices and . Specifically, we computed the genotype error , defined asand the clone frequency error ,
Note that, because we did not know which columns of correspond to which columns of , we compared to every permutation of the columns of and selected the permutation that resulted in the smallest genotype error. The selected permutation was then also used in the calculation of the clone frequency error.
Our simulation results (Figure 2) exhibit two primary trends. The overall error rate, as measured by either genotype or clone frequency error, decreases systematically as the number of subsections increases, and increases as the number of clones increases. Overall, both error rates are low, especially for . The observed trends are expected: for a fixed number of clones, the availability of more subsections leads to more accurate estimation of the true parameter values; and for a fixed number of subsections, the presence of more clones leads to a greater number of parameters that must be inferred, leading to greater error in estimation.
To assess the affect of sequencing error on the performance of Clomial, we added noise to the simulated data and repeated the above experiments. Specifically, we modeled noise by Bernoulli random variables with probability of success interpreted as the probability that a non-tumor allele is read as a tumor allele or vica versa. Running the EM algorithm on the noisy data revealed that Clomial is robust with respect to noise for all reasonable levels of sequencing error (Figure S6) in Text S1.
We obtained breast cancer tissue from a 44 year old premenopausal female with infiltrative ductal carcinoma (IDC) with ductal carcinoma in situ (DCIS), stage pT1c pN1, Grade II/III, estrogen receptor (ER) positive, progesterone receptor (PR) positive and Her2 negative. Axillary lymph node dissection revealed that one out of 13 nodes was positive for metastatic disease. A total of 6 tissue sections were obtained, including 2 sections from adjacent normal breast tissue, 3 from the primary breast cancer, and 1 from the positive lymph node. The tumor content, including both IDC and DCIS, ranged from 40% to 55% in the primary tumor and axillary lymph node tissue sections based on pathological examination. For subsequent analysis, each tissue section was subdivided into subsections (Figure 3).
To identify mutations and quantify allele frequencies, we performed two rounds of DNA sequencing. Initially, DNA was extracted from each individual subsection and subjected to exome capture followed by Illumina sequencing. Variants were detected independently in each subsection using the SeattleSeq Annotation Server. We focused on single nucleotide variants and short indels that exhibited a coverage of reads in at least one of the subsections, ranking them using DeepSNV [31] and Fisher's exact test (Methods). This analysis produced an initial set of 281 variants (Dataset S1).
To better quantify the allele frequencies at these loci, we designed primer pairs surrounding each locus and used these primers to perform a second round of targeted DNA sequencing. This experiment successfully sequenced 244 of the 281 loci, with a mean and median coverage of 1615 and 1118, respectively, reads per locus. Each of these loci was individually validated by visual inspection using the Integrative Genomics viewer (IGV). Manual inspection showed that many of the initially identified mutations were flanked by homopolymer repeats, suggesting that the alternate alleles were read calling errors, rather than true mutations [32]. For all downstream analysis we focused on a set of 17 confirmed somatic variants. For clarity of presentation, we refer to each somatic variant by the chromosome where it resides, appending a letter if more than one somatic variant occurred within a chromosome (Table S1 in Text S1). The targeted sequencing thus produced two 17-by-12 matrices containing, respectively, the total coverage and the tumor allele count at each locus (Table S1 in Text S1). Visual inspection of the allele frequency profiles shows, not surprisingly, a markedly different pattern of allele frequencies among the subsections from primary and metastatic sites (Figure 3). In addition, several of the samples (e.g., P1-4 and P3-1) exhibit consistently lower frequencies across all loci, presumably indicating a higher prevalence of normal cells within these samples.
We applied our EM optimization procedure to the two counts matrices, varying the number of assumed clones from C = 3 up to C = 6. For each value of C, we ran EM 100,000 times from different random initializations, and we selected the solution with the highest likelihood (Figure 4). The resulting three-clone solution identifies two mutations, chr4a and chr9b, that occur in both the primary and metastatic samples and segregate the remaining mutations into nine that occurred in the primary tumor and six that occurred in the metastatic lymph node. The four- and five-clone solutions further subdivide the primary tumor mutations, and the six-clone solution separates the two metastatic mutations into distinct clones.
To better understand the inferred clonal landscape, we investigated the relationship between clone frequencies and the anatomy of the three primary and one metastatic tumor sections. We hypothesized that clone frequencies should vary smoothly between adjacent subsections, reflecting the physical spread of successful clonal populations. This hypothesis is supported by the data (Figure 5 and Figure S1 in Text S1). The trends are most striking in sections P1 and P2, for which we obtained four separate subsections. In each case, the primary clone frequencies vary in a monotonic fashion as we traverse the sample. Given that the EM inference procedure was provided with no information about which subsection was derived from which section, nor the relative orientation of the subsections to one another, the smoothly varying frequencies among adjacent subsections provides evidence that the method has successfully identified true clonal variation.
Cancer progression is an evolutionary process in which clones accrue mutations over time, forming new clones. Accordingly, it should be possible to organize the clonal progression of a tumor into a phylogenetic tree with the founder clone at the root. We therefore investigated whether the clones inferred by our EM procedure obey some simple phylogenetic constraints, with two complementary goals. First, because our EM procedure makes no use of phylogenetic constraints, this analysis can provide further evidence for the validity of our inferred solutions. Second, the phylogenetic analysis has the potential to provide significant insights into the clonal and mutational history of this specific cancer.
We started with the C = 3 solution to our EM algorithm, manually constructing a phylogenetic tree in which each node is a clonal population, and edges are marked with the mutations that occurred in the evolution from the parent clone to the offspring (Figure 6A). This particular tree shows two founder mutations, chr4a and chr9b, occurring prior to metastasis, six mutations occurring along the metastatic lineage, and nine along the primary lineage. This is the only phylogenetic tree that is consistent with the inferred clonal genotypes.
In contrast, for the solutions inferred from the EM algorithm assuming C = 4 through 6, we found that it is not possible to construct a tree without requiring that the same mutation occur independently along multiple branches. We therefore considered all possible “nearby” trees (where “nearby” means that, among the distinct rows of the genotype matrix, the two trees differ by only one bit) that produce a valid phylogenetic tree with no repeated mutations. For example, for the C = 4 solution, we evaluated the likelihood of six nearby trees, yielding log-likelihoods of −28482, −21282, −7500, −6692, −5659, and −4333 (Table S2 in Text S1). The highest of these likelihoods is −4333, compared to −4244 for the solution initially inferred by EM. The selected solution requires changing only one bit in the genotype matrix from “0” to “1” (indicated by asterisks in Figure 4). The resulting phylogenetic tree (Figure 6B) closely resembles the C = 3 tree, except that one mutation initially assigned to the metastatic clone C3 is instead assigned to clone C2 in the C = 4 tree. Also, the nine mutations associated with the primary section in the C = 3 tree are further subdivided into three that occur shortly after metastasis and six that lead to clone C1. Reassuringly, the C = 5 and C = 6 solutions, constructed in a similar fashion (Figure 6C–D), are largely consistent with this story, each introducing a subdivision among the existing sets of mutations to produce a larger set of clones. Among these trees, the only inconsistencies concern (1) three mutations (chr5, chr9a and chr20b) that occur later according to the C = 4 solution than according to the C = 5 or C = 6 solutions and (2) two mutations (chr1 and chr4b) that are assigned their own branch, directly off the normal clone, in the C = 5 and C = 6 solutions. In practice, the chance that a randomly generated genotype matrix would produce a valid phylogenetic tree is vanishingly small (Note S3 in Text S1). Therefore, the fact that each of our inferred solutions very nearly produce a valid phylogenetic tree provides evidence for the validity of these solutions.
We also investigated the extent to which the observed mutation frequencies obey the phylogenetic tree. In principle, a mutation that occurs earlier in the evolution of the cancer should have a higher frequency than mutations that occur later along the same lineage because a child clone necessarily contains all of the mutations belonging to its parent clone. This investigation is hampered, however, by copy number variation. In practice, we cannot directly compare the allele frequencies of two distal sites because the observed allele frequencies are actually the product of mutation frequency and copy number. Empirically, we observe variation in copy number along the genome and differences in copy number variation from one subsection to the next (Figure S2 in Text S1). A consistent duplication of a large portion of chromosome 8 is known to occur commonly in breast cancer [33]. We were lucky, however, that two of our mutated loci occur quite close to one another on chromosome 9 (chr9a and chr9b, separated by only 3.3 Mbp). Given the observed data, the likelihood that a change in copy number occurring between these two loci is small, thereby allowing us to safely compare the corresponding mutation frequencies. Across all nine primary tumor subsections, we observe that the frequency of the parent mutation (chr9b) is higher than that of the child mutation (chr9a). Hence, these mutation frequencies are consistent with the inferred phylogeny.
To assess the stability of our inference, we performed leave-one-out analysis and compared the inferred phylogenies as follows. We held out each of the 12 tumor subsections one at a time and trained the model using the data from only 11 subsections for the case of C = 4. When samples p1-1 or p1-3 were excluded, the inferred genotypes were exactly the same as the genotype obtained from the full data. Excluding any of the other of 10 subsections resulted in a genotype which was different only in one bit; namely, the mutation chr4a was predicted to be present in all clones. However, this difference did not affect the inferred phylogeny because the change of this bit was in fact required to build a valid phylogenetic tree (Figure 4). In other words, by excluding any of the 12 tumor subsections, the inferred genotype always led to the same valid phylogenetic tree, which suggests that our algorithm is stable.
Once a tumor has been resected, clinicians pay a great deal of attention to characterizing its anatomy. Features such as necrosis, extension beyond normal anatomical boundaries, and microvascular invasion convey important prognostic information. In addition, the cancer cells within any given tumor are frequently heterogeneous with respect to features such as differentiation state, the fraction of cells undergoing mitosis (as determined by Ki67 staining), or (for breast cancer) the fraction of cells expressing HER-2 or estrogen receptor. The method described here provides a framework for linking a tumor's molecular anatomy to its structural anatomy as well as its phylogenetic evolution.
Several lines of evidence support the validity of the clonal genotypes and relative frequencies inferred by our model. One prediction from our phylogenetic reconstruction is that somatic variants at the trunk will be present at higher frequencies throughout all tumor subsections than variants appearing at the branches. While copy number variation across the somatic genome complicates these comparisons, one of two closely juxtaposed somatic variants (chr9b) is positioned at the trunk of our phylogenetic tree, while its neighbor (chr9a) arises in one of the branches. Consistent with this representation, the variant allele frequencies for chr9b are consistently higher than for chr9a in all ten tumor subsections examined.
Interestingly, phylogenies can be built from the inferred genotypes even given the relatively low purity of the tumor sections: contamination with normal tissue was in 9 out of 12 subsections in our data (Figure 4, ). In particular, although we estimate that the metastatic subsections contained tumor cells in M1-1 and in M1-2, the corresponding branch of the phylogenies is stable and consistent.
Similar to phylogenetic analysis, reassembly of the tumor subsections indicates that our assignment of mutations to clones produces spatial representations that are anatomically reasonable. With further refinements, our method should enable reconstructions that layer a tumor's phylogeny on top of its spatial organization.
While our results underscore the potential power of this new method, our study also has several limitations. Our assessments were confined to heterozygous somatic variants, and did not take into account the many chromosomal structural changes that were present in the tumor we examined. A comparison of exome copy numbers between primary tumor and lymph node indicates that the vast majority of these chromosomal changes preceded the divergence shown in our phylogenetic tree (Figure S2 in Text S1). In theory, one could imagine generalizing our generative model to take copy number variations into account by replacing the 2 in the denominator of Equation 2 with a hidden random variable for each locus, but without some form of aggressive regularization, this formulation would lead to a prohibitively complex and overfit model.
Additionally, a key characteristic of our method is the requirement to specify the number of clones prior to the EM inference procedure. It is important to recognize that this choice should depend upon properties of the data set itself, rather than fundamental properties of the cancer. After all, each cell division results in multiple mutations, such that every cancer cell constitutes a distinct clone. Consequently, a picture of the full clonal history of a cancer would consist of a phylogenetic tree with one leaf for each cancer cell. In practice, such a tree would be of limited utility and, more importantly, could not be accurately estimated from any reasonably sized data set. Perhaps the most useful definition of a tumor clone is a population of cells that exhibit distinct spatial or functional properties. Our approach allows the user to specify the number of clones and, hence, the resolution at which the clonal history is viewed.
Because Clomial does not impose any assumption on the distribution of mutation frequencies, the number of inferred clones may not exceed the number of samples; otherwise, the resulting optimization problem will be under-constrained.
In the particular cancer studied here, the three-clone solution appears to provide an inaccurate view of the clonal history. The placement of the chr17c mutation along the path leading to metastatic clone C2 is surprising, given that this particular locus has such low counts for both metastatic subsections (2 counts for subsection M1-1 and 0 counts for M1-2, Table S1 in Text S1). This apparent anomaly can be explained by the small counts associated with chr17c in four out of the 10 primary tumor subsections (3 counts in P1-3, 4 in P1-4, and 21 in each of P2-1 and P2-2). Faced with the choice of what genotype profile to assign to this particular locus, the inference procedure selected a solution in which only two subsections, rather than four, are inconsistent. However, given the flexibility of a 4-clone model, the anomaly is resolved, and chr17c defines a novel clone C2 that occurs in the primary tumor samples and is completely absent from the metastatic samples.
In practice, it may be possible to estimate how many clones the data set can resolve using a method such as the Bayesian Information Criterion (BIC), with a smaller BIC value indicating a better fit to the data [34]–[36]. This approach has been used previously for estimating tumor clonal composition [37], [38]. BIC analysis of our model on simulated data suggests that, on average, the BIC accurately estimates the true number of clones, even in the presence of sequencing noise (Figure S3A–B in Text S1).
We also computed the BIC for models trained on our real breast cancer data (Figure S3C in Text S1) and observed a large decrease in BIC (45%) when increases from 3 to 4, suggesting that the model is too simple to describe the data. However, the subsequent improvements of the BIC are smaller: 29%, 20%, 9%, and 3% respectively, as grows from 4 to 8. In general, one should avoid increasing the complexity of the model when the BIC improvement is small because, in such situations, adding to the number of free parameters of the model can potentially lead to over-fitting [39]–[47]. Note that, as an alternative to a BIC approach, one could instead take an approach motivated by cross-validation, as has been explored in the context of matrix factorization models [48]–[50].
Running the EM algorithm is very fast. In practice, using a 2.40 GHz processor with 2 GB memory, training a single EM instance on the real data set takes a few seconds up to several minutes, depending on the value of the hyperparameter (Figure S4 in Text S1). However, because the optimization problem in the M step is non-convex, many EM instances must be trained from different random initializations to avoid local optima.
We first noted that Clomial achieved good results on simulated data using only 10 random initializations when (Figure 2). Then, to further assess the appropriate number of EM instances to run, we revisited the solutions from all of our 100,000 EM instances, counting how many instances are required to achieve the best observed model (Figure S5 in Text S1). In practice, while 1000 EM instances is sufficient to find the optimum solution when or 3, a larger number of random initializations is required as the number of clones grows. This is an expected phenomenon because the complexity of the model grows significantly with , resulting in an optimization surface with many more local optima. Consequently, despite the highly parallel nature of the computation, scaling up to analysis of larger data set with larger numbers of clones will likely require improved EM training strategies, such as noise injection or regularization.
Finally, although we used a simple phylogenetic tree construction procedure to evaluate the quality of our inferred clonal genotypes, the EM inference procedure described here does not explicitly model tumor evolution. Ultimately, we aim to produce a model that automatically infers not only clonal genotypes and clonal frequencies, but also the number of clones and the phylogenetic tree relating them.
Our method differs significantly from other approaches. A recent characterization of 21 breast cancers defined clones by clustering mutations with similar variant allele frequencies [9]. The success of this strategy hinges on characterizing the frequencies of large numbers (hundreds or thousands) of somatic variants. In contrast, our method can reconstruct clonal phylogenies based on accurately measuring alleles of much smaller numbers of somatic variants. The view afforded by our method may provide novel insights into tumor biology. In particular, results from Nik-Zainal and colleagues [9] were interpreted to indicate that cancers become clinically apparent only after one of the competing clones has achieved clonal dominance. In contrast to this “winner takes all” hypothesis, our model suggests that some cancers might be more accurately regarded as ecosystems, in which clones may be subject to spatial influences that affect their competitive fitness, or may even collaborate to support tumor growth.
An important difference between our method and many other methods based on clustering [8], [9], [12] is our explicit probabilistic modeling of the random selection of normal and variant alleles during sequencing, according to a binomial distribution. By taking into account not just the relative frequency of the two alleles but the separate counts of normal and variant alleles, our model automatically assigns less importance to a locus with lower coverage, even if the locus yields the same variant allele frequency as a high-coverage locus.
While this manuscript was under review, two methods called PyClone [13] and PhyloSub [51] were published, which do model allele counts using a binomial distribution. These methods attempt to simultaneously infer not only clonal genotypes and frequencies, as Clomial does, but also infer the number of clones and their phylogeny. Furthermore, PyClone and PhyloSub are not limited, as Clomial is, to situations in which the number of inferred clones is less than or equal to the number of available samples. How is this possible? To make these inferences feasible, these clustering methods must make certain distributional assumptions about the data. Specifically, PyClone assumes a Dirichlet Process prior for clone frequencies, where the base distribution is Uniform and the concentration parameter is Gamma distributed with shape and scale parameters equal to 1 and , respectively. PhyloSub extends PyClone by using a tree-structured stick-breaking process [52] to directly account for phylogenetic relationships during the inference. In principle, these assumptions enable PyClone and PhyloSub to infer information about a large number of clones from only a single sample. On the other hand, when multiple samples are available, Clomial can draw accurate inferences without requiring these distributional assumptions. In practice, our comparison showed that Clomial and PhyloSub produce similar results on three previously described chronic lymphocytic leukemia (CLL) cases [53] (Tables S3–S5 in Text S1).
We note that if is the sequencing error rate at locus , then the probability of observing a variant allele at this locus in subsection is estimated by . In principle, sequencing noise could be incorporated into our model by replacing , defined in Equation 2, with in the likelihood and EM algorithm. However, given the robustness of the current method to noise (Figures S6 and S3C in Text S1), we opted to keep our model simple. In future applications, it may be beneficial to model noise in data produced by sequencing technologies that exhibit high error rates () such as PacBio RS [54].
The EM algorithm is not the only option for maximizing the log-likelihood for the observed data. In particular, one could instead treat both and as optimization variables and seek to maximize with respect to and . This would amount to iteratively updating and then updating until convergence, similar to the iterative algorithms typically used for matrix factorization models [23]–[25], [50]. However, this alternative approach would not have any computational advantage in terms of the update for , which would still not have a closed-form solution, and would need to be solved using BFGS-B or an equivalent approach. Furthermore, the update for would be very complicated under the constraint that is a binary matrix. Therefore, we developed a customized inference algorithm based on EM.
Whereas genetic testing for cancer patients today focuses on mutations affecting a relatively small number of cancer-associated genes, most cancers are sustained by networks of aberrantly regulated genes that collaborate to promote tumor growth. The ability to assign mutations to clones, and to layer a tumor's clonal content on top of its structural anatomy in space and over time, can provide new insights into the mechanisms that enable cancers to invade, metastasize and escape treatment.
This research was reviewed and approved by the Cancer Consortium Institutional Review Board (IRB) located at the Fred Hutchinson Cancer Research Center (FHCRC). The FHCRC has an approved Federalwide Assurance on file with the Office for Human Research Protections (number 00001920). The Federalwide Assurance is a formal written, binding commitment that assures that the FHCRC promises to comply with the regulations and ethical guidelines governing research with human subjects, as stipulated by the U.S. Department of Health and Human Services under 45 CFR 46. Because this study involved the use of de-identified specimens obtained from an IRB-approved repository, we did not interface with patients. Patient consent was administered, in compliance with 45 CFR 46, by investigators who maintain the repository. Patients gave their consent for their specimens to be stored in the repository and subsequently used for research in cancer. The FHCRC IRB deemed that our research was in concordance with the purpose of the registry and the patient informed consent.
We obtained breast cancer tissues from the Breast Cancer Biospecimen Repository of Fred Hutchinson Cancer Research Center after IRB approval. The patient was a 44 year old pre-menopausal woman diagnosed with infiltrative ductal carcinoma (IDC) and ductal carcinoma in situ (DCIS), stage pT1c pN1, Grade II/III, ER positive, PR positive and Her2 negative. Axillary lymph node dissection revealed that one out of 13 nodes was positive for metastatic disease. A total of 5 pieces were obtained from surgical samples including 1 tissue section from adjacent normal breast tissue (N1), 3 tissue sections from the primary breast cancer (P1, P2, P3), and 1 tissue section from the positive axillary lymph node (M1). Each section is about 1 cm by 1 cm by 0.5 cm. The tumor content, including both IDC and DCIS, ranges from 40% to 55% in the primary tumor and axillary lymph node tissue sections based on pathological examination (P1 55% IDC, P2 45% IDC, P3 40% IDC and 15% DCIS, M1 50% IDC).
Each individual section was subdivided into multiple subsections, and the anatomic locations of all the subsections were recorded (Figure 3). Using Qiagen AllPrep DNA/RNA Micro Kit, DNA was extracted from one normal subsection (N1-1), seven primary subsections (P1-2, P1-3, P1-5, P2-1, P2-3, P3-3, P3-4) and one metastatic subsection (M1-1). After quantification, all the DNA samples were subjected to exome capture followed by Illumina sequencing.
Next generation sequencing was carried out at the Northwest Genome Center at University of Washington on the normal subsection, seven primary subsections, and one metastatic subsection. For each subsection, one microgram of genomic DNA was used to construct the random-shearing library per standard protocol with Covaris acoustic sonication. Libraries then underwent exome capture using the Mb target from Roche/Nimblegen SeqCap EZ v2.0 ( exons and flanking sequence). Since each library was uniquely barcoded, samples were performed in multiplex. Massively parallel sequencing was carried out on the HiSeq sequencer.
Sequence reads were processed with a pipeline consisting of the following elements: (1) base calls generated in real-time on the HiSeq instrument (RTA 1.12.4.2); (2) Perl scripts developed in-house to produce demultiplexed fastq files by lane and index sequence; (3) demultiplexed BAM files aligned to a human reference (hg19) using BWA (Burrows-Wheeler Aligner; v0.5.9) [55]. Read-pairs not mapping within standard deviations of the average library size ( bp for exomes) are removed. All aligned read data were subjected to the following steps: (1) “duplicate removal” was performed, (i.e., the removal of reads with duplicate start positions; Picard MarkDuplicates; v1.14); (2) indel realignment was performed (GATK IndelRealigner; v1.0-6125) resulting in improved base placement and lower false variant calls; (3) base qualities were recalibrated (GATK TableRecalibration; v1.0-6125). All sequence data then underwent a previously described quality control protocol [56].
Variant detection and genotyping were performed using the UnifiedGenotyper tool from GATK (v1.0-6125). Variant data for each sample were formatted (variant call format) as “raw” calls that contain individual genotype data for one or multiple samples, and flagged using the filtration walker (GATK) to mark sites that are of lower quality/false positives, e.g., low quality scores (), allelic imbalance (), long homopolymer runs () and/or low quality by depth (QD ).
Most of the commonly used software for calling SNVs and indels, including SNVMix [57] and VarScan [58], requires tumor content . To allow identification of low frequency alleles that occur in only one or a few subsections, we did not pool all of the data together. Instead, we designed a method that is appropriate for multiple samples from one patient, with relatively low tumor content, ranging from 45% to 55%. At each chromosomal position (locus), we considered six mutually exclusive possible outcomes: A, C, G, T, deletion, and unknown. The counts of these six outcomes at each locus between normal and each of the multiple tumor subsections were compared with a Fisher's exact test. To correct for multiple testing, we used the qvalue R package to convert to . Only those chromosomal loci with in at least one comparison between normal and tumor samples were accepted for downstream analysis. This analysis identified 6310 loci.
For each accepted locus, we used a heuristic procedure to identify which of the six alleles differed between the tumor and normal sample. For each subsection, we carried out six Fisher's exact tests, one for each of the six possible alleles. Thus, each such test compared one allele's counts to the sum of the counts for the other five alleles. Using a p-value threshold of 0.01, an allele was declared to be increased, decreased, or unchanged in the tumor subsection as compared to the normal sample. The changes that were classified as “increased” and had a normal count of zero were called tumor-specific mutations. This procedure identified a total of 268 such tumor-specific mutations, with a mean and median sequencing depth of 92 and 75, respectively. Corresponding annotations were obtained from SeattleSeq (http://snp.gs.washington.edu/SeattleSeqAnnotation137).
In parallel, we also analyzed our data using deepSNV [31] by comparing the normal subsection to the 8 tumor subsections. We ran deepSNV on the loci with total coverage across all samples more than 50, which resulted in the identification of 29 loci with . The union of the two lists yielded 281 loci for further validation (Dataset S1).
Mutations were validated by targeted deep sequencing of DNA derived from one normal subsection (N1-1), 10 primary subsections (P1-1, P1-2, P1-3, P1-4, P2-1, P2-2, P2-3, P2-4, P3-1, P3-2) and two metastatic subsections (M1-1 and M1-2). The subsections were selected to have low normal content and to span the tumor anatomy. Genomic DNA was prepared as described for the initial exome sequencing. A HaloPlex probe capture library for selective capture of 281 target loci was generated with SureDesign (Agilent Technologies). Target enrichment for deep sequencing was carried out with the HaloPlexTM Target Enrichment System from Agilent Technologies following the manufacturer's protocol. Triplicate enrichments were performed for each sample. Target-enriched samples were sequenced using a MiSeq (Illumina). Of the 281 target loci, 244 were successfully sequenced with coverage more than 100 reads for the normal sample. The mean, median, and the standard deviation of the coverage were 1615, 1118, and 1600, respectively (Dataset S2).
All 244 loci were visualized using the Integrative Genomics Viewer [59], [60]. A set of 17 loci were selected based upon three criteria: (1) at least 3 reads cover the locus in the normal sample, (2) the variant allele is not present in the normal tissue (allowing for a few variant counts, which may reflect sequencing error) and (3) there are no nearby clustered mutations, indicative of sequencing or mapping error. Independently, the data were also analyzed using deepSNV. Applying a threshold of yielded 19 loci, including all 17 of the initially selected loci. The 17 loci were retained for downstream analysis (Table S1 in Text S1).
We computed BIC using the following formula:(14)where is the expectation of the complete-data log likelihood, which is maximized in the last M step (see Equations 7 and 13). Also, represents the total number of free parameters, and is the total number of counts.
|
10.1371/journal.pgen.1007012 | Genetic anticipation in Swedish Lynch syndrome families | Among hereditary colorectal cancer predisposing syndromes, Lynch syndrome (LS) caused by mutations in DNA mismatch repair genes MLH1, MSH2, MSH6 or PMS2 is the most common. Patients with LS have an increased risk of early onset colon and endometrial cancer, but also other tumors that generally have an earlier onset compared to the general population. However, age at first primary cancer varies within families and genetic anticipation, i.e. decreasing age at onset in successive generations, has been suggested in LS. Anticipation is a well-known phenomenon in e.g neurodegenerative diseases and several reports have studied anticipation in heritable cancer. The purpose of this study is to determine whether anticipation can be shown in a nationwide cohort of Swedish LS families referred to the regional departments of clinical genetics in Lund, Stockholm, Linköping, Uppsala and Umeå between the years 1990–2013. We analyzed a homogenous group of mutation carriers, utilizing information from both affected and non-affected family members. In total, 239 families with a mismatch repair gene mutation (96 MLH1 families, 90 MSH2 families including one family with an EPCAM–MSH2 deletion, 39 MSH6 families, 12 PMS2 families, and 2 MLH1+PMS2 families) comprising 1028 at-risk carriers were identified among the Swedish LS families, of which 1003 mutation carriers had available follow-up information and could be included in the study. Using a normal random effects model (NREM) we estimate a 2.1 year decrease in age of diagnosis per generation. An alternative analysis using a mixed-effects Cox proportional hazards model (COX-R) estimates a hazard ratio of exp(0.171), or about 1.19, for age of diagnosis between consecutive generations. LS-associated gene-specific anticipation effects are evident for MSH2 (2.6 years/generation for NREM and hazard ratio of 1.33 for COX-R) and PMS2 (7.3 years/generation and hazard ratio of 1.86). The estimated anticipation effects for MLH1 and MSH6 are smaller.
| Genetic anticipation is a phenomenon where symptoms of a hereditary disease appear at an earlier age and/or are more severe in successive generations. In genetic disorders such as Fragile X syndrome, Myotonic dystrophy type 1 and Huntington disease, anticipation is caused by the expansion of unstable trinucleotide repeats during meiosis. Anticipation is also reported to occur in some hereditary cancers though the underlying mechanism behind this observation is unknown. Several studies have investigated anticipation in Lynch syndrome, the most common hereditary colorectal cancer syndrome, yet there is a debate concerning whether anticipation occurs and what underlying mechanism there is. The objective of this project is to study if anticipation is part of the clinical picture in Swedish families with LS, with the long term goal to enable better prediction of age at onset in family members. Our results suggest that anticipation occurs in families with mutation in MSH2 and PMS2, while the evidence is equivocal for MLH1 and MSH6.
| Lynch syndrome (LS) is an autosomal dominant inherited syndrome that increases the risk of cancer, primarily in the colon, the rectum and the endometrial lining of the uterus, and to a lesser degree also in the stomach, the ovary, the hepatobiliary tract, the urinary tract, the small bowel and the brain [1,2]. LS is one of the most common heritable cancer syndromes, accounting for up to 4% of the total colorectal cancer burden in Europe, where patients have up to 70% lifetime risk of developing colorectal or endometrial cancer at an early age [1]. LS was formerly known as hereditary non-polyposis colorectal cancer (HNPCC), but when clinical criteria evolved to take into account not only colorectal cancer to identify families with LS [3,4] the name Lynch Syndrome became generally accepted [5]. Today the diagnosis LS is restricted to families with a known pathogenic germline mutation in one of the mismatch repair (MMR) genes MLH1, MSH2, MSH6 and PMS2 irrespective of family history [6,7]. The MMR system corrects indels or mismatches in the DNA, and is evolutionary conserved from bacteria to human [8]. In human the recognition of nucleotide mismatches is mediated by the protein heterodimers MSH2/MSH6 or MSH2/MSH3, while the removal and resynthesis of nucleotides is mediated by MLH1/PMS2 [9].
LS is a heterogeneous disease with regard to tumor spectrum and age at onset [10]. Part of this phenotypic variation has been linked to specific MMR gene mutation. For instance, MLH1 mutation carriers are suggested to have a higher risk for colorectal cancer (CRC) and earlier age of onset, compared to MSH2 and MSH6 mutation carriers [11–15]. In general, MSH6 mutation carriers tend to have a later age of onset and lower penetrance for LS associated tumors, apart from endometrial cancer, compared to MLH1 and MSH2 mutation carriers [16–20]. An older age of onset and a lower overall risk for CRC has also been suggested for PMS2 mutation carriers [21,22]. However, LS display phenotypic variation in age of onset also within families and between families with the same mutation [23–25]. This variation is attributed to individual genetic differences modifying the effect of an inherited MMR mutation [26–31]. Another factor proposed to influence age at onset is genetic anticipation, defined as progressive earlier onset and severity of disease in successive generations within a family. This phenomenon is closely related to the disease mechanisms in several genetic disorders, e.g the neurodegenerative diseases Fragile X syndrome, Myotonic dystrophy type 1 and Huntington disease, in which trinucleotide repeat expansion directly influence expressivity and penetrance of disease [32]. Anticipation has also been observed in hereditary cancer for example familial melanoma, Li-Fraumeni syndrome, breast, ovarian and pancreatic cancer, and recently in the renal cell cancer syndromes von Hippel-Lindau and HLRCC (hereditary leiomyomatosis and renal cell cancer) [33–39]. In LS, a progressive decrease of age at CRC onset was proposed already in 1925 when the syndrome was first described [40,41]. However, it is complicated to estimate genetic anticipation and there are contradictory reports regarding its existence in LS, though the majority indicate anticipation [42–52]. Previous studies have applied various statistical methodologies, compiled different mutations and included subjects with LS associated mutations as well as subjects with only a clinical diagnosis. In light of these studies, we analyzed affected and unaffected mutation carriers in LS families throughout Sweden, to investigate signs of anticipation using two regression models with adjustment for potential confounders, including gene-specific effects.
In Sweden, families with suspected LS are referred to the regional department of Clinical genetics in Umeå, Uppsala, Stockholm, Linköping, Göteborg or Lund, for counceling and genetic testing. Out of this population-based cohort, families identified with a LS-associated MMR mutation that received genetic counseling in Lund, Stockholm, Linköping, Uppsala or Umeå between January 1990 and December 2013 were enrolled in this study. This project was approved in accordance with the Swedish legislation of ethical permission (2003:469). All patients provided oral or written informed consent for genetic diagnostics as part of their routine clinical care. This anonymized genetic information may be used for research without further consent sought from the patients if approved by an ethical review board. Accordingly, this study was approved by the Regional ethical review board in Stockholm (dnr 2014/1320-31).
Patient and family cancer history was reported at the time of genetic counseling and cancer diagnoses were further confirmed from medical records or pathology reports. A total of 239 families with proven pathogenic MMR variants described in [53] (96 MLH1 families, 90 MSH2 families including one EPCAM-deletion family, 39 MSH6 families, 12 PMS2 families, and 2 MLH1+PMS2 families), comprising 1029 mutation carriers, were identified in the cohort. One individual whose parents were both mutation carriers was excluded. Additionally, the sex of 11 carriers was unknown, and the age at diagnosis for an additional 14 was missing. We excluded these 25 individuals, leaving 1003 at-risk carriers with available follow-up information and sufficient pathological and medical information to be included in the study. We grouped the EPCAM-deletion family within the MSH2 families, as it is reported that a partly deleted EPCAM gene (located upstream of MSH2) cause LS through reducing the expression of MSH2 in EPCAM-expressing tissues [54]. For statistical modeling purposes, we counted two families with mutations in both MLH1 and PMS2 as having mutations in PMS2 only (unreported auxiliary analyses that excluded these families altogether or counted them as MLH1 showed that our findings are not sensitive to this decision). The follow-up period was defined as the time from birth until age at onset, and for individuals who were diagnosed with multiple Lynch-related cancers, age of onset was recorded as the time of first diagnosis. Our first analytic approach was the normal random effects model (NREM) proposed by Larsen et al. [45], which has been used previously to test for anticipation in LS [43] and BRCA-mutation related cancers [55]. Let ni denote the number of carriers in the ith family, i = 1, 2, …, 239, and let j = 1, 2, …, ni index the jth individual in family i. The NREM is given by
Tij=μi+γZij+βXij+εij,
(1)
where Tij is the age of diagnosis in years for the jth member of family i (“person ij”), μi is the family-specific random intercept representing a typical age of diagnosis in the ith family, Zij is person ij’s generation (coded with respect to oldest observed generation in each family, as described in [41] and γ, the parameter of interest, is the mean change in age of diagnosis between consecutive generations, i.e. the anticipation effect. In the NREM, anticipation is indicated if γ < 0. Collectively the Xij term represents any other covariate(s) of interest for person ij, the effect of which is given by β. The final term εij is the residual error, assumed to be independently and normally distributed with mean zero and variance σ2. For each person who was not diagnosed with a Lynch-associated cancer during the follow-up period, the likelihood contribution is given by the normal survivor function, that is, the probability of being cancer-free at the age of last follow-up. We assume that the censoring mechanism is independent of the time to cancer diagnosis.
Our second analytic approach, which is also a regression strategy, extends the Cox proportional hazards model that was used in [56] to test for anticipation in lymphoproliferative tumors. Person ij’s hazard for cancer diagnosis at age t is modeled as:
λ(t|Zij,Xij)=λ0(t)exp(μ˜i+γ˜Zij+β˜Xij).
(2)
The function λ0(t) is the overall baseline hazard function. In Daugherty et al., the baseline hazard was assumed to be identical for all families, that is, μ˜i was not included in the model and within-family correlations were accounted for by robust sandwich-type covariance estimates. Here we add a random family-level effect μ˜i, similar to NREM, which makes the less restrictive assumption that all families’ baseline hazards are proportional to, rather than equal to, one another. We call this Cox model with family-level random effect COX-R. The remaining parameters are analogous to NREM. Specifically, γ˜ gives the generational effect of anticipation as a log-hazard ratio, with γ˜>0 indicating anticipation, and β˜ is the log-hazard ratio(s) for all other covariates.
In addition to adjusting for sex, we also included mutational status in NREM and COX-R. In Eqs (1) and (2), let person ij’s length-4 vector of covariates be given by Xij = {1[sexij = male], 1[genei = MSH2], 1[genei = MSH6], 1[genei = PMS2]}, where 1[y] is the indicator function, equal to 1 if y is true, sexij is the sex of person ij and genei is the mutational status of family i. MLH1 serves as the reference category.
We also investigated whether there were gene-specific effects of anticipation by substituting Zij in Eqs (1) and (2) with the four dimensional covariate vector.
All analysis was done in the R software package R Core Team [57]. Code for maximizing the integrated partial likelihood of model (2), marginalized over the random effects μ˜i, is provided in the R package COXME [58].
Table 1 presents the clinical characteristics of our data and Fig 1 plots the Kaplan-Meier estimate of the time to first Lynch-associated cancer diagnosis, to give an overview of the age at onset in our cohort. During the follow-up period, 719 carriers were diagnosed with at least one Lynch-associated cancer and 171 were diagnosed with multiple Lynch-associated cancers. Overall, the median age of first diagnosis was 51 years (95% CI: 50–53), but this varied with mutational status, being 49 years in both MLH1 and MSH2 patients and 58 and 67 years, respectively, for MSH6 and PMS2 patients.
Based on this, in addition to adjusting for sex, we also included mutational status in NREM and COX-R analyses. Table 2 gives the estimates, standard errors, and Wald-type p-values for the anticipation effects only, and Table 3 provides all parameter estimates. As shown in Table 2, the estimates of γ (NREM) and γ˜ (COX-R) are -2.1 (p = 0.0001) and 0.171 (p = 0.0013), respectively. Both suggest the presence of anticipation: a 2.1 year decrease in the age of diagnosis per generation and a hazard ratio of exp(0.171) = 1.19 between consecutive generations. Because mutational status appears to confound the age of diagnosis, we also investigated whether there were gene-specific effects of anticipation, yielding one estimated anticipation effect for each analyzed MMR-gene. These are given in the bottom rows of Table 2. In NREM, anticipation is estimated to be -1.8 (p = 0.044), -2.6 (p = 0.003), -1.1 (p = 0.366), and -7.3 (p = 0.014) years per generation respectively, for MLH1, MSH2, MSH6, and PMS2. In COX-R, the corresponding log-hazard ratios are 0.127 (p = 0.133), 0.284 (p = 0.001), -0.005 (p = 0.965), and 0.618 (p = 0.052), representing hazard ratios of 1.13, 1.33, 0.99, and 1.86, respectively. In both models, the confidence intervals (CIs) for the anticipation effects of MSH2 and PMS2 lie far from their null values (Table 2), whereas there is greater uncertainty with regard to any possible effect of anticipation in MLH1 and MSH6.
We investigated signs of anticipation in LS through the analysis of a large, Swedish population-based cohort and regression analyses suggest that anticipation exists in these families. The NREM analysis suggests that the age of diagnosis in families is decreasing by about 2 years per generation, and the COX-R analysis suggests a multiplicative increase in the rate of diagnosis of about 1.19 between generations. These regression analyses carry at least two important advantages over hypothesis testing approaches that compare the age of diagnosis between all parent-child pairs. First, they make use of the partial follow-up time from all at-risk carriers who have not yet been diagnosed; these individuals would otherwise be excluded from analysis. Second, the model-based structure allows for straightforward incorporation of genetic effects or other possible confounders. The underlying causal mutation evidently plays a role in the extent of anticipation, as our estimates varied between MMR genes. Among the MMR genes, the ordering of estimated anticipation effects was PMS2, MSH2, MLH1, and MSH6, with the largest effect in PMS2 (7.3 years/generation [NREM] or a hazard ratio of exp(0.618) = 1.86 [COX-R]). Although the small number of PMS2 families yielded correspondingly large uncertainty, these effects were still highly significant therefore this uncertainty does not invalidate the findings. For MSH2, the estimated effect of anticipation was 2.6 years/generation or a hazard ratio of exp(0.284) = 1.33.
The results are comparable to those reported in several earlier studies (for a review, see [41]). In an analysis of Lynch families from the Danish HNPCC Registry [45], an anticipation effect of about three years/generation was reported but no differences between mutational status was found. A version of the same data was considered again in Boonstra, et al. [43], who fit variants of both regression models considered here, reporting effects of 3.3 years/generation and hazard ratios of exp(0.22) = 1.25. Neither model in that study adjusted for mutational status. Also, the Cox model did not include family-level random effects, as we do here; our approach is arguably a more accurate, although still simplified, reproduction of the true underlying hazard process. A later report analyzed the same Danish HNPCC Registry data with Bayesian modeling techniques [59], allowing anticipation to be random between families. They estimated population-level gene-specific effects of anticipation, as performed here, for MLH1, MSH2, and MSH6. They found respective anticipation effects of 2.8, 2.5, and 1.0 years, consistent with our findings. Several other studies based on anecdotal observation or analyses of affected parent-child pairs have found effects of anticipation varying between 5.5 and 10 years [44,47,49,50]. A notable exception from previous studies is Tsai, et al. [46] who found no evidence for anticipation in 475 parent-child pairs from the Johns Hopkins Hereditary Colorectal Cancer Registry; in part this may be explained by differences in eligibility as only 14 of the 475 parent-child pairs analyzed had verified disease-predispoing germline MMR gene mutations.
The underlying mechanism for anticipation in heritable cancer is still unknown. However, it has been proposed that anticipation is caused by a progressive accumulation of germline mutations, due to the reduced DNA mismatch repair ability in mutation carriers [51]. Accordingly, haploid/monoallelic mutations in the MMR system affect the mutation load in the carrier prior to loss of the second allele, and accumulated alterations in the germ cells is transferred to the offspring [41]. Interestingly, there is an overrepresentation of mononucleotide repeats within and around the human MMR genes compared to other genomic regions, with an overrepresentation in the PMS2 gene [60,61]. It has been suggested that MMR proteins maintain the length of such microsatellites present within their own nucleotide sequences by an evolutionary mechanism operating by gene-protein interactions [60]. With the above arguments a deficient MMR system would propagate errors through generations and this would be most significant for mutations in the PMS2 gene, which is in accordance with our results. In addition, it has been shown that PMS2-deficient mice eggs forms embryos with an increased mononucleotide mutation rate, indicating that MMR mutations might affect germline mutation rate in a heterozygous state [62]. This also points to our results that PMS2 mutations carriers would display the most anticipation if the mutation load is inherited by the next generation.
Noteworthy, PMS2 and MSH2 are not part of the same protein complex involved in recognition, excision and resynthesis of mismatched nucleotides [63], nor does the MSH2 gene contain the same magnitude of mononucleotide repeats as PMS2 [60]. This together suggests a different underlying mechanism generating anticipation in MSH2 mutation carriers. For example, it is shown that MMR deficiency affect telomere shortening in human fibroblasts, and that this might influence heterozygous carriers of a MSH2 mutation in particular [64]. Moreover, in a recent study telomere shortening correlated significantly with age at onset in the MSH2 carriers, whereas the MLH1 carriers displayed longer telomeres and delayed age at onset [65]. Nevertheless, MMR mutation carriers with LS-associated cancer may have specific telomere-length dynamics but telomere shortening does not alone explain anticipation, as reported by Segui et al [66], indicating that gene-specific dynamics and different mechanisms are involved.
Despite a general concurrence with earlier studies, several caveats accompany our findings. Evidently, our study and previously published evidence that performed survival analysis for genetic anticipation in LS suggests that if genetic anticipation does exist, the effect is modest [42,43,45]. This makes anticipation a difficult problem statistically and challenges some of the clinical utility of our findings. At the population level, anticipation may well also be due to reasons other than genetic. For example, cohort effects arising from changes in treatment, diagnostic or environmental factors can also result in a decline in age at diagnosis. These effects should be visible both within family trees and in the entire population (which is a mix of mutation carriers from different generations). This is in contrast to genetic anticipation, which would only be seen within each unique family tree.
Voskuil, et al. found that the hazard ratio corresponding to generation decreased considerably in magnitude after adjusting for birth cohort [42], although their final estimated hazard ratios for the effect of generation were still very close to our estimate of 1.2. Statistically, birth cohort and generation are typically highly correlated, which can cause the resulting parameter estimates to be unstable. Boonstra, et al. [59] attempted to disentangle these effects by independently estimating secular trends in age of colorectal cancer diagnosis from a cancer registry of all colorectal cancers, and adjusting the Danish HNPCC Registry data for this estimated trend before analysis. Still, the results indicated as reported earlier in this section, population-level gene-specific effects of anticipation of 2.8, 2.5, and 1.0 years, respectively, for MLH1, MSH2, and MSH6. Our estimated effects of anticipation decrease by about 0.7 years when we directly apply the secular trends estimated in Boonstra, et al. [54].
Furthermore, it has been argued that anticipation may be falsely detected due to fecundity bias [48]. Through repeated simulations of parent-child pairs in which no anticipation exists (in truth) but the fertility rate was positively correlated with age of diagnosis, Stupart, et al. demonstrated in a particular scenario that an apparent anticipation effect of about 1.8 years can manifest. Noteworthy, the greatest reduction in fertility was predominantly among those diagnosed before age 29, affecting the fertility of the cohort as a whole. In our cohort, the Kaplan-Meier estimated proportion of patients free of diagnosis at age 29 was 96.5%, which suggests that fecundity bias due to these patients is likely to be small.
Taken together, our findings are in line with those of previous studies. That being said, the study of genetic anticipation is both complex and statistically challenging. The ideal setting in the continuing assessment of fine variations in LS phenotype, such as anticipation, would be prospective, population-based datasets, together with state-of-the-art statistical methods. Still, a number of promising findings have been reported previously, yet often the statistical methods or small sample sizes have been limiting. We believe that the analyses performed in our study properly consider familial, genetic, and clinical parameters and therefore give a representative measurement of anticipation in Lynch Syndrome.
|
10.1371/journal.pgen.1003051 | Transcription Elongation and Tissue-Specific Somatic CAG Instability | The expansion of CAG/CTG repeats is responsible for many diseases, including Huntington's disease (HD) and myotonic dystrophy 1. CAG/CTG expansions are unstable in selective somatic tissues, which accelerates disease progression. The mechanisms underlying repeat instability are complex, and it remains unclear whether chromatin structure and/or transcription contribute to somatic CAG/CTG instability in vivo. To address these issues, we investigated the relationship between CAG instability, chromatin structure, and transcription at the HD locus using the R6/1 and R6/2 HD transgenic mouse lines. These mice express a similar transgene, albeit integrated at a different site, and recapitulate HD tissue-specific instability. We show that instability rates are increased in R6/2 tissues as compared to R6/1 matched-samples. High transgene expression levels and chromatin accessibility correlated with the increased CAG instability of R6/2 mice. Transgene mRNA and H3K4 trimethylation at the HD locus were increased, whereas H3K9 dimethylation was reduced in R6/2 tissues relative to R6/1 matched-tissues. However, the levels of transgene expression and these specific histone marks were similar in the striatum and cerebellum, two tissues showing very different CAG instability levels, irrespective of mouse line. Interestingly, the levels of elongating RNA Pol II at the HD locus, but not the initiating form of RNA Pol II, were tissue-specific and correlated with CAG instability levels. Similarly, H3K36 trimethylation, a mark associated with transcription elongation, was specifically increased at the HD locus in the striatum and not in the cerebellum. Together, our data support the view that transcription modulates somatic CAG instability in vivo. More specifically, our results suggest for the first time that transcription elongation is regulated in a tissue-dependent manner, contributing to tissue-selective CAG instability.
| Several dominant genetic diseases, including Huntington's disease (HD) and myotonic dystrophy 1, are caused by the expansion of CAG/CTG repeats. These repeats are unstable in selective tissues. Repeat instability, resulting in the production of increasingly toxic mutant entities in the affected tissues, is proposed to accelerate disease progression. It is therefore essential to unravel the mechanisms contributing to tissue-selective somatic instability. In vitro and cell-based studies indicate that transcription is involved in CAG/CTG instability. However, the role of transcription in CAG/CTG instability in vivo has remained controversial, since mRNA tissue levels of CAG/CTG repeat-containing genes do not correlate with the tissue-specific pattern of instability. Moreover, it is unclear whether the transcriptional process would contribute per se to CAG/CTG instability or whether it is the increased chromatin accessibility associated with transcription that would promote instability. We addressed these issues using two HD transgenic mouse lines recapitulating HD tissue-specific instability. Our in vivo data indicate that high chromatin accessibility and transgene expression do not underlie tissue-selective CAG instability in HD, but suggest that the dynamics of transcription elongation is one mechanism contributing to this process.
| The expansion of trinucleotide repeats (TNRs) is the causative mutation of more than fifteen neurodegenerative, neurological and neuromuscular genetic diseases, including Huntington's disease (HD), several dominant spinocerebellar ataxias (SCAs), myotonic dystrophy 1 (DM1), Friedreich ataxia (FRDA) and fragile X syndrome (FXS) [1]. TNR expansions become unstable and toxic above a threshold of 30 to 50 repeat units. TNR instability leads to repeat size variation, often biased towards further expansion, in the germline and in somatic tissues over time [2]–[4]. Remarkably, the toxicity and the instability of TNRs are tissue-selective in a disease-specific manner. In some diseases, including HD, DM1 and FRDA, the affected and ‘unstable’ tissues partially coincide, thereby accelerating disease progression [1], [5]–[7]. In HD, somatic instability of CAG repeats is most important in the striatum, a tissue that selectively degenerates [4], [6]. In DM1, somatic CTG instability is also most extensive in affected tissues, including the muscle and some brain regions [8]–[10]. Similarly, somatic GAA instability in FRDA is important in dorsal root ganglia and cerebellum, two tissues primarily affected in the disease [11], [12]. Thus, it is essential to decipher the mechanisms underlying tissue-selective instability of TNRs to better define the relationship between toxicity and instability.
CAG/CTG repeats as well as other TNRs form stable DNA secondary structures in vitro, and several studies support the view that the error-prone repair of these structures by overwhelmed DNA repair machineries contributes to TNR instability [1], [13]–[19]. Studies based on cell models also suggest that transcription is involved in CAG/CTG instability [20], [21] and formation of stable DNA:RNA hybrids (R-loops) at repeat sites might contribute to the process [22]–[24]. Furthermore, all TNR loci associated with diseases are located in transcribed regions of the genome, supporting the idea that transcription and TNR instability are linked. However, the role of transcription in TNR instability in vivo remains unclear [1]. First, if transcription contributes to TNR instability, one would expect that transcriptional activity at TNR genes correlates with tissue-selective repeat instability, which seems to be contradicted by previous studies [25]–[27]. For instance, Huntingtin (HTT) mRNA in the striatum of HD patients or rodent models was not increased when compared to tissues exhibiting minimal CAG instability, such as the cerebellum [25], [26]. Similarly, somatic CTG instability in selective tissues did not correlate with increased transcription levels in DM1 mice [27]. Second, it is unclear whether the transcriptional process would contribute per se to TNR instability or whether transcription would only facilitate access of transacting-factors such as DNA repair proteins, due to increased accessibility of chromatin at transcribed regions [28]. In the latter hypothesis, chromatin structure rather than transcriptional activity is expected to influence TNR instability.
It has been shown that TNRs, together with specific cis-elements, have the capacity to modify the chromatin environment by inducing a heterochromatinization process, eventually leading to transcriptional silencing such as in FRDA and FXS, which are caused by GAA and CGG expansions, respectively [29]–[31]. Thus, heterochromatin rather than accessible chromatin has been associated with unstable TNRs, pointing to an apparent paradox.
Here, we have investigated the relationship between chromatin structure, transcription and tissue-selective CAG instability by using R6/1 and R6/2 HD transgenic mouse lines. These two lines have been generated using the same HD transgenic construct, which consists of ≈1000 bp of the human HTT promoter, the entire HTT exon-1, including a CAG repeat expansion, and 262 bp of HTT intron-1 [32]. The transgene is integrated in different genomic loci, recently mapped to chromosome 3 and 4 in R6/1 and R6/2 lines, respectively [33], [34]. R6/1 and R6/2 mice develop a comparable HD-like phenotype. However, progression of the disease is much faster in R6/2 than in R6/1 mice, dying at ≈12 and ≈40 weeks, respectively. Importantly, somatic CAG instability in R6 mice exhibits a similar tissue-selectivity profile as in HD patients, indicating that the mechanisms underlying tissue selectivity are conserved between the two species [35]. Specifically, CAG instability in R6 lines and in humans is most marked in the striatum and minimal in the cerebellum. We previously showed that the level of somatic instability was higher in R6/2 than in R6/1 mice of similar ages, indicating that the instability rate might be increased in R6/2 mice [36]. CAG instability was also higher in R6/2 mice, irrespective of the repeat size (100 or 160 CAG repeats), than in R6/1 mice (130 CAG repeats) [36], suggesting that it is the chromatin context at the site of transgene integration and possibly the expression level of HD transgene, rather than the initial expansion size that underlies the different somatic instability levels.
To test the hypothesis that the chromatin structure and/or transcriptional activity at the HD locus might contribute to somatic CAG instability, we have performed chromatin immunoprecipitation experiments and expression analyses from the striatum and the cerebellum of R6/1 and R6/2 mice. Our results indicate that the HD transgene is integrated in a more euchromatinized and transcriptionally active region in R6/2 mice than in R6/1 mice, supporting the view that an active chromatin structure results in increased CAG instability. However, chromatin accessibility and transcription initiation at the HD locus were similar in the striatum and cerebellum of R6 mice, indicating that tissue-selective CAG instability is unlikely to be simply explained by an accessible or transcriptionally active chromatin state. Strikingly, H3K36 trimethylation and elongating RNA Pol II were increased at the HD locus in the striatum of both R6/1 and R6/2 mice. Together, our in vivo data suggest that transcription elongation mediates tissue-specific somatic CAG instability in HD.
To correlate the chromatin state and the transcriptional activity at the HD locus with the level of CAG instability in R6/1 and R6/2 striatum and cerebellum, we determined CAG instability rates. The level of instability of CAG repeats in mouse striatum and cerebellum was assessed in this study and in previous work [36]. R6/1 mice with 130 repeats were analyzed at 3, 6, 12 and 38 weeks and R6/2 mice with 160 repeats at 3, 6 and 12 weeks (R6/1 and R6/2 mice die around 40 and 12 weeks, respectively). To determine instability rates, we represented the level of CAG instability in R6/1 and R6/2 striatum and cerebellum, as measured by the variation of CAG repeat size, according to age (Figure 1). Interestingly, instability rates over time could be modeled using linear regression analyses (Figure 1). CAG instability rates, as measured by the regression coefficient (slope), ranked as follows: R6/2 striatum>R6/1 striatum>R6/2 cerebellum>R6/1 cerebellum (Figure 1).
We investigated the hypothesis that the chromatin structure at the HD locus might be different in R6/1 and R6/2 tissues, which might underlie the different CAG instability rates. To explore this idea, we performed chromatin immunoprecipitation-PCR (ChIP-PCR) experiments and quantitatively assessed the level of histone marks associated with heterochromatin, euchromatin or transcriptionally active chromatin around the CAG expansion, as previously done for other TNRs associated to diseases [31], [37]–[39]. More specifically, we analyzed the striatum and the cerebellum of R6/1 and R6/2 mice. Tissue samples were sonicated to <500 bp to increase resolution of ChIP experiments (Figure S1). Primer pairs around the CAG tract, located ≈300 bp upstream of the repeats, within HTT proximal promoter (HD promoter), and ≈300 bp downstream of the repeats (HD intron-1), were selected to analyze the HD locus (Figure 2A). As controls, we included primers targeting regions expected to be euchromatic/transcriptionally active (the endogenous murine huntingtin gene Hdh, Hdh promoter and Hdh intron-1, Figure 2A) and heterochromatic/transcriptionally inactive (the ICRH19 locus). We used R6/2 mice carrying 160 CAGs at late-pathological stage (12 week-old) and R6/1 mice carrying 130 CAGs at both early- and late-pathological stages (13 and 36 week-old, respectively).
We first assessed the level of H3K9 dimethylation (Me2H3K9), a hallmark of heterochromatin. Me2H3K9 was 3- to 4-fold more elevated at HD intron-1 in R6/1 striatum and cerebellum when compared to R6/2 tissue-matched samples, regardless of age, and reached levels comparable to those measured at the ICRH19 locus (Figure 2B). Me2H3K9 at the HD promoter was also increased in R6/1 cerebella with respect to R6/2 cerebella, though it was similar in striata (Figure 2B). Globally, Me2H3K9 level at the HD locus, as assessed by calculating the mean level at the HD promoter and HD intron-1, was significantly increased in R6/1 tissues as compared to R6/2 tissues (Figure 2C). The increased level was not dependent upon age. Moreover, the higher level of Me2H3K9 did not result from increased nucleosome occupancy, as measured by the level of total H3 at the HD locus, which was similar in R6/1 and R6/2 tissues (Figure 3A). Noticeably, unmodified H3 was lower at the HD promoter when compared to HD intron-1, consistent with reduced nucleosome occupancy at the region upstream of the transcription start site (TSS), a feature shared by many promoters [40]. We also assessed Me2H3K9 at the HD locus and at a control heterochromatic region (the N-myc locus) in HD and control human lymphoblasts. The level of Me2H3K9 at the HD locus was low as compared to the level at the N-myc locus, and similar in HD and control cells (Figure 2E). In addition, it was comparable to the level measured using R6/2 tissues. Together, these results suggest that the 3′ end of the HD transgene is heterochromatinized in R6/1 mice, due to position effect imposed on the transgene.
We then examined the correlation between somatic CAG instability and Me2H3K9 levels at the HD locus (Figure 2D). The R2 value was only 0.31. This is because Me2H3K9 level was similar in striatum and cerebellum, despite the fact that it was increased in R6/1 tissues relative to R6/2 tissues. Thus, Me2H3K9 level at the HD locus inversely correlates with mouse-line dependent instability, but not with tissue-selective instability.
To further explore the chromatin structure at the HD locus of R6/1 and R6/2 mice, we assessed the level of the euchromatic mark acetyl H3K9 (AcH3K9) (Figure 3B). AcH3K9 was lowest at the ICRH19 locus, which is consistent with heterochromatin. AcH3K9 at HD intron-1 was lower in R6/1 cerebellum with respect to R6/2 cerebellum but not significantly different in R6/1 and R6/2 striatum. AcH3K9 was consistently low at the HD promoter, likely resulting from reduced H3 levels (Figure 3A). No significant difference between tissues and mouse lines was seen for AcH3K9 at the HD promoter (Figure 3B). Globally, AcH3K9 at the HD locus was similar in the different tissues and mice, though it was slightly decreased in R6/1 cerebellum as compared to R6/2 cerebellum, and did not correlate with CAG instability rates (R2 = 0.11, Figure 3C).
The euchromatic state is also characterized by histone post-translational modifications associated with transcription, including Me3H3K4, which is enriched at the TSS of active genes [41]. We assessed by ChIP the level of Me3H3K4 at the HD locus in the striatum and cerebellum of R6/1 and R6/2 mice (Figure 4). R6/1 mice were analyzed at 13 and 36 weeks of age and R6/2 mice at 12 weeks. To closely target the HTT TSS region, we included an amplicon corresponding to a fragment located between the TSS and CAG repeats (HD exon-1, Figure 4A). Consistent with the idea that nucleosome occupancy is low upstream of the TSS, Me3H3K4 was low at the HD promoter (Figure 4B). In contrast, Me3H3K4 at HD exon-1 and HD intron-1 was elevated, particularly in R6/2 tissues when compared to R6/1 tissues. Globally, Me3H3K4 at the HD locus was significantly increased in R6/2 tissues relative to R6/1 matched-tissues (Figure 4C), and this was not dependent upon mouse age. These results support the view that the chromatin at the HD locus is in a more transcriptionally active state in R6/2 than in R6/1 mice, and suggest that the transcriptional activity of the HD transgene is increased in R6/2 mice when compared to R6/1 mice. This is consistent with the results of Figure 2 suggesting that the chromatin at the HD locus is more heterochromatinized in R6/1 mice than in R6/2 mice. Thus, high Me3H3K4 and low Me2H3K9 levels at the HD locus correlate with high CAG instability levels of R6/2 mice. However, the correlation the R2 value between Me3H3K4 levels at the HD locus and CAG instability rates in striatum and cerebellum of R6/1 and R6/2 mice was only 0.40, due to comparable Me3H3K4 levels in mouse striatum and cerebellum (Figure 4C). Our results suggest that Me3H3K4 level does not correlate with tissue-selective instability, though it correlates with mouse line-dependent instability.
The above results suggest that HD transgene expression might be increased in R6/2 mice when compared to R6/1 mice. We measured the levels of HD transgene mRNA in R6 tissues by using quantitative RT-PCR (qRT-PCR) (Figure 5). The striatum and cerebellum of R6/1 mice and R6/2 mice were analyzed at both early- and end-pathological stages (i.e. 6, 13 and 36 weeks and 6 and 12 weeks, respectively). We selected primers targeting a region upstream of CAG repeats in HD exon-1 (Figure S2). Expression values were normalized relative to expression of Gapdh or 18S expression (Figure 5A and Figure S3), which led to comparable results. At 6 weeks, transgene expression was 3- to 5-fold higher in R6/2 tissues in comparison to R6/1 tissue-matched samples, in accordance with the results above showing that chromatin structure at the HD locus is transcriptionally more active in R6/2 mice than in R6/1 mice. The differential transgene expression between R6/1 and R6/2 tissues was reduced over time, since HD transcript levels showed a tendency for decrease in R6/2 tissues. Transgene expression was not significantly different between the striatum and the cerebellum of either R6/1 or R6/2 mice. As a result, the correlation value R2 between HD transgene mRNA levels and CAG instability rates was only 0.40 (Figure 5B).
We further assessed transgene expression in striatum and cerebellum of 6 week-old R6/1 and R6/2 mice using an absolute quantification method with the corresponding genomic DNA as standards. This allowed for comparison between PCR product levels amplified with the above primers (located upstream of the repeats) and with primers encompassing the CAG repeats (“CAG” primers) (Figure 5C and S4). The level of transgene transcripts were comparable using both sets of primers, and increased in R6/2 tissues when compared to R6/1 tissue-matched samples using both sets of primers.
To investigate the relationship between transcription and somatic CAG instability, we performed ChIP from the striatum and cerebellum of R6/1 and R6/2 mice at late-pathological stages with the RNA Pol II 7C2 antibody, which recognizes phosphorylated and non phosphorylated forms of the polymerase [42]. RNA Pol II was distributed along the HD locus, peaking downstream of the TSS, at HD exon-1 (Figure 6A). Interestingly, RNA Pol II at the HD locus was higher in the striatum than in the cerebellum of both R6/1 and R6/2 mice, suggesting that the level of RNA Pol II at the HD region is regulated in a tissue-specific manner. Moreover, as previously found with Me2H3K9 and Me3H3K4 marks, RNA Pol II at the HD locus was increased (2- to 3-fold) in R6/2 tissues when compared to R6/1 tissue-matched samples (Figures 6A and 6B). As a result, RNA Pol II at the HD locus was highly correlated with CAG instability in R6 mice (R2 = 0.91, Figure 6B). RNA Pol II at control regions, including at Hdh regions, was also more abundant in the striatum than in the cerebellum of R6/1 and R6/2 mice (Figure 6A), further indicating that increased RNA Pol II at genes in striatum relative to cerebellum is tissue-specific. Together, these results suggest that a high level of RNA Pol II at the HD locus might contribute to CAG instability in vivo.
The activity of RNA Pol II during transcription is regulated by phosphorylation [43]–[45]. Ser5 phosphorylation of RNA Pol II is involved in transcription initiation, while Ser2 phosphorylation promotes transcription elongation. To discriminate between initiating and elongating RNA Pol II at the HD locus, we performed ChIP with antibodies to phosphorylated Ser5 RNA Pol II and phosphorylated Ser2 RNA Pol II (Figure 7A). As for Me3H3K4, which is associated with transcription initiation, the level of initiating RNA Pol II at the HD locus was increased downstream of the TSS, was comparable in the striatum and cerebellum of HD mice and was globally increased in R6/2 mice when compared to R6/1 mice (Figure 7A and 7B). As a result, initiating RNA Pol II was poorly correlated with somatic CAG instability in R6 mice (R2 = 0.30; Figure 7B). Elongating RNA Pol II at the HD locus peaked downstream of the TSS, particularly at the HD intron-1, but was more elevated in striata than in cerebella of both R6/1 and R6/2 mice (Figure 7C). In addition, the increase was more pronounced in R6/2 striatum relative to R6/1 striatum (Figures 7C and 7D). As a result, CAG instability rates were highly correlated with the level of elongating RNA Pol II at the HD locus (R2 = 0.86; Figure 7D). These results suggest that the dynamics of transition from initiation to elongation is tissue-specific and might contribute to tissue-selective CAG instability. Elongating RNA Pol II at the HD locus appeared abnormally high because we did not detect such levels at similar regions of control active genes, including Hdh (Figure 7C and S5). This might suggest that CAG repeats impair the dynamics of transcription. Together, these results suggest that high elongating RNA Pol II levels, resulting from tissue-specific regulation of elongation dynamics, and high transcriptional activity at the HD locus, as seen in R6/2 striatum, increase the propensity for repeat instability.
We further examined the correlation between transcription elongation and somatic CAG instability by performing ChIP with an antibody to Me3H3K36, a histone mark associated with transcription elongation and termination [41], [46]. Me3H3K36 is particularly elevated at the 3′ ends of transcribed genes, and, accordingly, was elevated at the last Hdh exon (Figure 8A). Me3H3K36 at HD intron-1 was lower than at Hdh last exon, but interestingly was significantly increased in striatal tissues in both R6/1 and R6/2 mice when compared to corresponding cerebellar tissues. Globally, Me3H3K36 level at the HD locus was increased in striatum when compared to cerebellum (Figure 8B), and strongly correlated with somatic CAG instability in HD mice (R2 = 0.94; Figure 8B). Since the tissue selectivity of somatic CAG instability is conserved between R6 mice and HD patients, these results, taken together, suggest that transcription elongation regulation might contribute to HD somatic CAG instability in vivo.
In vitro or cell-based studies support a role for transcription in CAG/CTG instability [20]–[24]. However, the expression level of TNR genes does not correlate with tissue-selective somatic instability in vivo [26], [27], [47], [48]. Thus, the role of transcription in somatic CAG instability has remained unclear (reviewed in [1]). Our results show that despite HD transgene expression being similar between striatum and cerebellum from both R6/1 and R6/2 mouse lines, these tissues present very different levels of CAG instability (Figure 5). In addition, the level of histone modifications associated with heterochromatin (Me2H3K9), euchromatin (AcH3K9) and transcription initiation (Me3H3K4) at the HD locus was comparable in the striatum and cerebellum of HD mice (Figure 2, Figure 3 and Figure 4). Remarkably however, the level of a histone mark and RNA Pol II form associated with transcription elongation (e.g. Me3H3K36 and phosphorylated Ser2 RNA Pol II, respectively) correlated strongly with CAG instability in tissues of R6/1 and R6/2 mice (Figure 7 and Figure 8). Thus, our data suggest a specific role for transcription elongation regulation in modulating tissue-selective CAG/CTG instability in HD. Moreover, comparison between mouse lines shows that increased CAG instability in R6/2 tissues relative to R6/1 matched-tissues correlates with increased chromatin accessibility and transcription of the HD transgene (Figure 2, Figure 4 and 5). This suggests that integration site of the transgene influences repeat instability, affecting chromatin structure and transcriptional activity.
Previous studies have shown that TNRs can induce a heterochromatinization process, consisting in formation and spreading of heterochromatin at TNR loci. This process has been described in the case of DM1, FRDA and FXS [29], [31], [37], [38], [49]–[51]. Me2H3K9 was not significantly increased at the HD locus in lymphoblasts derived from HD patients, suggesting absence or limited heterochromatinization of the locus (Figure 2). CAG expansion implicated in HD rarely exceeds 80 repeats, whereas TNRs associated with DM1, FRDA and FXS are typically larger (often reaching several hundred to thousand units), which could contribute to their high propensity to induce heterochromatinization, and to transcriptional silencing in the case of FRDA and FXS [29], [51]–[53]. Consistent with transcriptional repression, the levels of histone marks and RNA Pol II associated with transcription initiation and elongation are decreased at the FRDA locus in FRDA cells [37], [38], [54]. However, the effect of heterochromatinization on TNR instability has remained unclear. Our results suggest that heterochromatinization of the HD locus, a situation mimicked in R6/1 tissues, correlates with decreased somatic CAG instability. Whether heterochromatinization would limit TNR instability at the expense of transcription in FRDA or FXS is an intriguing possibility. If true, reversal of heterochromatinization through histone deacetylase (HDAC) inhibition, a process currently considered to alleviate GAA-repeat induced gene silencing in FRDA [55], might in turn increase repeat instability. Surprisingly, depletion of HDACs reduced TNR expansion in yeast and a plasmid-based cell model [56]. However, the effect was likely mediated through alteration of non-histone enzyme activities, including the Sae2 nuclease.
Factors involved in various DNA repair mechanisms, including MMR, BER and NER, contribute to CAG/CTG instability in vivo, most likely through aberrant processing of stable secondary structures formed at repeats, such as hairpins or slip-outs [1], [14], [16], [17], [19], [57]–[60]. Accessibility of DNA to repair factors increases when chromatin is in an open conformation. We therefore tested the hypothesis that CAG/CTG instability and chromatin accessibility might be correlated. Although Me2H3K9 at the HD locus was increased in R6/2 tissues when compared to R6/1 tissues, it was similar in the striatum and in the cerebellum of R6 mice, indicating that chromatin accessibility is unlikely to underlie tissue-selective HD somatic CAG instability (Figure 2). Instead, our data suggest a role for transcription, and more specifically for transcription elongation, in somatic CAG instability in HD. Events associated with transcription initiation were only partially correlated with CAG instability rates in R6 mouse tissues, since Me3H3K4 and phosphorylated Ser5 RNA Pol II levels at the HD locus were increased in R6/2 when compared to R6/1 tissues, but were comparable in striatum and cerebellum (Figure 4 and Figure 7). In contrast, events associated with transcription elongation were strongly correlated with somatic CAG instability. Specifically, Me3H3K36 and phosphorylated Ser2 RNA Pol II levels at the HD locus were increased in striatum relative to cerebellum as well as in R6/2 tissues as compared to R6/1 tissues (Figure 7 and Figure 8). Interestingly, a previous study based on a cell model allowing for detection of contraction events at CAG/CTG repeats showed that the elongation factor TFIIS contributes to repeat instability [15]. In addition, inhibition of the proteasome and downregulation of BRCA1/BARD1 proteins, which modulate RNA Pol II activity and transcription elongation, decreased CAG/CTG instability in the treated cells [15]. It remains to be determined whether these proteins as well as other factors controlling the dynamics of transcription elongation, including positive transcription-elongation factor-b (P-TEFb), negative elongating factor (NELF) or DRB sensitivity-inducing factor (DSIF), a factor composed of SPT4 and SPT5 [61]–[63], contribute to tissue-selective CAG instability in HD. Interestingly, a recent study showed that depletion of SPT4 in yeast and mammalian cells selectively prevents transcription of CAG-expanded genes, thereby reducing toxicity of the gene products [64]. Thus, targeting transcription elongation might reduce both repeat instability and toxicity.
Total RNA Pol II at the HD locus and at the promoter-proximal region of the murine Hdh gene were increased in the striatum when compared to the cerebellum (Figures 5). This suggests that transcriptional activity at the Huntingtin gene is differently regulated in the striatum and in the cerebellum in both mice and humans. However, and in accordance with previous studies showing that normal and mutant Huntingtin expression is widespread in many tissues and comparable in striatum and cerebellum, HD transgene expression was not different in the striatum and cerebellum, irrespective of mouse age or line (Figure 5) [26], [27], [47], [48], [65]. HTT mRNA processing might be regulated in a tissue-specific manner, as previously suggested [66], [67]. As a result, despite increased transcriptional activity in the striatum as compared to the cerebellum, HTT mRNA levels could be similar in the two tissues, due to increased HTT mRNA stability in the cerebellum. Alternatively, the dynamics of HTT transcription might be different in striatum and cerebellum, which cannot be examined by measuring bulk mRNA levels. Increasing evidence shows that genes can be expressed through alternative modes, ranging from bursts of transcription to more continuous smooth mode [68], [69]. Whether transcriptional mode can be regulated in a tissue-specific manner is currently unknown and cannot be easily assessed. The differential levels of RNA Pol II at HTT promoter-proximal region might suggest that the regulation of the transition from initiation to productive elongation is different in striatum and cerebellum, supporting the view that the dynamics of HTT transcription is different in the two tissues [70]. Whether HTT is expressed through a bursting mode in the striatum -promoting CAG instability, and a more continuous mode in the cerebellum –limiting repeat instability, is an attractive hypothesis.
The dynamics of transcription across CAG expansion might be impaired, particularly in the striatum where elongating RNA Pol II at the HD locus is elevated (Figure 7). Using an in vitro approach, it was recently shown that G/C-rich sequences induce transcription stalling, due to formation of stable R-loops during elongation [71]. Cell-based studies showed that increased transcriptional activity at genes with TNRs or bidirectional transcription across CAG/CTG repeats promotes formation of R-loops, leading to increased TNR instability [22]–[24], [72]. How then would R-loops promote TNR instability is currently unknown, but it has been hypothesized that arrest of elongating RNA Pol II at non canonical DNA structures might result in gratuitous transcription-coupled nucleotide excision repair (TCR) [73]. Supporting this view, it was shown that proteins of the NER/TCR pathway, including CSB, XPG and XPA, reduce transcription-mediated instability of CAG/CTG repeats in a cell model allowing for detection of contraction events [15], [20]. Accordingly, somatic CAG instability was reduced in post-mitotic tissues, including brain tissues, in spinocerebellar ataxia type 1 mice deficient for Xpa [19]. The idea that the propensity for formation of stable R-loops is increased in somatic tissues presenting high CAG instability rates, such as the striatum, is an intriguing hypothesis.
In conclusion, our results indicate that increased chromatin accessibility and transcription provide a context favoring somatic CAG instability but do not underlie the tissue selectivity of the process. Our data suggest that transcription elongation specifically contributes to tissue-specific CAG instability. It is tempting to speculate that the dynamics of transcription elongation of the HTT gene is regulated in a tissue-specific manner, influencing the risk for CAG instability.
Hemizygous R6/1 (130 CAG) and R6/2 (160 CAG) mice from the Jackson Laboratory were both maintained on a mixed CBAxC57BL/6 genetic background. The experiments were approved by the ethical committee C.R.E.M.E.A.S (Comité Régional d'Ethique en Matière d'Expérimentation Animale de Strasbourg).
Lymphoblasts from control individuals (GM14907, GM06903) and HD patients (GM04868, GM04282, GM04724, GM04846) were obtained from the Coriell Cell Repository. The size of CAG repeats in HD cell lines is 49/13, 74/17, 68/18 and 50/48, respectively. Cell lines were grown in RPMI 1640 medium without Hepes supplemented with 15% fetal calf serum and gentamycin under standard conditions.
CAG repeat size was determined as previously described [35], [36]. Briefly, PCR products containing the CAG repeats were analyzed using the ABI Prism 3100 DNA analyzer instrument and GeneScan and Genotyper softwares. The variation of repeat size corresponds to the amplitude of Genescan profile and was determined by calculating the number of peaks above 10% of the maximum fluorescent peak intensity. The striatum and cerebellum of the same mice were analyzed at each time point and 4 to 8 different mice at each time point were analyzed.
ChIP experiments were performed essentially as described [36]. Briefly, for each experiment, striata and cerebella from 3 to 5 transgenic mice were pooled, cut into small fragments, fixed in 1% formaldehyde and incubated for 10 min at room temperature. Cross-linking was stopped by the addition of glycine to final concentration 0.125 M. Tissue fragments were washed with cold PBS supplemented with protease inhibitors [36]. Phosphatase inhibitors were included when using antibody to phospho Ser2 or Ser5 RNA Pol II. The tissues were then homogenized in sonication buffer, and lysates were sonicated to obtain DNA fragments <500 bp and centrifuged, as described [36]. The soluble chromatin fraction was pretreated with protein A Agarose/Salmon Sperm DNA (Millipore) for 1 h at 4°C. Subsequently, samples were incubated overnight at 4°C with antibodies to AcH3K9 (ab10812, Abcam), Me2H3K9 (ab1220, Abcam), H3 (ab1791, Abcam), Me3H3K4 (ab1012, Abcam), Me3H3K36 (ab9050, Abcam), phospho S5 RNA Pol II (ab5131, Abcam), phospho S2 RNA Pol II (ab5095, Abcam) and 7C2 RNA Pol II antibody [42]. Protein A Agarose/Salmon Sperm DNA was then added and the mixture was incubated for 3 h at 4°C. Agarose beads were washed, protein-DNA complexes were eluted from the beads and decrosslinked, and DNA recovered by phenol chloroform extraction and ethanol precipitation after treatment with ribonuclease A (Abcam) and proteinase K, as described [36]. Typically, 2 to 3 antibodies were used per experiment, and immunoprecipitation with each antibody was duplicated. 2 to 4 independent ChIP experiments were performed with each antibody. The results are expressed as percentage of input, and correspond to mean values calculated from the different experiments. ChIP experiments from cell lines were performed using 5×106 cells per ChIP and the procedure described above. A typical ChIP experiment was performed using 4 HD cell lines and 2 control cell lines. Immunoprecipitations with each cell line were performed with antibodies to Me2H3K9 (ab1220, Abcam) and RNA Pol II (7C2) in duplicates, and 2 independent ChIP experiments were performed. The results, expressed as percentage of input, correspond to a representative experiment and are calculated by averaging values from the 2 control cells and from the 4 HD cells. The PCR primers used are listed in Figure S2.
Striata and cerebella were dissected, snap frozen and stored at −80°C. Total RNA was prepared from frozen tissues using the RNeasy Mini Kit (Qiagen). cDNA were prepared using 1 mg of total RNA, random hexamers and SuperScriptII reverse transcriptase (Invitrogen) according to the manufacturer instructions. Quantitative RT-PCRs were performed on a Light-Cycler instrument (Roche) to assess expression levels of HD transgene. The SYBR PCR master mix (Qiagen) was used to assess the level of HD transgene mRNA using the primers upstream of CAG repeats. Specific conditions were used to assess the level of HD transgene mRNA using primers amplifying the repeats, as previously described [36]. Briefly, the Herculase Hotstart DNA Polymerase (Stratagene) was used according to recommendation of manufacturer, and 8% DMSO and Sybr green (Molecular probe) were included in the reaction. The PCR reactions were performed on a Light Cycler instrument (Roche). DNA was inactivated for 3 min at 98°C, and 45 cycles consisting in 40 seconds at 98°C, 30 seconds at 60°C and 2 min at 72°C were performed. As a control, PCR reactions were also performed in the absence of reverse transcriptase. Data were normalized to expression of Gapdh or 18S housekeeping genes. For absolute quantifications, standard curves were generated using genomic DNA and used to calculate amounts of PCR products. cDNA samples prepared from the striatum and cerebellum of R6/1 and R6/2 mice were analyzed using corresponding genomic DNA. All experiments were performed at least three times in biological triplicates. The PCR primers used are provided in Figure S2.
ANOVA followed by Newman Keuls tests were used for statistical analyses.
|
10.1371/journal.ppat.1003349 | Yersinia pestis DNA from Skeletal Remains from the 6th Century AD Reveals Insights into Justinianic Plague | Yersinia pestis, the etiologic agent of the disease plague, has been implicated in three historical pandemics. These include the third pandemic of the 19th and 20th centuries, during which plague was spread around the world, and the second pandemic of the 14th–17th centuries, which included the infamous epidemic known as the Black Death. Previous studies have confirmed that Y. pestis caused these two more recent pandemics. However, a highly spirited debate still continues as to whether Y. pestis caused the so-called Justinianic Plague of the 6th–8th centuries AD. By analyzing ancient DNA in two independent ancient DNA laboratories, we confirmed unambiguously the presence of Y. pestis DNA in human skeletal remains from an Early Medieval cemetery. In addition, we narrowed the phylogenetic position of the responsible strain down to major branch 0 on the Y. pestis phylogeny, specifically between nodes N03 and N05. Our findings confirm that Y. pestis was responsible for the Justinianic Plague, which should end the controversy regarding the etiology of this pandemic. The first genotype of a Y. pestis strain that caused the Late Antique plague provides important information about the history of the plague bacillus and suggests that the first pandemic also originated in Asia, similar to the other two plague pandemics.
| Plague is a notorious and fatal human disease caused by the bacterium Yersinia pestis that is endemic in many countries around the world. Three of the most devastating pandemics in human history have been associated with plague. The second pandemic originated in central Asia and peaked in Europe between 1348 and 1350 (a period of time known as the Black Death). The third pandemic began in the Yunnan province of China in the mid-1850s and subsequently spread to Africa, the Americas, Australia, Europe, and other parts of Asia. The second and third pandemics are well documented and scientifically proven. However, the first pandemic, which began in the 6th century and is also known as Justinianic Plague, is still a matter of controversy. Recently it has been suggested that Justinian's plague was not caused by Y. pestis. We detected Y. pestis DNA in samples obtained from multiple 6th century skeletons from Germany. This confirms that Justinianic Plague crossed the Alps and affected local populations there, including current day Bavaria. Furthermore, we used DNA fingerprinting approaches to determine Asia as the likely geographic origin for these strains.
| In 541 AD, eight centuries before the Black Death, a deadly infectious disease hit the Byzantine Empire, reaching Constantinople in 542 and North Africa, Italy, Spain, and the French-German border by winter 543 [1]. The so called “Plague of Justinian”, named after the contemporaneous emperor, led to mass mortality in Europe similar to that of the Black Death. It persisted in the territory of the Roman Empire until the middle of the 8th century and likely contributed to its decline, shaping the end of antiquity [1]. Based on historical records, this disease has been diagnosed as bubonic plague although discrepancies between historical sources and the progression of Y. pestis infections have led some authors to suppose that the Plague of Justinian was caused by a different pathogen (as discussed in [2]). This vivacious discussion was recently reinforced by an ancient DNA study of the second pandemic that also questioned whether Y. pestis was truly the causative agent of the first pandemic [3], [4].
Western scientists have traditionally subdivided Y. pestis strains into three biovars: Antiqua, Medievalis, and Orientalis; depending on their abilities to ferment glycerol and reduce nitrate [5]. However, this system ignores many other Y. pestis biovars that have been designated and described by other scientists [see 6,7,8]. Biovars, which are based upon phenotypic properties, do not always correspond directly to specific molecular groups because the same phenotype can result from different mutations [9]. As a result, it has been suggested that groupings within Y. pestis, or assignment of unknown strains to specific populations should be based upon molecular signatures and not phenotypes [9]. Fortunately, the recent construction of highly-accurate rooted global phylogenetic trees for Y. pestis [10], [11] (reproduced in Figure 1) have facilitated the assignment of isolates to distinct populations. The most recent global phylogeny is based upon single nucleotide polymorphisms (SNPs) identified from the genomes of 133 global strains [11]. All clones that caused the third pandemic belong to populations assigned to the molecular group 1.ORI [10], [11]; the basal node for this group is N14 (Figure 1).
Two recent studies [3], [12] have queried key SNPs in DNA samples obtained from victims of the second pandemic (14th century AD), facilitating the phylogenetic placement of these samples in the most recent global phylogeny [11]. These samples are along the branch between nodes N07 and N10 (Figure 1) close to the “big bang” polytomy at node N07, where major branches 1–4 split from major branch 0 [11]. Specifically, ancient Y. pestis DNA samples from two of these studies [3], [12], which were obtained in England and France, are along branch N07-N10 – just one SNP away from the polytomy at N07 [11]. An additional sample from one of these studies [12], which was obtained in the Netherlands, occurs farther along this same branch – three SNPs away from the polytomy at N07 [11].
Only a few previous studies [13]–[15] have described the isolation of Y. pestis DNA from victims of the Late Antique pandemic and only one work group [13], [14] attempted to genotype the samples, assigning them to biovar Orientalis, which is now also designated molecular group 1.ORI [9]. However, the authenticity of these results has been questioned repeatedly because current stringent ancient DNA anticontamination protocols (e.g. independent replication) were not utilized [16], [17]. In addition, the robustness of the genotyping approach utilized in one of these studies [13] has been questioned [18]. Finally, it has been suggested [12], [19] that the resulting phylogenetic assignment (i.e. membership in the 1.ORI group) could not have existed at the time of the Justinianic Plague. Indeed, it seems impossible that isolates from the 1.ORI group caused the first pandemic as this group likely evolved only over the last ∼200–210 years [10], [11].
Against this background, we analyzed and genotyped new material from putative Justinian plague victims dated to the 6th century A.D. from an Early Medieval graveyard in Bavaria, Germany. This cemetery, called Aschheim, contained 438 individuals in total and is characterized by a striking number of double and multiple burials clustering in the second half of the sixth century [20]. In an earlier study [15], we reported isolation of Y. pestis DNA from two individuals from Aschheim. However, this previous study failed to utilize all of the contamination controls and authentication of results that has been recommended for studies that describe the detection of pathogen DNA in human remains from archeological sites [12], [21].
In this current study we utilized these more stringent approaches and our results confirm that Y. pestis was indeed responsible for the Justinianic Plague. More importantly, we were able to genotype the Y. pestis DNA present in samples from one individual using five key SNPs from the recent global Y. pestis phylogenies [10], [11]. The genotyping results confirm that the Y. pestis strain from the Ascheim victim is more basal on the global phylogeny than the Y. pestis populations that caused the Black Death and the third pandemic (Figure 1).
Assuming that plague victims might have been buried together, we collected teeth from 19 individuals originating from 12 multiple burials from the 6th century at Aschheim (Table 1). All samples were tested for Y. pestis specific DNA in a newly built specialized aDNA laboratory in Munich using both quantitative Real-Time PCR (qPCR) and a conventional PCR approach; these approaches targeted a 70 nt portion and a 133 nt portion of the Y. pestis-specific plasminogen activator gene (pla), respectively. This gene, which is located on the multi-copy plasmid pPst that is specific to Y. pestis, has been used in several previous studies to test samples from plague skeletons dating to the time of the Black Death [e.g. 12], [22].
Using qPCR, we repeatedly obtained a specific pla amplification fragment from samples obtained from eight individuals although, with the exception of sample A120, the target copy number was extremely low in most of the analyzed DNA extracts (Table 1). In addition, via conventional PCR we repeatedly obtained a longer pla amplification fragment from samples from two of these individuals (A82 and A120; Table 1). These amplicons contained pla sequences (GenBank accession number KC170159) that were 100% identical to the type strain CO92.
Concurrently, four samples obtained from intact teeth from four different individuals were independently analyzed in a second DNA laboratory (Mainz; Table 1). This analysis involved amplification of a 148 nt pla fragment by conventional PCR [12]. Only one of the four samples (from individual A120) produced an amplicon (Table 1). The observable differences in pla amplification success across the three PCR approaches utilized in this study (Table 1) are likely a function of the target PCR amplicon sizes. In agreement with typical ancient DNA behavior [23], our amplification success decreased with increasing target length (Table 1).
We attempted to genotype all of the positive samples. However, likely due to differences in DNA preservation among the samples we were only able to gain reproducible results from samples from one individual, A120 (Table 2). Note that this was the only individual that was found to be Y. pestis-positive with all three PCR approaches (Table 1). We queried multiple samples from individual A120 with assays targeting five key SNPs from the most recent global phylogenies for Y. pestis [10], [11] and determined whether these samples possessed the ancestral or derived states for these five SNPs (Table 2). These five SNPs occur along specific branches in the Y. pestis phylogeny: s545 occurs along the branch between nodes N06 and N07; s87 and s89 occur along the branch between N04 and N05, s82 occurs along the branch between the phylogenetic branching point of Mongolian strain MNG 2972 (see below) and N04, and s463 occurs along the branch between the phylogenetic branching point of strain MNG 2972 and N03 (Figure 1). In the Munich aDNA laboratory we determined that Y. pestis DNA samples obtained from individual A120 possess ancestral states for SNPs s545, s87, and s89; and derived states for SNPs s82 and s463 (Table 2). In the second aDNA laboratory (Mainz) we confirmed these results for s82 and s87 (Table 2); assays for the other SNPs were not utilized in this laboratory. Partial alignments of selected SNP regions of sample A120 in comparison to the reference sequences of Y. pestis type strain CO92 and strain 91001 (var microtus) are shown in Table 3. In all cases, extraction and PCR negative controls never produced an amplicon when tested with Y. pestis specific primers. These results indicate that the phylogenetic position of sample A120 in the global Y. pestis phylogeny is along the branch between the phylogenetic branching point of strain MNG 2972 and N04, along branch N04-N05, along the branch from N04 to 0.ANT1, or along one of the sub-branches within 0.ANT.1 (Figure 1).
Our analyses conducted in two separate aDNA laboratories independently confirmed our previous results [15] that some humans buried in the 6th century Ascheim cemetery were infected with Y. pestis. These findings confirm that Y. pestis was the causative agent of the Justinianic Plague and should end the controversy over the etiological agent of the first plague pandemic. This outcome is contrary to a recent study [3] that questioned whether Y. pestis was indeed the causative agent of the first pandemic based upon the assumption that only strains from major branches one and two are pathogenic to humans, which they estimated to have emerged only in the 13th century AD. However, Cui et al. [11] recently determined that most Y. pestis lineages are capable of causing human plague and suggested that this capability has been present since Y. pestis evolved from its Y. pseudotuberculosis ancestor approximately 1,500–6,400 years ago. Thus, they concluded that Y. pestis strains pathogenic to humans already existed long before the beginning of the first pandemic.
Another important issue resolved by our study concerns the geographic origin of the Plague of Justinian. The phylogenetic position of our Y. pestis samples from the first pandemic (Figure 1) suggests all three plague pandemics were caused by Y. pestis strains that originated in Asia. Two recent studies placed the origin of the 1.ORI strains that caused the first pandemic in China [10], [11], and recent phylogenetic placement of samples from the second pandemic [3], [12] near extant strains from China [11] (Figure 1) suggests that strains that caused the second pandemic also originated in this region. The only extant Y. pestis strains assigned to the same portion of the global phylogeny (Figure 1) as our Justinian samples from individual A120 are members of group 0.ANT1, which has only been reported from western China [10], [11], and strains from Mongolia [8], such as MNG 2972 (Figure 1). Although multiple historical sources have pointed to an African origin for the Justinian Plague [1], [5], [24], including speculations based on genealogies of Y. pestis [11], they have not discussed the original sources of where the bacteria arose. Our results document that those original sources were in Asia.
Cui et al. [11] recently raised the possibility that the Angola strain (sole representative of group 0.PE3; Figure 1) might have spread from Africa to all of Europe and been involved in the first pandemic. They based this hypothesis on several points. First, the Angola strain contains more SNPs than any other known strain of Y. pestis, which is consistent with a history of involvement in epidemic waves. Second, their 95% confidence intervals for the age estimates of the nodes that flank Angola (0.PE3) in the global phylogeny, nodes N01 and N03 (Figure 1), are 2,775 BC – 590 AD and 932 BC – 806 AD, respectively, which overlap with the 541 AD date given for the beginning of the first pandemic. Third, they assume that the strain named Angola was actually isolated in Africa in the country of Angola. We do not dispute their first two points. However, we know of no published studies that describe the original isolation of strain Angola making its origins apocryphal. Additional contemporary Angola-like isolates would add insights into this single unique strain type. Although it remains possible that Angola-like strains (ancestors), regardless of its geographic origin, may have been involved in the first pandemic, this remains just a hypothesis until additional samples from the first pandemic are genotyped and found to be closely-related to the Angola strain.
Multiple independent age estimates for our samples are consistent with the timing of the first pandemic. The duration of occupancy of the row burial cemetery at Aschheim-Bajuwarenring has been determined by strong archaeological evidence to range from approximately 500–700 AD [20]. Radiocarbon dating, which has been carried out on three individuals analyzed in this study, including A120 (Table 1), is consistent with this range. Finally, the phylogenetic position of our samples on the global Y. pestis phylogeny is on main branch 0 between nodes N03 and N05, with node N04 occurring in between (Figure 1). In their Figure S8, Cui et al [11] provide the 95% confidence intervals for the age estimates for these three nodes. The date given for the beginning of the first pandemic, 541 AD, overlaps with the confidence intervals for nodes N03 and N04, although not with the confidence intervals for N05. Collectively, these various age estimates for our samples provide convincing evidence that they are of the correct age to have been involved in the first plague pandemic.
Our results also provide new stimulus to the discussion about simultaneous multiple inhumations in Europe during the Early Medieval period [25], [26]. It is often presumed that only mass graves are suggestive of a highly infectious disease [27], whereas our results indicate that epidemics can also be indicated by a clustering of simultaneous inhumations involving only two or three individuals (Table 1). This observation may help to identify additional potential victims of the Justinianic Plague. Genetic studies of additional skeletal remains from other plague sites in different geographic regions would not only enhance our knowledge regarding the evolution of the pathogen, but also improve our understanding of the epidemics and spread of the Justinianic Plague. In addition, as there is no known historical source indicating that the Justinianic Plague reached current day Bavaria, our results provide the only evidence that the disease crossed the Alps and affected local populations there [1].
The burial date of the individuals tested for Y. pestis in this study were previously estimated by archaeological methods [20] to fall in a range from 525 to 680 AD (Table 1). To confirm this, we carried out radiocarbon dating on three samples. For individual A58, calibration indicated cal. 431–544 AD (95.4% probability) as the most likely range. Individual A76 from a second burial pit was dated to cal. 443–566 AD (95.4% probability), and individual A120 from a third burial pit was dated to cal. 435–631 AD. (95.4% probability).
From all 19 individuals (Table 1) two or more teeth were taken and analyzed at the aDNA laboratory in Munich. For four individuals (A58, A76, A105, and A120), another intact tooth was sent directly to a second aDNA laboratory in Mainz where they were analyzed independently and blindly.
In Munich the pre-PCR DNA analyses, including the decontamination procedure, DNA extraction, and assembly of the reactions for PCR amplification; were carried out in the new aDNA laboratories at the ArchaeoBioCenter (Ludwig-Maximillians-University, Munich). This aDNA laboratory is located several kilometers from the laboratory used for the post PCR analyses, which included the actual amplification process and sequencing; the post PCR laboratory is situated at the Bundeswehr Institute of Microbiology in Munich. Movement of samples between the laboratories was always unidirectional: from the aDNA laboratories to the post PCR laboratory. The pre-PCR laboratories are dedicated solely to aDNA analysis and have specialized equipment, such as airlocks, HEPA filtered air, positive air pressure, and UV air flow cleaner. In addition, extensive cleaning protocols using bleach and UV irradiation are in place. All possible further methodological precautions were also taken, such as mock extractions, PCR blanks, and independent replications of extractions and amplifications.
In the first step, samples were subjected to decontamination procedures consisting of cleaning the outer surface with a 1% NaOCl solution and exposure to 15 min of UV irradiation on each side, with subsequent powdering using a ZrO2-coated mill. DNA extraction in Munich was performed as described previously [15] on powder aliquots of 0.4 g. In Mainz precautions for preventing contamination, pre-treatment of the samples and extraction protocols were as published previously [12].
Every sample analyzed in the Munich laboratory for Y. pestis specific DNA (pla) was tested at least for three times using the qPCR and conventional PCR approach before considering it negative. Samples that yielded amplification products in any of these PCR reactions were submitted to genotyping assays targeting five key SNPs from the most recent global Y. pestis phylogenies [10], [11]: s545 (qPCR approach); s87 (both qPCR and conventional PCR approach); and s82, s89, and s463 (conventional PCR approaches).
For qPCR assays (pla), or qPCR SNP endpoint genotyping assays (s87 and s545), we used 1× Platinum Quantitative SuperMix-UDG (Invitrogen), 6 mM MgCl2, (Applied Biosystems), 0.4 mg/ml BSA (Ambion/Life Technologies), assay specific primer and probe concentrations (Table 4) (TibMolbiol), and 2.0 to 4.0 µl of template DNA in a final reaction volume of 12 to 24 µl. Primer sequences are listed in Table 4. Cycling conditions comprised an initial step at 50°C for 2 min, an activation step at 95°C for 10 min, 50 cycles at 95°C for 10 sec, and an assay specific annealing temperature for 1 min (Table 4). Final cooling was carried out at 4°C for 30 sec. QPCR assays were carried out on a LightCycler 480 II platform (Roche, Mannheim, Germany). Quantification of pla-qPCR assays was possible by determination of the copy numbers per reaction by generating a standard curve using synthetic oligonucleotide constructs. Data analysis was performed using the LightCycler 480 II software version 1.5 (Roche, Mannheim, Germany).
For conventional PCR assays (pla, s82, s87, s89, s463), we used 1× Qiagen Multiplex PCR Master Mix, 0.4 mg/ml BSA, and 2 or 4 µl of DNA in a final volume of 50 µl. Primer sequences are listed in Table 4. The experiments were run on an Eppendorf Mastercycler Pro instrument. Cycling conditions started with an initial activation step at 95°C for 15 min. This was followed by 50 cycles at 94°C for 30 sec, an assay specific annealing temperature (Table2) for 30 sec, and 72°C for 1 min, ending with a final elongation step at 72°C for 10 min. Final cooling was carried out at 8°C until analysis.
Results (pla or SNPs) were only considered valid if they could be repeated at least three times from different extracts. Protocols for pla, s82, and s87 analysis in the second aDNA lab (Mainz) were carried out as previously published [12].
All amplified products were verified by DNA sequencing and BLASTN-analysis.
For the sequencing reactions in Munich we used 1× BigDye terminator v.3.1 Cycle Sequencing Ready reaction Mix (Applied Biosystems), 1 pmol/µl of the respective primers, and 3–5 µl of purified DNA template in a final volume of 10 µl. The reaction was run on a GeneAmp 9700 (Applied Biosystems) instrument, starting with an initial denaturation step for 1 min at 96°C, followed by 25 cycles at 96°C for 10 sec, 50°C for 5 sec and 60°C for 2 mins, and ending with cooling at 4°C until further processing. After purification using the Dye Ex 2.0 Spin Kit (Qiagen) sequences were generated on a Genetic Analyzer 3130 (Applied Biosystems) instrument. Sequences were further analyzed using the program CodonCodeAligner version 4.0. Analyses of the results of the SNPs assays were carried by aligning the amplicons to Y. pestis type strain CO92 (AL590842.1), which possessed the derived state for all of the queried SNPs, and Y. pestis microtus strain 91001 (AE017042.1), which possessed the ancestral state for all of the queried SNPs. Sequencing in Mainz was carried out as previously described [12]. If long enough, sequences were deposited in GenBank (Accession numbers KC170160-KC170162) and the alignments are shown in Table 3 (only partial sequences are shown for longer sequences).
The global Y. pestis phylogeny in Figure 1 is reconstructed from Figures 1A and S3B in Cui et al. [11]. Their phylogeny was constructed using SNPs discovered from the genomes of 133 modern isolates. We have indicated the main branches and molecular groups identified by Cui et al. [11] but not all of their sub-branches and sub-groups. The phylogenetic branching point for Mongolian Y. pestis strain MNG 2972 was determined using SNP information provided for this strain in Riehm et al. [8]. Note that, based upon the five SNPs queried in this study, this contemporary Mongolian strain possesses a distinct genotype when compared to the ancient Y. pestis DNA samples utilized in this study; the Mongolian strain possesses the ancestral state for s82.
The GenBank (http://www.ncbi.nlm.nih.gov) accession numbers for DNA sequences longer than 50 nt determined in this paper are KC170159-KC170163.
|
10.1371/journal.pgen.1002221 | Genetic Architecture of Aluminum Tolerance in Rice (Oryza sativa) Determined through Genome-Wide Association Analysis and QTL Mapping | Aluminum (Al) toxicity is a primary limitation to crop productivity on acid soils, and rice has been demonstrated to be significantly more Al tolerant than other cereal crops. However, the mechanisms of rice Al tolerance are largely unknown, and no genes underlying natural variation have been reported. We screened 383 diverse rice accessions, conducted a genome-wide association (GWA) study, and conducted QTL mapping in two bi-parental populations using three estimates of Al tolerance based on root growth. Subpopulation structure explained 57% of the phenotypic variation, and the mean Al tolerance in Japonica was twice that of Indica. Forty-eight regions associated with Al tolerance were identified by GWA analysis, most of which were subpopulation-specific. Four of these regions co-localized with a priori candidate genes, and two highly significant regions co-localized with previously identified QTLs. Three regions corresponding to induced Al-sensitive rice mutants (ART1, STAR2, Nrat1) were identified through bi-parental QTL mapping or GWA to be involved in natural variation for Al tolerance. Haplotype analysis around the Nrat1 gene identified susceptible and tolerant haplotypes explaining 40% of the Al tolerance variation within the aus subpopulation, and sequence analysis of Nrat1 identified a trio of non-synonymous mutations predictive of Al sensitivity in our diversity panel. GWA analysis discovered more phenotype–genotype associations and provided higher resolution, but QTL mapping identified critical rare and/or subpopulation-specific alleles not detected by GWA analysis. Mapping using Indica/Japonica populations identified QTLs associated with transgressive variation where alleles from a susceptible aus or indica parent enhanced Al tolerance in a tolerant Japonica background. This work supports the hypothesis that selectively introgressing alleles across subpopulations is an efficient approach for trait enhancement in plant breeding programs and demonstrates the fundamental importance of subpopulation in interpreting and manipulating the genetics of complex traits in rice.
| While rice (Oryza sativa) is significantly more Al tolerant than other cereals, no genes underlying Al tolerance in rice have been reported. Using genome-wide association (GWA) and bi-parental QTL mapping, we investigated the genetic architecture of Al tolerance in rice. Japonica varieties were twice as Al tolerant as indica and aus varieties. Overall, 57% of the phenotypic variation was correlated with subpopulation, consistent with observations that different genes and genomic regions were associated with Al tolerance in different subpopulations. Four regions identified by GWA co-localized with a priori candidate genes, and two highly significant regions co-localized with previously identified quantitative trait loci (QTL). Haplotype and sequence analysis around the candidate gene, Nrat1, identified a susceptible haplotype explaining 40% of the Al tolerance variation within the aus subpopulation and three non-synonymous mutations within Nrat1 that were predictive of Al sensitivity. Using Indica × Japonica mapping populations, we identified QTLs associated with transgressive variation where alleles from a susceptible indica or aus parent enhanced Al tolerance in a tolerant japonica background. This work demonstrates the importance of subpopulation in interpreting and manipulating complex traits in rice and provides a roadmap for breeders aiming to capture genetic value from phenotypically inferior lines.
| Aluminum (Al) toxicity is the major constraint to crop productivity on acid soils, which comprise over 50% of the world's arable land [1]. Under highly acidic soil conditions (pH<5.0), Al is solubilized into the soil solution as Al3+, which is highly phytotoxic, causing a rapid inhibition of root growth that leads to a reduced and stunted root system, thus having a direct effect on the ability of a plant to acquire both water and nutrients.
Cereal crops (Poaceae) have been a primary focus of Al tolerance research [2]. This research has demonstrated that levels of Al tolerance vary widely both within and between species [3]–[8]. Of the major cereal species that have been extensively studied (rice, maize, wheat, barley and sorghum), rice demonstrates superior Al tolerance under both field and hydroponic conditions [3], [8]. Although rice is 6–10 times more Al tolerant than other cereals, very little is known about the genes underlying this tolerance. Based on its high level of Al tolerance and numerous genetic and genomic resources, rice provides a good model for studying the genetics and physiology of Al tolerance.
In wheat, sorghum, and barley, Al tolerance is inherited as a simple trait, controlled by one or a few genes [9]–[11]. However, in maize, rice, and Arabidopsis, tolerance is quantitatively inherited [12], [13]. Al tolerance genes have been cloned in wheat and sorghum. The wheat resistance gene, ALMT1, encodes an Al-activated malate transporter [14]. The sorghum resistance gene, SbMATE, encodes a member of the multidrug and toxic compound- extrusion (MATE) family and is an Al-activated, root citrate efflux transporter [15]–[17].
Four mutant genes that lead to Al sensitivity in rice have recently been cloned, STAR1 (Sensitive to Al rhizotoxicity1), STAR2 (Sensitive to Al rhizotoxicity2), ART1 (Aluminum rhizotoxicity 1), and Nrat1 (Nramp aluminum transporter 1) [18]–[20]. The products of STAR1 and STAR2 are expressed mainly in the roots and are components of a bacterial-type ATP binding cassette (ABC) transporter. Both are transcriptionally activated by exposure to Al and loss of function of either gene results in hypersensitivity to Al. STAR1 and STAR2 are similar to two Al sensitive mutants in Arabidopsis, als1 and als3, also encoding ABC transporters [21], [22]. ART1 is a novel C2H2-type zinc finger transcription factor that interacts with the promoter region of STAR1. ART1 is reported to regulate at least 30 down-stream genes, some of which are involved in Al detoxification and serve as strong candidate genes controlling rice Al tolerance [19]. Nrat1 is one of the genes that is regulated by ART1 and was recently demonstrated to be an Al transporter that is localized to the root cell plasma membrane [18], [20]. It is hypothesized that Nrat1 confers Al tolerance by transporting Al into the cell and reducing the concentration of Al in the cell wall [20]. None of the four cloned rice genes described above have been demonstrated to be involved in natural genetic variation of Al tolerance in rice and only one (Nrat1) maps to a previously reported Al tolerance QTL [23], suggesting that these genes may be involved in basal Al tolerance [19], [20], [24]. A more thorough analysis is necessary to determine whether there might be natural variation associated with these loci that would help trace their evolutionary origins and clarify their contribution to the high levels of Al tolerance observed in rice.
Seven QTL studies on Al tolerance have been reported in rice using 6 different inter- and intra-specific mapping populations [13], [25]–[29]. Together, these studies report a total of 33 QTLs, located on all 12 chromosomes, with three intervals (on chromosomes 1, 3, and 9) being detected in multiple studies. In all of the QTL studies, Al tolerance was estimated based on relative root growth (RRG), and specifically on inhibition of the growth (elongation) of the longest root (elongation of the longest root in Al treatment/root growth of controls). Rice has a very fine and fibrous root system without dominant seminal roots. We recently showed that there is a weak correlation between rice Al tolerance based on RRG of the longest root and RRG of the total root system (R2 = 0.17) [8]. This raises the question whether mapping Al tolerance QTL using total root and longest root RRG indices independently might identify novel loci, helping to integrate QTL studies with studies based on induced mutations.
Historically, O. sativa has been classified into two varietal groups, Indica and Japonica, based on morphological characteristics, ecological adaptation, crossing ability and geographic origin [30]. These two varietal groups are believed to represent independent domestications from a pre-differentiated ancestral gene pool (O. rufipogon), followed by significant gene flow among and between subpopulations [17], [31]–[39]. These two varietal groups (names are italicized with an upper case first letter, i.e., Indica and Japonica) have been further divided into five major subpopulations (subpopulation names are italicized using all lower-case letters) (indica, aus, tropical japonica, temperate japonica, and aromatic [group V]) based on DNA markers (SSR, SNPs, indels) [40]–[42]. Genotypes that share <80% ancestry across subpopulations or varietal groups are classified as admixed varieties [42], while smaller groups adapted to specific ecosystems may be recognized as upland, deep water, or floating varieties [43], [44]. Upland varieties, which are generally grown at high altitudes on dry (non-irrigated) soils, are those most commonly exposed to acidic, Al-toxic soil conditions. These varieties are almost invariably of tropical japonica origin, suggesting a priori that the tropical japonica subpopulation would be a likely source of superior alleles for Al tolerance in rice.
Diverse panels of O. sativa are reported to have similar, or slightly elevated levels of linkage disequilibrium (LD) compared to species such as Arabidopsis, maize and human. The average extent of LD in rice has been estimated at between 50–500 kb [45]–[49], depending on the germplasm evaluated, compared to 10–250 kb in Arabidopsis and human [50]–[57], 100–500 kb in commercial elite maize inbreds and 1–2 kb in diverse maize landraces [58], [59]. The inbreeding nature of O. sativa, coupled with its demographic history, are major determinants of genome-wide patterns of LD. Strong selective pressure over the course of rice domestication has also lead to deep population substructure (Fst = 0.23 to 0.57) [40], [42], which sets it apart from Arabidopsis, in which population structure is gradual across geographic distances [60], [61]. Population substructure can lead to false-positives in association mapping studies, and must be taken into account [61]–[63]. The mixed-model has been demonstrated to work well in both maize and Arabidopsis [61], [63], and it has also shown its ability to greatly reduce the false positive rates in rice when used within a single subpopulation [64], though it may introduce false negatives when used on a diversity panel representing all domesticated subpopulations [65].
A diversity panel consisting of 413 O. sativa accessions, representing the genetic diversity of the primary gene pool of domesticated rice [66], was recently genotyped with 44,000 SNPs (∼10 SNPs/kb) [65], [67], [68] as the basis for GWA studies. The slow decay of LD, while facilitating GWA analysis, limits the resolution of association mapping in rice. The first targeted association mapping study in rice [45] demonstrated that LD decay in the aus subpopulation was approximately 90 kb (∼5 genes) in a region on chromosome 5 containing the xa5 resistance gene. LD is expected to decay more quickly in O. rufipogon (<50 kb, or 1–3 genes) [48], providing higher resolution for LD mapping, and more slowly in the japonica subpopulations [47]–[49]. Nonetheless, when compared to the resolution of a typical QTL study (250 lines) (∼10–20 cM resolution, where 1 cM = ∼250 kb), association mapping is expected to provide between 10–200 times higher resolution for a population of similar size as long as sufficient marker density is obtained to exploit the historical recombination. Thus, an association mapping study that uses markers densities similar to a QTL study will not have the increased resolution and will increase the risk of type-2 error. For both GWA and QTL analysis in rice, fine-mapping and/or mutant analysis is generally required to identify the gene(s) underlying a QTL of interest. However, the fine-mapping phase can generally be focused on a smaller target region following GWA analysis.
In this study, the genetic architecture of rice Al tolerance was investigated via bi-parental QTL analysis in two mapping populations using relative root growth of the longest root, the primary root system, and the total root system quantified with the digital root phenotyping methods described previously for rice Al tolerance [8]. Subsequently, genome wide association (GWA) analysis was undertaken using 36,901 high quality SNPs that had been genotyped on the rice diversity panel [65]. Regions identified by GWA were compared with regions identified as QTLs in bi-parental mapping populations for both this and previous studies, as well as with Al sensitive mutants and/or candidate genes. Phenotypic outliers identified in the diversity panel were further investigated to identify regions of subpopulation-admixture that accounted for extreme Al tolerance phenotypes.
Three hundred eighty three diverse O. sativa accessions from the rice diversity panel [42], [67](Table S1) were evaluated for Al tolerance using an Al3+ activity of 160 µM in a hydroponic nutrient solution. This Al3+ activity had been previously determined to be optimal for evaluating a wide range of Al tolerance in diverse rice germplasm [8]. In the diversity panel, Al tolerance, measured as the relative root growth of the total root system (TRG-RRG), was normally distributed around a mean of 0.59 +/−0.24(SD) and ranged from 0.03–1.35 (Figure 1A). Some varieties were inhibited by as much as 97%, while 16 varieties (representing three subpopulations) showed enhanced root growth in the presence of 160 µM Al3+ (Table S1).
When accessions were grouped based on varietal group (>80% ancestry) the Japonica varietal group (consisting of the temperate japonica, tropical japonica and aromatic subpopulations) was significantly more Al tolerant than the Indica varietal group (indica and aus subpopulations) (p<0.0001) (Figure 1B). The Japonica varieties had a mean Al tolerance value of RRG = 0.72, an interquartile range of 0.61–0.82, and ranged from 0.13–1.35. The Indica varieties had a mean Al tolerance value of RRG = 0.36, an interquartile range of 0.27–0.43, and ranged from 0.03–1.15 (Figure 1B). Eleven accessions were classified as “admixed” between varietal groups, and these had a mean Al tolerance equal to the mean of all 372 accessions (TRG-RRG = 0.59) with >80% ancestry to either varietal group. A one-way ANOVA demonstrated that subpopulation explained 57% of the phenotypic variation observed for Al tolerance (TRG-RRG) among the 274 accessions that carried a subpopulation classification. Despite the differences in mean TRG-RRG between subpopulations, considerable variation was also detected within each subpopulation (Figure S1).
Two immortalized QTL mapping populations were analyzed for Al tolerance. One consisted of 134 recombinant inbred lines (RIL) derived from the cross IR64/Azucena [69], and the other was comprised of 78 backcross inbred lines (BIL) derived from the cross Nipponbare/Kasalath//Nipponbare [70]. These populations were used to evaluate Al tolerance using three different indices of relative root growth (RRG), (1) longest root growth (LRG-RRG), (2) primary root growth (PGR-RRG) and (3) total root growth (TRG-RRG) (see Materials and Methods for details). The phenotypic distribution was approximately normal for each population, no matter which root screening index was used (illustrated for TRG-RRG in Figure S2A and S2B). The QTL mapping populations allowed us to determine which of the three root evaluation methods would be most useful for evaluating the diversity panel as a whole.
The method of phenotyping, specifically, the RRG index used to estimate Al tolerance, directly impacted the significance of QTLs detected by composite interval mapping (Figure 2A–2C and Figure S3A–S3C). In the RIL population, three Al tolerance (Alt) QTL were detected using total root growth (the TRG-RRG index), AltTRG1.1 on chromosome 1, AltTRG2.1 on chromosome 2, and AltTRG12.1 on chromosome 12 (Figure 2A–2C Table 1). The Azucena allele conferred increased tolerance at the loci on chromosomes 1 and 12 and reduced tolerance at the locus on chromosome 2. QTLs were detected in the same positions on chromosomes 1 and 12 using RRG based on primary root growth (the PRG-RRG index), although with lower LOD scores (Figure 2A–2C; Table 1). Using longest root growth (the LRG-RRG index), a single QTL was detected on chromosome 9, AltLRG9.1, and this QTL was not detected when the other root indices were used. The major QTL on chromosome 12 (AltTRG12.1), which explained >19% of the variation in Al tolerance based on TRG-RRG, is located between 2.69–5.10 Mb and encompasses the Al sensitive rice mutant art1, which is located at 3.59 Mb [19].
In the BIL population, two QTL were detected using the TRG index, AltTRG1.2 on chromosome 1, which co-localized with the AltTRG1.1 QTL identified in the RIL population, and AltTRG12.2 on chromosome 12, which did not overlap with the AltTRG12.1 identified in the RIL population (Figure 2A–2C, Figure S3A–S3C, Table 1). The Nipponbare allele conferred tolerance at the chromosome 1 locus and the Kasalath allele conferred tolerance at the AltTRG12.2 locus. No QTLs were detected on chromosome 2 in the BIL population. Using the PRG-RRG index, one QTL was detected on chromosome 6, where the Kasalath allele conferred resistance. No QTLs were detected using the LRG-RRG index in the BIL population.
The Al tolerance index used for evaluating the phenotype directly affected both the identity and the significance of the QTLs detected. Al tolerance index-specific QTLs were detected in both populations and no QTL locus was detected across all three indices. Based on number of QTL detected, significance of QTL, and variance explained by the QTL, total root growth (TRG) proved to be the single most powerful Al tolerance index. However, rice QTLs detected using different evaluation methods are likely to confer Al tolerance by different mechanisms, such as tolerance of primary, secondary, lateral, or all roots, and thus they are complementary and together provide a robust evaluation of the genetic architecture of Al tolerance than any single index alone.
To identify Al tolerance loci based on genome-wide association (GWA) mapping, we used an existing genotypic dataset consisting of 36,901 SNPs [65], and the total root growth (TRG-RRG) Al tolerance phenotype generated on 373 O. sativa accessions over the course of this study. GWA mapping was conducted, using SNPs with a MAF>0.05, across all 373 genotypes as well as independently within the indica, aus, temperate japonica, and tropical japonica subpopulations (Figure 3). The Efficient Mixed-Model Association (EMMA) [71] model was used in each analysis (both within and across subpopulations) to correct for confounding effects due to subpopulation structure and relatedness between individuals. As the subpopulation structure was highly correlated with Al tolerance, it was observed that analyzing all samples (373) together with the EMMA model resulted in an overcorrection (causing type 2 error) and a corresponding reduction in SNP significance (Figure S4). To address this problem, a PCA approach was also employed when analyzing all (373) samples together. However, the PCA approach resulted in a slight under-correction for population structure (Figure S4), demonstrating that results from each GWA method has limitations when used across all germplasm in this highly structured diversity panel.
A total of ∼48 distinct Al tolerance genomic regions were identified by GWA mapping (Figure 3). Twenty-one regions were detected (p<0.0001) across all (373) accessions using the PCA model (Figure 3), while only two SNPs were above the significance threshold when all (373) accessions were analyzed together using the EMMA model (Figure 3), both of which were also detected by PCA. The threshold of p<1.0E-04 was determined based on the upper-limit false discovery rate (FDR), determined from the candidate genes in the same approach as in Li et al. [72] (Table S2). Thirty-two regions were significantly associated with Al tolerance in the indica subpopulation (Figure 3), including five regions that were also detected across all (373) samples using the PCA model. In the aus subpopulation, a single, highly significant, region was detected on chromosome 2 that was unique to this subpopulation and contained the Nrat1 candidate gene LOC_Os02g03900 (Figure 3). No significant SNPs (MAF>0.05) were detected in the temperate japonica or tropical japonica subpopulations. The GWA mapping results indicate that the majority of significant loci are subpopulation-specific and that phenotypic variation for Al tolerance within given subpopulations is largely controlled by alleles that are unique to that subpopulation.
SNPs identified by GWA were also compared to a set of 46 a priori candidate genes as well as to positions of QTL regions identified through bi-parental mapping (this study and previous reports) (Table 1 and Figure 3). Two regions of highly significant SNP clusters, one within the aus (8 SNPs; p = 2.8E-07) subpopulation on chr. 2 and one within the indica (32 SNPs; p = 2.9E-07) subpopulation on chr. 3, co-localized to previously reported QTLs in populations in which an aus and indica parent served as the susceptible parents, respectively [17], [23]. The list of 46 a-priori Al tolerance candidate genes (Table 2) was compiled based on published information on Al sensitive mutants from rice and Arabidopsis [20]–[22], [24], cloned Al tolerance genes from wheat and sorghum [14], [15], expression profiles from Al treated maize and rice roots [19], [73], and an association study on specific candidate Al tolerance genes of maize [74]. Significant SNPs (p<1.0E-04) within a 200 kb window of the a priori candidate genes were enriched 2.4 times compared to other SNPs (p>0.0001) outside of the a priori and QTL regions. The 200 kb window was selected to fall within the estimated window of LD decay in rice (∼50–500 kb [45]–[49] and the upper-limit false discovery rate for the a priori genes was 42%. In addition, four of the 46 gene candidates (∼9%) were located within a 200 kb window enriched for GWA SNPs in this study (Figure 3 and Table 2). One of the candidate genes (Nrat1) on chr. 2, co-localized with both GWA SNPs and a previously reported QTL (Figure 3). The relationship between the four candidates that co-localized with GWA SNPs are discussed in order of their positions on the rice genome below.
A cluster of eight highly significant SNPs (p-values = 2.3×10−5–2.8×10−7) on chromosome 2 between 1.536 Mb–1.675 Mb was associated with Al tolerance within the aus subpopulation (Figure 3 and Table 2). Previously, a QTL had been reported in the same location (0.536–1.9 Mb) where the susceptible parent was of aus origin [26]. The LD decay in the aus subpopulation at this region was calculated to be 150 kb and a strong candidate gene was identified within the target region. The gene (LOC_Os02g03900 located at 1.66 Mb) encodes a Nramp6 metal transporter and was demonstrated to have altered expression patterns in Al-treated roots of the Al sensitive art1 rice mutant [19]. This Nramp6 metal transporter was recently reported as Nrat1, a plasma membrane-located transporter for Al with enhanced sensitivity to Al in the knockout mutant [20]. As was the case with the ART1 gene itself, the Nrat1 metal transporter has not been associated with natural variation for Al tolerance prior to this study.
On chromosome 5, a significant region was detected across all samples (373 genotypes) by PCA, co-localizing with the STAR2 gene (LOC_Os05g02750) (Figure 3 and Table 2). The LD decay across this region was estimated at >500 kb, and encompassed two significant regions detected across all samples (PCA), one of which was also detected within the indica subpopulation. STAR2 is the rice ortholog of the Arabidopsis Al sensitive mutant als3 [21]. It encodes the transmembrane domain of a bacterial-type ATP binding cassette (ABC) transporter and the star2 mutant is Al sensitive [24]. STAR2 was also found to be part of a gene network showing altered expression in response to Al in the art1 mutant compared to the ART1 wild type [19]. This study provides the first evidence that there may be natural variation for Al tolerance in rice at the STAR2 locus; however it is important to recognize that the PCA approach may under-correct for the effect of subpopulation in this study, thus it will be necessary to confirm the effect of the STAR2 alleles identified in this diversity panel.
A significant GWAS region identified in the indica subpopulation on chromosome 7 co-localized with LOC_Os07g34520, a rice ortholog of a maize isocitrate lyase a priori candidate gene associated with Al tolerance in maize [73], [74]. The LD decay across this region within the indica subpopulation was 250 kb.
Three highly significant regions detected within indica were further investigated to identify whether any clear Al tolerance candidate genes were located within these SNP clusters. The first region was a cluster of 32 significant SNPs (p = 3.0E-7) between 28.782–27.863 Mb on chr. 3 that co-localized with a previously reported QTL (Nguyen et al., 2002). Two clear candidates were identified among the 13 genes in this cluster; a nucleobase-ascorbate transporter (LOC_Os03g48810) and a chloride channel protein (LOC_Os03g48940). The second region was a 10 SNP cluster (p = 9.3E-12) between 26.986–27.479 Mb on chr. 7. Of the 80 genes in this region, 34 of which were retrotransposons, there were three strong candidate genes; a glycosyl transferase protein (LOC_Os07g45260), a cytochrome P450 protein (LOC_Os07g45290) and a zing finger RING type protein (LOC_Os07g45350). This region on chr. 7 was also identified in the introgression analysis as a localized introgressed region from Japonica into the highly tolerant Indica outliers (discussed below). The third region was an 8 SNP cluster between 4.892–5.164 Mb on chr. 11. Among the 48 genes in this region, there were two major classes of candidate genes observed, including 12 F-box proteins and a zinc finger CCHC protein.
We chose to further investigate the variation in and around the Nrat1 gene on chromosome 2 because multiple independent lines of evidence supported the existence of a gene(s) in this region responsible for a significant portion of the variation for Al tolerance in rice. Evidence included a strong GWA peak in the aus subpopulation, a previously reported QTL [26], and the localization of the Nrat1 Al transporter gene. Using the 44 K SNP data, LD in this region was calculated to be ∼150 kb in the aus subpopulation and 11 distinct haplotypes were observed in the entire diversity panel across a 139 kb region around the Nrat1 gene (1.536 Mb–1.675 Mb on chr. 2) (Figure 4A). Haplotype 1 (Hap. 1), which was unique to the aus subpopulation, was found in 8 Al sensitive aus accessions and one Al sensitive aus/indica admixed line. These 9 genotypes were among the least Al tolerant (7th percentile, mean RRG = 0.16) of the 373 accessions screened (Table S1). Haplotype 1 explained 40% of the phenotypic variation for Al tolerance within the aus subpopulation (Figure S5). In addition, four aus accessions that were highly or moderately Al tolerant were found to contain a tropical japonica introgression across this region (described in the section on Introgression analysis below).
Haplotype 2 (Hap. 2) was found in one aus and one indica accession, and was most similar to Hap. 1, differing at only 2/14 SNPs (Figure 4A). The two lines containing haplotype 2 had very different levels of Al tolerance; the aus variety, Kasalath (ID 85), was highly susceptible, with a RRG = 0.2, while the indica variety, Taducan (ID 163), was tolerant, with a RRG = 0.8, suggesting that this extensive 14-SNP haplotype across the 139 kb region was not predictive of Al tolerance. However, when the haplotype was built using only the four SNPs immediately flanking the Nrat1 gene, a group of 16 accessions sharing the same haplotype at these four SNPs was clearly identified. These 16 accessions, included the 10 susceptible aus accessions (including one aus/indica admixed line) carrying haplotype 1 and haplotype 2 and six indica accessions (of varying Al tolerance) carrying haplotype 2 and haplotype 3 (Figure 4A).
To determine if the four-SNP haplotype flanking the Nrat1 gene could be further resolved, we focused more deeply on the Nrat1 gene itself. We sequenced all 13 exons (including introns) of Nrat1 (1874 bp) in 26 susceptible and tolerant varieties representing the aus, indica, tropical japonica and temperate japonica subpopulations (Figure 4B). The accessions carried haplotypes 1, 2, 3, 6 and 11, as described in Figure 4A; where haplotype 1 was aus-specific and corresponded to the most sensitive group of accessions in the diversity panel; haplotype 2 was found in phenotypically divergent aus and indica accessions as described above; haplotype 3 was found in moderately tolerant indica varieties; haplotype 6, which appeared to be the ancestral haplotype, was the most common haplotype in all subpopulations and was associated with moderately high levels of tolerance; and haplotype 11, which was found in a majority of tropical japonica varieties, all of which were Al tolerant. Based on the 22 SNPs and/or indels identified across the 1,874 bp of Nrat1 sequence, highly resolved, gene haplotypes were constructed (Figure 4B). The gene haplotypes corresponded fairly well to the extended haplotype groups that had been constructed using the data from the 44 K SNP chip, except in the case of haplotype 2, where varieties differed at 10/22 (45%) of the SNPs across the Nrat1 gene. This fully resolved haplotype at the Nrat1 gene resulted in the susceptible Kasalath clustering with the other highly susceptible aus varieties and the tolerant Taducan clustering with other highly tolerant varieties (Figure 4).
Three non-synonymous SNPs (polymorphisms 4, 16, 17) were shared among the 9 highly susceptible aus accessions. When the Eukaryotic Linear Motif resource (http://elm.eu.org) was used to identify functional sites in the Nrat1 gene, polymorphism 16 was identified as a functional site where a C→T SNP caused an amino acid change from valine→alanine (amino acid 500). This protein site was predicted to be involved in PKA-type AGC kinase phosphorylation, with the functional site spanning amino acids 497–503. Thus, polymorphism 16 was identified as a strong functional polymorphism candidate underlying natural variation in Nrat1. The fact that polymorphism 16 was also observed in two Al tolerant temperate japonica and one moderately tolerant tropical japonica accession (haplotype 11) suggested that SNP 16 alone was not predictive of Al tolerance. However, a combination of polymorphisms 4, 16, and 17 was entirely predictive of Al susceptibility.
This study demonstrates the power of whole genome association analysis to integrate divergent pieces of evidence from independent bi-parental and mutant studies, enabling us to associate gene-based diversity with germplasm resources and natural variation that is of immediate use to plant breeders.
There is a clear difference in the degree of Al tolerance found in the Japonica varietal group and the Indica varietal group, with the 10th percentile of Al tolerance of Japonica (0.53) being nearly equal to the 90th percentile of Indica (0.55) (Figure 1B). However, there are clear outliers within each varietal group. Five Indica accessions are highly Al tolerant (ID 30, 66, 142, 163, 337), ranging from 2.1–3.2 times the mean Indica Al tolerance, and three Japonica accessions (ID 12, 52, 112) are highly susceptible, each approximately 0.19 of the mean Japonica Al tolerance (Figure 1B and Table S1).
To determine if these outliers were the result of introgressions across varietal groups, we calculated the allele ancestry of 5,467 SNPs distributed throughout the genome and identified specific genomic regions where historical Indica×Japonica admixture was detected only in the respective Indica or Japonica outlier lines. To do this, Japonica introgressions identified in highly Al tolerant Indica lines were used to query all other Indica accessions and only those Japonica introgressions that were uniquely present in the highly Al tolerant outlier Indica lines were considered as candidate regions underlying the outlier phenotype. When the five Indica outliers were used for this analysis, a few, well-defined regions comprising 2.4–4.9% of the genome corresponded to regions of Japonica introgression (Table 3). In the case of the three highly Al susceptible Japonica varieties, the genetic background was highly heterogeneous and the small number of lines precluded doing any admixture analysis. Therefore, the admixture analysis was conducted only on the five highly tolerant Indica outliers.
In the five outlier Indica accessions, 6 Japonica introgressions (median size = 780 kb) were identified that were specific only to these 5 lines. Three of these introgressions were present in two genotypes, two of the introgressions were present in three genotypes, and one introgression was present in four of the outliers (Table 3). Three introgressions encompass SNPs identified by GWA analysis and two co-localized with bi-parental QTL. The introgression that was present in four of the indica outlier genotypes was located on chromosome 7 between 27.05–28.62 Mb and contained 94 annotated genes. This introgression included a cluster of GWA SNPs that were highly significant within the indica subpopulation (p = 2.6×10−5, MAF = 0.10) and was one of the top 100 most significant SNPs identified when the diversity panel as a whole was analyzed.
In this study, we utilized bi-parental QTL mapping and GWA analysis to examine the genetic architecture of Al tolerance in rice and to identify Al tolerance loci. Phenotyping of the diversity panel provided valuable information about the range and distribution of Al tolerance in O. sativa and offered new insights into the evolution of the trait. The mean Al tolerance in Japonica was twice that of Indica (p<0.0001), and 57% of the phenotypic variation was explained by subpopulation. The relative degree of Al tolerance in the five subpopulations (temperate japonica>tropical japonica>aromatic>indica = aus) was consistent with the level of genetic relatedness among them [42], [44] and suggests that temperate and tropical japonica germplasm contain alleles that would be useful sources of genetic variation for enhancing levels of Al tolerance within indica and aus. This is supported by the identification of highly tolerant indica varieties from the rice diversity panel that contain introgressions from Japonica in regions characterized by GWA peaks. The highly tolerant Indica outliers demonstrate the feasibility of using a targeted approach to increase Al tolerance in Indica varieties by introgressing genes from Japonica.
While less obvious, our QTL analysis demonstrated the ability to increase Al tolerance in Japonica using targeted introgressions from Indica. This was demonstrated within both QTL populations by the identification of two loci in which alleles from the highly susceptible Kasalath parent conferred enhanced levels of Al tolerance in the Nipponbare genome (temperate japonica) and one locus where the moderately susceptible IR64 parent conferred enhanced tolerance in crosses with Azucena (tropical japonica) (Table 1). To date, only a few indica and aus accessions have been used in QTL mapping populations and the identification of a large number of GWA loci in indica, coupled with the fact that indica is significantly more diverse than all other O. sativa subpopulations [40], [42] suggests that there are likely to be many novel alleles that could be mined from the indica subpopulation. Further evidence of the value of this approach in the context of plant breeding comes from the transgressive variation observed in both QTL populations, where some RILs and BILs exceeded the Al tolerance observed in the tolerant tropical and temperate japonica parents, Azucena and Nipponbare, respectively, due to alleles derived from the susceptible indica (IR64) or aus (Kasalath) parents, respectively.
The significant differences in Al tolerance among varietal groups and subpopulations, and evidence that different genes and/or alleles contribute to Al tolerance within the major varietal groups, is consistent with Indica and Japonica domestication from pre-differentiated, wild O. rufipogon gene pools that differed in Al tolerance. Future experiments will test this hypothesis by comparing levels of Al tolerance found in wild populations of O. rufipogon. The inherently higher levels of Al tolerance found in the Japonica varietal group may help explain why tropical japonica varieties are so often found in the acid soils of upland environments.
Compared to QTL mapping, GWA significantly increases the range of natural variation that can be surveyed in a single experiment and the number of significant regions that are likely to be identified. Furthermore, GWA provides higher resolution than QTL mapping, facilitating fine-mapping and gene discovery. This was illustrated by the two highly significant regions detected by GWA that overlapped with previously reported QTLs. GWA detected a highly significant cluster of 32 SNPs (p = 2.9E-07) on chr. 3 within the indica subpopulation, defining the candidate region to 81 kb window containing 13 genes, while the previously reported QTL interval was 1,720 kb [17], containing 260 genes. Similarly, the Nrat1 locus identified within the aus subpopulation on chromosome 2 initially narrowed the target region to 139 kb containing 27 genes by GWA, while the previously reported [26] QTL interval was 1,360 kb and contained 234 genes.
Surprising, the Nrat1 region was not significant in the BIL population, in which the resistant parent (Nipponbare) contained a resistant haplotype at Nrat1 and the susceptible parent (Kasalath) contained the susceptible haplotype at Nrat1. The fact that a significant signal was not detected in the BIL population can likely be explained by one or more of the following: 1) the bias inherent in the small population size (78 BILs), 2) the backcross population structure in which only 11 individuals (14% of BILs) contained the Kasalath allele at the Nrat1 locus and/or 3) the effects of genetic background on the Nrat1 QTL region. The Nrat1 QTL region was detected in one previous QTL study by Ma et al. [23] where a BIL population consisting of 183 lines was used, with Kasalath as the susceptible aus parent and Koshihikari as the tolerant temperate japonica parent [23]. In that study, the Nrat1 QTL region was of minor significance (LOD = 2.81; R2 = 7%), and it is noteworthy that the two other (more significant) QTLs detected in that study were the two QTLs detected in our BIL population using only 78 lines. The fact that the Nrat1 QTL region was not detected in our BIL mapping population and was of low significance in the Ma et al. QTL study suggests that the effect of the Kasalath allele is likely to be influenced by genetic background effects (GXG). In an aus genetic background, the Nrat1 susceptible haplotype explains 40% of the phenotypic variation, and the diversity panel contains enough aus varieties for this to be statistically significant using GWA; however, in the BIL population where Nipponbare served as the recurrent parent, the aus alleles exist in a largely temperate japonica background. Given the extent of GXG observed in inter-sub-population crosses, and the small size of our BIL population, this appears to be the most likely explanation as to why the Nrat locus was not detected in our QTL experiment.
Although GWA significantly increased the power and resolution of QTL detection, nearly all the significant loci detected were subpopulation-specific. This is entirely consistent with the strong subpopulation structure in rice and the high correlation of Al tolerance with subpopulation, justifying our GWA analysis on each subpopulation independently. So the question might be asked as to why it is also necessary to conduct GWA in the diversity panel as a whole? The answer to this question lies in the complex biology and demographic or breeding history of O. sativa. In this study GWA was conducted both within and across subpopulations, and it demonstrated that GWA on the diversity panel as a whole leveraged power to detect alleles that were segregating across multiple subpopulations, even if they were rare within any one subpopulation group, while when used on independent subpopulations, it was useful in detecting alleles that segregated only within one or two subpopulations but tended to be fixed in others. This is what would be expected from what we know about the evolutionary history of rice with its examples of shared domestication alleles [35], [75] coupled with myriad subpopulation-specific alleles [41], [48], [76]–[78] that provide each subpopulation with its specific identity and spectrum of ecological adaptations.
There are cases in which QTLs discovered by bi-parental mapping are not detected by GWA analysis. One reason for this is that QTL mapping can readily detect alleles that are rare in a diversity panel, are subpopulation-specific, or where the phase of the allelic association differs across subpopulations, while GWA analysis has limited power to do so. This is important in the case of rice, because of the degree of differentiation between the subpopulations and the significant evolutionary differences between the Indica and Japonica varietal groups, as discussed above. Thus, while variation that is strongly correlated with subpopulation structure is undetectable by GWA analysis, these loci can be easily detected by QTL analysis if crosses between sub-populations are used. This is illustrated by the identification of the Al tolerance QTL, (AltTRG12.1) encompassing the ART1 locus on chromosome 12. This large-effect QTL (LOD = 7.85, R2 = 0.193) was clearly detected in the RIL population but was not detected by GWA analysis. The QTL mapping populations utilized in this study were of limited population size and thus largely underpowered [79]. As a result it is likely that some QTL effects were overestimated and that other small effect QTL were not detected. Although we cannot be certain of the exact amount of variance explained by a particular QTL, it is reasonable to conclude that the major QTL detected (AltTRG12.1) is, in fact, the most significant QTL in the population.
GWA mapping also provides a valuable link between functional genomics and natural variation, and in the case of rice, highlights the subpopulation-specific distribution of specific alleles and phenotypes. We implicate the involvement of the STAR2 (chr. 6)/ALS3 (Arabidopsis Al sensitive mutant) gene, previously identified as induced mutations in rice and Arabidopsis, respectively [22], [23], and document the detection of highly resolved, novel Al tolerance loci in the indica and aus subpopulations. This is a critical bridge for germplasm managers and plant breeders who look for alleles of interest in germplasm collections rather than as sequences in GenBank.
Our strongest example of the value of linking functional genomics and natural variation is illustrated by the GWA region on chromosome 2, where we demonstrate that the aus-specific susceptible haplotype in this region is functionally related to an Nramp gene. This gene was previously identified to have altered expression in the art1 (transcription factor) Al sensitive mutant [19] and was recently reported as Nrat1 (for Nramp aluminum transporter), an Al transporter localized to the plasma membrane of root cells, which when knocked out, enhances Al susceptibility. This is consistent with this transporter serving to mediate Al uptake by moving it directly into root cells, presumably into the vacuole, and away from the root cell wall [20]. Our haplotype analysis of the GWA region on chromosome 2 and sequence analysis of the Nrat1 gene identified putative sensitive and tolerant haplotypes that implicate the Nrat1 gene, and further identified two putative functional polymorphisms specific to the Al sensitive aus accessions. These data provides valuable information for identifying Nrat1 alleles that can be used to test the hypothesis put forth by Xia et al. [20], namely that Al tolerance is conferred by reducing Al concentrations in the cell wall. It will be interesting to see if the sensitive alleles of this gene encode an Nramp transporter that is less effective at mediating Al uptake. Furthermore, the observation that three of the four most Al tolerant aus accessions contain tropical japonica introgressions across this gene region strongly suggests that Al tolerance of aus genotypes can be increased by the targeted introgression of tropical japonica DNA at the Nrat 1 region.
One of the objectives of this study was to determine if the Al tolerance index employed (longest root growth [LRG], primary root growth [PRG], or total root growth [TRG]) affected the detection and/or significance of Al tolerance QTL. In a recent publication from our research team, it was demonstrated that significantly different Al tolerance scores were obtained with the different indices [8]. In all previous QTL studies, Al tolerance was determined based on relative root growth (RRG) of the longest root. This study demonstrated that the Al tolerance index has a direct effect on the detection and significance of QTLs. Total root growth (TRG) was the single most powerful Al tolerance index, based on number of QTL detected, significance of QTL and variance explained by the QTL. However, it is relevant to point out that LRG-RRG identified a large-effect QTL (AltLRG9.1) in the RIL population that was not detected using any other index, and PRG-RRG identified a unique QTL on chromosome 6 where the susceptible Kasalath variety carried the resistance allele. These observations suggest that different root evaluation methods are likely to identify Al tolerance QTLs that confer tolerance mediated by different types of roots, or possibly by different patterns of gene expression detectable only when specific phenotypic evaluation protocols are used.
The strongest example of the importance of utilizing the TRG-RRG index is demonstrated by the identification of the AltTRG12.1 QTL in the RIL mapping population. The ART1 gene, a C2H2-type zinc finger-type transcription factor that causes Al hypersensitivity when mutated, is located close to the center of the Alt12.1 QTL peak. When this gene was first identified, it was suggested that it was not involved in natural variation of Al tolerance in rice, as no QTL had ever been identified in the region [19]. Based on our results, it is likely that this QTL was not previously identified because relative root growth was measured only based on LRG, rather than on TRG-RRG. Further fine-mapping of this locus, along with sequence and expression analysis, is underway to determine whether the ART1 locus underlies this QTL and to understand the mechanism by which it contributes to natural variation for Al tolerance.
Previous studies in other cereals have reported that the correlation of Al tolerance between hydroponics and field conditions is >70% [80] and studies on rice Al tolerance mutants have demonstrated that tolerance/susceptibility observed in hydroponics screens is also observed under soil conditions [24]. To accurately assess the value of the loci detected in this study as targets of selection in rice breeding programs, we are currently developing experiments to determine the effect of the key loci detected in this work under Al-toxic field conditions. Furthermore, four sets of reciprocal NILs (8 NILs total) for the four QTLs detected in the RIL population are being developed to determine the effect of each QTL under both hydroponic and field conditions. Finally, field experiments will be conducted to determine which hydroponic root measurement phenotype (TRG, PRG, or LRG) is the best for predicting a genotypes Al tolerance under field conditions.
This study provides the most comprehensive analysis of the genetic architecture of Al tolerance in rice to date. It demonstrates the power of whole genome association analysis to identify phenotype-genotype relationships and to integrate disparate pieces of evidence from QTL studies, mutant analysis, and candidate gene evaluation into a coherent set of hypotheses about the genes and genomic regions underlying quantitative variation. By tracing the origin of Al tolerance alleles within and between rice subpopulations, we provide new insights into the evolution and combinatorial potential of different alleles that will be invaluable in breeding new varieties for acid soil environments. This work demonstrates how genetic and phenotypic diversity is partitioned by subpopulation in O. sativa and provides support for the hypothesis that the most efficient approach to enhancing many quantitative traits in rice is to selectively introgress genes/alleles from one subpopulation into another. Our study also lays the foundation for understanding the genetic basis of Al tolerance mechanisms that enable rice to withstand significantly higher levels of Al than do other cereals. It not only facilitates more efficient selection of tolerant genotypes of rice, but it points the way toward using this knowledge to enhance levels of Al tolerance in other plant species.
Plants were grown hydroponically in a growth chamber as described by Famoso et al. [8]. Al tolerance was determined based on relative root growth (RRG) after three days in Al (160 µM Al3+) or control solution. The hydroponic solution used in this study was chemically designed and optimized for rice Al tolerance screening; for a detailed comparison of the phenotypic procedures employed in this work compared to previously published rice Al tolerance work see Famoso et al. (2010). To obtain uniform seedlings, 80 seeds were germinated and the 30 most uniform seedlings were visually selected and transferred to a control hydroponic solution for a 24 hour adjustment period. After the 24 hour adjustment period, root length was measured with a ruler and the 20 most uniform seedlings were selected and distributed to fresh control solution (0 uM Al3+) or Al treatment solution (160 uM Al3+). Plants were grown in their respective treatments for ∼72 hours and the total root system growth was quantified using an imaging and root quantification system as described by Famoso et al. (2010). The mean total root growth was calculated for Al treated and control plants and RRG was calculated as mean growth (Al)/mean growth (control). The 373 genotypes screened for Al tolerance and used in the association analysis are part of a set of 400 O. sativa genotypes that have been genotyped with 44,000 SNPs as described by Zhao et al. [65].
The QTL populations consisted of a population of 134 recombinant inbred lines (RILs) derived from a cross between Azucena (tolerant tropical japonica) and IR64 (susceptible indica) [67], [70] and a population of 78 backcross introgression lines (BILs) derived from a cross between Nipponbare (tolerant temperate japonica) and Kasalath (susceptible aus) and backcrossed to Nipponbare. The Al3+ activity at which Al tolerance was screened was determined by identifying the Al3+ activity that provided the greatest difference in tolerance between the parents. The tolerant parent of the RIL population, Azucena, and the tolerant parent of the BIL population, Nipponbare, are similar in Al tolerance, whereas the susceptible parent of the RIL population, IR64, is significantly more tolerant than the susceptible parent of the BIL population, Kasalath (Figure 1A). To ensure that a normal distribution was obtained in each population, a different Al3+ concentration was used for each mapping population. The RIL population was screened at 250 µM Al3+ because the Azucena parent is very Al tolerant and the IR64 parent is only moderately susceptible. The BIL population was screened at 120 µM Al3+ because the Kasalath parent is extremely Al sensitive, though the Nipponbare parent is very Al tolerant. Figure 1 displays the Al tolerance of each mapping parent in reference to the 373 genetically diverse rice accessions screened at 160 µM Al3+. The genetic component of the phenotypic variance was calculated as VarG = VarG+Var(GxE)+error. QTL analysis was conducted using composite interval mapping (CIM) function in QTL Cartographer [81]. The significance threshold was determined by 1000 permutations.
Genome-Wide Association Analysis was performed using three approaches in all samples (373) with phenotypes. The first approach was the naïve approach, which is simply the linear regression of phenotype on the genotype for each SNP marker. The second approach was principle component analysis (PCA), where we obtained the four main PCs (principle components) that reflect the global main subpopulations in the sample to correct population structure estimated from software EIGENSOFT. [82]. The first four PCs are included as cofactors in the regression model to correct population structure: .
Here β and γ are coefficient vectors for SNP effects and subpopulation PCs respectively. and are the corresponding SNP vector and first 4 PC vectors, and is the random error term. The third approach was the linear mixed model proposed by [62], [63], implemented in the R package EMMA [71], which models the different levels of population structure and relatedness. The model can be written in a matrix form as: y = Xβ+Cγ+Zμ+e where β and γ are the same as above, both of which are fixed effects, and is the random effect accounting for structures and relatedness, is corresponding design matrices, and is the random error term. Assume μ∼N(0,σ2gK) and e∼N(0,σ2eI), and K is the IBS matrix, as in [62]. We also conducted GWA using both the naïve approach and the mixed model approach in each of the four main subpopulations (IND, AUS, TEJ, TRJ). For the mixed model, the model was changed to y = Xβ+Zu+e, since there was no main subpopulation division within each subpopulation sample. Linkage disequilibrium decay and haploblocks were calculated at specific chromosome/gene regions using Haploview software [83].
Population structure was analyzed employing Expectation-Maximization techniques on an HMM model of per-marker ancestry along a chromosome with a weak linkage model between adjacent markers on the same chromosome induced by the HMM's state dependence on the previous marker's subpopulation assignment (M. Wright, Cornell University, personal communication). The 5,467 SNPs used for admixture analysis were a subset of the 36,901 high quality SNPs on the 44 K chip, and were selected based on their information content and ability to distinguish genetic groups, rather than individuals. The two main criteria used to select the subset of SNPs were a) good genomic distribution and minimal LD among those used in the analysis, and b) MAF>0.05 in at least one subpopulation. The state of the HMM at each marker corresponds to the subpopulation of origin for the marker (and by extension, the region containing the marker and its adjacent markers). The number of a priori distinct subpopulations was K = 5, consistent with that reported previously by Garris et al. 2005 and Ali et al., 2011 [40], [66]. A set of 50 standard non-admixed “control” lines, 10 representing each of the Garris et al. subpopulations, that were genotyped on the 44 K rice SNP array were used to develop and evaluate the method. All 50 lines were correctly assigned to each of the subpopulations and concordant with previous results using STRUCTURE [84], with little or no admixture or introgressions detected. The EM/HMM method was favored over the corresponding “linkage model” of recent versions of STRUCTURE because the EM/HMM model explicitly modeled inbreeding and estimated the inbreeding coefficient for each line independently, permitting lines in various stages of purification or inbreeding to homozygosity to be analyzed. The lines phenotyped in this study that were also genotyped on the 44 K SNP array were then analyzed, combined with these 50 control lines and the local ancestry along chromosomes were assigned by maximizing the state path of the HMM while simultaneously estimating subpopulation specific allele frequencies using the forward-backward algorithm. Using this method, introgressions from a foreign subpopulation into a line with a vast majority of the genetic background originating from a single subpopulation were detected.
|
10.1371/journal.pgen.1002341 | Association of NCF2, IKZF1, IRF8, IFIH1, and TYK2 with Systemic Lupus Erythematosus | Systemic lupus erythematosus (SLE) is a complex trait characterised by the production of a range of auto-antibodies and a diverse set of clinical phenotypes. Currently, ∼8% of the genetic contribution to SLE in Europeans is known, following publication of several moderate-sized genome-wide (GW) association studies, which identified loci with a strong effect (OR>1.3). In order to identify additional genes contributing to SLE susceptibility, we conducted a replication study in a UK dataset (870 cases, 5,551 controls) of 23 variants that showed moderate-risk for lupus in previous studies. Association analysis in the UK dataset and subsequent meta-analysis with the published data identified five SLE susceptibility genes reaching genome-wide levels of significance (Pcomb<5×10−8): NCF2 (Pcomb = 2.87×10−11), IKZF1 (Pcomb = 2.33×10−9), IRF8 (Pcomb = 1.24×10−8), IFIH1 (Pcomb = 1.63×10−8), and TYK2 (Pcomb = 3.88×10−8). Each of the five new loci identified here can be mapped into interferon signalling pathways, which are known to play a key role in the pathogenesis of SLE. These results increase the number of established susceptibility genes for lupus to ∼30 and validate the importance of using large datasets to confirm associations of loci which moderately increase the risk for disease.
| Genome-wide association studies have revolutionised our ability to identify common susceptibility alleles for systemic lupus erythematosus (SLE). In complex diseases such as SLE, where many different genes make a modest contribution to disease susceptibility, it is necessary to perform large-scale association studies to combine results from several datasets, to have sufficient power to identify highly significant novel loci (P<5×10−8). Using a large SLE collection of 870 UK SLE cases and 5,551 UK unaffected individuals, we firstly replicated ten moderate-risk alleles (P<0.05) from a US–Swedish study of 3,273 SLE cases and 12,188 healthy controls. Combining our results with the US-Swedish data identified five new loci, which crossed the level for genome-wide significance: NCF2 (neutrophil cytosolic factor 2), IKZF1 (Ikaros family zinc-finger 1), IRF8 (interferon regulatory factor 8), IFIH1 (interferon-induced helicase C domain-containing protein 1), and TYK2 (tyrosine kinase 2). Each of these five genes regulates a different aspect of the immune response and contributes to the production of type-I and type-II interferons. Although further studies will be required to identify the causal alleles within these loci, the confirmation of five new susceptibility genes for lupus makes a significant step forward in our understanding of the genetic contribution to SLE.
| Systemic lupus erythematosus (SLE) is a relapsing-remitting complex trait which most commonly affects women of child-bearing age, with a ratio of 9∶1 in female to males. The disease prevalence varies with ethnicity, being more prevalent in non-European populations (approximately 1∶500 in populations with African ancestry and 1∶2500 in Northern Europeans) [1]. The condition is characterised by the production of a diverse range of auto-antibodies against serological, intra-cellular, nucleic acid and cell surface antigens [2]. The wide-ranging clinical phenotypes include skin rash, neuropsychiatric and musculosketal symptoms and lupus nephritis, which may be partially mediated by the extensive deposition of immune complexes. Today, thanks to improved treatments, the 10-year survival rate after diagnosis has increased to 90%, with lower survival rates being related to disease severity or complications from treatment [3]. Increased understanding of the underlying genetic basis for lupus is of key importance in improving the prognosis for lupus patients.
Until recently, the genetic basis of lupus remained largely undetermined, with only about ∼8% of the genetic contribution known [4]. However, within the last three years, tremendous progress has been made in defining novel loci, through three moderate-sized genome-wide association studies in European American cohorts and a replication study in a US-Swedish cohort [5]–[7]. The loci previously identified for SLE include genes involved in the innate immune response (eg. IRF5), T and B cell signalling (eg. STAT4, TNFSF4 and BLK), autophagy/apoptosis (eg. ATG5), ubiquitinylation (UBE2L3, TNAIP3, TNIP1) and phagocytosis (ITGAM, FCGR3A and FCGR3B). All of these pathways are of potential importance in lupus pathogenesis [8]–[10].
To date, a total of 1729 independent SLE cases have been subjected to genome-wide association genotyping using three genotyping platforms: Illumina 317 K BeadChip [5], Illumina 550 K BeadChip [6] and Affymetrix 500 K array [7]. There is currently no published meta-analysis of these datasets.
The aim of the current work was to perform a replication study using our UK SLE cohort on loci that showed some evidence for association in previous studies in order to extend the list of confirmed susceptibility genes for lupus.
To identify additional susceptibility loci for SLE, we first identified the independent genetic variants that showed moderate risk (5×10−3<P>5×10−8) in a combined US-Swedish dataset comprising 3273 SLE cases and 12188 controls [4]. We then genotyped 27 independent SNPs in a replication cohort of 905 UK SLE cases and 5551 UK control samples (Table 1), that included both British 1958 Birth Cohort samples and additional controls from the WTCCC2 project.
For the 27 genotyped SNPs, 10 variants which had not been genotyped by the WTCCC2 project, were imputed using IMPUTE2 [11]. This imputation was performed using CEPH HapMap samples as the phased reference sequence and the boundary of the surrounding haplotype blocks used to demarcate the imputation interval. The subsequent association analysis excluded two of these ten imputed SNPs because they had less than 95% certainty for the imputation (Table S2). In the US/SWE dataset, imputation of selected SNPs not genotyped previously [4] was performed using IMPUTE1 for HapMap. Phase II CEU sample haplotypes were used as reference with subsequent association analysis performed using SNPTEST and a genomic control factor (lambda-GC) values of: 1.05 (US dataset) and 1.10 (SWE dataset) after correction for population stratification.
In the UK replication sample by performing allelic association analysis using PLINK for the 23 SNPs passing QC (Tables S2 and S3), we demonstrated moderate association (P≤0.05) for twelve variants - with a lambda-GC of 1.01 following ancestry correction (see Table 2 and Table 3). Under the null hypothesis, only 1 of the 23 loci would be expected to have P≤0.05. The observed enrichment of associated SLE genes in the UK dataset suggested that many of these loci were likely to be true-positive associations.
We confirmed the similarity of odds-ratios (Het P value) and direction of the effect between the UK and US-SWE datasets (Table S4) and then performed a meta-analysis using Fisher's combined P-value (see Materials and Methods). This meta-analysis revealed five novel associated loci with P<5×10−8 (Table 2): NCF2 (neutrophil cytosolic factor 2) (rs10911363, Pcomb = 2.87×10−11, ORcomb = 1.19); IKZF1 (Ikaros family zinc-finger 1) (rs2366293, Pcomb = 2.33×10−9, ORcomb = 1.24); IRF8 (interferon regulatory factor 8) (rs2280381, Pcomb = 1.24×10−8, ORcomb = 1.16); IFIH1 (interferon-induced helicase C domain-containing protein 1) (rs1990760, Pcomb = 1.63×10−8, ORcomb = 1.15) and TYK2 (tyrosine kinase 2) (rs280519, Pcomb = 3.88×10−8, ORcomb = 1.17)(Table 1). The strength of these associations was similar to those found from a weighted meta-analysis, using the METAL programme (Table S4). A case-only analysis using PLINK in the combined UK/US/SWE dataset revealed no non-additive interactions between the five newly associated variants (P>0.05). These new SLE loci are discussed in more detail below and with additional information in Text S1.
Three of the SNPs tested were for loci that had shown genome-wide levels of significance in other SLE GWAS studies (Table S5). In the UK cohort we found further support for the association at JAZF1 (rs849142 PUK = 0.0243, ORUK = 1.13) and identified a third associated variant in the first intron of TNIP1 (rs6889239 PUK = 9.06×10−6, ORUK = 1.30), which is in strong LD (r2 = 0.895) with both the previous report in Europeans [4] and in perfect LD with a third SNP (rs10036748), first reported in a Chinese GWAS [12]. All three variants in TNIP1 are located within a 661 bp region of intron 1. We did not replicate the previous association with IL10 (rs3024505, PUK = 0.209 ORUK = 1.09) (Table S5).
These analyses increased the evidence of association for a number of additional loci that had shown borderline significance in the original US/SWE GWAS (Table 3), including CFB, C12ORF30, SH2B3, and IL12B. Genotyping of additional samples will be required to determine if the association signals shown in Table 3 represent confirmed genetic loci for SLE.
The work presented here confirms five new susceptibility loci for SLE at the level of genome-wide significance (P<5×10−8). Each of the associated variants lie within, or close to, the coding sequence for genes with known roles in immune regulation: NCF2, IKZF1, IRF8, IFIH1 and TYK2. Interestingly, each of these genes has been implicated in interferon signalling. While the interferons have classically been defined as anti-viral cytokines, recent studies have suggested an important role for interferon in the pathophysiology of SLE [13]. While most evidence points to the role of type I interferon in SLE [14] there is substantial data suggesting that type II interferon (IFNγ) is also involved in SLE pathogenesis [15].
NCF2 (neutrophil cytosolic factor 2) (1q25), is induced by IFNγ and specifically expressed in a number of immune-cell types, including B-cells. Our data suggest that the NCF2 association is independent from the previously reported signal in the neighbouring locus NMNAT2, [5] because we found no evidence of strong LD between the genotyped SNP within NMNAT2 (rs2022013) and that in NCF2 (rs10911363) (r2 = 0.136). Logistic regression in the UK replication cohort confirmed that NMNAT2 did not contribute to the association at NCF2 (P = 0.777).
NCF2, as a cytosolic subunit of NADPH-oxidase, may have a role in the increased production of the free radicals characterising B-cell activation [16] (Figure 1) which increases auto-antibody levels and may suggest a mechanism for the involvement of NCF2 as a susceptibility gene for SLE.
There are allele-specific significant expression differences for rs10911363, following a recessive model of basal expression for the risk T allele of rs10911363 in CEPH individuals but not in YRI and ASN (CHB+JPT) HapMap cohorts (PCEPH = 0.03) (Figure 2A). There is also a significant difference in gene expression for a variant (rs3845466) located 2 kb away from rs10911363 in intron 2 of NCF2 (Figure S2A), using lymphoblastoid cell lines (LCLs) from umbilical cords of 75 individuals which were taken from the GENEVAR collection (P = 0.0228). The population-specific nature of this correlation could be because of local differences in the pattern of LD within NCF2 between the CEU, YRI and ASN (CHB+JPT) HapMap cohorts. These population specific differences in LD may be between the genotyped SNP and an unknown causal allele(s) responsible for an expression difference seen in multiple ethnic backgrounds or between the genotyped marker and an unknown causal allele(s) exhibiting population-specific differences in gene expression itself. However, it will be necessary to confirm these findings in primary cells and tissues, because the EBV-transformed B cells model system may not entirely reflect the physiological conditions in peripheral B cells. Indeed a recent report showed that there may be systematic changes in gene expression within EBV-transformed B cells [17]. Nevertheless, with this caveat in mind, and taking each locus on a case-by-case basis, the model-based approach can provide important insights into measurement of transcript levels in ex vivo cells. For example, the increases in transcript levels that we initially observed in EBV-LCLs for OX40L, were also confirmed in peripheral blood B cells [18].
IKZF1 (Ikaros family zinc-finger 1) (17p14.3) is a transcription factor essential for dendritic cell and lymphocyte development. The association with rs2366293 is supported by a report of a second associated variant, rs921916 (Pcomb = 2.0×10−6) [4], found 860 bp away from rs2362293, which is in strong LD with rs2366293 (r2 = −0.746, D′ = 0.925) (Figure S2B). A third SNP, rs4917014, located ∼200 kb upstream of IKZF1, showed association with SLE in a Chinese GWAS (PGWAS = 2.93×10−06), but it was a separate signal from the European SNPs (r2<0.0002) [9], [12]. IKZF1 has a role in the production of IFNγ, by blocking the production of the Th1 master-regulator T-bet (Figure 1). The shifted Th1/Th2 equilibrium (in favour of Th1 cells) increases the levels of IFNγ directly [19] rather than indirectly as a result of cross-talk between the type-I and type-II IFN signalling pathways eg) via type-I interferon mediated activation of STAT1 homodimers, which are the primary means of signalling from IFNγ [20] and have recently been shown to be associated with SLE in a Swedish cohort [21].
The transcription factor IRF8 (interferon regulatory factor 8) (16q24.1), shows immune-cell restricted expression. rs2280381 is found 64 kb downstream of IRF8, and is in LD with the coding region (Figure S2C), but independent from a susceptibility allele for multiple sclerosis (rs17445836), 1 kb away [22]. The lupus variant influences IRF8 gene expression, since LCLs from three HapMap cohorts, showed a significant increase in IRF8 transcript levels in homozygotes for the risk allele (TT) compared to homozygotes for the non-risk allele (CC) (P = 0.045) (Figure 2A). IRF8 also has a key role in regulating the differentiation of myeloid and B-cells and in mice, IRF8 restricts myeloid cell differentiation but promotes B-cell differentiation [23](Figure 1).
IFIH1 (interferon-induced helicase C domain-containing protein 1) (2q24.3) is an ubiquitiously expressed, cytoplasmic sensor of dsRNA. The SLE risk allele for rs1990760 (Table 1) is identical to that previously reported in two organ-specific autoimmune diseases: T1D [24] and Graves' Disease [25]. Regression analysis using publically available genotype data from HapMap and expression data from GEO dataset GSE12526 revealed that individuals who were homozygous for the common risk T allele of rs1990760 had significantly higher IFIH1 transcript levels compared to individuals who were homozygous for the non-risk allele (P = 0.8.19×10−5) (Figure S3B). Furthermore, a recent paper showed that the presence of the risk T allele of rs1990760 was correlated with increased levels of IFN-induced gene expression, in lupus patients who were positive for anti-dsDNA antibodies [26]. Another report demonstrated that IFIHI was rapidly up-regulated by type-I IFNs (Figure 1), and that IFIH1 signalled downstream through NF-κB, to further increase IFN-α production [27].
TYK2 (tyrosine kinase 2) (19p13.2) phosphorylates the receptor subunits of cytokine receptors, including type-I IFN receptors which are found on all nucleated cells, leading to increased production of type I interferon responsive genes (Figure 1). The significant association in intron 11 TYK2 for rs280519 in our UK cohort (P = 5.24×10−4) crossed the threshold for genome-wide significance when combined with the US/Swedish cohort. The association for rs280519 increases the genetic evidence for the involvement of TYK2 reported in a smaller UK family-based SLE cohort [28]. There was an earlier report, using a Swedish/Finnish population, of association in TYK2. This Swedish/Finnish study showed association for a missense mutation in exon 8 (rs2304256) (Pcomb = 5.60×10−5, PSwe = 9.60×10−5) [29]. The Swedish individuals used in the earlier analysis are a subset of the Swedish individuals analysed for this current manuscript and rs2304526 is in moderate LD with the TYK2 SNP that we typed in this current study - rs280519 (r2CEPH-HapMap = 0.373). The association for rs2304256 was replicated in a second moderate sized European study [30], but not in the GWAS from the SLEGEN consortium [5]. In preliminary analysis in UK cases and controls, there are data to support the fact that rs280519 is enriched in SLE cases (n = 345) with renal disease compared to healthy controls (n = 5551) (P = 0.033).
There were variants in several loci for which we have found evidence of association (P<0.05) in our UK cohort, but which did not reach genome-wide significance in the combined analysis. One of these variants was rs17696736, located in intron 15 of C12ORF30 (MDM20). This protein is a subunit of N-acetyltransferase complex B (NatB), and may promote apoptosis by reducing cell cycle progression [31]. In the joint cohort, rs17696736 was in LD (r2 = 0.625) with a second variant on chromosome 12q24, a missense W262R allele (rs3184504) in the lymphocyte adaptor protein SH2B3. SH2B3 facilitates T-cell activation by mediating the interaction between the T-cell receptor and T cell signalling molecules [32]. Both MDM20 and SH2B3 are also associated with T1D [33], and SH2B3 is additionally associated with celiac disease [34] and both myocardial infarction and asthma [35]. The associated variant within IL12B, rs3212227, is located in the 3′ UTR region, and the SLE risk allele is the same as previously reported for psoriasis [36]. IL12B encodes for the larger subunit (p40) of two cytokines, IL12 and IL23, and thereby contributes to both Th1 [37] and Th17 [38] immune responses.
In summary, we have identified five new genes contributing to SLE risk: NCF2, IKZF1, IRF8, IFIH1 and TYK2. Dense fine-mapping and/or genomic re-sequencing of each locus will be required to reveal the functional alleles for each gene with respect to immune dysregulation in lupus. Taken together, these findings further support an important role of interferon pathway dysregulation in lupus pathogenesis.
The ethical approval for the study was obtained from the London Multi-Centre Research Ethics Committee (London MREC).
All of the 905 UK SLE cases conformed to the ACR criteria for SLE [39] with a diagnosis of SLE being established by telephone interview, health questionnaire and details from clinical notes. Written consent was obtained from all participants. Genomic DNA from the UK samples was isolated from anti-coagulated whole blood by a standard phenol-chloroform extraction.
Each of the 27 SNPs were genotyped on a custom Illumina chip, using the BeadXpress platform at the Oklahoma Medical Research Foundation (OMRF), Oklahoma. The panel of ancestry informative markers was typed independently on an Illumina platform at Gen-Probe, Livingstone.
Power calculations were performed in the UK case-control dataset for each of the markers tested, using the algorithm described by Purcell et al [40]. Taking into account varying minor allele frequencies for the risk alleles and the differences in effect size (OR), and by employing a population prevalence of 0.002 and D′ of 1, with an type I error rates, alpha = 0.05, each of the SNPs showing novel genome-wide significance in the meta-analysis showed a power of >48% (2) to detect an association in our cohort.
Markers were excluded from the analysis if they showed a genotyping success rate of less than 95% or had a Hardy-Weinberg P value in the B58BCC control samples of less than P = 0.001. A total of 21 cases were removed from the final analysis due to low percentage genotyping (<95%). All samples were filtered for cryptic relatedness and duplication using an identity by state test in PLINK (PI_HAT score >0.1). The full list of genotyped variants and the results of the QC analysis are shown in (Table S3).
A total of 35887 markers, distributed across each autosome, were selected for ancestry correction in the UK case-control cohort, these markers had all been typed as part of the HapMap project and on the WTCCC2 samples. The 35887 SNPs were chosen from a set of Illumina 317 K markers pruned for LD (r2<0.25) after removing regions of known extended LD, including the extended MHC and the region covering the inverted repeat on chromosome 8 (pers commun. David Morris, King's College and Kim Taylor, UCSF). This list of AIMs is available directly from the corresponding author, Professor Timothy Vyse.
The EIGENSTRAT PCA analysis was performed on the UK cases and also the control samples, both from the genotyped B58BCC and the WTCCC2 out-of-study controls. The eleven populations from HapMap3 were used as external references. Each SNP included in the PCA analysis showed >95% genotyping in the each dataset. Following EIGENSTRAT analysis, a graph was plotted of PC1 against PC2 for all the cases and controls in the UK study cohort (Figure S1). Individuals were only retained for association analysis if the values for their first two principal components fell within 6 SD of the mean for the CEPH HapMap samples. The genomic inflation factor (lambda-GC) for each population was calculated using PLINK.
All sample genotype and phenotype data was managed by, and analysis files generated with BC/SNPmax and BC/CLIN software (Biocomputing Platforms Ltd, Finland).
The imputation intervals for each imputed variant, defined as the bounds of the haplotype blocks, calculated using the Gabriel algorithm in Haploview, (for details of the intervals see Table S2). For SNPs which were not genotyped as part of the WTCCC2 project, we performed imputation using a method described by Marchini et al [11] to generate the missing genotypes for case-control association analysis. Each un-typed variant from our list of tested SNPs, was imputed in the WTCCC2 samples, using HAPMAP as the phased reference sequence. The LD pattern around each un-typed variant was examined using the CEPH cohort from HapMap. The boundaries of the haplotype blocks were determined using the default settings for the Gabriel et al algorithm in Haploview. For each imputed variant, these haplotype boundaries were used to define the boundaries of the imputation interval (Table S2). Only SNPs with greater than a 95% certainty in imputation, assessed using the quality score from the IMPUTE2 output file, were used for subsequent analysis.
Allelic association testing, using UK SLE cases with either genotyped control samples or imputed genotypes, was carried out using PLINK (http://pngu.mgh.harvard.edu/~purcell/plink/).
Prior to performing the meta-analysis, the heterogeneity of odds ratios was tested using METAL and the Cochran-Mantel-Haenszel test (Table S4). SNPs with P value<0.001 between the two studies were discarded. Combined analysis of P values generated in the UK samples with those from the US/SWE cohort in published data [4] was conducted using Fisher's combined P value and with a meta-analysis using the programme METAL, which weighted the effect size, based on the inverse of the standard error.
To determine whether there was any allele-specific effect on the level of gene expression, we used publically available genotype data on unrelated EBV-transformed B cells (CEU, YRI and CHB/JPT individuals which were part of the HapMap project) and expression data from the same individuals (GSE12526, GEO database) [41]. For each locus, which reached genome-wide significance by meta-analysis, we categorised the expression data based on the SNP genotype for the respective associated variant (homozygote risk allele, heterozygote and homozygous non-risk allele). The significance of the correlation between genotype and expression level was then calculated using logistic regression analysis in SNPTEST, using gender as a covariate.
Interactions between the five SNPs reaching genome-wide significance following meta-analysis, were assessed using the epistatic option in PLINK. To maximize the power of this test, we restricted our analysis to the SLE affected individuals from the combined US/SWE/UK cohort.
|
10.1371/journal.pgen.1004647 | GLD-4-Mediated Translational Activation Regulates the Size of the Proliferative Germ Cell Pool in the Adult C. elegans Germ Line | To avoid organ dysfunction as a consequence of tissue diminution or tumorous growth, a tight balance between cell proliferation and differentiation is maintained in metazoans. However, cell-intrinsic gene expression mechanisms controlling adult tissue homeostasis remain poorly understood. By focusing on the adult Caenorhabditis elegans reproductive tissue, we show that translational activation of mRNAs is a fundamental mechanism to maintain tissue homeostasis. Our genetic experiments identified the Trf4/5-type cytoplasmic poly(A) polymerase (cytoPAP) GLD-4 and its enzymatic activator GLS-1 to perform a dual role in regulating the size of the proliferative zone. Consistent with a ubiquitous expression of GLD-4 cytoPAP in proliferative germ cells, its genetic activity is required to maintain a robust proliferative adult germ cell pool, presumably by regulating many mRNA targets encoding proliferation-promoting factors. Based on translational reporters and endogenous protein expression analyses, we found that gld-4 activity promotes GLP-1/Notch receptor expression, an essential factor of continued germ cell proliferation. RNA-protein interaction assays documented also a physical association of the GLD-4/GLS-1 cytoPAP complex with glp-1 mRNA, and ribosomal fractionation studies established that GLD-4 cytoPAP activity facilitates translational efficiency of glp-1 mRNA. Moreover, we found that in proliferative cells the differentiation-promoting factor, GLD-2 cytoPAP, is translationally repressed by the stem cell factor and PUF-type RNA-binding protein, FBF. This suggests that cytoPAP-mediated translational activation of proliferation-promoting factors, paired with PUF-mediated translational repression of differentiation factors, forms a translational control circuit that expands the proliferative germ cell pool. Our additional genetic experiments uncovered that the GLD-4/GLS-1 cytoPAP complex promotes also differentiation, forming a redundant translational circuit with GLD-2 cytoPAP and the translational repressor GLD-1 to restrict proliferation. Together with previous findings, our combined data reveals two interconnected translational activation/repression circuitries of broadly conserved RNA regulators that maintain the balance between adult germ cell proliferation and differentiation.
| Throughout adulthood, animal tissue homeostasis requires adult stem cell activities. A tight balance between self-renewal and differentiation protects against tissue overgrowth or loss. This balance is strongly influenced by niche-mediated signaling pathways that primarily trigger a transcriptional response in stem cells to promote self-renewal/proliferation. However, the cell-intrinsic mechanisms that modulate signaling pathways to promote proliferation or differentiation are poorly understood. Recently, post-transcriptional mRNA regulation emerged in diverse germline stem cell systems as an important gene expression mechanism, primarily preventing the protein synthesis of factors that promote the switch to differentiation. In the adult C. elegans germ line, this study finds that the evolutionarily conserved cytoplasmic poly(A) polymerase, GLD-4, plays an crucial role in maintaining a healthy balance between proliferation and differentiation forces. This is in part due to translational activation of the mRNA that encodes the germ cell-expressed Notch signaling receptor, an essential regulator of proliferation. Moreover, GLD-4 activity is part of a redundant genetic network downstream of Notch that, together with several other conserved mRNA regulators, promotes differentiation onset. Given the widespread expression of these conserved RNA regulators in metazoans, cell fate balances that are reinforced by translational activation and repression circuitries may therefore be a general mechanism of adult tissue maintenance.
| During development, tissues grow to form functional organs. In adulthood, animal tissues remain constant in size, in part, as a result of the dynamic balance between self-renewal/proliferation and differentiation. Perturbation of this balance affects tissue homeostasis and, consequently, compromises organ function. While excess proliferation contributes to tumorigenesis, a deficit in proliferation leads to tissue degeneration. Hence, tight regulatory mechanisms are in place to control the balance between self-renewal/proliferation and differentiation. One prevalent cell-extrinsic regulatory mechanism of stem cells to self-renew/proliferate is their dependency on supporting niche cells, which trigger established signal transduction pathways that primarily lead to changes at the transcriptional level. However, to fine-tune proper tissue homeostasis and to provide tight feedback controls, additional cell-intrinsic gene expression mechanisms are likely to exist.
In recent years, invertebrate germline tissues emerged as powerful in vivo models to investigate the balance between proliferation and differentiation. One influential paradigm is the adult “female” germ line of C. elegans, which depends on a single somatic niche cell and maintains a strict spatio-temporal organization of proliferating and differentiating cells [1]. Undifferentiated germ cells proliferate exclusively in the distal end of the germ line, termed the proliferative zone (PZ) [2], [3], [4]. The PZ is proposed to contain a distal pool of germline stem cell-like cells (GSCs) and a proximal pool of transit amplifying cells that gradually mature to start differentiation at a defined distance from the distal end [1], [5], termed the mitosis-to-meiosis boundary. Germ cells crossing this boundary enter meiotic prophase, which is here defined as differentiation onset [2], [6], [7]. Germline proliferation relies on the Notch signaling pathway that is instructed by the somatic distal tip cell (DTC) [8], [9], [10]. Consistent with its continuous requirement for germ cell proliferation in the adult, the inactivation of Notch signaling leads to progressive loss of GSCs, due to differentiation of all germ cells [9]. Conversely, constitutive activation of Notch results in the expansion of the proliferative GSC pool at the expense of differentiation [11], [12]. In agreement with this, germ cells in the PZ express the Notch receptor GLP-1, while differentiating cells lose GLP-1 expression [13]. Hence, Notch-mediated transcriptional regulation of mitotic fate-promoting genes is suggested to directly maintain the proliferative fate [14], [15], [16]. However, germ cell-intrinsic mechanisms that promote niche-mediated germ cell proliferation are still widely unknown.
In nematodes and flies, germ cells also utilize conserved translational repressors to promote the undifferentiated state [17]. In C. elegans, two nearly identical translational repressors of the PUF RNA-binding protein family, FBF-1 and FBF-2, jointly referred to as FBF, are essential for adult GSCs [18]. FBF recognizes specific sequence elements (FBEs) in its mRNA targets and, by translationally repressing numerous meiosis-promoting genes, FBF is critical for maintaining the undifferentiated, proliferative state [7], [18], [19], [20]. Moreover, the fbf-2 locus is a proposed target of Notch-mediated regulation [14], [21], thus linking transcriptional activation with translational repression, the two dominant mechanisms used for sustained germ cell proliferation in different organisms [1].
Across species, differentiation onset of germ cells depends on translational control [17]. In nematodes, the STAR-type RNA-binding protein, GLD-1, inhibits GLP-1 protein accumulation [22], [23] and recognizes glp-1 mRNA by three GLD-1-binding motifs (GBMs) present in its 3′UTR [24], [25]. Meiotic prophase entry also requires the Nanos protein family member, NOS-3, a presumed translational repressor of yet unknown mitosis-promoting genes [21]. However, in the absence of GLD-1, NOS-3, or both, germ cells enter meiotic prophase in the adult [7], [26]. The cytoplasmic poly(A) polymerase (cytoPAP) complex GLD-2/GLD-3 is a proposed translational activator of meiosis-promoting mRNAs, envisioned to extend their poly(A) tail lengths. GLD-2 is a non-canonical nucleotidyltransferase, stimulated by the Bicaudal-C family member, GLD-3 [27], [28]. However, in the absence of GLD-2, GLD-3, or both, the PZ is expanded but meiosis is still initiated [7]. This complexity highlights that differentiation onset is in general a multi-pathway-regulated process [17].
In the current model of the core genetic network underlying differentiation onset in C. elegans, the four meiosis-promoting RNA regulators act in two parallel pathways. The two translational repressors (gld-1 and nos-3) form the first pathway; the two translational activators (gld-2 and gld-3) form the second pathway. This genetic redundancy is most apparent in germ cells that lack GLD-3 and NOS-3, as they do not enter meiotic prophase and continue to proliferate [7]. Importantly, tumorous proliferation of gld-3 nos-3 double mutant germ cells is independent of Notch signaling and dependent on cyclin E activity [4], [7]. Intriguingly, germ cells lacking GLD-2 and NOS-3 are able to start meiotic prophase [6], [7]. This suggests that the current pathway assignments are too simplistic and emphasizes that more meiosis-promoting regulators must exist [4], [6], [7]. Especially, commitment to female meiotic progression provides precedence for redundant translational activation activities in C. elegans. Here, in addition to GLD-2 cytoPAP-mediated GLD-1 expression [29], the GLD-4/GLS-1 cytoPAP complex has been identified to translationally activate gld-1 mRNA [30]. As a non-canonical poly(A) polymerase, GLD-4 is most similar to the conserved group of Trf4/5-type RNA modifiers that regulate RNA stability in the nucleus [30], [31], [32]. However, GLD-4 poly(A) polymerase is cytoplasmic and requires for its functions the nematode-specific protein, GLS-1 [30], [33]. In the absence of GLD-2, the GLD-4/GLS-1 cytoPAP complex is essential for female meiotic progression into pachytene [30].
In this study, we report that the GLD-4/GLS-1 cytoPAP complex has a dual role in regulating the balance between proliferation and differentiation. We find that the GLD-4/GLS-1 cytoPAP complex is crucial to maintain germ cell proliferation in the adult, in part by promoting robust translation of glp-1 mRNA. Moreover, to ensure that meiosis-promoting factors are inefficiently translated, GLD-2 cytoPAP levels are kept low in the GSC pool by FBF-mediated translational repression. Lastly, we also find that GLD-4/GLS-1 cytoPAP promotes meiotic prophase entry, in parallel to GLD-2 cytoPAP and independently of Notch. Our data suggest that two translational feedback loops limit the size of the proliferative germ cell pool and maintain a healthy balance of germ cell proliferation and differentiation in the adult germ line.
On average, the PZ of adult germ lines extends from the first germ cell row at the distal end further proximally to row 20, where germ cells start differentiation by entering meiotic prophase (Figure 1A,B). In wild type, the PZ is populated by about 225–250 germ cells (Figure 1C). Since there are no molecular markers for subpopulations of cells in the PZ, like stem cells, transit amplifying cells and cells in pre-meiotic S-phase, the start of meiotic prophase is commonly defined as the onset of differentiation [1], [34]. Differentiation is revealed by the germ cells' specific nuclear architecture and chromatin morphology, the combinatorial expression and localization of the meiotic cohesin REC-8, the synaptonemal protein HIM-3, and phosphorylated nuclear envelope protein pSUN-1 [34] (Figure 1A,B).
Germ cells in single mutants of meiosis-promoting genes (i.e. gld-1, nos-3, gld-2, gld-3) initiate meiotic prophase [7]. However, shifts in the position of the mitosis-to-meiosis boundary suggest a role in proliferation or differentiation. For example, in gld-2 single mutants, the PZ is extended and contains more germ cells than wild type [7] (Figure 1C,D), consistent with gld-2's function in promoting meiotic entry [35]. We found that gld-4 and gls-1 single mutants have smaller PZs with fewer germ cells (Figure 1C,D). The strength of the reduction appears to correlate with the reported allelic strengths of the individual mutations [30], [33] (Figure 1C). As the PZ of gld-4 gls-1 double mutants is similarly reduced (Figure 1C), these results argue for a common role of gld-4 and gls-1 in promoting mitosis. The PZ of the gld-2 gld-4 double mutant is similar to wild type in size and germ cell number (Figure 1C,D). Together, these results suggest that gld-2 and gld-4 have independent and opposing roles to set the mitosis-to-meiosis boundary in adults.
The PZ expands during larval development and is maintained during adulthood [8]. We measured the size of the PZ at the last larval stage (L4), and 24 hours (h), and 48 h later in young adults (Figure 1E–G). The difference between wild type and gld-4 is the smallest in L4 and greatest during adulthood, due to a large relative shrinkage of the PZ in gld-4 young adults (Figure 1E–G). Therefore, gld-4 activity is primarily important for the maintenance but not establishment of the PZ during early adulthood.
The documented presence of GLD-4 [30] and GLS-1 [33] in the distal end of the germ line and the single mutant phenotypes argue for a role of gld-4 and gls-1 in promoting germ cell proliferation. CytoPAPs are envisioned to regulate poly(A) tail metabolism of target mRNAs in a positive manner [31]. Biochemically, cytoPAPs elongate poly(A) tails, which in turn stabilize mRNAs and enhance their translation. We hypothesized that GLD-4 targets mRNAs encoding proteins important for proliferation in the PZ. An obvious, but not exclusive, candidate for this regulation is the Notch receptor-encoding glp-1 mRNA.
Notch expression is regulated at multiple levels in C. elegans [13], [23], [36]. To uncouple mRNA regulation from protein regulation, we used a translational reporter of GFP::H2B under the control of the glp-1 3′UTR [25] (Figure 2). The glp-1 3′UTR reporter is driven by a ubiquitous germ cell-specific promoter and encodes a translational fusion product of GFP and histone 2B (Figure 2A). This nuclear GFP signal reflects GLD-1-mediated regulation of the glp-1 mRNA [25].
In a wild-type background, reporter GFP expression is present in all animals analyzed and its pattern is similar to endogenous GLP-1 protein expression [13], [25] (Figure 2B). To assess whether reporter GFP expression is under the influence of GLD-4 cytoPAP activity, we crossed the glp-1 3′UTR reporter locus into the strong loss-of-function gld-4(ef15) mutant background (Figure 2C,D). To control for unexpected genetic background influences, we compared heterozygous and homozygous gld-4 siblings from the progeny of a heterozygous mother (see Materials and Methods). In the gld-4 heterozygous mutant, reporter GFP expression is similar to a wild-type background (compare Figure 2B with C). Strikingly, upon gld-4 removal, reporter expression was undetectable in almost all germ lines (Figure 2D). Consistent with a reduction in the GFP signal, we also observed lower GFP protein amounts by immunoblotting. When comparing gld-4 animals to wild-type background, we observed a reduction of >80% in protein abundance (Figure 2E). These results imply that the expression of the glp-1 3′UTR reporter depends on gld-4 cytoPAP activity.
GLS-1 and GLD-4 function together in meiotic progression [30], and in promoting differentiation onset (Figure 1C). Similar to the gld-4 mutant, reporter GFP expression is undetectable in most gls-1(ef8) mutant germ lines (∼87%, n = 220), suggesting that gls-1 promotes glp-1 3′UTR reporter expression similar to gld-4 activity.
To investigate whether gld-4 is the only known cytoPAP regulating reporter expression, we assessed GFP expression in gld-2 mutants and detected it in almost all germ lines (Figure 2F). Moreover, the amounts of GFP reach wild-type protein levels and are similar between gld-2 homozygous and heterozygous mutants (Figure 2E). Importantly, reporter GFP expression is still dependent on gld-4 activity in the gld-2 mutant background, as its expression is undetectable in all gld-2 gld-4 homozygous double mutants (Figure 2G). These results suggest that glp-1 3′UTR reporter expression is largely independent of gld-2 cytoPAP activity, and specifically dependent on gld-4 cytoPAP activity.
To further investigate at which level GLD-4 cytoPAP may regulate glp-1 3′UTR reporter expression, we made use of GLD-1, a known translational repressor of glp-1 mRNA [23]. In gld-1 single mutants, reporter GFP is expressed in the PZ and in differentiating germ cells (100%, n = 140). To test, whether loss of GLD-1 would de-repress reporter GFP expression in the gld-4 mutant, we analyzed GFP::H2B expression in the gld-1 gld-4 double mutant background. Most germ lines weakly express GFP when compared to gld-4 mutants (compare Figure 2H with 2D). A similar weak de-repression is observed in gld-4 mutant germ lines that contain mutated GLD-1-binding site reporter mRNAs (glp-1 3′UTR mut) (Figure 2I). Taken together, these results confirm that the glp-1 3′UTR reporter can be translated in a gld-4 mutant background and that expression of the glp-1 3′UTR reporter is partly dependent on the GLD-4 cytoPAP even when GLD-1-mediated repression is removed.
Several mechanisms may account for reduced glp-1 3′UTR reporter expression in the absence of gld-4. To confirm that the effects on GFP::H2B expression are due to translational and not transcriptional regulation of the glp-1 3′UTR reporter, we examined the mRNA levels of the wild-type glp-1 3′UTR reporter by RT-qPCR (Figure 2J). Compared to wild type, we noticed a reduction of ∼4-fold in both gld-4 and gld-2 mutant backgrounds (Figure 2J), suggesting that glp-1 3′UTR reporter mRNA is less abundant in either cytoPAP mutant. Importantly, the glp-1 3′UTR reporter mRNA levels are similar to each other in both cytoPAP homozygous mutants, yet they give rise to different amounts of reporter protein (compare Figure 2D with 2F, and Figure 2E). Hence, we conclude that the major reduction in reporter GFP expression in gld-4 mutants is primarily at the translational and not at the transcriptional level.
To further investigate whether endogenous GLP-1 protein expression is one likely candidate of gld-4-mediated regulation, we measured GLP-1 protein expression in gld-4 mutants and compared it to wild type (Figure 3). By quantifying GLP-1 intensities in distal germ lines of L4+24 h and L4+48 h animals, we observed a significant decrease in the gld-4 mutant background in the PZ over time (Figure 3A,B). When we measured endogenous glp-1 mRNA levels in L4+24 h animals we observed a mild increase in gld-4(ef15) mutants compared to wild type (Figure 3C). Together these observations suggest that gld-4 promotes GLP-1 expression post-transcriptionally.
A prerequisite for GLD-4/GLS-1-mediated glp-1 mRNA regulation is that they form an mRNP complex. To test for a possible association of GLD-4 and GLS-1 with glp-1 mRNA, we performed several RNA co-immunoprecipitation (RIP) experiments, using GLD-4-specific, GLS-1-specific, and non-specific antibodies. Subsequent RT-PCR (Figure 4A) and RT-qPCR (Figure 4B) analysis of different RIP experiments revealed a specific enrichment of endogenous glp-1 mRNA, which is similar to the positive control, gld-1 mRNA (Figure 4B). These results demonstrate an association of GLD-4/GLS-1 cytoPAP complex with endogenous glp-1 mRNA and establish a potential physical link for glp-1 mRNA translational regulation.
Cytoplasmic polyadenylation affects RNA stability and translational efficiency [37]. To test whether ribosomal engagement of the endogenous glp-1 mRNA requires GLD-4 cytoPAP, we performed sucrose gradient sedimentation experiments. In theory, the more ribosomes are attached to an mRNA, the further the mRNA migrates into the gradient during ultra centrifugation. Therefore, efficiently translated mRNAs will be in the heavier, polysome fractions of the gradient, while poorly or non-translated mRNAs tend to sediment to lighter, non-polysomal fractions. Due to the large amounts of material needed, we compared control RNAi and gld-4(RNAi) knockdown worms (Figure 4C), knowing that gld-4(RNAi) efficacy is less robust than using mutants. In extracts of wild type and control RNAi (Figure 4C), about 50% of the endogenous glp-1 mRNA resides in the polysome fraction, suggesting that half of the glp-1 mRNA population is actively translated, consistent with the known germline and embryonic translational repression of glp-1 mRNA [23], [38]. Upon knockdown of gld-4, but not in control RNAi, we observed a shift of glp-1 mRNA into lighter fractions of the gradient (Figure 4D). This reflects a specific decrease in translational competence of endogenous glp-1 mRNA as rpl-11.1, a germ line-enriched mRNA that encodes a protein of the large ribosomal subunit [39], is unaffected (Figure 4D).
CytoPAPs modify the 3′ends of RNAs [31]. To investigate whether GLD-4 affects the length of the glp-1 mRNA poly(A) tail, we performed a poly(A) test (PAT) assay [40], and compared endogenous glp-1 mRNA poly(A) tails, using sucrose gradient fractioned mRNA and non-fractionated input as our starting material. To obtain enough RNA material for the PAT assay and to discriminate translationally active from inactive mRNA pools, we combined several samples of the non-polysomal and polysomal fractions. While all three samples show reduced glp-1 poly(A) tail lengths in gld-4(RNAi) compared to control RNAi knockdowns, we observe no clear difference between the respective non-polysomal and polysomal fractions (Figure 4E,F). The observed poly(A) tail differences are consistent with the contribution of gld-4 activity to gld-1 mRNA [40]. This data suggests that GLD-4 cytoPAP activity has an overall impact on glp-1 poly(A) tail status. Together, our combined results suggest that GLD-4 association with endogenous glp-1 mRNA may stimulate its efficient translation.
GLD-4 and GLD-2 cytoPAP expression patterns are distinct in “female” germ lines. GLD-4 is expressed equally strong within the entire PZ and in meiosis [30] (Figure 5A). By contrast, GLD-2 is poorly expressed in the distal half of the PZ, becomes more abundant further proximal, and is most abundant in cells that have entered meiosis [28] (Figure 5A). Hence, the differential expression of the two proteins in the PZ may form the basis of GLD-4's unique role in mitosis and GLD-2's role in meiotic entry.
Intriguingly, the protein expression pattern of GLD-2 does not match its ubiquitous mRNA expression pattern in the distal PZ [28]. This suggests translational regulation of GLD-2 expression. An obvious translational repressor in this region is FBF, which represses two mRNAs encoding meiosis-promoting regulators, GLD-1 [18] and GLD-3 [7]. To test for FBF-mediated gld-2 mRNA repression, we knocked down fbf by RNAi and assessed GLD-2 protein abundance in the distal-most germ line by indirect immunofluorescence, using GLD-4 as a reference, and quantified the amounts (Figure 5B,C). While GLD-4 levels are not significantly different between fbf(RNAi) and control RNAi germ lines, GLD-2 expression levels in the PZ are higher in fbf(RNAi) than in wild type (compare Figure 5A and 5B) or control RNAi experiments (Figure S1A–C). The GLD-2 protein increase is largely limited to the distal half of the PZ: ∼2.2-fold more in cells most distal (Figure 5C, area 1), compared to ∼1.5-fold more in cells most proximal to the PZ (Figure 5C, area 2). Such a restriction to the proliferative zone is consistent with previous reports on FBF activity [18], [20], [41] and suggests that GLD-2 but not GLD-4 is a specific target of FBF regulation.
FBF interacts with mRNAs through the conserved FBF-binding element (FBE) [18]. We identified five putative FBEs in the 3′UTR of gld-2 mRNA (Figure 5D) and tested each element for binding to FBF protein in a yeast 3-hybrid assay. Only FBE4 in its wild-type sequence was consistently and specifically bound by FBF (Figure 5E). Neither element was bound by PUF-5 (Figure 5E), a different C. elegans PUF protein that is abundantly expressed in differentiating female gametes [42]. Intriguingly, the bound FBE sequence is also present in two closely related Caenorhabditis species, suggesting that gld-2 mRNA translational repression may be conserved (Figure 5D). Moreover, RIP experiments of GFP-tagged FBF-2 confirmed a physical association of gld-2 mRNA with FBF in worm lysates, which appears to correlate with the number of active FBEs in the tested mRNAs (Figure 5F); the positive control, gld-1 mRNA, possesses two functional FBEs and was enriched strongest [18]. Taken together, we conclude that consistent with published FBF-1 RIP-Chip experiments [19], GLD-2 but not GLD-4 is most likely a direct target of the central mitosis-promoting translational repressor, FBF. Consistent with previous genetic findings [7], an evolutionary conserved translational repression of GLD-2 cytoPAP in undifferentiated cells might be pivotal for the robustness of the balance between proliferation and differentiation.
The current framework of the core regulatory network underlying meiotic entry appears incomplete and a third meiosis-promoting activity is likely to exist (Figure 6A) [4], [6], [7], [34]. Even though both meiosis-promoting pathways are inactive in the gld-2; nos-3 double mutant, germ cells enter meiosis [7], [26] (Figure 6B,D; Table 1). Intriguingly, GLD-2 and GLD-4 have a combined function during late meiosis when germ cells are past the onset of differentiation [30]. Hence, it seemed plausible that a further biological overlap of those two enzymes may exist at differentiation onset. Indeed, we find that the triple mutant gld-2 gld-4; nos-3 lacks any signs of differentiation and it is tumorous (Figure 6C,E; Table 1). This demonstrates that gld-4 activity promotes meiotic entry in the absence of gld-2 and nos-3.
GLS-1 stimulates GLD-4 enzymatic activity and the GLD-4/GLS-1 cytoPAP complex promotes late meiosis [30]. To test if gld-4 activity requires gls-1 function for promoting meiotic entry, we generated the gld-2 gls-1; nos-3 triple mutant. Similar to the gld-2 gld-4; nos-3 triple mutant, no meiotic entry was observed (Figure 6F; Table 1), indicating a shared function of gld-4 and gls-1. Together this suggests that in addition to a requirement for proliferation, the GLD-4/GLS-1 cytoPAP complex promotes the onset of differentiation in combination with GLD-2 cytoPAP.
A prediction of this model is that the function of a single cytoPAP is enough to promote entry into meiosis in the absence of nos-3. Hence, we generated the gld-4; nos-3 and the gls-1; nos-3 double mutants. In either double mutant, in comparison to the triple mutant with gld-2, we found robust entry into meiosis (Figure 6G,H; Table 1). In conclusion, gld-4 and gls-1 promote meiotic entry in parallel to gld-2 and nos-3, suggesting that gld-4 and gls-1 might be additional pathway components that promote differentiation onset. Moreover, the striking similarity between the gld-3 nos-3 double and gld-2 gld-4; nos-3 triple tumorous germ lines suggest that gld-2 and gld-4 or gls-1 activities are largely equivalent to gld-3 activity with regard to the meiotic entry process.
NOS-3 and GLD-1 are assumed to act in a pathway parallel to the GLD-2/GLD-3 cytoPAP pathway (Figure 6A). To complete our analysis of the genetic interactions between the NOS-3/GLD-1 and the GLD-4/GLS-1 cytoPAP pathways, we generated triple mutant strains that had either one of the GLD-4/GLS-1 cytoPAP complex components removed in a gld-2 gld-1 double mutant background (Figure 7; Figure S2; Table 1).
Germ cells, double mutant for gld-2 gld-1, enter meiosis in the majority of germ lines (Figure 7A; Figure S2; Table 1). Germ cells, triple mutant for gld-2 gld-1 gld-4 (Figure 7B) or gld-2 gld-1 gls-1 (Figure 7C), failed to enter meiosis and all germ lines are tumorous (Table 1). Importantly, germ cells in the gld-1 gld-4 (Figure 7D) and the gld-1 gls-1 (Figure 7E) double mutants enter meiosis (Table 1). Surprisingly, the gld-1 gls-1 double mutant did not stain for HIM-3 (Figure 7E). However, gld-1 gls-1 germ cells entered meiosis, as judged by their nuclear architecture, chromosome morphology, and the expression of pSUN-1 (Figure 7F). Our combined results are consistent with the previous triple mutant results, in which a nos-3 mutant gene replaced gld-1 (Figure 6), and establish a role of gld-4 and gls-1 in the onset of differentiation, suggesting that both genes operate in parallel to gld-2, gld-1 and nos-3.
Notch signaling promotes proliferation, upstream of the meiosis-promoting network [35]. To investigate whether proliferation of tumorous triple mutant gld-2 gld-1 gld-4 and gld-2 gld-1 gls-1 germ lines depends on Notch activity, we investigated GLP-1 protein expression and genetically ablated glp-1 function (Figure S3). In either triple mutant, GLP-1 remains expressed throughout the tumorous germ lines (Figure S3A,C). Consistent with their proliferative activity, dividing cells are scattered throughout the germ line and stain positively for phospho-histone-3 (PH-3) (Figure S3A,C), a marker for cells in prometaphase [43]. Loss of glp-1 in either triple mutant neither abolishes proliferation nor leads to meiotic entry and cells remain undifferentiated (Figure S3B,D). These results suggest that Notch is not required for proliferation in germ cells that are fully compromised in all meiosis-promoting pathways.
Germ cell proliferation in gld-3 nos-3 tumorous germ lines is independent of Notch signaling but depends on cyclin E [4] (Figure 8A; Table 1). In cye-1 RNAi knockdown experiments, we found that also gld-2 gld-1 gld-4 tumorous proliferation requires cyclin E activity (Figure 8B; Table 1). Moreover, consistent with gld-3 nos-3; glp-1 cye-1(RNAi) germ lines [4], an additional removal of Notch activity in gld-2 gld-1 gld-4; cye-1(RNAi) animals increases the ability of germ cells to start meiotic prophase more distally (Figure 8C). In either case, however, differentiation onset is aborted immediately after zygotene/very early pachytene and germ cells do not commit to meiosis. Together, these similarities among the gld-3 nos-3 and gld-2 gld-1 gld-4 tumorous germ lines suggest that gld-4 and, most likely gls-1, are components of a meiosis-promoting pathway that acts on the gld-2 side of both known meiosis-promoting pathways, rather than in a separate, third meiotic entry pathway (summarized in Figure 9) [4].
Our summed findings highlight that translational control, in the combined form of translational activation and repression, serves as a key regulatory mechanism to maintain adult tissue homeostasis in the C. elegans germ line (Figure 9). Central to our findings is the dual activity of the GLD-4/GLS-1 cytoPAP complex, which has a major role in promoting germ cell proliferation and a minor role in differentiation onset. By focusing on the activity of numerous key RNA regulators, this work expands the known core genetic circuitry downstream of niche-mediated Notch signaling that governs the balance between proliferation and differentiation (summarized in Figure 9A). At the molecular level, we propose a rheostat that consists of two translational control modules, one specific for proliferation (Figure 9B) and one specific for differentiation onset (Figure 9C). Both modules are interconnected via their mRNA targets, and this reciprocal translational activation and repression of either proliferation or differentiation factors may fine-tune the size of the proliferative zone.
The GLD-4/GLS-1 cytoPAP complex has multiple roles in germ cell development [30], [33]. In this work, we demonstrate that both complex members contribute to the maintenance of the size of the proliferative zone by primarily influencing adult germline proliferation and secondarily differentiation onset. This dual role is consistent with the ubiquitous expression of both proteins in the respective regions of the adult germline tissue [30], [33]. GLD-4 is the enzymatic component of the GLD-4/GLS-1 cytoPAP complex and is evolutionarily most similar to nuclear Trf4/5-type polymerases, which add short poly(A) tails to nonproductive RNA molecules to initiate exosome-mediated RNA degradation [32]. By contrast, GLD-4 and its enzymatic activator GLS-1 are cytoplasmic proteins implicated in translational control [30], [33]. The notion that translational activation of mRNAs is coupled to cytoplasmic poly(A) tail extension or maintenance is primarily shaped by the work on poly(A) polymerases, such as members of the conserved GLD-2 family [31]. By analogy, GLD-4 cytoPAP's role in proliferative germ cells may therefore translationally activate mitotic-fate promoting mRNAs. We provided four pieces of evidence that the Notch receptor-encoding glp-1 mRNA is a likely mRNA target of GLD-4/GLS-1 cytoPAP activity: (1) GLD-4 associates with glp-1 mRNA, and (2) positively influences its poly(A) tail length. (3) Furthermore, we found that expression of a glp-1 3′UTR translational reporter and that of endogenous GLP-1 protein depends on GLD-4 presence. (4) Lastly, translational efficiency of endogenous glp-1 mRNA requires gld-4 activity. Therefore, these results combined support the idea that abundant GLP-1 expression is maintained by GLD-4-mediated translational activation. However, the partial reduction in glp-1 poly(A) tail length might either reflect an intrinsic enzymatic difference between Trf4-type PAPs and GLD-2, or suggests that other, yet undiscovered cytoPAPs may work redundantly to GLD-4. Alternatively, additional poly(A) tail-independent mechanisms for GLD-4-mediated translational activation may exist. Regardless of the precise molecular function of GLD-4, the translational repressor of glp-1 mRNA, GLD-1 protein, starts to accumulate in the proximal part of the PZ, prior to the mitosis-to-meiosis boundary [44], suggesting that glp-1 mRNA may already be subject to translational repression in the proximal PZ. Therefore, to ensure robust GLP-1 protein expression, GLD-4-mediated translational activation of glp-1 mRNA may help to counteract GLD-1-mediated translational repression to maintain the size of the PZ in the adult (Figure 9A,B). However, glp-1 mRNA is presumably not the only target of the GLD-4/GLS-1 cytoPAP complex, and others are likely to exist.
In the balance between proliferation and differentiation, the two translational activators, GLD-4 and GLD-2, seem to have antagonistic roles that may also constrain their regulation and function in the PZ. A loss of GLD-4 shrinks the PZ and a loss of GLD-2 expands the PZ. Therefore, GLD-2 may promote meiosis at the expense of mitosis in the gld-4 single mutant. Conversely, GLD-4 may be responsible for the expansion of the PZ in the gld-2 single mutant. Importantly, upon loss of both cytoPAP activities, the PZ re-adjusts to an intermediate size, arguing that they form an antagonistic pair. In particular, the distinct expression profile of either cytoPAP presumably reflects and affects their divergent roles in regulating mRNA-specific gene expression. The delay of GLD-2 protein expression in the PZ correlates with its genetic requirement for the onset of differentiation and a putatively required absence in undifferentiated cells. Moreover, its 2–3 fold lower abundance in the distal half of the PZ may selectively favor and functionally constrain GLD-4-mediated germ cell proliferation. Hence, a healthy balance between GLD-2 and GLD-4 functions appears to be perpetuated to maintain the size of the adult PZ.
To maintain adult germ cell proliferation and prevent progressive shrinkage of the PZ, gld-2 mRNA translation is delayed by FBF, a dominant translational repressor of several meiosis-promoting genes [7], [18], [19], [45]. We found that GLD-2 but not GLD-4 cytoPAP accumulation in the PZ appears to be inhibited by FBF, and that gld-2 mRNA associates with FBF most likely at least through one FBF-binding site in its 3′UTR. Therefore, a translational repressor (FBF) that turns off the activities of mRNAs encoding meiosis-promoting proteins (e.g. GLD-2) is combined with a translational activator (GLD-4) that turns on mRNA activities that encode mitosis-promoting proteins (e.g. GLP-1) to maintain germ cell proliferation (Figure 9A,B).
Conversely to germ cell proliferation, the onset of differentiation requires translational repressors (GLD-1 and NOS-3) that presumably turn off mRNA activities encoding mitosis-promoting proteins and translational activators (GLD-2, GLD-3, GLD-4, and GLS-1) that presumably turn on mRNA activities encoding meiosis-promoting proteins [6], [7], [26], [35] (Figure 9A,C). Previous genetic work established two parallel pathways, which either indirectly or directly promote differentiation onset (Figure 6A). However, not all components are equal in their potential to contribute to meiotic prophase entry. In this regard, the synergism of NOS-3 and GLD-3 is of equal strength, as is NOS-3 with both GLD-2 and GLD-4/GLS-1, or, GLD-1 with both GLD-2 and GLD-4/GLS-1. Hence, our findings of a dual role for GLD-4 cytoPAP strengthens the role of translational control even further, highlights the importance of translational activation for the balance of proliferation and differentiation, and clarifies the many levels of redundancy within the two, major, parallel pathways of the current genetic circuitry (Figure 9A).
Differentiation onset deploys two translational activators of presumed meiosis-promoting mRNAs (Figure 9C). In this regard, GLD-2 cytoPAP performs a more prevalent role in activating meiosis-promoting mRNAs as its combined loss with genes of the first, translational repressor pathway (i.e. gld-1 gld-2 or gld-2; nos-3 doubles) causes more germ cell overproliferation than is observed in the respective gld-4 double mutant germ lines. Importantly, germ cells of gld-2; nos-3 or gld-2 gld-1 double mutants enter meiosis in a gld-4- and gls-1-dependent manner, as triple mutant germ cells (e.g. gld-1 gld-2 gld-4 or gld-2 gld-4; nos-3) do not enter meiosis. Consistent with previous findings that germline proliferation in tumorous gld-2 gld-1 or gld-3 nos-3 double mutants is glp-1-independent [4], [35], tumorous triple mutant gld-2 nos-3 germ cells that lack in addition either gld-4 or gls-1 do not require GLP-1 activity to remain in mitosis either, arguing for their genetic position downstream of Notch and in parallel to each other for meiotic entry (Figure 9A). Intriguingly, the similarities between the gld-3 nos-3 double and gld-2 gld-4; nos-3 or gld-2 gls-1; nos-3 triple mutants suggest further that gld-3 activity equals the combined activities of gld-2 and gld-4/gls-1 with respect to the loss of nos-3, which positions gld-4/gls-1 within the second, translational activator pathway at the level of gld-2 (Figure 9A). These genetic behaviors appear to parallel the known molecular protein interactions. The multi-KH domain protein GLD-3 binds directly to GLD-2 cytoPAP and GLS-1 [30], [33], illustrating that GLD-3 may serve as an integral regulatory factor for both GLD-2 and GLD-4/GLS-1 cytoPAPs to promote differentiation onset.
Redundancy of cytoPAP-mediated translational activation has been previously reported in a later step of meiotic prophase of female germ cells that require abundant GLD-1 expression for meiotic commitment [30]. Intriguingly, gld-2 gld-4 double mutant germ cells enter meiosis [30], suggesting that the remaining low GLD-1 amounts might be sufficient to promote meiotic entry. Consistent with this idea, gld-2 gld-4 gld-1 triple mutant germ cells never enter meiosis, arguing that in the absence of cytoPAP activity, the remaining gld-1 activity/GLD-1 amount is indeed crucial for meiotic entry. In agreement with previous findings [4], [46], [47], our work suggests that for differentiation onset in gld-2 gld-4 double mutants, cyclin E represents an important target of GLD-1-mediated translational repression. However, we expect additional differentiation onset-promoting mRNA targets to be positively regulated by GLD-2 and GLD-4, either in a combinatorial manner or separately. Alternatively, other RNA-directed molecular functions, such as miRNA stability described for GLD-2 orthologs in mammals [48], might be relevant. Future research on the RNA-regulatory repertoire of GLD-2 and GLD-4 will be required to better resolve these issues.
We propose that two modules of translational activation and repression, interconnected via their mRNA targets, establish a molecular rheostat that leads to a reciprocal expression of either proliferation or differentiation factors. Together they maintain adult germline proliferation in adult C. elegans animals. Translational repression, in particular, is an established mechanism in Drosophila and C. elegans development. Translational control is also an essential mechanism of the transition from self-renewal/proliferation to differentiation in Drosophila germ cells [49], [50]. Our work suggests that the regulation of turning translation on is equally important for maintaining a healthy balance between proliferation and differentiation as turning translation off. With this work, we begin to fill this obvious gap in our understanding of adult tissue maintenance.
C. elegans strains were handled according to standard procedures [51]. Worms were grown at 20°C and used for most experiments at an age of 24 hours (h) past mid-L4. Bristol N2 served as the wild-type strain.
Mutations used: LGI: gld-2(q497), gld-1(q485), fer-1(b232), gls-1(ef4), gls-1(ef8), gld-4(ef9), gld-4(ef15). LGII: gld-3(q730), nos-3(q650). LGIII: glp-1(q175). Transgenes used: rrrSi117[Pmex-5::GFP::H2B::glp-1(wt 3′UTR) unc-119(+)] II, rrrSi118[Pmex-5::GFP::H2B::glp-1(GBM1,2,3 mut 3′UTR) unc-119(+)] II; both are Mos1-mediated single copy gene insertions and their sequences are described in the supplemental text of [25]. JH2929 expresses the LAP-tagged FBF-2 [52].To generate the new gld-2(q497) gld-1(q485) double mutant, we crossed heterozygous gld-2 males with heterozygous gld-1 hermaphrodites. Next, we crossed F1 non-green worms, containing gld-1 and gld-2, and green siblings, containing the hT2[qIs48] I;III balancer. In the F2 progeny heterozygous balancer animals were screened for a recombination event between the gld-2 and the gld-1 locus by genomic PCR for the q485 deletion and sequencing for the q497 point mutation. Homozygote hT2[qIs48] I;III animals are embryonic lethal and cannot be analyzed as a wild-type sibling control.
All other double and triple mutants on LGI were generated in a similar manner and balanced by hT2[qIs48] I;III and are listed in Table 1. Quadruple mutants containing genes on LGI and LGIII were balanced by hT2[qIs48] I;III and validated by PCR for deletions and by sequencing for gld-2(q497) and glp1(q175). Double and triple mutant combinations on LGI and LGII were maintained by a closely linked GFP transgene (ccIs4251) to unc-15(e73) on LGI and mIn1[mIs14 dpy-10(e128)] on LGII. The presence of all mutations was validated by PCR for deletions or by sequencing. Primer sequences are available on request.
RNAi experiments were performed according to published feeding RNAi procedures [53]. The fbf RNAi construct corresponds to fbf-1 (nts 1040–1845), cye-1 is described elsewhere [54], and the empty pL4440 vector served as a control. L4-staged N2 animals were placed on RNAi plates and analyzed 24 h later. The efficiency of fbf knockdown was confirmed by a loss of anti-FBF immunoreactivity 24 h past L4, and after continued feeding at 48 h past the L4 stage by phenotypic changes of the germ lines, i.e. the shrinkage of the PZ and even later the appearance of male-fated germ cells [55].
Antibodies against the following proteins were used as described: anti-HIM-3 1∶200 [56], anti-pSUN-1 1∶1000 [57], anti-GLD-4 1∶20 [30], anti-GLP-1 1∶10 [13], anti-PH-3 1∶500 (9706, Ser-10, 6G3, Cell Signaling), anti-FBF-1 1∶100 [55], anti-GLH-2 1∶200 [58], anti-GLS-1 1∶20 [33]. Monoclonal anti-REC-8 1∶20 (mo560-G25-1, at 10 ng/ul) and anti-GLD-2 1∶20 (A4-4, at 10 ng/ul) antibodies were generated against recombinant HIS-REC-8(aa330–525) and GST-GLD-2(aa959–1113) fusion peptides. The antibodies are specific to the respective proteins; no immunocytochemistry signal was observed in corresponding null mutants and the protein expression patterns in wild type match published ones [28], [59]. Secondary antibodies (1∶500) were coupled to FITC, CY3 and CY5 (Jackson Labs).
Extruded germ lines were prepared in solution as described [33]. The correct localization and comparable intensities of GLH-2 served as a tissue penetration control for all immunofluorescence experiments. Images were acquired with Axiovision Software (Zeiss) on a wide-filed Imager Z1 (Zeiss) microscope, equipped with an AxioCam MRm (Zeiss) camera. Raw images were processed in Photoshop CS5 (Adobe) and assembled in Illustrator CS5 (Adobe). For quantification of immunofluorescent intensities, all images for comparison were taken with identical settings. A median focal plane was chosen where the syncytium was at its maximum width. The pixel intensities were measured in Fiji (ImageJ). To compare GLP-1 intensities, a line scan was performed as is indicated in Figure 3B, ranging from the distal germline tip to the beginning of pachytene. Then all values were binned into the 10 fractions whose positions are displayed in Figure 3B. Averages of those fractions between all analyzed germ lines were calculated and normalized to GLH-2 intensities (measured in the same way). To compare cytoplasmic GLD-2 and GLD-4 intensities, four identical circles were placed over the rachis of the distal arm along the distal-to-proximal axis as indicated in Figure 5C (five germ cell diameters (GCD) proximal of the distal tip, at the end of the PZ, at the beginning of pachytene, and ten GCD into pachytene) and averaged for all germ lines per genotype. The GLD-2/GLD-4 intensities given are not normalized to the GLH-2 signals, which were in these cytoplasmic regions very low. To ensure equal penetration, we independently measured the peri-nuclear GLH-2 signal in neighboring nuclei and found it very similar among all analyzed germ lines.
Immunoblots were performed according to standard procedures with a mixture of two anti-GFP antibodies, at a final dilution of 1∶1000 (11814460001, clones 7.1 and 13.1, Roche) and 1∶200 (sc-9996, B-2, Santa Cruz), anti-tubulin 1∶100000 (T 5168, clone B-5-1-2, Sigma), and HRP-conjugated anti-mouse secondary antibodies (Jackson Labs).
Three-hybrid experiments were performed as described [60]. gld-2 RNA sequences were cloned into the XmaI and SphI sites of the vector pIIIA/MS2-2, using either PCR-amplified fragment (FBE4) or annealed synthetic oligonucleotides (remaining FBE sites). Their nucleotide positions in relation to the first nucleotide of the gld-2 3′UTR are as follows: FBE1 (nts 298–335); FBE2 (nts 354–394); FBE3 (nts 460–493); FBF4 (nts 683–763); FBE5 (nts 952–986). For binding specificity, a mutation (UG to AC, see Figure 4D) was engineered by site-directed mutagenesis using Quikchange (Stratagene).
For the sucrose gradient experiments, whole-worm extracts of L1 synchronized adult animals (L4+24 h) grown in comparable feeding-RNAi conditions were prepared by pulverizing frozen worms and adding lysis buffer [50 mM HEPES pH 7.5, 125 mM KCl, 5 mM MgCl2, 1 mM DTT, 0.005% NP-40, 2× Protease Inhibitor Cocktail without EDTA (Roche), 100 U/ml Ribolock (Fermentas), 2 mM PMSF, 4 mM Benzamidine, 2 µg/ml Leupeptin, 2 µg/ml Pepstatin, 0.1 µg/ml Pefabloc, 2 mM NaF, 2 mM Na3VO3 and 200 µg/ml Cycloheximide], followed by a low speed centrifugation to removed insoluble components. The clear supernatant of three biological replicates was layered onto a 17–50% w/v sucrose gradient and processed as previously described [61] with the only exception that the gradients were spun for 210 min. For the mRNA distributions analysis, 10 fmole of a polyadenylated in vitro transcribed luciferase mRNA was added to each fraction prior to RNA isolation as an internal RNA standard for extraction efficiency. The Trizol (Invitrogen) isolated RNA from individual fractions was resolved in equal volumes of water and further analyzed by qRT-PCR or pooled for splint-mediated poly(A) tests [40].
RIP experiments were performed from mixed-stage animals as described [62] using anti-GFP (MPI-CBG), anti-GLD-4 [30], or anti-GLS-1 [33] antibodies. Semiquantitative RT-PCR samples of three independent RIP experiments were resolved on ethidium bromide-stained agarose gels. The control samples without Reverse transcriptase were negative and are not shown. For the qRT-PCR experiments, equal volumes of gradient fractionated RNA, total RNA, or RIP material of three biological replicates was used as input material. cDNA synthesis was performed using Revert Aid Premium Transcriptase (Fermentas) in combination with oligo dT primers, following the manufactures protocol. qPCRs were performed on a Mx3000P qPCR system (Stratagene) using ABsolute qPCR SYBR Green mix (Thermo) under standard conditions.
For measuring poly(A) tail length, we pooled five non-polysome and polysome fractions as indicated in Figure 4. Together with 4 µg total RNA of the input material, we processed the pooled sucrose gradient experiments by ligating an RNA anchor to the 3′ends, preformed a gene-specific RT-PCR, and resolved the DNA samples on a high-resolution agarose gel, according to [40]. Lane quantifications were performed using Fiji (ImageJ), as described in [40].
|
10.1371/journal.pgen.1004681 | Selection on a Variant Associated with Improved Viral Clearance Drives Local, Adaptive Pseudogenization of Interferon Lambda 4 (IFNL4) | Interferon lambda 4 gene (IFNL4) encodes IFN-λ4, a new member of the IFN-λ family with antiviral activity. In humans IFNL4 open reading frame is truncated by a polymorphic frame-shift insertion that eliminates IFN-λ4 and turns IFNL4 into a polymorphic pseudogene. Functional IFN-λ4 has antiviral activity but the elimination of IFN-λ4 through pseudogenization is strongly associated with improved clearance of hepatitis C virus (HCV) infection. We show that functional IFN-λ4 is conserved and evolutionarily constrained in mammals and thus functionally relevant. However, the pseudogene has reached moderately high frequency in Africa, America, and Europe, and near fixation in East Asia. In fact, the pseudogenizing variant is among the 0.8% most differentiated SNPs between Africa and East Asia genome-wide. Its raise in frequency is associated with additional evidence of positive selection, which is strongest in East Asia, where this variant falls in the 0.5% tail of SNPs with strongest signatures of recent positive selection genome-wide. Using a new Approximate Bayesian Computation (ABC) approach we infer that the pseudogenizing allele appeared just before the out-of-Africa migration and was immediately targeted by moderate positive selection; selection subsequently strengthened in European and Asian populations resulting in the high frequency observed today. This provides evidence for a changing adaptive process that, by favoring IFN-λ4 inactivation, has shaped present-day phenotypic diversity and susceptibility to disease.
| The genetic association with clearance of Hepatitis C virus (HCV) is one of the strongest and most elusive known associations with disease. The genetic variant more strongly associated with improved HCV clearance inactivates the recently discovered IFNL4 gene, which encodes for antiviral IFN-λ4 protein, and turns it into a polymorphic pseudogene. We show that functional IFN-λ4 is conserved and functionally important in mammals. In humans though the inactivating mutation appeared in Africa just before the out-of-Africa migration and quickly became advantageous, with the strength of selection (the degree of advantage) varying across human groups. In particular, selection became stronger out of Africa and was strongest in East Asia, raising the frequency of the pseudogene and resulting in the virtual loss of functional IFN-λ4 protein in several Asian populations. Although the environmental force driving selection is unknown, this process resulted in variable clearance of HCV in modern human populations. The complex selective history of IFNL4-inactivating allele has thus shaped present-day heterogeneity across populations not only in genetic variation, but also in relevant phenotypes and susceptibility to disease.
| Interferon-lambda (IFN-λ) proteins induce antiviral effectors in host target cells and have a crucial role in immune defense against pathogens [1]. The IFNL family classically included three genes (IFNL1, IFNL2, and IFNL3; formerly IL29, IL28A, IL28B, respectively) located within a 50 kb region of chromosome 19 [2], [3]. Several intergenic variants within the IFNL cluster had been identified as showing remarkable association with clearance of hepatitis C virus (HCV) [4]–[6], which is worldwide responsible for ∼170 million infections and over 350,000 deaths per year [7], [8]. The underlying functional basis of this association remained unclear despite numerous efforts to identify functional consequences of these variants [9]–[16].
An additional member of the IFN-λ family has recently been discovered: IFN-λ4, which bears only 30% amino acid identity with the other IFN-λs and is encoded by the IFNL4 gene, also located within the IFNL locus [2], [17]. IFN-λ4 shows similar antiviral activity like IFN-λ3 but as it shows limited secretion it might also act intracellularly, unlike the other IFN-λs [18]. A compound di-nucleotide exonic variant (rs368234815, ΔG>TT) in IFNL4 causes a frame-shift of its open reading frame and results in the polymorphic pseudogenization of IFNL4 - the polymorphic loss of IFN-λ4 protein [17]. The existence of IFNL4 was not even computationally predicted because the human reference genome contains the TT allele and lacks the IFNL4 open reading frame [17]. Remarkably, the derived TT allele not only eliminates IFN-λ4, but it also shows the strongest genetic association reported to date with improved spontaneous and treatment-induced HCV clearance [17], [19], [20].
The function of IFN-λ proteins is crucial for response to pathogens and this locus has evolved under natural selection, with signatures of positive selection being described in the three classical IFNL genes (IFNL1-3) [21]. However, that analysis did not cover the IFNL4 gene, nor the frame-shift rs368234815 variant, which were then unknown [21]. Therefore, the evolutionary history of this interesting functional variant and its influence on the local signatures of selection remained unknown.
Here we report an in-depth comparative and population genetic analysis that focuses on IFNL4 and the rs368234815 polymorphism. We show that the functional IFN-λ4 protein is under purifying selection in mammals, while in humans the IFNL4 pseudogenizing TT allele carries strong signatures of positive selection. We use a new Approximate Bayesian Computation (ABC) approach [22], [23] to provide evidence of a complex selective history of the TT allele, which involves changes in selective strength across human populations. This selective process had important implications in present-day phenotypic diversity and susceptibility to disease.
The IFNL4 gene is present in most mammals analyzed, although it is absent in mouse and rat (Methods). To understand the evolutionary conservation of IFNL4 we performed a comparative analysis of the IFNL4 coding sequences from a representative set of mammals (N = 12). The overall dN/dS (non-synonymous to synonymous substitution ratio) is 0.23 across mammals and 0.22 across primates (Figure 1), indicative of purifying selection maintaining the sequence and function of the protein. Notably, all individual branches except squirrel monkey have dN/dS<1 and no model of protein evolution supported dN/dS>1 in specific branches or sites (Table S1). This reveals strong evolutionary conservation of IFN-λ4 in mammals, reflecting its functional relevance.
The selective constraint on IFN-λ4 contrasts with the pseudogenization of the gene in humans through the derived TT allele [17]. The multiple-species alignment shows that ΔG is the conserved, ancestral allele and TT is the derived human-specific allele. The mutational process from ΔG to TT in humans is unclear, but only these two forms have been observed, so they should be considered as two alleles of a di-nucleotide variant (Methods). The TT allele shows considerable frequency variation across human groups. The 1000 Genomes data [24] reveals a gradient in frequency that rises from Africa (0.29–0.44) to Europe (0.58–0.77) and the New World (0.51–0.65), and reaches near fixation in East Asia (0.94–0.97) (Figure 2, Table S2, full population names in Methods).
Population differentiation can be quantified with the fixation index FST [25], a measure of the pairwise level of differentiation in allele frequencies. We used Yoruba (YRI) as the background population because it has the lowest frequency of the derived TT allele in Africa. To put these values in the context of genome-wide population differences, FST was also calculated for every SNP in the 1000 Genomes dataset. For the TT allele the largest FST, 0.63, corresponds to Southern Han Chinese (CHS) versus YRI, which places the TT allele in the 0.5% tail of the empirical genomic distribution of CHS-YRI FST (Fig. 3A, Table 1). FST is also in the 0.8% tail of the genomic distribution for the other East Asian populations (CHB, JPT, Fig. 3B and C, Table 1), and in the 4% tail for Europeans (CEU) and one African population, Luhya (LWK) (Table 1, Fig. S1). These results remain significant when other populations were used as background and in continental comparisons, and when the genome-wide distribution was restricted to SNPs with the lowest frequency in Yoruba (Table S3). Therefore, rs368234815 is among the 0.8% most differentiated SNPs between African and East Asians, and among the 12% most differentiated SNPs between African and European populations.
The unusually high population differentiation of the TT allele is compatible with a scenario of recent population-specific natural selection. Under certain selection models such high differentiation should be accompanied by extended haplotype homozygosity in the populations experiencing selection, but not in other populations. We evaluated such a signature with the cross population extended haplotype homozygosity test [26] (XP-EHH), which was calculated across the genome relative to the Yoruba population. The XP-EHH value for the TT allele is in the 0.5% tail of the empirical distribution for East Asian populations (p = 0.003, 0.005, and 0.003, for CHS, CHB, and JPT, Fig. 3A–C, Table 1), and the signal remains significant when calculated relative to a European population (GBR) and in analyses at the continental level (Table S4). In addition, some non-Asian populations show marginally significant signatures of positive selection too (CEU, PUR, LWK, Table 1, Fig. S1). Similar results were obtained with iHS, a statistic that explores haplotype homozygosity within a single population [27] (Table 1) (although iHS lacks power when population frequency is very high, like in Asia). The unusual allele-specific haplotype homozygosity is evident in Figure 3D, which shows the haplotype structure of the locus in one African, one European, and one East Asian population (for all populations see Fig. S3). We note that FST shows very weak correlation with both XP-EHH and iHS in the genome (r = 0.12 and r = −0.08 Spearman rank test, respectively, although the large number of data points makes these weak correlations significant, P-value<2.2e-16). Therefore, the FST and XP-EHH/iHS observations can be considered largely independent.
Finally, not only rs368234815 itself but also its genetic locus shows signatures of recent positive selection, with significant Fay and Wu's H test [28] (FW), which detects an excess of high-frequency derived alleles in the region (Table 1). Together, the combined signatures of FST, XP-EHH, iHS and FW provide strong evidence for the action of natural selection rapidly increasing the frequency of the TT allele in East Asia. The signature outside Asia is less clear, with most populations showing significant signatures of selection for a subset of the tests performed.
A classical problem in population genetics is the identification of the genetic variant responsible for a selection signal. High linkage disequilibrium (LD) in the region surrounding IFNL4 (Table 2, Fig. 3D, Fig. 4, Fig. S7) hampers the distinction of signatures across all the linked variants, making it difficult to identify the causal variant. We conclude that rs368234815 is the most likely variant driving the signatures of selection, based on three lines of evidence: (1) its functionality and phenotypic consequences, (2) its genetic association with viral clearance, which reflects its effect on fitness, and (3) its signatures of selection.
First, the TT allele has a clear phenotypic consequence as it leads to abrogation of IFN-λ4. This is in contrast with other variants in the locus for which no conclusive functional data has been reported despite numerous efforts [9]–[14]. Second, of all variants in the IFNL region, rs368234815 shows the strongest genetic association with spontaneous and treatment-induced HCV clearance in African Americans [17], [19]; in Europeans and Asians the strong LD across the region results in comparable associations for many variants [15]–[17], [20], [29] (Table 2, Fig. S2, Fig. S7). Third, of all protein-coding or HCV-associated variants in this locus, rs368234815 shows the strongest combined signatures of positive selection in East Asians (Fig. 3A–C, Fig. S1 and S2, Table 2). Only one other polymorphism (intergenic rs8109886, located upstream of IFNL4, Fig. 4), shows signals of selection comparable to rs368234815 (Fig. S2, Table 2). No function has been ascribed to this variant despite a moderate HCV association that is likely due to linkage to TT [6], [17], [30] (Table 2 and Fig. 4), making it a priori a less likely candidate for selection. Indeed, simulations of the evolutionary process showed that the large frequency change of rs8109886 can be explained by linkage to the TT allele alone (Note S1).
We also put rs368234815 in the context of the signatures of selection in the larger genomic region. IFNL4 is located upstream of IFNL3 in a region of moderate LD that is separated from the IFNL1/IFNL2 locus by a recombination hotspot (Fig. S7, Table 2). Manry et al. [21] identified signatures of recent positive selection in all three original IFNL genes (IFNL1-3) but neither IFNL4 nor the rs368234815 variant were known at that time and thus they were not considered. The recombination hotspot breaks LD between the IFNL1/INFL2 locus and the IFNL3/IFNL4 locus (Fig. S7, Table 2), showing that these signatures are in all likelihood independent, as suggested by Manry et al. [21]. There is moderate LD between IFNL4 and IFNL3, with an average r2 between rs368234815 and IFNL3 SNPs of 0.18 in CEU and of 0.44 in CHS (see also Fig. S7). So, the selection signatures in IFNL3 and IFNL4 may not be independent. In fact, the seven SNPs identified by Manry et al. [21] (detailed in Table 2 and Fig. 4) have (1) weaker signatures of selection; (2) unclear functional effects, and (3) weaker association with HCV clearance than the TT allele in Africa (Table 2, Figure 4). Also, those that show some signatures of selection have high to moderate LD with rs368234815 (Table 2), with LD broken mostly by a few recombination events in the ancestral haplotype (Note S1). Taken together, these lines of evidence confirm that IFNL1/2 and IFNL3/IFNL4 have likely been independently targeted by positive selection in recent human history, as suggested by Manry et al. [21], and highlight rs368234815 TT as the most likely selected allele in its region.
The classical model of positive selection involves selection on a de novo mutation (SDN), a so-called hard sweep, where a new mutation immediately becomes beneficial and selected (reviewed in [31]). This scenario is difficult to reconcile with our observations, because unequivocal signatures of selection are observed only in East Asians but the TT allele is common worldwide. The TT-carrying haplotype harbors the highest genetic diversity in Africa indicating that it arose there before the out-of-Africa dispersion (Note S2, Table S5), a result that is consistent with the IFNL4 haplotype network (Fig. S4). Under SDN, only a model where selection begins weak in Africa and becomes stronger outside of Africa could explain our observations (Fig. 5A). An alternative model is selection from standing variation (SSV), also known as a soft sweep (reviewed in [31]). In this scenario an existing neutral or nearly neutral allele becomes advantageous, for example upon environmental change (Fig. 5A).
To disentangle the most likely model of selection for the TT allele we applied a modified version of a recently published ABC approach [23], which we extended to be able to analyze two-population models. In brief, we match millions of simulations under the different models to a summary of the observed genetic data in the IFNL4 region, and use the best matching simulations for further inferences. Under reasonable assumptions we expect the most realistic selection model to produce the closest simulations to real data, and thus simulations can be used to make inferences about the selective history of the allele [23] including the model of selection and relevant parameters (Note S3). While the method relies on some assumptions (e.g. correct demographic and dominance models) this approach has been shown to be robust and to have high power to recover the correct selection scenario [23]. We assess that we overall have high power to recover the correct model, with 76% of the SSV and 95% of the SDN simulations being assigned correctly under the East Asian demographic model, and 70% of the SSV and 97% of the SDN simulations being assigned correctly under the European demographic model (Note S3). The slight bias observed was considered when interpreting the results. For our analysis we consider three models: neutrality (no selection), selection from a de novo mutation (SDN) and selection from standing variation (SSV) (Fig. 5A).
In East Asian populations we obtain negligible support for neutrality and very strong support for the SDN model (Fig. 5B, Table 3, Table S6). Results in Europeans are also consistent with the SDN model, although the weaker signals of selection and the slight bias observed above make these results less conclusive (Table 3, Note S3, Fig. S5, Table S6). The posterior probability for the SDN model is ∼95% in East Asia and ∼80% in Europe, corresponding to Bayes factors (Bayesian measures of relative model support [32]) of ∼10 and ∼4, respectively. This provides substantial and robust evidence for the SDN model, compared to the SSV and NTR models for East Asian and European populations according to Jeffrey's interpretation [33]. Therefore, we conclude that the TT allele was likely positively selected upon appearance. The ABC-based parameter estimates are less reliable than the model choice [23] because they always have large credible intervals (Bayesian measures of confidence). However, the posterior distributions have modes that differ from the modes of the prior distributions, indicating that they are determined by information from the data and not by the prior (Fig. S5). Also, the estimates are quite concordant within and between continental groups (Fig. S5, Table 3). So while they should be interpreted with appropriate caution, the estimates do provide additional useful information about the model and timing of selection. We infer that the TT allele emerged before the out-of-Africa migration (estimated tmut≈55,900 years ago (41,360–68,640)) and was immediately, or shortly thereafter, targeted by moderate positive selection (selection coefficient, sA, ≈0.58% (0.17–1.23)); we estimate that selection intensified substantially outside of Africa, with the selection strength nearly quadrupling in Europe and in Asia (sNA≈2.6% (0.6–4.8); Table 3, Fig. S5).
One important aspect of the simulations is the mode of dominance (also known as the genetic model), and the ABC analysis above was performed on simulations under a perfectly additive model where heterozygotes have half the fitness effect of homozygotes (dominance coefficient h = 0.5). This model is reasonable because in TT/ΔG heterozygotes only one IFNL4 copy is truncated, and because genetic studies show that the odds ratios (ORs) for HCV clearance in heterozygotes are intermediate to those in the two homozygotes [17]. These two arguments argue strongly against a model of complete dominance for TT as realistic, but other models are more difficult to discard a priory. We thus compare three dominance models: (1) a fully recessive model for the TT allele (h = 0), (2) the perfectly additive model used above (h = 0.5), and (3) a supra-additive model where the additive effect is non-linear and heterozygotes are closest in fitness to ΔG homozygotes. This model has been proposed based on the ORs for the intronic IFNL4 variant, rs12979860 which is in high LD with rs368234815 and is thus a good proxy for the dominance effects of TT (Table 2) [34]. Based on those results we use a dominance coefficient h = 0.38 (see Note S4). When we compare the three dominance models in East Asia, regardless of the selection model, the fully recessive model has marginal support (4%), with the two additive models showing similar posterior probabilities (slightly higher for additive: 56%, than supra-additive: 44%, Fig. 5C and Note S4). When we compare the ABC results in the two additive models, they both strongly support the SDN model over the SSV model (95% in the additive model and 90% in the supra-additive model, corresponding to a Bayes factor of ∼12), and both models provide virtually no support for the neutral model (Figure 5B–D and Note S4). Parameter estimates also agree well among these two models (Fig. S5, Fig. S6 and Note S4). Therefore our results in East Asia validate the use of an additive model and show that the ABC inferences are not sensitive to the particularities of the additive model used. In European population the results are less clear, just as in the original ABC analysis and as expected given the weaker signatures of selection. Still, these results also support the two additive models (36% support for additive and 38% for supra-additive; Fig. 5C) as well as the SDN model (∼81% support for SDN in both the additive and the supra-additive, corresponding to a Bayes factor ∼4, Fig. 5B and D, Note S4).
These results show a complex selection history for the TT allele, with selection starting upon appearance of the allele but with intensity changing over time and geographic range. The model is consistent with all our observations, including the marginal evidence for selection observed in non-Asian populations (Table 1). It is interesting that we infer selection on the TT allele even in Yoruba, where the signature is undetectable with classical methods likely because of weak selection and lower frequency although the TT allele shows clear signatures of homozygosity (Fig. 3D). Interestingly, and in agreement with this model, we do observe some signatures of positive selection in another African population, the Luhya. It remains possible that the advantage of the TT allele was counteracted by additional selective forces in Africa that maintained the TT allele at an intermediate frequency, such as balancing selection, although we note that the locus lacks classical signatures of long-standing balancing selection (Note S5, Table S7).
Here we show that functional IFN-λ4 is under purifying selection throughout the mammal clade while positive selection has favored the elimination of IFN-λ4 through pseudogenization in humans. Selection on the TT allele has been particularly strong in specific populations, leading to extremely high frequency of the pseudogene and subsequent virtual loss of IFN-λ4. This event is phenotypically relevant: not only is IFN-λ4 biologically important [17], [18] and evolutionary conserved, but the loss of IFN-λ4 through pseudogenization shows remarkable association with improved HCV clearance [17], [19], [20].
The precise reason behind the advantage of IFN-λ4 elimination is unknown, but its immunological role and clear antiviral activity against HCV [17] make exposure to pathogens (and in particular viral agents) the most likely selective force. However, due to its slow progression into fatal disease [35] HCV is unlikely to have exerted such strong selective pressure, although we cannot completely discard this possibility. Besides HCV, it has been shown that functional IFN-λ4 has antiviral activity against coronaviruses [18], while the IFN-λ4 pseudogene increases susceptibility to cytomegalovirus retinitis among HIV-infected patients [36]. Suggesting that IFN-λ4 pseudogenization is likely associated with several phenotypic traits. It is perhaps surprising that suppression of an antiviral protein results in improved viral clearance, although it has for example been shown that during chronic infection blockage of persistent signaling of IFN I (a different type of interferon) can improve viral clearance [37], [38].
We showed that a complex selective regime, with variation in selection strength in different geographical areas, best explains the history of the IFNL4 locus. Signatures of non-neutral evolution have been detected in other interferons, including at least one other IFNL family member (IFNL1 or IFNL2) [21]. Although the mode and tempo of selection in these other IFNL genes are not well understood, together these observations suggest that IFN-λ proteins have played an important role in recent human adaptation, probably as a consequence of their role in individuals' constant fight with pathogens. It is likely, though, that only the selective history of the IFNL4-TT allele had a strong influence in the rate of clearance of some viruses, at least HCV, across human groups.
It has been proposed that gene loss may exert an important role in evolution, including human evolution [39], and the loss of otherwise conserved regulatory elements may play a role in the acquisition of human-specific phenotypes [40]. Loss-of-function mutations show global signatures of purifying selection [41]–[43] and tend to carry detrimental effects [44]. A few exceptions exist, though, where truncating polymorphisms show signatures of positive or balancing selection [45]–[50]. Still, as with other targets of selection, most of these cases lack biological interpretation. In fact, IFNL4 joins a small group of known genes where a striking signature of local adaptation is coupled with a clear molecular phenotype (e.g. [46], [47], [51]), which in this case is also associated with disease risk. As such, it contributes to our understanding of how recent human evolution has shaped genetic and phenotypic human diversity, including present-day heterogeneity in susceptibility to disease.
In order to explore the level of functional constraint in IFNL4, we estimated the level of protein conservation in primate and non-primate mammals. Specifically, we assessed the ratio (dN/dS) of non-synonymous substitutions per non-synonymous site (dN) to synonymous substitutions per synonymous site (dS) across gene orthologs. Since purifying selection eliminates deleterious protein-coding changes, dN/dS decreases with negative selection and increases with relaxed constraint and positive selection.
We used human IFNL4 reference sequence NM_001276254.2 to BLAT genomes of other species and generate multiple-species sequence alignment of IFNL4 coding exons 1 through 5 (Table S8). The panda-predicted IFNL4 ortholog was subsequently used as BLAT query to extract coding exons for additional non-primate species (Table S8). Further, we sequenced IFNL4 (exons and introns) in genomic DNA and reconstructed complete IFNL4 cDNA sequences of chimpanzee (Genbank accession JX867772), baboon (Genbank accession KC525947) and crab-eating macaque (Genbank accession KC525948). The whole IFNL4 genomic region is absent in mouse or rat. All discovered functional IFNL4 sequences (Table S8) where used for a multiple-sequence alignment which was created using ClustalW [52] and annotated with Jalview [53].
The alignment was analyzed with codeml (part of PAML4 [54]) to test various models of selection. We estimated the overall dN/dS for the complete tree and compared likelihoods for models that allowed: i) free dN/dS for each branch (i.e., lineage heterogeneity); ii) a primate-specific dN/dS; and iii) a human-specific dN/dS. Additionally, we performed tests aimed to detect site-specific signatures of positive selection across the phylogeny (branch models): i) model 1a (neutral) vs. model 2 (positive selection); ii) model 7 (neutral) vs. model 8 (with dN/dS>1); and iii) model 8a (with dN/dS = 1) vs. model 8 (with dN/dS>1).
We analyzed genome-wide data from the 1000 Genomes release (2010/11/23; phase I) [24]. We considered (1) autosomal variants detected in the low coverage sequencing, and (2) populations with information for at least 50 unrelated individuals, which was met by 13 populations from four different continents [African ancestry: YRI (Yoruba in Ibadan, Nigeria), LWK (Luhya in Webuye, Kenya), ASW (African Ancestry in Southwest US); European ancestry: GBR (British from England and Scotland), CEU (Utah residents (CEPH) with Northern and Western European ancestry), FIN (Finnish from Finland), TSI (Toscani in Italia); East Asian ancestry: CHS (Han Chinese South), CHB (Han Chinese in Beijing, China), JPT (Japanese in Toyko, Japan); American ancestry: MXL (Mexican Ancestry in Los Angeles, CA), CLM (Colombian in Medellin, Colombia), PUR (Puerto Rican in Puerto Rico)]. As an exception PUR (Puerto Rico) was analyzed although it contains only 44 individuals. Some analyses were performed both by population and by continent; in these cases the continental groups contain 150 randomly selected, unrelated individuals (America 144) with an equal contribution from each population within the continent.
For the rs368234815 ΔG/TT frameshift-substitution variant the 1000 Genomes dataset only contains the T insertion/deletion variant rs11322783 (-/T, chr19:39739154, dbSNP b138), while the substitution rs74597329 (G/T, chr19:39739155, dbSNP b138) is absent. This is due to the automatic variant caller failing to correctly identify an insertion and a substitution in the same genomic position. We sequenced an amplicon containing rs368234815 in 153 individuals included both in the 1000 Genomes and HapMap sets (CEU, YRI and CHB/JPT). Sequencing confirmed the presence of only two alleles (ΔG and TT) and showed good concordance with the 1000 Genomes data between our ΔG/TT genotypes and 1000 Genomes genotypes for the overlapping insertion/deletion variant rs11322783 (4 individuals of 153 tested were discordant, providing an estimated 97.4% genotype and 98.7% allele concordance rate). This validated the use of 1000 Genomes dataset for our subsequent analyses. We used the ancestral allelic state annotated in the 1000 Genomes data, which is based on the Ensembl 59 comparative 32 species alignment [55]; only SNPs with a high-confidence ancestral inference were used, and indels were excluded due to their cryptic variation patterns [56].
We used FST, iHS and XP-EHH to explore the signatures of selection of rs368234815 TT allele. FST is a measure of population differentiation and unusually high FST can indicate population-specific positive selection that drastically increases allele frequency in the population under selection [57]. To calculate FST we used the Weir and Cockerham [25] estimator implemented in vcf-tools [58].
Positively selected alleles rapidly increase in frequency with recombination having little chance to break their association with nearby variants. If the selected allele was originally in few haplotype backgrounds and it has not reached fixation, it will be associated with extended haplotype homozygosity (EHH), a pattern that will be absent for the non-selected allele. We used two statistics to explore this expectation. First, iHS [27] measures the allele-specific decay of EHH within a population by comparing the associated EHH of ancestral and derived alleles. Second, XP-EHH [26] that detects alleles that are under selection in one population only, by comparing EHH patterns both among allelic types and across populations; as such XP-EHH has higher power to detect population-specific selection. Low frequency variants break the EHH signal, so following [59] we considered only SNPs with derived allele frequency ε 5% for XP-EHH or minor allele frequency ε 5% for iHS. Local recombination rate estimates were obtained from a combined recombination map based on HapMap data [60] from Africa, European, and Asian populations. Both statistics were standardized to a mean of zero and a standard deviation of one; for iHS, scores were then binned by frequency (1%) as previously suggested [27]. Correlation of FST with XP-EHH (CHS vs. YRI) or iHS (CHS) was calculated for all variants present in the respective dataset with Spearman's rank correlation test implemented in R [61].
We used each of these statistics to analyze every non-African population; for between-population comparisons we used Yoruba as background, unless noted otherwise. To assess the putative effects of this choice of populations we repeated the analyses for continental groups, for different background populations, and for SNPs that have their lowest allele frequency in Yoruba. In all cases the empirical P-values were obtained by comparing the score for rs368234815 to the whole-genome empirical distribution of the respective statistic. Since this is a hypothesis-driven analysis with a single variant analyzed within a single locus, no multiple testing or genome-wide corrections are needed.
We also applied tests that analyze the signatures of selection in the IFNL4 genetic region (∼2.5 kb). Here we show results for Fay and Wu's H test [28], which detects the excess of high-frequency derived alleles expected after a recent sweep with recombination. Significance was estimated using 10,000 standard neutral coalescent simulations [62]. Because demography affects the SFS and can cause spurious results if not properly accounted for, our simulations are run under a demographic model which includes inferred parameters for populations of African [63], European [63], Asian [63] and American [64] ancestry. A custom made perl program (Neutrality Test Pipeline) was used to calculate the statistic and corresponding P-value.
To infer the model of selection that best fits IFNL4 data and estimate the timing and selection strength of the TT allele, we used an Approximate Bayesian Computation (ABC) approach [22]. In particular, we followed a published approach [23], which has been previously shown to discriminate well between SDN, SSV and neutrality (NTR) [23]. In brief, this approach is based on performing a large number of simulations under different selection models, with random parameters drawn from some probability distribution (called the prior distribution). Real data and simulations are compared based on summary statistics, and through a rejection scheme the simulations that most closely resemble real data help inform inferences about the best-fitting model. The parameter values that generate these simulations are then used to obtain the posterior distribution of each parameter, whose mean and standard deviation are used to perform the parameter inferences. We extended the method to consider more than one population, since two-population statistics are most informative in our case.
Specifically, the approach uses msms [65] to simulate data, custom python scripts to calculate all summary statistics, and ABCtoolbox [66] for all ABC inferences. Under both selection models, we started with uniform priors with a range as follow (see Fig. 5A):
Because simulations with the selected allele fixed are likely to be very different from the observed data, we conditioned on the selected allele segregating in both populations. This resulted in non-uniform prior distributions presented in Figure S5 and S6. We used 104 simulations to distinguish between the neutral model and the two selection models, and a larger set of 8×105 simulations for the more subtle distinction between the two selection models and for parameter estimation. For the simulations, we used the population history model estimated by Gravel et al. [63] and assumed a constant recombination rate of 1.76 cm/Mb throughout the region (average recombination rate in the IFNL locus [60]), and a perfectly additive model of dominance (h = 0.5). Lack of an appropriate demographic model for American and non-Yoruba African populations precludes analysis for those populations. The following single-population statistics were calculated: the average number of pairwise differences π, Watterson's θ, Fay and Wu's H [28] and Tajima's D [67], all for both 4 kb around the site and a 8 kb (6 kb upstream and 2 kb downstream of the site) interval around the TT allele. The between-population statistics employed were: FST [68] for the selected site, FST in 4 kb around the site, FST for the whole region, and XP-EHH on the selected site [26]. In addition, we also included the frequency of the selected allele in both populations. This resulted in a set of 16 summary statistics, which, following Wegmann et al. [69] and Peter et al. [23], was reduced to seven summary statistics using PLS-DA [70] for model choice and regular PLS for parameter inference [71]. Performance of the ABC model choice and parameter distribution for the SDN model has been assessed for each particular model (Note S3). Confidence in the choice of selection models has been supported with Bayes factors.
In addition, we investigated the influence of the dominance model in our inferences. We analyzed a recessive model for TT (h = 0), the perfectly additive model above (h = 0.5), and a supra-additive model (h = 0.38), using 500,000 simulations for each model. We run an ABC analysis for model selection with all simulations (from all three dominance models and the three selection models NTR, SDN, and SSV). We then assess the posterior probability of each dominance model regardless of selection model, and the posterior probability (and parameter estimates) of each selection model for the additive and supra-additive dominance models (see Note S4).
1000 Genomes, ftp://ftp.1000genomes.ebi.ac.uk/vol1/ftp/phase1/; GENCODE, http://pseudogene.org/psidr/; HapMap, http://hapmap.ncbi.nlm.nih.gov; XP-EHH and iHS executables, http://hgdp.uchicago.edu/Software/; VCFtools, http://vcftools.sourceforge.net; ABCtoolbox: http://www.cmpg.iee.unibe.ch/content/softwares__services/computer_programs/abctoolbox/index_eng.html; msms: http://www.mabs.at/ewing/msms/
|
10.1371/journal.pgen.1002870 | Fine-Mapping and Initial Characterization of QT Interval Loci in African Americans | The QT interval (QT) is heritable and its prolongation is a risk factor for ventricular tachyarrhythmias and sudden death. Most genetic studies of QT have examined European ancestral populations; however, the increased genetic diversity in African Americans provides opportunities to narrow association signals and identify population-specific variants. We therefore evaluated 6,670 SNPs spanning eleven previously identified QT loci in 8,644 African American participants from two Population Architecture using Genomics and Epidemiology (PAGE) studies: the Atherosclerosis Risk in Communities study and Women's Health Initiative Clinical Trial. Of the fifteen known independent QT variants at the eleven previously identified loci, six were significantly associated with QT in African American populations (P≤1.20×10−4): ATP1B1, PLN1, KCNQ1, NDRG4, and two NOS1AP independent signals. We also identified three population-specific signals significantly associated with QT in African Americans (P≤1.37×10−5): one at NOS1AP and two at ATP1B1. Linkage disequilibrium (LD) patterns in African Americans assisted in narrowing the region likely to contain the functional variants for several loci. For example, African American LD patterns showed that 0 SNPs were in LD with NOS1AP signal rs12143842, compared with European LD patterns that indicated 87 SNPs, which spanned 114.2 Kb, were in LD with rs12143842. Finally, bioinformatic-based characterization of the nine African American signals pointed to functional candidates located exclusively within non-coding regions, including predicted binding sites for transcription factors such as TBX5, which has been implicated in cardiac structure and conductance. In this detailed evaluation of QT loci, we identified several African Americans SNPs that better define the association with QT and successfully narrowed intervals surrounding established loci. These results demonstrate that the same loci influence variation in QT across multiple populations, that novel signals exist in African Americans, and that the SNPs identified as strong candidates for functional evaluation implicate gene regulatory dysfunction in QT prolongation.
| The QT interval (QT) provides a measure of a ventricular action potential, and its prolongation is associated with sudden death and ventricular arrhythmias. Genome-wide association studies performed in European populations have identified common genetic variants that influence QT. However, it is unclear whether these variants are relevant in other populations, including African Americans. The increased genetic diversity in African Americans also provides opportunities to narrow association signals and identify candidates for functional evaluation. We therefore used data from 8,644 African Americans to further characterize previously identified QT loci. Of the fifteen known independent QT variants at the eleven previously identified QT loci, six were associated with QT in African Americans. We also identified three variants that were independent from previously reported signals and narrowed intervals flanking association signals using patterns of linkage disequilibrium. Finally, bioinformatic-based characterization pointed to candidates located outside protein coding regions. Our results underscore the utility of genetic studies in African ancestral populations to identify novel variants and narrow intervals surrounding established loci. These results suggest that known QT loci are important in African Americans and that further characterization of these loci in other populations may provide additional insights into the genetic and molecular mechanisms underlying QT.
| The QT interval (QT), as measured by the resting 12-lead electrocardiogram (ECG), reflects the duration of ventricular depolarization and repolarization, providing a non-invasive assessment of an average ventricular action potential. QT prolongation is an established risk factor for ventricular tachyarrhythmias [1], coronary heart disease [2], and sudden cardiovascular as well as all-cause death [3]. Although numerous factors influencing QT have been identified, including heart rate [4], structural heart disease [5], [6], gender [7], [8], age [9], and medication use [10]–[12], a large portion of the variance in QT remains unexplained.
Several lines of evidence support a genetic contribution to QT. Initial evidence was provided by studies of inherited cardiac arrhythmias including long- and short-QT syndromes, which identified rare and highly penetrant mutations in ion channel and ion channel associated genes associated with QT [13]. Family studies have also reported that ventricular repolarization (as measured by QT) is heritable [14]–[17]. In addition, recent genome-wide association (GWA) studies performed in populations of predominantly European descent have identified common SNPs in twelve loci, including NOS1AP, KCNH2, and PLN [18]–[23], that influence the distribution of QT.
To date, the majority of published GWA studies examining QT have been performed in populations of European descent, although one study also examined an Indian Asian population [23]. It is therefore unclear whether previously identified QT loci are relevant in other racial groups such as African Americans or whether there are population-specific SNPs influencing QT. Furthermore, the increased genetic diversity in populations of African ancestry provides opportunities for the narrowing and fine-mapping of loci identified in European and Indian Asian populations [24]. Fine-mapping, which includes the dense genotyping of common and rare SNPs at already established loci, is a helpful next step in the identification of functional polymorphisms underlying the QT distribution. For example, dense genotyping can capture rarer SNPs that may be inadequately represented on frequently used genome-wide genotyping arrays, and which may have large effects, thereby potentially helping to explain a larger fraction of the QT heritability [25], which is estimated to range from 35% to 52% [14]–[17].
In this study, we evaluated eleven QT loci previously identified in populations of European and Indian Asian descent in 8,644 African American participants from the Atherosclerosis Risk in Communities study (ARIC) and Women's Health Initiative clinical trial (CT). In addition to testing the previously reported QT index SNPs at the known loci, we also searched for stronger markers of the index signal and investigated evidence for independent, novel SNPs influencing QT in African Americans. For loci associated with QT in African Americans, we also investigated whether patterns of linkage disequilibrium (LD) within African Americans could narrow the regions likely to harbor the biologically relevant variant. Finally, we queried bioinformatic databases and performed related in silico analyses to identify potential candidate polymorphisms for follow-up functional evaluation.
Participants are drawn from two separate studies (Methods), which together comprise 9,702 individuals. After applying the exclusion criteria (Methods), phenotypic data from 8,644 participants are included in this analysis. The majority of participants are female (86%) and ages range from 45–79 years (Table S1). Estimates of mean QT and heart rate are consistent across the studies. Approximately 60% of participants are directly genotyped on the Metabochip; the remaining 40% have genotypes imputed from the Affymetrix 6.0 panel (Methods).
The Metabochip, a custom array containing approximately 195,000 SNPs, is designed to facilitate fine-mapping of loci associated with cardiovascular and metabolic traits, including QT, blood pressure, cholesterol, type 2 diabetes, and anthropometrics. To identify and fine-map signals associated with QT in African Americans, 6,670 SNPs from eleven previously identified QT loci represented on the Metabochip are examined (Table S2; Methods). The number of SNPs at each locus with minor allele frequency (MAF) estimates ≥1% ranges from 51 to 1,371, corresponding to regions spanning 67 Kb to 664 Kb in size (median size: 275 Kb).
On average, imputation quality for metabochip SNPs imputed in WHI SNP Health Association Resource (SHARe) participants (n = 3,531) is high across the 11 loci, with the exception of RNF207, KCNH2, and KCNQ1 (Table S2). As described in the Methods, SNPs with imputation quality scores <0.95 are discarded.
The six published GWA studies of QT (five studies of European ancestral populations, one study of European and Indian Asian ancestral populations) reported 25 index SNPs (P≤5.0×10−8, Table 1; D' estimates provided in Table S3) across eleven loci, which together represent fifteen independent signals (at r2≥0.20 in European ancestry populations). The fifteen signals include three independent signals at NOS1AP and two independent signals at both PLN and KCNH2.
We first test the fifteen independent signals for association with QT in African Americans. Specifically, we evaluate the index SNPs for each of the fifteen independent signals as well as all SNPs in LD with the index SNPs (r2≥0.20 using European ancestral LD patterns) to determine whether any of the fifteen independent signals generalize to African Americans. The significance criterion, αa = 1.20×10−4, is based on the number of tag SNPs in African Americans that capture (r2≥0.80, using African American LD patterns) all SNPs that are correlated with the index SNPs (r2≥0.20; determined using European ancestral LD patterns; see Methods).
Six of the fifteen independent signals are significantly associated with QT in African Americans (Table 1): NOS1AP independent signals 1 and 2, ATP1B1, PLN independent signal 1, KCNQ1, and NDRG4. Of those that are not significantly associated with QT (P-value>1.20×10−4), estimates for index SNPs representing the two KCNH2 independent signals as well as the RNF207 index SNP show a consistent magnitude and direction of effect when compared to published estimates (Table S4).
The best marker in African Americans for the first NOS1AP independent signal (rs12143842) and KCNQ1 (rs12296050) is identical to the European index SNP. For the ATP1B1 and NDRG4 index SNPs as well as the second NOS1AP independent signal, the best marker in African Americans shows only a slightly more significant association than the index signal (<1 order of magnitude change in the P-value; Table 1). In contrast, at the first PLN independent signal, the three index SNPs are not significantly associated with QT in African Americans (P-value range: 0.034–0.60), although a substantially stronger marker of the index signal is detected (rs56403768; P-value = 3.8×10−5). In Europeans, rs56403768 is correlated with the three index SNPs (LD r2 range: 0.68–0.88). However, patterns of correlation are weaker in African Americans (LD r2 range: 0.39–0.50).
We then evaluate evidence for additional independent signals at the eleven previously identified QT loci, focusing on SNPs that are uncorrelated with the index signals in European populations. Here, statistical significance is defined using an efficient Monte Carlo approach that accounts for LD between SNPs at the eleven previously identified QT loci (αb = 1.37×10−5; see Methods). Eleven SNPs at two loci that are uncorrelated with the index SNPs (r2≤0.20; evaluated using both African American and European ancestral LD patterns) exceed our significance threshold. Conditional analysis confirms that these eleven SNPs represent three novel associations, one flanking the NOS1AP locus and two residing within or nearby ATP1B1 (Table 2; Figures 1 and S1). Of note, the novel NOS1AP SNP is monomorphic in populations of European ancestry. Together, the six best markers in African Americans and the three novel SNPs explain 1.6% of the variance in heart rate-corrected QT.
Next, we examine whether African American LD patterns can assist with the narrowing of association signals at the six loci that generalized to African Americans. An example of variation in LD patterns by ancestral population is shown by the first NOS1AP independent signal. Among African Americans, 0 SNPs are correlated (r2≥0.50) with rs12143842, which is the best marker in African Americans and the index SNP reported by three prior QT GWA studies (Figure 1; Table 3) [20]–[22]. This is in contrast to LD patterns estimated in Europeans for NOS1AP independent signal 1, where 87 SNPs are correlated with the three index signals that characterize this independent signal (r2≥0.50), representing a region spanning 114.2 Kb. For the five remaining markers (Figures 1 and S1, S2, S3, S4), African American populations exhibit lower levels of LD as compared to European populations (mean narrowing = 48.2 Kb). Likewise, fine-mapping in African Americans helped to narrow the association signals for all six loci.
Bioinformatic analysis of the six best markers in African Americans and the three novel and independent SNPs (Table S5; Methods) does not identify any correlated non-synonymous coding variants; instead, all signals harbor variants that occur solely within non-coding regions with the potential for influencing cis-regulation (Table S6). Several variants occur within candidate regulatory elements (promoter regions and DNase hypersensitive sites in human cardiomyocytes), including three (rs3864884, rs1646010, and rs27097) that are predicted to have allele-specific binding affinities for various transcription factors of relevance to cardiac function (Table S7). Though further functional characterization is outside the scope of this study, rs3864884, rs1646010, rs27097, and rs37036 represent compelling candidates for follow-up evaluation.
In this study composed of approximately 8,600 African American participants, we evaluated fifteen independent signals across eleven loci that were previously associated with QT in populations of European and Indian Asian ancestry at genome-wide significant levels. For five independent signals – the two NOS1AP independent signals, ATP1B1, NDRG4, and KCNQ1 – the best markers in African Americans were either identical to or only slightly more significant than the index signal. These five markers are therefore not considered better signals than the index SNP. However, for the first PLN locus, the three previously identified index SNPs were not significantly associated with QT in African Americans. This result suggests that rs56403768, the best marker in African Americans, is a better proxy of a biologically important PLN allele and may help improve localization of the true association.
In addition to generalizing six previously characterized QT loci, we also identified three novel and independent signals for NOS1AP and ATP1B1. Notably, rs79163067, the novel NOS1AP signal, was monomorphic in European populations. When these three novel and independent variants were combined with the fifteen independent QT loci previously identified in populations of European and Indian Asian ancestry, our results suggest that to-date at least 18 independent variants influence QT.
Finally, we showed that evaluating LD patterns in admixed populations such as African Americans assisted with the narrowing of intervals flanking the putative causal variants. This narrowing was particularly evident for the first NOS1AP locus that included index SNP rs12143842, the most frequently reported SNP in the QT GWA study literature to-date. Rs12143842 also was the SNP that explained the majority of variance in heart rate-corrected QT as well as the best marker in African Americans for the first NOS1AP independent signal. Rs12143842 has yet to be evaluated in any functional studies, and although our bioinformatics characterization did not identify any compelling functional candidates, the SNP resides less than five Kb from the annotated NOS1AP promoter. Future studies evaluating the functional relevance of rs12143842 are clearly indicated.
Of the nine QT SNPs we identified (i.e. the six best markers in African Americans and the three novel SNPs), all resided in or were in LD with SNPs residing in candidate long-range regulatory elements in human cardiomyocytes, annotated promoter regions, and highly conserved non-coding elements, and as such, strongly implicate gene regulatory dysfunction in QT prolongation. One of the more striking predictions was with rs3864884, where the major allele [C] is predicted to bind an entirely different set of transcription factors (TFs) than the minor allele [T]. Notably, only the major allele is predicted to bind Hairy-related TFs, which are involved in regulating cardiac morphogenesis [26]. The minor allele is predicted to bind TBX5 and AhR; the former has been linked to a number of cardiac phenotypes including Holt-Oram syndrome [27] and atrioventricular conduction [28] and the latter regulates cardiac size [29], a known risk factor for QT-prolongation and cardiac sudden death [30]. Consistent with this prediction, the minor allele is associated with an increased QT in this population.
We were unable to identify associations with RNF207, SCN5A, KCNH2, LITAF, LIG3, and KCNJ2 based on our statistical significance threshold. RNF207 was reported in two previous QT GWA studies. However, because of design priorities, only 51 SNPs with MAF≥1% were available for analysis. Future fine-mapping efforts in African Americans or other admixed populations that include denser genotyping of this locus may therefore be useful.
SCN5A, KCNH2, and LITAF were all reported by two prior GWA studies of QT. SCN5A also has also been implicated in GWA studies of PR in populations of European [31] and African ancestry [28] as well as QRS interval duration in populations of European descent [32]. Our inability to detect signals at these three loci may simply reflect inadequate power, especially for KCNH2, for which estimates of the index SNPs in African Americans were consistent with the published literature. KCNJ2 is a biologically plausible locus influencing QT, as it harbors mutations causing rare, familial forms of long QT syndrome [33]. Yet, the high P-values and the dense genotyping coverage of SNPs suggest that KCNJ2 does not influence QT in African American populations.
The genetic architecture of African Americans and other admixed populations is on average characterized by lower correlation between SNPs when compared to European populations. Such populations therefore are valuable for the fine-mapping of previously identified loci, as fewer SNPs are expected to be correlated with the underlying functional variant, which is expected to be the same in populations of different ethnicity. We therefore anticipate that future fine-mapping efforts that include populations with different ethnic backgrounds will be useful for the further refinement of loci influencing QT as well as the identification of population-specific variants, as demonstrated by the current report.
Several limitations of the present study warrant further consideration in order to inform future efforts for fine-mapping and functional characterization of QT loci. First, although the Metabochip includes dense genotyping of most QT loci, it is possible that the causative variants are not included on the Metabochip. Second, our functional characterization is based on in silico analyses and requires experimental validation. Third, the majority of study participants were female. It is unclear how a predominantly female population may have influenced the results presented herein, considering the well-known dependence of QT on gender [7], [8]. Finally, our results, which are consistent with prior studies [22], show that common SNPs only explain a very small fraction of the variance in QT, although heritability estimates suggest a substantial genetic component. These modest effect sizes corroborate the multifactorial etiology of QT and demonstrate that substantially greater efforts are required to explain the “missing heritability”. Future efforts with increased sample sizes that examine rare variants, gene-gene and gene-environment interactions, and structural variants poorly captured on existing arrays are clearly needed [34].
In conclusion, our findings provide compelling evidence that the same genes influence variation in QT across ancestral populations and that additional, independent signals exist in African Americans. Moreover, all SNPs identified as strong candidates for functional evaluation implicate gene regulatory dysfunction. Further characterization of these loci, including direct sequencing and large-scale genotyping in African Americans and other admixed populations, may provide more information on the genetic and molecular mechanisms underlying QT.
The Institutional Review Board at all participating institutions approved the study protocol. This study was conducted according to the principles expressed in the Declaration of Helsinki.
The Population Architecture using Genomics and Epidemiology (PAGE) study is a National Human Genome Research Institute funded effort examining the epidemiologic architecture of common genetic variants that have been reproducibly associated with human diseases and traits [35]. The PAGE study consists of a coordinating center and four consortia with access to large, diverse population-based studies including three National Health and Nutrition Examination Surveys, the Multiethnic Cohort, the WHI, the ARIC study, the Coronary Artery Risk Disease in Young Adults study, the Cardiovascular Health Study, the Hispanic Community Health Study/Study of Latinos, the Strong Heart Study, and the Strong Heart Family Study.
This PAGE Metabochip study included African American participants from the ARIC and WHI CT studies. Participants from the other PAGE studies were excluded from this effort due to the unavailability of ECG measures and/or genotype data. Genotypes of WHI CT participants were obtained in three phases: two sets of women were directly genotyped on the Metabochip platform by PAGE investigators during wave 1 (n = 797) and wave 2 (n = 1,128) and women (n = 3,531) with Metabochip variants imputed from previous genome-wide SNP data provided by the WHI SHARe [36]. Participants meeting the following criteria were excluded from the study: QT unavailable, atrial fibrillation/atrial flutter on ECG, left or right bundle branch block on ECG, QRS duration >120 milliseconds, intraventricular conduction delay on ECG, pacemaker implant antedating ECG, ancestry outlier, excessive heterozygosity, low call rate, or second member of first degree relative pair. Further details on the ARIC and WHI CT studies are provided in Text S1 (Participating Studies).
For each study, certified technicians digitally recorded resting, supine (or semi-recumbent), standard 12-lead ECGs at study baseline for each participant using Marquette MAC PC machines (GE Healthcare, Milwaukee, WI, USA). The ARIC and WHI CT studies used comparable procedures for preparing participants, placing electrodes, recording, transmitting, processing, and controlling quality of the ECGs. QT was measured electronically using the Marquette 12SL algorithm.
The Metabochip was a custom Illumina iSELECT array that contained approximately 195,000 SNPs and was designed to support large scale follow up of putative associations for cardiovascular and metabolic traits including QT, blood pressure, cholesterol, type 2 diabetes, and anthropometrics. Approximately 33% of the Metabochip SNPs were included as replication targets and 62% for fine-mapping. In total, 257 loci were selected for fine-mapping, with the surrounding regions totaling 45.5 Mb accounting for overlaps (14.2 Mb for the densest fine-mapping regions). Eleven QT loci identified in previous GWA studies in populations of European and Asian ancestry were represented on the Metabochip (Table 1). The only published QT locus that is not represented on the Metabochip is an intergenic region on 13q14 reported by Marroni et al [19], but not replicated by other published GWA studies of populations with similar ancestral backgrounds. SNPs reported in the literature but not genotyped on the Metabochip (NOS1AP, rs10494366; NDRG4, rs7188697, rs37062) were represented by proxies, defined as SNPs in high LD (r2≥0.90) with the index SNP using HapMap YRI data.
Samples were genotyped at the Human Genetics Center of the University of Texas-Houston and the Translational Genomics Research Institute for ARIC and WHI, respectively, following each genotyping center's standard procedures. HapMap YRI (Yoruba in Ibadan, Nigeria) samples were also genotyped independently by each study to facilitate cross-study quality control. Genotypes were called separately for each study, albeit with a common protocol and common personnel, with GenomeStudio using the GenCall 2.0 algorithm. Because the Metabochip includes SNPs with much lower MAFs than are usually called with GenCall, SNPs were recalled using the GenoSNP genotype-calling algorithm [37]. SNPs with call rates <95%, Hardy-Weinberg equilibrium P<10−6, >1 Mendelian error (in 30 YRI trios), >2 replication errors, or >3.3% discordant calls in YRI across genotyping centers or against the HapMap database were considered quality control failures. Samples with call rates <0.95 or an inbreeding coefficient F>0.15 were excluded from further analysis [38].
Prior to analyses, related participants were identified using PLINK [39] by estimating identical-by-descent statistics for all pairs. When apparent first-degree relative pairs were identified, the member from each pair with the lower call rate was excluded from further analysis. Principal components of ancestry were determined using the Eigensoft software [40], [41] and apparent ancestral outliers were excluded from further analysis.
Briefly, n = 1,962 WHI participants who were genotyped on both the Affymetrix 6.0 and Metabochip genotyping platforms were used to infer Metabochip genotypes to the n = 8,421 population of WHI participants genotyped on the Affymetrix 6.0 array [36]. Before phasing and imputation, Affymetrix 6.0 SNPs with genotype call rates <90%, Hardy-Weinberg P-values<10−6, or MAF<0.01 were removed. Participants with call rates <95%, those who demonstrated excess heterozygosity, were part of a first-degree relative pair, or who were identified as an ancestry outlier were excluded. This yielded a set of 987,749 SNPs for the 1,962 reference participants. Mean concordance rates for the 23,703 SNPs in common was 99.7%. Haplotypes were reconstructed using MaCH and were used as a reference to impute Metabochip data into the 6,459 WHI participants with only Affymetrix 6.0 data. Liu et al., (2012) demonstrated the ability to impute 99.9% (97.5%, 83.6%, 52.0%, 20.5%) of SNPs with MAF≥0.05 (0.03–0.05, 0.01–0.03, 0.005–0.01, and 0.001–0.005) with average dosage r2 = 94.7% (92.1%, 89.0%, 83.1%, and 79.7%), respectively. For this analysis, all imputed SNPs with r2<0.95 were excluded.
To interpret fine-mapping results, LD in our African American PAGE Metabochip sample was calculated in 500 Kb sliding windows using PLINK. In addition, Metabochip LD and frequency information (but not individual-level information) was provided by the Malmö Diet and Cancer Study on 2,143 control participants from a Swedish population [42] to facilitate LD and MAF comparisons to populations of European ancestry. HapMap CEU LD data were used for previously published GWA studies in European populations, as not all European index variants were represented on the Metabochip. Regional association plots use positions from NCBI build 36. Recombination rates were estimated from HapMap phase II data.
Linear regression models were used to study the association between QT and 6,670 SNPs from 11 regions fine-mapped for QT assuming an additive genetic model and including age, sex, study center, ancestry principal components, and heart rate as covariates. Study-specific association results were combined using an inverse variance meta-analysis approach as implemented in METAL [43].
For each QT locus, it is expected that SNPs associated with QT in African Americans will be correlated with the index SNP reported in Europeans. Therefore, we first identified and tested SNPs that are correlated (r2≥0.20) with the index signals in Europeans using LD statistics estimated in the Malmö Diet and Cancer Study. In order to determine the appropriate multiple testing threshold for declaration of whether the previously identified signals were significantly associated with QT in African Americans, i.e. generalizability, we then estimated the number of tag SNPs needed to capture all common alleles (r2≥0.80) using African American LD patterns. The multiple testing threshold for declaring generalization was αa = 0.05/415, where 415 = the total number of tags identified using African American LD patterns.
To identify significant population-specific SNPs influencing QT that were not correlated with the index signal in Europeans (i.e. r2<0.20, which was estimated in the Malmö Diet and Cancer Study), we used an efficient Monte Carlo approach that accounts for LD between SNPs at the previously identified QT loci (αb = 1.37×10−5) [44]. Conditional analyses were then performed to determine the number of independent signals the population-specific SNPs represent. Specifically, analyses were repeated for each locus including the SNP with the smallest P – value as a covariate. This approach was performed adjusting for successively less significant SNPs until no SNPs with P –values lower than αb = 1.37×10−5 were identified. To facilitate comparability with previous reports examining the proportion of variance in QT explained by common SNPs, heart rate-corrected QT [45] was regressed on the six best markers in African Americans and the three population-specific variants assuming an additive genetic model and including age, sex, study center, ancestry principal components as covariates.
For each of the nine QT SNPs (i.e. the six best markers in African Americans and the three novel SNPs), we identified all SNPs in LD (r2≥0.5) using the genotypes from the African American population described in this study. We refer to these SNP sets as Trait Associated SNP (TAS) blocks. We assigned each TAS to one or more of the functional annotation datasets listed in Table S5. These datasets are not mutually exclusive. For example, a TAS can reside in both a candidate regulatory element (dataset #7) and a CTCF binding site (dataset #10). For TASs that occur within predicted transcription factor binding sites (datasets #3 and #8), we calculated transcription factor binding affinity for each TAS allele using PWM-scan [46], as described previously [47]. For TASs that occur within 3′ untranslated regions, we used the TargetScanS algorithm to determine whether they disrupt likely microRNA target sites (dataset #5). To define candidate non-promoter regulatory elements of greatest relevance to QT (dataset #7), we restricted the analysis of DNase I hypersensitive sites (open chromatin loci) to only those present in human cardiomyocytes.
|
10.1371/journal.pmed.1002635 | Benzodiazepine prescribing for children, adolescents, and young adults from 2006 through 2013: A total population register-linkage study | Pharmacoepidemiological studies have long raised concerns on widespread use of benzodiazepines and benzodiazepine-related drugs (BZDs), in particular long-term use, among adults and the elderly. In contrast, evidence pertaining to the rates of BZD use at younger ages is still scarce, and the factors that influence BZD utilisation and shape the different prescribing patterns in youths remain largely unexplored. We examined the prevalence rates, relative changes in rates over time, and prescribing patterns for BZD dispensation in young people aged 0–24 years in Sweden during the period January 1, 2006–December 31, 2013, and explored demographic, clinical, pharmacological, and prescriber-related attributes of BZD prescribing in this group.
Through the linkage of 3 nationwide Swedish health and administrative registers, we collected data on 17,500 children (0–11 years), 15,039 adolescents (12–17 years), and 85,200 young adults (18–24 years) with at least 1 dispensed prescription for a BZD during 2006–2013, out of 3,726,818 Swedish inhabitants aged 0–24 years. Age-specific annual prevalence rates of BZD dispensations were adjusted for population growth, and relative changes in rates were calculated between 2006 and 2013. We analysed how BZD dispensation varied by sex, psychiatric morbidity and epilepsy, concurrent dispensation of psychotropic medication, type of dispensed BZD, and type of healthcare provider prescribing the BZD. Prescribing patterns were established in relation to duration (3 months, >3 to ≤6 months, or >6 months), dosage (<0.5 defined daily dosage [DDD]/day, ≥0.5 to <1.5 DDD/day, or ≥1.5 DDD/day), and “user category” (“regular users” [≥0.5 to <1.5 DDD/day for ≥1 year], “heavy users” [≥1.5 DDD/day for ≥1 year], or otherwise “occasional users”). Multinomial regression models were fitted to test associations between BZD prescribing patterns and individual characteristics of study participants. Between 2006 and 2013, the prevalence rate of BZD dispensation among individuals aged 0–24 years increased by 22% from 0.81 per 100 inhabitants to 0.99 per 100 inhabitants. This increase was mainly driven by a rise in the rate among young adults (+20%), with more modest increases in children (+3%) and adolescents (+7%). Within each age category, overall dispensation of BZD anxiolytics and clonazepam decreased over time, while dispensation of BZD hypnotics/sedatives, including Z-drugs, showed an increase between 2006 and 2013. Out of 117,739 study participants with dispensed BZD prescriptions, 65% initiated BZD prescriptions outside of psychiatric services (92% of children, 60% of adolescents, 60% of young adults), and 76% were dispensed other psychotropic drugs concurrently with a BZD (46% of children, 80% of adolescents, 81% of young adults). Nearly 30% of the participants were prescribed a BZD for longer than 6 months (18% of children, 31% of adolescents, 31% of young adults). A high dose prescription (≥1.5 DDD/day) and heavy use were detected in 2.6% and 1.7% of the participants, respectively. After controlling for potential confounding by demographic and clinical characteristics, the characteristics age above 11 years at the first BZD dispensation, lifetime psychiatric diagnosis or epilepsy, and concurrent dispensation of other psychotropic drugs were found to be associated with higher odds of being prescribed a BZD for longer than 6 months, high dose prescription, and heavy use. Male sex was associated with a higher likelihood of high dose prescription and heavy use, but not with being prescribed a BZD on a long-term basis (> 6 months). The study limitations included lack of information on actual consumption of the dispensed BZDs and unavailability of data on the indications for BZD prescriptions.
The overall increase in prevalence rates of BZD dispensations during the study period and the unexpectedly high proportion of individuals who were prescribed a BZD on a long-term basis at a young age indicate a lack of congruence with international and national guidelines. These findings highlight the need for close monitoring of prescribing practices, particularly in non-psychiatric settings, in order to build an evidence base for safe and efficient BZD treatment in young persons.
| Benzodiazepines and benzodiazepine-related drugs (BZDs) are among the most widely used psychotropic medications in the world and have long raised public health concerns due to the risk for their users of developing dependence and severe adverse effects, in particular among long-term users.
The prevalence rates and attributes of BZD use have been well described in adults and the elderly.
In contrast, evidence regarding BZD prescribing for younger populations is scarce and fragmented, and factors influencing BZD utilisation in children, adolescents, and young adults remain largely unexplored.
We linked nationwide Swedish health and administrative registers on 3,726,818 individuals aged 0–24 years living in Sweden between 2006–2013, among whom 17,500 children (0–11 years), 15,039 adolescents (12–17 years), and 85,200 young adults (18–24 years) were dispensed a BZD prescription at least once during that period.
The prevalence rate of BZD dispensation has increased by 22% between 2006 and 2013, and has increased particularly in young adults.
For nearly 65% of all study participants with 1 or more dispensed BZD prescriptions, the first prescription was issued outside specialised psychiatric services. Most first prescriptions originated from primary care (41%) and non-psychiatric specialist settings (24%) such as paediatrics and internal medicine.
Over 75% of study participants were dispensed other psychotropic medication concurrently with a BZD, raising concerns about the potential risk of drug interactions.
Almost 30% were prescribed a BZD for longer than 6 months, contradicting international and national guidelines that advise against BZD use beyond 2–4 weeks for adults and generally discourage prescribing BZDs for ages below 18 years.
There is a need for clinicians, particularly those working in non-psychiatric services, to implement strategies to avoid potentially harmful patterns of prescribing BZDs to young people.
Young individuals who are prescribed BZDs should be closely monitored, in particular, with respect to duration of BZD treatment and concurrent prescription of other psychotropic drugs.
| Widespread use of benzodiazepines and benzodiazepine-related drugs (BZDs) has long raised public health concerns, owing to the risks of developing tolerance, dependence, withdrawal syndromes, and severe adverse effects, particularly among long-term users [1–3]. A recently published perspective piece makes parallels between the contemporary epidemic of overprescribing of opioids and that of BZDs, with the former phenomenon being well acknowledged by clinicians and policymakers, while for the latter one a lesser effort is being made to address today’s prescribing practices [4]. Current knowledge on BZD prescribing, incidence and prevalence rates, and patterns of use mainly rely on data on adults and the elderly, for whom BZDs are often prescribed for managing anxiety symptoms and insomnia [1,3]. International and national guidelines recommend using BZDs for no longer than 2–4 weeks since the risk–benefit ratio beyond that period is debatable [5–8]. Evidence pertaining to the corresponding issues in individuals at younger ages is limited, not least due to BZDs not being recommended as a pharmacological option for treatment of any psychiatric disorders in persons below 18 years of age [9–11]. Indeed, the only firmly established indications of BZDs in this age group are the control of different types of seizures and the treatment of status epilepticus [12–15]. Studies on BZD prescribing in children and adolescents consistently report low annual prevalence rates (0.3%–0.5% in North America [16] and 0.2%–0.9% in Europe [17–20]), while the results of time-trend analyses appear to vary between countries. Over the last 2 decades, BZD prescription rates have increased in children and adolescents in various Western countries [16,21,22], while remaining stable or decreasing in others [17,18,21]. Depending on how each study defines the age boundaries for the category of young adults, BZD prescription rates are reported to range between 1% and 5%, and to mainly increase over time [21,23,24]. Noteworthy is that recent European data on new BZD users show a low and decreasing incidence in the age group below 18 years [25]. This, in light of stable or increasing prevalence rates, points towards a risk for chronic BZD use in this population.
Despite being scarce, studies on paediatric BZD use raise a series of concerns, in particular related to inappropriate prescriptions, concurrent use of psychotropic drugs, changes in the characteristics of the prescribers towards a higher involvement of non-psychiatrists and primary care physicians, and long-term use [16–18,21,26–29]. Few studies have investigated the patterns of BZD use among people below 18 years of age; existing studies used definitions of long-term use ranging from over 1 month [30,31] to over 3 months of cumulative BZD prescribing [17], and the proportion of long-term BZD users varied widely, from 3.3%–5.9% [17,31] to 14.7% [30]. A higher likelihood of long-term use is reported for children and adolescents with psychiatric disorders, epilepsy, and BZD prescriptions initiated by psychiatrists, with less clear results on the role of sex [31]. Given the frequent “off-label” BZD prescribing (i.e., outside of approved indication or age category [32]), and that the risk–benefit ratio of BZD use has not yet been fully clarified for children and adolescents [33–35], it is important to establish predictors and attributes of BZD prescribing patterns in these age groups to serve as an evidence base for guiding clinicians in their prescribing practices.
To fill this important gap in the knowledge base, we aimed to explore the annual prevalence rates of BZD dispensations and relative changes in rates over time among individuals aged below 25 years during 2006–2013. Subsequently, we sought evidence on how BZD dispensations in children, adolescents, and young adults vary by individual characteristics (sex, psychiatric morbidity and epilepsy, and concurrent dispensation of psychotropic medication), type of BZD drug, and type of healthcare provider prescribing the BZD. Furthermore, we aimed to describe BZD prescribing patterns with regard to duration of prescription and prescribed dosage and to explore associations between different patterns and characteristics of the patients.
The study was set in a population-based cohort constructed through the record linkages of 3 Swedish registers with complete national coverage. The Swedish Prescribed Drug Register (PDR) encompasses data on prescribed medications dispensed across all pharmacies in Sweden from July 2005 onwards, registered using Anatomical Therapeutic Chemical Classification System (ATC) codes, along with dosage, dispensed amount, dispensation date, and type of healthcare provider prescribing the BZD [36]. The PDR does not include treatment indication and medications administered in hospitals. The National Patient Register (NPR) comprises data on clinical diagnoses, coded using the International Classification of Diseases (ICD), from inpatient care (1964 onwards) and specialist outpatient services (2001 onwards), with complete (national) coverage since 1987 and 2001, respectively [37]. The register was validated for an array of diagnoses, with a positive predictive value of 85%–95% overall and up to 97% for psychiatric disorders [37–41]. The Total Population Register (TPR) contains demographic characteristics of Swedish inhabitants from 1968 onwards [42]. The PDR and the NPR are held by the Swedish National Board of Health and Welfare, and the TPR is maintained by Statistics Sweden. The linkage was performed through the unique personal identification number assigned to all Swedish citizens and residents [43]. The study was approved by the Regional Ethical Review Board (2013/5:8) in Stockholm, Sweden. The requirement for informed consent was waived because the study was register-based and the included individuals were not identifiable at any time.
The study participants comprised all individuals aged 0–24 years with at least 1 dispensed BZD prescription between January 1, 2006, and December 31, 2013, according to the PDR. Sex and birth year were collected from the TPR and linked to the data from the PDR. The birth year was used to calculate age at the first and each consecutive BZD dispensation. The participants were categorised into children (0–11 years), adolescents (12–17 years), and young adults (18–24 years). For the analyses over the whole period 2006–2013, categorisation was based on age at first BZD dispensation. For assessing BZD dispensations within each year (e.g., annual prevalence), the actual age of participants in the specific year was used.
The study was performed in correspondence with a prespecified analysis plan (see S1 Text). Annual prevalence of BZD dispensations was calculated separately for children, adolescents, and young adults as the proportion of individuals who were dispensed a BZD at least once during the year, for the years 2006 to 2013, out of the total number of inhabitants of the same age category in Sweden in the corresponding year (as reported by Statistics Sweden [47]). Individuals with multiple BZD dispensations within the same year were counted only once. The rates are reported per 100 inhabitants to be interpreted as age-specific annual prevalence of BZD dispensation in the general population of Sweden in 2006–2013.
Subsequently, the attributes of BZD dispensations (i.e., sex, concurrent dispensation of psychotropic drugs, psychiatric diagnoses and epilepsy, type of BZD drug, and healthcare provider category) were analysed among all individuals (0–24 years) with dispensed BZD prescriptions and stratified by age categories. The results are reported in percentages to be interpreted as the proportion with a certain attribute per 100 study participants. The analyses were performed within each year (the denominator included individuals with at least 1 dispensed BZD prescription in a given year) and across the study period (the denominator included all individuals with at least 1 dispensed BZD prescription between 2006 and 2013).
To study changes over time in annual prevalence rates of BZD dispensations in Swedish youths or in proportions of various attributes of BZD dispensations among the study participants, the relative change in measures in 2013 from those in 2006 (the referent value) was calculated: The value estimated in 2006 was subtracted from that in 2013, and the result was divided by the value in 2006. The relative change is reported as a percentage, with positive quantities corresponding to increases in values over time, and negative ones corresponding to decreases. Multinomial logistic regression models were fitted to obtain odds ratios and corresponding 95% confidence intervals for associations of the prescribing patterns with age, sex, lifetime psychiatric diagnoses and epilepsy, and concurrent dispensation of psychotropic drugs. In the modelling strategy, variables were included in the multivariate analysis if they were found significant in the univariate models or if they fulfilled the criteria for being confounders [48].
A series of sensitivity analyses were conducted by repeating the main analyses in a sub-population restricted to individuals without a lifetime diagnosis of epilepsy to capture the BZD prescribing for indications other than seizures. In reporting, we followed the STROBE guidelines for cohort studies (see S1 Checklist). All statistical analyses were performed using SAS, version 9.4 [49].
Out of 3,726,818 individuals aged 0–24 years living in Sweden during 2006–2013 (as reported by Statistics Sweden [47]), 117,739 (3.16%) collected at least 1 BZD prescription before their 25th birthday, with prevalence rates ranging from 0.81 per 100 inhabitants in 2006 to 0.99 per 100 inhabitants in 2013, corresponding to a relative increase of 22.2%. Among the study participants, at the time of the first BZD dispensation, 14.8% were children (aged 0–11 years, n = 17,500), 12.8% were adolescents (aged 12–17 years, n = 15,039), and 72.4% were young adults (aged 18–24 years, n = 85,200). The total number of individuals collecting BZDs increased steadily from 2006 to 2013, with sizeable difference between age groups (Fig 1). Throughout the study, the annual prevalence of BZD dispensations in young adults was 5- to 8-fold higher than that in children and adolescents, with the relative change in rate indicating a 20.1% increase from 2006 (prevalence rate of 1.99 per 100 inhabitants) to 2013 (2.39 per 100 inhabitants) for the young adult group. Prevalence rates remained considerably lower and more stable in children (in 2006: 0.28 per 100 inhabitants; in 2013: 0.29 per 100 inhabitants) and adolescents (in 2006: 0.43 per 100 inhabitants; in 2013: 0.46 per 100 inhabitants), with a relative increase of 3.5% and 6.9%, respectively.
Table 1 presents the distribution of demographic and clinical characteristics of the study sample—individuals aged 0–24 years with at least 1 dispensed BZD prescription between 2006 and 2013—for the total sample and within age categories. Among the individuals with dispensed BZD prescriptions, the proportion of females was higher than that of males overall and among adolescents and young adults, but lower in children. Three out of 4 individuals from the total sample collected at least 1 additional class of psychotropic drugs within 6 months of the dispensation of a BZD, and 24.7% of the study participants received 3 or more classes of concurrent medications. The proportions varied across ages, with less than 50% of children and over 80% of adolescents and young adults collecting other psychotropic drugs concomitantly with a BZD. Antidepressants, non-BZD anxiolytics, and non-BZD hypnotics/sedatives prevailed among concurrent drugs in adolescents and young adults, while in children mood stabilisers were collected most frequently.
Among individuals who collected a BZD up to June 30, 2013 (n = 111,182; 94.4% of the study sample), 45.3% had a record of a psychiatric disorder diagnosed within 6 months of BZD dispensation, with anxiety and depression being the most frequent diagnoses (20.3% and 20.1%, respectively). In total, over 11% of study participants were diagnosed with epilepsy within 6-month proximity to BZD dispensation (Table 2). The results for adolescents and young adults corroborated the ones seen in the total sample, while in children only 16.6% had a psychiatric diagnosis, with disruptive behavioural disorders being the most frequent diagnoses (11.8%). The proportion of individuals diagnosed with epilepsy close to BZD dispensation was highest in children (39.7%), followed by that in adolescents (22.7%) and young adults (3.1%). In addition, we assessed lifetime diagnosis of psychiatric disorders and epilepsy using data for all 117,739 study participants. As anticipated, the proportions of participants with lifetime diagnosis were higher than those with diagnosis within 6 months of BZD dispensation, although the most common diagnoses remained unchanged in all age groups.
As presented in Table 3, throughout the study, BZD anxiolytics were collected by 57.8% of all study participants dispensed a BZD, with marginal change between 2006 and 2013 (relative decrease of 3.1%). Diazepam and oxazepam were the top BZD anxiolytics dispensed (32.5% and 25.5%, respectively), showing reverse relative changes over time (relative decrease of 17.8% and relative increase of 41.3%, respectively). A similar proportion of study participants collected BZD hypnotics/sedatives, including Z-drugs (60.9%; relative increase of 9.8%), with zopiclone and zolpidem being dispensed most frequently (39.3% and 27.5%, respectively; relative increase of 36.4% and relative decrease of 24.5%, respectively). Over the study period, less than 3% of study participants were dispensed clonazepam (relative decrease of 18.2%).
The abovementioned results differed between age groups (Table 3). Children were mainly dispensed BZD anxiolytics (97.4%), but uncommonly dispensed BZD hypnotics/sedatives, including Z-drugs (8.4%), and clonazepam (6.7%). Over the whole study period, diazepam was the dominant BZD drug dispensed to children (95.5%). Adolescents prevailed over young adults in collecting BZD anxiolytics (59.1% versus 49.4%) and clonazepam (3.8% versus 1.8%), while BZD hypnotics/sedatives and Z-drugs were collected at higher rates among young adults compared to among adolescents (71.7% versus 60.5%). Notwithstanding these differences, within each age category overall dispensation of BZD anxiolytics and clonazepam decreased over time, while dispensation of BZD hypnotics/sedatives, including Z-drugs, showed an increase between 2006 and 2013.
In the total sample of study participants dispensed a BZD, the first BZD prescription was most commonly issued within non-psychiatric healthcare services (i.e., in primary care, 40.7%, or specialised care other than psychiatry, 24.1%) (Table 4). The healthcare provider category in which BZD prescriptions were initiated differed between age categories, although services outside of psychiatry prevailed across all ages. Only 8.1% of the first BZD prescriptions for children and 40% of such prescriptions for adolescents and young adults were issued at a psychiatric service. Regardless of the age category, the share of BZD prescriptions initiated in psychiatric healthcare services appeared to decrease between 2006 and 2013 (Fig 2).
Throughout the study, 15.4% of the study participants dispensed a BZD were prescribed a BZD for a period of >3 to ≤6 months, while 29.3% of the participants were prescribed a BZD for longer than 6 months (Table 5). The results of univariate multinomial logistic regression indicated that age above 11 years at the first BZD dispensation, a lifetime diagnosis of any psychiatric disorder, and concurrent dispensation of psychotropic drugs were associated with a higher likelihood of being prescribed a BZD for >3 to ≤6 months and for >6 months, while lifetime history of epilepsy was associated with lower odds of having a duration of prescription of >3 to ≤6 months, but with higher odds of being prescribed a BZD for longer than 6 months. When the potential confounding effects of demographic and clinical covariates was accounted for, the observed associations mostly decreased in strength, but all remained significant. Furthermore, in the multivariate model, male sex was associated with a higher likelihood of being prescribed a BZD for >3 to ≤6 months, but no association appeared for a duration of prescription of longer than 6 months.
In addition, over 13% of the participants were prescribed a BZD with an average daily dosage of ≥0.5 to <1.5 DDD/day, while an average daily dosage of ≥1.5 DDD/day was observed in 2.6% of individuals (see S3 Table). Regardless of the adjustment strategy, both categories of prescribed dosage were associated with male sex, age at first BZD dispensation above 11 years, a lifetime diagnosis of any psychiatric disorder, and concurrent dispensation of psychotropic medication. In the analysis that controlled for the effect of other covariates, lifetime history of epilepsy had a significant association with a higher likelihood of being prescribed a BZD with an average daily dosage of ≥1.5 DDD/day.
Furthermore, 6.1% of study participants fulfilled the criteria for being regular users, while 1.7% were considered heavy users (see S4 Table). In the multivariate model, male sex was associated with higher odds of being a heavy user, but not a regular user. The odds of being a regular or heavy user increased with age at first dispensation, in the presence of lifetime psychiatric disorders or epilepsy, and with concurrent dispensation of other psychotropic drugs.
Because treatment of seizures is the most common indication for BZDs in early life [9], we restricted the analyses to study participants without lifetime diagnosis of epilepsy (n = 102,548). The results were largely unchanged (see S5–S8 Tables). The only exceptions were the analyses of BZD prescribing patterns, where a sizeable reduction in the proportion of individuals in the highest category of each pattern was observed (see S9–S11 Tables). The differences, however, appeared only among children, while the proportions of adolescents and young adults with a BZD being prescribed on a long-term basis (> 6 months), high average daily dosage (≥1.5 DDD/day), and heavy use remained similar to the ones detected by the main analyses.
This total population register-based study is among the first to systematically evaluate BZD dispensation and its attributes among children, adolescents, and young adults at a nationwide level. There were 5 principal findings. First, the study indicated a 22% increase in the prevalence rate of BZD dispensation between 2006 (0.81 per 100 inhabitants) and 2013 (0.99 per 100 inhabitants) in the population aged 0–24 years. This increase was mainly driven by a steady rise in the rate among young adults, with more modest increases in children and adolescents. Second, in all age categories of those with dispensed BZD prescriptions, a high proportion of polypharmacy was observed, with almost half of children and over 80% of adolescents and young adults having been dispensed other psychotropic drugs concomitantly with a BZD. Third, off-label BZD prescription was common. This was particularly notable in adolescents, among whom a substantial proportion were dispensed zopiclone and zolpidem—drugs that are not approved for ages below 18 years, according to international and Swedish pharmaceutical guidelines [5,50]—with a marked increase in dispensations of both drugs between 2006 and 2013 in this age group. Fourth, although the type of healthcare provider that initiated BZD prescriptions varied between the age categories, the prescribers were mainly outside of specialised psychiatric services; approximately 65% of all prescriptions originated either in primary care or non-psychiatric specialist services. Fifth, the most alarming results came from the analyses of prescribing patterns, with an unexpectedly high proportion of individuals across all ages being prescribed a BZD on a long-term basis—nearly every fifth child and every third adolescent and young adult among those who received a BZD in 2006–2013 were prescribed such medication for longer than 6 months. An elevated likelihood of long-term prescribing was associated with age above 11 years at first BZD dispensation, lifetime diagnosis of any psychiatric disorder or epilepsy, and concomitant dispensation of other psychotropic medication. Sensitivity analyses provided additional evidence of long-term prescribing being a common phenomenon, as the proportion of adolescents and young adults who were prescribed a BZD for longer than 6 months remained large even when individuals with lifetime diagnosis of epilepsy were excluded. Among children, although the proportion of those prescribed a BZD on a long-term basis dropped by 80% after excluding those with lifetime diagnosis of epilepsy, BZD prescribing for longer than 6 months remained present in nearly 4% of study participants in this age group.
A general data scarcity on BZD prescribing in younger populations and differences in methodological approaches between studies makes it difficult to find a direct comparison for the present findings, which integrate a wide spectrum of attributes of BZD dispensation in young people. Nonetheless, the rates of BZD dispensations in children, adolescents, and young adults from our study corroborated the results of cross-national and national projects from Europe and Canada [16,17,21], although they were higher than the rates reported for Norway and Iceland [18,28]. Likewise, our findings of increasing rates over time were supported by cross-European [21] and Canadian data [16], but differed from the decreasing rates seen among children and adolescents in Norway, Ireland, and Denmark [17,18,26]. While interpreting the results on BZD dispensations in young populations, it is important to keep in mind that there is no firmly established indication for BZD treatment in child and adolescent psychiatry [9]—that is also true for Sweden [8,51,52]. In Swedish clinical guidelines for adults (i.e., age 18 years and above), the use of, for example, oxazepam is limited to alcohol withdrawal and delirium tremens and as a second-line drug for temporary management of anxiety symptoms, and zopiclone is a first-line drug for insomnia [8,51,53]. Additionally, guidelines recommend diazepam and midazolam to be used for all ages for treatment of acute epilepsy, while for status epilepticus, midazolam is listed for patients aged below 18 years and diazepam for older individuals [51]. In relation to the abovementioned, our results on concomitant dispensation of psychotropic drugs add to the concerns previously raised about a high prevalence of psychotropic polypharmacy in children and adolescents [17,20,25,27,28,54]. The concerns stem, particularly, from the evidence of enhanced adverse effects of BZDs, including central nervous system depression and respiratory depression, in the presence of other central nervous system depressants (e.g., antipsychotics, antidepressants, and non-BZD anxiolytics) or respiratory depressants (e.g., opioid analgesics) [20]. BZD concurrency with prescription opioids (which in our study appeared in 13% of adolescents and 18% of young adults) gives another cause for concern due to the increased risk of overdose and death among polydrug users [55,56]. Given the restrictions posed by the guidelines on clinical utilisation of BZDs, our findings of dispensed prescriptions for zopiclone and zolpidem in individuals aged below 18 years are undoubtedly worrisome, but not uncommon, and corroborate the results from studies in other countries with smaller sample sizes but comparable age categories [17,34]. In addition, our finding of a large proportion of BZD prescriptions being initiated outside psychiatric services corresponds to similar practices reported for patients aged 0–20 years in North American studies and aged 0–17 years in Icelandic studies, although those studies cover periods only up to 2010 [24,28,57]. The authors of those studies put forward the finding of the increase in visits to non-psychiatrists that appears to coincide with an increase in prescribing of BZDs to young individuals [24,28,57]. The issue was further addressed in a systematic review that demonstrated a substantial diversity in knowledge, attitudes, and awareness about balancing the risks and benefits of BZD treatment among primary care specialists [58]. It is worth mentioning that, over the study period, the number of adolescents and young adults seeking psychiatric care in Sweden increased significantly, with anxiety and depression being the top diagnoses [59,60]. The same age groups also demonstrated a recent increase in receiving psychiatric diagnoses in primary care services and specialised care other than psychiatry [59]. Additionally, there has been a shift in the Swedish healthcare services towards expecting primary care to act as a “first line of psychiatry” [61].
Our finding of an unexpectedly high proportion of individuals who were prescribed BZDs for longer than 6 months deserves special attention. Although there is no consensus between pharmacoepidemiological studies on the definition of long-term use [2], international clinical guidelines strongly advise limiting BZD treatment to the shortest possible period of 2–4 weeks due to side-effects and dependence and withdrawal concerns [1,3,6,7]. A recent systematic review on long-term BZD use reported an average prevalence of 24% (95% CI 13% to 36%) among BZD users of all ages, though the evidence mostly came from studies on adult and elderly users, with the definition of long-term use varying from 1 month to over 12 months [2]. As discussed above, the Swedish guidelines do not recommend BZDs for treatment of psychiatric disorders among individuals aged below 18 years, and the maintenance therapy to prevent seizures in patients with genuine epilepsy or other convulsive disorders cannot explain the observed prescribing patterns. Given that in our study nearly 30% of the participants were prescribed a BZD for longer than 6 months (even after excluding patients with lifetime epilepsy diagnosis, who represented 11% of the study population), further examination is required of potential predictors and mediators of this long-term BZD prescribing pattern. As we assessed only the likelihood of such a pattern, it was not possible to infer causality, and, therefore, no risk or protective factors could be clearly established in this study. However, the results pointed out several attributes of long-term BZD prescribing, including age of onset older than 11 years, psychiatric comorbidities and epilepsy, and concomitant dispensation of other psychotropic medications. These findings are consistent with previous research [2,31,62] and highlight the need for thorough monitoring of BZD prescribing practices for drug–drug interactions, adverse effects, and guideline adherence.
The study has several strengths, including the use of the official Swedish registers with complete national coverage, extensive collection of clinical and demographic data, and timely and routine updating. The robust and comprehensive nature of the present data minimises the potential influence of sampling and reporting error and recall bias. The PDR covers all dispensed drugs regardless of reimbursement status and service provider characteristics, which makes our data representative of the prescribing practices across all healthcare levels in Sweden. In addition, the use of population-based registers ensures generalisability of the results at the national level. Our study has, however, some limitations. First, the PDR does not contain data on the indications for prescriptions, which may have affected our ability to fully analyse a relation between comorbid diagnoses and BZD dispensations among the study participants. In addition, we were unable to retrieve data on disorders diagnosed at primary care services, which could possibly have resulted in an underestimation of the proportion of individuals with comorbidity if they were diagnosed outside of hospitals and specialised care services. Second, we lacked data on the diagnosis of febrile convulsions (ICD-10 code: R56.0) in the NPR. Febrile convulsions are rather common below 6 years of age, and such data, if available, might have given an additional insight into the associations between clinical diagnoses and BZD dispensations, particularly among young children. Third, the data collection was limited to the diagnoses coded with the ICD 10th revision (i.e., starting from 1997). This may have led to missing diagnoses of psychiatric disorders or epilepsy in people born between 1982 and 1997 if these diagnoses were recorded in the NPR before 1997 and were not mentioned again later. Fourth, although the initial analyses included 3 types of BZD prescribing patterns, only the duration of prescription was discussed as a principal finding. It was difficult to draw conclusions about the other 2 patterns as many BZD substances have several indications and, hence, different dosages may be recommended for different conditions. In the absence of information on the exact indication, we wanted to avoid misinterpreting the patterns of prescribed dosage and user category, and, therefore, the results reported on these patterns should be interpreted as suggestive. Fifth, our analyses rest on dispensation data, and, hence, we cannot be certain that the medication was used in proximity to the date of dispensation and by the person it had been prescribed to. However, it is unlikely that the proportion of individuals who collected but did not use BZDs varied across the study period [25]. Lastly, the PDR only covers the period starting July 2005, making it impossible to analyse the rates and patterns of BZD prescriptions dispensed prior to initiation of the register.
Notwithstanding the limitations, to our knowledge, this is the first study to explore BZD prescribing patterns from different perspectives, namely duration and drug dosage, and to identify individual characteristics that are associated with different prescribing patterns in young people. Our study delineates the directions for further research since less consistency still exists in relation to other potential attributes of BZD prescribing, in particular long-term prescribing, including sex, pharmacological features of BZD substances, concurrent prescription of distinct BZD classes, and combination of BZD treatment with certain modalities of psychotherapy. Further work is needed to get insight on the underlying mechanisms and risk and protective factors that shape the patterns of BZD prescription and use among youths. Our findings have clear implications across a range of clinical settings as well as for public policy. The results highlighted a need to devise and implement strategies to avoid potentially harmful patterns of prescribing BZDs to young people. This can only be achieved by involving primary care—the source of the majority of BZD prescriptions—but also other categories of prescribers, including but not limited to specialised care in paediatrics, internal medicine, and psychiatry. Prescribing practices need to be addressed using education, feedback, and peer support. For patients with problematic use, the UK National Institute for Health and Care Excellence guidelines recommend using stepped care models, where patients are given the least intrusive recommended treatment [63]. Furthermore, given the increased availability of addictive drugs via the internet, reductions in harmful prescribing need to be coordinated with availability policies involving, for example, customs and policing.
The annual prevalence rate of BZD dispensation in the Swedish population aged 0–24 years increased substantially between 2006 and 2013. This increase was mainly driven by a steady rise in the rate of BZD dispensation to young adults, accompanied by an increasing proportion of off-label prescriptions. Polypharmacy was the norm rather than the exception, pointing to the potential risk of drug interactions. Contradicting clinical guidelines, almost 30% of the study participants were prescribed BZDs on a long-term (>6 months) basis. During the study period, growing shares of prescriptions to individuals in all age categories were initiated in primary care or non-psychiatric specialty settings. It is critical to improve prescribing practices through close monitoring of BZD utilisation in young people, strengthening the adherence to BZD prescribing guidelines among primary care practitioners and other non-psychiatric specialists, and improving communication between non-specialised care and specialised psychiatric services.
|
10.1371/journal.pgen.1002510 | Rewiring of PDZ Domain-Ligand Interaction Network Contributed to Eukaryotic Evolution | PDZ domain-mediated interactions have greatly expanded during metazoan evolution, becoming important for controlling signal flow via the assembly of multiple signaling components. The evolutionary history of PDZ domain-mediated interactions has never been explored at the molecular level. It is of great interest to understand how PDZ domain-ligand interactions emerged and how they become rewired during evolution. Here, we constructed the first human PDZ domain-ligand interaction network (PDZNet) together with binding motif sequences and interaction strengths of ligands. PDZNet includes 1,213 interactions between 97 human PDZ proteins and 591 ligands that connect most PDZ protein-mediated interactions (98%) in a large single network via shared ligands. We examined the rewiring of PDZ domain-ligand interactions throughout eukaryotic evolution by tracing changes in the C-terminal binding motif sequences of the PDZ ligands. We found that interaction rewiring by sequence mutation frequently occurred throughout evolution, largely contributing to the growth of PDZNet. The rewiring of PDZ domain-ligand interactions provided an effective means of functional innovations in nervous system development. Our findings provide empirical evidence for a network evolution model that highlights the rewiring of interactions as a mechanism for the development of new protein functions. PDZNet will be a valuable resource to further characterize the organization of the PDZ domain-mediated signaling proteome.
| Rewiring of interactions is a powerful tool for the evolution of organism complexity. Rewiring among preexisting proteins provides a simple mechanism for the development of new signaling circuits by redirecting information flows without a gain or loss of genes. Particularly, interactions mediated by short linear motifs can be easily changed by mutations during evolution, resulting in a rewiring of interactions. However, how interaction rewiring of linear motif interactions facilitates the emergence of new protein function during evolution is poorly understood. Here, we systematically investigated the rewiring of interactions mediated by PDZ domains, which are one of the most commonly found peptide recognition modules. We found that PDZ domain-ligand interactions are frequently rewired by C-terminal sequence mutations in PDZ ligands during evolution. Especially, rewiring of PDZ domain-ligand interactions was involved in neuronal function development, occurring concurrently with the emergence of vertebrates and suggesting that reorganization of signaling pathways by rewiring PDZ domain-ligand interactions significantly contributed to the evolution of nervous systems in vertebrates. Our findings highlight the rewiring of interactions as an effective means for functional innovation, providing new insight into eukaryotic evolution, which has not been fully explained by only the expansion of protein families.
| PDZ domains are linear motif-mediated protein-protein interaction modules. PDZ domain-ligand interactions have been greatly expanded in metazoans and are widely used to assemble signaling complexes, including those found in neuronal synapses [1]. Thus, an understanding of how PDZ domain-ligand interactions have evolved would help elucidate the design principle of the eukaryotic signaling network. Many studies have revealed the evolutionary history of PDZ domain families and their functional roles [2], [3]. However, it remains poorly understood how PDZ domain-mediated interactions are rewired during the evolution of the protein interaction network.
Systematic analysis of interaction rewiring will provide new insights into eukaryotic evolution, which is not fully explained via only the expansion of protein families. Recently, it was suggested that rewiring of interactions is an important mechanism for the evolution of biological systems. Network comparison studies showed that protein interactions frequently change after gene duplication [4], [5]. In particular, linear motifs were suggested to have great potential to rewire interactions because of their high rate of change [6], [7]. Indeed, phosphorylation sites in one species are often lost in other species [8], [9]. Moreover, human-specific phosphorylation sites are recently examined to identify novel phenotypes in humans because the interaction rewiring of kinase interactions may contribute to the emergence of novel biological functions [10].
Structural information of interacting cellular components (i.e., structural interactome) would provide a more complete picture of a cell and help elucidate the evolutionary principle of the protein interaction network [11]. Recently, structural information of protein complexes were mapped onto protein interaction networks [12], [13]. Indeed, such interface information of protein interactions would more clearly explain evolutionary principles, such as the network evolution model by gene duplication [12] and the role of residues surrounding linear motifs in terms of binding specificity [14]. Therefore, to understand the underlying design principle of the PDZ domain-ligand interaction network, detailed interface information at the amino acid level is needed.
In this work, we attempted the first systematic investigation of interaction rewiring in the PDZ domain-ligand interaction network and its role in eukaryotic evolution. We constructed a comprehensive human PDZ domain-ligand interaction network and traced the changes in interaction rewiring during evolution. We developed position weight matrices (PWMs) of human PDZ domains from the experimental data of PDZ domain-ligand interactions. The binding motif information of PDZNet helped to elucidate the changes in PDZ domain-ligand interactions. We found that PDZ domain-ligand interactions are frequently rewired throughout evolution via mutations of C-terminal PDZ ligand sequences. Particularly, interaction rewiring occurred concurrently with emergence of vertebrates whose rewired interactions were largely involved in neuronal signaling, suggesting that nervous system evolution might be achieved by the interaction rewiring of signaling components, such as PDZ protein-ligand interactions. Furthermore, the broad specificity of PDZ domains contributes to interaction rewiring by increasing the chance of acquiring PDZ binding motifs by sequence mutations. Our findings will prompt a new approach for the study of eukaryotic evolution by considering the rewiring of interactions as a major evolutionary process of domain-ligand interactions.
To elucidate how PDZ domain-ligand interactions have evolved, an accurate and detailed understanding of their interactions is essential. Furthermore, a network approach is useful to understand how evolution of PDZ domain-ligand interactions contributed to eukaryotic evolution, because protein functions may not be encoded in an individual protein but rather be encoded in the relationships between proteins in a protein-protein interaction network [15]–[17]. Therefore, we constructed a comprehensive network of PDZ protein-ligand interactions by integrating the experimental data of PDZ domain-ligand interactions and protein-protein interaction databases (Figure 1).
We developed a quantitative model of PDZ domain binding strengths from the experimental data of PDZ domain-ligand interactions, including interactions between 81 PDZ domains and 217 peptides from a protein array [18], the phage display of 86 PDZ domains [19], [20], interactions between 147 PDZ domains and 219 ligands from a database of in vivo PDZ domain-ligand interactions (PDZBase) [21], and literature mining [22], [23] (Figure 1A). This model converts binary interactions between PDZ domains and ligands into PWMs, which can quantify the binding strengths of a given PDZ domain and peptide sequence based on the physical and chemical properties of binding pocket residues as well as the frequencies of amino acids found in the bound peptides. To capture the binding strengths of the PDZ domain-peptide interactions, we combined a machine-learning algorithm and an information theory-based PWM method. We provide the PWMs of human PDZ domains as a resource (Table S1). In this study, we focused on the C-terminal motifs of ligands for the analysis of PDZNet. Although several internal PDZ binding motifs have been reported, most PDZ domain-ligand interactions are mediated by C-terminal residues, owing to the structural constraint on the internal motifs that require the β-hairpin fold [24], [25].
We found that the binding scores of PWMs well represent the experimental affinities of PDZ domain-ligand interactions (Figure 2). The large-scale binding affinities (Kd) of PDZ domain-peptide interactions are available for SNA1 and ERBIN PDZ domains [26]. The PWMs provided the binding scores of the interactions, which showed a strong positive correlation with the experimental affinities for both SNA1 (R2 = 0.76) and ERBIN (R2 = 0.85) PDZ domain-peptide interactions. Moreover, in vivo binding affinities of PSD-95_1 (the first PDZ domain of PSD-95) with its ligands correlated well with its binding scores from the PWMs (Figure 3A and 3B). The Kds of Kv1.4 and GluR6 to PSD-95_1 were measured experimentally [27]. Kv1.4 bound to PSD-95_1 with high affinity (Kd = 1.5 µM), whereas GluR6 bound to PSD-95_1 with low affinity (Kd = 160 µM). When we measured the binding scores of the PSD-95_1 ligands based on PWMs, the binding score of Kv1.4 was found to be higher (binding score = 14.72) than that of GluR6 (binding score = 6.00; Figure 3B). Next, we measured how precisely the in vivo ligands of PDZ domains can be rediscovered by the binding scores obtained from PDZ domain-ligand interactions. We found that although the interaction data for the target PDZ domain were excluded from the training set, 290 of 320 (91%) of the known PDZ ligands were found in the top 10 percentile of binding scores (Figure 3C). We also found that our PWMs provided reliable predictions for PDZ domains derived from various species (Table S2). Furthermore, we found that our predicted PWMs agreed well with experimental data-based PWMs [20]. We compared the PWMs derived from phage display experiments with the predicted PWMs of the MAGI1_2, DLG1_2, and PTN13_2 PDZ domains and confirmed that they were nearly identical (Figure S1).
To construct PDZNet with high-confidence interactions, we prioritized the experimentally validated PDZ protein-ligand interactions from the prediction results of the PWMs. It is a challenge to correlate the occurrence of amino acids in a linear motif to the binding specificity of peptide-binding domains [28]. The PWM method treats each amino acid position in a linear motif independently; thus, predicted interactions may include a fraction of false-positive results. Therefore, we only included interactions supported by experimental evidence. To assemble experimentally validated protein interactions, we integrated 22 different PPI databases containing 101,777 interactions among 11,043 proteins (Figure 1B and 1C).
PDZNet is composed of 97 PDZ proteins and 596 partners with 1,212 interactions (Figure 4A), which can be accessed in Table S3. PDZ proteins interact with a various number of ligands (Figure S2) and most (98%) PDZ proteins are connected in a large single network via shared ligands. Beginning with PDZNet, we generated two network projections (Figure 4B), which displayed both PDZ protein-PDZ protein and ligand-ligand connections via common interacting partners. In the “PDZ protein network” (PPN; Figure 4B, left panel), nodes represent PDZ proteins; two PDZ proteins are connected if they share at least one ligand. Meanwhile, in the “PDZ ligand network” (PLN; Figure 4B, right panel), nodes are PDZ ligands; two PDZ ligands are connected if they share at least one PDZ protein. On average, a PDZ protein interacts with 17 partners, and a PDZ ligand interacts with three PDZ proteins. We further examined whether this multispecificity is also found at the domain level. For proteins with multiple PDZ domains, PDZNet specifies the interactions mediated by individual domains, yielding 2,643 PDZ domain-ligand interactions (Figure S3). On average, a PDZ domain interacts with 14 ligands, and a ligand interacts with four PDZ domains, suggesting that the complexity of PDZNet originated from the multispecificity of PDZ domains.
We discovered that an interface similarity exists among PDZ domains that share the same ligands. In the PPN, PDZ protein pairs connected by the same ligands tend to have similar pocket residues (Figure S4). For example, SAP97_1, SAP97_2, PSD-93_1, PSD-93_2, SAP102_2, and PSD-95_1 have similar binding pocket residues that bind the same ligand (AT2B4), suggesting that gene duplications contribute to the multispecific interactions in PDZNet and increase network complexity because interaction partners from gene duplication events tend to share the same interface [12]. Indeed, we found that SAP97, PSD-93, SAP102, and PSD-95 PDZ proteins were paralogs, the products of gene duplication events. Interestingly, we also found cases of non-paralogous proteins that have similar binding specificities, which suggest that convergent evolution might also play a role in the development of network complexity. For example, two PDZ proteins, LAP2 and MAGI2, were found to interact with the same ligand, CTND2, although they are evolutionarily unrelated. Meanwhile, the PLN provides a complementary ligand-centered view of PDZNet. We found that the connected ligands in the PLN tend to have similar C-terminal sequences. As shown in Figure 4B, two PDZ ligands, ARVC and CTND2, interact with the LAP2 PDZ protein and have same binding motif (DSWV).
We then asked how PDZ domains and ligands obtained multiple partners during evolution. Gene duplication and subsequent diversification events are considered major factors for network growth. Although gene duplication played a significant role in PDZ proteins and ligand evolution [29], it may not explain how a PDZ domain can interact with multiple, non-homologous ligands.
We found that sequence mutations played an important role for the attachment of non-homologous ligands to PDZ domains. On an evolutionary time scale, the compendium of PDZ ligands expands via two processes: (1) the introduction of new PDZ ligands by gene duplication of existing partners, or (2) the de novo evolution of new interactions via the acquisition of PDZ-binding motifs (Figure 5A). To examine the extent of gene duplication in PDZNet growth, we calculated the paralog fractions of PDZNet because gene duplication products usually remain as homologous genes [30]. We discovered that the relatively small fraction of PDZ ligands that share a common partner were paralogs (33.6%), whereas a significantly larger portion of PDZ proteins that share a common partner were paralogs (54.5%; Wilcoxon's rank-sum test; p = 1.24×10−4; Figure 5B).
Next, we examined the sequence evolution of the binding motifs of human PDZ ligands and discovered that a large portion of PDZ ligands acquired their binding motifs via sequence mutations. We examined the C-terminal sequences of PDZ ligands in each PDZ domain-ligand interaction pair across 16 representative species. We found that nearly one-third of human PDZ ligands gained their PDZ domain interactions by C-terminal mutations during evolution (Table S4; experimental evidence of human PDZ domain-ligand interactions are provided in Table S5). For example, NOS1AP obtained a PDZ-binding motif via sequence mutation and became an interaction partner with the NOS1 PDZ protein from vertebrates. We discovered that NOS1AP has orthologs in a wide range of species from yeast to human (Figure 5C). To examine whether the PDZ-binding motif of NOS1AP emerged from vertebrates, we compared the C-terminal sequences and binding scores of NOS1AP from invertebrate and vertebrate orthologs. The binding of mouse NOS1AP with NOS1 PDZ protein has been confirmed experimentally [31]. The C-terminal sequences of the vertebrate orthologs of NOS1AP are identical, whereas the C-terminal sequences of the invertebrate orthologs of NOS1AP vary across species and differ from those of the vertebrate NOS1AP orthologs. Moreover, we searched for evidence of a NOS1AP–NOS1 interaction in invertebrate PPI databases, including Databases of Interacting Proteins (DIP) [32], BioGrid [33], and Comprehensive Drosophila Interactions (Droidb) [34], but none was found. When we compared the binding scores of NOS1AP to NOS1, all invertebrates orthologs showed low binding scores (average binding score = −3.03), whereas the binding scores of vertebrate orthologs were high (average binding score = 5.27; Figure 5D), indicating that NOS1AP was an invertebrate non-binder but gained the ability to bind the NOS1 PDZ protein in vertebrates.
Interaction rewiring is an effective evolutionary mechanism given that it reconfigures molecular systems without a gain or loss of genes [35]. We hypothesized that the rewiring of PDZ domain-ligand interactions via sequence mutation contributed to the evolution of the vertebrate nervous system, in which PDZ proteins and ligands play an important role [1]. To determine whether the rewiring of PDZ domain-ligand interactions had a significant impact on vertebrate nervous systems, we calculated the rewiring rates between species from Escherichia coli to humans and examined the changes in the C-terminal sequences of PDZ ligands and the binding specificities of PDZ domains (Figure S5). We found that PDZ domain-ligand interactions were most frequently rewired between invertebrates and vertebrates (Figure 6).
We also examined the types of biological processes that are significantly affected by the rewiring events of PDZ domain-ligand interactions. We found that the PDZ ligands that arose in invertebrates and gained their PDZ-binding motifs in vertebrates participated significantly in the process of neurological system development (Table S6). For example, with the emergence of vertebrates, the breakpoint cluster region protein, BCR, acquired a PDZ-binding motif and began to interact with the AFADIN PDZ protein by changing its C-terminal sequence from ARLK (binding score = −2.03) to STEV (binding score = 5.87). The binding of the PDZ domain interaction sites of BCR and the AFADIN PDZ domain was also confirmed by immunoprecipitation and NMR chemical shift perturbation experiments [36], [37]. In vertebrates, BCR controls the interaction between AFADIN and RAS GTPase [36]. AFADIN also interacts with the vertebrate-specific receptors EPHA7 and EPHB3 of the Eph-receptor family, which regulate the morphology and motility of neuronal cells through the RAS GTPase [38], [39]. Thus, the interaction between BCR and AFADIN may evolve to control EPH receptor signaling, which is greatly diversified in vertebrates. Meanwhile, we found that proteins that arose and acquired PDZ domain interaction sites in invertebrates tend to be involved in various cellular processes, such as vesicle-mediated transport, cell cycle, and RNA splicing (Table S6). These results suggest that the emergence of the vertebrate nervous system integrated preexisting functional units during the evolution of synapse complexity.
We found that metazoan-specific PDZ proteins adopted their ligands from proteins of premetazoan origin. The phylogenetic profile shows the origin of the PDZNet proteins (Figure S6). Many of the human PDZ ligands were detected in premetazoans, whereas most human PDZ proteins were only found in metazoan species. The binding scores of nearly one-half of the premetazoan orthologs of the human PDZ ligands were less than zero, indicating that these proteins were not the interaction partners of PDZ proteins in premetazoan species. However, these proteins acquired PDZ-binding motifs in metazoans and began to interact with metazoan PDZ protein partners. For example, EXOC4 is found in yeast and gained its PDZ-binding motif in vertebrates (Figure 7; the multiple sequence alignment of EXOC4 orthologs is shown in Figure S7). The binding of mouse EXOC4 with SAP102 via its PDZ domain was validated in a yeast two-hybrid system and by pull-down assays [40]. The yeast ortholog of EXOC4 is a component of the exocyst complex, which transports vesicles to the plasma membrane. After gaining a PDZ-binding motif recognized by the SAP102 PDZ domain in vertebrates, it participates in NMDA receptor trafficking [40]. This finding suggests that the evolution of metazoan functions required the rewiring of functional modules that existed in premetazoans and contributed to network growth. Indeed, previous studies have noted that proteins of premetazoan origin played important roles in metazoan-specific functions, such as synaptic signaling [29]. Together, the premetazoan ancestry of PDZ ligands highlights the de novo occurrence of PDZ domain-ligand interactions in the rewiring of metazoan evolution.
Next, we asked which physiological system was most affected by the mutations of PDZNet proteins. Mutations could affect the binding specificity of PDZ-ligand interactions via the replacement of interfacial residues or the destabilization of PDZ domain and ligand structure. If an interaction gained from the evolution of PDZNet had contributed to the development of a certain physiological system, an alteration of the interaction could be associated with genetic diseases caused by a malfunction of the system.
We investigated the disease associations of the PDZNet components and found that many PDZNet proteins are significantly associated with neurological diseases (Figure 8). Human genetic diseases were mapped to the components of PDZNet using disease-gene association data from the Online Mendelian Inheritance in Man (OMIM) [41]. Genetic diseases were classified into 20 disease classes based on the physiological system affected [42]. We examined whether a certain disease class was more enriched in the PDZNet components than the other proteins in the human interactome. Of the 20 disease classes examined, the neurological disease class was the most highly associated with mutations of the PDZNet components (Table S7). For example, a mutation in the PDZ protein, NLGNX, perturbed its PDZ domain interaction with the ligand protein, SNTG2, which is suggested to be a cause of mental retardation and autism [43]–[45]. This finding reconfirms the importance of PDZ domain-ligand interactions in the evolution of the nervous system. A morbid map of PDZNet components with the classification of genetic diseases is provided in Table S8 as a resource.
In this study, we describe the first PDZ protein-ligand interaction network coupled with quantitative binding strength. Our network approaches elucidated how PDZ domains have diversified their binding partners in the organization of various signaling complexes from receptors to downstream signaling relays. Moreover, we showed that de novo evolution of PDZ domain-ligand interactions played an important role in the growth of PDZNet. These findings provide empirical evidence for a network evolution model that highlights the rewiring of interactions as a mechanism of functional innovation.
PDZNet provides information beyond just the state of interaction binding. First, PDZNet provides information regarding the binding interface. High-throughput experiments provided large-scale PPI information; however, the identification of which amino acids were used in the interactions has been difficult. The quantitative model of PDZ domain-ligand interactions provides sequence information on domains and linear motifs, enabling a deeper understanding of the mechanisms involved in their interactions. Second, PDZNet provides the binding strengths of the interactions. The quantitative binding strengths of PDZ domain-ligand interactions enable us to understand the competition among interaction partners for switching between signaling flows.
The multispecificity of PDZ domain-ligand interactions has unique advantages in the evolution of PDZ domain function in the cell signaling network. First, the multispecificity of PDZ domains contributes to the frequent rewiring of PDZ domain-ligand interactions and broadens the extent of recognizable sequences, thus increasing the chance that a protein gains a suitable sequence to interact with its partners. Indeed, we found that PDZ domain pockets prefer multiple amino acids for interactions. We analyzed amino acid preference patterns from the PWMs of human PDZ domains (Figure S8) and found that the degeneracy of binding motifs facilitate the binding of different PDZ ligands to the same PDZ domain. This finding is consistent with those of a recent study that revealed the specificities of PDZ domains lie on a continuum [18]. Second, the multispecificity of PDZ domains enables the combinatorial assembly of signaling complexes that control signaling processes. PDZ proteins interact with many signaling proteins and form preassembled complexes, which are important for the precision of information flow and the fidelity of cell signaling events [46]. An interesting observation from our network approach is that a PDZ protein is connected to many ligands. These ligands may interact with a PDZ protein in a tissue-specific manner; the subsequent cell type-dependent expression of the PDZ ligands may lead to an alternative assembly of signaling complexes, thus enabling cell type-specific responses for extracellular signals. Indeed, we observed that the ligands of the SAP97 PDZ protein showed tissue-specific expression patterns, allowing the formation of tissue-specific cell signaling complexes (Figure S9). Third, the multispecific interactions of PDZ domains may enhance the robustness of the signaling processes mediated by PDZ domains. The robustness of the cell signaling network is known to increase because several means often exist to achieve one function as the failure of one can be compensated by others [47]. In PDZNet, PDZ domains tend to interact with a series of homologous proteins, particularly cell surface receptors. This interaction may ensure reliable transmission of signals mediated by PDZ proteins to the plasma membrane.
We found that almost one-third of human PDZ ligands obtained their PDZ-binding motifs via C-terminal sequence mutations, providing evolutionary advantages to the PDZ domain-mediated interactions. First, the formation of linear motifs is an efficient mechanism to increase the number of interactions. Emergence of short linear motifs rarely disrupts the protein structure and can be accompanied by few amino acid changes [6]. Second, the de novo evolution of interactions via sequence mutation provides an effective means for functional innovation. Gene duplication is known to have a limited role in the molecular innovation of biochemical function but facilitates the modularization of functional networks by specialization [30]. In contrast, the de novo evolution of interactions allows connections between evolutionarily unrelated functional modules, thus enabling the reconfiguration of the molecular system. For instance, gain of the PDZ domain-ligand interaction between the EXOC4 PDZ ligand and the SAP102 PDZ protein demonstrated an innovation by bridging two different functional modules. We examined species-specific functional annotations of PDZ ligands and found that yeast EXOC4 participates in vesicle transport with other exocyst complex members, but vertebrate EXOC4 regulates NMDAR transport to the postsynaptic membrane by interacting with the SAP102 PDZ domain [40]. Third, when a PDZ protein gains ligands by sequence mutation, it avoids a loss of fitness caused by an increase in dosage. The de novo evolution of PDZ domain-ligand interactions does not increase the copy number of the PDZ ligand genes, avoiding an unfavorable increase in protein concentration. In contrast, gain of interactions by duplication may cause a loss of fitness because proteins that contain linear motifs tend to be intrinsically disordered and dosage sensitive [48].
We were also interested in whether new PDZ domain interaction sites were acquired via C-terminal point mutations or DNA insertions. After careful observation of DNA modifications in newly acquired PDZ ligands, we found instances of both. For example, protein PBK of Macaca mulatta acquired PDZ domain interaction motif “ETDV” via C-terminal point mutations in which a single nucleotide substitution (T→C) changed Ile to Thr and another mutation (C→T) changed the codon for Gln to a stop codon (Figure S10A). On the other hand, EXOC4 acquired new PDZ domain interaction sites via DNA insertion in Oryzias latipes (Figure S10B). A large section of DNA inserted near the C-terminus of EXOC4 caused a frame shift mutation, which in turn became the PDZ domain-binding motif “ITTV.”
We found that the rewiring of PDZ domain-ligand interactions most frequently occurred between invertebrates and vertebrates. This massive rewiring may be connected to repeated rounds of whole-genome evolution in ancestral vertebrates. According to Ohno's model [49], when a gene is duplicated, mutations freely accumulate in the redundant duplicate copy under no selection. Therefore, the duplicate copy has a greater chance of developing new functions without altering existing functions. This evolutionary mechanism may facilitate network rewiring in early vertebrates.
We found that the components of PDZNet are largely associated with neurological diseases. We then asked whether we could identify mutations affecting PDZ-ligand binding, which causes genetic diseases. The disruption of the PDZ domain interaction between PICK1 and GluR7 is known to cause seizures, a chronic neurological disease [50]. Mutations in the C-terminal sequence of GluR7 disrupted its PDZ domain interaction with PICK1. To examine whether our quantitative model can predict the effects of mutations in GluR7, we generated the PWM of the PICK PDZ domain and calculated the binding scores for both the wild-type and mutant forms of GluR7 (Figure S11). We found that the wild type had a high binding score (5.98), and the mutant had a much lower binding score (−0.02). This example illustrates how our method can be applied to characterize genetic diseases that are caused by mutations affecting PDZ domain-ligand interactions.
An important issue of the present biological network study is its incompleteness [51]. PDZNet has room for improvement regarding network coverage in two respects: shortage of nodes and links in the current network. To test whether the conclusions obtained in this work are sufficiently robust with regard to both, we constructed smaller random networks from PDZNet and repeated the analyses. In each trial, 20% of the proteins or interactions were randomly removed from PDZNet. We found that in all tests, the overall organization of the rescaled PDZNet remained largely unaltered, and the conclusions and the differences between the paralog fractions of the PDZ proteins and ligands were retained (Figures S12, S13, S14, S15, S16, S17), supporting the robustness of our findings to the future expansion of PDZ domain-ligand interactions.
Due to the incompleteness of the interactome networks, expansion of network coverage is of significant value. PDZ domain-ligand interactions were relatively difficult to detect using current experimental techniques because transient interactions are often lost during experimental washing steps. Furthermore, a PDZ domain-ligand interaction often depends on phosphorylation [24], so it can be missed when screening for protein interactions preformed in a single condition. Therefore, many PDZ domain-ligand interactions remain to be discovered. We anticipate that putative PDZ domain-ligand interactions with high-binding scores from PWMs, expression correlations, and similar phylogenetic profiles may be used to uncover novel interactions. Therefore, we provide a candidate list of PDZ domain-ligand interactions to assist in the discovery of novel PDZ domain-ligand interactions (sbi.postech.ac.kr/pdz).
We assembled experimentally confirmed PDZ domain-ligand interactions from various data sources. In detail, we obtained PDZ domain-peptide binding data from a high-throughput binding assay between 81 mouse PDZ domains and 217 peptides derived from genome-encoded receptors by protein array [18]. We collected in vivo PDZ domain-ligand interactions from the published literature, including peptide binding data of Drosophila INAD [22], [23] and human AF-6 [52] PDZ domains. Additionally, a PDZ domain-ligand interaction database, PDZBase [21], which currently lists 339 in vivo interactions between 145 PDZ domains and 217 ligands, was used. Finally, we obtained 54 human and 28 worm PDZ domains in a high-throughput binding assay [20] and four N-terminal PDZ domains of human INADL using phage display [19]. The collection of these data resulted in 4,467 experimentally confirmed PDZ domain-ligand interactions.
We collected 563 human PDZ domain sequences from the Pfam repository [53]. After eliminating redundancy, we obtained 268 sequences. We then examined pocket residues of the PDZ domains using hidden Markov model (HMM) alignment, removed the sequences that did not align in the pocket region, and finally obtained 241 distinct human PDZ domains.
We developed a two-step approach to quantify the strength of binding between the PDZ domains and ligands. Using this approach, the binding affinity between each PDZ pocket and its corresponding ligand position was predicted individually based on the idea that the contribution of each ligand position to the binding affinity is additive [54], which is a widely accepted view in the modeling of linear motif interactions [55], [56]. The workflow of our approach is summarized in Figure 9.
In the first step, we designed the selectivity space of each pocket (Figure 9, left panel) to contain 20 axes, representing preferences for the corresponding amino acids in the peptide ligand (Figure S18). To obtain the ligand selectivity of the PDZ domains, three types of interaction data were used, namely protein arrays of mouse PDZ domains against synthesized peptides [18], collections of individual studies of PDZ domain-ligand interactions [21]–[23], and high-throughput binding assays using phage display [19]. We made a multiple sequence alignment (MSA) of PDZ domains using a HMM for the PDZ domain. We then extracted pocket residues from the MSA and encoded them into feature vectors based on their physicochemical properties. With the feature vectors, we constructed 20 training sets. In each training set, the feature vectors from specific amino acid-preferring pockets were used as a positive set, and the remainder was used as a negative set. We then applied Fisher's Linear Discriminant (FLD) analysis to these training sets such that discriminative axes were trained to distinguish specific amino acid-preferring pockets, resulting in a projection matrix composed of the axes' direction vectors. By multiplying the feature vectors with the projection matrix, we located the pockets of PDZ domains in the selectivity space. Thus, the selectivity spaces for each pocket capture intrinsic amino acid preferences from binary interaction data.
In the second step, to build a PWM of a query PDZ domain, we generated an affinity profile that represents the relative affinity contributions of 20 amino acids to the PDZ domain pocket (Figure 9, right panel). Based on the assumption that closely residing pockets in the selectivity space are similar in their amino acid preferences, we gathered the nearest neighbors of a query domain in a selectivity space to establish an affinity profile from their preferred amino acid sets. Pocket residues of a query PDZ domain were encoded into feature vectors using physicochemical properties and then located on a selectivity space using the projection matrix described above. We gathered ligand sets preferred by the nearest neighbors of the query pocket and estimated the binding affinity contributions of each position.
We converted the pocket residue sequences of a PDZ domain into vector representations by replacing all 20 amino acids with 10 physicochemical properties (amino acid indices) that describe the number of hydrogen bond donors [57], polarity [58], volume [58], bulkiness [59], hydrophobicity [59], [60], isoelectric point [59], positive charge [57], negative charge [57], electron ion interaction potential [61], and free energy in water [62]. We normalized the values such that the standard deviation is 1 and the average is 0.
Our goal was to predict the specificities of a PDZ domain without knowledge of its structure. As such, a method to extract pocket residues from the sequences of PDZ domains was designed. To identify the positions of pocket residues within the PDZ domain sequence, an MSA was constructed, and the known structure of the PSD-95_1 domain was referenced. We performed a multiple alignment of the PDZ domain sequences using a HMM [63] and an HMM that was optimized for PDZ domains from Pfam [53] and aligned the secondary structure profile of the PSD-95_1 domain with the sequence alignment. Pocket residues were subsequently extracted according to the pocket definitions described in Wiedemann et al. [26] (Figure S19).
To estimate the PDZ domain-ligand binding affinity, we adopted an information theory-based PWM method that is widely used to estimate protein-DNA binding affinities [64], [65]. In each selectivity space, 40 preferred amino acids of the neighboring pockets of the query were gathered. A PWM was calculated using the four sets of collected amino acids in which amino acid frequencies were calculated at each ligand position; these frequencies were compared to the background frequency that we expected to observe for the C-terminal sequences of the ligands.PWM(a, i) is the affinity contribution of amino acid a at the ith position, fa,i is the frequency of amino acid a at the ith position in the collected set, and pa,i is the background frequency, defined as the probability of observing amino acid a at the ith position in any ligand protein. The constant k was empirically determined to be 1.921, so that PDZNet includes all experimentally confirmed PDZ domain-ligand interactions as positive binding scores.
A PWM was used to calculate the binding score of a potential interaction partner with a given sequence by summing the corresponding amino acids for the affinity contribution of each position. The binding score of each peptide was calculated according to the following formula:where PWM(Si,i) is the affinity contribution of the amino acid Si at the ith position in the matrix and Si is the amino acid at the ith position of the peptide.
Affinity values of the 5,257 peptides against both the SNA1 and ERBIN PDZ domains were obtained from Wiedemann et al. [26], who assessed the affinity values of the peptides with a combination of experiments (i.e., surface plasmon resonance and Boehringer light unit) and statistical analyses.
To evaluate the performance of our method, we measured the ability to identify the 217 known binding partners of 145 PDZ domains in the PDZBase [21]. Using a standard leave-one-out procedure, we generated PWMs and genome-wide rank lists of interaction candidates for each domain using their corresponding PWMs. Our method successfully predicts the binding partners of PDZ domains for which no interaction data are available. When we examined the percentile ranks of experimentally confirmed PDZ domain-ligand interactions, most were enriched at high positions in the rank lists (90th∼100th percentile of the binding score; Figure 3C).
We compiled human protein interactions from a total of 22 existing protein interaction databases: the Bio-molecular Interaction Network Database (BIND), the Human Protein Reference Database (HPRD), the Molecular Interaction database (MINT), DIP, IntAct, BioGRID, Reactome, the Protein-Protein Interaction Database (PPID), BioVerse, CCS-HI1, the comprehensive resource of mammalian protein complexes (CORUM), IntNetDB, the Mammalian Protein-Protein Interaction Database (MIPS), the Online Predicted Human Interaction Database (OPHID), Ottowa, PC/Ataxia, Sager, Transcriptome, Complexex, Unilever, protein-protein interaction database for PDZ-domains (PDZBase), and a protein interaction dataset from the literature [66]. We removed low-confidence interactions that were not supported by direct experimental evidence. The final network comprises 101,777 interactions between 11,043 human proteins.
We collected all physical interactions mediated by the PDZ proteins from the integrated PPI network. This PDZ protein-mediated interaction set may have some interactions that are mediated by interaction domains other than PDZ domains, because many PDZ proteins have various domains other than PDZ domains. Therefore, we removed such interactions that were connected by domain-domain interactions rather than PDZ domain-ligand interactions. First, we confirmed that PDZ domain-mediated interactions are rarely augmented by other interaction domains. We found that domain-domain interactions are not present in the experimentally confirmed PDZ protein-ligand interactions from the PDZBase [21]. Furthermore, we found that domain-domain interactions are only enriched in low-scoring PDZ protein interactions (Figure S20). Based on these observations, we removed domain-domain interactions from PDZNet.
We also removed interactions that could be mediated by other peptide-binding domains, such as SH3 and WW domains, rather than PDZ domains. We searched the known peptide-binding motifs and removed interactions mediated by peptide-binding domains that had low binding scores. The cut-off binding score was set to the lowest binding score of the experimentally confirmed PDZ domain-peptide interactions from the PDZBase [21]. The binding score represents the predicted binding strength between a PDZ domain and the C-terminal sequence of its partner. Subcellular localization information was taken from Swiss Prot and consensus localization annotations [67].
Let two species, i and j, be in a common tree with humans, and species i is more distant from humans. If a human PDZ interaction is absent in species i and present in species j, we define the interaction as rewired. Thus, the rewiring occurred during the time interval between the emergence of species i and j. We also consider that all proteins in species i could be rewired to PDZ proteins in species j. Thus, we define the rewiring rate as the following:where nj is the number of rewired interaction found in species j; t is the divergence time from human; pall,i is the number of proteins orthologous to human protein in species i; and ppdz,j is the number of proteins orthologous to human PDZ proteins. Divergent time was obtained from the timetree (http://www.timetree.net).
To analyze the interactions between orthologous PDZ domains, we calculated the binding scores of the C-terminal sequences of orthologous PDZ ligands and the predicted PWMs with orthologous PDZ domain sequences.
We examined whether particular protein functions were enriched for protein categories that were defined based on the time of protein emergence and PDZ-binding motif acquisition. We systematically classified PDZ ligands into two categories: (1) proteins arose in invertebrates and acquired PDZ domain interaction sites in vertebrates; (2) proteins arose and acquired PDZ domain interaction sites in invertebrates; we then analyzed the overrepresented functional terms of each group. We used DAVID [68] for gene set enrichment analysis. All ligand proteins were used as background.
Mutations of PDZNet proteins were mapped to genetic diseases using disease-gene association databases from OMIM. The OMIM database lists gene-disease associations between 2,929 disease types defined by Morbid Map (MM) and 1,777 genes associated with particular disease types. Disease types were further categorized into 1,340 distinct diseases by joining disease subtypes into a single disease if similar disease names were used. These disease types were further classified into 20 disease classes based on the physiological system affected [42]. The p values for over- or under-representation of the disease-associated genes in PDZNet were obtained using a hypergeometric distribution. We independently calculated the probability of the disease-associated genes in each class.
We created a user-friendly web service that provides a PWM and rank list of interaction candidates of a given PDZ domain sequence (Figure S21). The automated pipeline of the web service extracts pocket residues from the query PDZ domain sequence, predicts binding specificity (represented as a PWM), and generates a genome-wide rank list of potential ligands. The web service can handle various exceptions. For example, if a query is an incorrect PDZ domain sequence or an incorrect alignment was made in the pocket residues, the web service provides messages with explanations.
|
10.1371/journal.pcbi.1006894 | The logic of ionic homeostasis: Cations are for voltage, but not for volume | Neuronal activity is associated with transmembrane ionic redistribution, which can lead to an osmotic imbalance. Accordingly, activity-dependent changes of the membrane potential are sometimes accompanied by changes in intracellular and/or extracellular volume. Experimental data that include distributions of ions and volume during neuronal activity are rare and rather inconsistent partly due to the technical difficulty of performing such measurements. However, progress in understanding the interrelations among ions, voltage and volume has been achieved recently by computational modelling, particularly “charge-difference” modelling. In this work a charge-difference computational model was used for further understanding of the specific roles for cations and anions. Our simulations show that without anion conductances the transmembrane movements of cations are always osmotically balanced, regardless of the stoichiometry of the pump or the ratio of Na+ and K+ conductances. Yet any changes in cation conductance or pump activity are associated with changes of the membrane potential, even when a hypothetically electroneutral pump is used in calculations and K+ and Na+ conductances are equal. On the other hand, when a Cl- conductance is present, the only way to keep the Cl-equilibrium potential in accordance with the changed membrane potential is to adjust cell volume. Importantly, this voltage-evoked Cl--dependent volume change does not affect intracellular cation concentrations or the amount of energy that is necessary to support the system. Taking other factors into consideration (i.e. the presence of internal impermeant poly-anions, the activity of cation-Cl- cotransporters, and the buildup of intra- and extracellular osmolytes, both charged and electroneutral) adds complexity, but does not change the main principles.
| We have developed software that calculates membrane potential and cell volume that result from redistribution of principal ions (K+, Na+, and Cl-) during normal cellular activity and experimental manipulations. Calculations in the model are done by an iterative charge-difference method that makes few assumptions about governing equations. Most of the features that were considered to be important for volume and voltage regulation were incorporated in the model, including the unique capability to perform calculations with different values of transmembrane water permeability. We have used the program to reexamine interactions between ionic fluxes, membrane potential, and cell volume and found that there was a previously unappreciated difference in the way that the distribution of cations and anions affect the cell. Na+ and K+, which are distributed unevenly across the membrane by the Na+/K+-ATPase, are primarily responsible for the membrane potential, but, contrary to popular belief, do not directly participate in volume regulation. On the other hand, the Cl- conductance determines the extent of volume changes, because Cl- has to follow the changes of membrane potential, which inevitably leads to changes in cell volume. The software is available to download and use for other investigations.
| The transmembrane movements of ions during neuronal activity are inevitably associated with changes in ionic concentrations, both intracellularly and extracellularly. This activity-dependent redistribution of ions can be osmotically imbalanced and consequently can lead to changes in volume of the cells and of the extracellular space (ECS). It is usually assumed that three principal ions–Na+, K+, and Cl-—are responsible for the link between transmembrane conductance, voltage and volume alterations associated with neuronal activity. Among them, extracellular K+ experiences the largest relative activity-dependent changes. The increase of the extracellular K+ concentration ([K+]o) during neural activity is the easiest to detect, and was recorded first [1, 2], using ion-selective microelectrodes [3]. Soon it was discovered that the increased [K+]o induced by electrical stimulation was associated with a 50% reduction of ECS volume ([4] in the cortex of cat). However, it also was shown that the relationship between ions, voltage and volume was not simple, since in some other layers of the cortex, shrinkage of ECS was not detected in spite of considerable elevation of [K+]o [4].
The correlation between the increase of [K+]o and decrease of ECS volume was found in different parts of the nervous system under various conditions (honey bee eye, light stimulation—[5]; optic nerve, electrical stimulation—[6]; spinal cord, electrical stimulation—[7]; cortex during spreading depression—[8]). However, when extracellular concentrations of Na+ and Cl- ([Na+]o and [Cl-]o) were measured to obtain a full picture of the anticipated osmotic imbalance of ions, the results were sometimes confusing. For instance, it was shown that the decrease of ECS volume evoked by electrical stimulation was indeed accompanied by a decrease of [Na+]o, since Na+ enters cells during stimulation, but this decrease had the same amplitude as the [K+]o increase, and [Cl-]o started to change only after the stimulation, during so called “self-sustained neuronal afterdischarges” [9]. During spreading depression, when ECS was reduced to one fourth of its original volume [8], a large increase of [K+]o was indeed exceeded by even larger decrease of [Na+]o [10, 11]. But the decrease of [Cl-]o was markedly larger (by 12–31 mM) than needed for electrical compensation of extracellular cation deficiency. As discussed further below, these changes still have not been fully explained with a mechanistic model.
The vertebrate retina presents a case of special interest in this respect because it consists of cells that respond to light differently–with predominantly depolarization in proximal layers (ganglion cells, amacrine cells; see S1 Text), but with hyperpolarization in distal layers (photoreceptors, horizontal cells). Accordingly, during illumination the ECS volume decreases in proximal retina, but increases in distal retina [12, 13]. In proximal retina, similarly to brain, ECS shrinkage is associated with a [K+]o increase and also with a larger [Na+]o decrease and with a compensating [Cl-]o decrease [14]. In distal retina the changes are reversed: ECS expansion is associated with a [K+]o decrease and with a larger [Na+]o increase [14]. However, [Cl-]o still decreases in the outer retina when it is expected to increase in order to compensate for total ECS cation excess.
It should be noted that although measuring extracellular ion concentrations with ion-selective microelectrodes is the best available method to obtain data on ionic redistribution during neuronal activity, these measurements are technically difficult, and results must be interpreted with caution. In the case of [Na+]o and [Cl-]o changes, the measured voltage changes of the ion-selective electrode are small (except in spreading depression) and can be partly compromised by possible electrical artifact arising from a combination of changes in field potential and the huge resistance of the ion-selective microelectrode [14, 15]. Also, the sensors are not absolutely selective; for instance, some Na+-selective sensors are influenced by changes in extracellular Ca2+ [9], and some Cl--selective sensors respond to pH fluctuations [14]. Most importantly, the changes of extracellular ions and volume evoked by neuronal activity often stimulate reactions of ever present nearby glial elements, which in turn can alter those ionic changes and affect ECS volume [16, 17].
Nevertheless, it is possible to conclude that measurements with ion-selective microelectrodes revealed a certain pattern of related changes in voltage, volume and distribution of the principal ions (K+, Na+, and Cl-) that more or less consistently (within the limits of the method) repeats itself in various parts of the nervous system. Neuronal excitation, which is in most cases associated with depolarization due to an increase of Na+ conductance, results in Na+ influx responsible for the [Na+]o decrease. This Na+ influx is electrically compensated partly by outward K+ flux and partly by inward Cl- flux, which leads to an increase of [K+]o and a decrease of [Cl-]o. The consequent transmembrane transfer of NaCl into the cells evokes osmotically obliged movement of water, decreases ECS volume, and increases the cell volume. When the neuronal activity is associated with hyperpolarization (like light-induced responses of photoreceptors and horizontal cells in the distal part of the vertebrate retina) the changes of ion concentrations and volumes have opposite signs.
However, the reasons for the voltage-dependent, osmotically imbalanced redistribution of the main ions that lead to volume changes are not entirely clear. Movement of a very small amount of ions is enough for changes of the membrane potential, and that amount has no practical effect on ionic concentration, so ion and volume changes must be more complex than expected from membrane potential considerations alone. The ionic concentrations are changed when the precise balance between influxes and effluxes through membrane passive and active transport systems that existed at rest is temporarily disturbed during activity. For instance, the light-induced decrease of [K+]o in the distal retina is a result of a temporal inequity between the passive K+ leak out of the photoreceptors, which is quickly reduced by the hyperpolarization, and the active K+ pumping into the cell by Na+/K+-ATPase, which needs time for adjustment [18]. It is natural to assume that the osmotically imbalanced ionic changes that lead to changes in volume could be a result of the unequal exchange of Na+ for K+ by the Na+/K+-ATPase (3 Na+ out of a cell for 2 K+ in). Because of this, the decrease of [Na+]o during neuronal depolarization could be expected to be larger than the increase of [K+]o. That cation imbalance must be electrically compensated by the extracellular decrease in an anion (most probably by Cl-) concentration, further diminishing the extracellular osmolarity compared to the intracellular. On the other hand, the passive transmembrane Cl- flux itself is directly affected by changes in membrane potential–attracted into the cell by depolarization and repelled out of the cell by hyperpolarization. In this case, appropriate Na+ and/or K+ flux is needed for electroneutrality, creating osmotically active NaCl/KCl transfer. Thus, there is currently no simple and single explanation for the link between voltage, volume and ions in the nervous system. Is unequal exchange of Na+ for K+ by the Na+/K+-ATPase responsible for osmotically imbalanced redistribution of ions? Or is ion redistribution the consequence of the direct influence of the changing membrane potential on Cl- flux? Maybe both factors play their roles; in this case, what is the contribution of each?
Here we will use computational modeling to provide answers for those questions. Numerous computational models that aim to understand the interrelations among ions, voltage and volume have been developed recently [19–25], (for review see [26, 27]). But most of them are based on modifications of the Goldman equation, and the limitations of this approach (particularly for a dynamic, changing system) have been well described [27]. Alternatively, a much more attractive “charge-difference” method for calculations of simultaneously changing ionic concentrations, membrane potential, and cell volume was introduced [28].
In this work a charge-difference computational model was used for further understanding of the link between ions, voltage, and volume. The list of new and improved features that make our program different from that published earlier is presented in the Methods. The focus was on the specific roles for cations and anions, which is probably the most important feature of the work distinguishing it from current literature, where Na+, K+, and Cl- were usually treated together. Special attention was directed to the energy requirement to support voltage and volume changes. It also was demonstrated that, contrary to intuitive assumptions, changes of the membrane potential do not necessarily lead to changes in volume, and changes of volume can have no effect on the cation concentration and the membrane potential. Additionally, both Donnan and Double Donnan equilibrium were reexamined. Since the water permeability of the membrane is critically important for Donnan equilibrium, some simulations with various values of this parameter were performed. Our program is the only one that is capable of such calculations.
It should be noted that bicarbonate ion, which is probably the second most important anion after Cl-, was omitted here. Although HCO3- can move across cell membranes through numerous GABA and glycine channels and by certain cotransporters and exchangers, it is involved in the fundamentally important CO2/HCO3- buffering system, and its concentration is mostly determined by highly diffusible CO2. Accordingly, it is tightly linked to energy metabolism, generation and evacuation of acidic metabolic wastes and other processes that require separate study, and it is beyond the scope of this paper.
The software is offered to share (https://sites.northwestern.edu/ralcomputational/) and significant efforts were directed toward making it flexible and user friendly.
The model calculates membrane potential (Em) and cell volume (Vol) depending on extra- and intracellular concentrations of K+, Na+, and Cl-, their transmembrane conductances, and the activities of the Na+/K+ ATPase, Na+,K+,2Cl--cotransporter and K+,Cl--cotransporter. These two cation-Cl- cotransporters were included in the model because they directly link Cl- with cations and are prevalent in the nervous system. We also take into account the concentration ([An-]i) and mean charge valence (z) of intracellular membrane-impermeant osmolytes, which comprise a substantial amount of the intracellular anions. In this respect our model is similar to the most advanced and recent models [23–25, 28]. Additionally, our model has some capabilities which previous models do not have. First, the calculations can be performed with different values of transmembrane water permeability. In existing models on volume regulation very high water permeability is accepted by default, so water instantly follows the ions and other osmolytes, changing the cell volume, but preventing any osmotic difference between the internal and external solutions. In most of our simulations the water permeability was also assumed to be large in order to focus on other aspects of ion-dependent volume-voltage regulation. But in one series of simulations we used the widest possible range of water permeability (from infinity to zero) to investigate its consequences for the cell volume and internal osmolarity. In the extreme case of zero water permeability we can simulate the development of Donnan equilibrium. Second, electrically neutral impermeant osmolytes (like glucose) can be added to extracellular fluid to simulate Double Donnan effects. Third, to simulate certain experimental environments, the model can perform calculations in conditions where osmolyte concentrations change with time. In this part of the work, we modeled the buildup of four different substances: a) external electrically neutral impermeant osmolytes, b) extracellular NaCl, c) internal electrically neutral impermeant osmolytes, and d) internal electrically charged impermeant osmolytes with the addition of an appropriate amount of Na+ to maintain electroneutrality. Besides, the model permits various stoichiometries of the Na+/K+-ATPase, which has been done [24] or can be done [23] in previous work. In most simulations a normal stoichiometry of 3 Na+ to 2 K+ was used, but we also examined a hypothetical case when an electroneutral pump (3 Na+:3 K+) was responsible for redistribution of the cations, and some other more exotic stoichiometries were also tested.
With the exception of mentioned above buildups of external NaCl and electrically neutral impermeant osmolytes, all extracellular concentrations during all other simulations were assumed to be constant (as for an isolated cell in a Petri dish).
The parameters are expressed in an easily appreciable physical form and the values are biophysically realistic as far as possible. The conductance of ions is expressed in ions/(sec*V) and can be converted to the usual electrophysiological measure of conductance in Siemens. For instance, the value of 2*1010 ions/(sec*V) for total ionic conductance, which was often used in this work, is equal to 3.2*10−9 coulombs/(sec*V), i.e. 3.2*10−9 Siemens (or 3.2 nS). This corresponds to an input resistance of 312.5 MΩ, a reasonable value for a small sized neuron (for instance, the input resistance of a starburst amacrine cell in the rabbit retina is in the range of 200–250 MΩ [29]. Similarly, the basis for many parameters that were chosen here is experimental work performed mostly on vertebrate retina. For example, a relatively low ratio of gK/gNa is characteristic of the photoreceptors [30, 31], which respond to light with a decrease of gNa [32], as in the Results: “Conductance of Cl- and cell volume.”. The transporter activities are presented in cycles/sec for better comparison to fluxes through the conductances, since the number of cycles/sec is proportional (and in some cases–equal) to the number of ions transferred per second.
The same principles for calculations were used here as in the “charge-difference” method of Fraser and Huang [27, 28]. There was no attempt to derive an equation that describes Em and Vol from ionic concentrations and conductances and activities of transporters. Instead, our program 1) counts transmembrane fluxes through the channels and transporters for each ion during a short time period, when (important!) the conditions are assumed to be unchanged, 2) calculates the resulting changes in intracellular ionic concentrations, osmolarity and electrical charge at the end of this time period, and 3) makes appropriate adjustments of the intracellular concentrations, Em and Vol. These three steps in the calculation are described in the sections below (e.g. “The 1st set of calculations”). This cycle repeats over and over again. If a balanced (resting) state was reached, the combined fluxes through all appropriate channels and transporters for each ion (K+, Na+, and Cl-) would equal 0 and ionic concentrations, Em and Vol would not change. When the conditions (a conductance or activity of a transporter) change, the balance will be disturbed, and a new balanced state will be found with potentially new values for Em and Vol, as well as for ionic concentrations. Consequently, at the end of the calculation the activities of the transporters that depend on ionic concentrations can be different from their initial values.
The discretization of a continuous process, which is the core of this method, is an apparent idealization, but it does not affect the precision of results when the resting state is found. And during the dynamic phase, any desired level of accuracy can be achieved by choosing an appropriately short discrete time step. In this work the duration of the time step was 0.1, 0.5, or 1.0 millisecond.
After choosing an initial set of concentrations and pump rates, transmembrane fluxes for each ion (in ions/sec) through each transport system are calculated. Inward fluxes are assumed to have a positive sign, and outward fluxes are considered to be negative, in accordance with the way they affect intracellular ionic concentrations.
First, fluxes are added separately for each ion and the sums are multiplied by the time step to produce the amount of ions that were moved in or out of the cell. Then the amounts are converted into concentrations. Buildups of extra- or intracellular osmolytes defined by the investigator are also taken into account. Buildups are distinguished from fluxes because these are exogenous substances that are added de novo to one side of the membrane at a specified rate. They then can affect the concentrations of substances, but never affect electrical charge, since they are either electrically neutral substances or an electrically balanced combination of cations and anions. The concentrations at the beginning of the current time step are marked with index b and the concentrations at the end of the time step (yet before possible adjustments for volume changes) are marked with the index e. To obtain the changes in total intracellular electrical charge, the change in the amount of Cl- is subtracted from the change in the amount of cations and the result is multiplied by the charge of one cation.
dNa=(NaFc+NaFp+NaFnkc)*st;
(12)
dK=(KFc+KFp+KFnkc+KFkc)*st;
(13)
dCl=(ClFc+ClFpc+ClFnkc+ClFkc)*st;
(14)
[osm]o,e=[osm]o,b+bOso*st;
(15)
[Na+]o,e=[Na+]o,b+bNaCl*st;
(16)
[Cl‑]o,e=[Cl‑]o,b+bNaCl*st;
(17)
[osm]i,e=[osm]i,b+bOsi*st;
(18)
[An‑]i,e=[An‑]i,b+bAn*st;
(19)
[Na+]i,e=[Na+]i,b+dNa/(vole*L)+bAn*st*(‑z);
(20)
[K+]i,e=[K+]i,b+dK/(vole*L);
(21)
[Cl‑]i,e=[Cl‑]i,b+dCl/(vole*L);
(22)
dQ=(dNa+dK–dCl)*e;
(23)
where dNa, dK, and dCl are changes in the intracellular amount of respective ions; st is the duration of the discrete time step, [osm]o and [osm]i are extra- and intracellular concentrations of impermeant neutral osmolytes; [An-]i is the concentration of internal impermeant anion; z is the mean valence of An; bOso, bOsi, bNaCl, and bAn are “buildups” of extra- and intracellular impermeant neutral osmolytes, external NaCl and of internal impermeant anion, respectively; vole is the cell volume at the beginning of the time step; L is Avogadro’s number (6.02*1023 mol-1); dQ is the change in internal electrical charge (in coulombs); and e is the electrical charge of one cation (1.6*10−19 coulomb). “Buildups” are inputs to the model to allow gradual changes in applied concentrations over some period of time.
Notes: All fluxes are added to each other algebraically. For instance, KFp and KFnkc are positive and increase the amount of intracellular K+, but KFc and KFkc are negative and decrease it (Eq 13). Addition of Cl- (dCl) increases the intracellular osmolarity, but it decreases the total intracellular electrical charge; so dCl is subtracted from total charge (Eq 23). The buildup of internal impermeant anion (bAn) is associated for electroneutrality with buildup of internal Na+, which must be multiplied by the mean valence of An- (-z, Eq 20).
The changes in the ionic concentrations obtained in the 2nd set of calculations can affect the intracellular osmolarity. Buildup of extra- or intracellular osmolytes, if defined by the investigator, will result in an additional imbalance in osmolarity. In the 3rd set of calculations, water is allowed to move across the cell membrane to restore osmotic equilibrium by changing the cell volume. The calculations can be performed with different transmembrane water permeability, i.e. with different rates of osmotically driven adjustment of the cell volume. This is an important difference of our program from previous models concerning volume regulation. The change in Em is found from the change in intracellular charge (dQ) divided by the cell membrane capacitance (c). The concentrations, Em and vol at the end of the current time step, but before the final volume adjustments, are marked with index e and after it with the index f.
Em,f=Em,e+dQ/c;
(24)
osV=([Na+]i,e+[K+]i,e+[Cl‑]ie+[An‑]i,e+[osm]i)/([Na+]o,e+[K+]o,e+[Cl‑]o,e+[osm]o);
(25)
VoR=1/tau;
(26)
chV=1‑(1‑osV)*VoR*st;
(27)
volf=vole*chV;
(28)
[K+]i,f=[K+]i,e/chV;
(29)
[Na+]i,f=[Na+]i,e/chV;
(30)
[Cl‑]i,f=[Cl‑]i,e/chV;
(31)
[An‑]i,f=[An‑]i,e/chV;
(32)
[osm]i,f=[osm]i,e/chV;
(33)
where tau is the time constant of exponential changes of the cell volume in response to a sudden change in osmolarity, VoR is a constant inversely related to tau, but VoR = 1/st if tau < st and VoR = 0 if tau > 108 sec (more than 3 years), osV is a coefficient of osmosis driven volume changes assuming instant transmembrane water movement, chV is a coefficient of osmosis driven volume changes corrected for limited water permeability of the membrane.
Notes: The calculations in this part are straightforward, but the section concerning the water permeability and the time constant of volume changes (Eqs 26 and 27) should be explained. Imbalance between intra- and extracellular osmolarity will induce transmembrane movement of water that changes the cell volume and restores osmotic equilibrium. If membrane water permeability is assumed to be infinite, the water shift and volume changes would happen instantly. In reality, the water permeability is high, but not infinite. It is also different in different cell types. The higher the water permeability, the faster the volume change. The water flux, and consequently, the speed of volume changes also depends on the osmotic gradient, and both the flux and the speed decrease with time as the gradient diminishes. This is similar to the well described discharge of a capacitor in a simple RC circuit. In our case the cell volume is analogous to the voltage of the RC circuit, and membrane water resistance (which is the inverse of water permeability)–to the electrical resistance. Thus, the dynamics of cell volume changes evoked by a sudden shift in osmolarity can be described, by analogy with an RC circuit, with the following equation of exponential decay:
vol(t)=vol1+(vol2–vol1)*(1–e−t/tau)
(34)
where vol(t) is the cell volume in time t, vol1 is the initial cell volume, vol2 is the final cell volume with osmolarity equilibrated, and tau is a time constant, which is inversely proportional to the membrane water permeability. Thus, tau is a convenient measure of water permeability, and if water permeability is such that the time constant of the volume change is 1 minute, it means that after 1 minute the cell experiences 63.2% of the expected volume change. Twofold smaller water permeability will correspond to twofold larger time constant of the volume changes, and the cell will need 2 minutes for 63.2% of the volume change.
Eq 25 determines the coefficient osV, by which the cell volume and consequently intracellular concentrations must be corrected in order to return to osmotic equilibrium. But due to limited water permeability another coefficient (chV), which is a fraction of osV and determined by Eq 27, is used later for correction of the cell volume and the intracellular concentrations. The smaller tau, the larger chV (and closer to osV). For practical purposes, in Eq 27 we used VoR, which is the inverse of tau. It permits a pair of convenient conditions. If tau determined by the investigator is 0 or just shorter than the time step (st), VoR is assumed to be equal to 1/st; as a result osV = chV, i.e. water instantly corrects the cell volume and osmolarity. If extremely low water permeability with tau > 108 sec is chosen, VoR is assumed to be equal to 0; as a result osV = 1, i.e. water does not move across the membrane at all as in simulation of Donnan equilibrium.
After completing the 3rd set of calculations, the program moves forward in time by one time step and repeats all the cycles.
Our modeled cell is assumed to have an effective volume of 7.5*10−13 L, equivalent to a cube 10x10x10 μm, with 25% of the volume occupied by organelles (like in rat rod photoreceptors [38]). All calculations are done with respect to this effective cell volume of water containing ions and other osmolytes. Since the surface area of a cube 10x10x10 μm is equal to 6*10−6 cm2, the total membrane capacitance of the cell is 1.2*10−11 F (given a specific capacitance for a neuronal membrane of 2 μF/cm2), and remains the same in all calculations. During volume changes cells usually alter their shape but keep the same surface area (for a recent reference see [39]). We specifically pointed out that our model cell is a cube, which permits an increase in cell volume by almost 40% by turning it into a sphere without any changes to surface area, and consequentially with no change in capacitance. During our simulations the volume increases were usually within this range, except in the catastrophic occasion of a swelling cell whose membrane was permeable to Na+, Cl- and water, but lacked Na+/K+-ATPase (Figs 1D and 2C). Nothing prevents the cell from keeping the same surface area when the volume decreases.
The initial conditions and manipulations of the examined system (concentrations, conductances, transporter activities) varied from simulation to simulation, and their exact values are listed in the figure legends. It should be noted that initial conditions that describe the starting point of the system before manipulations can refer to equilibrium states (Figs 1–5) or nonequilibrium but steady states with stable concentrations and Em (Figs 6–10). In the second case ionic conductances and the rate of the Na+/K+-ATPase were chosen for reasons explained in the text, but concentrations and Em that are given as initial conditions in the figure legends were the results of preparatory calculations that led to the state of the system at t = 0. Preparatory calculations were also necessary to determine initial conditions in some equilibrium states. One particular set of conditions includes 6 mOsm external impermeant osmolyte and a valence of -1.5 for internal impermeant anions, which exemplifies an osmolarity-charge asymmetry, i.e. conditions when the quantity of equivalent charge is not the same as the quantity of osmotically active molecules. Since such asymmetry is expected to be common in real cells, this set is called the “realistic” conditions and often will be compared to “simplified” conditions, which are symmetrical in osmolarity-charge respect, when the valence of internal impermeant anions is equal to -1 and there are no other external osmolytes besides NaCl and KCl.
The aim of this work is to examine the interactions between ionic concentrations, membrane potential, and cell volume in a complex system containing most of the components which are considered to be important for the matter. However, in order to determine the specific roles of those components, it is convenient to dissect the system into simpler subsystems. Accordingly, we will first look at systems based on Donnan (Figs 1 and 2) and Double Donnan (Fig 3) equilibrium which can keep the membrane potential stable (and different from 0) without energy expenditure. Then we will focus on a system in which the membrane is permeable only to cations (Na+ and K+) and contains the Na+/K+-ATPase, with the additional complexity of an uneven transmembrane distribution of neutral osmolytes and a deviation of the mean valence of the impermeant internal anions from -1 (Figs 4 and 5). The Cl- conductance (gCl) and cation-Cl- cotransporters (NKCC and KCC) will be added in the third group of simulations (Figs 6, 7 and 8). Finally, we will examine the changes in cell volume and the membrane potential which are associated with the buildup of osmolytes, internal and external, electrically neutral and charged, permeant and impermeant, that roughly simulate regulatory volume increases (RVI) (Figs 9 and 10).
The classical example of a system that generates a considerable transmembrane potential without spending energy is the one based on Donnan equilibrium. The conditions which can lead to Donnan equilibrium—unequal distribution of Cl- across the cell membrane and selective permeability of the membrane to cation(s) and to Cl-, but not to other intracellular anions–are typical for living cells, including neurons. Those impermeant “other intracellular anions” consist of a diverse group of large and small molecules which contribute noticeably to the voltage-volume regulation, and they are addressed specifically later. For the current simulation of Donnan equilibrium let us just assume that the impermeant anions ([An-]i) account for most of the internal anion concentration (135 mM) and have a mean valence = -1. At the beginning (before time = 0) the membrane is permeable to nothing, the membrane potential is 0 mV, and for the osmotic and electrical balance the extracellular Cl- concentration ([Cl-]o) is 150 mM, the intracellular Cl- concentration ([Cl-]i) is 15 mM, and cation concentrations (in this case Na+) are 150 mM, both inside and outside. The membrane remains impermeable to water, so the cell volume does not change.
The results of the simulations are presented in Fig 1A and 1B. At time = 0 the cell begins to be permeable to Cl- and to Na+. In the simulation, gNa and gCl are set to be equal. Cl- immediately starts to move into the cell because [Cl-]o is 10 times larger than [Cl-]i. Rapidly the influx of Cl- brings negative charge into the cell and hyperpolarizes the membrane. This creates an electrical driving force for Na+ influx, and the entering Na+ partly neutralizes intracellular negativity. As a result, Cl- continues to enter because its strong inward concentration-dependent driving force exceeds its outward electrical driving force. Na+ also continues to enter because its inward electrical driving force exceeds its weak outward concentration-dependent driving force. In 20 minutes both [Na+]i and [Cl-]i are increased by 81.8 mM and practically stabilized. At these concentrations (231.8 mM for [Na+]i and 96.8 mM for [Cl-]i) their Nernst potentials are both equal to the membrane potential Em = - 11.6 mV (Fig 1B), which means that the concentration-driven fluxes are counter-balanced by electrically-driven fluxes for both ions and the system has reached a stable equilibrium. This is Donnan equilibrium, and as expected [Na+]i * [Cl-]i = [Na+]o * [Cl-]o. Indeed, 231.8 * 96.8 = 22438.24 ≈ 22500 = 150 * 150. Additional time is required to achieve a more accurate fit between the results of the calculation and the Donnan expectations. With a chosen accuracy of 15 significant digits, the concentrations are finally stabilized after about 90 min at 231.986477689649 mM for [Na+]i and 96.9884116601443 mM for [Cl-]i; their product is 22499.9999977506.
Special calculations were not performed for validating the model but results like these (as well as some others that will be presented later) demonstrate the model’s validity. It is also worth mentioning that the initial change of Em, which looks instantaneous in Fig 1B, in reality decreases exponentially with a time constant of 3.7 ms (see S1 Fig), and this is exactly what is expected in our cell, which was set to have an input resistance of 312.5 MΩ and a membrane capacitance of 12 pF.
Increasing both [Na+]i and [Cl-]i, of course, increases the total intracellular osmolarity. In this example the intracellular osmolarity increased from 300 to 463.6 mOsm ([Na+]i = 231.8 mM, [Cl-]i = 96.8 mM, and [An-]i = 135 mM), creating a strong osmotic gradient. Thus, another fundamental condition that is necessary for Donnan equilibrium is the prevention of transmembrane movement of osmotically obliged water. Otherwise the water enters the cell following ions, increasing the cell volume and diluting ionic concentrations. Exactly that happened when the same simulations as for Fig 1A and 1B were repeated, but under the assumption that the membrane was highly water-permeable, and water instantly compensated for the potential osmotic imbalance associated with ionic transfer (Fig 1C and 1D). As in the previous simulation, opening of Cl- and Na+ conductances permits both ions to enter the cell (Cl- due to the concentration gradient, and Na+ because of the intracellular negativity created by the influx of Cl-). But water also enters, increasing cell volume and diluting intracellular concentrations. One of the effects of the dilution is a decrease of the impermeant anion concentration [An-]i. The other is that the [Na+]i remains the same and equal to [Na+]o because the increase in the amount of Na+ is precisely compensated by the increase in volume. As a result, the equilibrium potential for Na+ is 0, and Na+ will continue to enter the cell as long as the membrane potential stays negative. Cl- also will continue to enter (and [Cl-]i will continue to increase in spite of the dilution) because its equilibrium potential is twice as negative as the membrane potential. Speaking of the membrane potential, in a system of two unevenly distributed ions with different equilibrium potentials, the Em will obviously be somewhere in between those potentials. In this case, when the membrane is equally permeable to both Cl- and Na+, Em is the arithmetic mean of their Nernst potentials: Em = (ECl + ENa) / 2. Since ENa = 0, Em should be equal to half of ECl. Again, the simulations show exactly what is theoretically expected.
Fig 1C and 1D show that the addition of water permeability to a system which consists of permeable Na+ and Cl- and impermeant An- makes equilibrium unattainable. The equilibrium requires that Em = ENa = ECl, and since ENa = 0 it is only possible if [Cl-]i = [Cl-]o and accordingly [An-]i = 0. It should be remembered that in the simulation of Fig 1C and 1D the water permeability was assumed to be extremely large, permitting no osmotic imbalance. Although this idealization is not too far from reality and is usually accepted as true in computational models of volume regulation, it is still unrealistic. So Fig 2 shows a series of simulations with the same initial conditions as for Fig 1 but using various values of water permeability expressed for simplicity as a time constant of osmotic volume adjustment (see Methods). At a reasonable water permeability (time constant = 1 min, Fig 2A) the changes of ionic concentrations were similar to the data presented in Fig 1C (instant transmembrane water transfer), and at very low water permeability (time constant = 1 hour, Fig 2B) the concentrations were similar to the data presented in Fig 1A (no transmembrane water transfer at all). The larger the water permeability (shorter time constant), the faster the volume increase (Fig 2C). At a time constant of 1 sec the volume increase is indistinguishable from that under the assumption of instant water movement, and it is not much different at a realistic time constant of 1 min. Importantly, even with unrealistically low water permeability (time constant = 1 hour) the cell volume will increase slowly but steadily, theoretically, to infinity, and practically until the cell blows up.
Also, the smaller the water permeability (longer time constant), the larger the transmembrane osmotic gradient (Fig 2D). Bacterial and plant cells can counteract the osmotic pressure with hydrostatic pressure because they have rigid cell walls to preserve the cell volume. Animal cells have no such walls, and their ability to withstand osmotic pressure is limited. Thus, in animals, osmotically imbalanced transmembrane transfer of ions is inevitably associated with changes in cell volume. The time constant of volume adjustment of neuronal and glial cells is in the range of tens of seconds to a minute. For instance, the osmotically evoked volume increase of retinal Muller cells has a time constant of 0.5–1.0 minute [40]. The light-induced volume changes of vertebrate photoreceptors have approximately the same dynamics, judging by changes of ECS in the retina [13, 14].
For all other calculations in this work, simulations were done with a highly water permeable membrane (volume time constant = 1 sec) in order to focus on other aspects of voltage-volume regulation. This will affect the dynamics (although slightly, as indicated by the similarity of the 1 min and 1 sec curves in Fig 2C), but not the overall conclusions.
As we can see, Donnan equilibrium is not applicable for animal cells, including neurons. However, it is theoretically possible to achieve equilibrium in conditions described above if an impermeant osmolyte were added to the external solution to create a so-called Double Donnan system [41]. Let us assume that the extracellular solution contains 135 mM of an impermeant neutral osmolyte, the same concentration as the internal impermeant anion (An-). To keep the same external osmolarity, [Na+]o and [Cl-]o have to be reduced from 150 mM to 82.5 mM. Once again, after opening the gNa and gCl, Na+, Cl-, and water will enter the cell and volume will increase. Again, [Na+]i does not change, and [Cl-]i increases when [An-]i decreases, creating the illusion that An- is being replaced by Cl- (Fig 3A). And again, for equilibrium, both ENa and ECl should be equal to Em. But in this case it is possible because [Na+]o ≠ [Na+]i, and ENa = -15.96 mV. After 30 min [Cl-]i increases to 45.35 mM which corresponds to ECl = -15.98 mV, and the system is approaching equilibrium (Fig 3B). This is the Donnan equilibrium and when [Cl-]i reaches 45.375 mM, the equation [Na+]i * [Cl-]i = [Na+]o * [Cl-]o will be true. Of course, entering Na+, Cl-, and water will increase the cell volume, but it stabilizes at 129% of the initial value.
Thus, in a Double Donnan system equilibrium can be achieved without compromising osmotic balance if an impermeant external neutral osmolyte is present in a sufficient concentration. The problem is that in order to keep [Cl-]i low (and make osmotic room for important internal anions like proteins and nucleic acids) the concentration of the external neutral osmolyte must be high. However, in reality the total concentration of all external impermeant neutral osmolytes is quite low. Glucose is by far the most concentrated neutral osmolyte in ECS (5–6 mM). Numerous others have concentrations of small fractions of mM, and the total concentration of all external impermeant neutral osmolytes in normal conditions hardly ever exceeds 10 mOsm. But with only 10 mOsm of the neutral osmolyte, [Na+]o and [Cl-]o will be 145 mM, so according to the Donnan equation [Cl-]i will be 140.17 mM, leaving almost no osmotic room (less than 10 mM) for other internal anions including vitally important proteins and nucleic acids.
The most concentrated substance outside of the cell is Na+. If the membrane were impermeant to Na+, the Double Donnan system could be created by utilizing Cl- and another permeant cation, K+ (Fig 3C and 3D). Initial concentrations for these simulations are: [Na+]o = [Na+]i = 145 mM, [K+]o = [K+]i = 5 mM, [Cl-]o = 150 mM, [Cl-]i = 15 mM, and [An-]i = 135 mM. When gCl and gK are open, both ions will enter the cell as in previous simulations. Water will follow and increase the cell volume and decrease the concentration of impermeant substances, which are [Na+]i and [An-]i in this case (Fig 3C). After approximately 5 minutes, the equilibrium potential for K+ decreases and the equilibrium potential for Cl- increases to the same level as membrane potential, EK = ECl = Em, (Fig 3D), and equilibrium is reached. The equilibrium potential for Na+ is different from the membrane potential, but it has no consequence since we assumed that the membrane is not permeable to Na+. Needless to say, this is purely theoretical and an absolutely unrealistic case, because every neuron has a significant Na+ conductance, and the presence of even the smallest Na+ conductance makes this equilibrium unachievable (see S2 Fig).
Concluding this part, it should be noted that the equilibrium conditions described above were determined completely by concentrations of the ions involved. The values of their conductances influence the time to achieve equilibrium but have no effects on the equilibrium potentials. Accordingly, when equilibrium is reached it cannot be changed by alterations of conductances.
The simulations in this section deal with two cations, Na+ and K+, and consequences of their nonequilibrium distribution across the cell membrane due to activity of the Na+/K+-ATPase. It has been recognized for more than a half century that “the big triad” of Na+ conductance, K+ conductance and Na+/K+-ATPase not only determines the membrane potential but forms the basis of the whole system of ionic homeostasis. Accordingly, the previous models concerning cell volume regulation starting from the early works [41–43] and up to the most recent [23–25] paid considerable attention to the cation conductances and Na+/K+-ATPase. Nevertheless, some features of the big triad were overlooked, and some others were misinterpreted.
The transmembrane redistribution of Na+ and K+ by the Na+/K+-ATPase is illustrated in Fig 4A. Before time = 0, [Na+]o = [Na+]i = 145 mM and [K+]o = [K+]i = 5 mM. Also 150 mM of an impermeant monovalent anion was present both inside and outside of the cell for electrical and osmotic balance. Since in these calculations the membrane is permeable to nothing but Na+ and K+, it does not matter what kind of anion it is (Cl- or something else), but it is important that this anion is monovalent to preserve osmolarity-charge symmetry. In an attempt to make the cation transfer electrically neutral, an imaginary electroneutral Na+/K+-ATPase, which exchanges 3 Na+ for 3 K+ was used in this simulation. Also, gNa was equal to gK. At time = 0 the pump starts to transfer Na+ out of the cell and K+ into the cell against growing concentration gradients for both cations. Na+ starts to leak back to the cell and K+ - out of it, and with changes of the cation intracellular concentrations these leaks increase. At the same time the active transport of Na+ and K+ by the pump slows down because it strongly depends on [Na+]i (see Methods), which is decreasing. The pump rate was set high (24 billion transfers/sec) and the pump activity with initially high [Na+]i was 20.4 billion transfers/sec. (For explanation of our definitions of the pump rate and the pump activity see Methods). The pump activity decreased to 3 billion transfers/sec in about 1.5 seconds when [Na+]i decreased to 8 mM and after about 5 seconds it stabilized at 331.4 million transfers/sec when [Na+]i decreased to 2.52 mM. At this time the active transport of the pump and passive leaks of Na+ and K+ were equal, and steady state (when the concentrations, and consequently Em, remain the same) was achieved. It is important to emphasize that this is not an equilibrium state like Donnan equilibrium, because both ENa and EK are different from Em, and this stable state must be supported by constant energy expenditure.
Interestingly, in spite of the effort to make transport electrically neutral, Em also changed, first decreasing to -27.35 mV and then increasing and stabilizing at +8.90 mV. The reason for Em changes in a system with seemingly equal exchange of Na+ for K+ is the asymmetrical effect of the same absolute changes in [Na+]i and [K+]i on their respective equilibrium potentials. Indeed, when [K+]i increases by 5 mM (from 5 to 10 mM), [K+]i is doubled and ΔEK = -18.5 mV; the simultaneous decrease of [Na+]i by 5 mM (from 145 to 140 mM) means a relative change of only 3.4% and ΔENa < 1mV. The different time courses of EK and ENa (Fig 4A) illustrate this asymmetry, contrasting with symmetrical changes of [Na+]i and [K+]i.
In the model, the activity of the pump is conveniently expressed in cycles per second and, accordingly, in ATP spent per second. So, we can directly connect the energy spent for the active transport of Na+ and K+ with the final intracellular concentrations of these ions in steady state and the resulting resting Em. The steady state values of concentrations and Em that are reached after a few seconds in the simulation of Fig 4A were the result of a particular rate of steady state ATP utilization (331.4 million ATPs/sec). Additional simulations with larger or smaller initial pumping rates were done and yielded different sets of values of steady state concentrations and Em. This allowed the creation of the graph (Fig 4B) on which steady state values of [Na+]i, [K+]i and Em were plotted against an energy cost expressed in ATP spent per second. In subsequent figures lines through symbols will be used for graphs of this type; each set of points associated with a certain abscissa value represents a separate simulation. For instance, the concentrations and Em from simulations presented in Fig 4A are included in Fig 4B and marked by the arrows.
As more energy is spent, stronger electro-chemical gradients for Na+ and K+ are created. But as shown above, the stronger electro-chemical cation gradients do not necessary mean more negative Em. With the pump activity at 149.8 million ATP/sec, [K+]i = [Na+]i = 75 mM and Em reached its most negative value (-27.35 mV) for this condition. With the pump activity at 299.7 million ATP/sec, the cation concentrations are exactly reversed from the extracellular values ([K+]i = 145 mM and [Na+]i = 5 mM); accordingly, ENa = -EK, and Em = 0 mV. We regard the expenditure of ATP on the rising side of the U-shaped Em curve to be wasteful, because the same Em could be achieved at a lower ATP cost at a point on the falling side of the Em curve. Thus, we call the range of ATP utilization above the point of minimum Em “overpumping.”
In reality, the Na+/K+-ATPase is of course electrogenic, transferring 3 Na+ for 2 K+, and that directly influences Em (Fig 4C). Simulations with all the same conditions as previously, but with an electrogenic 3Na+/ 2K+ pump, show that Em is equal not to 0, but to -17.98 mV at the point of the reversal of the cation concentrations. To reach this point, activity of the pump must be 359.6 million ATP/sec, i.e. the pump generates a current of 57.5 pA (3.596*108 multiplied by the charge of one cation, which is 1.6*10−19 coulomb). Multiplication of this current by the input resistance of the modeled cell (312.5 MΩ) gives us the same voltage (-17.98 mV) that was calculated by the model using only the transmembrane movements of mass and charge. ENa and EK calculated from [Na+]i and [K+]i that were determined in simulations for Fig 4C, the voltage generated by the pump (Epump), and the final Em, are plotted against the pump activity in Fig 4D. To calculate Em (under the modeled condition of equal gNa and gK) we used the equation:
Em=(ENa+EK)/2+Epump
This analytically calculated Em is equal to the Em calculated by the model with precision better than 3*10−5 mV. Since the program calculates Em from the total intracellular electrical charge and the membrane capacitance (see Methods, Eq24) this fit provides another validation of the program.
Obviously, the cost of Na+ and K+ electro-chemical gradients and the resulting Em depends on the intensity of cation leakage. The cell with half the gNa and gK (input resistance = 625 MΩ instead of our usual 312.5 MΩ) will spend half the energy for the same gradients and voltage, and this is true for both an electroneutral and electrogenic pump (see S3 Fig).
The next step toward a more realistic system is considering the fact that the mean valence of impermeant anions is probably never equal to -1. The value of the mean valence of impermeant anions is difficult to determine experimentally; it probably varies in different cell types and possibly also in the same cells in different conditions. It also can be defined differently (more on this in Discussion). In this work we define the mean valence of the impermeant anions as the total charge of all impermeant intracellular anions divided by their total osmolarity. An exception will be made for Cl-, which is only temporarily, in this set of simulations, assumed to be impermeant; so, Cl- is counted separately from the impermeant anions. For the simulations presented in Fig 4E we assumed that the mean valence = -3 to show what is probably the largest effect. When the mean valence of impermeant anions is larger than -1 (here and later, when we say “larger” in respect to mean valence of an anion we refer to the absolute value, ignoring the sign), fewer anion molecules are necessary to electrically compensate the cations. For instance, 150 mM of monovalent cations from the previous simulation can be neutralized by 45 mM of anions with valence = -3 in addition to 15 mM of Cl-. However, this would result in an intracellular osmolarity of only 210 mOsm/L. Then, to keep internal and external osmolarity equal, all intracellular concentrations should be proportionally increased by a ratio of 300/210, so the conditions for Fig 4E when the pump rate is zero are [Na+]i = 207.14 mM, [K+]i = 7.14 mM, [Cl-]i = 21.43 mM and [An-]i = 64.29 mM. Since intracellular concentrations of Na+ and K+ are higher than their respective extracellular concentrations, the membrane is hyperpolarized to -9.5 mV; this is an equilibrium state because Em = ENa = EK, and support of those unequal transmembrane distributions of Na+ and K+ as well as negativity of Em does not cost any energy.
It is notable, comparing the results of calculations presented in Fig 4E (when the mean valence of the impermeant anions was = -3) with that in Fig 4C (when the mean valence of impermeant anions was = -1), that increasing the mean valence of impermeant anions shifts down the whole curve of Em by the same value (-9.5 mV in this case) regardless of the pump activity. In both cases the most negative Em was achieved at the pump activity of 201.5 million ATP per second. Also, in both cases the same amount of energy (359.6 million ATP per second) is necessary to reverse intracellular Na+ and K+ concentrations, although those reversed concentrations were different (see Fig 4C and 4E).
Fig 4F shows that similar effects were found in a condition when 15 mM of neutral impermeant osmolyte was added externally (the mean valence of intracellular impermeant anions was = -1 in this simulation). This increase of external osmolarity by 5% requires a proportional increase of all intracellular concentrations (including Na+ and K+), which in turn leads to generation of a small ENa = EK = Em = -1.3 mV in the equilibrium state with no pump activity, as well as a downward shift of the Em curve by -1.3 mV throughout the whole range of pump activity. Again, the same amount of energy as in previous simulations was needed to reach the critical point of the most negative Em and reversed [Na+]i and [K+]i. (see Fig 4C, 4E and 4F).
In all simulations of this part so far, Na+ and K+ have had the same conductance of 1*1010 ions/(sec*V), which is equal to 1.6 nanosiemens, resulting in a realistic input resistance of our modeled cell (312.5 MΩ). But it is typical for neurons that their gK is several times larger than gNa. For next set of simulations gK = 1.8*1010 and gNa = 2*109 ions/(sec*V), so the input resistance remains the same, but the ratio gK : gNa is 9 : 1. The results of calculations performed with “simplified” conditions (mean valence of internal impermeant anions = -1, concentration of external impermeant osmolyte = 0) are presented in Fig 5A. They are clearly different from the results of calculations performed with the same conditions, but with gK = gNa (Fig 4C). First, as expected, Em is significantly more negative. According to the chord conductance equation [44]
Em=(gK*EK+gNa*ENa)/(gK+gNa)
an increase of the K+ contribution makes Em more negative. Interestingly, even in this case Em can be still “overpumped” i.e. the negativity of Em diminishes when the pump activity grows too large. Presumably there is little or no advantage in spending this energy for the pump when it leads to a smaller value of Em. Second, a high gK : gNa ratio enables the pump to spend less energy for creating ionic gradients. For instance, 179.8 million ATP per second is needed to equilibrate [Na+]i and [K+]i (both equal to 75 mM) when gK = gNa (Fig 4C), but only about 55 million ATP per second is sufficient when gK : gNa = 9:1 (Fig 5A).
Repeating these calculations with “realistic” conditions (mean valence of internal impermeant anions = -1.5, concentration of external electrically neutral impermeant osmolyte = 6 mM) gives results presented in Fig 5B. After necessary osmotic adjustments, [Na+]i and [K+]i in the equilibrium stage (pump activity = 0) increased to 174 and 6 mM, respectively, and the membrane hyperpolarized to Em = -4.87 mV. Accordingly, compared to “simplified” conditions (Fig 5A), the range of cation concentration changes is wider and the whole Em curve is shifted down by -4.87 mV. The terms “simplified” and “realistic” appear in quotation marks as a reminder that we use them only with respect to different concentration configurations that lead to symmetrical (convenient in calculations) or asymmetrical (usual in nature) osmolarity-charge configurations, respectively. Here both conditions are tested in a purely theoretical case where there is no Cl- conductance, and later they will be applied to much more real situations with Cl- conductance present and intracellular Cl- concentration affected by cation-Cl- cotransporters.
With many differences described above, all graphs of this part have one thing in common–the changes of [Na+]i were always mirrored by the changes of [K+]i, i.e. all removed Na+ was replaced with an equal quantity of K+. This is true even if the only job of ATPase was to remove Na+ (3Na+ per ATP in this simulation) without transferring any K+. As soon as both [Na+]i and Em decrease due to the electrogenic 3Na+-ATPase, Na+ begins to leak back to the cell and K+ also enters the cell attracted by the negativity. After some time a steady state will be established when K+ will be in equilibrium (EK = Em) and the passive leak of Na+ into the cell will be equal to the active Na+ pumping. The larger the activity of the pump, the stronger the electro-chemical Na+ gradient, and the more Na+ is replaced with K+ in the cell, as illustrated by Fig 5C (“simplified” conditions) and 5D (“realistic” conditions). Here all energy was spent for the Na+ gradient; K+ was distributed passively. But still K+ plays key role - replacing Na+ it makes possible creating the Na+ gradient. If gK = 0, i.e. K+ cannot enter the cell, [Na+]i would remain practically the same regardless of the activity of a pump moving only Na+. In this case all energy of the pump will be spent on generating a negative Em, which drives back all Na+ that was actively removed.
For better comparison of the effects that different parameters of the modelled cell have on the relation between Na+/K+-ATPase activity and Em, the results of several calculations are presented together in Fig 5E. A pair of simulation conditions—“simplified” and “realistic”–was used in each of four general settings. In Setting 1 an electrogenic 3Na+/2K+ pump was used and gK = gNa with the input resistance = 312.5 MΩ (as in Fig 4C). Other settings were different from Setting 1 in one of the following respects - imaginary electroneutral 3Na+/3K+ pump in Setting 2 (as in Fig 4B), ratio gK : gNa = 9:1 in Setting 3 (as in Fig 5A and 5B), and imaginary only Na+ pump (3Na+ per ATP) in Setting 4 (as in Fig 5C and 5D). Thus, it is proper to compare Setting 1 with the three others. As expected, an electroneutral 3Na+/3K+ pump (Setting 2) had a smaller effect on Em than an electrogenic 3Na+/2K+ pump (Setting 1); what was interesting is that the difference was not large before the pump activity reached the level of overpumping. From the energy point of view the most interesting result comes from comparing Setting 1 with Setting 3 - high gK relative to gNa enabled the pump to create stronger ionic gradients and promote a more negative Em while spending less ATP. A strongly electrogenic 3Na+ pump (Setting 4) significantly hyperpolarized the membrane at high pump activities, but it came with a high ATP price if gK = gNa. It is also notable that under every setting and with every pump activity the difference between the “simplified” and “realistic” conditions was always the same and equal to what was it was in the passive, no pump condition, i.e.-4.87 mV.
As we already pointed out, in all simulations in this section the internal Na+ was replaced with an almost exactly equal quantity of the external K+ regardless of relative conductance of these ions and the stoichiometry of the pump. Here we would like to emphasize the word “almost,” since the exchange of Na+ for K+ was not exact. The negative Em means that there is some deficiency of internal cations. Similarly, when ions were overpumped sufficiently for Em to become positive, it is due to some surplus of internal cations. Subtraction or addition of an intracellular substance leads to osmotic imbalance, compensatory water movement, and consequently, appropriate cell volume changes. The volume changes associated with Em changes from Fig 5E are presented in Fig 5F. As expected, dependence of the cell volume on the pump activity closely followed the dependence of Em on the pump activity in every setting. Since all volumes were normalized to the initial cell volume in the passive state with no pumping (defined as 100%), the pre-existing cation deficiency associated with initial negativity of Em in the “realistic” conditions (i.e. the -4.87 mV) was already counted; so, the curves for the volume changes in the “simplified” and “realistic” conditions are identical in all four settings.
The main result related to the cell volume, however, is that its changes are extremely, unnoticeably small. A tiny quantity of ions is needed to recharge the cell membrane significantly. A simple calculation shows that 9.6*10−13 coulomb of charge will hyperpolarize our modelled cell, with membrane capacitance 1.2*10−11 farad, to -80 mV. This charge is carried by 6*106 ions, which equals 13.3 μM in our cell with volume 7.5*10−13 L, i.e. only 0.0044% of the total internal osmolarity. The results of simulations presented in Fig 5C and 5D demonstrate precisely that: when the cell was hyperpolarized to about -80 mV (setting 3), the cell volume change was a bit more than 0.004%. Thus, we can conclude that the Na+/K+ pump replaces internal Na+ with practically the same quantity of K+. Accordingly, the pump has practically no measurable effect on the cell volume, and this is true with any stoichiometry of the pump and conductance of Na+ and K+.
To summarize this part, we can conclude that in the “only cation” system tested above three properties - Na+ conductance, K+ conductance and Na+/K+-pump activity - determine with certainty the values of three features - [Na+]i, [K+]i and Em. When the properties are constant, the determined features also stay unchanged in steady or resting state. But it is not an equilibrium state, and a constant expenditure of energy is required to keep it. Changes in the pump activity lead to changes in [Na+]i, [K+]i and Em until a new steady state is reached. The same is true for changes of Na+ or K+ conductances. It is also important to note that changes in the pump activity as well as changes in cation conductances have no practical effects on the cell volume.
In this part, the cation system described above will be enriched by addition of Cl- conductance and later by Cl--cation cotransporters. Cl- is by far the most concentrated external anion and numerous Cl- permeable channels and Cl- transferring transporters make this ion unavoidably important for the nervous system. Here we will show that Cl- conductance is the reason for voltage-related cell volume changes.
The results presented in Fig 6 illustrate the changes in concentrations, voltage and cell volume when Cl- conductance was opened at time = 0 to disturb the resting state achieved with open cation channels and active Na+/K+-ATPase (gNa = 8*109 ions/(sec*V), gK = 1.2*1010 ions/(sec*V), pump activity 2.64*108 ATP/sec, Em = -43.3 mV). This resting state, with a relatively high gNa and a moderate Em, was chosen so that changes in gCl could potentially cause either depolarization or hyperpolarization. As a result of this disturbance [Cl-]i changed significantly, but [Na+]i, and [K+]i, changes were barely noticeable (Fig 6A) and only transient (Fig 6B). Regardless of the initial [Cl-]i, (15 mM for solid and dashed lines, and 45 mM for the dotted line) or the value of gCl (1011 ions/(sec*V) for the solid line and 1010 ions/(sec*V) for the dashed and dotted lines), [Cl-]i in the new resting state was 29.6 mM. At this concentration ECl = -43.3 mV, i.e. ECl is equal to the Em established by the cations. To electrically compensate transmembrane movement of Cl-, Na+ and K+ move together with it, entering the cell when [Cl-]i increases from 15 to 29.6 mM and leaving the cell when [Cl-]i decreases from 45 to 29.6 mM. These ionic movements cause the cell volume to increase in the former case and decrease in the latter case (Fig 6A) keeping cation concentrations constant. The intracellular concentration of impermeant anion An- experiences opposite changes to Cl- as a result of the volume change, keeping [Cl-]i + [An-]i almost constant. The small deviations of total anion concentration (~0.25 mM and ~0.55 mM with gCl of 1010 ions/(sec*V) and 1011 ions/(sec*V), respectively) from the initial value of 150 mM that peaked in the first 2–3 seconds after increasing gCl was a result of restricted transmembrane water movements assumed in this modeling. There was no such deviation if we assumed instant water transfer.
Some small changes of [Na+]i and [K+]i, as well as of Em, also happened after opening of Cl- conductance, but, in contrast to anion concentrations and volume, both cation concentrations and Em recovered to their original values in 5–6 minutes when the system reached the new steady state (Fig 6B). In the case of low initial [Cl-]i (15 mM, dashed lines) Em was temporary hyperpolarized, Cl- and both cations entered the cell, and the cell volume increased. Initially Na+ entered the cell faster than the volume increase and [Na+]i slightly increased; at the same time [K+]i decreased (in spite of the influx of K+) and the sum of [Na+]i and [K+]i remained practically equal to 150 mM, with small deviations in the first seconds due to limited water transfer, as in the case of the anions. The transient increase of [Na+]i and the decrease of [K+]i are due to the stronger diluting effect of increasing volume on the ion in higher concentration, which is K+. Opposite transient changes in Em, [Na+]i and [K+]i occurred if [Cl-]i started at 45 mM (Fig 6B, dotted lines). If the steady state condition before opening gCl had been set to achieve [Na+]i = [K+]i = 75 mM (at a pump activity 1.66*108 ATP/sec), opening of gCl would have led to a decrease of [Na+]i and an increase of [K+]i, because K+ conductance in these simulations was set to be larger than Na+ conductance. If the initial steady state condition had been [Na+]i = [K+]i = 75 mM and gNa = gK (at the pump activity 1.80*108 ATP/sec), opening of gCl would have had no effect on the cation concentrations, besides small and short-lived deviations (both increases) related to delayed water movements, and it would have been the same for electrogenic and neutral pumps. gCl affected nothing except the time necessary for equilibration; it took 6–7 minutes with gCl = 1010 ions/(sec*V) (dashed and dotted lines in Fig 6) and 3–4 minutes with gCl = 1011 ions/(sec*V) (solid lines in Fig 6).
Calculations presented in Fig 6, parts A and B were done in “simplified” conditions, i.e. assuming that the mean valence of internal impermeant anions = -1 and that there were no external osmolytes besides Na+, K+ and Cl-. Largely similar results were obtained in “realistic” conditions (the mean valence of internal impermeant anions = -1.5 and the concentration of external electrically neutral impermeant osmolytes = 6 mM) with the most notable difference being in the cation concentrations (Fig 6C and 6D). Smaller amounts of polyvalent intracellular anions were needed to electrically compensate intracellular cations and that, together with the addition of impermeant extracellular osmolyte, demanded certain osmotic adjustments that affected the initial ionic concentrations. As a result, the resting [Na+]i and [K+]i were different from those in “simplified” conditions described above with the same rate of the Na+/K+-ATPase. Moreover, [Na+]i and especially [K+]i were different depending on how much of the initial anion concentration was due to [Cl-]i (10% and 30% of the total charge of intracellular anions for the dashed and dotted lines, respectively). Accordingly, Em also was different under low or high initial [Cl-]i, although the difference was only ~1 mV (Fig 6D). In this “realistic” condition, just as in the previously described “simplified” condition, opening of gCl led to changes of [Cl-]i toward a new value that was the same regardless of initial [Cl-]i. Again, the transmembrane movement of Cl- was accompanied by co-directed movements of Na+ and K+ that led to appropriate changes in the cell volume and consequently [An-]i (Fig 6C). However, the absolute change of [An-]i was 1.5 times smaller than the change of [Cl-]i, since each An- was carrying 1.5 times more charge. In this condition, [Na+]i and [K+]i, and accordingly Em, were also shifted to new levels (Fig 6D). When a new resting state was established, Em and all four cation and anion concentrations stabilized at new values which were not dependent on initial [Cl-]i. And the new [Cl-]i was again exactly what was required to make ECl = Em (-46.15 mV).
Since Cl- is in equilibrium in this new resting state, alterations in gCl cannot change anything in the system. But alterations in conductances of nonequilibrated cations can, and some of the results are different depending on gCl. Fig 7 illustrates how temporal changes of gNa affect the ionic concentrations, cell volume and voltage under various gCl. The cell modelled here is permeable to Na+, K+, and Cl- and is at rest under “realistic” conditions. So, it is similar to the one presented in Fig 6C and 6D, but with one difference in Fig 7A and 7B: it uses the imaginary electroneutral 3Na+/3K+-ATPase that does not transfer any net charge or mass and consequentially cannot directly influence the cell volume or voltage.
At time = 0 gNa was reduced by a factor of 4 (from 8*109 to 2*109 ions/(sec*V)) and 10 seconds later gNa returned to its original value. The temporary decrease of gNa leads to a hyperpolarization of Em, moving it closer to EK. This reduces the driving force for K+, and passive K+ efflux decreases. Passive Na+ influx also decreases (in spite of the increased driving force for Na+) due to the reduction in gNa. But the Na+/K+-ATPase continues to pump K+ in and Na+ out of the cell initially with the same activity. As a result, [K+]i increases and [Na+]i decreases during the temporary decrease of gNa. These cation changes are almost identical when gCl is negligible (108 ions/(sec*V), i.e. less than 0.5% of the total transmembrane conductance, dashed lines) or considerable (1010 ions/(sec*V), i.e. more than 30% of the total transmembrane conductance, solid lines). Em is more sensitive to gCl (Fig 7B), demonstrating the well-known “shunting inhibition” (see S2 Text), supported in many neurons by the Cl- permeable GABA- and glycine-gated channels. But it is the cell volume that is affected most by the value of Cl- conductance. The hyperpolarization evokes efflux of Cl-, which was in equilibrium before the time = 0. In order to electrically compensate it, an efflux of cations is required. The model shows that under these conditions Na+ influx decreases more than K+ efflux, causing a net efflux of cations that is equal to Cl- efflux, and when the ions leave the cell, the cell volume decreases. Depending on the value of gCl, the ionic fluxes could be large or small, determining the size of volume changes. Thus, during electrical activity associated with changes of cation concentrations the cell may or may not experience detectable changes of its volume, depending on the value of gCl.
It should be noted that the decrease of [Na+]i slows down the Na+/K+-ATPase. The simulations in Fig 7A and 7B used an electroneutral 3Na+/3K+-ATPase, and alterations in the pump activity had no consequences for the cell volume and voltage. But the real electrogenic 3Na+/2K+-ATPase does transfer both charge and mass, and it is intuitively expected that the electrogenic pump should contribute to changes of Em and the cell volume. The results of modeling with an electrogenic Na+/K+ pump are presented in Fig 7C and 7D. As expected, decrease of the pump activity associated with the decrease in [Na+]i clearly manifested itself in a slow reduction of the hyperpolarization (Fig 7D). However, the volume changes were again completely under control of the gCl. Independently of the stoichiometry of the pump, changes in its activity cannot change the cell volume if the membrane is not permeable to Cl-, and on the other hand, when Cl- conductance is considerable the volume changes happen irrespective of whether the pump is electroneutral or electrogenic.
One more point should be made concerning the relation between the Na+/K+-ATPase and the cell volume. Activity of the pump, and consequently the expenditure of ATP, follows [Na+]i. Since the changes of [Na+]i were practically identical with high and low gCl (Fig 7A and 7C), the expenditure of ATP was the same and irrelevant to the volume changes. This also can be seen in the simulations of Fig 6. Changes of [Na+]i tell us that some extra energy was spent during the transition when the cell volume increased (Fig 6 dashed lines) and some energy was saved when the cell volume decreased (Fig 6 dotted lines), but after reaching the resting state, the cell spent exactly the same amount of energy to keep the larger volume as to keep the smaller volume. And it also was equal to the amount of energy the cell spent before gCl opening, precisely in “simplified” conditions and with precision of a fraction of 1% in “realistic” conditions.
Certain molecular mechanisms can influence [Cl-]i, shifting it away from equilibrium, so that ECl ≠ Em. The most important of these for the nervous system are two cation-Cl- cotransporters - the Na+,K+,2Cl- cotransporter and the K+,Cl- cotransporter (NKCC and KCC, respectively). Fig 8A shows steady state values of intracellular ion concentrations, Em, and cell volume as a function of the rate of the NKCC, in the condition when Cl- conductance is very small (108 ions/(sec*V)). The lower row of numbers under the x-axis represents the corresponding activity of NKCC, which gives information on the actual quantity of ions transferred across the membrane. As for the Na+/K+-ATPase, the activity, in distinction to the rate, is dependent on ionic concentrations and will change together with them. The activity of the transporter is k1 times its rate (see Methods), where
k1=log10(([Na+]o*[K+]o*[Cl‑]o2)/([Na+]i*[K+]i*[Cl‑]i2)).
(35)
The NKCC pumps Na+, K+, and Cl- into the cell, and the direct result of that is an increase in cell volume. The higher the cotransporter rate (and activity), the larger the cell volume (Fig 8A). [Cl-]i also increases with the cotransporter rate, but, interestingly, [Na+]i and [K+]i remain almost exactly the same (note: there is a 3:2 Na+/K+-ATPase in these simulations). [Na+]i increased by only 0.006 mM and [K+]i actually decreased by 0.005 mM. Em also changed very little–from -43.3 mV to -43.2 mV, reflecting a subtle depolarizing influence of Cl-, which is not in equilibrium in this case. At the highest activity of the cotransporter, [Cl-]i reached 82.98 mM; at this concentration ECl = -15.8 mV. But the conductance of Cl- in this simulation was very low, so the Cl- contribution to Em is negligible.
Fig 8A also clearly demonstrates that the capability of NKCC to elevate [Cl-]i (and the cell volume) is limited. When [Cl-]i is increasing, kl approaches 0. Because the cation concentrations are constant (Fig 8A) and so is [Cl-]o, this limiting value of [Cl-]i can be obtained by setting kl = 0 and solving the Eq 35 for [Cl-]i. In our calculations [Na+]o = 145 mM, [K+]o = 5 mM, [Na+]i = 17.9 mM, [K+]i = 132.1 mM, and [Cl-]o = 150 mM, so the largest [Cl-]i that can be achieved is 83.06 mM. [Cl-]i approaches this level when the cotransporter rate is 108 cycles/sec and activity = 1.35 million cycles/sec. At that point the driving force of the transporter is almost exhausted, and increasing its rate by 100 times means increasing activity only to 1.37 million cycles/sec, i.e. only by 1.5%.
Fig 8B shows how the activity of NKCC influences concentrations, voltage and volume when Cl- conductance is high (1010 ions/(sec*V)). In such conditions the cotransporter is similarly capable of elevating [Cl-]i and increasing the cell volume to about the same values as in the case of low gCl, but the activity of the cotransporter has to be roughly 100 times higher because it has to overcome a leakage of Cl- that is 100 times larger through the high conductance. More importantly, the high gCl makes Cl- a noticeable contributor to Em. Thus, the cotransporter-generated increase of [Cl-]i is accompanied not only by an increase in cell volume, but also by a depolarization from -43.3 mV to -34.6 mV. [Na+]i and [K+]i were again almost unaffected, although their small changes were larger than in the case of low gCl: +0.42 mM for [Na+]i and -0.42 mM for [K+]i.
The KCC uses the strong K+ outward concentration gradient to extract Cl- from the cell. The coefficient k2 that links the activity of the cotransporter to its rate is expressed as follow:
k2=log(([K+]o*[Cl‑]o)/([K+]i*[Cl‑]i)).
(36)
According to this equation, when k2 = 0 the lowest [Cl-]i which can possibly be achieved in our conditions ([K+]o = 5 mM, [Cl-]o = 150 mM, [K+]i = 132.1 mM) is 5.68 mM, and the calculations show that our modeled cell approaches this limit with a cotransporter activity of 4.13 million cycles/sec when gCl is low (Fig 8C). Together with lowering of [Cl-]i, KCC decreased the cell volume, but the cation concentrations remained remarkably similar (only +0.040 mM for [Na+]i and -0.037 mM for [K+]i), in spite of the fact that the cotransporter removed exactly the same amount of K+ as Cl-. Em also was very little affected - the cotransporter at its maximal activity produced only -0.27 mV of additional hyperpolarization.
As expected, increasing gCl 100 times demanded much higher activity of the cotransporter for lowering of [Cl-]i toward the limit (Fig 8D). Also, as expected for a high gCl, the decrease of [Cl-]i caused by KCC was accompanied by significant hyperpolarization (from -43.3 mV to -59.7 mV). The cation concentrations were affected as well, although not as much as [Cl-]i: [K+]i decreased by 2.6 mM and [Na+]i increased by the same 2.6 mM.
It is also noteworthy that activity of both cation-Cl- cotransporters is associated with increased expenditure of energy. Moving one more element of the system (Cl-) out of the equilibrium state obviously should cost some extra energy, regardless of the direction of this movement - an increase or decrease of [Cl-]i and, consequently, an increase or decrease of the cell volume and depolarization or hyperpolarization of the cell membrane. In this respect, it is surprising how little extra energy was needed in the case of the NKCC. When [Cl-]i increased by 177% and cell volume increased by 77% with a very active cotransporter, ATP consumption increased only by 2.3%. And this was under the high gCl condition. When gCl was low, even larger increases of [Cl-]i and the cell volume were achieved with a tiny cost of 86,000 extra ATPs per second, which is 0.03% of the total energy. The KCC is more demanding. A decrease of [Cl-]i to 19.7% of its initial concentration with a cell volume reduction to 83.5% required an additional 12% of ATP when gCl was high, and an extra 2% of ATP did comparable work when gCl was low.
All simulations of cation-Cl- cotransporters above were done in the “simplified” condition. We performed the same series of calculations in the “realistic” condition and high gCl for NKCC (Fig 8E) and for KCC (Fig 8F). Changes of [Cl-]i in an asymmetric concentration-charge system, like the “realistic” condition, is associated with some additional complications in cation concentrations. The [Na+]i + [K+]i is not constant anymore; contrarily, the sum must change due to changes in the ratio “total anion charge”/”total anion concentration” resulting from opposite changes in the concentrations of monovalent Cl- and polyvalent An- (more on this in Discussion). As a result, the logic of cation concentration behavior induced by cation-Cl- cotransporters is not obvious. Specifically, NKCC noticeably decreased [K+]i, but did not change [Na+]i, in spite of pumping both cations into the cell. In its turn, KCC increased [Na+]i, although it did not transfer this ion; the cotransporter also had a biphasic increase-decrease effect on [K+]i. The movement of water and changes in Em are important in understanding these unintuitive changes.
Effects of cation-Cl- cotransporters on [Na+]i and [K+]i are intriguing and deserve a more detailed analysis in the future, but for the purpose of this paper it should be stressed that those effects were small. In most cases there were almost no cation changes compared with changes of Cl-, which was transferred simultaneously with cations and in equal amount. And [Cl-]i changes were always accompanied by changes of the same sign in the cell volume. Low gCl allowed to the system to achieve large effects on [Cl-]i and the cell volume at a small activity of the cotransporters, but high gCl was needed to influence Em. It seems that those cotransporter-evoked Cl--dependent alterations of Em are responsible for the disturbances in cation concentrations.
In the last part of this work we will examine how changes of external and internal osmolarity affect the cell volume, [Na+]i, [K+]i, [Cl-]i, [An-]i, and Em. The reason-consequence chain in this section will be different from the previous sections. Up to this point changes in concentration of permeant ions evoked changes in Em and possibly in cell volume, if the redistribution of ions was not osmotically balanced. Here, the initial event was an alteration of osmolarity that directly and predictably influences the cell volume. When external osmolarity increases, the cell shrinks; when internal osmolarity increases, the cell swells. These changes of cell volume may or (surprise!) may not lead to changes in intracellular ionic concentrations, as will be shown. Finally, changes of ion concentration, if they occur, will inevitably change Em.
Simulations in this part resemble what happens or may happen during a regulatory volume increase (RVI). The initial set of calculations simulates the first phase of RVI in which extracellular osmolarity is increased by adding some impermeant neutral osmolyte. The set includes simulations with high gCl (1010 ions/(sec*V), solid lines in Fig 9) and with low gCl (108 ions/(sec*V), dashed lines in Fig 9), both in “realistic” conditions. The first phase of RVI lasts a few seconds to minutes [45], so 30 mM of an external osmolyte was added at the rate of 0.5 mM/sec for 1 minute (gray area in Fig 9A and 9B). Qualitatively all changes of the ion concentrations, the volume and the voltage were in accord with expectations. The cell shrank, and [Na+]i, [K+]i, [Cl-]i, and [An-]i increased in proportion to their initial level, at least at first glance. Em hyperpolarized, which was anticipated because the increase of [K+]i enhanced the K+ transmembrane gradient and its hyperpolarizing effect, and the increase of [Na+]i diminished the Na+ transmembrane gradient and its depolarizing effect. But the quantitative picture is more complicated. First, the cation increases were not proportionally equal. At the end of the first minute, when external osmolarity reached its maximum (336 mOsm. i.e. 9.8% more than the initial value of 306 mOsm), [K+]i increased by more than 10% and [Na+]i increased by a little more than 2%. This is an apparent deviation from the simplistic volume-induced increases of concentrations that were expected to be proportionally equal, and it points toward a redistribution of ions during osmosis-related changes.
Redistributions of Cl- are the most interesting because Cl- is tightly connected with cell volume. There were no cation-Cl--cotransporters in this simulation, so Cl- was in equilibrium before the increase of external osmolarity. Osmosis-related shrinkage of the cell increased [Cl-]i and diminished the concentration-dependent inward-directed component of the driving force for Cl-. At the same time a cation-induced hyperpolarization enhanced the voltage-dependent outward-directed component of the driving force for Cl-. Thus, Cl- left the cell, and the increase of [Cl-]i was smaller than expected from the volume decrease itself. The difference, of course, depended on the value of gCl. When gCl was low, the increase of [Cl-]i was close to the expected change from volume alone (9.5% vs 9.7%), but when Cl- conductance was high, those two numbers were very different (3.1% vs 11%). After 1 minute of increased external osmolarity [Cl]i was noticeably out of equilibrium. In the case of low gCl, [Cl-]i after 1 minute was 28.10 mM; accordingly, ECl = -44.7 mV, i.e. 4.6 mV more positive than Em (-49.3 mV). High gCl shunted the membrane, so the hyperpolarization was smaller (to -48.3 mV). The rise of [Cl-]i also was smaller due to Cl- leakage (to 26.47 mM), so ECl (-46.3 mV) was 2 mV more positive than Em. When external osmolarity stabilized, [Cl-]i began to decrease, leading to both further hyperpolarization and a further volume decrease. If Cl- conductance was high, Cl- quickly equilibrated and after 10 minutes [Cl-]i = 23.70 mM, i.e. about 2 mM less than the initial concentration. That new [Cl-]i was at equilibrium and fit with new electrical conditions (ECl = Em = -49.3 mV). If gCl was low, no changes in [Cl-]i, volume or voltage were visible from 1 to 10 minutes, except a small and quick depolarization that reflected cation adjustment after the disturbance. However, the calculations showed that after several hours the cell would come to the same equilibrium state as in the case of high gCl.
Knowing that the ions were redistributed during the osmotic shock, it is not surprising that the cell volume changes were themselves different from expectations. An increase of external osmolarity by 8.9% should decrease volume in a cell that is impermeant to anything but water by 8.2%, and that was close to the volume reduction at the end of osmotic shock when gCl was low (8.8%). But when Cl- conductance was high, the cell volume decreased by 9.8% at 1 minute and by 11.8% in the eventual resting state.
Fig 9C and 9D demonstrates that fundamentally similar changes happen when NaCl, i.e. a substance that easily can cross the cell membrane, was used instead of a neutral impermeant osmolyte. As in the previous case, the external osmolarity was slowly elevated by 30 mOsm by the end of 1 minute; to do this NaCl was added at a rate of 0.25 mM/sec. Addition of extracellular NaCl directly influenced not only external osmolarity, but also the transmembrane gradients of Na+ and Cl-. Increasing the Na+ gradient enhanced the depolarizing effect of this cation on Em, which resulted in a smaller osmotic-dependent hyperpolarization (if compared to the case of an increased external neutral osmolyte described above), but only when gCl was low (compare dashed lines in Fig 9B and 9D). When gCl was high the hyperpolarization during the osmotic shock was even slightly larger. The increased transmembrane Cl- gradient had more recognizable effects. Now Cl- did not move very far from equilibrium, as in the case of a neutral osmolyte. The difference between ECl and Em never exceed 0.8 mV with high gCl, and 1.5 mV with low gCl. As a result, [Cl-]i experienced much smaller changes after the osmotic disturbance on its way to equilibrium. Accordingly, smaller further hyperpolarization and volume decreases happened after adding NaCl than a neutral osmolyte (Fig 9B and 9D). Again, a long time (many hours) is needed to reach equilibrium if gCl is low, which leads to the illusion that nothing changed in this case after the disturbance.
It also should be noted that some extra energy is needed to support steady state in a smaller cell volume after the external osmotic increase, although the cost is not high. ATP expenditure increased by about 1.5% when the system stabilized after the neutral osmolyte-induced disturbance, and about 3% extra ATP was needed in case of NaCl.
In the next two sets of simulations the internal osmolarity was increased. This was similar to the second, active phase of RVI. First, a neutral impermeant osmolyte was added to intracellular space (Fig 10A and 10B). Such an increase could happen when some macromolecules were broken down to many smaller molecules (like glycogen to glucose) or some osmolyte (for instance, taurine) was transported into the cell from the extracellular space with an appropriate transporter. For the simulation, we assume that a neutral impermeant osmolyte increases with a rate of 0.05 mM/sec for 10 minutes until its concentration reaches 30 mM. This might be too fast to be real, but slowing down the process does not change the results (except diminishing the difference between simulations with different gCl).
Most (but not all) changes of concentrations, voltage and volume associated with elevation of internal osmolarity are just opposite to those observed with a simulated increase of external osmolarity (Fig 10A and 10B). The cell swelled, and intracellular concentrations decreased, with the important exception of [Cl-]i, which increased when gCl was high. Again, decreases of ionic concentrations were not proportionally equal, in spite of the proportionality that would be expected as the direct effect of the increasing volume. [K+]i decreased by more than 10%, while [Na+]i decreased by less than 2%. Em depolarized due to decreases in both cation concentrations. The depolarization forced Cl- to enter the cell, but it had little effect, and [Cl-]i still decreased when gCl was low. However, when gCl was high, this depolarization-driven Cl- influx was substantial and [Cl-]i increased. Accordingly, the difference between ECl and Em was small (maximum 0.44 mV), and likewise the effect of Cl- on Em was also small. Still, Cl- was out of equilibrium, and when the buildup was complete, [Cl-]i continued to increase to equilibrate with the new Em, initiating a further depolarization and swelling. The new resting state was reached much more slowly with low Cl- conductance. Finally, although it might cost energy to build up an intracellular osmolyte, the cell actually saved about 1.5% of the ATP required to support ionic balance with the new larger volume.
The last set of simulations dealt with the curious case of the buildup of an intracellular impermeant anion with a mean valence equal to -1.5. It should be noted that synthesis of a new organic anion must be accompanied by a cation for electroneutrality. The most probable cation in such a case is H+, so the addition of an anion would also cause the addition of an acid. The regulation of intracellular pH is an undoubtedly important, but complicated, problem that goes beyond the scope of this paper. So, we assume that our modeled cell is capable of resolving the problem of stabilization of pH. For instance, the cell could exchange each new internal H+ for external Na+ using a Na+/H+ exchanger. Thus, in our simulations the buildup of impermeant anion will be supplemented by appropriate addition of Na+. In the “realistic” conditions of our simulation 3 Na+ were needed to electrically compensate 2 An-, which have a mean valence = -1.5. Accordingly, buildup of the anion with rate of 0.02 mM/sec for 10 minutes was accompanied by addition of Na+ at 0.03 mM/sec, to produce osmotically the same increase as the electrically neutral osmolyte in the previous set of simulations.
The results of this buildup were a bit surprising (Fig 10C and 10D). Besides the inevitable increase of the cell volume, there were no other changes of significance. Addition of 12 mM of An- with 18 mM of Na+ was largely compensated by the cell swelling, so [An-]i and [Na+]i increased only by 0.54 mM (0.5% of initial) and by 0.3 mM (1.6% of initial), respectively. [K+]i decreased as expected, but only by 0.13 mM (less than 0.1% of initial). [Cl-]i experienced the largest relative changes (2.5%), which still was less than 1 mM. The small changes in ion concentrations produced small changes in Em (-0.46 mV). And most importantly, all concentrations, including [An-]i and [Na+]i, returned to their initial values in a few minutes after the end of the buildup. Together with ions, Em also returned to its initial value. Thus, the lone result of addition of AnNax was increase of the cell volume. Some extra energy was spent during swelling, but when ionic gradients were restored, exactly the same amount of ATP could support the resting state at a larger cell volume.
The model presented here is very flexible, and allows calculation of both dynamic and steady state values for cell volumes, concentrations, membrane potentials and energy requirements resulting from changes in ion and water conductances, concentrations of permeant and impermeant ions, net valence of anions, and transporter and pump rates. Thus, it is quite general and can be used to investigate many situations. We have used it to investigate the influence of anions, cations and the transporters on cell volume and membrane potential. Some aspects of data obtained during our computational simulations have been discussed in the Results section. Here we will address three major points that deserve special attention.
The first part of this statement brings no news, but the second does. Long ago the existence of a pump that actively extruded Na+ against its concentration gradient was postulated to explain cell volume [42, 43] and until now the key role of the Na+/K+-ATPase in volume regulation was not questioned [19, 23–25] (for review see [26]). Our simulation, however, demonstrated that in a system when only cations were concerned (and it is obvious that Na+/K+-ATPase deals only with cations) the pump is responsible for the voltage, but not for the volume. Any changes in the principal cation triumvirate - Na+ conductance, K+ conductance, Na+/K+-pump activity - always and inevitably lead to changes in Em (Figs 4 and 5), even in theoretical conditions specifically designed to make an equal exchange of Na+ for K+ (electrically neutral 3Na+/3K+-pump, gNa = gK). But, as was well known, the cell voltage is much more sensitive to transmembrane movement of ions than the cell volume. The same amount of ions that is sufficient to charge the membrane capacitance and create a considerable change in Em is negligibly small for cell volume and is associated with practically undetectable volume changes (Fig 5F). As a result, the slightest imbalance in total cation transfer across the membrane, which is irrelevant for the cell volume, can be important for Em and will stimulate strong negative feedback to prevent further imbalance. In this respect Em ensures osmotically balanced changes in Na+ and K+, and this is true in all conditions with any combination of the stoichiometry of the Na+/K+-ATPase, its activity, and cation conductances, including the case when the pump only removes Na+ from the cell (Fig 5C and 5D). Thus, the cations are not directly involved in cell volume regulation. It would be reasonable to say that Na+ and K+ are not for volume, but for voltage. Importantly, they have to pay for this privilege with ATP.
The ability of our program to show the definite ATP cost of the cation nonequilibrium appears to be useful for better understanding of relationships between ions, voltage and volume. For instance, the electrogenic pump needs more energy than our hypothetical electroneutral one to create the same concentration gradients. But the reason for the increased energy requirement is the different quantity of transferred ions per one cycle of the pump, not electrogenicity as is intuitively expected. To equilibrate [Na+]i and [K+]i (both equal to 75 mM, under the condition where gNa = gK), a 3Na+/2K+-pump which transfers 5 ions per cycle needs to spend 20% more ATP (179.8 million/sec) than a 3Na+/3K+-pump (149.8 million/sec) which transfer 6 ions per cycle. In both cases an equal number of cations is transferred per second (899 million) by the ATPase, and because these are resting states the same amount of ions passively leak back (Na+ into the cell, and K+ out of it). Of course, Em is more negative with an electrogenic pump than with an electroneutral one (-36.34 vs. -27.35 mV), but since a very small amount of ions produces this voltage shift, it is practically not reflected in the energy expenditure. The stoichiometry of the pump has a great influence on Em, and simulations, which are not shown, revealed that pumps that all transferred the same amount of charge per ATP, with Na+:K+ ratios of 6:0, 4:2, 3:3, 2:4 and 0:6 will generate -72.30, -42.34, -27.35, -12.37, and +17.60 mV of Em, respectively, in order to achieve [Na+]i = [K+]i, but they all spend the same amount of energy with a precision of less than 0.002%. Thus, the energy is spent for cation electro-chemical gradients, which of course influence Em, but not for voltage itself via electrogenicity. Half the energy would be enough to reach this resting stage if the cell had half the cation conductance (S3 Fig). Interestingly, a cell can save a lot of energy supporting the cation electrochemical gradients and negative Em if its membrane is preferentially permeable to K+ (Fig 5C). The same electrogenic 3Na+/2K+-pump in the cell with the same total cation conductance needs 3.2 times less energy to equilibrate [Na+]i and [K+]i if gK/gNa = 9 compared to gK/gNa = 1 (55.8 million/sec instead of 179.8 million/sec). This is because the dominance of gK over gNa results in a smaller leak of the cations in spite of a much more negative Em (-66.11 vs. -36.34 mV).
It should be remembered that the importance of creating the transmembrane cation electro-chemical gradients by the Na+/K+-ATPase goes far beyond of generation of Em. These gradients (particularly the strong Na+ gradient) are heavily utilized by a cell for transport of metabolites, supporting Ca2+ homeostasis, controlling pH, and clearing neurotransmitters from the extracellular space, among other functions. These gradients can be used to cause non-equilibrium transmembrane distribution of Cl-, with all the following consequences. Changes in Em associated with changes in activity of the Na+/K+-ATPase (as well as changes in Na+ and K+ conductances) are also a prerequisite for possible cell volume changes. But in an “only cation system” all ionic transmembrane movements are osmotically balanced to satisfy macro electroneutrality. It is the addition of a membrane permeant anion (Cl-) what makes possible electrically neutral and osmotically significant ionic fluxes that lead to cell volume changes.
In the absence of specialized cotransporters (mainly the cation-Cl- transporters that were modeled here and to some extent the bicarbonate-Cl- transporter, which was beyond the scope of this work) Cl- is distributed passively across the cell membrane. This means that Cl- has to adjust its intracellular concentration to be in equilibrium with the cation-controlled Em (Fig 6A). When Em changes due to changes in cation conductance or Na+/K+-ATPase activity, [Cl-]i is forced to change also in order to fit the new Em. How fast these changes in [Cl-]i occur depends on the magnitude of the Cl- conductance (Fig 7). If gCl is low, [Cl-]i changes will develop slowly and will not be noticeable in a short time. But if gCl is high, [Cl-]i changes will be comparable to those of [Na+]i and [K+]i. Since Cl- “shares the room” with impermeant intracellular anions (An-), all changes in [Cl-]i must be associated with opposite sign changes in [An-]i which is only possible if the cell volume is changed. Thus, alterations of Em induced by changes in cation (mostly Na+) conductance during normal neuronal activity will inevitably be accompanied by volume changes if gCl is substantial, or will have no visible volume effects, if gCl is low.
The presence of cation-Cl- cotransporters complicates the matter. NKCC elevates [Cl-]i above equilibrium and KCC lowers it below equilibrium. A nonequilibrium distribution of Cl- enables this anion to contribute to Em (Fig 8). Regulation of Na+ conductance is still by far the more common (and more effective) way to manipulate Em of neurons, because Na+ is much further from equilibrium than Cl-, but the contribution of unequally distributed Cl- to Em should not be ignored (for review see [37]). For instance, gCl in combination with a nonequilibrium distribution of Cl- plays an important role in such complex neuronal process as direction selectivity in the retina [29].
Still, the cation-Cl- cotransporters do not disrupt the link between Cl- and the cell volume. The cotransporters determine not the absolute value of [Cl-]i, but the Cl- electro-chemical driving force, i.e. the difference from the concentration that would be equilibrium at current Em. When Em changes due to changes in gNa, gK, or the pump activities, [Cl-]i is forced to adjust accordingly, leading to cell volume changes, just as in the case with no cation-Cl- cotransporters.
To conclude this part, Cl- may or may not influence Em, depending on its transmembrane distribution (nonequilibrium or equilibrium). But the presence of substantial gCl is absolutely necessary for activity-dependent cell volume changes. The importance of Cl- in cell volume regulation was discussed theoretically and demonstrated experimentally in the literature (recently [46, 47]). What is stressed in this paper is the fact that gCl, not Na+/K+-ATPase, is responsible for volume changes. The apparent dependence of cell volume on the activity of the Na+/K+- ATPase can be explained by the following chain of events: changes in the activity of Na+/K+-ATPase lead to osmotically balanced (and therefore volume-irrelevant) changes in [Na+]i and [K+]i, that in turn affect Em. With the presence of significant gCl, the changes in Em force [Cl-]i to adjust accordingly. The transmembrane flux of Cl- is electrically neutralized by a co-directed flux of cations and the resulting transfer of NaCl and KCl is osmotically noteworthy, leading to changes of the cell volume. So, swelling of the cell following suppression of the Na+/K+-ATPase could be avoided if it would be possible to completely block gCl.
Cl- is “sharing room” with impermeant anion, An-, and the sum of [Cl-]i and [An-]i must be constant if extracellular concentrations remain unchanged. This undisputable fact has led to the suggestion that changes in [An-]i should induce compensatory changes in [Cl-]i and, consequently, that [An-]i can be the key factor to determine [Cl-]i, making possible a nonequilibrium distribution of Cl- [48]. This work was criticized from both theoretical [49] and experimental [50] points of view. And actually it also was shown computationally ten years earlier that a slow leak of An- out of the cell diminishes the cell volume, but eventually does not change [Cl-]i nor does it change [Na+]i, [K+]i, Em, and [An-]i itself [28]. The same results were obtained during influx of An- [25] as well as a buildup of An- in this work (Fig 10C and 10D); only cell volume in those two cases increased because of An- addition. Also, when [Cl-]i was changed due to, for instance, the activity of cation-Cl- cotransporters, the problem of keeping the sum of [Cl-]i and [An-]i constant was resolved by appropriate adjustment of “the room,” i.e. the cell volume (Fig 8).
However, when the mean valence of impermeant anions was altered, not only [Cl-]i, but also cation concentrations and Em are changed [25]. [Na+]i, [K+]i, [An-]i, [Cl-]i and Em were also changed in our simulation of intracellular buildup of the neutral osmolyte (Fig 10A and 10B), which could be viewed as analogous to a decrease of mean valance of the impermeant anion (if the impermeant anion were defined as everything inside the cell except Na+, K+, and Cl-). In this work neutral osmolyte is treated separately from other internal impermeants, and they all are considered to be parts of a broader concept—the osmolarity-charge asymmetry.
The osmolarity-charge asymmetry inevitably arises when the internal impermeant anion (An-) has a mean valence (z) different from -1. The equation for internal macro electroneutrality is:
[K+]i+[Na+]i=‑z*[An‑]i+[Cl‑]i,
(37)
The complication here is that An- is not a certain anion or even a set of anions of the same kind, but rather a collection of very different small and large molecules. In a cell with a membrane that is permeable only to Na+, K+ and Cl-, An- can be defined as “everything which is internal, charged and impermeable” and it definitely must be an “anion” to compensate for the deficiency of negative charge of the main inorganic ions inside the cell. In this case the mean valence of An- (z) is the ratio of all electrical charges that belong to An- to the concentration of An-. Osmotically active proteins that carry multiple negative charges per molecule support the case for z < -1 (i.e. larger negative charge), but they are responsible for less than 10% of total cytoplasmic osmolarity [51]. Immobile proteins incorporated in the cell membranes, which can represent half of all proteins [52] provide some more negative charge without any osmotic contribution. However, the majority of [An-] is made of small organic molecules [53] that are mostly monovalent under physiological pH, such as creatine phosphate (about 40 mM in frog muscle [54]) and other phosphates and sulfates. Also, a significant part of the internal osmolytes is comprised of amino acids (up to 37 mM in rat brain, [55]), but among them only glutamate and aspartate are negatively charged, and some others are positively charged. Taking into account the wide diversity of components from which An- is comprised, it is not surprising that the mean valence of An- could be very different from cell to cell. In the case of myocytes a reasonable value of z is -1.65 [28], while for lymphoid cells it can be as large as -2.5 [23, 56]. Dusterwald and coworkers have used z = -0.85 [25].
Whatever z is, as long as z ≠ -1, it creates an asymmetrical osmolarity-charge setting, when the quantity of internal anions is not equal to the quantity of internal cations in an osmotic sense. The presence of internal neutral osmolytes, such as the just mentioned neutral amino acids, contributes to the asymmetry. A special place among them is occupied by sulfonic amino acid taurine, a zwitterion which is neutral at physiological pH. The concentration of taurine can be as high as 60 mM or even more, but varies significantly from species to species and from cell type to cell type, with higher concentration in mammals than in amphibians or reptiles, in retina than in brain or muscle, and in photoreceptors than in other retinal cells [57–59]. Addition of an external neutral osmolyte (although in the case of neural systems it is only a few mM, mostly from glucose) completes the equation of osmotic balance:
[K+]o+[Na+]o+[Cl‑]o+[osm]o=[K+]i+[Na+]i+[An‑]i+[Cl‑]i+[osm]i
(38)
where [osm] is the concentration of uncharged osmolytes.
To quantify the extent of the osmolarity-charge asymmetry a new parameter - the coefficient of asymmetry (ka) - is introduced as follows:
ka=(‑z*[An‑]i+[Cl‑]i)/([An‑]i+[Cl‑]i+[osm]i–[osm]o)
(39)
Because charge of monovalent Na+ and K+ is equal to their osmolarity, both intracellular and extracellular, the charge-osmolarity imbalance, quantified by ka, results from anions and uncharged species. Thus, ka is the ratio of all intracellular negative electrical charges to all osmotically active intracellular molecules except cations. The equation also includes external osmolyte, because addition of [osm]o is the same for net osmolarity as subtraction of an equal amount of [osm]i. It is convenient to replace [osm]i−[osm]o with d[osm], which is the difference in concentrations of internal and external neutral osmolytes and can be positive or negative. Accordingly, the equation for internal macro electroneutrality (Eq 37) can be rewritten in terms of ka:
[K+]i+[Na+]i=ka*([An‑]i+[Cl‑]i+d[osm])
(40)
Since [K+]o + [Na+]o = [Cl-]o, Eq 38 for osmotic balance can be rewritten with regard to cations as:
2*[cat]o=[cat]i+[cat]i/ka
(41)
where [cat]i and [cat]o are total intracellular and extracellular concentrations of the cations. From here one can derive the equation of osmolarity-charge asymmetry that describes the uneven, yet equilibrium cation distribution:
[cat]i/[cat]o=2*ka/(ka+1);
(42)
Accordingly, Em in this equilibrium state with no Na+/K+-ATPase activity will be determine by ka:
Em=(RT/F)*ln(2*ka/(ka+1))
(43)
In this condition, both ENa and EK must be equal to Em, and both cations separately follow Eq 42:
[Na+]i/[Na+]o=2*ka/(ka+1);
(44)
[K+]i/[K+]o=2*ka/(ka+1);
(45)
This is true equilibrium when the concentration driving forces for both Na+ and K+ are countered by the electrical driving force, and osmotic balance also holds. In this respect it is similar to Double Donnan equilibrium, but with one important difference - the equilibrium based on the osmolarity-charge asymmetry is possible only if the cell membrane is permeable to cations, but not to Cl-. Opening of gCl will lead to inevitable and unlimited swelling. Of course, the system can be stabilized if the cation gradients are supported by the Na+/K+-ATPase. Thus, there are two factors of different nature that determine cation concentrations: one active, dependent on the cation conductances and the pump activity, and the other passive, dependent on ka. Accordingly, there are two parts of cation-dependent Em. The active part of Em is much larger than the passive (Fig 5E), and also the active part can be quickly, in a small fraction of a second, changed by manipulating cation conductances. There is no doubt that this active component, which is determined by gNa, gK and the activity of the Na+/K+-ATPase, dominates Em. On other hand, although the passive osmolarity-charge asymmetry dependent component is probably present in Em of every cell (since it is very unlikely that ka = 1 in any cell), its contribution will be insignificant in most conditions. Changes in ka, are probably common since they can be a result of changes in z, [An-]i, [osm]i, or [osm]o, but in most cases with small effect on Em. As illustrated in Fig 11B, a decrease of ka from 2 to 1.5 (which was achieved by an increase of d[osm] by 32 mM with constant z) depolarized Em by only 2.8 mV, since it was associated with relatively insignificant changes in [K+]i and [Na+]i (from 6.9 to 6.0 mM and from 193.3 to174 mM, respectively, Fig 11A). Also accompanying this was a modest 8.3% increase in cell volume (normalized to the volume in osmolarity-charge symmetric conditions when ka = 1). The small changes in Em (about 2 mV) demonstrated in simulations of osmotic disturbances (Fig 9A and 9B, Fig 10A and 10B) are a consequence of changes in osmolarity-charge asymmetry, when the cell was slightly hyperpolarized due to an increase of ka resulting from the buildup of external osmolyte and slightly depolarized due to a decrease of ka resulting from the buildup of internal osmolyte. But when Em was changed, [Cl-]i had to change too, assuming the presence of substantial gCl. In this respect Glykys and coauthors [48] were right in claiming that impermeant anion An- can influence [Cl-]i, although it happened not because of a direct link, but due to a chain of events including changes in osmolarity-charge asymmetry, cation concentrations and Em. However, if Cl- was in equilibrium with Em, it will continue to be in equilibrium.
Importantly, osmolarity-charge asymmetry is also affected by changes in [Cl-]i. if the distribution of ions was already asymmetric. As was mentioned in the results concerning Fig 8E, the elevation of [Cl-]i by NKCC in “realistic” (i.e. osmolarity-charge asymmetric) conditions led to a decrease of the total internal anion charge because An- with z = -1.5 was replaced by the monovalent Cl-. The inevitable result of that were changes of cation concentrations and Em. Effects of cation-Cl- cotransporter-induced changes in [Cl-]i on [Na+]i, [K+]i, and Em can be revealed by comparing “realistic” conditions with “simplified” (i.e. osmolarity-charge symmetric) conditions, especially when gCl was low (108 ions/(sec*V)) and Cl- practically had no direct contribution to Em. When Cl- was pumped by cation-Cl- cotransporters in symmetric “simplified” conditions, it replaced (or was replaced by) an equal quantity of monovalent An-. As a result, ka continues to be 1, [Na+]i and [K+]i remain almost the same, and slight changes of Em did not exceed 0.3 mV (Fig 8A for NKCC and Fig 8C for KCC). But in asymmetric “realistic” conditions an increase of [Cl-]i by NKCC resulted in a reduction of ka from 1.47 to 1.28, a decrease of [Na+]i and [K+]i and a depolarization by 1.42 mV (see S4A Fig); a decrease of [Cl-]i by KCC led to smaller changes in ka (from 1.47 to 1.56) and a subsequent increase in both cation concentrations and hyperpolarization by -0.78 mV (S4B Fig).
Finally, it should be noted that this asymmetry-dependent voltage is completely determined by ka and is independent of the internal ionic and osmotic compositions as long as they result in the same ka. The data for Fig 11 were obtained during manipulation of the neutral osmolytes (see explanation in the figure legend), but the stars with numbers were from our previous simulations with different internal compositions (star 1: Fig 4E, star 3: Fig 4F, and star 2: Fig 5B, “realistic” conditions; all for the points at the left of the graphs where there is no Na/K pumping). The results of computational simulations were exactly the same as predictions from Eqs 43, 44, and 45.
To summarize, [An-]i and its mean valence play an important role in determination of cell volume. It was shown earlier, and it was confirmed here. [An-]i and its mean valence also, together with other factors ([Cl-]i, internal and external neutral osmolytes), contribute to creation of osmolarity-charge asymmetry, which passively influence cation distribution and Em, although the effect is small compared to the active Na+/K+-ATPase dependent cation voltage.
|
10.1371/journal.ppat.1002992 | The Polymorphic Pseudokinase ROP5 Controls Virulence in Toxoplasma gondii by Regulating the Active Kinase ROP18 | Secretory polymorphic serine/threonine kinases control pathogenesis of Toxoplasma gondii in the mouse. Genetic studies show that the pseudokinase ROP5 is essential for acute virulence, but do not reveal its mechanism of action. Here we demonstrate that ROP5 controls virulence by blocking IFN-γ mediated clearance in activated macrophages. ROP5 was required for the catalytic activity of the active S/T kinase ROP18, which phosphorylates host immunity related GTPases (IRGs) and protects the parasite from clearance. ROP5 directly regulated activity of ROP18 in vitro, and both proteins were necessary to avoid IRG recruitment and clearance in macrophages. Clearance of both the Δrop5 and Δrop18 mutants was reversed in macrophages lacking Irgm3, which is required for IRG function, and the virulence defect was fully restored in Irgm3−/− mice. Our findings establish that the pseudokinase ROP5 controls the activity of ROP18, thereby blocking IRG mediated clearance in macrophages. Additionally, ROP5 has other functions that are also Irgm3 and IFN-γ dependent, indicting it plays a general role in governing virulence factors that block immunity.
| The ability of microorganisms to cause disease in their hosts is often mediated by proteins that are secreted by the pathogen into the host cell as a means of disarming host signaling. Previous studies with the protozoan parasite Toxoplasma gondii have revealed that secretion of parasite protein kinases into the host cell mediates virulence in mouse, a natural host for transmission. Curiously, some of these virulence factors are active protein kinases, while other related pseudokinases lack enzymatic activity; hence, it was unclear how they functioned in promoting virulence. In the present work we demonstrate that ROP5, an inactive member of this protein kinase family, regulates the active protein kinase ROP18, which normally prevents clearance of the parasite in interferon-activated macrophages. Allosteric regulation of enzymes is a common theme in biology, but this is the first example of such a mechanism regulating a pathogen virulence factor. The potential advantage of such a layered process is that it might allow greater temporal or spatial control and perhaps protect the parasite from disabling strategies by the host.
| Toxoplasma gondii is an obligate intracellular parasite that infects a wide range of vertebrate animal hosts and causes zoonotic infection in humans, leading to potentially severe congenital infections and risk of reactivation in immunocompromised patients [1]. In North America and Europe, T. gondii exists as four distinct clonal lineages that show marked virulence differences in laboratory mice, which serve as a model for infection [2], [3]. Forward genetic analyses have been used to map the genes responsible for virulence in laboratory mice [4], [5]. Remarkably, this complex trait is largely mediated by a few members of a large family of polymorphic serine threonine (S/T) protein kinases secreted from rhoptries (ROP) into the host cell during invasion [6], [7]. The ROP kinase family consists of ∼20 active members, as well as a similar number of putative pseudokinases that are predicted to lack kinase activity [8]. The structures of several ROP pseudokinases reveal they contain a typical kinase fold and yet they are structurally and phylogenetically diverse [9], [10].
Most strains of T. gondii survive within naïve macrophages; however, when previously activated by exposure to IFN-γ macrophages acquire the ability to kill or inhibit parasites [11]. During primary infection, inflammatory monocytes are recruited to the site of infection where they are critical for control of intracellular T. gondii [12], [13]. Macrophages control T. gondii through induction of iNOS, which leads to stasis [14], reactive oxygen intermediates, which leads to killing of opsonized parasites [15], and upregulation of immunity related GTPases (IRGs), which destroy intracellular parasites [16], [17]. Recruitment of IRG effectors to the parasite containing vacuole results in destruction of the parasite residing within it [18], [19]. Compared to most mammals, the IRG gene family is highly amplified in rodents [20], where it plays a major role in natural resistance to T. gondii [21], [22].
Not all strains of T. gondii are susceptible to clearance in IFN-γ-activated macrophages: highly mouse virulent type I parasites resist IRG recruitment and consequently avoid clearance, while intermediate virulent type II and avirulent type III parasites are unable to block IRG recruitment and are destroyed [14], [23]. Recent studies have revealed the mechanism for this escape: the S/T kinase ROP18 phosphorylates a number of IRGs on key threonine residues in switch region I of the GTPase domain, thereby preventing assembly on the vacuole and blocking clearance in activated macrophages [24], [25]. The IRG system is also dependent on the autophagy protein Atg5, although the molecular basis for this requirement remains unclear [23], [26].
Genetic mapping studies have also implicated the pseudokinase ROP5 in acute virulence [27], [28], a result that is unlikely to be due to kinase activity as ROP5 lacks a key catalytic residue and binds ATP in an unconventional manner [29]. Comparison of several pair-wise genetic crosses suggest that ROP5 interacts with ROP18 [27]; however, the molecular basis for the dramatic effects of ROP5 on virulence remains uncertain. Pseudokinases in other systems have recently been shown to perform regulatory roles [30], raising a similar possibility for ROP5. Herein, we explore the mechanism of action for ROP5 and demonstrate that it controls ROP18 activity, while also serving a separate and essential role in acute virulence.
Previous studies have shown that ROP5 deficient (RHΔku80Δrop5) parasites grow normally in vitro, but are highly attenuated in laboratory mice [27]; however, the molecular basis of this phenotype is not understood. To examine growth in vivo, CD-1 outbred mice were infected with either wild type (RHΔku80) or ROP5 deficient (RHΔku80Δrop5) parasites expressing luciferase and growth was followed over time. Virulent RHΔku80 parasites expanded rapidly, as detected by luciferase activity, until the mice succumbed to infection between days 6–8 (Figure 1A, Figure S1). In contrast, ROP5 deficient (RHΔku80Δrop5) parasites expanded normally for the first few days and then dramatically decreased by the end of the first week, resulting in survival of the mice (Figure 1, Figure S1).
To examine innate immune responses, we measured inflammatory cytokines in serum during the first week post infection. The levels of IFN-γ, IL-6, and MCP-1 increased at day 3 and were all significantly higher at day 5 (P≤0.01) in mice infected with wild type (RHΔku80) parasites as compared to ROP5 deficient (RHΔku80Δrop5) parasites (Figure 1A, B). Consistent with a rise in IFN-γ, we also observed an increase in IL-12p40 at day 3 and 5 in mice infected with wild type (RHΔku80) parasites. In contrast, IL-10 and TNFα were essentially unchanged. Overall, changes in cytokine levels appeared to track closely with parasite burden.
To test whether ROP5 modulates host cell transcription, we infected human foreskin fibroblasts (HFF) with either wild type (RHΔku80) or ROP5 deficient (RHΔku80Δrop5) parasites and examined gene expression using the Affymetrix HG-U113A_2 Human Array. There were no significant differences in gene expression in HFF infected with wild type (RHΔku80) or ROP5 deficient (RHΔku80Δrop5) parasites (NCBI GEO record GSE32104), indicating ROP5 does not directly modulate host gene expression, at least under the conditions tested in vitro.
The failure of ROP5 deficient parasites to expand in vivo could emanate from some alteration in the immune response. For example, if ROP5 were normally immunosuppressive, in its absence, a stronger or more potent immune response might provide more effective control of infection. To test whether ROP5 deficient parasites lack a normally suppressive function, mice were infected separately, or coinfected with wild type (RHΔku80) and ROP5 deficient (RHΔku80Δrop5) parasites, and followed for 30 days to assess survival. Consistent with previous studies, mice infected with wild type (RHΔku80) or ROP5 complemented (RHΔku80Δrop5 Complement) parasites succumbed in 9–10 days (Figure 2A). Although ROP5 deficient parasites failed to cause lethal infection, coinfection with wild type parasites led to rapid death of all mice by day 10 (Figure 2A). Moreover, mice immunized with ROP5 deficient parasites were able to generate a normal adaptive immune response and survive a secondary challenge with a normally lethal dose of wild type parasites (Figure 2B). These findings indicate that it is unlikely that ROP5 is globally immunosuppressive.
Alternatively, it was possible that ROP5 deficient parasites were metabolically restricted in vivo, similar to the previously reported pyrimidine biosynthesis mutants (i.e. cpsII mutants), which are unable to propagate in vivo due limitations in uracil for salvage [31]. cpsII mutants fail to expand in both wild type and Ifng−/− mice [31], consistent with their metabolic limitation. To determine if ROP5 deficient (RHΔku80Δrop5) parasites were controlled by an IFN-γ-dependent mechanism, we tested the virulence of ROP5 deficient (RHΔku80Δrop5) parasites in Ifngr1−/− mice, which are unable to respond to IFN-γ and hence, are highly susceptible to toxoplasmosis [32]. Ifngr1−/− mice were completely susceptible to infection with ROP5 deficient parasites, succumbing in the same time frame as wild type C57BL/6 or Ifngr1−/− mice infected with virulent wild type parasites (Figure 2C). To further clarify the role of ROP5 in immune evasion we tested the virulence of ROP5 deficient (RHΔku80Δrop5) parasites in mice lacking the inducible nitric oxide synthase (iNOS) (Nos2−/−mice), the superoxide-generating NADPH-oxidase gp91phox (X-CGD mice), or the recombination activating gene 1 (Rag1−/− mice), important for B and T cell function. iNOS−/− and X-CGD mice survived infection with ROP5 deficient (RHΔku80Δrop5) parasites (Figure 2D) and did not present symptoms of illness or weight loss (data not shown), similar to C57BL/6 control mice (Figure 2C). Although Rag1−/− mice succumbed to ROP5 deficient (RHΔku80Δrop5) infection, death was delayed compared to wild type (RHΔku80) infection (Figure 2D).
Collectively these findings indicate that IFN-γ signaling is necessary for controlling ROP5 deficient parasites and suggest that ROP5 is critical to the parasites' ability to resist innate immune effectors generated by the IFN-γ response.
Previous studies have shown that inflammatory monocytes, which are recruited to the peritoneal cavity following i.p. infection [13], are critical for controlling toxoplasmosis in mice [12]. To determine if the altered growth kinetics of ROP5 deficient (RHΔku80Δrop5) parasites in mice were due to differences in cellular recruitment, we examined the frequency of myeloid cells in the peritoneum at intervals after infection by FACS. Infection with wild type (RHΔku80), ROP5 deficient (RHΔku80Δrop5) and ROP5 complemented (RHΔku80Δrop5 Complement) parasites all induced robust recruitment of inflammatory monocytes to the peritoneal cavity by day 3 (Figure 3A, B middle gate). Although the numbers of inflammatory monocytes were similar at day 3, the number of resident monocytes drastically decreased in mice infected with all three strains of T. gondii compared to uninfected mice (Figure 3B, Figure S2), a result that is likely due to cell lysis as a consequence of high parasite replication at this time point (Figure 1A). The number of inflammatory monocytes at day 5 also correlated with parasite growth, declining in mice infected with wild type or ROP5 complemented parasites (Figure 3A), while remaining elevated in mice infected with ROP5 deficient parasites. Increased numbers of neutrophils were also observed in infected mice (Figure S2), a phenomenon that has been seen previously in association with high parasite burdens [12]. Collectively these findings indicate that ROP5 deficient parasites initially expand in resident macrophages, but that following recruitment of inflammatory monocytes, the infection is controlled.
Based on the differences in initial growth in vivo (Figure 1A), we tested the ability of ROP18 deficient and ROP5 deficient parasites to survive in resident peritoneal macrophages (Gr1− F4/80+) vs. inflammatory monocytes (Gr1+ F4/80+). Naïve macrophages isolated from the peritoneal cavity of normal mice showed limited ability to clear either wild type (RHΔku80), ROP18 deficient (RHΔku80Δrop18), ROP5 deficient (RHΔku80Δrop5) or ROP5 complemented (RHΔku80Δrop5 Complement) parasites (Figure 4A). In contrast, parasites that were deficient in either ROP18 or ROP5 were efficiently cleared by Gr1+ monocytes in vitro (Figure 4B), a result that was not accompanied by loss of cells from the monolayer. Survival was completely restored to wild type levels in a strain genetically complemented for ROP5 expression (Figure 4B).
Since ROP18 is known to enhance survival through disrupting IRG recruitment to the parasite containing vacuole [24], we examined the cellular localization of Irgb6 after infection with ROP18 or ROP5 deficient parasites. Although Irgb6 remained diffuse in cells infected with either wild type (RHΔku80) and ROP5 complemented (RHΔku80Δrop5 Complement) parasites, both the ROP18 and ROP5 deficient parasites readily accumulated Irgb6 on the vacuole membrane. The level of Irgb6 recruitment was lower than the extent of clearance in the overnight assay, a result that may be due to kinetic differences (Figure 4C, D). Collectively, these results show that inhibition of recruitment of IRGs by virulent strains of T. gondii requires both ROP18 and ROP5.
Given that ROP18 has previously been shown to be both necessary and sufficient to subvert IRG clearance in murine macrophages infected with T. gondii [24], it became important to determine how ROP5 influences this pathway. The dependence of ROP18-dependent functions on ROP5 might result from either altered expression or localization of ROP18. Western blot analysis demonstrated that ROP18 was expressed at near wild-type levels in the absence of ROP5 (Figure 5A,B). Immunofluorescence analysis detected ROP18 properly localized at the vacuolar membrane in both wild type cells (RHΔku80) and ROP5 deficient (RHΔku80Δrop5) parasites (Figure 5C,D). Collectively, these findings indicate that absence of ROP5 is not responsible for altered expression or localization of ROP18.
Some mammalian pseudokinases have been reported to allosterically regulate active kinases [30], suggesting that ROP5 may regulate ROP18. To test this hypothesis in vitro, recombinant ROP18 was used to phosphorylate the artificial substrate dMBP (Figure 5E), or the natural substrate Irgb6 (Figure 5F), in the presence or absence of recombinant ROP5. To determine whether a similar interaction occurs in vivo, ROP18 was immunoprecipitated from wild type (RHΔku80) parasites or those that were ROP5 deficient (RHΔku80Δrop5) (Figure 5G,I) and used to phosphorylate dMBP (Figure 5H) or Irgb6 (Figure 5J) in vitro. In the absence of ROP5, both recombinant ROP18 (Figure 5E,F) and endogenous ROP18 (Figure 5H,J), demonstrated a greatly diminished capacity to phosphorylate both dMBP and Irgb6 substrates based on 32PO4 labeling. Activity of recombinant ROP18 increased ∼12 fold (Figure 5 E,F), while a ∼35 fold increase in endogenous ROP18 activity was observed in the presence of ROP5 (Figure 5J). Restored expression of ROP5 in the complemented clone resulted in phosphorylation of dMBP or Irgb6 by immunoprecipitated ROP18 (Figure 5 H,J). The enhanced activity of ROP18 in the presence of ROP5 did not result from a stable complex between these proteins, as they failed to co-immunprecipitate in lysates of infected, IFN-γ activated cells (Figure S3A). Enhanced activity of ROP18 in the presence of ROP5 also did not result from an interaction between ROP5 and Irgb6 (Figure S3B), nor did the previously demonstrated interaction between ROP18 and Irgb6 [24], require the presence of ROP5 (Figure S3C). Although ROP5 activated ROP18, it failed to demonstrate catalytic activity of its own in vitro (Figure 5E,F), consistent with the prediction that it encodes a pseudokinase [27], [28]. Taken together, these results indicate that the catalytic activity of ROP18 is regulated by the predicted pseudokinase ROP5, and that this pathway is required for avoidance of IRG clearance in inflammatory monocytes.
Our results indicate that ROP5 and ROP18 are both required for escape from the IRG pathway, consistent with the finding that ROP5 regulates the activity of ROP18. We sought to determine whether the defects in ROP5 and ROP18 could be compensated by defects in the IRG pathway. However, of the two IRG proteins that are shown to be targeted by ROP18, there is presently no knockout available for Irgb6, and the phenotype of Irga6 mutants challenged with T. gondii is very modest, especially in cell-autonomous control of parasite survival [33]. Nevertheless, Irgm proteins are known to regulate the proper recruitment of Irga6 and Irgb6 to the vacuole surrounding susceptible strains of T. gondii [34], [35]; absence of Irgm1 or Irgm3 alters the targeting and function of Irgb6 and Irga6 and effectively cripples the IRG system. However, the use of mice lacking Irgm1 is complicated by pleomorphic effects of Irgm1-deficiency on T cell development [36] and macrophage motility [37]. In contrast, Irgm3-deficient mice have shown normal immune cell development, yet are highly susceptible to infection by T. gondii [17]. Heterologous expression of tagged proteins indicates that Irgm proteins are necessary for proper recruitment of Irga6 and Irgb6 to the vacuole surround susceptible strains of T. gondii [35], suggesting the same requirement might be true for endogenous proteins.
To explore the interaction between ROP5 and ROP18 and the IRG pathway, we examined the recruitment of Irgb6 and parasite clearance in IFN-γ-activated, bone-marrow-derived macrophages derived from wild type C57BL/6 and Irgm3−/− mice. To provide a more complete set of parasite strains for this comparison, we complemented the RHΔku80Δrop18 mutant described previously [24] by reintroducing a single copy of ROP18 that restored normal expression and reversed the virulence defect seen in outbred mice (Figure 6A,B, Figure S4, S5). Parasites deficient in ROP18 or ROP5 demonstrated decreased survival in IFN-γ activated wild type macrophages compared to their respective complemented lines or the wild type strain (RHΔku80), which is resistant to clearance as previously reported [24] (Figure 6C). The enhanced clearance of ROP5 or ROP18 deficient parasites was largely reverted in IFN-γ activated Irgm3−/− macrophages (Figure 6D). Vacuoles containing both ROP18 deficient (RHΔku80Δrop18) and ROP5 deficient (RHΔku80Δrop5) parasites showed enhanced Irgb6 accumulation that was restored to normal in the complemented parasite strains in wild type macrophages (Figure 6 E,F). Additionally, the enhanced recruitment of Irgb6 seen in ROP5 or ROP18 deficient mutants was restored to normal in the absence Irgm3 (Figure 6 E,F). Deletion of Irgm3 also affected the abundance and pattern of Irgb6, which tended to aggregate in clusters in the absence of Irgm3 (Figure 6F). These studies reinforce the model that recruitment of Irgb6 is dependent on Irgm3 and establish that ROP18 and ROP5 have indistinguishable phenotypes when it comes to survival in activated macrophages in vitro.
To examine survival in vivo, wild type and Irgm3−/− mice were challenged with parasites and luciferase activity and survival were recorded. Following s.c. challenge of C57BL/6 mice, wild type (RHΔku80) parasites rapidly expanded while ROP5 deficient (RHΔku80Δrop5) parasites were controlled as shown by luciferase imaging studies (Figure 7A). Interesting, ROP18 deficient (RHΔku80Δrop18) parasites expanded with delayed kinetics in C57BL/6 mice and tissue burdens had begun to recover when animals succumbed to infection (Figure 7A). A similar response was also seen in CD1 outbred mice challenged with ROP18 deficient (RHΔku80Δrop18) parasites (Figure S4). In the absence of Irgm3, both wild type (RHΔku80) and ROP18 deficient (RHΔku80Δrop18) parasites underwent rapid expansion and reached high tissue burdens as reflected by luciferase activity, while ROP5 deficient (RHΔku80Δrop5) parasites showed a delayed expansion and reached lower total levels (Figure 7A). The expansion of parasites observed by bioluminescence imaging mirrored survival outcomes. Challenge with wild type strain (RHΔku80) parasites led to rapid and complete mortality of both wild type and Irgm3−/− mice (Figure 7B). Wild type mice infected with ROP18 deficient (RHΔku80Δrop18) parasites exhibited a delayed death phenotype similar to that seen in outbred mice (Figure 6B), while Irgm3−/− mice succumbed rapidly, similar to wild type parasite infection (Figure 7B). In contrast, ROP5 deficient (RHΔku80Δrop5) parasites were completely avirulent in wild type C57BL/6 mice, while they caused 100% mortality in Irgm3−/− mice, albeit with a delay in time to death (Figure 7B). The susceptibility of Irgm3−/− mice infected with ROP5 deficient (RHΔku80Δrop5) parasites was more rapid when injected i.p. with a similar time to death as wild type parasites (Figure S6).
Although the pseudokinase ROP5 was previously shown to be essential for acute virulence of T. gondii in laboratory mice, the basis for this was initially unclear, especially given the predicted lack of catalytic activity of this protein. Here we demonstrate that ROP5 regulates the activity of ROP18, an active S/T kinase that phosphorylates IRGs, thus blocking their accumulation on the parasite containing vacuole. ROP5 was necessary for the full enzymatic activity of ROP18, although it was not required for stable expression or normal trafficking to the parasite-containing vacuole. Studies using Irgm3 deficient macrophages revealed that the inability of ROP5 and ROP18 deficient parasites to avoid IRG recruitment was fully reverted in vitro. Moreover, the attenuation of the ROP deficient mutants was fully reversed in Irgm3 deficient mice. These findings reveal that ROP5 is a multifunctional pseudokinase that regulates acute virulence in T. gondii in part by governing the active kinase ROP18, and by affecting additional effectors that are both IFN-γ and Irgm3-dependent.
ROP5 is a member of a polymorphic family of secretory S/T kinases that are highly divergent from human kinases and which have been amplified in the genome of T. gondii [8]. Forward genetic mapping revealed that ROP5 is primarily responsible for differences in mouse virulence between highly virulent type I strains and intermediate virulent type II strains [27], and also between type II and avirulent type III strains, although paradoxically the type III ROP5 locus is positively associated with virulence [28]. In all strains, the ROP5 locus encodes a cluster of predicted pseudokinases all of which lack the central conserved Asp residue of the catalytic triad typical of S/T kinases [38]. Our findings demonstrate that the virulence defect in ROP5 deficient parasites is completely reversed in mice lacking Ifngr1−/− or Rag1−/−, indicating that ROP5 mediates escape from IFN-γ-dependent effector mechanisms. Alternative models, such as ROP5 deficient parasites being auxotrophic for nutrients that may be limiting in vivo, or ROP5 being a global suppressor of immune responses, are not supported.
Previous studies have shown that IFN-γR1 is required for control of T. gondii in both hematopoietic and non-hematopoietic cells [32], and both compartments likely contribute to IRG-mediated clearance, which in the mouse provides one of the most effective means of control [21], [22]. At the level of survival in macrophages in vitro, ROP5 and ROP18 were both required for avoidance for recruitment of IRGs and clearance. In previous studies we have shown that ROP18 deficient parasites exhibit normal survival in naive macrophages, but are restricted in IFN-γ activated peritoneal macrophages and that survival correlates with avoidance of Irgb6 recruitment [24]. Here we extend these findings to show that ROP18 or ROP5 deficient parasites show enhanced Irgb6 recruitment and clearance in Gr1+ monocytes, and in bone marrow derived macrophages activated in vitro with IFN-γ. In separate studies, others have shown that ROP18 or ROP5 deficient parasites also fail to block recruitment of Irga6 and Irgb6 in IFN-γ-activated MEFs [39]. In the present study, the increased susceptibility of ROP18 and ROP5 deficient parasites to clearance by IFN-γ-activated macrophages was completely dependent on Irgm3, a regulatory protein required for homeostasis of IRGs. Moreover, deficiency in Irgm3 reverted the phenotype of both the Δrop18 and the Δrop5 mutants in vivo.
Previous genetic analyses of acute virulence in type I strains of T. gondii indicated that ROP5 and ROP18 interact to control virulence [7], [27]. We now demonstrate that the basis for this relationship is that ROP5 controls the kinase activity of ROP18, thus affecting its ability to phosphorylate substrates. ROP5 activation of ROP18 activity contributes to avoidance of IRG recruitment in IFN-γ activated macrophages, hence promoting survival. ROP18 actively phosphorylates a number of IRG proteins in a common motif in switch region I [24], thereby affecting GTPase activity, and oligomerization [25]. Following phosphorylation, IRG proteins are unable to load onto the parasite containing vacuole, thus blocking this interferon-mediated clearance pathway [40]. Previous studies have indicated that Irgm3 is recruited to the PVM [19], [41], and it contains the conserved motif [24], and therefore is a potential substrate of ROP18. Additionally, the activity of ROP18 in phosphorylating ATF6β [42], is likely to also dependent on ROP5. ATF6β has been proposed to affect a later step in resistance mediated through dendritic cell activation of T-cells, a process likely important in adaptive immunity [42]. At an earlier stage, ROP18 is essential for controlling avoidance of the IRG pathway, a process that participates primarily in innate immunity [43]. Collectively, these two pathways likely control the ROP5-dependent activities of ROP18 in mediating virulence.
Our findings are consistent with a model whereby ROP5 acts as an allosteric regulator of ROP18. ROP5 was required for full catalytic activity of ROP18 using endogenous enzyme immunoprecipitated from cells and in vitro testing against the heterologous substrate dMBP or the natural substrate Irgb6. ROP5 also directly activated the kinase activity of recombinant ROP18 in vitro against both dMBP and Irgb6. These results are reminiscent of recent reports of a role for pseudokinases in mammalian cells regulating their active partners [30], [44]. For example, the pseudokinase STRADα regulates LKB1, a S/T protein kinase that regulates AMP activated kinase and acts as a tumor suppressor [45]. Although catalytically inactive, STRADα adopts a closed conformation typical of an active kinase and together with the adaptor MO25α promotes the active conformation of LKB1 [46], [47]. Unlike the situation with LKB and STRADα, the activation of ROP18 occurs despite there not being a strong interaction with ROP5, which does not coIP from cell lysates [48], and present report. However, the failure to observe a stable complex under these conditions does not preclude ROP5 and ROP18 from interacting at a lower affinity, or in a complex that depends on local interactions on the PVM. It is conceivable that transient binding of ROP5 to ROP18 facilitates auto-catalytic activation, which results from phosphorylation in helical extensions of the N-lobe of the kinase domain [10]. In separate studies, using a more sensitive approach based on tandem-affinity purification (TAP) [49] of ROP5, ROP18 was one of the major components to copurify in a complex with ROP5, and this was validated by reciprocal TAP-tagging of ROP18 (Etheridge, Sibley unpublished). Moreover, we observed that ROP5 is found in a complex with other ROP kinases, suggesting it may activate other kinases similar to ROP18 (Etheridge, Sibley unpublished). Although the precise mechanism of regulation is yet unclear, our data indicate that ROP5 is an allosteric activator of ROP18, thus establishing a new role for pseudokinases in controlling pathogen virulence factors.
Our findings differ from a recent report that also examined the interaction of ROP5 and ROP18 [48]. This prior study reported that co-expression of a cosmid contain the locus of ROP5 from the type I strain was not able to enhance the activity of ROP18 from a type II strain, when co-transfected into a recombinant strain called S22 [48]. Interpretation of this experiment is complicated by the fact that S22 is the product of recombination between types II and III and it also contains the type II ROP5 locus, which has previously been associated with avirulence [27]. Either due to this complex mixture of ROP5 alleles, or another undefined locus, this strain may harbor an epistatic activity that suppresses activation of ROP18. In contrast we demonstrate that the major allele of ROP5 from the type I strain activates the kinase activity of ROP18 from a type I strain in vivo, using isogenic strains, and in vitro using purified recombinant protein to phosphorylate both heterologous and endogenous substrates. This later result was also observed using GST-Irga6 as a substrate in vitro [39], confirming the ability of ROP5I to enhance the activity of ROP18I.
Two recent studies also reported that ROP5 binds directly to some IRGs, notably Irga6, affecting its oligomerization and GTPase activity in vitro [39], [48]. This result suggests that ROP5 may also inhibit oligomerization in vivo, thus decreasing IRG accumulation on the parasite-containing vacuole. However, this activity alone is unlikely to be sufficient for parasite survival, given the inability of type III strains to avoid IRG recruitment despite expressing the same complement of ROP5 alleles as seen in type I strains [27]. Additionally, although ROP5 binds reasonably well to Irga6, it binds less efficiently to other IRGs such as Irgb6 ([39] and present study). As such, the ability of ROP5 to directly activate the kinase activity of ROP18 may be more important for targets such as Irgb6. Collectively, the binding of ROP5 to IRGs and inhibition of oligomerization is expected to work cooperatively with its ability to enhance the catalytic activity of ROP18, thus disrupting IRG function.
The IRG pathway has been described as a major immunity mechanism in the murine system due to the expansion of this family of innate immune effectors in the rodent lineage [20]. The nearly exclusive expression of IRGs in the murine system has led some to question its relevance to human infection. However, this view overlooks the obvious importance of rodents in the natural transmission of toxoplasmosis, which is a zoonotic disease that humans acquire from infected food animals and cats, although not directly from rodents. Additionally, there are several reasons to believe IRGs are also directly relevant to humans. Humans express only two IRG family members: IRGC, which is testis specific and unlikely to be involved in immunity, and IRGM, which is truncated [20]. Despite likely not being a functional GTPase, IRGM has been implicated in autophagy-mediated control of Mycobacterium tuberculosis [50] and Salmonella typhimurium [51]. Additionally, both humans and rodents express a second family of related GTPases called guanylate binding proteins (GBPs), which are also strongly upregulated following treatment with IFN-γ [52]. GBPs have recently been shown to be required for the control of Listeria monocytogenes and M. tuberculosis in the mouse [53]. GBPs are also recruited to vacuoles containing T. gondii in a strain-dependent manner [52], [54], suggesting a role in pathogen control. Direct evidence for such a role was recently provided by a study reporting deletion of a locus on chromosome 3 in the mouse, encoding 5 GBPs, impairs immunity to challenge with a type II strain of T. gondii [55]. In separate studies, we have shown that these effects are partially dependent on Gbp1 and that recruitment of Gbp1 to the PVM is mediated in a ROP5 and ROP18-dependent manner (Selleck, Sibley submitted). Homeostasis of GBPs requires Irgm proteins in the murine system [56], hence the dramatic reversal of the ROP5 deficient parasites in Irgm3−/− mice may be due to defects in both the IRG and GBP systems. Further studies will be needed to determine the relationship between these different IFN-γ induced systems and to define the role of ROP kinases and IRG-dependent immunity mechanisms in control of human infection.
Previous genetic crosses have implicated only a few loci in controlling acute virulence in the mouse model [6], [22], [28], [57]. Notably, only type I parasites are efficient at blocking IRG clearance [14], [23] and they have a combination of a type I allele at ROP18 and a type I allele at ROP5, the latter of which is also expressed by type III parasites [7], [27]. Consistent with their extremely low level of ROP18 expression, type III parasites are avirulent, a phenotype that is fully reverted with transgenic expression of type I [7] or type II [6] ROP18. Although type II strain parasites have a functional ROP18, they have a type II ROP5 locus that is associated with avirulence. Although we have not tested the ability of type II ROP5 to regulate the activity of the type II allele of ROP18, genetic studies indicate that this interaction is not sufficient to promote full virulence [28], [57], nor is it sufficient to mediate avoidance of IRG recruitment and clearance [14], [23].
The more severe defect of ROP5 deficient parasites vs. ROP18 deficient parasites in wild type mice, suggest that in addition to regulating ROP18, it has other functions, perhaps serving as scaffold for regulating other important kinases in T. gondii. ROP kinases are highly polymorphic and have expanded in the T. gondii genome under strong selective pressure [8]. Similarly, the IRG pathway is highly amplified in rodents where it plays a major role in resistance to intracellular pathogens such as T. gondii. Placing the ROP5 pseudokinase at the center of this pathway may be an evolutionary strategy to divert attention from the active kinases, in which diversity is constrained to preserve catalytic activity.
All animal experiments were conducted according to the U.S.A. Public Health Service Policy on Humane Care and Use of Laboratory Animals. Animals were maintained in an AAALAC-approved facility and all protocols were approved by the Institutional Animal Care and Use Committee (School of Medicine, Washington University in St. Louis).
CD-1 and C57BL/6 mice were purchased from Charles River Laboratories. Ifnγr1−/−, Rag1−/−, and Nos2−/− mice,, all on a C57/BL6 background were obtained from Dr. Herbert Virgin (Washington University). Mice deficient in the superoxide-generating NADPH-oxidase gp91phox subunit (NOX2), referred to as X-CGD mice, on a C57BL/6 background [58] were obtained from Dr. Mary C. Dinauer (Washington University). Irgm3−/− mice [17] from Dr. Greg Taylor (Duke University) were bred locally at Washington University. Age and sex- matched mice were challenged by i.p. or s.c. injection with parasites and survival followed for 30 days post injection, as described previously [27].
Mice inoculated with luciferase expressing parasites were weighed at intervals after infection and imaged by bioluminescence as described previously [59]. For cytokine measurements, blood was obtained from the saphenous vein. Sera were obtained using microtainer serum separators (BD Bioscience) by centrifugation at 3,000 g and stored at −20°C. IL-12p40 was measured using a mouse OptEIA ELISA kit (BD Biosciences). The other cytokines were quantified using the Cytometrix Bead Array Mouse Inflammation Kit (BD Biosciences), detected using a FACS Canto II flow cytometer (BD Biosciences), and analyzed using FCAP ARRAY (Soft Flow, Inc.).
Fully virulent, type I RH strain parasites that are deficient in Ku80 (RHΔku80), attenuated Δrop18 (RHΔku80Δrop18), avirulent Δrop5 (RHΔku80Δrop5), and virulent ROP5 complemented (RHΔku80Δrop5Complement) parasite strains were described previously [24], [27]. Generation of the ROP18 complement (RHΔku80Δrop18Complement) is described in Figure S4. Parasites were serially passaged in human foreskin fibroblasts (HFF) monolayers, as described previously [24]. Cultures were negative for mycoplasma contamination using the e-Myco plus mycoplasma PCR detection kit (Boca Scientific). Peritoneal macrophages harvested from naïve CD1 mice, bone marrow derived macrophages, and RAW 264.7 cells (American Type Culture Collection (ATCC)) were cultured as described previously [24]. Cells were activated by treatment overnight with either 10 or 50 units/mL murine IFN-γ (R&D Systems) and 0.1 ng/mL LPS from E. coli O55:B5 (Sigma-Aldrich). Gr1+ monocytes were harvested at day 4 from the peritoneal cavity of mice that have been primed 4 days previously by inoculation with 200 CTG strain parasites, as described previously [24]. Macrophages were characterized by expression of the cell surface markers using mAb RB6-8C5 against Gr1 that was directly conjugated to Alexa Fluor 594 and/or mAb HB-198 (ATCC) against F4/80 that was directly conjugated to Alexa Fluor 488, using commercially available coupling kits (Invitrogen).
The pDestR4R3-UPRTKO-Clickluc plasmid was constructed with the MultiSite Gateway 3-fragment pDest-R4R3 system (Invitrogen). Flanking fragments (1 kb 5′ pDONR-P41r and 3′ pDONR-P2rP3) for the T. gondii uracil phosphoribosyl transferase gene (UPRT) gene were amplified from RH strain lysate with iProof High Fidelity DNA polymerase (Bio-Rad). The middle fragment (pDONR-P1P2) contained the Click Beetle luciferase (Clickluc) gene driven by the dihydrofolate reductase (DHFR) promoter as described [60] (Table S1). RHΔku80, RH Δku80Δrop18, or RHΔku80Δrop5 parasites (1), were electroporated with pDestR4R3-UPRTKO-Clickluc linearized with BsiWI, selected with fluorodeoxyuridine (FUDR)(1×10−5 M), and luciferase positive single cell clones were identified by positive bioluminescence.
A 5-fragment Gateway clone was generated to integrate the ROP18 gene under control of the IMC1 promoter into the UPRT locus (Figure S5). The targeting construct was amplified by PCR as described previously [61], and electroporated into RHΔku80Δrop18 strain parasites, stable transformant clones isolated and verified by immunofluorescence staining and western blotting.
Recruitment of cells to the peritoneal cavity was analyzed by FACS, using protocols described previously [59]. In brief, mice were sacrificed at intervals after infection, peritoneal cells were isolated, and stained and analyzed by FACS. Cells were incubated at 4°C in Fc block (clone 2.4G2, BD Bioscience) in MACS buffer (PBS, 0.5% BSA, 2 mM EDTA, pH 7.2) and negative cells excluded by staining with AmCyan-Aqua Fixable Dead Cell Stain (Invitrogen). Labeled antibodies V450 anti-Gr1 (RB6-8C5, BD Bioscience), APC-anti-F4/80 (Invitrogen), PE-Cy7 anti-CD11b (M1/70, BD Bioscience), and APC-eFluor 780 anti-B220 (RA3-6B2, eBioscience) were incubated for 15 min at room-temp, washed and re-suspended in MACS buffer. Samples were detected using a FACS Canto II flow cytometer (BD Biosciences), and analyzed with FlowJo software (Tree Star, Inc). Absolute cell numbers were calculated using the total cell count multiplied successively by the percentages for the appropriate gates obtained through flow cytometry. A total of 100,000 cells were analyzed for each sample.
Infected HFF or macrophage monolayers cultured on coverslips were fixed in 4% formaldehyde and permeablized in 0.05% Triton X-100 in PBS or 0.05% saponin for 10 min. Samples were blocked with 10% FBS, incubated with primary antibodies for ∼20 min, washed 3 times with PBS, and incubated with species-specific secondary antibodies conjugated to Alexa Fluors (Invitrogen) for ∼20 min. Samples were rinsed in PBS, mounted in ProLong Gold with DAPI (Invitrogen) and examined with a Zeiss Axioskop 2 MOT Plus microscope (Carl Zeiss, Inc.). Images were acquired with an AxioCam MRm camera (Carl Zeiss, Inc.) and processed with Photoshop CS4.
Macrophage monolayers were challenged with parasites that were propagated in HFF cells as described above. Clearance was assessed by comparing the percentage of cells infected following a 30 min infection pulse vs. those remaining after 20 h, as described previously [24]. The numbers of infected cells were determined by counting of 10 fields using a 40× objective lens from 3 replicates per condition. Two or three replicate experiments were performed for each assay.
Gr1+ monocytes or IFN-γ treated bone marrow derived macrophages were challenged with parasites for 30 min, fixed in formalin buffered saline, and processed for immunofluorescence, as described previously [24]. T. gondii containing vacuoles were stained with mAb Tg17–113, which recognizes dense granule protein 5 (GRA5) [62]. Irgb6 was localized with rabbit anti-Irgb6 [34], and the numbers of positive vacuoles were determined by counting of 10 fields using a 40× objective lens from 3 replicates per condition. Two or three replicate experiments were performed for each assay.
Parasite lysates were resuspended in denaturing Laemmli sample buffer, resolved in 10% acrylamide gels, transferred to nitrocellulose and probed with rabbit anti-ROP18 [24], rabbit anti-ROP5 [27], or rabbit anti-actin [63]. Blots were washed, incubated with goat anti-rabbit IgG conjugated to HRP (Jackson ImmunoResearch), and detected using the ECL Plus western blotting system (GE Healthcare) and FLA5000 phosphorimager analysis (Fuji Life Sciences).
Immunoprecipitations were performed as described previously [24]. In brief, cells were lysed 1% NP-40, 50 mM Tris–HCl, 150 mM NaCl, pH 8.8 plus protease inhibitors, centrifuged at 1,000 g 4°C, and pre-cleared by incubation with protein G sepharose (Pierce Biotechnology Inc.) for 1 h at 4°C. ROP18 was immunoprecipitated using either the mAb BB2 against the Ty-1 tag, or polyclonal rabbit anti-ROP18. Irgb6 was immunoprecipitated from IFN-γ activated bone marrow derived macrophages with rabbit anti-Irgb6, and purity was confirmed by MS/MS, as described previously [24]. Protein G sepharose was charged with antibodies for 1 h at room temperature, washed with PBS, incubated with cell lysates overnight at 4°C, followed by washing with PBS. The efficiency of ROP18 precipitation was assessed by probing with rabbit anti-ROP18 that was directly conjugated to NHS-biotin (Pierce, Thermo Scientific), washed in PBS, incubated with streptavidin conjugated to HRP, and detected as described above for western blotting.
ROP18-kinase domain (ROP18-KD) was expressed and purified as described previously [10]. Genomic DNA from the type I RH strain of T. gondii was used to amplify the genes encoding full length ROP18 (ROP18-FL) (starting from Glu83 based on the second ATG of GenBank protein CAJ27113) or ROP5 (starting from Val25 of GenBank protein AAZ73240.1) using iProof high-fidelity polymerase (Bio-Rad) and primers listed in Table S1. Amplicons were cloned into pGEX-6P-1 using primers that introduced a C-terminal His6 tag. ROP18-FL was expressed in BL21 (DE3)-V2R-pACYC LamP, as described previously [10]. ROP5-FL was expressed in Rosetta (DE3)pLysS (Novagen). Cells were induced with 1 mM IPTG, grown overnight at 15°C, and soluble proteins were purified using Glutathione Sepharose 4B (GE Healthcare) according to the manufacturer's recommendations. Protein purity and concentration were assessed by SDS-PAGE and SYPRO Ruby staining.
Kinase activity of recombinant or immunoprecipitated ROP18 was tested on the heterologous substrate dephosphorylated myelin basic protein (dMBP) (Millipore) (0.5 µg/reaction), or separately on Irgb6 that was immunoprecipitated from IFN-γ activated RAW cells using rabbit anti-Irgb6. Kinase reactions were conducted in 25 mM Tris–HCl pH 7.5, 15 mM MgCl2 and 2 mM MnCl2. containing 10 µCi of 32P γ-ATP (specific activity: 3,000 Ci/mmol) (Perkin Elmer, Inc) in addition to 33 µM unlabelled ATP (Sigma-Aldrich). Reactions were allowed to proceed at 30°C for 30 min, samples were heated to 95°C in Laemmli sample buffer, resolved on 12% or 10% SDS–PAGE gels, dried, and imaged using a FLA5000 phosphorimager.
Statistical calculations were performed in Excel. Student's t tests were performed under the assumption of equal variance and using a two-tailed test,, where P≤0.05 was considered significant.
Confluent HFF monolayers grown in T75 flasks were infected with 12×106 wild type (RHΔku80) or ROP5 deficient (RHΔku80Δrop5) parasites (MOI of 4), or mock infected and allowed to grow for 24 hours. Cells were washed in PBS lacking divalent cations, removed by trypsinization, washed in DMEM containing 10% FBS, pelleted by centrifugation at 400 g for 10 min, and the pellets stored at −80°C. To extract RNA, pellets were thawed and processed with the Qiagen RNeasy kit supplemented with β-mercaptoethanol and DNase I treatment (Qiagen). Total RNA was processed and labeled into cRNA using the Ambion Message Amp Premier (Ambion) according to the manufacturer's protocol using 500 ng starting RNA. A total of 10 µg cRNA was hybridized to the HG-U113A_2 Affymetrix Human Genome Array (Affymetrix) using standard manufacturer's hybridization and scanning protocols. Data was processed using GeneSpring 7.2 (Agilent Technologies) with the following normalizations: Robust Multi-array Average (RMA), Data transformation: Set measurements less than 0.01 to 0.01, Per Chip and Per Gene: Median Polishing. There was only one gene >2.5 fold different between RHΔku80 and RHΔku80Δrop5 infected host samples, a serine protease inhibitor - BC005224 (Genbank id for the probe set specific for this gene; http://www.ncbi.nlm.nih.gov/nuccore/BC005224) and the variance in the RHΔku80Δrop5 data for this gene resulted in the difference being non-significant. Data was submitted to NCBI GEO record GSE32104.
|
10.1371/journal.pcbi.1000823 | Transat—A Method for Detecting the Conserved Helices of Functional RNA Structures, Including Transient, Pseudo-Knotted and Alternative Structures | The prediction of functional RNA structures has attracted increased interest, as it allows us to study the potential functional roles of many genes. RNA structure prediction methods, however, assume that there is a unique functional RNA structure and also do not predict functional features required for in vivo folding. In order to understand how functional RNA structures form in vivo, we require sophisticated experiments or reliable prediction methods. So far, there exist only a few, experimentally validated transient RNA structures. On the computational side, there exist several computer programs which aim to predict the co-transcriptional folding pathway in vivo, but these make a range of simplifying assumptions and do not capture all features known to influence RNA folding in vivo. We want to investigate if evolutionarily related RNA genes fold in a similar way in vivo. To this end, we have developed a new computational method, Transat, which detects conserved helices of high statistical significance. We introduce the method, present a comprehensive performance evaluation and show that Transat is able to predict the structural features of known reference structures including pseudo-knotted ones as well as those of known alternative structural configurations. Transat can also identify unstructured sub-sequences bound by other molecules and provides evidence for new helices which may define folding pathways, supporting the notion that homologous RNA sequence not only assume a similar reference RNA structure, but also fold similarly. Finally, we show that the structural features predicted by Transat differ from those assuming thermodynamic equilibrium. Unlike the existing methods for predicting folding pathways, our method works in a comparative way. This has the disadvantage of not being able to predict features as function of time, but has the considerable advantage of highlighting conserved features and of not requiring a detailed knowledge of the cellular environment.
| Many non-coding genes exert their function via an RNA structure which starts emerging while the RNA sequence is being transcribed from the genome. The resulting folding pathway is known to depend on a variety of features such as the transcription speed, the concentration of various ions and the binding of proteins and other molecules. Not all of these influences can be adequately captured by the existing computational methods which try to replicate what happens in vivo. So far, it has been challenging to experimentally investigate co-transcriptional folding pathways in vivo and only little data from in vitro experiments exists. In order to investigate if functionally similar RNA sequences from different organisms fold in a similar way, we have developed a new computational method, called Transat, which does not require the detailed computational modeling of the cellular environment. We show in a comprehensive analysis that our method is capable of detecting known structural features and provide evidence that structural features of the in vivo folding pathways have been conserved for several biologically interesting classes of RNA sequences.
| RNA molecules play diverse roles in many of the most basic cellular processes. In the translation process, for instance, the protein coding ‘message’ is encoded in a messenger RNA (mRNA) and transfer RNAs (tRNAs) and ribosomal RNAs (rRNAs) are involved in this catalytic process. Micro RNAs are implicated in regulating mRNA availability. A range of other non-protein-coding RNAs (ncRNAs) have been identified [1], [2]. Moreover, studies of mammalian transcriptomes have found, rather surprisingly, that the majority of the genome is transcribed, and that the vast majority of transcripts do not overlap with known protein-coding regions, hinting at the possibility that many functionally important classes of ncRNAs remain to be discovered [2], [3].
For many classes of ncRNA molecules studied so far, RNA structure plays a crucial part in defining its functional role in the cell. We know, for example, that tRNAs assume a distinct three-dimensional conformation in order to function properly during translation and that the functional configuration of the ribosome complex relies both, on properly folded rRNAs as well as many proteins binding to the respective rRNAs. In contrast to proteins, we can typically learn a lot about an RNA's functionality by studying only its secondary structure, i.e. the set of base-pairing nucleotide positions in the RNA sequence. This is the case because most RNA sequences studied so far fold in a hierarchical manner, with the secondary structure emerging first and the tertiary contacts between secondary structure elements emerging later.
In vivo, an RNA molecule is synthesized during transcription and will immediately start to fold [4], [5]. A succession of cellular events — involving, for example, splicing, RNA editing, the binding of proteins, metabolites or other RNA molecules — may influence the kinetic, co-transcriptional folding pathway in vivo which yields one or more biologically active, i.e. functional structural confirmations.
The view that one RNA sequence has one functional RNA structure turns out to be too simplistic. We know by now of several cases, where a given RNA sequence has more than one functionally important RNA structure, e.g. ribo-switches [6]–[8] which change their structure upon binding a metabolite, as well as cases, where a transient RNA structure is functionally important [9], [10]. We therefore propose to develop a method which allows us to identify evolutionarily conserved structural elements which are likely to be required for the formation of the functional structures in vivo.
There exist by now a wide range of computational methods that can predict an RNA secondary structure given an RNA sequence. Many of these methods [11]–[14] in particular earlier methods, aim to predict the thermodynamically most stable RNA secondary structure. Many biological systems, however, are not in thermodynamic equilibrium. The predictions of these so-called minimum-free energy (MFE) methods depend on the underlying energy parameters which in turn depend on the temperature, the ion concentration and other parameters. Theoretical studies of RNA molecules [15] have shown that the thermodynamic structure of even moderately long RNA molecules often does not correspond to the functional RNA structure that has been conserved during evolution, i.e. the RNA structure that exerts the biological function in vivo. This may, at least partly, be due to co-transcriptional folding [4], [5], [16]. More recent structure prediction methods use a comparative approach which simultaneously analyzes several evolutionarily related RNA sequences from different organisms [17]–[32]. Detailed structural studies employing several dedicated evolutionary models [21] find that the substitution rate in base-paired regions is reduced by a factor of and in loop and bulges by a factor of with respect to the substitution rate in un-structured regions, i.e. that loops and bulges tend to evolve significantly slower than un-structured regions and only slightly faster than base-paired regions at least in set of RNA structures investigated in [21]. This is in line with our expectation that loops and bulges are on average more likely to be bound by other molecules (RNAs, DNAs or proteins) than unstructured regions. These comparative methods aim to detect the RNA secondary structure that has been conserved during evolution. The implicit assumption made by these methods is that evolutionarily conserved structures are likely to be functionally important which has been shown to be a reasonable assumption. The performance of these comparative methods is – generally speaking – higher than that of non-comparative methods [33] provided the input data are high-quality multiple-sequence alignments or the method is capable of generating a multiple-sequence alignment as part of its predictions [27]–[30], [32]. All of the above computational methods, however, only aim to predict a single RNA structure and cannot be used to detect the presence of transient RNA structures or the presence of multiple functional RNA structures such as, for example, ribo-switches which are known to have two distinct functional structures.
The program RNAsubopt [34] takes a single RNA sequence as input and predicts a list of all structures below a certain energy cutoff. Enumerating enough structures to capture most of the structure probability, however, is only possible for short sequences. Moreover, since the total number of possible structures is so vast, the probability of any particular structure is not a reliable indicator for identifying potential alternative structures. Rather, one would like to group similar structures together, and identify groups with a high overall probability. Voss et al. [35] formalize this grouping process by defining abstraction functions in order to map structures to ‘RNA shapes’, and are capable of calculating the total probability for a given shape. The runtime for this method grows exponentially with sequence length, making it impractical for sequences longer than about 400 nucleotides [36]. It is possible, however, to sample structures from the Boltzmann distribution in polynomial time [37], and to then apply the RNA shapes abstraction in order to estimate the shape probability for longer sequences. This approach is capable of recovering alternative structures for some ribo-switches [35]. All of the above approaches assume the RNA sequence to be in thermodynamic equilibrium and are thus limited to identifying alternative structures which occupy a significant portion of the Boltzmann distribution. For co-transcriptionally folding RNA sequences (which may become kinetically trapped), this assumption does not necessarily hold and the time-averaged probabilities for different structural configurations encountered during the kinetic folding may differ markedly from their respective probabilities derived from the Boltzmann distribution.
In vivo, RNA molecules are known to fold co-transcriptionally [4], [5], i.e. while they emerge during transcription. The resulting kinetic folding pathway can depend on a variety of events during and after transcription such as the speed of transcription [9], [38], [39], splicing [40], RNA editing [41], the binding of proteins [42], metabolites [43] and other RNA molecules [44], the temperature and the concentrations of monovalent and divalent ions [45]. The co-transcriptional folding pathway can differ significantly from the re-folding one [46], [47], both in terms of time line and structural features.
The increasing interest in RNA folding pathways has spurred the development of computational methods for RNA structure prediction which take the folding kinetics explicitly into account. These methods try to model the physical process by which an unfolded RNA folds into its functional conformation(s) as a continuous-time Markov process which allows only local rearrangements of secondary structures. If we knew all entries of the transition rate matrix containing the transition rates between all pairs of possible structures, the vector of probabilities for all structures at a given time could be calculated as . As the state space of all possible secondary structures can be very large for RNAs of biological interest, it is generally not feasible to calculate the full transition matrix. However, folding trajectories can be sampled using Monte Carlo stochastic simulation of the Markov process. Several programs, including RNAkinetics [48]–[50], Kinfold [51] and Kinefold [52]–[54], employ this method, though they differ significantly in their implementation.
Mironov and Lebedev [49] were the first to model the co-transcriptional folding of an emerging RNA sequence and to allow entire helices not only to form, but also to disintegrate [55]. The transition probabilities of their Markov chain Monte-Carlo method correspond to the chemical rate constants for forming and disintegrating helices [48] and thus have a clear physical interpretation. Their theoretical framework could be readily extended to also deal with pseudo-knotted RNA secondary structures [49].
Kinfold [51] defines legal transitions as the formation, disruption, or shifting of a single base-pair. The folding trajectories it generates are therefore very fine-grained, specifying when each base-pair is added or removed. In Kinefold [52]–[54], transitions add or remove entire helices, a simplification which reduces the number of legal transitions from any state, but which also requires a more complex estimation of the transition state energy. The program assumes that the energy barrier is the energy required to nucleate three base pairs of a new helix, plus the energy required to displace any helices blocking the formation of the new helix. Kinefold also allows for pseudo-knotted structures, which requires a more complicated energy model than the standard Turner model [56] used by Kinfold (which ignores pseudo-knots). Kinefold also takes into account some topological constraints induced by pseudo-knots which may kinetically trap other helices [54]. Both programs can simulate the folding from an unfolded state as well as the co-transcriptional folding of an emerging sequence. The latter is done by dividing the sequence into transcribed and un-transcribed regions whose boundary shifts 5′ to 3′ at a certain rate, and restricting legal moves to those that form no base-pairs in the un-transcribed region. Neither of these two programs can model dynamic transcription speeds, although there is experimental evidence that transcriptional pausing influences the folding [57].
Other computational approaches for predicting kinetic folding pathways consider energy landscapes in order to reduce the size of the state-space. The energy landscape can be viewed as a barrier tree, where the local minima are leaves in the tree which are connected to one or more gradient basins via saddle-points. Saddle-points are the lowest energy structures that connect the gradient basins around these local minima [51], [58]. Constructing such a barrier tree representation of the energy landscape requires the consideration of all possible structures. Barrier trees constructed from a list of the lowest-energy structures (generated with RNAsubopt [34]) typically capture the most relevant features of the energy landscape for sufficiently short sequences ( base pairs (bp)). In order to reduce the state space, Wolfinger et al. [59] define the state-space as the basins around local minima of the energy landscape, and calculate the transition rates between adjacent basins using a variation of the so-called flooding algorithm used to construct barrier trees. Barrier trees are also useful for interpreting folding trajectories sampled with Monte Carlo simulations [51]. A similar approach is taken by Tang et al. [60], [61], where the folding landscape is approximated by a probabilistic road map which defines the allowed transitions between states. They restrict the state-space to a set of secondary structures probabilistically sampled from the Boltzmann distribution. Transitions are only allowed to the nearest neighbors, with energy barriers estimated heuristically. Ideally, these states should capture the main features of the folding landscape while being few enough to solve the master equation (though it is also possible to do Monte Carlo simulation here). Zhang and Chen [62]–[64] partition the structure space into clusters based on the presence or absence of certain (somewhat arbitrarily chosen) rate-limiting base-stacks, which have particularly high energy barriers to their formation or disruption. The distribution of structures within clusters is assumed to be at thermodynamic equilibrium, so the transition rates between clusters can be calculated by summing the rates of transition between the structures at the boundaries of clusters, adjusted for the probability of the boundary structure in its cluster. All of these thermodynamic-landscape-based methods, however, are not applicable to the analysis of co-transcriptional folding, since an RNA's energy landscape changes while it is being transcribed. By calculating energy landscapes for all partially transcribed subsequences and then mapping the local minima from each landscape onto its successor, however, one could – in theory – adapt landscape-based methods to co-transcriptional folding [65].
Long sequences are problematic for all the above methods since the number of possible secondary structures, and therefore the worst-case complexity of the energy landscape, grows exponentially with the sequence length. The Kinwalker program [66] was designed to allow the analysis of the folding kinetics for long sequences (around 1000 bp). For this, it dispenses with simulation and instead deterministically predicts a potential co-transcription folding pathway which is pieced together from heuristically chosen combinations of pre-computed minimum free-energy (MFE) structures for short sub-sequences and assumes (similar to MFE methods for RNA structure prediction) the pseudo-knot free MFE structure to be the final RNA structure. The method can be considered kinetic in that it allows the incorporation of an MFE sub-structure only if the energy barrier between the current structure and the resulting merged structure that the transition can occur within a reasonable time, i.e. before the next transcription step. Calculating the energy barrier between two arbitrary structures, however, has been shown to be NP-complete [67]. Kinwalker thus employs a further heuristic for estimating these barriers. In summary, Kinwalker aims to find the MFE structure at each transcription step, subject to the constraint that the transitions between structures be kinetically feasible.
All of the above prediction methods take at most the RNA sequence itself, the temperature, the concentration and a constant transcription speed into account, but do not capture any potential interactions with other molecules or other features of the biological environment which may influence the folding pathway in vivo. The latter is difficult to do, not only because we typically lack information on the interaction partners and the mechanisms and timing of their interactions, but also because we cannot easily capture the wealth of relevant details of the complex cellular environment in a computationally tractable model. The performance of the existing computational methods can strongly depend on the sequence length and other features of the individual input sequence. This is not surprising given that any errors in the early stages of the folding pathway prediction are magnified as the folding progresses. A precise knowledge of the transcription start site i.e. the 5′ end of the RNA sequence is thus crucial. The prediction performance of the existing methods has thus only been evaluated on very small data sets.
It is also challenging to study kinetic folding pathways experimentally. There exist by now a range of powerful experimental techniques for studying large sets of RNA sequences in an ensemble-averaged way such as UV melting, isothermal titration calometry, circular dichroism, chemical foot-printing and, more recently, single-molecule techniques such as fluorescence correlation spectroscopy [68], single-molecule fluorescence resonance energy transfer [69] and force spectroscopy [70], [71]. These experimental methods, however, still await to be taken from the test tube to the cell in order to explore how RNA sequences fold in vivo [72].
We propose a conceptually new computational approach for studying RNA folding pathways in vivo. Rather than trying to replicate the folding kinetics of a single RNA sequence in vivo — which is very difficult to do — we introduce a comparative approach which takes several evolutionarily related RNA sequences as well as an evolutionary tree relating these sequences as input. Our main goals in devising Transat can be summarized as follows:
Our data set comprises four sub-sets which have been chosen to represent (a) data, where multiple functional RNA secondary structures are known, (b) data, where only one functional RNA secondary structure is known and (c) artificially generated data which allows us to investigate some features of Transat in greater detail. Our aim was to compile a large and diverse data set and to include as many examples of known functional and transient RNA structures as possible. Taken together, our data set comprises 1126 multiple-sequence alignments whose lengths ranges from 100 to 1247 bp and which comprise between 6 and 712 sequences.
The performance of new prediction methods is best benchmarked by comparing the set of predicted to the set of known structure for an, ideally, large and diverse data set that has been carefully and completely annotated (the test set). If the prediction method depends on free parameters, these parameters should have been trained or manually derived from a training set which should have no overlap with the test test (and which has to be large and diverse enough to minimize the risk of parameter over-fitting). Typically, training and test sets are permuted several times in cross-evaluation experiments in order to show that both, the parameter training and the resulting performance are fairly independent of the particular choice of training and test set. This careful benchmarking is comparatively easy to accomplish for some applications, e.g. methods for RNA secondary structure prediction, but more difficult for others.
The conclusive benchmarking of computational methods for predicting kinetic folding pathways has, so far, been difficult. This is due to several reasons. First, detailed experimental results on folding kinetics, usually done via temperature- or pH-jump kinetic trapping procedures [95] or single-molecule ‘optical-tweezer’ manipulation [96], are only available for a small number of sequences which are typically quite short ( bp) and may, moreover, correspond to artificial sequences. Second, the assumptions made explicitly or implicitly by the prediction methods may not apply to the experimental setting. Third, there are no standard metrics for comparing experimental results with output from computational prediction methods (whose type of output varies greatly from method to method). Fourth, many computational methods (especially more heuristic ones [66], [97]) rely on a number of free parameters which require a dedicated test set in order to train them reliably and to avoid overlap with the test set. Consequently, most methods for predicting the RNA folding kinetics have been evaluated via a qualitative rather than quantitative comparison and only by considering a few chosen experimentally investigated sequences.
Transat has been devised to detect conserved RNA helices of statistical significance. Using Transat, we thus hope to not only detect the helices of the known functional RNA structure, but also new helices of functional importance which may be involved in defining the RNA's folding pathway in vivo.
We devised Transat as a method to detect the statistically significant, conserved helices of functional RNA structures, including the helices of transient, pseudo-knotted and alternative structures as they are known to exist in vivo. As we explain in detail in the introduction, it is currently not possible to model the kinetic folding of RNA structures in vivo as function of the time as we not only lack many crucial details on the cellular environment that may influence the folding pathways (i.e. which molecules bind the RNA sequence in question when and where), but also because we currently have no adequate theoretical framework that would allow us to efficiently simulate the complex cellular environment using computational methods. We circumvent these conceptual problems by devising Transat as a comparative prediction method which takes a fixed multiple-sequence alignment of homologous RNA sequences and a tree quantifying their evolutionary relationship as input and detects evolutionarily conserved helices and estimates their statistical significance. By employing this comparative approach, we lose the ability to predict structural features of the cellular folding pathway(s) as function of the time and to detect species-specific structural features which may also be functionally important, but gain the ability to highlight statistically significant, functional helices that have been conserved without actually having to model the cellular environment nor its evolution over time.
Our comprehensive performance evaluation of Transat for a large and diverse data set (comprising 1126 multiple sequence alignments ranging from 100 to 1247 bp and comprising between 6 and 712 sequences) shows that Transat not only reliably detects the helices of known unique RNA reference structures, but that it also able to capture known pseudo-knotted structures as well as known alternative structural configurations. In addition to these known structural features, Transat predicts a number of distinct, novel helices of statistical significance. These may, for example, correspond to well-conserved structural features of a co-transcriptional folding pathway in vivo supporting the notion that homologous RNA sequence not only assume similar functional RNA structures, but also fold in a similar way. For some examples, the additional helices suggest a pseudo-knotted functional configuration, where only a pseudo-knot free RNA structure has been annotated so far. As we show for two examples, the predictions by Transat can also help identifying regions of an RNA sequence that are bound by other molecules and thus single-stranded because these are regions which are devoid of statistically significant helices. Detailed investigations show that Transat's predictions are robust with respect to alignment errors and modifications of the input tree and that its performance is fairly independent of the alignment length. Transat's performance is more correlated with the length of the input tree which is not surprising given that a certain degree of evolutionary diversity is required to observe pairs of co-varying alignment columns, where the base-pairing potential, but not necessarily the nucleotides forming the base-pairs has been conserved. We also find that the dominant structural features predicted by Transat typically do not coincide with those of the Boltzmann distribution of (pseudo-knot free) RNA secondary structures if we assume thermodynamic equilibrium. In particular, we find that the presence of known pseudo-knotted reference structures and of known alternative, functional RNA structures cannot be inferred from the Boltzmann distribution, i.e. by assuming thermodynamic equilibrium. This discrepancy may be partly attributed to the fact that the Boltzmann distribution does not include pseudo-knotted RNA structures, but overall confirms our expectation that there is a priori no good reason to assume that RNA sequences in vivo are in thermodynamic equilibrium or unbound by other molecules.
The Transat software is available from people.cs.ubc.ca/∼irmtraud/transat/. This web-page also contains information on the input and output files of this analysis as well as detailed documentation on how to use Transat. Users of Transat can rank the predicted helices according to their p-values with lower values implying higher statistical significance. Lab scientists seeking to confirm specific helices in dedicated experiments can prioritize their experiments by starting with the statistically most significant helices.
We hope that the predictions by Transat will enable more comprehensive and systematic studies of RNA folding pathways and alternative structural configurations and that they will provide useful input to the design and interpretation of future experiments. Whether the near future will bring more experimental insight into how RNA sequences fold in vivo depends to a large extent on the development of new experimental techniques that would allow us to observe an RNA sequence in its cellular environment.
Transat currently focuses on highly conserved structural features that are statistically significant, but ignores those that are functional, but only present in a small fraction of the input sequences. One possibility for future work is thus to extend Transat in order to also capture structural features that are only present in a few of all input sequences. As Transat already explicitly models the evolutionary relationship between all input sequences and the evolution of unpaired and base-paired nucleotides, this should be relatively straightforward to do. Another, more challenging possibility for future work is to take Transat beyond the required fixed input alignment. This is partly what the program SimulFold [32] addresses, but would need to done for individual helices and complemented by a corresponding procedure for estimating p-values.
|
10.1371/journal.ppat.1007605 | Cyclical adaptation of measles virus quasispecies to epithelial and lymphocytic cells: To V, or not to V | Measles virus (MeV) is dual-tropic: it replicates first in lymphatic tissues and then in epithelial cells. This switch in tropism raises the question of whether, and how, intra-host evolution occurs. Towards addressing this question, we adapted MeV either to lymphocytic (Granta-519) or epithelial (H358) cells. We also passaged it consecutively in both human cell lines. Since passaged MeV had different replication kinetics, we sought to investigate the underlying genetic mechanisms of growth differences by performing deep-sequencing analyses. Lymphocytic adaptation reproducibly resulted in accumulation of variants mapping within an 11-nucleotide sequence located in the middle of the phosphoprotein (P) gene. This sequence mediates polymerase slippage and addition of a pseudo-templated guanosine to the P mRNA. This form of co-transcriptional RNA editing results in expression of an interferon antagonist, named V, in place of a polymerase co-factor, named P. We show that lymphocytic-adapted MeV indeed produce minimal amounts of edited transcripts and V protein. In contrast, parental and epithelial-adapted MeV produce similar levels of edited and non-edited transcripts, and of V and P proteins. Raji, another lymphocytic cell line, also positively selects V-deficient MeV genomes. On the other hand, in epithelial cells V-competent MeV genomes rapidly out-compete the V-deficient variants. To characterize the mechanisms of genome re-equilibration we rescued four recombinant MeV carrying individual editing site-proximal mutations. Three mutations interfered with RNA editing, resulting in almost exclusive P protein expression. The fourth preserved RNA editing and a standard P-to-V protein expression ratio. However, it altered a histidine involved in Zn2+ binding, inactivating V function. Thus, the lymphocytic environment favors replication of V-deficient MeV, while the epithelial environment has the opposite effect, resulting in rapid and thorough cyclical quasispecies re-equilibration. Analogous processes may occur in natural infections with other dual-tropic RNA viruses.
| Key questions in infectious disease are how pathogens adapt to different cells of their hosts, and how the interplay between the virus and host factors controls the outcome of infection. Human measles virus (MeV) and related animal morbilliviruses provide important models of pathogenesis because they are dual-tropic: they replicate first in immune cells for spread through the body, and then in epithelial cells for transmission. We sought here to define the underlying molecular and evolutionary processes that allow MeV to spread rapidly in either lymphocytic or epithelial cells. We discovered unexpectedly rapid and thorough genome adaptation to these two tissues. Genome variants that cannot express functional V protein, an innate immunity control protein, are rapidly selected in lymphocytic cells. These variants express only the P protein, a polymerase co-factor, instead of expressing P and V at similar levels. Upon passaging in epithelial cells, V-competent MeV genome variants rapidly re-gain dominance. These results suggest that cyclical quasispecies re-equilibration may occur in acute MeV infections of humans, and that suboptimal variants in one environment constitute a low frequency reservoir for adaptation to the other, where they become dominant.
| RNA virus populations are quasispecies. Quasispecies, also known as mutant spectra, clouds or swarms, are genome distributions that are generated upon replication of RNA viruses in infected cells and organisms [1]. Quasispecies can adapt to dynamic environments and evade selective pressures exerted by antibodies or antiviral drugs [2]. Next-generation sequencing, which greatly expands the capacity to capture low frequency variants within virus quasispecies is beginning to reveal the mechanisms driving mutant spectra adaptation during some RNA virus infections, including those of HIV and HCV [3, 4]. Inter-host adaptation of RNA viruses, for example of arboviruses to arthropods and vertebrates, or of influenza viruses to birds and mammals, are well characterized [5, 6], but insights about genetic diversity and adaptation of viruses that replicate sequentially in two tissue niches of the same host are rare.
MeV provides an important model of pathogenesis due to its dual-tropic nature: it replicates first in lymphatic tissues and then in epithelial cells. Receptors determine MeV tropism [7]. After contagion, the signaling lymphocyte activation molecule (SLAM) [8] mediates MeV entry in alveolar macrophages and dendritic cells that ferry the infection through the airway epithelium and spread it to local lymph nodes [9]. MeV then spreads in immune tissues, causing immunosuppression [10, 11]. Immune cells deliver the infection to columnar epithelial cells that express nectin-4, the MeV epithelial receptor [12, 13]. Nectin-4 expression in the upper airway epithelia accounts for efficient MeV replication at a location facilitating extremely efficient contagion [7].
MeV is a negative strand RNA virus of the genus Morbillivirus in the family Paramyxoviridae [14]. Morbillivirus genomes are organized into six contiguous, non-overlapping transcription units separated by three untranscribed nucleotides and coding for eight viral proteins, in the order (positive strand): 5’-N-P/V/C-M-F-H-L-3’ [15]. The second transcription unit codes for two non-structural proteins, C and V, that are expressed in non-traditional ways. C is translated from an alternative reading frame accessed from a downstream internal start codon [16]. V is translated from mRNAs in which the viral polymerase inserts one pseudo-templated guanosine after a highly conserved poly-purine stretch, a process of co-transcriptional RNA editing that results in translation of a unique cysteine-rich 68-amino acid carboxyl-terminal domain [17]. Both V and C interfere with the host immune response [18–21].
MeV populations are quasispecies. An early analysis of infections of HeLa cells with a vaccine-lineage MeV estimated the intra-population diversity at 6–9 positions per genome [22]. However, this and many other studies of MeV biology are based on vaccine-lineage MeV strains that, during the attenuation process, were adapted for growth in stable cell lines of disparate origin. Insights about the quasispecies composition of a wild-type MeV replicating in lymphocytic or epithelial cells, the most relevant cell types for infection, are not available.
To address this gap in knowledge, we performed cell-specific adaptation studies. We discovered that MeV genomes that cannot express functional V protein are rapidly selected in lymphocytic cells. Upon passaging in epithelial cells, V-competent MeV genomes rapidly re-gain dominance, indicating that suboptimal variants in lymphocytes can serve as a low frequency reservoir of alleles for adaptation to epithelial cells.
We sought to obtain virus inocula of consistent quasispecies composition, and to passage them in set cellular environments. Towards obtaining consistent inocula, we generated MeV from a cDNA copy of its genome, using standardized procedures based on the overlay of “rescue” 293-derived cells with Vero-hSLAM cells [23]. In particular, we operated with the infectious cDNA from the Ichinose-B (IC-B) strain that has been derived from a wild-type virus [24] and was extensively used for pathogenesis studies in primates [25–29]. To facilitate monitoring the progression of infection during passaging, we used a virus expressing a reporter protein, MeV-IC323-mCherry. We chose lymphocytic Granta-519 cells [30, 31] and lung epithelial H358 cells [12, 32] as model environments.
To assess whether MeV adapts to these environments, we passaged the MeV-IC323-mCherry inoculum (passage 1, p1) 14 times in either lymphocytic Granta-519 cells, or in airway epithelial H358 cells. We used a multiplicity of infection (MOI) of 0.1 for the first infection, followed by inoculations with 20% of the cell-associated inoculum for each subsequent infection. This strategy, while based on variable MOIs across passaging, allowed to maintain sufficient levels of infectious output (S1 Fig, top panel). The passaged viruses were named L14 and E14, respectively.
An indication of adaptation to Granta cells came from a growth kinetics analysis of L14 in either Granta (Fig 1A) or H358 (Fig 1B) cells. In lymphocytic Granta cells, the L14 virus replicated to higher titers than either p1 or E14 (Fig 1A). In contrast, in H358 cells all three viruses reached similar titers (Fig 1B). These data suggest that significant changes in quasispecies composition may have occurred during passage on Granta cells.
This experiment was part of a plan to document the kinetics of quasispecies adaptation to different cell types (Fig 2A). In addition to lymphocytic (L) and epithelial (E) adaptations, the plan included sequential (S) adaptation, as a model of MeV replication within a host. Sequential adaptation consists of 7 passages on lymphocytic cells, followed by 7 passages on epithelial cells. We sought to obtain sequence information on the inoculum (p1), lymphocytic-adapted virus (L7 and L14 passages), epithelial-adapted virus (E7 and E14 passages), and sequentially adapted virus (S14 passage).
To focus our analyses on MeV genomic RNAs, we purified encapsidated genomes (ribonucleocapsids, RNP) of p1, L7, L14, E7, E14 and S14 by isopycnic centrifugation. The sequence of the purified nucleic acids was then analyzed by RNAseq. We obtained 1.5–3 million MeV specific reads for each virus RNP preparation (S2A Fig). Average coverages exceeded 10,000 reads per nucleotide over the length of all six genomes, and coverage never dropped below 1,000 reads per nucleotide (S2B Fig). S1 Table lists all alleles in any sample that differ by >10% from the reference sequence.
We initially focused on L14 because its replication kinetics suggests adaptation. Fig 2B (top) visualizes all variants detected at >10% frequency that were expanded selectively in lymphocytic cells. Strikingly, several of these variants are located near the middle of the P gene (Fig 2B, colored arrows) surrounding the G-insertion site. In particular, substitutions at positions -10, -9, -7, and +1, as numbered from the G-insertion site (position 2499 on the MeV genome), were represented at a higher level in the genomic population (Fig 2B, bottom). At position +1, two different variants were enriched: +1(A) and +1(G). Altogether, the five editing-proximal variants accounted for about 90% of the reads covering this site (Fig 2B, top). Notably, no read with more than one of these mutations was detected. This suggests that five different genomes with a single point mutation near the editing site were positively selected during lymphocytic passaging. In contrast, in epithelial-adapted virus E14, no editing site-proximal variants, but two variants present at >10% frequency were detected (Fig 2C). Both variants resulted in F protein amino acid changes. In addition, three non-cell-specific M gene variants were detected at different levels in all six analyses (S1 Table).
We then analyzed the evolution of the editing site-proximal sequences during quasispecies adaptation. We noted that in L7 and L14, the total editing-proximal variant pool accounted for about 90% of the genomes (Fig 2D). However, the pool composition changed: the -10 variant displaced the -9 variant between passages 7 and 14; the -7 variant, present at low levels in L7, faded into background by passage 14; yet the +1(G) variant remained relatively constant. On the other hand, no mutations surrounding the editing site were selected in epithelial-adapted MeV (Fig 2D, E7 and E14). Strikingly, re-adaptation to epithelial cells resulted in elimination of the editing-proximal variants (Fig 2D, S14). These data suggest that opposing selective pressures are exerted on RNA editing in lymphocytic and epithelial cells.
We then sought to assess whether the editing site-proximal mutations impacted the efficiency of G-nucleotide insertion in P mRNA. Towards this, we infected HeLa-hSLAM cells with p1, L7, L14, E7, E14 and S14 viruses, purified their mRNAs, amplified the relevant P gene segment, and performed dideoxy-sequencing. Fig 3 (left panel) shows the p1 analysis: a homogeneous sequence extending over the conserved poly-purine tract AAAAAGGG becomes heterogeneous after the G-insertion position (vertical dotted line). The similar height of the G- and C-signals at this position indicates an approximately 1:1 ratio of edited (G) and unedited (C) transcripts.
In contrast, in lymphocytic cell-passaged viruses (L7 and L14), editing efficiency was strongly reduced. In Fig 3 (upper two panels), the sequence of L7 shows heterogeneity at positions -9 and +1, as expected from the corresponding RNAseq data, and reduced RNA editing, indicated by the smaller secondary peaks from position +2 onwards. Similarly, the sequence of L14 shows the expected -10 and +1 heterogeneity, and reduced RNA editing. Since the +1 variants are downstream of the G insertion site, which complicates interpretation of the chromatograms, we performed a complementary analysis with a reverse primer that confirmed the above conclusions (S3 Fig, top three panels).
Contrastingly, in epithelial cell-passaged viruses editing efficiency remained near 50% (Fig 3, bottom two panels E7 and E14). Re-adaptation of L7 virus to epithelial cells restored efficient RNA editing (Fig 3, middle row, panel S14). These data directly confirm that opposing selective pressures are exerted on RNA editing in lymphocytic and epithelial cells.
We then assessed whether reduced P mRNA editing negatively impacts V protein expression. Towards this, we infected cells with p1, L7, L14, E7, E14 and S14 viruses, extracted their proteins, and estimated the relative abundance of P and V by immunoblot. The analyses of Fig 4A show that in p1 the V signal was stronger than the P signal. In contrast, the V and P signals in L7 and L14 were of similar intensities. The V signal of the S14 virus, which was re-adapted to epithelial cells, was stronger than the P signal, as were the V signals of the viruses exclusively passaged on epithelial cells. Fig 4B shows quantification of signal strength in three repeat experiments. The P:V ratio of p1 was set at 1. This ratio was between 6 and 8 for L7 and L14, while there was no statistical significance between p1 and the epithelial-adapted or sequentially passaged MeV. Thus, lymphocytic cell-adapted viruses express proportionately less V protein.
To assess whether selection of V-deficient MeV variants is reproducible, we infected the same lymphocytic (Granta) and epithelial (H358) cell lines with MeV-IC323-mCherry-uN. This virus expresses the two polymerase subunits P and L in the same ratio as the parental, non-recombinant, MeV because the added transcription unit with mCherry is inserted upstream of the N gene, rather than downstream of H. As in the first experiment, we passaged this virus either 14 times on lymphocytic cells, or 14 times on epithelial cells at an initial MOI of 0.1 This was followed by inoculations with a fraction (10%) of the cell-associated inoculum for each subsequent infection, which allowed to maintain sufficient levels of infectious output (S1 Fig, bottom panel). We then performed RNAseq on purified viral RNP, and identified genomic variants represented at or above the 10% level in any passage examined (S2 Table).
Fig 5A illustrates the results of this passaging experiment. This time we examined viral genomes at an early time point after passage 1 (L1-2nd or E1-2nd) and again after passage 14 (L14-2nd or E14-2nd). In passage L1-2nd, we noted a significant over-representation of the +1(G) substitution (4.4% allelic frequency), and of the -9 variant (2.1% allelic frequency). At the end of lymphocytic selection (L14-2nd), the -9 variant approached 90% of the population, and in combination the editing site-proximal variants accounted for nearly 93% of the population. In contrast, after either 1 or 14 epithelial passages (E14-2nd), editing site-proximal variants accounted for about 2% and less than 1% of the population, respectively.
Other mutations were selected in the repeat passaging experiment. In L14-2nd, one M gene variant and the H gene variant causing the amino acid change L526F were found in >90% of the reads, and one N gene variant in about 14% of the reads (Fig 5B). In contrast, in E14-2nd two M and one L gene variants accumulated in 10–30% of the reads. Since these mutations were not enriched in the first passaging experiment, they could be adventitious early events that have no impact on adaptation. On the other hand, in the first experiment the H protein variant L526W emerged, suggesting reproducible selective advantage for an aromatic residue at H position 526. Interestingly, L526 is adjacent to a hydrophobic groove relevant for receptor interactions [33].
In combination, the results of the first and second experiment indicate that the key adaptation to the Granta cell environment is the selection of V-deficient mutants. Selection is fast, and independent of the ratio of P and L protein expression. On the other hand, different types of V-deficient genomes are selected in repeat experiments.
To assess whether V-restriction occurs in more than one cell line, we passaged the same virus used for the first Granta cells adaptation experiment in other lymphocytic cell lines, Raji and JVM-2. After two passages in Raji cells the +1(G) variant became dominant, and by passage 6 the wild-type C allele faded into background (Fig 6). A complementary analysis with a reverse primer confirmed this conclusion (S3 Fig, bottom panel). Interestingly, editing efficiency steadily decreased over time, but lagged selection of +1(G) (Fig 6). This indicates that editing efficiency is not uniquely determined by the +1(G) mutation. Dideoxy-sequencing confirmed that the P gene mRNA of these passages was identical to the reference sequence and no detectable minor alleles were present. Conversely, RNA editing efficiency remained constant after MeV passaging in JVM-2 cells, even after 14 or 23 passages (Fig 6). Thus, V restriction occurs in at least two lymphocytic cell lines.
To assess by which mechanisms the editing site-proximal variants impact V expression, we generated viruses differing from parental MeV-IC323-mCherry by a single base. The genomes of these viruses bear the following substitutions: -10, -9, -7 and +1(G) (Fig 7A, top to bottom). We note that the +1(G) virus does not contain a G insertion, rather a C-to-G substitution at position 2499, which is the “+1” position, counting from the editing site. Three of these mutations result in the amino acid changes listed above the corresponding P genes, whereas the -9 mutation is silent.
We then assessed whether the four recombinant viruses have similar characteristics as the parental virus with a growth assay based on permissive Vero cells expressing the MeV receptor SLAM (Vero-hSLAM). Fig 7B indicates that, while the editing site-proximal variants grow to higher titers than wild type at an earlier time point, the wild-type MeV eventually reaches the highest titers.
We then assessed whether each mutation impacts RNA editing, and to which extent. As done previously with the passaged virus mixtures, RNA was extracted from HeLa-hSLAM cells infected with the editing-site proximal viruses, the relevant P gene segment amplified, and dideoxy-sequencing performed. The chromatograms of Fig 7C (second to fourth panel from top) indicate that in mutants -10, -9, and -7, RNA editing was not detectable (no secondary peaks after the dotted line). In contrast, in mutant +1(G) (bottom panel) editing was only slightly reduced as compared to the parental virus (top panel).
To assess whether mutations elsewhere in the genome could impact editing, we generated a recombinant virus with one standard and one mutated P gene (S4 Fig, top drawing). We then determined the editing efficiency of both P gene copies. We confirmed that only the mutated P gene copy had reduced editing capacity (S4 Fig, bottom). Thus, editing site-proximal variants directly govern editing efficiency.
We also characterized the impact of each mutation on V protein expression. For this, proteins were extracted from HeLa-hSLAM cells infected with the four viruses and expression levels of P and V compared to those of the viral N protein, and with cellular actin. The immunoblots of Fig 7D indicate that in mutants -10, -9, and -7, V protein expression was strongly reduced compared with that of the parental virus, whereas in mutant +1(G) V protein expression was maintained. We then quantified signal strength in three repeat experiments (Fig 7E). Relative to the parental P-to-V expression ratio, the -10, -9, -7, and +1(G) expression ratios were 15, 22, 55 and 2, respectively. These results are consistent with the levels of V mRNA expressed by the respective viruses.
These results indicate that the +1(G) virus edits P mRNAs and expresses V protein efficiently. However, the corresponding H232D mutation may inactivate V protein function: histidine 232, which together with three cysteines coordinates a Zn2+ ion, is essential for innate immunity interference by the V protein of MeV and other paramyxoviruses [34].
Both lymphocytic adaptation experiments yielded 10-to-1 mixtures of V-deficient and V-competent genomes. To model what may occur when genome mixtures are transferred to host epithelial cells, we performed a competition assay. We inoculated H358 cells with an excess of V-deficient virus mixed with either 1%, 3%, 10% or 30% V-competent wild-type virus at a MOI of 0.1. The -9 variant was used as a V-deficient genome model. We followed how the two genomes competed by purifying mRNA from infected cells, amplifying a P gene segment, and performing dideoxy-sequencing.
We compared mRNA frequencies after 1, 2, or 3 passages of V-deficient variants either alone (top chromatograms), or mixed with increasing amounts of wild-type genomes (second to fifth row) (Fig 8). The relative amounts of the genomes are proportional to the heights of the C (V-deficient) or T (wild type) signals at position -9 (asterisks). After one passage, the fraction of wild-type genomes in the four mixtures increased, reaching about 10%, 25%, 50% and 70%, respectively (Fig 8, left column). At passage 2, V-deficient genomes were in the minority in all mixtures (Fig 8, right column). Consistently, at passage 2 secondary peaks due to RNA editing were prominent in all mixtures (positions downstream of dotted line). At passage 3, wild-type genomes constituted more than 90% of the population even in the lowest initial dilution (1%). On the other hand, in pure V-deficient virus infections neither at passage 3 (top row, right column) nor at passage 7 (not shown), V-competent genomes were detected. Thus, provided that they constitute at least 1% of the initial population, V-competent genomes rapidly out-compete V-deficient genomes in H358 epithelial cells. Taken together, these data demonstrate that the sequential adaptation of the MeV genome to lymphocytic and epithelial cell lines results in cyclical selection of V-deficient and V-competent genomes.
Seeking to characterize processes that facilitate MeV adaptation to its two cellular environments, we discovered fast and thorough quasispecies re-equilibration. Our observations beg the question of why such a striking phenomenon was not previously described. Previous studies of tropism have mainly been focused on the attachment protein, rather than being approached with an unbiased genetic method. Another important consideration is that the standard procedure for MeV isolation relies on nasal secretions, and thus yields virus of epithelial origin. These viruses are isolated and grown on interferon-defective Vero cells expressing the primary MeV receptor human SLAM [35], the same cells used to grow our viruses. In this cellular environment, the standard “wild-type” MeV genome sequence is indeed dominant.
We asked what may happen in the initial phases of a host infection, when MeV replicates in lymphocytic tissue. To model this phase, we infected lymphocytic cell lines Granta or Raji and observed selection of MeV genome variants that cannot express functional V protein. These V-deficient genomes arise based on several different point mutations and after a few passages account for 90% of MeV genomes. Upon passaging in epithelial cells, V-competent wild-type genomes rapidly out-compete the V-deficient variants, and the quasispecies composition returns to the original equilibrium. Thus, in our experimental system V-competent genomes, which are sub-optimal variants in lymphocytes, constitute a low frequency variant pool for adaptation to epithelial cells.
Quasispecies re-equilibration is based on differential V protein expression. V proteins, which are conserved among most Paramyxoviridae, are polyvalent innate immunity controllers. The MeV V protein reduces both type I interferon signaling by inactivating STAT1 and STAT2 [18, 28, 36] and interferon production by inhibiting the cytoplasmic RNA sensor MDA5 [34, 37, 38]. While the residues interacting with STAT1 are located in the shared P and V amino-terminal half [39], those interacting with STAT2 and MDA5 are located in the V-unique carboxyl-terminal domain [21, 34]. Thus, V-deficient MeV are unable to control either STAT2-dependent or MDA5-dependent interferon activation.
Complete V-deficiency would be perplexing, even considering that Granta and Raji cells have reduced antiviral innate defenses [30, 40, 41]. However, the quasispecies growing in Granta cells include about 10% V-competent genomes. Thus, “just right” levels of V-protein expression may be required for efficient MeV spread in lymphocytic cells. Reduced innate immune defenses cannot be the only determinant of V-restriction because Vero-hSLAM cells do not select for V-defective mutants [42] while being interferon-defective [43].
In three of the four recombinant MeV with editing site-proximal mutations, RNA editing was minimal, and P protein expression enhanced at the expense of V protein expression. The exception was the +1(G) recombinant virus, which maintained near wild type RNA editing, and standard P-to-V protein expression ratio. However, the +1(G) mutation, which is silent for P, for V alters a histidine involved in Zn2+ binding [34]. This interferes with V-protein interactions with both STAT2 and MDA5 [34]. Positive selection of this mutant, which during the first experiment accounted for about half of the V-deficient genomes, suggests that inactivation of V protein function, rather than enhanced P protein expression, is key for adaptation to lymphocytic cells.
While editing efficiency is in the 30–50% range in different MeV strains propagated on Vero-hSLAM cells [42], a recent study revealed 5–20% editing efficiency in MeV mRNA extracted from brain autopsy materials of seven subacute sclerosing panencephalitis cases [44]. This indicates that during infection of humans editing efficiency can vary. It could also reflect rapid quasispecies adaptation to the neuronal environment.
Rapid adaptation to Granta and Raji cells could be explained by a negative effect of V on MeV replication in lymphocytic environments. The MeV V protein can limit viral RNA synthesis [45, 46], and a cellular co-factor differentially expressed in epithelial and lymphocytic cells could regulate this effect. Alternatively, the V-protein interaction with either STAT2, or MDA5, may be key for adaptation to lymphocytic cells. This hypothesis can be tested through the generation of selectively STAT2- or MDA5-blind MeV [28], or by passaging MeV in Granta cells that do not express STAT2 or MDA5 [47].
Rapid quasispecies re-equilibration may occur during acute MeV infections. This would be facilitated by the mode of MeV spread within hosts, which is based on the intercellular transfer of multiple genomes [48–50]. Experimental primate infections indicate that MeV spreads primarily through infected cells within a host [51]. In particular, MeV can spread between lymphatic cells through the formation of synapse-like interfaces [52] that are likely to transfer large numbers of genomes. Simultaneous transfer of genome packets may also occur in airway epithelia, based on the formation of intercellular pores [49]. Moreover, infectious MeV particles contain multiple genomes [48]. Thus, genome mixtures, rather than individual genomes, may spread through the host.
In conclusion, genomic adaptation of a dual-tropic RNA virus to its two natural cellular environments is unexpectedly rapid and thorough. Similar cyclical quasispecies re-equilibration processes may occur during natural infections with other dual-tropic RNA viruses. These include noroviruses, which infect epithelial and non-epithelial cell types [53], and HIV, which infects T-cells and macrophages. We suggest that the virulence of these dual-tropic RNA viruses may reflect the combined activity of distinct cell-specific quasispecies.
Vero-hSLAM (kindly provided by Y. Yanagi,[35]), 293-4-46 [23], and HeLa-hSLAM [54] cells were cultivated in Dulbecco's high-glucose modified Eagle's medium (D-MEM; HyClone; GE Healthcare Life Sciences, Logan, UT) supplemented with 10% (vol/vol) fetal bovine serum (FBS; Gibco; Life Technologies, Carlsbad, CA), and 1% Penicillin-Streptomycin Solution (Pen-Strep; Corning; Tewksbury, MA). Vero-hSLAM and 293-4-46 cells were grown in the presence of Geneticin (G418; Corning; Fisher Scientific; Hampton, NH) at final concentrations of 0.5 mg/ml and 1.2 mg/ml, respectively. HeLa-hSLAM cells were grown in 0.1 mg/ml Zeocin (Gibco; Invitrogen; Carlsbad, CA). The mantle cell lymphoma cell lines Granta-519 [55, 56] (Cat. # ACC 342; DSMZ; Braunschweig, Germany) and JVM-2 [57] (Cat # CRL-3002; ATCC; Manassas, Virginia), the Burkitt’s lymphoma cell line Raji [58] (Cat # CCL-86; ATCC) and the bronchioalveolar carcinoma cell line H358 [59] (Cat. # CRL-5807; ATCC) were cultivated in RPMI 1640 (HyClone) supplemented with 10% (vol/vol) FBS and Pen-Strep.
Recombinant MeV constructs were generated in the IC323 background (similar to the wild-type IC-B strain [24]). p(+)MV323(mCherryNLS)uN and p(+)MV323(mCherryNLS)H contain an additional transcription unit with mCherry fused to a triple repeat nuclear localization signal (NLS) either upstream of N or downstream of H, respectively. p(+)MV323(mCherryNLS)H was generated by transferring mCherry-NLS from pB(+)MVvac2(mCherryNLS)H [31] into p(+)MV323(GFP)H, replacing GFP, using the restriction sites MluI and AatII. Rescued MeV generated from p(+)MV323(mCherryNLS)H (named MeV-IC323-mCherry) was used for the first passaging experiment, and MeV from p(+)MV323(mCherryNLS)uN (named MeV-IC323-mCherry-uN) was used for the second passaging experiment.
p(+)MV323-eGFP-P(-9) was constructed as described previously (Singh et al, submitted). The editing site was modified in the -9 position (T to C) using complementary primers spanning CCAGCACTTCCGAGACACCCATCAAAAAGGGCACAGACGCGAGAT (mutagenized nucleotide is underlined) on the intermediate pCG-eGFP-P323 plasmid. The mutagenized eGFP-P323 construct was then cloned into an additional transcription unit of p(+)MV323(GFP)H downstream of H using MluI and AatII sites, replacing eGFP to generate the final construct.
MeV-IC323-P(-10)-mCherry (referred simply as mutant -10), and the (-9), (-7) and +1(G) mutants were generated first by performing site directed mutagenesis on the P gene plasmid, pCGPmeI-MVwtIC323-PmeI. Complementary primers used for mutagenesis span the follow sequence: CCAGCACTTCCGAGACACCCATTAAAAAGGGCACAGACGCGAGAT; which contained single mutated nucleotides (underlined) for each of the four viruses (T to A for -10, T to C for -9, A to G for -7, and C to G for +1G). The mutagenized pCG-Pmel-MVwtIC323-Pmel plasmids were then ligated into p(+)MV323 [24] using BstEI and BssHI restriction sites. An additional transcription unit containing an mCherry-NLS reporter was inserted downstream of H using MluI and AatII, generating p(+)MV323(-10)(mCherryNLS)H, p(+)MV323(-9)(mCherryNLS)H, p(+)MV323(-7)(mCherryNLS)H, and p(+)MV323(+1G)(mCherryNLS)H. All plasmids were verified using dideoxy-methods.
Recombinant viruses were produced as reported previously [23], generating passage 0 (p0) stocks. These stocks are amplified from a single syncytium to a 10 cm2 plate with 5x106 Vero-hSLAM cells. Passage 1 (p1) stocks were generated by infecting 2x108 Vero-hSLAM cells with p0 at 37°C until extensive cytopathic effect (2–4 days). Cells were harvested by scraping into Opti-MEM (Gibco) and then lysed by three freeze-thaw cycles (liquid nitrogen and 37°C). Cleared lysates were aliquoted and stored in -80°C for future experiments. Viral titers were determined using the 50% tissue culture infectious dose (TCID50) method [60].
For passaging, p1 stocks were used to infect 1-2x107 Granta-519 cells or H358 cells at MOI 0.1 for 3 days. Infected cells were collected in 1 ml opti-MEM and cell-associated MeV was released by 3 freeze-thaw cycles. In the two experiments, either 10% or 20% of the cleared lysate (100 or 200 μl) was used to infect the next dish of 1-2x107 Granta-519 or H358 cells. Constant volume was used for simplicity. Because volume was standardized for passaging, the MOIs were different for each passage, with most MOI in the 0.002 to 0.2 range. Infections were carried for either 3 days or until cell lysis began. We initially attempted passaging at consistent MOI, but could not always maintain sufficient levels of infectious output, especially after host cell type switching.
Either 106 Granta-519, H358, or Vero-hSLAM cells were infected at the indicated MOI in triplicate. Infected cells were harvested at the indicated time points, and then lysed by three freeze-thaw cycles. Titers of cell-associated MeV were measured with the TCID50 method on Vero-hSLAM cells.
To generate sufficient viral genomic material for sequencing, 2x108 Vero hSLAM cells were infected with either 500 μl of p1 stock or half of the passaged MeV inoculum. To prevent premature cell lysis, 20 μg/ml of fusion inhibitory peptide (Z-D-Phe-Phe-OH) (Bachem California Inc., Torrance, CA) was added 24 h post infection and the infection was moved from 37°C to 32°C until harvest. Purification of MeV ribonucleocapsids (RNP) was carried out by isopycnic centrifugation as described previously [47], with one variation after pelleting of RNP through CsCl gradient [61]. The RNP pellets were solubilized in 2 ml LEH (10 mM HEPES pH 7.5, 100 mM LiCl, 1 mM EDTA) containing 1% wt/vol sodium dodecyl sulfate (SDS). RNA was extracted twice with phenol:chloroform:isoamyl alcohol (25:24:1, vol/vol/vol; Thermo Fisher Scientific; Waltham, MA) and once with chloroform:isoamyl alcohol (24:1). LiCl was added to achieve 150 mM concentration and the RNA was precipitated with two volumes 95% ethanol at -20°C overnight. RNA was resuspended in 25 μl diethyl pyrocarbonate-treated water.
RNP RNA (0.5 μg) was incubated in 10 μl of a buffered zinc solution (Thermo Fisher Scientific) for 7 minutes at 70°C, according to the manufacturer’s protocol. Fragmented RNA was purified by phenol:chloroform:isoamyl alcohol phase separation and ethanol/sodium-acetate precipitation. The concentration and integrity of the RNA was assessed on an Agilent Bioanalyzer DNA 100 chip (Agilent, Santa Clara, CA). cDNA library prep was conducted using Illumina TruSeq Stranded Total RNA Sample Prep Kit (Illumina, San Diego, CA) according to the manufacturer’s protocol. The 300 x 2 paired end sequencing of each library was performed on an Illumina MiSeq using MiSeq v2 sequencing kit and MCS v2.6.2.1 collection software. Base-calling was performed using Illumina’s RTA version 1.18.54.
The raw BAM files from Illumina sequencing were uploaded into the Galaxy web platform [62]. We used the public server (http://usegalaxy.org/) for downstream processing and analysis of the data. Briefly, .bam files were converted into .fastq files using SamToFastq version 1.126.0, generating two FASTQ files for each data set (split by read group). Illumina adapter sequences were clipped using the FASTX-Toolkit. Low quality reads were filtered using FASTQ Quality Trimmer [63] by trimming reads from the 3’ end that had quality scores below or equal 20. Additional filtering was performed using the FASTX-Toolkit to eliminate reads that did not contain 95% or greater nucleotides having a quality score above 30. We used Bowtie2 version 2.2.6.2 [65] to process reads and align them to the Chlorocebus sabaeus genome (GCA_000409795.2 [64]), ribosomal RNA (hsa-45S-pre-rRNA, accession: NR_046235.3) and a MeV genome identical to IC323-EGFP (accession LC420351.1), in which the EGFP additional transcription unit sequence was either replaced with mCherry-NLS (MeV-IC323-mCherry-uN) or replaced and moved downstream of H (MeV-IC323-mCherry). IdxStats version 2.0 from the SAMTools software package [66] was run to determine read count distributions across the reference sequences. The IC323 genome aligned .bam files were then loaded in Integrative Genomics Viewer 2.3.98 (IGV; Broad Institute; Cambridge, MA) [67, 68] and aligned reads were visualized. Read count tables were generated using IGVTools [68]. Allelic frequencies were calculated and additional analyses were performed after by uploading allelic frequencies into Microsoft Excel.
Immunoblotting was performed as described previously [69]. Briefly, HeLa hSLAM cells were cultured in 6-well plates and infected with the viruses indicated. At the times indicated cells were lysed as described previously [70], incubated on ice for 30 min, and nuclei pelleted by centrifugation at 16,000 x g at 4°C for 30 min. Supernatant was collected and protein was quantified by biocinchoninic acid assay and read using the Tecan Infinite M200 Pro reader (Männedorf, Switzerland). Each lane was loaded with 20 μg total protein, fractioned by 10% SDS-PAGE, and transferred to Immobilon-P membranes (Merck; Darmstadt, Germany) using a wet transfer protocol. Membranes were blocked with 5% (wt/vol) nonfat milk (BioRad, Hercules, CA) in Tris-buffered saline (TBS), pH 6.8 for 1 h and incubated with primary antibodies at 4°C overnight. Membranes were washed three times with TBS with 0.5% (vol/vol) Tween 20 (TBST) for 5–10 min each, incubated with horseradish peroxidase (HRP)-conjugated secondary antibody at room temperature for 1h, washed three times with TBST, and incubated with Supersignal West Pico chemiluminescent substrate (Thermo Fisher Scientific). Membranes were exposed to Hyblot CL autoradiography films (Denville Scientific, Holliston, MA) or scanned using BioRad ChemiDoc Imaging System (Hercules, CA).
A rabbit antiserum was raised against the peptide sequence KRNKDKPPITSGSGGAIRGIKH, corresponding to amino acids 12 to 33 of the MeV N protein, coupled to keyhole limpet hemocyanin, as described previously [71]. MeV P [70] and V [71] antisera were used at dilutions of 1:5,000. Rabbit polyclonal anti-GFP (Abcam, Cambridge, United Kingdom) was used at 1:1,000 dilution. Mouse monoclonal anti-actin (HRP) (Sigma-Aldrich, St. Louis, MO) was used at 1:25,000 dilution. Rabbit secondary antibodies conjugated with HRP were used at 1:10,000 dilution.
Total RNA was extracted using Trizol reagent (Thermo Fisher Scientific) and precipitated with isopropanol according to the manufacturer’s instructions. Precipitated RNA was resuspended in 20 μl of DEPC-treated water and stored at -80°C. RNA (100 ng) was reverse transcribed using Superscript III reverse transcriptase (Invitrogen; Carlsbad, CA) and oligo (dT) (Promega; Madison, WI) to prime the reaction, according to the manufacturer’s protocol. PCR was performed using Phusion HF kit (New England Biolabs; Ipswich, MA) on the reverse transcribed product. For MeV-IC323-eGFP-P(-9), primers F1 5’AACCAACCATCCACTCCCAC and R 5’GAGGATCGGAAGCGTTACCT were used to amplify endogenous P; F2 5’GAGGATCGGAAGCGTTACCT and R were used for eGFP-P amplification. For amplification of P in all other viruses: 2001F 5’CTCAGCAATTGGATCAAC and P rev 5’AGGTAACGCTTCCGATCCTC [69] were used. PCRs were carried out for 35 cycles (98°C 10s, 50°C 30s, 72°C 1:20m). The PCR product was then sequenced by dideoxy methods with either 2001F or 2401F (AGAGGCAACAACTTTCC) forward primers and 2801R 5’GATTCTAGCTTGGAGATTA as a reverse primer. Chromatograms were analyzed using MEGA7 [72].
Student’s unpaired t-tests were performed to determine significance compared to parental or p1 viruses in growth curves and in western blots. P values are marked **, P < 0.01.
|
10.1371/journal.pcbi.1006672 | Enzymatic and non-enzymatic pathways of kynurenines' dimerization: the molecular factors for oxidative stress development | Kynurenines, the products of tryptophan oxidative degradation, are involved in multiple neuropathologies, such as Huntington's chorea, Parkinson's disease, senile dementia, etc. The major cause for hydroxykynurenines's neurotoxicity is the oxidative stress induced by the reactive oxygen species (ROS), the by-products of L-3-hydroxykynurenine (L-3HOK) and 3-hydroxyanthranilic acid (3HAA) oxidative self-dimerization. 2-aminophenol (2AP), a structural precursor of L-3HOK and 3HAA, undergoes the oxidative conjugation to form 2-aminophenoxazinone. There are several modes of 2AP dimerization, including both enzymatic and non-enzymatic stages. In this study, the free energies for 2AP, L-3HOK and 3HAA dimerization stages have been calculated at B3LYP/6-311G(d,p)//6-311+(O)+G(d) level, both in the gas phase and in heptane or water solution. For the intermediates, ionization potentials and electron affinities were calculated, as well as free energy and kinetics of molecular oxygen interaction with several non-enzymatically formed dimers. H-atom donating power of the intermediates increases upon the progress of the oxidation, making possible generation of hydroperoxyl radical or hydrogen peroxide from O2 at the last stages. Among the dimerization intermediates, 2-aminophenoxazinole derivatives have the lowest ionization potential and can reduce O2 to superoxide anion. The rate for O-H homolytic bond dissociation is significantly higher than that for C-H bond in non-enzymatic quinoneimine conjugate. However, the last reaction passes irreversibly, reducing O2 to hydroperoxyl radical. The inorganic ferrous iron and the heme group of Drosophila phenoxazinone synthase significantly reduce the energy cost of 2AP H-atom abstraction by O2. We have also shown experimentally that total antioxidant capacity decreases in Drosophila mutant cardinal with L-3HOK excess relative to the wild type Canton-S, and lipid peroxidation decreases in aged cardinal. Taken together, our data supports the conception of hydroxykynurenines' dual role in neurotoxicity: serving as antioxidants themselves, blocking lipid peroxidation by H-atom donation, they also can easily generate ROS upon dimerization, leading to the oxidative stress development.
| Kynurenines, the products of tryptophan catabolism, are compounds with the multiple neuroactive properties. Hydroxykynurenines are redox modulators with a dual role in the oxidative stress development: being able to block lipid peroxidation, they also produce toxic free radicals during the oxidative self-dimerization. The later can occur both enzymatically and non-enzymatically. Computational study of the dimerization process may reveal the stages and intermediates being the source of the reactive oxygen species. This would help to provide therapeutic strategies aimed at the decrease of kynurenines' dimerization toxic effects. We have modeled oxidative conjugation of kynurenines using quantum-chemical approaches, performing calculations for the energies of the dimerization stages, as well as the redox properties of the intermediates. Phenoxazinone synthase provides a structural basis for the full oxygen reduction to water, prohibiting generation of toxic free radicals. The final non-enzymatic dimerization stages are the most probable source of hydroperoxyl and superoxide anion radical. The accumulation of L-3HOK in Drosophila mutant cardinal with impaired phenoxazinone synthase correlates with decrease in both total antioxidant capacity and lipid peroxidation in an age-dependent manner. Our results support the concept of protective role of proteins converting hydroxykynurenines to their dimeric forms, preventing free radicals overproduction and oxidative stress development.
| The kynurenine pathway (KP) is a primary route of tryptophan degradation in mammals. Its metabolites, collectively called kynurenines, possess diverse neuroactive properties, being involved in the development of numerous neuropathologies, such as Alzheimer's disease (AD), Parkinson's disease (PD), Huntington's disease (HD), etc. [1–4]. Indoleamine-2,3-dioxygenase (IDO) or tryptophan-2,3-dioxygenase (TDO) governs the initial rate-limiting stage of KP, converting tryptophan to N-formylkynurenine, which is then non-enzymatically hydrolyzed to kynurenine. Till the late 1970th, NAD+ synthesis was thought to be the only biological role for KP, however, now its neuro- and immunomodulatory role is well-established [5]. Kynurenines are deeply involved in neurotoxicity, neuroinflammation and excitotoxicity. These three phenomena result from interrelationship of metabolic disturbances leading to increase in reactive oxygen and nitrogen species (ROS, RNS) [6].
There are two chief molecular mechanisms by which kynurenines act on nervous system: the ligand-receptor interaction and the regulation of oxidative stress. Kynurenine (KYN), the first of KP metabolites, is a ligand of the aryl hydrocarbon receptor, suppressing antitumor immune response, both in humans [7] and Drosophila [8]. Quinolinic acid (QUIN) is an agonist and kynurenic acid (KYNA) is an antagonist of ionotropic glutamate receptors [9,10], modulating neurodegenerative processes development [11]. KYNA is also a ligand for G protein-coupled receptor GPR35, regulating Ca2+ mobilization and inositol phosphate production [12]. L-3-hydroxykynurenine (L-3HOK) and 3-hydroxyanthranilic acid (3HAA) are antioxidants inhibiting peroxyl radical-mediated oxidation of phosphatidylcholine and plasma lipid peroxidation [13,14]. At the same time, their autoxidation leads to the hyperproduction of noxious ROS, damaging cellular structures and causing apoptotic cell death [15,16].
Both prooxidant and antioxidant properties of hydroxykynurenines are due to the easiness of one-electron or H-atom abstraction. 2-aminophenol (2AP), a structural precursor of L-3HOK and 3HAA, has a significantly lower homolytic O-H bond dissociation enthalpy (BDE) compared to that for phenol [17]. Antioxidants should have O-H BDE low enough to provide free radicals quenching by H-atom. However, too low BDE may cause the antioxidant H abstraction by molecular oxygen, converting both of them to free radicals able to initiate lipid peroxidation [18]. There must be an optimal balance between the ability to donate H-atom and chemical inertness of the compound to remain harmless to a cell. The antioxidant protection systems, including specific enzymes, govern ROS and other free radicals conversion to the non-active low-energy compounds, such as water molecule.
2AP dimerization can occur enzymatically [19] and non-enzymatically [20,21] being catalyzed by several organic or inorganic compounds. In both cases, the condensation of two 2AP (or their structural analogues R-2AP) includes three consecutive stages of two H-atom abstraction equal to six-electron oxidation (Fig 1). Likewise, 3HAA oxidative dimerization by laccase leads to cinnabaric acid (CIN) formation [22]. Non-enzymatic dimerization may occur via several redundant stages. Here, we present the scheme of 2AP non-enzymatic conversion including both experimentally found [20] and some hypothetic intermediates (Fig 2). Being different in their intermediates and the number of stages, both enzymatic and non-enzymatic pathways lead to formation of 2-aminophenoxazinone (2APX) or its structural analogue 2R-2APX.
The summary equation is the following:
R−2AP+O2+oxidant→R−2APX−R+2H2O+H2−oxidant
(1)
where the last two H-atom abstractions are considered to be a non-enzymatic process [19,21,23]. For L-3HOK, there is an additional stage of two H-atom abstraction from R-2APX-R leading to the fourth planar ring cyclization and formation of xanthommatine (XAN), the insect brown pigment [24].
Phenoxazinone synthases (PHSs, class EC 1.10.3.4) belong to various structural types of proteins having different functions in bacteria, plants, fungi, and animals. phsA, the first reported PHS involved in actinomycin production by Streptomyces antibioticus, is an oligomer Cu2+-containing enzyme that shares structural similarities to laccase [19,21]. It uses 2AP and O2 as substrates, hydrogen peroxide may be an intermediate of the reaction. In Drosophila melanogaster, PHS is a phsA functional analogue encoded by cardinal (cd), showing significant similarity to the heme-containing peroxidase [25]. It catalyzes the conversion of 3HOK to XAN and produces CIN from 3HAA, its most specific substrate [26]. Some other proteins, such as laccase, tyrosinase, peroxidase and human hemoglobin are able to catalyze similar reactions [21].
In Drosophila cd mutant, PHS activity is decreased to ~40% of the wild type [24]. Enzymatic dihydroxanthommatin (DXAN) synthesis from 3HOK is disrupted in cd demonstrating 2.9-fold 3HOK increase in heads [24,27], as well as age-dependent memory loss [28]. The oxidative dimerization of 3HOK and 3HAA to XAN and CIN can occur both enzymatically and non-enzymatically. Spontaneous XAN and CIN formation in Drosophila homogenates is catalyzed by themselves serving as electron acceptors, while DXAN is a powerful inhibitor of the oxidation process. Unexpectedly, the enzymatic activity, possibly associated with some unspecific oxidase, was shown to be similar in cd and two other kynurenine pathway mutants, vermilion and cinnabar, and non-enzymatic oxidation activity was depleted in cd [29]. The last one was probably associated with XAN which autocatalyzes its own formation.
Both activities seem to produce an impact on age-dependent progressive cd-specific loss of middle-term memory accompanied by neurodegeneration [28] and male courtship song distortions [30,31]. Drosophila cd mutant is a model object to study senility processes and ROS-induced neurodegeneration mechanisms at the molecular level [28]. To develop the therapeutic strategy implying natural and synthetic antioxidants, we should properly understand the impact of enzymatic and non-enzymatic R-2AP conversion pathways to oxidative stress progression. The by-products of enzymatic pathway are two chemically inert water molecules and some unknown reduced oxidant. At the same time, non-enzymatic oxidation likely leads to toxic ROS formation due to single electron or H-atom abstraction by molecular oxygen. PHS dysfunctions should shift the balance to the non-enzymatic oxidation, which in turn may occur via several pathways.
Studying the redox properties of dimerization intermediates would reveal the most prone to produce ROS. The complex nature of dimerization processes makes it hard to study all its separate stages experimentally. This requires the usage of computational and modelling approach. Previously, the ability of hydroxykynurenines to donate H-atom and inhibit phenoxyl or methyl peroxy radicals was computationally investigated [32]. Both L-3HOK and 3HAA are powerful antioxidants, rapidly quenching peroxy radicals by hydroxyl H-atom transfer. In this study, we use similar approach to estimate Gibbs free energies (ΔG) of R-2AP dimerization stages for enzymatic and non-enzymatic pathways. The redox properties of R-2AP dimerization products were estimated by calculating their ionization potentials (IP) and electron affinities (EA). ΔG and rates of H-atom abstraction by molecular oxygen were calculated for some intermediates of non-enzymatic pathway. D. melanogaster PHS structure in complex with R-2AP was modeled and the energies of 2AP interaction with PHS oxy-heme group were calculated. For all compounds, single-point energies were calculated both in the gas phase and in heptane or water solution modeling hydrophobic lipid and aqueous surrounding, respectively. Finally, the age-dependent total antioxidant capacity (TAC) and lipid peroxidation level (LP) were estimated in Drosophila cd[1] strain with L-3HOK accumulation.
The enzymatic and non-enzymatic R-2AP dimerization pathways share the common stages of initial H-atom dissociation from R-2AP (N1, 2) leading to quinoneimine (R-2APq) production. Enzymatic pathway includes the following stages: conjugation (N3, 9); H-atom abstraction (5, 6, 11, 12); isomerization/ H-atom migration (N4, 10), proton association (N7, 13); proton dissociation (N8, 14) (Table 1). We called the “classic” pathway of dimerization the way including stages N1 –N14, firstly described for 2AP in [19]. Non-enzymatic pathway may exert via several redundant stages, such as conjugation (N15, 20, 23), H-atom abstraction (N16, 17, 19, 21, 22), isomerization/ H-atom migration (N18, 24, 25) (Table 2). For L-3HOK-D10, there is an additional non-enzymatic conjugation stage (N26) leading to XAN formation and two H-atom abstractions. Noteworthy, not all the stages of the “enzymatic pathway” are catalyzed by PHS. For instance, D7 → D10 reaction may occur spontaneously [19]. Hence, the very distinction of “enzymatic” and “non-enzymatic” dimerization pathways is rather conditional.
For the majority of the stages, the optimal conformations of dimer skeleton including two connected planar rings with OH and NH2 groups are similar among all the studied compounds, though the side chain conformations may vary (Fig 3). Hence, we consider our models of R-2AP dimerization intermediates to be generally correct. However, for some structures the difference between R-2AP-D is pronounced, e.g. for D1, D5 and D2'-D4'. Mainly this is due to hydrogen bonds (HBs) formation by side chains of several compounds. 2AP lacking R side chains has a lower opportunity to form intramolecular HBs or electrostatic interactions, positively or negatively shifting ΔG values. Noteworthy, at physiological pH L-3HOK is in zwitterion (zi) form. This may significantly affect its side chains interactions, e.g. leading to NH3+ and COO- groups’ spatial approach. We found L-3HOKzi and its dimers to be unstable in the gas phase and poorly converging to the local minimum in aqua that made impossible to use them in our computations.
The total ΔG for two 2APs conversion to 2APX (D10) is the same for the both pathways (shown in Table 1). In the gas phase, its values are close for the uncharged compounds (2AP, L-3HOK, 3HAA; 325.739±12.137 kcal/mol), being significantly lower than that for the negatively charged 3HAAi. The aforementioned difference becomes lower in heptane and actually disappears in water solution. Here, ΔG is almost identical for three kynurenines (332.798±4.960 kcal/mol), being higher compared to that for 2AP. The conjugation stages (N3, 15) are slightly endothermic or exothermic (for 3HAA, N15). Therefore, they may occur spontaneously. For 3HAAi, N3 conjugation is highly endothermic in the gas phase or in heptane. Seemingly, this is due to the repulsion of two negatively charged COO- groups, which becomes more energetically favorable in water solution. The intramolecular conjugation stages N9 and N20 are also slightly endothermic, with the highest ΔG for 3HAAi-D, while N24 stage is less energetically favorable than N20. Taking into account BSSE correction, the dimerization energies should be ~3–7 kcal/mol higher, and 3HAAi dimerization is even more energetically unfavorable compared to the uncharged form (see Materials and Methods). The tautomerization stages (N4, 10, 18) following the conjugation stages and leading to the second ring aromatization are mainly exothermic. For 2AP, L-3HOK and 3HAA, ΔG values are strongly correlated both in the gas phase and in heptane or water solution, while for 3HAAi the correlation is strong only in aqua (Table I in S2 Table).
H-atom homolytic bond dissociation stages have the highest ΔG values (BDG): the average BDG is 50.056±5.136 kcal/mol for the gas phase, 51.032±5.365 kcal/mol for heptane solution, and 52.335±5.229 kcal/mol for water solution (the differences are non-significant; n = 44). H-atom abstraction BDG widely vary for the different stages, being maximal for N22 (~76 kcal/mol) and minimal for N21 (~23 kcal/mol) (S1 Fig). The only exception is the hypothetical reaction D6' → D10 (N24) where BDG is highly negative. D6' is a biradical and triplet in its ground state, so its spontaneous transition to the closed shell D10 is spin-forbidden. Hence, the whole D3' → D4' → D6 → D10 pathway (N22, 23, 24) seems to be complicated. The alternative pathway includes D5' intermediate which is either directly converted to D10 (N21), or isomerized to D8 (N25), joining the “classic” pathway (see Fig 2).
For N-H bond dissociation leading to quinoneimine structures formation, BDG is significantly higher than that for O-H bond dissociation (compare N22 –N19, N2 –N1, N6 –N5, N12 –N11). Partially, this may be caused by the addition of a diffuse orbital to O-atom at level III (see Materials and Methods). However, the average difference between N-H and O-H BDG calculated at level II is 11.081±4.854 kcal/mol, while for the bond dissociation enthalpy (BDE) it is 12.798±6.773 kcal/mol (n = 12, enzymatic pathway). Hence, N-H bond is more difficult to break, especially for 3HAAi-D. BDG is lower for dimers compared to monomers (compare N5 –N1, N6 –N2, N19 –N1) and for the oxidized dimers compared to the reduced dimers (compare N11 –N5, N12 –N6). This fact corresponds to proposal of [19]: 2-aminophenols become more electron rich as the oxidation progresses, allowing for non-enzymatic oxidation at the last stage. At physiological pH, D4 and D9 are mainly in the protonated (NH2+) form. We modelled D3 → D5 and D8 → D10 conversions as consisting of three stages: a. aromatic N-H bond dissociation (N6, 10); b. quinoneimine NH group protonation (N7, 11); c. charge migration and amine bridge NH group deprotonation (N8, 12). For b. and c., the summary ΔG is negative, except for 3HAAi-D9(H+).
For non-enzymatic pathway stages where C-H bond dissociation is accompanied by the second ring aromatization (N16) or intramolecular cyclization (N17), BDG is lower than that for O-H BDG values, except for 3HAAi-D in N17. However, C-H bond dissociation rate may be low due to the high value of the activation barrier. In this case, N16 can proceed in two stages: the exothermic keto-enol tautomerization of D1' second ring (N18) and subsequent enol O-H bond dissociation (N19) with BDG close to those for the similar dissociation step of enzymatic pathway (N5). As mentioned before, N17 is likely to proceed via intramolecular cyclization (N20) and subsequent C-H bond dissociation (N21). N21 has the lowest BDG, making D5' a powerful H-atom donor in reactions with molecular oxygen or some other ROS generators. However, D5' C-H bond dissociation may be kinetically inhibited, as well as that for D1'.
O-H BDG slightly differs for the kynurenines' derivatives: typically, it is somewhat higher for L-3HOK-D compared to that for 3HAA-D (N1, N5, N11), except in N19. In this case, low BDG value may be explained by the relaxation of the distorted L-3HOK aromatic ring (see below). Contrary, N-H BDG is higher for 3HAA compared to all the other compounds. For 3HAAi-D, O-H BDG is significantly lower compared to the non-ionized form in the gas phase, the difference becoming less pronounced in water solution. This corresponds to our previous data that negatively charged carboxylic group decreases O-H bond dissociation energy in kynurenines–an effect reduced in media with the high dielectric constant [32]. This is similar for N-H bond: BDG becomes the same or even higher for 3HAAi-D compared to 3HAA-D in water solution. On the contrary, C-H BDG is higher for 3HAAi-D than for 3HAA-D (N17, N21). However, in all cases BDG differences for kynurenines' derivatives are statistically insignificant (S1 Table).
The other probable factors causing ΔG differences for 2AP and kynurenines' dimerization are:
The effects of polar interactions are decreased in heptane and are nearly extinguished in water solution.
In living organism, 2AP derivatives form complexes with metal ions such as Cu2+ affecting their redox properties [33]. In this study, we are focused on H-atom donation ability of R-2AP derivatives in apo form, without inorganic ions, seeking stages with the lowest H-atom abstraction BDG, which may partly redox molecular oxygen (O2) to toxic ROS. In addition, we studied the electron-donating and electron-accepting capacities of R-2AP derivatives in free form, without metal ions and any other substances modulating their redox properties.
The total energy effect of R-2AP dimerization is determined by Eq 1: six H-atoms are abstracted by O2 and some oxidant to produce two H2O molecules and some reduced oxidant. In Table 3, ΔG is given for the summary reaction while considering O2 to be an oxidant abstracting the last two H-atoms. In this case, the summary dimerization reaction is thermodynamically favorable, with a total energy output of ~50–100 kcal/mol. However, O2 may be only partly reduced at some stages, forming ROS, such as hydrogen peroxide (H2O2), hydroperoxyl radical (HO2*) or superoxide anion radical (O2*-). O2*- formation is a result of an electron abstraction by O2 (see below). HO2* and H2O2 can be formed through a single- and paired H-atom abstraction, respectively. Likely, this occurs in compounds where the aromatic OH and NH2 groups are oriented towards O2 by their H-atoms. The appropriate pair half-reactions would be N1+2, 5+6, 11+12, and 19+22. The half-reactions of R-2AP-D oxidation corresponding to the summary reactions with ΔG<0 are shown in Table 3.
For L-3HOK and 3HAAi, H2O2 can be formed at the stages N11+12, believed to pass spontaneously or outside the catalytic center of PHS [19]. In the non-aqueous phase, H2O2 formation may occur via non-enzymatic oxidation of D2' to D4' (N19+22). In the gas phase, it is also possible for the reactions N5-6, which normally pass within the catalytic center. For 2AP-D and non-ionized 3HAA-D, the same reactions are thermodynamically unfavorable. For L-3HOK-D10 in water solution, the final two H-atom abstraction leading to XAN (N26) can be accompanied by H2O2 production. HO2* formation is the least sterically hindered one, as it occurs via a single H-atom abstraction. However, for the “classic” pathway, it is energetically favorable only for N11 with 3HAAi (in gas and heptane). For the non-enzymatic pathway, both C-H (N16, 17, 21) and O-H (N19, in gas and heptane) dissociation energies are low enough to provide H-atom migration to O2. To sum up, non-enzymatically formed R-2AP-D are more thermodynamically prone to interact with O2 leading to ROS formation.
The redox properties of R-2AP-D are not restricted to H-atom abstraction power, but also related to their ability to donate and receive electrons. Here, we limited our calculations to forms that were shown to exist experimentally (see Fig 1), except for the protonated forms which IP are significantly higher due to their positive charge.
IP widely varies for different R-2APs, depending on both their chemical nature and dimerization form (Table 4 and S2 Fig). In the gas phase, 3HAAi-D have significantly lower IP compared to the other compounds, such as 3HAA-D (S1 Table), in accordance to previous data that the ionized acidic group decreases kynurenines' IP [32]. This effect is partly compensated in heptane solution and only to a small extent remains in water solution. IP is slightly lower for 2AP-D compared to L-3HOK-D, as well as to 3HAA-D. Thus, in the aqueous surroundings, the influence of R-2AP-D chemical nature on its electron-donating power becomes stronger and the influence of its acid-base forms on IP values weakens.
For the most compounds, R-2APq and 2-aminophenoxazinole derivative D7 have the highest and the lowest IP, respectively. Quninoneimines possess lower ability to abstract an electron compared to the radical forms of R-2AP and their dimers (compare R-2APq–R-2AP*, D9 –D8, D4 –D3). Forms with H-atom attached to the second ring disrupting its aromatic system have higher IPs compared to forms with two aromatic rings (compare D1 –D2, D6 –D7), possibly, due to the lower capacity for the unpaired electron delocalization. The oxidation of O-H and N-H groups disrupting the ring aromaticity decreases the ability to abstract an electron (compare D10 –D7, D9 –D7, D4 –D2, D1'–D1, D3'–D3). At the same time, IP decreases when rings conjugate keeping their aromaticity (R-2AP–D2 –D7).
For the same compounds, EA were calculated (Table 5 and S3 Fig). Predictably, the negatively charged 3HAAi derivatives have the most negative EA, which is lower for the dimers with two ionized carboxylic groups than for the monomer. This difference decreases in heptane and virtually vanishes in water solution. EA is slightly less for 2AP-D compared to L-3HOK-D, as well as to 3HAA-D. Thus, in the aqueous surroundings, EA variations for different compounds become less pronounced. Among the uncharged compounds, EA is maximal for D3'. The trend for electron accepting power change is mirror to that for the electron donating ability, as EA decreases within the following groups: D3'–D3, D1'–D1, D4 –D2, D10 –D7, D9 –D7; D1 –D2, D6 –D7; R-2APq–R-2AP*–R-2AP, D9 –D8, D4 –D3. However, the uncharged R-2AP has lower EA compared to D2, and D7 is intermediate between them. All these three substances preserve 2AP moiety, being the worst electron acceptors among the studied compounds.
To reveal the summary trend, we have calculated the Mulliken electronegativity (χ) which is a half-sum of IP and EA (Fig 4). Again, we see that 3HAAi-D have a lower χ compared to the uncharged 3HAA-D. 2AP-D is slightly less electronegative compared to L-3HOK-D, as well as to 3HAA-D. R-2APq and D7 have the highest and the lowest χ, respectively. The trends of χ change are similar to that for EA: 1. χ is higher for the oxidized forms of compounds compared to their restored forms; 2. χ is higher for tautomers with disrupted aromaticity compared to those with the aromatic ring. For the compounds containing 2AP moiety, χ decreases in the raw: 2AP–D2 –D7, same as for IP values.
Summing up, L-3HOK-D and 3HAA-D are slightly better electron acceptors and worse electron donors compared to 2AP-D. However, as 3HAA derivatives are mainly in ionized form at physiological pH, their electronegativity should be more like as that for 2AP-D. Among all the studied compounds, 2-aminophenolic derivatives (D7, D2, 2AP) are the worst electron acceptors, while D7, D2 and their radical forms are the best electron donors. Hence, they should be the most powerful electron-donating antioxidants among all R-2AP derivatives. However, in case of their redox potential being too low, they are able to convert molecular oxygen to toxic ROS. Most 3HAAi–D can reduce O2 to superoxide anion radical O2*- in the gas phase and in heptane solution (see Table 3). However, this reaction is thermodynamically forbidden in water solution, even for the most powerful electron donor 3HAAi-D7 (IP 94.892 kcal/mol). Possibly, O2*- can be formed in the presence of metal ions such as Cu2+, which are known to increase 3HAAi electron-donating power [33]. In such case, 3HAAi-D7/ D8 are the most probable O2*- generators.
For different R-2AP derivatives, there is a strong Pearson correlation among their BDG, IP, EA, or χ values in water solution (Tables II-IV in S2 Table). In the gas phase, these values for 3HAAi D do not correlate with those for the other dimers, because of the high influence of the charged COO- groups on the redox properties and H-atom donating power. This effect is extinguished to some part in heptane. Here, R-2AP-D and especially 3HAAi-D demonstrate higher ability to produce toxic ROS compared to that in aqua. However, R-2AP derivatives should have different solubility in the lipid phase, which affects their ability to penetrate through the hydrophobic lipid bilayers. The additional factors should be considered, such as the affinity to enzyme and metal ions within the cell, as well as to specific membrane transporters, which in complex determine R-2AP-D availability for interaction with O2 and membrane lipids.
As it has been said before, C-H homolytic bond dissociation in D1' (stage N16) and D5' (N21) is more energetically favorable compared to O-H bond dissociation in D2' (N19). However, N19 conversion due to D2' interaction with O2 may occur faster than that for N16 and N21. To check this hypothesis, we have calculated the reaction rates for the corresponding stages, as well as for the reverse stages (Fig 5 and Table 6). For 3HAAi, only D1'-O2 transition state (TS) occurred to be stable with our computational approach.
N16 is significantly slower than N19. At the same time, N16 is virtually irreversible, while the reverse reaction is significantly faster for N19 stage, except for L-3HOK-D3'. In N16, ΔG is lower for the products of reaction, whereas in N19, it is lower for the reactants, making D2'-O2 complex more stable compared to D3'-HO2*.Thus, D1' can be directly converted to D3' via H-atom abstraction by O2, making N16 reaction a potential source of HO2*. In N16, HO2* interacts with C = O group of the second aromatic ring, in N19, it remains in complex with C-O* group. This may prevent HO2* diffusion into the surrounding medium, partially decreasing the reaction rate.
N21 is dramatically slow, the reverse reaction being several orders of magnitude faster. The spin value is 1.5 for D5'–O2 complexes (including reactants, TS and products), corresponding to three unpaired electrons in the orbitals. Hence, N21 reaction is spin-forbidden, as O2 in the ground state has a spin value 1. Thus, to be oxidized, D5' should firstly be converted to D8: H-atom migrates from C- to N-atom, thereafter being abstracted by molecular oxygen.
In N16, the reactants, TS and products are in similar conformations for all 2AP derivatives, except for 3HAAi-D3'–HO2* where HO2* is away from the first aromatic ring. Logs of the reaction rates in gas, heptane and water are highly correlated (Table V in S2 Table). The direct reaction rate is higher for 2AP-D1' compared to 3HAA-D1' and L-3HOK-D1'. 3HAAi-D1' reacts most quickly with O2 except in water solution, the reverse N16 reaction for 3HAAi-D1' is very slow. In N19, O2 and HO2* positions differ for the studied complexes, albeit their TS geometries are quite similar. However, their reaction barriers and rates are approximately the same, except for L-3HOK-D2' where the products–reactants ΔG is negative, possibly due to some relaxation of steric strain within D2' structure. In N21, the conformations are the same for all the complexes. However, as it has been aforementioned, this reaction in unlikely to occur.
In Drosophila PHS, the putative catalytic site is similar to that in lactoperoxidase containing heme-binding residues, as well as the axial His residue. We can assume that PHS heme group in complex with O2 is in ferric form, similar to that in globins and TDO. Upon 2AP binding to PHS active site, its hydroxyl and amine hydrogens should be oriented towards O2. Automatic docking of R-2AP derivatives to PHS with oxo heme in the active site generated several structures, some of which meets the above criteria (Fig 6A and 6B). The binding pocket is wide enough for R-2AP and R-2AP-D2 to contact with O2 by their 2AP moiety. For R-2AP-D7 with a plane tricyclic group, there are no docked structures with such contacts to O2. The only exception is 3HAAi-D7 where OH and NH2 groups are oriented in the opposite way to that in 3HAAi-D2. Here, there are no interactions between enzyme and COO- groups, and the total binding energy is relatively low. These data correspond to the experimental fact that PHS does not catalyze the last two H-atoms abstraction from R-2AP-D7.
The aromatic group of the substrate is nearly parallel to the heme plane, being ~3.2–3.5 Å above the pirrolic ring of heme. For 3HAAi, L-3HOKzi and their D2 derivatives, the PHS-substrate complex is stabilized by the ionic bond of COO- group and Arg230. For L-3HOKzi and L-3HOKzi-D2, there is an ionic bond between αNH3+ and the heme COO- group. His89 is above O2 and R-2AP aromatic NH2 group, being able to contact with them. The axial His336 is under the heme group (Fig 6B). The average distance between R-2AP O/N atoms and corresponding O atom of O2 is ~1.8/2.1 Å. The binding energy is minimal for 2AP and maximal for L-3HOKzi (Table 7). The second OH group in 2AP-D2 contacts with Thr94 C = O, and the second αNH3+ group in L-3HOKzi-D2 contacts with Asp229, that must additionally stabilize the enzyme-substrate complexes.
O2 in the ground state in complex with porphyrin group of heme and imidazole group of the axial His has the total spin value 0, corresponding to quantum mixture of two configurations: Fe(II)-O2 (ferrous heme) and Fe(III)-O2– (ferric heme) [34]. Fe(II) and Fe(III) ions significantly enhance the rate of DOP H-atom abstraction by O2 and H2O2 formation [35,36]. To check the influence of ferrous iron on 2AP H-atom abstraction by O2, we calculated the free energies of 2AP interaction with: a. O2 (triplet state); b. O2 in complex with FeCl2 (singlet and triplet states); c. O2 in complex with porphyrin group of heme and the axial imidazole (singlet state) (Fig 7C and Table 8). In the absence of iron, H-atom abstractions from 2AP leading to HO2* and H2O2 is thermodynamically unfavorable. In the presence of FeCl2, 2AP conversion to HO2* is favorable for the complex in triplet state. 2AP conversion to H2O2 remains unfavorable, but ΔG significantly decreases. Thus, inorganic iron can facilitate two-step 2AP oxidation to H2O2. For 2AP complex with Fe-oxo-heme-imidazole in singlet form, all oxidation steps are only slightly endothermic, which makes PHS-2AP complex a potential source of ROS, including HO2* and H2O2. Both inorganic and heme iron induce the partial negative charges on O2 atoms that may facilitate H-atom abstraction.
To estimate the correlation between L-3HOK accumulation and oxidative stress development in D. melanogaster, we measured TAC and LP in male heads at different stages of adult life. cd strain (2.9X L-3HOK accumulation [24,27]) was compared to the wild-type strain Canton-S (CS) (Fig 7). To check the possible influence of the fly keeping temperature and the quality of samples purification (centrifugation strength) on TAC, we performed several experiments varying these parameters (a.–e.). TAC is decreased in cd relative to CS in 5 day-old flies kept at 22°C (e.) and 29-day-old flies kept at 25°C (a.). There is a significant relative TAC decrease in cd at the 13th day (a., d.) and the 21th day (b.–e.). We do not see any significant age-dependent changes in CS TAC, while cd shows TAC decrease at the 5th–13th day interval (a.) or TAC increase from midlife to the 29th day (c.–e.). Thus, a strong tendency to TAC decrease in cd is observed, being the most reliable for middle-aged (13–21-day-old) flies. LP is somewhat lower in cd than in CS, however, there are no statistically significant LP differences between two strains (f.). At the same time, LP decreases in cd after the 5th day, remaining comparatively low during the total studied period, whereas in CS LP decrease is observed only at the 21th day. Thus, two effects can be observed simultaneously: 1. the age-dependent decrease in cd TAC, probably corresponding to the toxic effects of L-3HOK accumulation in fly heads; 2. the age-dependent increase in cd ability to inhibit lipid peroxidation, possibly connected to the peroxy radical scavenging activity of L-3HOK and its dimers.
Kynurenines, the products of tryptophan catabolism, are compounds with a broad spectrum of neurobiological activities, involved in the development of various neuropathologies, such as HD, PD, AD, schizophrenia, depression, dementia-like disorders, and suicide [2,28,37]. Most of kynurenines are able to modulate cell receptors, but there are no known specific targets for them. For all the aromatic amino acids, their major catabolism pathway is biochemically separated from that leading to neuromediator production, which makes it possible to be tightly regulated in organism. The neuromediator synthesis includes the stage of decarboxylation: dopamine (DOP), the first catecholamine, is formed from dioxyphenylalanine (DOPA), histamine–from histidine, and serotonine–from 5-hydroxytryptophan. Some kynurenines, like KYN and L-3HOK, can be decarboxylated in liver and brain to kynuramines, a specific class of biogenic amines with rather poorly studied biological effects. Kynuramine itself competes with tryptamin receptor in the rat brain, also being an antagonist of α1-adrenergic receptor, while the melatonin products, 5-methoxylated kynuramines, are radical scavengers [38].
The main physiological function of kynurenines is to regulate the organism response to adverse environmental effects, such as infection or stress, providing defensive reactions on molecular and behavioral levels. While the hepar IDO is constutively active, being responsible for basic tryptophan catabolism, the brain IDO and some other KP enzymes are activated by different harmful agents like bacterial and viral products, as well as inflammatory cytokines [2,39,40]. L-3HOK is mainly produced by microglia, whereas KYNA is formed in astrocytes. Kynyrenine 3-monooxygenase is one of the key enzyme affecting the balance of neurotoxic and neuroprotective kynurenines, shifting KP from predominant KYNA production towards the generation of L-3HOK and QUIN, responsible for neurodegeneration and depression [41].
The neurotoxic effects of hydroxykynyrenines are caused by ROS hyperproduction, while QUIN is an excitotoxin activating NMDAR. QUIN is not synthesized in Drosophila. This makes cd, a Drosophila mutant with enzymatically disrupted XAN synthesis, a useful model to study specific effects of L-3HOK accumulation on neuropathology development [28]. 5-day-old cd male demonstrate a middle-term memory decrease after heat-shock–an effect prevented by adding phenolic antioxidants in the food media of developing flies [42]. Nowadays, there is a renaissance of interest in science to chemistry and physiology of ommochromes, the invertebrate pigments, including Drosophila brown pigments XAN and DXAN produced by L-3HOK dimerization, which also participate in many biological processes, such as electron transport and free radical trapping [43]. In mammals, different proteins can catalyze R-2AP oxidative dimerization, such as hemoglobin and peroxidase, including heme or copper ions in their catalytic sites [21]. Thus, dimerization of hydroxykynurenines is involved in many physiological processes connected to ROS metabolism.
It is still not completely clear which ROS are directly generated during kynurenines' dimerization, as well as what is the role of enzymatic and non-enzymatic pathways in their production. H2O2 and iron play an important role in 3HOK-induced cytotoxicity [44]. H2O2 seems to be the chief product of non-enzymatic 3HOK oxidation, as catalase but not superoxide dismutase blocks 3HOK-induced cell death [45]. H2O2 then generates highly toxic hydroxyl radicals that damage nervous cells [16]. However, H2O2 also can be formed due to O2*- disproportionation [36]. 3HOK and 3HAA generate H2O2 in a Cu(II)-dependent manner, likely through anilino or phenoxyl radicals formation [46]. 3HOK adducts with lens protein are also generated upon the ultraviolet illumination, probably contributing to age-related development of cataract [47]. These processes are difficult to study experimentally, as most ROS are very active and easily convert to each other. The computational approach may be useful to reveal the mechanism underlying kynurenines’ dimerization leading to ROS formation.
In our studies, we were specially focused on the oxidative dimerization of hydroxykynurenines L-3HOK and 3HAA, as well as their structural precursor 2AP. The dimerization pathways, including both enzymatic and putative non-enzymatic stages, were computationally studied, considering the free energies of conversion of intermediates and their electron donation/acceptor power. For simplicity, the process of H-atom abstraction was modeled as a proton-coupled electron transfer (PCET) without separation of a proton and an electron migration stages.
In general, both H-atom and electron dissociation energies decrease as the dimers oxidation progresses, in agreement with [19]. However, BDG and IP values are minimal for the compounds with intact 2AP moiety, such as D7, being higher for the radical and/or quinoneimine forms. BDG is extremely low for C-H bond in non-enzymatically formed R-2APq conjugate D1', which dissociation seems to be energetically favorable due to the concomitant aromatic radical formation. The rate of C-H bond dissociation is lower compared to that for O-H bond in D1' tautomer D2'. However, D1' reaction with O2 is irreversible that makes it a potent source of hydroperoxyl radicals. The ionized COO- groups in 3HAA dimers potentiate both electron- and H-atom donation, in agreement with [32]. 2AP moiety within the non-enzymatically formed dimers D7 and D2' can react with O2 generating H2O2 via the consecutive O-H and N-H bonds dissociation. In addition, H2O2 can be formed as a by-product of L-3HOK-D10 oxidation to XAN. The electron donation is more thermodynamically favorable in the lipid surrounding than in aqua.
The chemistry of kynurenines’ dimerization can be better understood on comparing them with catecholamines, the structurally related compounds with similar redox properties. Catecholamines are widely involved in multiple neurological processes in brain and periphery. Like L-3HOK, DOPA and DOP inhibit lipid peroxidation in complex with α-tocopherol at low pH, becoming pro-oxidants at pH 8.0 due to the interaction of semiquinone anion radical with O2 [48]. Catecholamines are susceptible to the oxidative dimerization to melanin producing ROS, toxically affecting the nervous system. Structures in brain accumulating catecholamines are prone to age-dependent or disease-induced degeneration, such as substantia nigra in PD. DOP interacts spontaneously with O2 in two steps, firstly producing O2*- and o-semiquinone, which is then oxidized to 2APq analogue o-quinone. o-Quinone produces aminochrome and the eumelanin precursor 5,6-indolequinone, which both can form adducts with proteins involved in PD development, such as α-synuclein, parkin, actin, and tubulin [49,50]. The redox-active metals play a special role in the aforesaid processes: Fe(II) increases the rate of DOP oxidation to o-semiquinone by several orders of magnitude. O2*- interaction with DOP is energetically unfavorable; however, O2*- is much more powerful oxidant for o-semiquinone compared to molecular oxygen. H2O2 can be formed by O2*- oxidation or disproportionation, as well as directly by leucoaminochrome or 5,6-dihydroxyindole oxidation [35,36].
For 2AP and its derivatives, L-3HOK and 3-HAA, we can suppose similar two-step oxidation by O2, leading firstly to semiquinoneimine (R-2AP*) and then to quinoneimine (R-2APq). O2 is triplet in the ground state, having two unpaired electrons. Hence, it is unable to react directly with most organic molecules with all paired electrons, except for those that can produce free radicals via one-electron transfer [51]. Upon binding O2, the ions of transition metals, such as iron and copper, easily change its spin state, making possible two-electron transfer to O2 from a substrate. One such activating compound is the heme group of globins, as well as of some peroxidases and dioxygenases. Fe-oxo-heme complex is formed in globins, which carry O2 in blood and muscle tissue. Similar complex is thought to catalyze the oxidation of L-tryptophan to N-formylkynurenine in the active site of TDO. Heme-containing enzymes reduce O2 using PCET mechanism, leading to formation of Fe(III)-OOH* heme complex (ferrix-hydroperoxy intermediate, also known as Compound 0). In peroxidases, this complex is formed upon H2O2 binding to ferric heme. Being protonated, it rapidly abstracts a water molecule, turning to Fe(IV) = O+* heme (Compound I). After that, the active site should be restored by some appropriate enzyme [52,53]. Compound I is believed not to be formed in TDO active site [52]. Hence, the cleavage of O-O bond should occur after O2 binding to substrate.
L-3HOK, 3HAA and 2AP can bind to ferrous oxy-hemoglobin converting it to ferric met-hemoglobin, and vice versa [54,55]. Drosophila PHS seems to form similar complexes with R-2AP. In the case of two simultaneous H-atoms abstraction by O2, R-2APq and H2O2 would form and immediately dissociate from the active site. However, PHS is known to produce two water molecules upon 2AP enzymatic oxidation [21]. Hence, only one of two hydrogens, likely hydroxyl H-atom, should be transferred to O2 at the first stage. In the catalytic site of PHS, the oxidation seems to pass step-by-step without leaving unreduced oxygen forms from the enzyme to the extra medium. According to our data, both FeCl2 and ferrous heme significantly decrease the energies of two-step 2AP H-atoms abstraction by O2, making HO2* and H2O2 production energetically favorable or low-cost. HO2* protonation by some residue within PHS catalytic site probably assists O-O heterolysis in Compound 0, producing water and Compound I, likewise in cytochrome c peroxidase and horseradish peroxidase [56]. This prevents H2O2 formation by the enzyme, which therefore serves as a component of antioxidant system, prohibiting ROS formation at the first stages of R-2AP dimerization. At the same time, H2O2 formation seems to occur non-enzymatically upon R-2AP interaction with O2, especially in the presence of iron.
The hypothetic scheme of the whole PHS cycle is shown in Fig 8.
Here, the first H-atom abstraction by O2 generates 2AP* and Compound 0. Being protonated, possibly by His89, Compound 0 abstracts a water molecule and turns to Compound I, which is then partly reduced by the amine H-atom, converting 2AP* to 2APq. Then another 2AP molecule attaches 2APq and restores its amine and hydroxyl groups (D2 formation). The second hydroxyl H-atom abstraction by D2 generates the second water molecule, reducing PHS active site to ferric form in one or several steps. The ferric form is then reduced to the initial ferrous form, possibly with an electron donated by D3. Hence the enzymatic stage of 2AP cyclization is ended with D4(H+), which is then undergoes a non-enzymatic cyclization to D10.
The previously mentioned scheme assumes that a substrate remains in the active site of PHS until the full reduction of the O2-heme complex. PHS thereby prevents formation of toxic non-reduced oxygen forms during 2AP and R-2AP oxidative dimerization. As it has been proposed earlier [19], 2AP moiety is regenerated after conjugation via tautomerization including two H-atom migration to the quinone group. According to our data, D2 has the lowest BDG among the enzymatically converted 2AP derivatives, and D2/D3 have the lowest IP, which makes them easy to donate H-atom and electron, respectively. The following non-enzymatic stages seem to be the main source of ROS, such as HO2* and O2*-. Thus, PHS prevents ROS formation at the first stages of kynurenines’ oxidation, which are greatly facilitated in presence of transition metals. Some other unknown antioxidant protection system should be involved at the further stages. In cell, PHS may be in complex with the other enzymes, which can assist R-2AP final oxidation or detoxify ROS, the by-products of XAN and CIN formation. Some oxidant or oxidants should participate at the final stages, specifically attaching the last two H-atoms. For example, XAN in Drosophila can catalyze its own formation, serving as H-atoms and/or electrons acceptor [29].
Little is known about the age-dependent changes of kynurenines’ level in adult Drosophila, as well as about the molecular mechanisms of L-3HOK accumulation in cd. An interesting fact experimentally shown by us is that TAC is decreased in cd heads compared to the wild type CS, being lowest at the 13th-21th days of life, while LP is at the same level for both strains and specifically decreases in cd after the 5th day. These two effects seem to partially offset each other, as they are not too pronounced and not stably observed for all fly ages in all experiments. Changing the temperature of flies keeping from 22 to 25°C accelerates flies ageing, which can move the age of major TAC differences between two strains from the 21th to the 13th day. TAC is almost equal for both strains at 29th day when flies are kept at 22°C, remaining lower for cd at 25°C. Thus, high temperature may contribute to the oxidative stress development in old Drosophila. LP decreases in the aged cd and remains low throughout the studied period of life compared to 5-day-old flies. This might be caused by chronic accumulation of L-3HOK and/or L-3HOK-D, which antioxidant activity is primarily connected to H-atom donation power. The same accumulation would increase the possibility of HO2* and H2O2 formation, which in turn triggers the complex reactions leading to oxidative stress. The causal connection between L-3HOK accumulation in cd and TAC or LP decrease remains to be confirmed experimentally.
To summarize, we have computationally shown that H-atom and electron donation power of 2AP and hydroxykynurenines progressively increases upon their oxidative dimerization. Being regulated by PHS or the other enzymatic systems, this process does not lead to harmful effects associated with ROS hyperproduction. However, when the enzymatic activity is disturbed or L-3HOK level is too high, the same process begins to pass non-enzymatically, being accompanied by free radicals production, especially in the presence of transition metal ions. The total effect of L-3HOK accumulation seems to be noxious, though it can decrease some aspects of the oxidative stress, connected to lipid peroxidation. Thus, we can use two principle strategies to prevent the neurotoxicity caused by hydroxykynurenines’ self-oxidation: 1. To block the whole KP or some its stages in the affected tissue using synthetic inhibitors, as it had been proposed earlier [1,2]; 2. to elaborate the therapeutic approach to stimulate the activity of enzymes specifically converting kynurenines and ROS to non-active forms. Studying PHS function in Drosophila, its mechanisms of action, regulation, age-dependent activity changes, interaction with the other proteins, and the molecular nature of possible PHS damages in cd can help us in this.
To facilitate the readability of dimer names in text, all of them were coded as following: “X-Dn”, where “X” is the monomer name (2AP, L-3HOK, 3HAA, 3HAAi, and R-2AP–for the last three monomers), “D” stands for “dimer” or “dimers” and “n” is the dimer number according to scheme of dimerization (see Figs 1 and 2).“D'n” indicates dimers converted via the non-enzymatic pathway. The reaction numbers are shown with the letter “N”. The enzymatic dimerization was modeled according to the schemes given in [19, 21], with the following modifications: 1. R-2AP quinoneimine monomers were modeled in the uncharged form; 2. two consecutive stages of H-atom abstraction were added after D2 (D3, D4) and D7 (D8, D9) (Fig 1). The non-enzymatic dimerization was modeled according to the scheme in [20]: 2APq → D1' → D3' → D10, with the addition of several hypothetical intermediates (Fig 2). The influence of hydrophobic (heptane) and aqueous surroundings on R-2AP dimerization energies, IP, and EA was studied.
The structures of 2-aminophenol (2AP), 2-aminophenoxazin-3-one (2APX), L-3HOK, 3HAA, XAN and DXAN were taken from the PubChem Compound database [57]. The structures of all dimerization intermediates were constructed on the base of PubChem structures using Vega ZZ 3.0.3 [58]. The initial energy minimization and systematic conformational search of low-energy geometry for all non-radical compounds was performed with the help of Avogadro [59] and Confab [60] using MMFF94 force field [61]. The carboxylic groups of 3HAA and its dimerization products are in ionized form at physiological pH (7.4), while L-3HOK and its dimers are mainly in zwitterionic form [62]. The ionized 3HAA (3HAAi) and 3HAAi-D were modeled with total charge -1 and -2, respectively, as well as their uncharged forms. L-3HOK zwitterions are not stable in the gas phase, and their optimization is hardly to converge in water solution. Hence, L-3HOK and L-3HOK-D were modeled in the neutral form. The initial geometries of radical structures (after H atom abstraction or electron addition/ abstraction) were set equal to that for the paternal closed shell structures. For quinoneimine structures, there are two possible conformers with N-H bond orientations: a. towards C = O bond, b. away from C = O bond. The B3LYP-optimized conformer a. had the lower energy compared to conformer b., except for 3HAAi and its derivatives. We performed all the calculations for the respective conformers with the lowest B3LYP energy. D1 and D6 have a chiral C atom. Since dimer conjugation is not stereospecific [19], all the calculations have been performed for a single optical isomer of D1 and D6, as well as their derivatives D1' and D5'/D6'.
All quantum chemical calculations were performed using Firefly 8.1.0 partially based on the GAMESS (US) [63] source code. Firefly 8.1.0 was kindly provided by Alex A. Granovsky [64]. For 2AP complexes with ferrous forms, the geometries were fully optimized using density functional theory (DFT) at B3LYP/6-31G(d) level (I). For the other compounds, the optimization was performed at B3LYP/6-311G(d,p) level (II) [65–67]. B3LYP1 version of B3LYP was used. All closed shell molecules were calculated in a singlet state, using restricted DFT method. Doublet state was used for radical structures and triplet–for D6' and complexes with molecular oxygen, using the unrestricted DFT method. All R-2AP-D radicals were optimized using the corresponding closed shell form as the initial conformation. For the most complexes with ferrous forms, the restricted open-shell DFT method was used. The symmetry point group was set as C1 for all compounds. Hessian matrix, vibrational frequencies and thermal corrections to the total energy at 298.15 K were calculated using the same method as for the geometry optimization. The nature of all stationary points was determined by evaluating the vibrational frequencies. Several of the ion-radicals, mainly those with the ionized carboxylic groups, are not stable in the gas phase, and their optimization was performed in water solution using DPCM model [68], without cavitation, dispersion and repulsion free energies.
For the optimized structures, single point energy calculations were performed at B3LYP/6-311+(O)+G(d) level (III). The calculations were performed both in the gas phase and in heptane or water solution at 298.15 K using DPCM. Due to the bad SCF convergence, p polarization functions at H atoms were omitted, and the diffuse functions were added only to O and H atoms. To check whether it affects the relative order for X-H bond dissociation energies, we performed the linear regression analysis for 36 X-H dissociation energy values calculated at level II and III in the gas phase, where X is O, N or C-atoms: energy (III) = 0.977 x energy (II)-0.916. The energy difference are maximal for O-H dissociation energies (-4.476±0.787 kcal/mol). Hence, at level III, O-H bond dissociation energies are somewhat decreased relative to those for N-H and C-H bonds. This may be caused by the greater opportunity for the unpaired electron to be delocalized on O-atom diffuse orbital. For proton, GFREE was set to be -6.28 kcal/mol in the gas phase and -272.18 kcal/mol in water solution, as in [69]. In literature there are no data on proton GFREE in heptane. Hence, we calculated the summary ΔG for the reaction pairs N7, N8 and N13, N14, including the sequential protonation and deprotonation of R-2AP-D.
Gibbs free energy was obtained from the vibrational frequency calculations at 298.15 K, using unscaled frequencies:
G=ET+GCORR
(2)
where G is Gibbs free energy, ET is a total energy of the optimized molecule (for the gas phase) or a total free energy in solution (for heptane and water solution) calculated at level III, GCORR is a thermal correction to G calculated at level II.
The free energy change (ΔG) was calculated as follows:
ΔG=GRAD+GH–GW(H-atom dissociation)
(3)
ΔG=GDIM−GM1–GM2(dimerization energy)
(4)
where DIM is dimer, H is hydrogen, M1 is monomer 1, M2 is monomer 2, RAD is radical, W is the whole molecule (before H abstraction).
Ionization potential (IP) and electron affinity (EA) were calculated as follows:
IP=GCAT–GW
(5)
EA=GW–GAN
(6)
where AN is anion-radical (one-electron addition), CAT is cation-radical (one-electron abstraction).
The Mulliken electronegativity (χ) was calculated as follows:
χ=(IP+EA)/2
(7)
For complex with molecular oxygen, transition states (TS) and corresponding local minima were optimized at level II. Intrinsic reaction coordinates (IRC) calculations [70] were performed for all TS species at the same level to confirm that anticipated reactants (R) and products (P) are connected to TS.
ΔG is corrected reaction activation energy:
ΔG#R=GTS−GR
(8)
ΔG#P=GTS−GP
(9)
ΔGP−R=GP–GR
(10)
where GTS, GR and GP are free energies of transition state, reactants and products of reaction; ΔG#R and ΔG#P are reaction barriers for reactants and products; ΔGP-R is the free energy change for reactants conversion to products, respectively. The rate of reaction (M-1s-1) between antioxidant and radical was calculated as in [32,71] using conventional TS theory:
k(T)=Ix(kBT/h)x[exp(−ΔG#/RT)]x24.3xA(T)
(11)
where I is the reaction pathway degeneracy (1 for the all compounds), kB is Boltzmann's constant, h is Planck's constant, ΔG# is ΔG#R or ΔG#P, 24.3 is a multiplier used to convert the units from 1 atmosphere standard state to 1 M standard state, and A(T) is a temperature-dependent factor corresponding to quantum mechanical tunneling, approximated by the Wigner method [72]:
A(T)=1+(1/24)x(1.44νi/T)2
(12)
where νi is the imaginary frequency (cm-1) whose vibrational motion determines the direction of the reaction.
Atom coordinates of the optimized structures are given in S1 Dataset. G and GCORR values calculated at level II for ΔG, IP, EA, and k(T) are given in S3 Table.
For 2AP, L-3HOK and 3HAA, ΔG, IP and EA values calculated at level II and III in the gas phase are highly correlated (R2 > 0.99; Table VI in S2 Table). The same is true for 3HAAi ΔG (II) and (III). Hence, both methods give the same relative order of energy values. R2 is ~0.97 for 3HAAi IP (II–III) and ~0.77 for 3HAAi EA (II–III). The correlation decrease is possibly connected to facts that many 3HAAi forms are unstable in the gas phase and were optimized in water solution, but their single-point energies are given for the gas phase. Solvation effects significantly affect the relative order of IP and EA values, as it were shown here (Figs 4 and S2 and S3). For log k(T), there is a strong correlation between values calculated at level II and level III in gas, heptane, and water (Table VI in S2 Table).
For IP and EA, the average GCORR is rather small, being 0.241±0.415 and 1.949±0.458 kcal/mol, respectively (n = 61). For X-H bond dissociation, the average GCORR is -13.391±0.477 kcal/mol, and the average thermal enthalpy correction (HCORR) is -7.713±0.317 (n = 44). Hence, X-H BDG was ~5 kcal/mol lower than homolytic bond dissociation enthalpy (BDE). Here, we calculated X-H BDG instead of BDE for the unification of all energetic effects of dimerization reactions. For 2AP, L-3HOK and 3HAA, O-H BDE calculated at level II is ~71.0 kcal/mol (S3 Table), whereas the experimental 2AP O-H BDE is 81.3 kcal/mol [73]. The calculated O-H BDE of hydroxykynurenines may differ from the experimental absolute values, but their relative order is satisfactory reproduced by computational approach [32].
BSSE correction [74] was performed at level II for several calculations, including reactions with H-atom abstraction, as well as D1 formation from two monomers (A, B) separated by N-C bond (S3 Table). The negative BSSE correction decreases X-H BDG making it less positive and increases dimerization energy making it less negative. For D1 formation, BSSE correction is -5.373±2.904 kcal/mol (n = 4), being higher for the negatively charged D1-3HAAi form than for the neutral D1-3HAA form. For H-atom abstraction, BSSE correction values and data spread are significantly lower (-1.543±0.299 kcal/mol, n = 6). Thus, we did not consider BSSE while comparing X-H BDG values.
Homology modeling of D. melanogaster PHS in complex with the rigid heme group was performed with the help of MODELLER 9.14 software [75] using goat lactoperoxidase as a template (PDB ID: 2ojv, residues 13–596; 35% sequence identity). After the loops modeling, the model structure was dynamically optimized. The quality of the 10 model structures was estimated using SaliLab Model Evaluation Server [76–78]. The structure with minimal predicted RMSD (5.352 Å) between Cα atom coordinates in the native structure and the model, which also had the most negative value of MODELLER energy, was selected for docking. O2 position relative to heme was set the same as in human oxy-hemoglobin (PDB ID: 6bb5 A), except Fe-O distance was 1.95 Å instead of 1.85 Å. The partial charges for heme group were taken from [79]. For O2, the partial charges on atoms proximal and distal to Fe were -0.07 and 0.07, respectively. All the substrates were in the appropriate ionic form: 3HAA and its derivatives–the ionized COO- group (3HAAi), L-3HOK and its derivatives–the ionized αCOO- and αNH3+ groups (L-3HOKzi, zwitterion form). The optimal conformations were found using MMFF94 force field. The automatic docking of the flexible R-2AP substrates to the rigid PHS catalytic site was performed using Autodock 4.2 software [80] with the help of Lamarckian genetic algorithm [81]. PHS model and docked substrates are presented in S2 Dataset.
The illustrations were prepared using ChemSketch [82] and VMD [83].
Pearson correlation and linear regression were calculated using Social Science Statistics online resource [87]. Statistical significance was estimated using two-sided randomization test at p <0.05 [88].
|
10.1371/journal.pcbi.1004788 | H3 Histone Tail Conformation within the Nucleosome and the Impact of K14 Acetylation Studied Using Enhanced Sampling Simulation | Acetylation of lysine residues in histone tails is associated with gene transcription. Because histone tails are structurally flexible and intrinsically disordered, it is difficult to experimentally determine the tail conformations and the impact of acetylation. In this work, we performed simulations to sample H3 tail conformations with and without acetylation. The results show that irrespective of the presence or absence of the acetylation, the H3 tail remains in contact with the DNA and assumes an α-helix structure in some regions. Acetylation slightly weakened the interaction between the tail and DNA and enhanced α-helix formation, resulting in a more compact tail conformation. We inferred that this compaction induces unwrapping and exposure of the linker DNA, enabling DNA-binding proteins (e.g., transcription factors) to bind to their target sequences. In addition, our simulation also showed that acetylated lysine was more often exposed to the solvent, which is consistent with the fact that acetylation functions as a post-translational modification recognition site marker.
| Post-translational modification (PTM) of histone tails is an important component of epigenetics. Acetylation of histone tails generally functions to activate gene expression, though the molecular mechanism is not well understood. We used enhanced sampling simulation to examine the impact of acetylation on the structure of the histone H3 tail within the nucleosome. The results suggest acetylation makes the H3 tail conformation more compact and enhances dissociation of nucleosomal DNA from the histone core. Further, the acetylated lysine was more exposed to the solvent, which is consistent with its role as a PTM recognition site marker. These findings increase our understanding of the impact of PTM on nucleosome stability and dynamics and on the higher order structure of chromatin.
| In eukaryotic cells, the genome is compactly stored within the nucleus as a complex with proteins. The basic structural unit of the complex is the nucleosome, which is composed of 146 or 147 base pairs of DNA wrapped around a histone octamer consisting of two copies each of histones H3, H4, H2A and H2B[1]. Posttranslational modification (PTM) of histones, including acetylation, methylation, phosphorylation and ubiquitination, has been studied since the early days of epigenetics. PTM occurs most often in the N-terminal regions (tails) of histones, and the precise locations of PTM are closely linked to specific DNA functions and biological events[2]. This prompted Strahi and Allis to propose the histone code hypothesis[3]: “multiple histone modifications, acting in a combinatorial or sequential fashion on one or multiple histone tails, specify unique downstream functions.” For example, acetylation of lysine residues on the H3 and H4 tails generally activate transcription[4, 5]. However, because the tails are intrinsically disordered with no static conformation, details of the molecular mechanism by which PTM exerts its effects remain unclear[6, 7], and there are no decisive clues as to which extent PTM affects the conformations of histone tails, nucleosomes or chromatin.
Histone tails play a key role in partial charge neutralization of nucleosomes and contribute to nucleosome aggregation. For example, Allan et al.[8] found that nucleosomes without histone tails do not aggregate, and Krajewski et al.[9] reported that the H3 and H4 tails were especially important for chromatin folding. Histone tails contain a large number of positively charged lysine and arginine residues, and thus preferentially contact DNA via electrostatic interactions. The positive charges partially neutralize the negative charges of nucleosomal DNA and reduce the electrostatic repulsive forces among nucleosomes, thereby mediating nucleosome aggregation. On the other hand, lysine acetylation neutralizes the positive charge on lysine residues, which weakens the interactions between the tails and the DNA. Lee et al.[10] suggested that acetylation leads to dissociation of the tails from the DNA and/or induces a change in the DNA configuration within the histone core, which allows transcription factor binding. However, the impact of acetylation on nucleosome structure or the higher order structure of chromatin is not yet known.
The difficulty in understanding how the conformation of nucleosomes or chromatin is changed through PTM is attributable to a lack of conformational information about histone tails. Although the crystal structures of nucleosomes have been solved, the conformations of the histone tails could not be determined because they are structurally flexible unless in complex with a specific protein. In addition, the existence of DNA strongly affects conformation of the tails, because positively charged histone tails favorably interact with negatively charged DNA. Conformational sampling for the histone tails should be investigated in the presence of nucleosomal DNA. Therefore, previous simulation studies using an enhanced sampling method in explicit water obtained conformational ensembles for histone tails isolated from the nucleosome[11]. These simulations showed that histone tails were not just flexible chains, but chains that had intrinsic conformational preferences. To the best of our knowledge, however, there are no simulation reports in which tail conformation was studied in the context of the nucleosome, and so the effects of histone tails on the structure and stability of nucleosomes remain unclear.
In the present study, we used computer simulations to ask how the H3 tail behaves within the nucleosome and how acetylation of H3 on K14 induces changes in the chromatin conformation. Because histone tails are intrinsically disordered and thus difficult to experimentally characterize, molecular dynamics (MD) simulation was used to investigate the flexible peptide. Using an enhanced sampling method, adaptive lambda square dynamics (ALSD)[12], we carried out conformational sampling of H3 histone tails, with and without acetylation at K14, and determined which conformation is preferred in the presence of nucleosomal DNA. In addition, we investigated how a single acetylation affects tail conformation and the solvent accessibility of the acetylated residue.
In conventional MD simulations, the conformations of histone tails cannot be sufficiently sampled within a practical simulation time, because attractive electrostatic interactions strongly bind the tails to the DNA. We therefore applied ALSD enhanced sampling to obtain various conformations of the H3 tail within the context of a nucleosome (Fig 1A). We assessed the interaction of the H3 tail and the DNA based on the contact surface area (CSA) between the two. The CSA is the surface area of the DNA covered by the H3 tail and was computed as the difference between the solvent accessible surface area (SASA)[13] of the DNA, with and without the H3 tail. For the SASA calculation, atomic radii were set at 1.7 Å for carbon, 1.625 Å for nitrogen, 1.48 Å for oxygen, 1.87 Å for phosphate and 1.4 Å for water molecules. Hydrogen atoms were excluded from the calculation. In the ALSD simulation, a scaling factor, λ, was introduced to scale the potential energy of the H3 tail. As λ decreased, the potential energy was scaled down, and changes in the conformation of the tail were enhanced. Fig 1B and S1 Fig show the relation between the CSA distribution and λ in the unacetylated and K14-acetylated (K14ac) systems. Lower CSA values indicate dissociation of the tail from the DNA and elongation in the solvent. Our simulation results show that the CSA varied widely (0 to 1880 Å2) and declined with decreasing λ, demonstrating that ALSD sampled a broad range of H3 tail conformations. A sampled conformation at CSA = 0 is shown in S2 Fig.
Ensembles reweighted to λ = 1 only give conformations seen under realistic conditions (corresponding to a temperature of 300 K) and have high CSA values. The average and standard error from 256 ALSD simulation runs with trivial trajectory parallelization (TTP)[14] were 1135 ± 10 and 1062 ± 9 Å2 for the unacetylated and K14ac system, respectively. These large CSA values indicate that the H3 tail is in the vicinity of the DNA, and does not adopt an elongated conformation (S2 Fig). The elongated dissociated H3 tail conformations are thermally unstable in the canonical ensemble at λ = 1, with and without K14 acetylation. Note, however, that a nucleosome model (PDB ID: 1KX5[15]) often used in the literature and textbooks has much lower CSA values (209 and 281 Å2 for chains A and E of the H3 histones) than those obtained in the present study, which indicates the tail is protruding from the nucleosome and might raise the possibility of giving an incorrect impression of the H3 tail conformation within the nucleosome. Our simulation results show that with or without K14 acetylation, H3 tails bind to the DNA, though K14 acetylation decreases the CSA slightly. We suggest that a conformation in which the H3 tail is dissociated from the nucleosomal DNA is not thermodynamically stable, even with K14 acetylation. Addition of acetylation at other sites on the H3 tail or other binding factors may be required to stabilize the dissociated state.
Our simulation showed that, unlike the structured histone core domain, the H3 tail had no specific native conformation within the nucleosome, with or without K14 acetylation. Nonetheless, the H3 tail was not just a disordered region; it did have conformational preferences. To characterize the conformations of the H3 tail, we used the program DSSP[16] to analyze the secondary structure in the reweighted ensembles. Fig 2A shows the average α-helix content ratio with the standard error for each residue. Residues 2–12 and 17–28 had high helix content in the H3 tail, indicating the tail has the ability to form α-helix within the nucleosome, which is consistent with CD results from Baneres et al.[17]. It has also been reported that a tail segment isolated from H3 has the ability to form α-helix[11]. However, ours is the first MD simulation of the H3 tail within the nucleosome and confirms the H3 tail has the ability to form α-helix. Dreveny et al. [18] reported that residues 4–11 in the H3 tail form α-helix within a complex composed of the tail and histone acetyltransferase (PDB ID: 4LK9, 4LKA and 4LLB for unmodified, K9ac and K14ac H3 tail, respectively[18]). Our simulation suggests the same region to which the enzyme binds adopts an α-helical conformation, irrespective of the acetylation. In addition, the α-helix content was clearly increased with K14ac, particularly at residues 14 to 19 (Fig 2A). The increase of helical content by acetylation is consistent with previous experimental results by CD spectra [17, 19]. We depict the α-helix obtained from the conformational ensembles in S3 Fig, where two α-helices are formed in the tail.
We also calculated the radius of gyration (Rg) of the H3 tail. The averages and standard errors for the unacetylated and K14ac systems were 13.90 ± 0.14 and 13.14 ± 0.10Å, respectively. Although the difference was small, the K14ac H3 tail tended to be more compact than the unacetylated tail. In particular, the ratio of Rg values larger than 15 Å was lower in the tail with K14ac (Fig 2B). These results indicate that K14ac causes the H3 tail to assume a more compact conformation, which is consistent with the increase in α-helix content.
Contact between the linker DNA and histone tail was investigated and is shown in S4 Fig, parts A and B. The unacetylated tail made contact with the DNA at the −9 to −6 and 8 to 12 bp positions (the root of the H3 tail was defined as 0 bp) more frequently than did H3 K14ac, but less frequently at the 1 to 4, and 6 bp positions (S4 Fig, part B). The K14ac tail contacts the DNA nearer to the root of the H3 tail. This is consistent with our finding that the K14ac tail adopts more compact conformations. The contour maps of the spatial distributions of the H3 tail and nucleosomal DNA (S5 Fig) were constructed as follows. Conformations from the reweighted ensembles were structurally aligned to the histone core region of a corresponding reference structure taken from the 1KX5 model without the H3 tail. We then calculated the frequencies that heavy atoms in the H3 tail, and DNA appeared in predefined 3D space grids. Consistent with the CSA result, the distributions clearly showed that the H3 tail was always located near the nucleosomal DNA in both systems, with and without K14 acetylation. In the simulations, we also observed conformations in which the linker DNA was tightly wrapped around the histone core, as was seen in the crystal structure, as well as conformations in which the linker DNA was partially dissociated from the histone core.
To determine the degree to which the H3 tail conformations differ in the presence and absence of K14 acetylation, we calculated differential maps for the spatial distributions of the DNA and H3 tail with and without acetylation of the tail (Fig 3). The difference in the distributions of the H3 tail indicates the preferences of the conformation with and without acetylation. The unacetylated tail distributed broadly along the DNA (blue region in Fig 3 top), while the K14ac tail was more likely to be distributed more compactly around the root of the tail (red region in Fig 3 top), as indicated by differences in the Rg distributions shown in Fig 2B. In addition, dissociation of the linker DNA was enhanced in the K14ac system (Fig 3 bottom, red region) as compared to the unacetylated system (Fig 3 bottom, blue region).
We evaluated the exposure of each residue in the H3 tail to the solvent to assess the accessibility of K14ac. Generally speaking, positively charged residues such as lysine and arginine energetically favor exposure to the solvent or contact with negatively charged molecules such as nucleosomal DNA. For acetylated lysine residues which are charge-neutralized, solvent exposure might be less energetically favorable, but they provide a binding site for transcription factors and so must be exposed to solvent.
To investigate lysine exposure with and without acetylation, we used two indices for each side chain terminal atom in lysine, arginine, and K14ac residues: the DNA contact ratio and the solvent exposure ratio. In these analyses, we considered only heavy atoms, and atomic radii used were the same ones used for the CSA analysis. The exposure ratio was defined as the relative SASA of each terminal atom to those when the side chain was fully extended. We used as terminal atoms, the NZ atom for lysine, NH1 and NH2 for arginine, and OT and CT for K14ac. Fig 4A shows that the contact ratio for K14ac was obviously lower than that for the arginine and other lysine residues, indicating that acetylation greatly reduced the contact with the DNA. Fig 4B shows the solvent exposure ratios. The exposure of K14ac was slightly higher than that of unacetylated lysine, although the difference is smaller than was observed for the contact ratio. Our simulation suggested that this is because the terminal atoms of K14ac are sometimes buried in the tail. Note that the contact ratio analysis considered contacts only with DNA atoms; other atoms in the tail were not counted. These results show that the solvent exposure of the terminal atoms of K14ac is greater than or at least equal to that of unacetylated lysine residues, which barely contacted the nucleosome DNA. Thus acetylation of K14 affects the conformations of the tail and the DNA and mediates K14 exposure to the solvent, making it available for recognition by regulatory proteins (e.g., transcriptional factors).
These analyses also revealed that there are differences between lysine and arginine residues. Fig 4 and S6 Fig show that arginine residues have consistently higher DNA contact ratios and lower solvent exposure ratios than lysine residues, irrespective of their location in the tail: arginine’s contacts with DNA are stronger than lysine’s, and lysine is easier to expose to the solvent than arginine. These characteristics may be important for the interaction between proteins and DNA. For example, it was previously shown that substituting lysine for arginine in H3 histone induces unwrapping of the nucleosome DNA[20]. Although it is generally thought that arginine and lysine have similar characteristics and lysine is sometimes substituted with arginine to see the suppressive effect of acetylation, our results suggest that such substitution might induce unexpected results due to difference in the behavior of the two amino acids.
Although histone tails are biologically important regions, details of their conformations in the context of nucleosomal structure are not well understood, in large part because of their conformational flexibility. This lack of structural information limits our understanding of how the tails contribute to gene regulation. In this work, we used ALSD simulations to elucidate some of the molecular details of H3 tails, with and without acetylation, and clarified the differences between their ensembles.
NMR studies [6, 7] suggested that the flexible H3 tail extended at least to residue 35 of H3 histone. Further, Gao et al.[7] showed that the side chains of two residues, H39 and Y41, were likely immobile, and K36 to P38 were experimentally indeterminable. Our analysis of phi-psi backbone distributions agreed well with these experimental results. It suggested that residues 1 to 36 except A15, P16, A29 and P30 were flexible, while residues after K36 were immobile (S7 Fig). In addition, phi-psi plots in the simulation clearly showed an increase in α-helical content at positions 14 to 18 upon K14 acetylation.
Wang et al. showed that acetylation increases the α-helical content of histone tails both in the nucleosome and in solution, irrespective of their interaction with the DNA [19]. Note that in their experiment, all four core histones were acetylated to some degree. They also speculated that the increase in α-helix mediated by acetylation would limit the region of the DNA with which the tail could interact, as we inferred from our simulations. Interestingly, they showed that the increase in α-helix content in the tails at residues 14 to 18 was independent of their interaction with DNA. Consistent with those findings, our simulation demonstrates that the tail conformation causes a change in the DNA dynamics, and suggests that acetylation increases α-helical content. We inferred this increase is due to the favorable, electrostatic interaction of K14ac and K18 when adopting the α-helix. In H3 tail, Lys residues reside every four or five in sequence. These Lys residues are aligned on one side and stacked closely when the tail forms an α-helix. Acetylation of Lys removes unfavorable, electrostatic interaction between Lys residues, thereby increases the α-helical content.
Our simulation and FRET data [21–23] both suggest that acetylation of H3 tail enhances DNA dissociation. A FRET experiment showed the acetylation of H3 tails is associated with DNA unwrapping while that of H4 tails not [21], which is consistent with our simulation result. Controversially, another FRET experiment on nucleosome core particle with a 147 bp DNA fragment showed that acetylation did not cause change in relative FRET efficiency [22]. This is also consistent with our simulation, since corresponding average distances between the bases at positions -68 and 7 in our sampled conformations with and without acetylation did not show significant difference (S8 Fig, parts A and C). Nevertheless, the distances between the linker DNA and inner DNA (at positions -83 and 0) were significantly different, suggesting that the H3 K14ac still affects the dynamics of linker DNA (S8 Fig, parts B and C).
Here, we try to speculate the dissociation mechanism based on our atomic model simulation results. In the unacetylated system, the H3 tail distributed broadly along the DNA, and it seemed to function as a scaffold between the linker DNA and the nucleosome core region, preventing the unwrapping. In the acetylated system, the H3 tail became more compact in the region around the root of the tail, and the tail formed a conformational cluster with positive charges on the surface of the nucleosome. This cluster may elevate the linker DNA, inducing the unwrapping from the histone core. Our simulation results provide a molecular evidence that the neutralization of the K14 charge by acetylation affects the DNA configuration by changing the tail conformation and the point of contact with the DNA, though the H3 tail itself does not dissociate from the DNA.
Our simulation, as well as other NMR and FRET studies, suggests the structural impact of acetylation is subtle at the single nucleosome level, but it is sufficient to change the specificity of the PTM recognition site. The next step in simulation is to study chromatin aggregation and dissociation. What happens when hyper-acetylation or accumulation of acetylation occurs within chromatin? A coarse-grained simulation is one way to study such phenomena.
MD simulation is a promising approach to understanding the details of conformational states, but it remains beyond the computational power currently available to apply all-atom MD to study the entire chromatin structure. Instead, MD simulations based on coarse-grained models simplify the atomic representation to reduce the computational cost [24–28]. When carrying out such simulations, it is key to determine appropriate force field parameters for histone tails. There are no perfect coarse-grained parameters that can apply to all of biological systems. To stabilize the native conformation in a coarse-grained MD, parameters like Go-potential specialized for each system are often introduced into the simulations. Our simulation results provide information that could be utilized to determine such parameters using, for example, the inverse Monte Carlo method [29].
In our simulation, the tail did not behave as a Gaussian chain and always had contacts with the nucleosomal DNA. This indicates that histone tails do not behave like a random chain, but instead have conformational preferences. In addition, there are constraints on where the tail can be located. Development of a model that takes these features into account is necessary. Schlick’s group has already moved in that direction with their coarse-grained model [24–26]. We also suggest development of a force field that reflects the difference between lysine and arginine residues, as our simulation showed arginine interacted with DNA to a greater degree than Lys, and Lys was more exposed to the solvent. In most coarse-grained MD employing a one-residue one-particle model, lysine and arginine are represented as similar residues with the same positive charge and different vdW radii. The expression of this difference is important in an epigenetic point of view.
We performed conformational sampling simulations of histone H3 tails, with and without K14 acetylation, within the nucleosome structure to characterize the conformational states and study the impact of acetylation. The obtained conformational ensemble showed that 1) H3 tails with or without acetylation are located nearby the nucleosomal DNA and maintain contact with the DNA, and 2) parts of the H3 tail formed α-helix irrespective of K14 acetylation, but the tendency to form helix is stronger in the acetylated system. Acetylation slightly weakened the interactions between the tail and the DNA due to the charge neutralization and made the tail conformation more compact. We suggest the compaction induces unwrapping of the linker DNA so that transcriptional factors can access the DNA. In addition, we demonstrated differences in the characteristics of arginine, lysine and acetylated lysine, suggesting the necessity to develop a new force field for coarse-grained simulation that reflects these differences.
To obtain a conformational ensemble of histone H3 tails within a nucleosomal structure, we performed a simulation with a system that included both the H3 tail and the linker DNA. No nucleosome structures that included the atomic coordinates of both regions have been deposited in the PDB. We therefore constructed the system using two nucleosome models, 1KX5[15] and 1ZBB[30]. 1KX5 is a mono-nucleosome model with modeled histone tails. 1ZBB is a tetra-nucleosome model with linker DNA. Using those models, we constructed the hybrid nucleosome model shown in Fig 1. Histones from 1KX5 and DNA from 1ZBB were combined by superimposing their common histone core regions. In this way, we extended the DNA by 10 bp at the end, as compared to 1KX5. Then to reduce computational cost, only atoms within a sphere of 54 Å radius at the root of the H3 tail (a nitrogen atom in the 40-th residue) were considered in the simulations. The N- and C-termini of the trimmed amino acid residues were capped with acetyl and N-methyl groups, respectively. The system was then immersed in a sphere of explicit water: the center was the same as the sphere described above, the radius was 60 Å and the water molecules were equilibrated at 300 K and 1 g/cc in advance. Water molecules overlapping histones and DNA were removed. We exchanged some water molecules with Na+ and Cl- ions to neutralize the net charge and bring the ion concentration close to physiological (0.153 M). Ultimately, the system consisted of 90306 atoms (3554 atoms for DNA, 5886 atoms for histone, 34 chloride ions, 111 sodium ions, and 80721 atoms for water) for the unacetylated system (Fig 1). A second system for K14ac was also constructed in the same manner (total 90306 atoms, 3554 atoms for DNA, 5890 atoms for histone, 33 chloride ions, 111 sodium ions, and 80718 atoms for water).
We used force field parameters taken from an AMBER-based hybrid force field (ω = 0.75) [31], AMBER bsc0[32], TIP3P [33], and ion08[34] for the proteins, DNA, water molecules and ions, respectively. We used point charge parameters published by Papageorgiou on the Web (http://pc164.materials.uoi.gr/dpapageo/amberparams.php) for acetylated lysine. To maintain the trimmed nucleosome conformation, weak harmonic potentials as position constraints referring to the 1KX5 model were applied to heavy atoms in the vicinity of the water sphere boundary. Note that the position constraints were not applied to the linker DNA. To maintain hydrogen bonds for base pair formation, we also applied weak harmonic potentials as distance constraints. To avoid evaporation of water molecules from the water sphere boundary shown in Fig 1, a harmonic potential (force constant = 100 kcal/mol/Å2) was applied to the water oxygen atoms only when they were outside the boundary. We used the MD simulation program PRESTO ver. 3[35] extended by the authors. A time step of 2 fs was used. The SHAKE algorithm[36] was used to constrain the geometry of atom groups X-H (X is a heavy atom). The cell-multipole expansion method[37] was used to compute long-range electrostatic interactions, and the constant-temperature method[38] was applied to control the simulation temperature. We performed ALSD simulations[12] to realize an efficient conformational sampling of histone H3 tail. To speed up the sampling, we used trivial trajectory parallelization (TTP)[14], which searches a conformational ensemble with N (= 256, in this work) multiple, independent simulations starting from different initial conformations. We carried out iterative and productive runs for 36 ns × 256 runs = 9.216 μs and 30 ns × 256 runs = 7.680 μs, respectively (see the next section for the details of the iterative and productive runs).
ALSD[12] is a simulation technique to enhance conformational sampling in a predefined partial system, while conformations in the rest of the system were allowed to thermally fluctuate. For the ALSD procedure, we divided a system into two regions: a tail region (N-terminal 40 residues of a histone H3) and the rest (the nucleosome without the H3 tail, ions and water molecules). The total potential energy of the system, E, was decomposed into three terms,
E=Etail+Etail−rest+Erest,
(1)
where Etail, Etail-rest and Erest denote potential energy terms for intra-tail, inter tail-rest, and intra-rest regions, respectively. ALSD simulation is a canonical MD simulation with ALSD Hamiltonian,
HALSD=λ2Etail+λEtail−rest+Erest+K+mλλ˙2/2+RTlnP(λ,T),
(2)
where λ is an extra dynamic variable with the fictitious mass mλ, K is the kinetic energy of the system, λ is the velocity on the λ axis, R is the gas constant, T is the constant simulation temperature (in this work, T = 300 K), and P(λ, T) is a canonical probability distribution at λ. During the ALSD simulation, the variable λ moves on the λ axis, obeying HALSD and scales only the potential energy terms related to the tails. When 0 < λ < 1, interactions only for the tail are reduced, and conformational changes in the tail are enhanced. The last energy term is an umbrella potential to regulate the sampled λ range (0.6 < λ < 1.03 in this work). ALSD can realize a random walk on the λ axis if a priori unknown function P(λ, T) in the umbrella potential term is accurately estimated. The random walk facilitates overcoming energy barriers such as strong attractive electrostatic interactions between the tail and the DNA. In practical ALSD simulations, iterative runs of simulations are carried out to accurately estimate P(λ, T) before productive runs to obtain a conformational ensemble. In this work, we did 19 iterative runs each for the unacetylated and K14ac systems. A canonical ensemble at λ = 1 can be reconstructed using a reweighting scheme[12]. Unless otherwise stated, we used canonical ensembles reweighted at λ = 1 for conformational analyses.
|
10.1371/journal.pbio.1001532 | Reduction of the Cholesterol Sensor SCAP in the Brains of Mice Causes Impaired Synaptic Transmission and Altered Cognitive Function | The sterol sensor SCAP is a key regulator of SREBP-2, the major transcription factor controlling cholesterol synthesis. Recently, we showed that there is a global down-regulation of cholesterol synthetic genes, as well as SREBP-2, in the brains of diabetic mice, leading to a reduction of cholesterol synthesis. We now show that in mouse models of type 1 and type 2 diabetes, this is, in part, the result of a decrease of SCAP. Homozygous disruption of the Scap gene in the brains of mice causes perinatal lethality associated with microcephaly and gliosis. Mice with haploinsufficiency of Scap in the brain show a 60% reduction of SCAP protein and ∼30% reduction in brain cholesterol synthesis, similar to what is observed in diabetic mice. This results in impaired synaptic transmission, as measured by decreased paired pulse facilitation and long-term potentiation, and is associated with behavioral and cognitive changes. Thus, reduction of SCAP and the consequent suppression of cholesterol synthesis in the brain may play an important role in the increased rates of cognitive decline and Alzheimer disease observed in diabetic states.
| Diabetes is associated with an increased risk of Alzheimer disease, depression, and cognitive decline, but the causal link underlying these associations is unclear. We previously showed that in diabetic mice there is a reduction in brain synthesis of cholesterol, which is required for normal formation of synapses between neurons. Here we show that this deficit is caused, in part, by a reduction in the levels of SCAP, a protein known to help regulate cholesterol synthesis by promoting the relocalization, cleavage, and liberation of the key transcription factor SREBP2. These changes in cholesterol biosynthesis are rescued by treatment of the diabetic mice with insulin. When the level of SCAP in the brains of non-diabetic mice is lowered by genetic manipulation, there is a decrease in cholesterol synthesis in the brain, and this results in impaired signaling between neurons, memory deficits, and abnormal responses to stress. These findings indicate that the reduction in SCAP associated with diabetes can contribute to changes in cognitive function in this disease.
| The brain is the most cholesterol rich organ in the body, containing more than 20% of the sterol pool and almost all of the cholesterol is produced in situ [1]. Multiple in vitro studies have indicated that cholesterol in the brain is important for synapse biogenesis and vesicle formation [2],[3]. Cholesterol synthesis is a highly regulated process controlled by the master transcriptional regulator SREBP-2. SREBP-2 is transcribed and translated into an inactive precursor that is sequestered in the endoplasmic reticulum (ER). However, when sterol levels are low, the sterol sensor SCAP is able to chaperone SREBP-2 to the Golgi apparatus where it is cleaved to release a transcriptionally active form that can enter the nucleus [4]. Conversely, in times of sterol abundance, SCAP is bound by sterols and remains sequestered in the ER along with the unprocessed SREBP-2 [4].
Diabetes mellitus is a multifactorial disease due to deficient insulin secretion and/or action, resulting in hyperglycemia, alterations in lipid metabolism, and a variety of complications in tissues throughout the body. These complications extend to the central nervous system (CNS), including cognitive dysfunction and behavioral changes, and are observed both in type 1 and type 2 diabetes [5]. Studies have shown altered information processing, psychomotor efficiency, attention, and mental flexibility in type 1 diabetes, whereas type 2 diabetes more often affects memory, psychomotor speed, and executive function [6]. In addition, changes have been observed on imaging of the hippocampi of both type 1 and type 2 diabetics [7],[8].
We recently demonstrated that in mouse models of diabetes there is a broad reduction in the expression of genes in the cholesterol biosynthetic pathway throughout the brain, resulting in decreased brain cholesterol synthesis [9]. This is accompanied by decreased expression of SREBP-2, which is more pronounced at the protein than the mRNA level [9]. Simultaneous reduction in cholesterol content and cholesterol biosynthesis suggests that, in the brain, diabetes causes a defect at the level of the sterol-sensing molecules, which constitute a negative-feedback system on SREBP-2 processing and maintenance of cellular cholesterol content. Here we show that the sterol sensing protein SCAP, which plays a key role in the post-transcriptional regulation of SREBP-2, is decreased in the brains of diabetic mice. Brain-specific heterozygous SCAP knockout mice have reduced levels of cholesterol genes and exhibit a 30% reduction in brain cholesterol synthesis, mimicking what is observed in diabetes. This results in defects at the electrophysiological and cognitive level. Thus, diabetes results in a reduction of SCAP expression, which contributes to a reduction of brain cholesterol synthesis. Reduction in SCAP leads to an impairment in neuronal transmission and cognitive dysfunction that may contribute to the neurological complications observed in diabetic states.
Recently we demonstrated that diabetes produces a global suppression of the enzymes of cholesterol biosynthesis and their master transcriptional regulator SREBP-2 in the brain; this results in reduced cholesterol biosynthesis and altered ability of neurons to form synapses and synaptic vesicles [9]. Since reduction of cholesterol and its precursors would normally drive a compensatory increase in SREBP-2 and cholesterol biosynthesis, we speculated that sterol-sensing might be impaired in the brains of diabetic mice. SCAP, a binding partner and a chaperone protein of SREBP-2, plays an essential role in sterol feedback regulation, serving as the major sterol-sensing molecule [4]. Western blotting of extracts of brains from streptozotocin (STZ)-treated mice (a model of type 1 diabetes) revealed an ∼50% decrease in SCAP protein in the mouse cerebral cortex (Figure 1A). This reduction of SCAP protein in the STZ diabetic mouse was not limited to the cortex and was observed throughout the brain, with a 33% decrease in the hypothalamus (Figure 1B) and a 31% decrease in the thalamus (unpublished data). There is also a significant, but smaller, reduction of SCAP protein in the cerebral cortices of db/db mice (a model of type 2 diabetes), consistent with a modest down-regulation of genes of the cholesterol synthesis pathway in the brains of these animals (Figure 1C). The decrease in Scap mRNA in STZ diabetic mice was completely rescued by three injections of insulin into the intracerebroventricular (ICV) space over the course of 30 h (Figure 1D). ICV insulin also resulted in a partial rescue of SCAP protein during this short treatment period (Figure 1E). This occurred with no change in blood glucose levels [9]. By contrast, there were no changes in Scap message after treatment of leptin deficient ob/ob mice with leptin (unpublished data).
To determine the consequences of reduced SCAP content in the brain in the absence of diabetes, we created mice with homozygous and heterozygous loss of SCAP specifically in the brain by crossing mice carrying a modified Scap gene with LoxP sites surrounding exon 1 [10] with mice expressing Cre-recombinase under control of the rat nestin promoter and enhancer [11]. This results in inactivation of the Scap gene in both neuronal and glial cells beginning on day e11.5 [11]. Homozygous Scap-ablated embryos (nestin-Cre+/−:Scaplox/lox, designated as N-Scap [−/−]) were obtained at normal Mendelian frequency; however, all of the homozygous N-Scap (−/−) mice died immediately after birth on postnatal day 0 (P0). Analysis of brains of these mice revealed an almost complete (>95%) absence of Scap mRNA and SCAP protein (Figure 2A and 2B). The N-Scap (−/−) neonates exhibited almost normal gross morphology, but had flat skulls and uninflated lungs (Figure 2C). Likewise, when mice were delivered by Cesarean section on embryonic day 18.5 (E18.5), all of the control (nestin-Cre−/−:Scaplox/wt or nestin-Cre−/−:Scaplox/lox) and N-Scap (+/−) (nestin-Cre+/−:Scaplox/wt) littermates initiated breathing, whereas N-Scap (−/−) failed to do so. The flat skull in N-Scap (−/−) was due to microcephaly; brains from N-Scap (−/−) mice exhibited an overall reduction in size of 43% by weight compared with those from the control littermates (50±10 mg versus 87.5±15 mg in controls, p = 0.003) (Figure 2D). There were no gross morphologic changes in the brain other than the decrease in size.
Gene expression analysis of the brains of N-Scap (−/−) mice revealed a pattern consistent with a loss of SCAP activity. Thus, SREBP-2 (Srebf2) and its downstream genes (Hmgcr, Sqle, Fdps, Idi1, and Ldlr) were all markedly down-regulated by 70%–90% (Figure 2E). Expression of the SREBP-1c (Srebf1c) gene was also reduced by 88%, whereas expression of its downstream genes (Fasn and Acaca) were only decreased by half (Figure 2E). The reduction in cholesterol synthetic enzymes was accompanied by a 50% reduction in total cholesterol content per gram of brain tissue (2.3±0.6 versus 4.7±1.2 µg of cholesterol/mg of brain in controls, p<0.01) (Figure 2F). In contrast, total brain triglyceride content was increased in N-Scap (−/−) mice compared to controls (19.9±0.4 versus 15.9±1.7 µmol of triglycerides/mg of brain in controls, p<0.05) (Figure 2G). Measurement of triglycerides in adult N-Scap (+/−) brains showed no difference when compared to controls (26.8±3.3 versus 27.4±1.3 µmol of triglycerides/mg of brain in controls) (Figure 2G).
Homozygous deletion of the Scap gene in the brain also caused marked changes in brain histology and cell type distribution. Expression of glial fibrillary acidic protein (Gfap), an astrocyte marker, was up-regulated by more than 3-fold in the whole brain of N-Scap (−/−) mice, whereas the microglia marker F4/80 (Emr1) was reduced by 59% and the neuronal marker MAP-2 (Matp2) was unchanged (Figure S1). The increase in Gfap mRNA correlated with a significant increase in GFAP protein in the N-Scap (−/−) brain (Figure 2H), which was confirmed by immunohistochemistry showing a robust increase of GFAP- positive fibers in N-Scap (−/−) brains resulting in a histological picture resembling gliosis (Figure 2I).
To better mimic the 50% reduction in SCAP in the brains of STZ-diabetic mice we characterized mice with haploinsufficiency of the Scap gene in the brain. qPCR analysis of Scap mRNA in extracts of brain regions dissected from N-Scap (+/−) mice revealed an approximately 40% reduction of Scap gene expression in all parts of the brain examined, including cerebral cortex and hypothalamus (Figure 3A). By Western blotting of extracts of the cortex, expression of SCAP protein was reduced slightly more (60%, when normalized with β-tubulin) in N-Scap (+/−) mice (Figure 3B).
Previous studies have shown that knockout of the Scap gene in the liver causes a reduction in the amounts of both the precursor and the nuclear forms of SREBP-2 and SREBP-1 [10]. We found a similar significant reduction of both SREBP-2 and SREBP-1 precursors in the cytoplasmic extract of cerebral cortices of N-Scap (+/−) mice (Figure 3C), as well as a reduction in the mature nuclear forms of these transcription factors (Figure 3D). This resulted in a decrease in the expression of genes downstream of SREBP-2 (Hmgcr, Sqle, Idi1, and Ldlr) by 10%–27% in the cerebral cortex and hypothalamus, as well as some decrease in Srebf2 mRNA (Figure 3E). The reduction of squalene epoxidase at the protein level was also confirmed by Western blotting (Figure 3C). The decrease in SREBP-1 did not result in a decrease in expression of the downstream protein fatty acid synthase but did cause a 20% reduction in acetyl-CoA carboxylase in the N-Scap (+/−) mice. There were large decreases in these proteins in the N-Scap (−/−) mice (Figure S2). The reduction in cholesterol synthesis was not accompanied by a similar decrease in cholesterol catabolism. Levels of Cyp46a1 mRNA were the same in the cerebral cortices and hypothalami of control and N-Scap (+/−) mice, and whole brains of control and N-Scap (−/−) mice (Figure S3).
To determine if the down-regulation of cholesterologenic genes could affect in vivo cholesterol synthesis, we assessed cholesterol synthesis in the brains of N-Scap (+/−) mice in vivo using tritiated water. This revealed a 29% reduction in cholesterol synthesis in the brains (p<0.05) of N-Scap (+/−) mice (Figure 3F), closely paralleling the change observed in STZ-diabetic mice [9]. Synapse marker proteins, including syntaxin-1A (STX1A) and post-synapse density 95 (PSD95) were also decreased in the N-Scap (+/−) mouse cerebral cortices (Figure 3G); similar to what has been observed in STZ-diabetic mice [12]. This occurred with no change in glucose levels (after a 4-h fast 182±6.7 versus 175±8 mg/dl in controls) or fed insulin levels (0.32±0.02 versus 0.33±0.04 ng/ml in controls). Thus, the heterozygous N-Scap (+/−) mice mimic the reduction of SCAP protein, the impaired cholesterol biosynthesis, and the decrease in synapse markers observed in the brains of diabetic mice, but without the concomitant changes in glucose or insulin levels.
To determine whether changes in cholesterol metabolism induced by the decrease of brain SCAP protein might alter synaptic transmission, we performed intracellular and extracellular electrophysiological recordings from neurons in the area CA1 region of the hippocampus, an area of the brain with demonstrated abnormalities on imaging in humans with type 1 and type 2 diabetes [7],[8]. Intracellular recording of miniature excitatory postsynaptic current (mEPSC) events were monitored by continuous voltage-clamp to assess differences in frequency and amplitude of presynaptic events at CA1 neurons. The basal frequency of mEPSCs of area CA1 neurons was reduced by more than 60% in N-Scap (+/−) mice (0.5±0.1 Hz; n = 15 cells; p<0.01) compared to control mice (1.4±0.3 Hz; n = 15 cells) (Figure 4A and 4B), but no differences in mEPSC amplitudes were observed between the two genotypes (control: 8.9±0.2 pA; N-Scap (+/−): 8.7±0.1 pA; n = 1500 events) (Figure 4C). These results strongly implicate an effect of SCAP reduction on presynaptic neurotransmitter release, with no differences at the postsynaptic membrane.
To further assess the effect of reduced SCAP levels on synaptic transmission, we tested the synaptic efficacy of the Schaeffer collateral pathway between control and N-Scap (+/−) mice via paired-pulse facilitation and long-term potentiation (LTP). Paired-pulse facilitation was elicited with interpulse intervals of 50, 100, 150, and 200 ms, and then calculated as the paired-pulse ratio of field excitatory postsynaptic potential (fEPSP) amplitude elicited by the second to first test stimuli. The paired-pulse ratio was greatest when elicited with an interpulse interval of 50 ms and was significantly reduced in N-Scap (+/−) hippocampal slices (control: 1.73±0.08; n = 17; N-Scap (+/−): 1.45±0.04; n = 12; p<0.01) (Figure 4D and 4E), suggesting significant alterations in synapse plasticity in N-Scap (+/−) hippocampi. Further studies will be required to define the mechanism of this defect at the synaptic level.
We also tested the efficacy of neurotransmitter release following high frequency stimulation to determine if a robust and inducible LTP can be generated at CA1 synapses of N-Scap (+/−) hippocampal slices. High frequency stimulation (4×100 Hz) reliably induced a robust and reproducible LTP in hippocampal slices from both control and N-Scap (+/−) mice. However, the change in fEPSP slope was markedly reduced in N-Scap (+/−) slices (Figure 4F and 4G). The mean fEPSP slopes measured 10, 30, 60, and 120 min after LTP induction in control slices were 255±1, 189±4, 196±1, 188±1, and 183±1, respectively (n = 10) (Figure 4H). All corresponding slope values were significantly lower in slices from N-Scap (+/−) mice (198±3, 147±1, 139±2, 127±2, and 125±2, respectively; n = 8; p<0.001) (Figure 4H). Taken together, these findings indicate a significant impairment of synaptic transmission at CA1 neurons of the hippocampus in the brains of SCAP heterozygous knockout mice.
Memory is impaired in patients with type 2 diabetes and to a lesser extent in patients with type 1 diabetes [13],[14]. To assess memory in our mouse model we used the novel object recognition test, which takes advantage of the natural bias of preference for novelty in rodents [15]. As expected, after a training session, control mice spent significantly more time exploring the novel object as compared to the training object (16.4±3 s versus 9.4±2 s, p<0.01), resulting in a 63% preference for the novel object (Figure 5A and 5B). By contrast, N-Scap (+/−) mice failed to demonstrate any significant preference for the novel object over the training object (13.8±2.5 s versus 12.9±2.1 s), spending almost identical amounts of time with each object (Figure 5A and 5B).
Diabetic patients have an increased prevalence of anxiety disorders [16],[17]. We used two different tests to assess anxiety in the N-Scap (+/−) mice. We performed the elevated open-platform test, which is used to assess psychological stress [18]. While controls spent 33.9±7.7 s in a frozen posture on the open platform, the N-Scap (+/−) mice exhibited a much shorter duration of freezing behavior (14.3±2.7 s, p<0.05) induced by this stress (Figure 5C). The novelty-suppressed feeding test assesses the effects of conflicting motivations of the drive to eat food after fasting and the fear of venturing into a novel arena of white paper on which the food pellets have been placed [19]. N-Scap (+/−) mice showed a markedly reduced latency to enter the arena compared to control mice, (17.1±6.4 versus 61.5±19 s, p<0.05) i.e., were more adventuresome or reckless in their behavior than controls (Figure 5D). Once they entered the novel area, however, there was no difference in the latency to start eating the food pellets (151.2±34.1 versus 150.7±18.8 s) (Figure 5D), demonstrating that hunger was not the driver of the reduced latency.
Interestingly, in unstressed conditions, circulating levels of the stress-related hormones corticosterone, epinephrine, and norepinephrine did not differ between control and N-Scap (+/−) mice. However, when the mice were isolated in individual transport containers for an hour, there was a robust elevation of stress hormone levels in N-Scap (+/−) mice compared with control mice with elevations in corticosterone (1.8-fold) and norepinephrine (2.2-fold) and a trend towards increased epinephrine (15-fold) (Figure 5E). The surprising increase in stress hormones in these mice despite an apparent decrease in anxiety behaviors may indicate an inappropriate reaction to stress rather than a true anxiolytic effect.
To further explore the possible relationship between behavior and the abnormal recordings found in the hippocampi of these mice we performed two behavioral tests which specifically use hippocampal function: contextual fear conditioning and the Morris water maze. In the contextual fear conditioning experiment N-Scap (+/−) mice showed impaired acquisition of the fear conditioning in response to a foot shock, as demonstrated by a decrease in freezing (Figure 5F). In addition, on the second day there was reduced freezing during the first minute in the context and the first and second minute in the altered context (Figure 5G), with later time points maintaining the same trend but losing statistical significance. This testing suggests impairments in hippocampal functioning, although we cannot rule out the possibility that there is some difference in pain thresholds between the mouse genotypes. We also performed a Morris water maze. We were unable to see significant differences between the two genotypes in this test though the N-Scap (+/−) mice showed a trend towards spending more time in the incorrect quadrants during the probe trials (Figure S4). Similar discordance between abnormal recordings in the hippocampus and abnormalities in the water maze have been previously observed in other models [20].
Finally, we evaluated feeding behavior and energy expenditure as measures of hypothalamic function in these mice. The N-Scap (+/−) mice consumed more food than the control mice, especially during the dark phase (Figure 6A and 6B). This phenotype is consistent with mice following SREBP-2 gene silencing in the hypothalamus shown in our previous study [9]. This correlates with mildly reduced expression of the anorexigenic neuropeptide CART (Cartpt), and increased expression of the orexigenic neuropeptide Agrp in the hypothalamus of the N-Scap (+/−) mice (Figure 6C). Energy expenditure, as represented by oxygen consumption in a metabolic cage, was about 20% higher in the N-Scap (+/−) mice than in control mice (Figure 6D). Consistent with this, locomotor activity of N-Scap (+/−) mice during this period indicated a significant, parallel increase (Figure 6E and 6F). The increase in energy expenditure compensated for the increase in food intake, resulting in similar body composition (Table S1). The respiratory exchange ratio (RER) was also high in the N-Scap (+/−) mice during the first 6 h of the dark cycle (Figure 6G), suggesting increased utilization of carbohydrate as a fuel.
In our previous study we showed that cholesterol synthesis is significantly impaired in multiple mouse models of diabetes due to a down-regulation of most of the genes in the cholesterol biosynthesis pathway [9]. In this study we show that this is due, at least in part, to a significant reduction of the sterol sensor protein SCAP in the brains of diabetic mice. Here we have used genetically modified mice with a defect in cholesterol synthesis to determine the consequences of decreased brain cholesterol, in the absence of the hyperglycemia and impaired insulin signaling found in diabetic mice. Homozygous deletion of Scap in the brain causes microcephaly, gliosis, and early postnatal lethality. Mice with a brain-specific Scap heterozygous knockout, on the other hand, closely mimic the decrease in sterol synthesis observed in the diabetic brain, and this leads to significant attenuation of synaptic transmission and cognitive dysfunction, even in the absence of any systemic metabolic derangement. Because Scap is knocked out during the prenatal period we cannot rule out a developmental contribution to the changes observed.
Cholesterol comprises a significant component of the neuronal membrane and is thus critical for proper neuronal transmission. Reduction in cholesterol synthesis decreases the formation of synaptic contacts between neurons [9]. We show here that the reduction of cholesterol synthesis in the brains of N-Scap (+/−) mice produces adverse effects on neuronal transmission by decreasing the efficacy of synaptic transmission. Diabetic rodents exhibit impaired performance in the novel object recognition test and contextual fear conditioning test—functional measures of memory in rodents [21],[22]—and this is mimicked in the N-Scap (+/−) mouse. Impairment of LTP expression in the CA1 region of the hippocampus in diabetic animals has also been reported [23]–[25] and is also seen in the N-Scap (+/−) mouse. The impairment in LTP provides an electrophysiologic correlate for the impaired memory observed in the novel object recognition test and the contextual fear conditioning test. Coupled with our previous studies showing altered brain cholesterol metabolism in diabetes [9], the current results provide the first evidence indicating that failure in regulation of cholesterol homeostasis in the brain may contribute to the cognitive impairment seen in diabetes (see model in Figure 7). While the N-Scap heterozygous mouse model system does not allow for the precise dissection of the relative contributions of SREBP-2 processing and its effects on cholesterol synthesis versus SREBP-1 processing and its potential effects on fatty acid and triglyceride synthesis, it seems likely that SREBP-2 and cholesterol are more important. Although the knockout of SCAP produces changes to SREBP-1 processing, we find no defect in free fatty acid synthesis in the brains of N-Scap (+/−) mice (unpublished data), nor is there a decrease in triglyceride content in the brain. This is in contrast to a conditional knockout of SCAP in the liver, which causes decreases in serum free fatty acids and triglycerides of 46% and 53%, respectively, but only a 24% decrease in serum cholesterol [10]. This may be explained by the fact that while cholesterol is unable to cross the blood-brain barrier, fatty acids and glycerol, the building blocks of triglycerides, are able to pass. While fatty acid and glycerol transport across the blood-brain barrier may not fully compensate for all of the defects created by the decrease of SREBP-1 in the brain, it seems to significantly mitigate the effects given the normal to increased triglyceride levels observed in this manuscript. Nonetheless, the SREBP-1 pathway may have a greater impact in diabetes as there is a 15% decrease in fatty acid synthesis in brains of STZ diabetic mice (unpublished data). Further, there is also the possibility that components of the SREBP-2 pathway beyond what are explored here, such as isoprenoid production, may be contributing to the observed phenotype.
Exactly how lowering cellular cholesterol in the brain might affect brain function is unclear and likely multifactorial. Cholesterol depletion has been shown to retard or prevent clathrin-mediated endocytosis, including internalization of acetylcholine receptors [26] and block vesicle biogenesis in neurosecretory cells, consistent with a role for cholesterol in regulating membrane fluidity and the changes in curvature necessary for full vesicle formation [2],[27]. Moreover, cholesterol binds the synaptic vesicle protein synaptophysin and modulates interaction with the essential vesicular SNARE protein, synaptobrevin, to regulate exocytosis [28]. Knockdown of SREBP-2 in primary neuronal cultures produces decreases in expression of the synaptic vesicle marker VAMP2 [9]. Interestingly, deletion of low-density lipoprotein receptor-related protein 1 (LRP1) in forebrain neurons in mice leads to a decrease in brain cholesterol levels and memory loss with selective reduction of glutamate receptor subunits, which is partially rescued by restoring neuronal cholesterol [29]. These reports are consistent with the phenotypes of attenuated synapse transmission and cognitive impairment with decreased synaptosomal markers in N-Scap (+/−) mice.
Activation of glial cells, as demonstrated by increased GFAP staining, is a common feature of many types of neural insults including trauma, toxins, neurodegenerative triggers, or infection [30]. Multiple studies have reported increased GFAP expression or astrocyte content in cerebral cortices and hippocampi of diabetic animal models [21],[31],[32]; however, the mechanisms producing these changes have not been elucidated. The brains of N-Scap (−/−) mice show increased GFAP expression and gliosis. Whether this is a compensatory or reactive response induced by cholesterol deficiency in the brain remains to be determined.
Several neurodegenerative diseases have been associated with alterations in cholesterol metabolism. Reduction of cholesterol synthesis in astrocytes is thought to contribute to neurodegeneration in multiple models of Huntington disease [33]. Alzheimer disease has been associated with both diabetes and cholesterol metabolism [34]–[37]. The ε4 allele of ApoE, a cholesterol transport protein, is the only known risk factor, other than aging, for late onset Alzheimer disease [35]. On the other hand, studies examining the relationship of serum cholesterol levels in the elderly to dementia have differing conclusions, and several clinical trials in humans using the statin class of lipid lowering drugs have yielded conflicting results (reviewed in [36]). Part of this may reflect the broad range of human physiology. For example, in our previous study we showed a 3-fold range of expression of Srebf2 and Hmgcr and a 2-fold range in synaptosomal cholesterol content in brains of 16 elderly humans with and without diabetes and dementia [9]. Based on our studies and others [36], one might predict that too much cholesterol, as well as too little, could have detrimental effects on neuronal function. However, because cholesterol in the brain is controlled independently of serum cholesterol, we do not currently have a clinically useful tool for assessing cholesterol levels in the human brain. Interestingly, a recent study used intranasal delivery of insulin to the brain as a therapy for human dementia patients with some improvement seen in memory and self care tasks [38]. The source of this benefit is unknown, but one of the responses to this therapy may be an increase in cholesterol synthesis in the brain.
Taken together with our previous work [9] we propose a model for how diabetes may affect cholesterol in the brain, leading to changes in behavior and brain function (Figure 7). In this model, diabetes mellitus reduces expression of SCAP in the brain, and this reduction of SCAP causes a defect in SREBP-2 processing, leading to a reduction in active SREBP-2. This leads to a down-regulation of genes involved in brain cholesterol synthesis, which leads to impaired synaptic transmission and abnormal cognitive function. These findings provide a novel view on the role of cholesterol regulation in the brain in diabetes, and open the possibility of therapeutic strategies for reversing the effects of diabetes on the brain and nervous system.
All animal studies followed the US National Institutes of Health guidelines and were approved by the Institutional Animal Care and Use Committees at the Joslin Diabetes Center and Beth Israel Deaconess Medical Center.
Scap-floxed mice [10], nestin-Cre mice [11], C57BL/6 mice, and db/db mice (C57Bl/Ks background) were purchased from the Jackson Laboratory. Scap-floxed mice with the nestin-Cre transgene were maintained on a C57BL/6×129Sv mixed genetic background; therefore for all studies littermates were used for analysis. For STZ-induced diabetes experiments, 7-wk-old C57BL/6 mice were treated with a single intraperitoneal injection (200 µg/g body weight) of STZ (Sigma). Some of the STZ diabetic mice were treated ICV with insulin. For these experiments 7-wk-old C57Bl/6 mice were placed in a stereotactic device under anesthesia, and a 26-gauge guide cannula (Plastics One Inc) was inserted into the right lateral cerebral ventricle (1.0 mm posterior, 1.0 mm lateral, and 2.0 mm ventral to the bregma). A dummy stylet was inserted into each cannula until used. After 1 wk of recovery, the mice received a single IP injection of STZ to induce diabetes. Twelve days later, the mice received three ICV injections of insulin (3 mU in 2 µl) or the same volume of PBS (9am, 7pm, and 9am the following day) through an internal cannula using a Hamilton microsyringe. Four hours after the last injection the hypothalami were collected. All mice were maintained on a 12-h light/12-h dark cycle and fed a standard mouse chow diet (LabDiet Mouse Diet 9 F, PMI Nutrition International). All mice used for experiments were male. For analysis of gene and protein expression, the brain was quickly removed under anesthesia with 2.5% Avertin (15 µl/g body weight, IP), placed on ice, and dissected into the hypothalamus and cerebral cortex using a mouse brain matrix (ASI Instruments Inc.).
RNA from murine brain tissue was isolated using an RNeasy kit (Qiagen). As a template, 1 µg of total RNA was reverse-transcribed in 50 µl using a High Capacity cDNA Reverse Transcription kit (Applied Biosystems) according to the manufacturer's instructions. Three microliters of diluted (1∶4) reverse transcription reaction was amplified with specific primers (300 nM each) in a 25 µl PCR reaction with a SYBR Green PCR Master Mix (Applied Biosystems). Analysis of gene expression was done in the ABI PRISM 7000 sequence detector with initial denaturation at 95°C for 10 min followed by 40 PCR cycles, each cycle consisting of 95°C for 15 s and 60°C for 1 min, and SYBR green fluorescence emissions were monitored after each cycle. For each gene, mRNA expression was calculated relative to Tbp expression as an internal control.
Brains were immersed in 4% paraformaldehyde (PFA) overnight at 4°C then embedded in paraffin, and 8 µm coronal sections were collected. Sections were deparaffinized and blocked with a Mouse Ig Blocking Reagent (M.O.M. Immunodetection kit, Vector Laboratories) containing avidin for 1 h, washed with PBS, treated with M.O.M diluent for 5 min, and incubated with a mouse monoclonal antibody recognizing GFAP (1∶500, Millipore) and biotin solution for 30 min at room temperature. After washing with PBS, the sections were incubated with a biotinylated anti-mouse IgG secondary antibody for 20 min. The samples were washed with PBS, treated with a streptavidin/peroxidase complex reagent, washed with PBS again, and stained with a VIP Substrate kit (Vector Laboratories). Hematoxylin was used for counter-staining.
Nuclear and cytoplasmic extracts of brain tissue were prepared per the manufacturer's directions (NE-PER kit, Pierce). Whole tissue extracts were prepared using RIPA buffer containing 1% SDS and protease inhibitor cocktail (Sigma). Protein concentrations were measured using a BCA assay (Pierce). Immunoblotting was performed with antibodies against SCAP (Santa Cruz), β-tubulin, CREB, fatty acid synthase, acetyl-CoA carboxylase (Cell Signaling Technology), SQLE (ProteinTech), SREBP-1/SREBP-2 (gifts from Jay Horton and Guosheng Liang), STX1A and PSD95 (Abcam), and SYP and MBP (Millipore).
Brain tissue was homogenized in 50 mM NaCl. Lipid fraction was then extracted through multiple washes with a 2∶1 chloroform∶methanol solution. Samples were dried down with 10% triton-X 100/acetone. Cholesterol content was assayed by enzymatic assay (Wako Chemicals). Triglycerides were measured by colorimetric assay (Abnova).
Each anesthetized animal was injected intraperitoneally with 50 mCi of [3H]water in 0.2 ml of PBS. One hour after injection (which is a time long enough to reflect endogenous rates of synthesis and short enough to avoid significant inter-organ redistribution [39]), blood was collected by retro-orbital puncture, and the [3H]water specific activity in the plasma was measured. The brain was removed, and the whole cerebrum (250–290 mg) was saponified with 2.5 ml of 2.5 M KOH (75°C, 2 h). Sterol-containing lipids were extracted with 10 ml hexane and 5 ml 80% ethanol. Cholesterol was isolated by thin layer chromatography (hexane∶diethyl ether∶glacial acetic acid = 80∶20∶1), and the incorporated tracer was measured by a scintillation counter. The synthesis rates were calculated as nmol of [3H]water incorporated into cholesterol per gram of tissue per hour.
Transverse hippocampal slices (400 µm) were prepared from the brains of 4–6-wk-old male control or N-Scap (+/−) littermates that were submerged in an ice-cold high sucrose solution containing (in mM): 250 sucrose, 2.5 KCl, 1.24 NaH2PO4, 10 MgCl2, 10 glucose, 26 NaHCO3, 0.5 CaCl2 that was aerated with 95% O2/5% CO2. Slices were then maintained at 30°C in artificial cerebrospinal fluid (ACSF) containing (in mM): 124 NaCl, 2.5 KCl, 1.24 NaH2PO4, 1.3 MgCl2, 10 glucose, 26 NaHCO3, 2.5 CaCl2, and allowed to recover for at least 1 h before being transferred to the recording chamber for the start of experiments.
Voltage-clamp whole-cell recordings of mEPSC events at area CA1 neurons were recorded at a holding potential of −65 mV in the presence of 500 nM tetrodotoxin (TTX) using a glass microelectrode (9–10 MΩ) filled with an internal pipette solution containing (in mM): 137 K-gluconate, 2 KCl, 5 HEPES, 5 MgATP, 0.3 NaGTP, 10 creatine (290 mOsm/l [pH 7.4]).
fEPSPs were elicited by stimulating the Schaeffer collaterals with a concentric bipolar electrode (FHC) and recorded with an ACSF-filled glass microelectrode (1–2 MΩ) positioned in the stratum radiatum of area CA1. Baseline test stimuli were applied once per min at a test stimulation intensity (0.1 ms pulse width) adjusted to evoke fEPSP amplitudes that were 40% of maximal size. Paired-pulse stimulation was elicited by two consecutive test stimuli; the interpulse interval was varied from 50–200 ms, at 50 ms increments. LTP was induced by four trains of 100 Hz (1 s train duration) elicited 5 min apart. fEPSPs were not monitored between each high frequency train but were monitored with the test stimuli for up to 120 min after the induction of LTP.
All recordings were acquired with a Multiclamp 700B amplifier (Molecular Devices) via a Digidata 1440A (Molecular Devices) digitizer interface then recorded with pCLAMP 10.2 software. Offline analysis of mEPSCs and fEPSPs were performed with MiniAnalysis (Synaptosoft) and Clampfit 10.2 (Molecular Devices), respectively. Statistics and graphs were produced using Prism 5 (GraphPad Software). Statistical significance for mean comparisons was determined by the unpaired Student's t test at p<0.05. Data were represented as mean ± SEM, where appropriate, and n refers to the number of hippocampal slices, unless indicated otherwise.
The novel object recognition test was performed as described previously [40] with slight modifications. Mice were maintained in group housing over the course of the experiment. Briefly, the mice were habituated for 1 h to a plastic cage box on the first day. The floor was covered with 1 cm of wood bedding. On the second day, a familiarization trial was performed for 5 min, allowing each mouse to explore the two identical objects (objects A and B) in the same box. The two objects were placed along the long axis of the box. Mice were filmed while exploring objects A and B and this was quantified to ensure that the mice did not show preference for the object on one side of the cage over the other. There was not a significant difference between exploration of object A and object B. Then, the mouse was removed from the trial box and placed in its home cage for 1 h. After each exposure, the objects were wiped with 70% ethanol to eliminate odor clues. One hour after the familiarization, each mouse was placed in the box with one of the old objects (object A) and a new one (object C). The position of object C was the same as object B in the familiarization trial, and the time to explore them was again 5 min. A mouse was considered to be engaging in exploratory behavior if the animal touched the object with its forepaw or nose or sniffed at the object within a distance of 1 cm. Activity was monitored and calculated from a timed video of the experimental field. In a separate cohort of mice both the old object (object A) and the new object (object C) were presented to mice for 30 s to determine if one object was more interesting to the mice than the other. There was no difference in exploratory behavior between the two objects when presented simultaneously.
The novelty-suppressed feeding test was performed as described previously [19] with slight modifications. Twenty-four hours before the test, food was removed from the cages. At the time of testing, a pair of food pellets (regular chow) were placed on a white paper disk positioned in the center of a trial box without wood bedding. An animal was placed in the corner. The latency to enter the arena and then to chew the pellets were recorded within a 5 min period. Activity was monitored and calculated from a timed video of the experimental field.
The elevated open platform test was performed as described previously [18] with slight modifications. In brief, a transparent glass cylinder (12 cm diameter, 21 cm high) was placed upside-down and each mouse was positioned at the top (open platform). Freezing behavior was defined as no movement, excluding respiratory movement, while in a crouching posture. The duration of freezing was the total amount of time that the animal showed freezing. If the mouse slipped off the platform, it was immediately placed back on the platform and the experiment was continued. The mouse behavior on the elevated open platform was video recorded for 5 min.
During the acquisition phase mice were randomly dropped at one of four points; N, S, E, or W. A hidden platform was located in the southwest (SW) quadrant. Mice were given 1 min to find the hidden platform and latency was recorded. Mice were led to the hidden platform if they did not reach it within 1 min. Each mouse went through eight trials per day. Latency times leveled off on day 4. On day 5 the probe trial was conducted. The hidden platform was removed and the mouse was dropped off at the N drop point. Time spent in each quadrant was recorded over 1 min. Data was analyzed using TopScan software from Cleversys, Inc.
On day one each mouse was placed in a novel box for 2 min and freezing activity was recorded (baseline). At the end of the 2 min the mouse received a 0.5 mA shock for 2 s. Freezing time was again measured over 2 min (post shock 1). At the end of the 2 min a second 0.5 mA shock was delivered over 2 s and freezing was recorded for 1 min. The mouse was then returned to its home cage. On day 2 each mouse was placed back in the cage from day 1 and freezing was recorded for 3 min (context). The mice were then placed in an unfamiliar box and freezing was again measured for 3 min (altered context). Data was analyzed using TopScan software from Cleversys, Inc.
Data are expressed as the mean ± SEM. Statistical significance was calculated using an unpaired Student's t-test for comparison between two groups, and by an analysis of variance (ANOVA) for multigroup comparison. Fisher's PLSD was used for the Morris water maze and contextual fear conditioning.
|
10.1371/journal.pbio.1001692 | Fitness Trade-offs Restrict the Evolution of Resistance to Amphotericin B | The evolution of drug resistance in microbial pathogens provides a paradigm for investigating evolutionary dynamics with important consequences for human health. Candida albicans, the leading fungal pathogen of humans, rapidly evolves resistance to two major antifungal classes, the triazoles and echinocandins. In contrast, resistance to the third major antifungal used in the clinic, amphotericin B (AmB), remains extremely rare despite 50 years of use as monotherapy. We sought to understand this long-standing evolutionary puzzle. We used whole genome sequencing of rare AmB-resistant clinical isolates as well as laboratory-evolved strains to identify and investigate mutations that confer AmB resistance in vitro. Resistance to AmB came at a great cost. Mutations that conferred resistance simultaneously created diverse stresses that required high levels of the molecular chaperone Hsp90 for survival, even in the absence of AmB. This requirement stemmed from severe internal stresses caused by the mutations, which drastically diminished tolerance to external stresses from the host. AmB-resistant mutants were hypersensitive to oxidative stress, febrile temperatures, and killing by neutrophils and also had defects in filamentation and tissue invasion. These strains were avirulent in a mouse infection model. Thus, the costs of evolving resistance to AmB limit the emergence of this phenotype in the clinic. Our work provides a vivid example of the ways in which conflicting selective pressures shape evolutionary trajectories and illustrates another mechanism by which the Hsp90 buffer potentiates the emergence of new phenotypes. Developing antibiotics that deliberately create such evolutionary constraints might offer a strategy for limiting the rapid emergence of drug resistance.
| The evolution of drug resistance in human pathogens is considered an inevitable consequence of the selective pressures imposed by antimicrobial drugs. Yet resistance to one antifungal drug, amphotericin B (AmB), remains extremely rare despite decades of widespread use. Here we explore the biological mechanisms underlying this conundrum. By examining natural and experimental populations of Candida albicans, we identify multiple mutations that confer resistance to AmB in vitro. As with the evolution of resistance to other antifungals, we find that the chaperone protein Hsp90 is involved in enabling the evolution of resistance to AmB. We also discover, however, that mutations that confer AmB resistance impose massive costs on other aspects of fungal pathogenicity; strains that are resistant to AmB are hypersensitive to attack by the host immune system and are unable to invade and damage host tissue. Thus, the evolution of resistance to AmB is restricted by a tradeoff between tolerance of the drug and the ability to cause disease. We propose that developing new antibiotics for which resistance presents such dire tradeoffs may be a promising strategy to prevent the evolution of resistance.
| Understanding how organisms rapidly evolve novel traits is a central problem in both evolutionary biology and the treatment of infectious diseases. The emergence of drug-resistant pathogens not only provides a model for studying the evolution of new phenotypes but also poses a grave challenge to human health. Antibiotic treatment selects for rare mutations that alter cellular processes and, thereby, either mitigate the toxicity of the drug or bypass it altogether. Sometimes resistance mechanisms are completely orthogonal to the normal biology of the cell, as is the case for the amplification of efflux pumps or the horizontal acquisition of drug-detoxifying enzymes in bacteria. But often the mutations that confer resistance alter basic cellular processes in such a way as to create a variety of new stresses. The latter is especially relevant in eukaryotic pathogens, where the rarity of genetic exchange within and between populations necessitates the de novo evolution of resistance [1],[2].
Candida albicans is the leading fungal pathogen of humans and the fourth most common cause of all hospital-acquired infections [3],[4]. Normally a harmless commensal, changes in host immune status allow C. albicans to become pathogenic. Infections range from superficial thrush to life-threatening systemic disease. Mild to moderate infections are currently treated with triazoles, which inhibit Erg11 (lanosterol 14α-demethylase), preventing ergosterol biosynthesis [5]. Life-threatening systemic infections require treatment with echinocandin or polyene agents. Echinocandins inhibit the synthesis of the cell wall polymer (1,3)-β-D-glucan, resulting in loss of cell integrity [5].
The third most commonly employed antifungal is the polyene drug amphotericin B (AmB), which was the standard of care for ∼40 years [6]. Its potent fungicidal activity derives from its ability to selectively bind the major sterol of fungal membranes, ergosterol [7],[8]. Among other effects, this binding induces pore formation in the plasma membrane and results in rapid cell death. While AmB is extremely effective at killing fungi, its clinical utility is impaired by several liabilities. First, pharmacokinetics and distribution are poor, allowing some fungi to hide in niches where drug exposure is limited [9]. Second, AmB induces idiosyncratic systemic reactions involving fever and tremors. Third, and still more problematic, AmB's cumulative, dose-dependent renal toxicity limits use in many patients.
Despite these limitations, a remarkable advantage of AmB is that it has been exceptionally refractory to the evolution of resistance. After 50 years of use as monotherapy, the acquisition of AmB resistance in C. albicans remains extremely rare. For comparison, the antifungal drug 5-flucytosine was introduced several years later than AmB, but resistance rendered this drug obsolete against Candida in less than 20 years [10]. In a recent study of 9,252 clinical isolates of C. albicans, 99.8% remained AmB-sensitive [11]. Although the less toxic triazoles and echinocandins have recently replaced AmB as the frontline therapy, AmB still retains frequent use in many settings, particularly when an infection resists treatment with other drugs [12]. Indeed, resistance to triazoles emerges frequently and, although echinocandins are relatively new to the clinic, resistance to echinocandins is also already arising [13]–[18].
The best-validated mechanism of resistance to AmB observed in clinical isolates of C. albicans to date involves a double loss of function in both ERG3 and ERG11 (C-5 sterol desaturase and lanosterol 14α-demethylase, respectively), identified by biochemical analysis of membrane sterol composition [19]–[21]. In other fungal pathogens, sterol analysis of rare AmB-resistant isolates has identified resistant strains lacking ERG2, encoding C-8 sterol isomerase, and ERG6, encoding C-24 sterol methyltransferase [22]–[24]. However, there has been no systematic analysis of AmB resistance mutations in Candida using matched isogenic strains. More importantly, the consequences of these mutations upon the biology and pathogenicity of Candida remain largely unexplored.
Here we thoroughly explore mutations that can confer AmB resistance in C. albicans with the goal of understanding why resistance emerges so rarely in the clinic. Our results establish that the evolutionary constraints imposed by AmB are distinct from those of other antifungals. They provide insights into the mechanisms by which external and internal biological stresses restrict evolutionary trajectories.
In addition, our work broadens the role of protein homeostasis regulators as potentiators for the emergence of new traits. Finally, our findings suggest a general strategy for antimicrobial drug development that might be broadly useful in limiting the emergence of resistance.
As a first step towards understanding the evolution of resistance to AmB in Candida, we sought to broaden and validate the list of mutations that allow the fungus to tolerate this drug. As AmB-sensitive parental strains from which rare AmB-resistant isolates evolved are not available, the identification of mutations conferring resistance has proven challenging. Nevertheless, we sequenced the entire genome of two independent clinical isolates that had evolved resistance to AmB, one from C. albicans and one from the closely related species C. tropicalis. For comparison, we also resequenced the AmB-sensitive C. albicans reference strain SC5314.
Using paired-end reads, we achieved over 50-fold coverage of these genomes, which allowed us to detect simple polymorphisms as well as complex genome rearrangements. As expected, the strains differed from each other and from the reference strain at more than 20,000 sites. To identify the variants responsible for resistance, we took advantage of previous work and inspected candidate genes acting in the ergosterol biosynthesis pathway.
In the C. albicans AmB-resistant isolate, we detected a high density of mispaired reads at the ERG2 (ORF19.6026) locus (Figure 1A). Further analysis revealed that both copies of the ERG2 gene in this strain carried an insertion of the TCA2 retrotransposon (Figure S1A). Whole-genome analysis of polymorphisms indicated that the strain carried a high level of heterozygosity across its entire genome, with only two small regions of homozygosity. Strikingly, one of these included the transposon insertion in ERG2 (Figure S1B).
In the C. tropicalis isolate, the sequence of ERG2 was identical to that of the AmB-sensitive reference strain, MYA-3404. However, a mutation was observed in ERG3 (CTRG_04480), another enzyme involved in sterol synthesis (Figure S2A). Specifically, phenylalanine replaced serine 258, a residue that is absolutely conserved in this protein from fungi to mammals (Figure S2B). In addition, the ERG11 (CTRG_05283) ORF of this isolate harbored a deletion of 170 nucleotides (Figure S2C). Again, despite generally high levels of heterozygosity in other regions of the genome, the regions surrounding both ERG3 and ERG11 had become homozygous (Figure S2D). These results suggest that selective sweeps had operated to fix new mutations to a homozygous state in both clinical isolates. But, of course, the sequencing of many more AmB-sensitive and resistant isolates would be necessary to establish this conclusion.
To validate that either loss of ERG2 or the combined loss of ERG3 and ERG11 function is sufficient to confer resistance to AmB, we created them anew in a wild-type background. Because C. albicans is an obligate diploid, we used auxotrophic markers to sequentially delete the loci, creating homozygous mutations in ERG2 and double homozygous mutations in ERG3 and ERG11.
To confirm the inactivation of these genes in both lab strains and clinical isolates, we exploited the unique spectral characteristics of ergosterol. These result from conjugation of the double bonds C5–C6 and C7–C8, formed and isomerized by Erg3 and Erg2, respectively. We prepared sterol extracts from both of the clinical isolates and both of the laboratory mutants along with the isogenic wild-type control. Extracts from the wild-type laboratory strain exhibited the double-peaked spectrum between 240 and 300 nm characteristic of the conjugated bonds (Figure 1C). Extracts from both the laboratory mutants and the clinical isolates did not.
Next, we determined the minimal inhibitory concentration (MIC) of AmB in the knockout strains and compared it to the wild-type laboratory strain and the clinical isolates. AmB resistance levels in the two newly created laboratory strains matched those of the corresponding clinical isolates. For the erg2 mutants this was a 10-fold increase in MIC (Figure 1D), and for the erg3 erg11 mutants it was a 20-fold increase. We were unable to obtain further transformants in these mutants for technical reasons, thus precluding any attempt at complementation of the phenotype. However, we verified our results with an additional independent mutant in each background (Figure S5). Thus, laboratory-generated mutants successfully reproduce the resistant phenotypes observed in clinical isolates.
To discover other mutations that could confer AmB resistance in Candida, we next employed in vitro evolution. The drug-sensitive reference strain SC5314 was not suited to this analysis because we repeatedly found that the resistance that emerged in selections with this strain was highly unstable. Another strain, C. albicans ATCC-10231, proved to be less susceptible to this problem. To isolate resistant variants, this strain was inoculated in liquid media containing a low concentration of AmB and serially passaged seven times into media with a 2-fold higher concentration of AmB at each step. Surviving cells from each passage were isolated and saved. We then used whole genome sequencing and alignment of the parental strain with strains from the third, fifth, and seventh passages to identify the mutations emerging as strains developed resistance.
Alignment of genome sequences and algorithmic detection of novel polymorphisms emerging within the series revealed a trajectory of mutations in the ERG6 (ORF19.1631) gene, encoding Δ(24)-sterol C-methyltransferase (Figure 1B). Notably, the parental strain had been heterozygous for a premature stop codon that replaced the codon for Glu70. A mutation in the other allele of ERG6, Asp180Gly, appeared at the second step in the series. In the third step, the copy number of the Q70Stop allele increased to a 3∶1 ratio relative to the Asp180Gly allele. Finally, a loss of heterozygosity event, unique to the left arm chromosome 3 where ERG6 resides, resulted in homozygosity of the nonsense allele (Figure S3A–B).
Several independent selections with this strain—with either a similar gradual selection process or with selection regimes employing immediate shifts to high drug concentrations—all involved additional mutations in ERG6 (unpublished data). Finally, we validated the capacity of ERG6 mutations to create AmB resistance by using auxotrophic markers to delete the gene in the SC5314 reference strain background. This resulted in a 12-fold increase in the AmB MIC (Figure 1D).
To systematically define genes whose inactivation might confer AmB resistance in vitro, we used homologous site-directed recombination to generate isogenic diploid deletion mutants for all seven nonessential genes acting in the latter half of the ergosterol biosynthesis pathway (the steps after cyclization of squalene to lanosterol). Only the deletion of ERG2, ERG6, or of ERG3 and ERG11 together conferred more than a 3-fold increase in the AmB MIC (Figure 1D). While other mutations conferring resistance to AmB may exist, these three are the most critical, as they have all been detected in the clinic and validated in the laboratory. Thus, we focused our efforts to explain the exceptional rarity of clinical AmB resistance on understanding the broader biological consequences of mutations in ERG2, ERG6, or ERG3 and ERG11.
We previously reported that the emergence and maintenance of resistance to triazole and echinocandin antifungals critically depends on the molecular chaperone Hsp90 (ORF19.6515) [25]. Hsp90 is one of the most abundant proteins in eukaryotic cells, constituting ∼1% of total cellular protein, and acts as a protein homeostasis buffer. Although Hsp90 is an essential protein, its activity can be reduced up to 10-fold without impairing normal growth [25]–[27]. Previous work has suggested that phylogenetically diverse organisms use this excess reservoir of protein-folding capacity to promote the rapid evolution of new traits through a litany of mechanisms [25],[27],[28]. These include binding and stabilizing mutant proteins with novel activities, promoting the folding and maturation of metastable signal transduction proteins that respond to harsh environmental conditions, and allowing for the release of cryptic genetic variation upon stress. In pathogenic fungi, we have shown that mild compromise of Hsp90 function prevents the de novo emergence of resistance to triazoles and, in fact, reverses the resistance of strains of which had previously evolved triazole and echinocandin resistance in the clinic [25],[29],[30]. Hsp90 promotes antifungal drug resistance by stabilizing calcineurin and protein kinase C, two signal transducers that promote resistance by mitigating the stress to the cell wall and membrane that is induced by these drugs [25],[30],[31].
We asked if Hsp90 also plays a role in the evolution of resistance to AmB, taking advantage of two structurally unrelated natural products with high specificity for Hsp90 (geldanamycin and radicicol). We examined mutants that had evolved AmB resistance under drug pressure in the clinic (Figure 2A–B), mutants created deliberately by targeting genes in the sterol pathway (Figure 2C), and mutants arising at each step in our in vitro selection (Figure 2D). We spotted drug-resistant and control isolates on media containing no drug, fluconazole, or AmB, in either the presence or in the absence of the Hsp90 inhibitors.
As previously described, modest inhibition of Hsp90 completely blocked fluconazole resistance (Figure 2A, middle panel). It did not affect growth of the fluconazole-resistant isolate in the absence of fluconazole (Figure 2A, top panel). Modest inhibition of Hsp90 also abrogated AmB resistance (Figure 2A, bottom panel). Surprisingly, however, low concentrations of either of the Hsp90 inhibitors completely blocked the growth of all of the AmB-resistant strains, even in the absence of AmB (Figures 2A–D and S4B).
Perturbations in ergosterol biosynthesis can lead to general increases in the accumulation of diverse small molecules. This raised the possibility that the hypersensitivity to Hsp90 inhibitors might simply be due to an increase in their intracellular accumulation. To investigate, we determined the MICs of a panel of seven chemically and mechanistically distinct cytotoxic agents (that do not act through Hsp90) in all of the resistant strains. These MICs were compared to the MICs of geldanamycin and radicicol, as well as two synthetic Hsp90 inhibitors from completely different chemical scaffolds. AmB-resistant mutants were, indeed, generally more sensitive than wild-type cells to many of the cytotoxic compounds. The decrease from the wild-type in the MIC of any of these cytotoxic agents ranged from 2- to 8-fold in the erg2 or erg6 mutants and 4- to 16-fold in the erg3 erg11 mutant (Figure S4A–C). But the decreases in MIC of the four Hsp90 inhibitors were 18- to 48-fold for the erg2 mutants, 85- to 109-fold for the erg6 mutants, and 222- to 480-fold for the erg3 erg11 mutants (Figure S4B–C). Thus, the hypersensitivity of AmB-resistant strains to Hsp90 inhibitors cannot simply be attributed to a general increase in drug accumulation. Rather, the growth of these mutants must critically depend on maintaining very high levels of Hsp90 function even in the absence of AmB.
Was the effect of the Hsp90 inhibitors restricted to growth inhibition, or did they actually cause cell death? We previously showed that Hsp90 inhibition renders the typically cytostatic drug fluconazole cytocidal, killing the fungus instead of simply blocking its growth [29]. Indeed, low concentrations of Hsp90 inhibitors were cytocidal to the AmB-resistant mutants (Figure 2E). But, once again, in contrast to the fluconazole resistant strains, Hsp90 inhibition killed AmB-resistant cells even in the absence of AmB. Thus, the mutations that confer AmB resistance cause a novel and critical dependence on Hsp90 for the simple maintenance of normal viability.
Hsp90 promotes the maturation of a diverse array of metastable signal transduction proteins, including kinases, phosphatases, and ubiquitin ligases (known as Hsp90 clients) [32]. These function in many stress response pathways. Thus, the simplest explanation for the extreme dependence of AmB-resistant strains on Hsp90 is that these normally nonessential client proteins are required to tolerate the perturbations in cellular homeostasis caused by mutations in ergosterol biosynthetic enzymes.
To investigate, we first tested our AmB-resistant strains for constitutive transcriptional activation of a variety of stress response genes. These include targets of the known HSP90 client calcineurin [30],[33], as well as genes involved in the response to iron starvation or oxidative stress, two stresses tightly linked to membrane sterol homeostasis. To provide a point of comparison, we exposed wild-type strains to external stresses known to induce these responses.
The AmB-resistant mutants indeed exhibited a constitutive activation of diverse stress responses (Figure 3A, left panel). Pathways of iron starvation were constitutively active, most strongly in the erg3/erg11 and erg6 mutants, as evidenced by the high expression of RBT5 (ORF19.5636), FET34 (ORF19.4215), FTR1 (ORF19.7219), and FTH1 (ORF19.4802), and SIT1 (ORF19.2179). Genes responding to general plasma membrane and oxidative stressors also showed generally broad elevation [including CAT1 (ORF19.6229), GPX1 (ORF19.86), CRH11 (ORF19.2706), and DDR48 (ORF19.4082)]. Induction of calcineurin targets [UTR2 (ORF19.1671), RTA2 (ORF19.24), ECM331 (ORF19.4255)] was observed at varying levels as well, most strongly in the various lab strains [33],[34]. The level of constitutive activation of these pathways in AmB-resistant strains was in many cases comparable to levels seen in wild-type strains exposed to severe external stresses (Figure 3A, right panel). Intriguingly, the AmB-sensitive, fluconazole-resistant erg3 mutant did not show dramatic upregulation of any of the responses tested, but only a weak induction of several iron starvation genes. These data suggest that the mutations that confer resistance to amphotericin concomitantly exert an array of stresses to cellular membrane and redox homeostasis.
Hsp90-dependent stress-response pathways are not essential for growth of wild-type strains. In pathogenic fungi, validated Hsp90 clients include calcineurin, the MAP-Kinase Hog1, and Protein Kinase C (PKC) [30],[31],[35]. To test whether they become essential in the resistant mutants, we took advantage of the high conservation of these proteins: highly selective drugs targeting their human homologs have been developed for diverse therapeutic purposes, and these are active on the fungal proteins as well. These chemical probes allowed us to selectively reduce the activities of these proteins in the genetically intractable clinical isolates. In laboratory strains, these compounds allowed us to bypass the difficulties inherent in maintaining mutations expected to exhibit synthetic lethalities.
To inhibit calcineurin, we used FK-506 and Cyclosporin A, two structurally and mechanistically distinct inhibitors of the phosphatase. To inhibit PKC, we used enzastaurin, a synthetic PKC inhibitor with high selectivity for this kinase, and confirmed our findings with cercosporamide, a natural product fungal-specific inhibitor of PKC [31],[36]. Treatment with either calcineurin inhibitor inhibited growth of all of the Amphotericin-resistant strains, with complete growth inhibition of erg2 and erg6 mutants (Figure 3B). The erg3 erg11 strains were slightly less sensitive to calcineurin inhibitors, but showed a dramatically increased sensitivity to PKC inhibition.
We also asked if wild-type cells rely on Hsp90-dependent stress responses to defend themselves from the toxic effects of AmB. To do so, we created genetic knockouts of these normally nonessential genes in a wild-type background. Indeed, hog1 mutants were hypersensitive to AmB, while strains lacking calcineurin [cnb1(orf19.4009)] were not (Figure 3C). Mild inhibition of Hsp90, which is sufficient to impair calcineurin activity, did not change the AmB MIC. However, more extensive inhibition of Hsp90, which would destabilize Hog1 [35], did sensitize cells to the antifungal. Thus, Hog1 is required to tolerate the stress imposed by drug treatment, while the calcineurin and PKC pathways are required to tolerate the stress imposed by resistance mutations.
Next, we tested the role of Hsp90, Hog1, and calcineurin in the de novo emergence of AmB resistance. To do so, we generated ERG2/erg2Δ heterozygotes in a wild-type background and in strains lacking Hog1 or calcineurin. We then selected for loss of the remaining allele of ERG2 by plating on media containing AmB. As expected, ERG2/erg2Δ heterozygotes that were otherwise wild-type produced resistant colonies at a rate of ∼10−5 (Figure 3D). Low concentrations of the Hsp90 inhibitor geldanamycin completely eliminated the emergence of such colonies. Strains lacking calcineurin also failed to produce resistant colonies, and hog1 strains produced only a few small, slow-growing colonies. We conclude that Hsp90-dependent stress responses are required to enable the de novo emergence of AmB resistance.
Although we have successfully validated the ability of several ergosterol biosynthesis mutations to confer resistance to AmB in vitro, resistance rarely evolves during the treatment of infected patients. We wondered if the phenotypic benefit of AmB resistance might be undermined by fitness costs imposed by the mutations. That is, the high levels of internal stress that burden the AmB-resistant mutants might make them unable to withstand the additional external stresses imposed by the host. To investigate, we tested the ability of the resistant mutants and wild-type control to tolerate a range of stresses encountered in host environments, including (1) elevated temperatures (fevers are a universal response to systemic fungal infection and a common side-effect of AmB treatment); (2) hydrogen peroxide, hypochlorous acid, and nitric oxide (used by neutrophils to kill Candida); and (3) serum, iron deprivation, and antimicrobial peptides, as these are other common sources of stress in the host.
While AmB-resistant strains grew similarly to wild-type strains at 37°C and 39°C, they grew more poorly at 41°C (Figure 4A). Resistant mutants were hypersensitive to the presence of peroxide, hypochlorous acid, and the nitric oxide donor DPTA-NONOate (Figure 4A–C). Resistant mutants also proliferated moderately more slowly than wild-type cells in the presence of an iron chelator (Figure 4D). Resistant strains were also sensitive to growth in 100% bovine serum (Figure 4F) at elevated temperature, but were not more sensitive to the neutrophil-associated antimicrobial peptide Calprotecin/S100A (Figure 4E).
Next, we tested their susceptibility to attack by neutrophils, the most critical component of the innate immune system in combating acute fungal infection. We isolated human neutrophils from whole blood to 99% purity and activated them by treatment with recombinant TNF-α. Wild-type or AmB-resistant C. albicans strains were co-cultured with neutrophils for 6 h, at which point neutrophils were lysed and fungal growth was measured. The hog1 mutant, previously reported to be hypersensitive to neutrophil attack [37],[38], was included as a positive control. AmB-resistant mutants were indeed significantly hypersensitive to neutrophil attack, exhibiting at least as strong of a defect as the hog1 mutant (Figure 4G).
Another potential fitness cost of AmB-resistance mutations could be a compromise of pathogenic virulence mechanisms. C. albicans responds to several stimuli in the host environment by undergoing dramatic morphological changes, including the adoption of filamentous hyphal forms. Filamentation enables penetration and invasion of host tissue, and is a highly validated virulence factor for this pathogen [39],[40]. We asked if our AmB-resistant mutants can filament effectively when exposed to 10% fetal bovine serum in RPMI culture media. After 4 h of incubation, wild-type strains exhibited long and robust filaments (Figure 5A). The fluconazole-resistant erg3 mutant exhibited a mild delay in hyphal protrusion but still formed substantial filaments. However, the erg2 mutant could only form short and amorphous filaments. The erg6 and erg3 erg11 strains and clinical isolates were entirely unable to form hyphal extensions, and remained mainly in the yeast form.
We then asked if this defect in filamentation reduced the capacity of the pathogen for tissue invasion. Monolayers of primary human endothelial cells were established in culture, and infected with the C. albicans strains. Lysis of the endothelial cells was monitored by assaying the release of cytosolic lactate dehydrogenase (LDH). All amphotericin-resistant mutants showed dramatic defects in their ability to damage the monolayer (Figure 5B).
Finally, we compared the virulence of our AmB-resistant laboratory and clinical strains with that of wild-type and fluconazole-resistant (erg3) strains in a mouse model of Candida fungemia. To provide a rigorous test, we used a relatively high intravenous inoculum of 4×106 fungal cells in young Balb/c mice. At the time of sacrifice, both kidneys were isolated from each mouse. One was analyzed for fungal burden by homogenization and plating of CFU; the other was submitted for histological analysis.
The wild-type strain killed all infected mice within 1–2 d (Figure 6A). Necropsy revealed high viable fungal burden in the kidney, extensive filamentous fungal morphology, and moderate tissue damage (Figure 6B–C). The erg3 mutant demonstrated reduced virulence as previously reported [41]–[43], but still killed all mice in an average of 3–4 d. At the time of death, mice infected with this mutant had extremely high kidney fungal burdens, filamentous fungal morphology, and extensive kidney necrosis.
All AmB-resistant mutants, including the clinical isolate, were completely avirulent. Some mice showed mild weight loss in the first day after inoculation. But within a few days all infected mice recovered and appeared healthy. Kidney fungal burdens at 12 d postinoculation were at least three orders of magnitude lower than those of mice infected with wild-type Candida or the fluconazole-resistant strain (Figure 6B). Thus, the resistant strains failed to tolerate the host environment or immune attack and colonize this organ, let alone damage it. Histological analysis demonstrated healthy kidneys with some signs of resolving acute inflammation, suggesting that innate immune attack may have contributed to the clearance of these strains (Figure 6C). The lack of morbidity in mice infected with the clinical isolate suggests that this strain was recovered from the patient harboring it not because it was virulent but simply because it had survived AmB treatment.
The emergence of drug resistance has diminished the utility of nearly every class of antimicrobial drug. Yet, 50 years after its introduction, AmB remains as effective as ever. The failure of fungi to evolve resistance to AmB presents a considerable evolutionary puzzle with important consequences for human health. Our work offers a mechanistic solution. In a comprehensive search for mutations that can produce AmB resistance, we sequenced rare resistant clinical isolates, evolved resistant strains in the laboratory, and targeted candidate genes by site-directed recombination. Every mutation that was capable of conferring robust AmB resistance came at great cost to the pathogen. They all diminished Candida's ability to survive the diverse array of stresses that are inherent to growth in a mammalian host, crippled a major virulence factor required for invasive disease (filamentation), and eliminated the capacity to kill mice.
Certainly, AmB therapy often fails, but not because the fungus acquires resistance to the drug [44],[45]. Instead, treatment failure is linked to other factors, including the inability of the drug to penetrate certain niches of the body, dose-limiting renal toxicity, or the underlying disease of the patient [6],[7],[9]. While a small number of AmB-resistant clinical isolates have been reported in large surveys of clinical strain collections, we suspect that many (if not all) of these will prove to be avirulent, as was the one we tested. That is, they may have survived in a superficial niche less exposed to the stresses of the bloodstream but would not be capable of mounting a virulent systemic infection. Such strains might persist in patients with extreme immune system deficiencies. However, resistance to AmB is rare even in patients receiving myeloablative therapies that eliminate immune function [6],[7]. Even in the absence of stress from the immune system, the lack of filamentation and hypersensitivity to other aspects of the host environment (such as fever or iron deprivation) likely restricts the virulence of resistant strains.
Certainly, other explanations have been put forth for the rarity of AmB resistance and may also be relevant. For example, drugs that target lipids are not susceptible to resistance caused by substitutions in drug-binding pockets, a common occurrence with drugs that target proteins. In addition, because AmB acts on the plasma membrane, it is not susceptible to resistance mediated by increased drug efflux. Nevertheless, our work indicates that several deletion or loss of function mutations can readily arise in Candida and confer resistance to AmB in vitro. But these mutants do not become prevalent in the clinic. It might be argued that AmB resistance is rare because it is dosed intravenously and is not often employed in the types of long-term prophylaxis that breed resistance. However, the structurally similar polyene nystatin, which has the same ergosterol-binding mechanism of action as AmB, is widely used as a topical agent in the prevention of thrush in immunocompromised patients and the treatment of superficial rashes in neonates. Although there has been ample opportunity for resistance to emerge and become a clinical liability, it has not.
Our work also elaborates on the central role played by Hsp90 in potentiating the evolution of new phenotypes, but here it takes on a novel character. As previously reported, Hsp90 plays a critical role in the evolution of drug resistance in Candida and Aspergillus [25]. Hsp90 allows drug-resistant mutants to survive stresses imposed by triazoles and echinocandins, but is not required to tolerate the mutations conferring resistance to those drugs alone. Such is not the case for AmB. The alterations in sterol structure that confer AmB resistance cannot be achieved without causing a high level of constitutive stress. High levels of Hsp90, then, become essential to simply support viability, even in the absence of the drug. Thus, our findings illustrate yet another way that Hsp90 enables the acquisition of dramatic new phenotypes. Given its conservation, we suggest its role in supporting stress responses operates very broadly in the evolution of phenotypic diversity, allowing organisms to acquire mutations that confer novel phenotypes but simultaneously create stresses that otherwise would not be tolerated.
The clinical experiment of 50 years of AmB use indicates that the emergence of drug resistance, which widely plagues antimicrobial therapeutics, is not inevitable [46]. By elucidating the mechanisms that restrict the evolution of virulent AmB-resistant Candida, our work suggests a strategy that might be applied more broadly to prolong the ever-shortening window of efficacy encountered with new antibiotics: the development of compounds that exploit the high costs of resistance mechanisms. This strategy need not require the targeting of lipids. Advances in structural biology and medicinal chemistry have enabled the design of enzyme inhibitors that are less susceptible to resistance mediated by point mutations or by drug efflux [47]–[50]. Resistance to these agents may require the microbe to make more complex changes to its physiology and these, too, may come at a high cost. How might we discover targets that could induce such constraints upon resistance? One possibility is to focus on essential genes that also play critical roles in stress responses or virulence processes, for which rewiring of pathways may fundamentally alter pathogenicity.
In any case, fungal-selective inhibition of Hsp90 presents an attractive mechanism to prevent the emergence of drug resistance to all three antifungal classes in clinical use. Rooted in ancient and conserved biological processes, a similar strategy may prove useful in cancer, where resistance has greatly limited the efficacy of targeted therapeutics. Indeed, pharmacological inhibition of Hsp90 is now being explored as a strategy to forestall the emergence of resistance in diverse malignancies [28]. Investigating the mechanisms that support rapid evolutionary change with an eye to the constant challenges that cells face in their host environment presents a problem of broad biological interest with important clinical implications.
All animal protocols were conducted in accordance with the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. All animals were maintained according to the guidelines of the MIT Committee on Animal Care (CAC). These studies were approved by the MIT CAC (protocol #0312-024-15). All efforts were made to minimize suffering.
C. albicans and C. tropicalis strains were routinely maintained at 30°C in YPD (2% Bacto peptone, 2% dextrose, 1% yeast extract). Stocks were maintained in 15% glycerol at −80°C. For generation of deletion mutants, transformants were selected on synthetic medium (2% dextrose, 0.67% Difco yeast nitrogen base with ammonium sulfate) with an amino acid dropout mixture. RPMI 1640 media (Gibco) was buffered with 165 mM MOPS, pH 7.0, and supplemented with 2% dextrose.
Using an Illumina HiSeq platform with paired-end reads, we obtained an average coverage of 50-fold. After quality control filtering, reads from each sequenced genome were aligned against the Candida albicans SC5314 reference sequence, unless otherwise specified (Assembly 21, downloaded from the Candida genome project on June 27, 2011, available here: http://www.candidagenome.org/download/sequence/C_albicans_SC5314/Assembly21/archive/C_albicans_SC5314_version_A21-s01-m01-r03_chromosomes.fasta.gz) using the BWA aligner [51]. This was followed by variant calling with respect to this Candida albicans reference using the UnifiedGenotyper from version 1.0.5974 (the version we used throughout these analyses) of the Genome Analysis Toolkit (GATK) [52]. (To ensure that lower quality SNPs that are present in both a parental and derived strain were correctly identified as being common, we disabled the maximum deletion fraction in the call to the UnifiedGenotyper module. The specific parameters used were: “-dcov 1000 -stand_emit_conf 10.0 -stand_call_conf 50.0 –max_deletion_fraction 1.0.”) For the Candida tropicalis genome (strain OY5), only a preliminary assembly consisting of 24 scaffolds for the reference example, Candida tropicalis MYA 3404, is currently available. We downloaded assembly “ASM633v1” from NCBI (https://www.ncbi.nlm.nih.gov/nuccore?term=GG692395:GG692418[PACC] on December 20, 2012). Alignment and SNP calling for the C. tropicalis genome was performed as per the C. albicans genomes. All reads will be available at NCBI under BioProject accession numbers PRJNA194436 (http://www.ncbi.nlm.nih.gov/bioproject/194436) for C. albicans and PRJNA194439 (http://www.ncbi.nlm.nih.gov/bioproject/194439) for C. tropicalis; the umbrella project accession number is PRJNA195600 (http://www.ncbi.nlm.nih.gov/bioproject/195600).
To identify variants (including SNPs and indels) unique to a strain, we compared the “parental” strain to individual “derived” strains. We used a combination of custom code and the GATK's CombineVariants and SelectVariants features to locate, and then rank by quality, the SNPs and indels detected in open reading frames that were present only in derived strains. From the previously generated VCF files for each of the parental and derived strains as described above, the CombineVariants module was used to create a single list of SNPs (specifically we set the module options “-priority” to the name of the derived strain and the option “-genotypeMergeOptions” to “UNIQUIFY”). With this output VCF file, we employed the SelectVariants module to detect variants unique to the derived strain via the option: “-select "set = = <derived-strain>" ”. Additionally, to find cases where heterozygotes become homozygotes, or vice versa, we again used the SelectVariants by using the intersection feature: “-select "set = = Intersection" ”. The merged genotype calls within each common SNP were then further filtered to find high-quality calls where the zygosity changed. Alignments of the reads for the ranked SNPs and indels were then visually inspected in the Integrative Genomics Viewer (IGV) for quality control [53]. In the case of the clinical isolates, we did one pairwise comparison: between the SC5314 wild type and the ATCC 200955 AmB-R clinical isolate (“derived”). For the in vitro selection experiments, we defined the first isolate (#1) as parental, and compared the subsequent three serially derived isolates (#2, #3, #4) back to this parental isolate.
To visualize the LOH events in the in vitro selection series (BV01-BV05), we performed multisample SNP calling using the GATK UnifiedGenotyper module to generate a VCF file containing all SNPs in all four strains (UnifiedGenotyper options: “-glm SNP -nt 1 –downsample_to_coverage 100000”). Following quality filtering of SNPs, from this VCF file, we created a Python script to generate a list of positions of SNPs (relative to reference), if present in any of the four strains. For each SNP position in each of the strains, we were then able to classify whether it was homozygous for the reference base (blue), homozygous for the variant base (red), or heterozygous (white). The resulting “heterogram” for all chromosomes was visualized using the quilt.plot function from the R “fields” package [54]. Regions where a LOH event is likely to have occurred show up as blocks of blue and red SNPs (regions of high homozygosity) against the backdrop of white (heterozygous SNPs).
For each SNP in a given strain, we extracted the counts of reads containing the reference and variant base from the allele depth (“AD”) VCF annotation at that position. Using only SNPs of quality 1,000 or more, we then computed the “base ratio” at each position by dividing the count of reads for the minor allele base by the total number of reads for both bases at that SNP position, resulting in values of between 0 (complete homozygosity) and 0.5 (complete heterozygosity). We then averaged these base ratios over a 1 kb sliding window for each chromosome. We performed these analyses for the in vitro evolution series (BV01-BV04) of C. albicans as well as the single C. tropicalis isolate (OY5).
Sterols were extracted and analyzed as previously described [55]. Equal weights of cell pellets were used across strains. Absorption patterns were recorded by scanning between 240 and 300 nm at 0.5 nm intervals.
Strains and primers are listed in Tables S1 and S2. Deletion strains were constructed as described in [56], using HIS1, LEU2, and ARG4 markers (the URA3 marker was not used, and all strains used were URA3 wild type). Briefly, PCR products containing approximately 350 nucleotides of upstream and downstream homology for each gene were generated, and fusion PCR was used with a selectable auxotrophic marker to generate the knockout construct. Proper insertion of the auxotrophic marker and loss of the endogenous gene were confirmed by PCR. Two heterozygotes were selected from each initial knockout transformation for the knockout of the second allele, from which two knockout strains were tested for each (four total strains). The erg3 erg11 mutant was constructed as previously described [21]. These four strains were compared for phenotypic concordance in filamentation and stress resistance to minimize the effect of secondary mutations; additional data on a second mutant strain for key mutants is presented in Figure S5.
Antifungal susceptibility was determined in flat bottom, 96-well microtiter plates (Costar) using a broth microdilution protocol as described [25]. Overnight cultures were grown at 30°C in YPD for at least 16 h, and cell density was measured by OD600 before seeding approximately 103 cells per well in YPD or RPMI media, at 30°C or 37°C, as indicated in figure legends. Growth was measured at 24 or 48 h postincubation by alamar blue (Invitrogen) fluorescence with excitation at 550 nm and emission at 590 nm, and in certain cases confirmed by measurement of OD600 after agitation using a spectrophotometer (Tecan). MIC80 was defined as the concentration of drug reducing growth by 80% relative to the wells containing no drug. For susceptibility to AmB and stress response inhibitors (Figures 1D, 3B–C, and S5A), growth scores were determined by normalization of the values for each sample to the values obtained for the wild-type strain in the absence of AmB (or in the DMSO negative control for Figure 3B). For these assays (Figures 1D, 3B–C, and S5A), each condition was tested in duplicate and repeated on at least two different days. Relative growth data were quantitatively displayed in color using Java TreeView 1.1.3 (http://jtreeview.sourceforge.net). Sensitivity to stressors (Figure 4B–E) was similarly determined by microplate dilution and reading of alamar blue dye fluorescence; here, growth scores were calculated by dividing the growth value of the mutant by the growth value of the wild type in each particular stress condition. For stress assays (Figures 4B–E and S5B–C), data represent the mean of six wells, pooled from experiments performed on two separate days. Statistical significance was determined in Graphpad Prism 5.0 using the Student's t test function. Error bars represent the SEM for each group. Serum sensitivity was performed as previously described [57].
Cidal and static effects of Hsp90 inhibition were tested generally as previously described [29]. Cells were grown overnight in YPD and diluted to a concentration of 104 cells/mL at 30°C in YPD containing Hsp90 inhibitors at the indicated concentrations. After 24 h, cells were plated onto YPD at two dilutions and colonies counted.
Drugs used in growth assays included AmB (Fungizone, Invitrogen), fluconazole (TCI chemicals), radicicol (A.G. Scientific), geldanamycin (A.G. Scientific), Cyclosporin A (CalBiochem), FK-506 (A.G. Scientific), Cercosporamide (Sigma-Aldrich), and Enzastaurin (LC Labs). All drugs were dissolved in DMSO, with the exception of fluconazole (H20) and AmB, which was obtained as an aqueous suspension with sodium deoxycholate.
Spotting assays were performed by growing overnight cultures of strains in YPD at 30°C, washing in PBS, and resuspending in PBS at a concentration of 5×106 cells/mL. Four 5-fold serial dilutions were performed before spotting. All RPMI-agar plates were used within 6 h of pouring due to the potential instability of AmB and peroxides; it is recommended to test a range of AmB concentrations in agar plates due to potential chemical instability. Selection of AmB-resistant colonies was performed using ERG2/erg2Δ heterozygotes in wild-type, cnb1Δ/Δ, or hog1Δ/Δ backgrounds on RPMI-agar plates containing 0.4 µg/mL AmB; the wild-type was also selected on media containing AmB and 2.5 µM geldanamycin. Strains were grown overnight in YPD at 30°C, washed in PBS, and plated at a density of 8×106 cells per plate, and incubated for 2 d at 37°C before photographs were taken.
ATCC 10231 was grown overnight at 30°C in YPD and 2×108 cells were inoculated into one liter of YPD containing 0.25 µg/mL AmB at 30°C. Cultures were grown shaking for 24–48 h until turbidity was observed; cells were then removed, washed in PBS, and split for freezing glycerol stocks or reinoculation into media with a 2-fold higher concentration of AmB at the same cell density. The process was repeated until the concentration of AmB reached 32 µg/mL. The process was repeated in three independent selections. Strains were then thawed from glycerol stocks and struck to single colonies for future MIC assays.
Neutrophils were prepared fresh from the blood of a healthy human donor following standard protocols, using Histopaque 1077 density gradient centrifugation and hypotonic erythrocyte lysis [58]. After isolation, neutrophils were activated by treating with recombinant TNF-α (10 ng/mL). Killing assays were performed essentially as described in [38],[59]. Briefly, neutrophils were co-cultured with log-phase C. albicans strains at a 1∶1 ratio, with both cell types at a concentration of 104/mL. Control wells were inoculated in the identical conditions but without neutrophils added. Plates were incubated at 37°C in a humidified incubator for 6 h, at which point neutrophils were lysed by adding one volume of water containing 0.1% Tween-20 and a 1∶40 dilution of Alamar blue; wells were vigorously pipetted up and down. Alamar blue fluorescence was measured after 90 min of incubation at 37°C. Relative growth was measured by dividing values obtained in the presence of neutrophils by those obtained in their absence for each strain. Control wells lacking C. albicans were included to verify that this treatment does not quantify growth of the neutrophil cells. Results from three separate plates are shown; growth of each mutant strain is presented as a fraction of the wild-type growth from the same plate. Statistical analysis was performed using paired Student's t test in Microsoft Excel.
Hyphal induction was performed by growing C. albicans overnight at 30°C in YPD, washing in PBS, and diluting 1∶100 into RPMI+10% fetal bovine serum at 37°C (Sigma-Aldrich). After 2 or 4 h, cultures were briefly concentrated by centrifugation and visualized by DIC microscopy.
Endothelial cell invasion assays were performed with human umbilical vein endothelial cells (HUVEC's, Lonza) as previously described [60]. Monolayers were infected with C. albicans strains at a 1∶1 HUVEC∶fungus ratio and assayed for cytotoxicity with the CytoTox-96 Lactate dehydrogenase assay (Promega) after 6 h of co-incubation. Cytotoxicity was quantitated as the fraction of LDH release relative to a 100% value of wells treated with 1% Triton X-100 and a baseline value of HUVEC cells not infected with Candida. Error bars are indicative of the standard error of the mean for each group. Error bars indicate SEM. Results pooled from two experiments are displayed (six replicate wells per experiment). Statistical analysis was performed using unpaired Student's t test in Microsoft Excel.
For measurement of the expression of stress response genes, strains were grown overnight in YPD 30°C and diluted to OD600 of 0.15 in YPD, then grown for 5 h to mid-log phase and either centrifuged without treatment or subjected to different stresses. For stressed wild-type cells (SN250 strain), the conditions were as follows: AmB treatment with 1 µg/mL AmB for 15 min, Osmotic shock with 0.3M NaCl for 10 min, Nitrosative stress with 2 mM DPTA-NO (Cayman Chemical) for 15 min, calcium shock with 150 mM calcium chloride for 10 min, oxidative stress with 10 mM tert-butyl peroxide for 10 min, and iron chelation with 500 µM bathophenanthroline sulfonate for 4 h. Cultures were centrifuged at 1,500 g for 5 min and quickly flash frozen in liquid nitrogen. Total RNA was isolated with an RNeasy column kit (Qiagen), normalized to equal amounts of total RNA across samples, and reverse transcribed for 120 min with the high capacity reverse transcription kit (Applied Biosystems). qPCR was performed with SYBR green mastermix (Applied Biosystems) on an Applied biosystems ABI7900 thermal cycler, using oligonucleotides described in Table S2. Each measurement was obtained from an average of four wells, including two biological replicates and two technical replicates. Expression analysis was performed by the comparative ΔCt quantitation method, comparing mutant or stress-treated strains to wild-type untreated strains, using normalization to four internal control genes: TDH3, ACT1, TEF3, and RPP2B; the mean value obtained from the four normalizations was used. For representation by heatmap (Treeview), relative expression levels were determined by dividing each value by the maximum expression level for that gene in any tested condition, with the untreated wild-type samples set as the baseline (as certain genes were induced over 50-fold and others were not induced greater than 4-fold in any condition).
We utilized 7–9-wk-old female Balb/c mice from Charles River laboratory (n = 8 mice for WT, 10–14 mice for mutant strains). Each strain was tested in two independent experiments (performed at different times), and data were pooled. Strains to be injected were grown overnight in YPD, diluted, and grown for 5 h into mid-log phase at 30°C, then washed twice in phosphate buffered saline (PBS), counted by hemocytometer and plating of dilutions, and resuspended in PBS at a concentration of 4×107 cfu/mL. We used 100 µL of each suspension to inject mice by lateral tail-vein injection. Mice were weighed daily and monitored for signs of morbidity and sacrificed when body weight decreased by more than 20%. Kidneys were removed and either homogenized in PBS and plated for viable colony units (in duplicate) or submitted for fixation and staining with Periodic-acid Schiff stain. A veterinary pathologist was consulted for histological analysis. All experimental procedures were carried out according to NIH guidelines and MIT protocols for the ethical treatment of animals.
|
10.1371/journal.ppat.1002627 | Small-Molecule Inhibitors of Dengue-Virus Entry | Flavivirus envelope protein (E) mediates membrane fusion and viral entry from endosomes. A low-pH induced, dimer-to-trimer rearrangement and reconfiguration of the membrane-proximal “stem" of the E ectodomain draw together the viral and cellular membranes. We found stem-derived peptides from dengue virus (DV) bind stem-less E trimer and mimic the stem-reconfiguration step in the fusion pathway. We adapted this experiment as a high-throughput screen for small molecules that block peptide binding and thus may inhibit viral entry. A compound identified in this screen, 1662G07, and a number of its analogs reversibly inhibit DV infectivity. They do so by binding the prefusion, dimeric E on the virion surface, before adsorption to a cell. They also block viral fusion with liposomes. Structure-activity relationship studies have led to analogs with submicromolar IC90s against DV2, and certain analogs are active against DV serotypes 1,2, and 4. The compounds do not inhibit the closely related Kunjin virus. We propose that they bind in a previously identified, E-protein pocket, exposed on the virion surface and although this pocket is closed in the postfusion trimer, its mouth is fully accessible. Examination of the E-trimer coordinates (PDB 1OK8) shows that conformational fluctuations around the hinge could open the pocket without dissociating the trimer or otherwise generating molecular collisions. We propose that compounds such as 1662G07 trap the sE trimer in a “pocket-open" state, which has lost affinity for the stem peptide and cannot support the final “zipping up" of the stem.
| Fusion of viral and cellular membranes is necessary to establish infection by an enveloped virus. This process is facilitated by rearrangement of protein(s) present on the virion surface in response to molecular cues from the compartment from which fusion occurs, such as low pH of an endosome. Dengue virus is an enveloped virus in the flavivirus family; its “E" (for envelope) protein is the fusion mediator. We previously showed that peptides derived from the membrane proximal “stem" of the E protein bind a form of E that represents a late-stage fusion intermediate. We used this assay to screen for small-molecule inhibitors that compete for stem-peptide association with E. We describe one such inhibitor and its analogs that block viral fusion. These inhibitors also block infectivity if added to dengue virus before infection. Withdrawing the inhibitor before fusion reverses the blockage. We propose that these small molecules bind a hydrophobic pocket on the virion surface and that the virus carries them into the endosome, where they prevent viral fusion by stabilizing an intermediate conformation of the E protein that cannot complete the fusion-promoting conformational change. Identification of these fusion inhibitors shows that viral entry is a possible target for anti-flavivirus drugs.
| Enveloped viruses penetrate into the cytosol of their target cell by fusion of viral and cellular membranes [1], [2]. Flaviviruses, such as dengue, penetrate from endosomes, following uptake by clathrin-mediated endocytosis [3], [4]. At endosomal pH, proton binding by their envelope protein, E, triggers a fusion-promoting conformation change [5], [6].
The flavivirus envelope fusion protein, E, forms a well-ordered lattice of 90 dimers on the surface of a mature, infectious virus particle [2], [7]. Crystal structures of soluble forms of E (“sE"), which include the first ∼395 of ∼445 ectodomain residues but lack a conserved, membrane-proximal “stem" region, have contributed to molecular descriptions of flavivirus fusion [8]–[12]. The three domains (DI–III) of the E protein reorient with respect to one other during the fusion-promoting conformational transition, which includes dissociation of the prefusion dimer and reconfiguration of the subunits into trimers [2]. At an intermediate stage a hydrophobic “fusion loop" at one end of the extended E subunit inserts into the outer leaflet of the target bilayer [2], [13]. The driving force for pinching the two membranes together appears to come from contacts made by domain III, as it folds back against domain I, and by the stem, as it “zips" up along adjacent domain II monomers [1], [2].
Molecular understanding of the fusion pathway and the proteins involved has enabled discovery of small-molecule and peptide inhibitors that target intermediates in these fusion-inducing rearrangements. The best-known example of the latter type of entry inhibitor is T-20/enfuvirtide, a peptide used to treat HIV-1 infection [14]–[17]. The T-20 peptide interferes with a late stage in the fusion-inducing conformational transition of HIV-1 gp41. Certain small molecules block HIV-1 fusion by a similar mechanism, binding in a conserved pocket on the gp41 inner core [18]. Inhibitors that target the fusion glycoprotein, F1, of respiratory syncytial virus (RSV) also prevent infection by blocking a conformational transition [19], [20].
Targeting the HIV-1 and RSV glycoproteins is possible, because fusion occurs at the plasma membrane, where exposure of the relevant fusion intermediates allows straightforward access to the specific inhibitors. For viruses such as flaviviruses that fuse from endosomal compartments, however, targeting an intermediate of the rearranging fusion protein requires concentrating the inhibitor within the endosome, as its potential binding sites are not available until reduced pH has induced their exposure.
Kielian and co-workers have reported reconstitution of an sE trimer for both alpha- and flavivirus envelopes, suggesting that one might use reconstitution strategies to identify inhibitors that block steps in fusion [21], [22]. We found recently that we could target a fusion intermediate of dengue virus E with peptides derived from its ectodomain stem [23], [24]. These peptides bind the postfusion form of DV2 sE trimer, mimicking late steps in stem rearrangement. They inhibit in vitro fusion and DV2 infectivity. C-terminal modification with membrane targeting sequences increases their inhibitory strength [23]. A series of experiments support a two-step mechanism, in which a reversible, non-specific interaction with the viral membrane brings virion-associated peptides into the low-pH endosome, where full exposure of the peptide site on the E-protein conformational intermediate leads to tight, specific binding, which interfere with the final “zipping" of the stem [24].
Can small molecules also inhibit this step in the fusion pathway? We have adapted the assay we used to study interaction of stem-derived peptides with stem-less sE trimer, to screen for small-molecule inhibitors that target this fusion intermediate. We have identified compounds that compete for stem peptide association, and we show that they reversibly inhibit DV infectivity. We further show, using model liposomes, that these molecules specifically block viral fusion. They appear to bind the virion before adsorption to cells, by interacting with the prefusion, dimeric E protein, possibly in a previously identified, hydrophobic pocket. This association presumably permits their virus-associated transfer into endosomes. Limited structure-activity relationship studies have yielded compounds that inhibit DV2 (NGC isolate) with IC90∼1 µM. Our competition screen has thus identified a group of potent, small-molecule inhibitors of DV entry validating experimental screens for small molecules that block viral entry from internal compartments.
We have described a fluorescence polarization (FP) assay to identify stem-derived peptides that bind the trimeric postfusion conformer of DV2 sE [24]. We found peptides from the C-terminal stem region that bind tightly to this proposed fusion intermediate. With one such peptide, DV2419–447, tagged at its N-terminus with FITC, we adapted the FP assay to screen for small molecules that compete for binding to trimeric sE (Figure 1). We screened ∼30,000 compounds in a 384-well format and found several that were active, as measured by a reduction in the fluorescence polarization signal. We chose to pursue further work on one such “hit", 1662G07 (Figure 2), for which several structurally similar compounds were commercially available.
We tested a set of related compounds, both from the screening libraries and from commercial vendors, to obtain preliminary structure activity relationships (SAR). Several modifications to the parental scaffold affected competition with peptide (Table S1). Modification or removal of the nitrile moiety (R1 position) impaired or abolished activity. Removal of the halogen in the m-position on the R2 (as phenyl) decreased competition; addition of strongly electron withdrawing groups (either trifluoro- or trifluoromethoxy) in the o- or m-positions increased it. A thiophene at R3 was equivalent to the furan in 1662G07, but a cyclopropane in that position eliminated competition with peptide.
From preliminary SAR analysis with commercial analogs described above, we chose to synthesize two series of compounds based on the parental scaffold. Series 3-148, 3-149 and 3-151 varied the R2 position while the 3-110 varied at the R3 positions. (Figures 2 and 4, Tables S2 and S3). Among the 51 compounds from these series, we found that sixteen competed with stem-derived peptide for binding to the DV2 sE trimer. Strongly electron withdrawing groups in the o-, m-, or p- positions on R2 (as phenyl) enhanced activity; substitution with methyl or methoxy-methyl groups at these positions yielded inactive compounds. Some larger, heterocyclic rings impaired activity (Figures 2 and 4. Tables S2 and S3).
Do these small molecules, identified in a screen for binding to a late-stage fusion intermediate, inhibit DV2 viral infectivity? We used a standard plaque forming assay to test the effect of the analogs from the 3-148, 3-149, 3-151 and 3-110 series at a single concentration on growth of DV2 NGC. The virus inoculum was preincubated with compound for 15 minutes and then adsorbed to BHK-21 cells, at a multiplicity of infection (MOI) of 1, for 1 hour at 37°C. Supernatants were harvested after 24 hours and titred by standard plaque assay [23]. A subset of the compounds from both series reduced DV2 infectivity (Figures 2 and 4, Tables S2 and S3). Compounds were inactive against vesicular stomatitis virus, an unrelated enveloped virus and were noncytotoxic at the concentrations tested (Figure S7, Figures 2 and 4, Tables S1, S2 and S3). Comparison of the activity profiles from the 3-148, 3-149 and 3-151 series in the peptide-competition and viral infectivity assays revealed a clear concordance between active compounds in one assay and actives compounds in the other (Figure 3). The likelihood that this degree of concordance could be random is less than 10−4.
Compounds in the 3-110 series were particularly active against DV2. We therefore tested this series at a single concentration (5 µM) against isolates from the other three dengue serotypes: DV1 WP74, DV3 THD 3 and DV4 TVP360. The compounds of series 3-110 inhibit DV1, 3 and 4 infectivity to varying degrees (Figures 4 and S1 and Table S3). DV3 was particularly insensitive to inhibition, as we had also found when testing its response to stem-derived peptides. From this initial screen, we looked more closely at three analogs that appeared to have the strongest antiviral effect, 3-110-5, 3-110-14 and 3-110-22. We examined their inhibition of DV2 and DV4 viruses to determine IC90s. As seen in Figure 4 these compounds had strong antiviral activity against DV2 and DV4, with IC90s in submicromolar and micromolar ranges, respectively. At the same concentration used with the DV serotypes, none of the compounds had detectable activity against Kunjin, a subtype of West Nile virus (Figure S2); the analogs most potent for inhibiting dengue, 3-110-5, 3-110-14 and 3-110-22, had no effect on Kunjin, even at 20 µM.
The small-molecule inhibitors were selected in a screen that detects formation of an E-protein conformation adopted only after a virion has arrived in the low-pH environment of an endosome – an intracellular compartment presumably inaccessible to the free compounds. A series of order-of-addition experiments using the 3-148, 3-149 and 3-110 series show that to have a significant inhibitory effect, the compounds must be preincubated at 37°C with the viral inoculum before adsorption to cells (Figure 5). We observed the same level of inhibition using a direct plaque assay as a readout (Figure S8). When compound and virus inoculum were added to cells at the same time, we detected an approximately tenfold drop in viral titre compared with the DMSO control for compounds in the 3-110 series and little or no effect for compounds in the 3-148 and 3-149 series. Postinfection treatment of cells with compound one hour after initial adsorption of virus did not reduce viral titre, nor did pretreatment of cells for one hour before virus adsorption (Figure 5 and data not shown). These results imply a direct association of compound and virion before endocytosis of the virus.
To detect DV2 fusion with liposomes, we used the content-mixing assay we previously applied to characterize peptide inhibitors of DV [24]. Selected compounds from the 3-148, 3-149 and 3-110 series were incubated for 15 minutes at 37°C with virus, which was then added to trypsin-loaded liposomes. We adjusted the pH of the medium to ∼5.5 for 10 minutes, back-neutralized the samples, and incubated for an additional 45 minutes at 37°C to allow trypsin to act. Digestion of the viral core protein, which would have been exposed to protease only after fusion of virions with liposomes, was assessed by SDS-PAGE and immunoblotting. Protection of the core protein from proteolysis with retention of the envelope protein indicated an effective fusion inhibitor. A subset of the compounds we tested specifically blocked content mixing of virus with trypsin-loaded liposomes (Figure 6). We used stem peptide DV2419–447, previously shown to inhibit fusion, as a positive control.
The assay we used to find inhibitory compounds detects an interaction with trimeric E, which forms on virions only after exposure to low pH. Yet the inhibitory small molecules appear to bind virions at pH 7. Thus, the inhibitors can associate with both the pre- and postfusion E-protein conformers. Unless their binding site is at an interface between adjacent subunits in one of the two conformational states, we expect the compounds also to bind a monomeric form of E. We expressed and purified DI/DII, a soluble, monomeric fragment of E comprising only the first two domains (Figure S3) and showed by surface plasmon resonance (SPR) that the inhibitory small molecules from both series indeed bind directly and reversibly to DI/DII (Figure 7). Control compounds, which do not inhibit viral infectivity and do not compete for peptide binding to trimeric sE, do not bind DI/DII.
To rule out the possibility that the small molecules inactivate virions nonspecifically, we tested whether addition of DV2 DI/DII to an inoculum preincubated with a small-molecule inhibitor could restore infectivity. We incubated a virus inoculum for 10 minutes with selected compounds from the 3-148, 3-139 and 3-110 series at inhibitory concentrations and then added DI/DII in molar excess. Exogenous DI/DII indeed reversed the small-molecule inhibition (Figure 8). DV2 DI/DII alone did not affect viral titre. WNV DI/DII did not restore infectivity in the presence of these compounds (Figure S9), consistent with their failure to inhibit Kunjin virus.
We have shown that 1662G07 and it analogs inhibit growth of DV2 and that they block low-pH triggered fusion of virus with liposomes. We identified the parent compound, 1662G07, in a high-throughput screen that detected competition by active molecules with binding to trimeric sE of a fluorescein-tagged, stem-derived peptide. We designed this assay to represent the final step(s) in the low-pH triggered, E-protein conformational change – the process that induces penetration from endosomes through fusion of viral and endosomal membranes. Nonetheless, the inhibitory activity of these small molecules depends on binding to a virion before the virus encounters a cell, indicating that the compounds can also associate with E in its dimeric, pre-fusion conformation. Indeed, they bind in solution to a monomeric, DI/DII fragment of E, and addition of this fragment to an inoculum preincubated with one of the compounds restores infectivity, presumably by sequestering the inhibitor.
How can small molecules that bind the prefusion E dimer and the DI/DII fragment also block association of a stem-derived peptide with the sE trimer? From the known structures of sE dimers (prefusion) [9] and trimers (postfusion) [25] and from accurate docking of the former into subnanometer-resolution cryoEM reconstructions of virions [26], we can propose both an answer to this question and a model for the mechanism of action of the small-molecule inhibitors we have studied. The activity of these molecules in an assay for infectivity correlates well with their capacity to compete with a stem-derived peptide for binding to sE trimer. The most straightforward explanation for this correlation is that the compounds bind at a site accessible on both prefusion and postfusion E conformers. The most obvious site on E for small-molecule binding is a pocket, adjacent to the hinge between domains I and II, which accepts a β-octyl-glucoside (β-OG) molecule when sE dimers are crystallized in the presence of the detergent. This pocket closes down in the trimer conformation seen in the crystal structure, and the closed pocket is incompatible with occupancy by a bulky ligand (e.g., 1662G07). A dimer-to-trimer conformational transition will then require expulsion of the ligand, imposing a barrier to completion of the fusion process. For this reason, several groups have used in silico screens to find potential pocket-binding compounds, and in at least two cases, the results of those screens have yielded active inhibitors [27]–[31]. It has not yet been shown, however, whether the compounds found in this way indeed bind in the pocket as predicted. One of those computational screens used the Maybridge library for its search, and one of two active inhibitors identified is related to 1662G07, including most of the scaffold in Figure 2 [27]. That compound was not represented, however, in the version of the Maybridge library we used in our experimental screen. We have docked several of our compounds, using the GLIDE program [32]. We obtain fits consistent with the crystallographically observed interactions of β-OG (Figure S6).
Although the β-OG pocket is closed in the trimer, its mouth is fully accessible. Examination of the E-trimer coordinates (PDB 1OK8) shows that conformational fluctuations around the hinge could open the pocket without dissociating the trimer or otherwise generating molecular collisions (Figure 9). We suggest that compounds such as 1662G07 inhibit peptide binding by trapping the sE trimer in a “pocket-open" state, which has lost affinity for the stem peptide and cannot support the final “zipping up" of the stem (Figure 9). Binding of a compound in the β-OG pocket can explain how it accompanies virions into endosomes. Then, even if structural rearrangements of the E protein as it transitions at low pH from dimer to trimer expel the compound from the pocket, it would still remain at relatively high concentration in the endosomal space and be able to rebind rapidly. Effective inhibition would simply require that the rate of rebinding be higher than the rate at which the stem zips up along domain II.
One potent inhibitor of infectivity for all four dengue serotypes, compound 3-110-22, failed to inhibit stem-peptide binding in the competition assay (Figure 4). A likely explanation is that modification of the parental scaffold to produce 3-110-22 gave a compound with high affinity for the β-OG pocket in the E dimer, but lower affinity for the pocket-open state of the E trimer. The region that surrounds the mouth of the pocket differs in the two conformations, because of the fold-back of domain III (Figure 9), and 3-110-22 has a bulky substituent group. It appears that compound 3-110-22 indeed binds tightly to the E dimer, because unlike a number of others, it blocks the dimer-to trimer-transition of sE in vitro (Figure S4).
There are less likely alternative explanations for the inhibitory action of the compounds we have studied. One postulates a binding site, other than the β-OG pocket, that is present on monomeric DI/DII, on dimeric sE and on trimeric sE and that overlaps the peptide site on the trimer; another is a multi-site and multi-step mechanism. There is no evident candidate site for the former mechanism. The latter requires correlated affinities and properties of multiple sites. We therefore suggest that our peptide-competition, high-throughput screen has identified a large set of molecules that bind the β-OG pocket and that we have devised a useful assay for pocket-binding inhibitors, potentially applicable to any flavivirus for which one can prepare a stable, trimeric sE.
Stem peptide 419–447, with the DV2 NGC sequence and an RGKGR solubility tag appended at its C-terminus, was synthesized using standard Fmoc chemistry on an ABI 431 Peptide Synthesizers at the Tufts University Core Facility (Boston, MA), purified using reverse phase HPLC, and analyzed by mass spectrometry. Fluorescein-isothiocyanate (FITC) was conjugated to the N-terminus of the peptide through a β-alanine linker.
Liposomes [made with 1-palmitoyl-2-oleoyl-sn-glycero-3-phosphocholine (POPC), 1-palmitoyl-2-oleoyl-sn-glycero-3-phosphoethanolamine (POPE) (Avanti Polar Lipids) and cholesterol (Sigma-Aldrich) in a 1∶1∶2 molar ratio in TAN buffer (20 mM triethanolamine, 100 mM NaCl, pH 8.0)] were prepared by freeze-thaw extrusion through a 0.2 µ filter as described previously [24].
The postfusion sE trimer was produced as described [24]. Purified sE (Hawai'i Biotech) was incubated at 37°C in the presence of liposomes of the composition described above and acidified with MES buffer. Liposomes were solubilized with n-octyl-β-D-glucoside (β-OG) and n-undecyl-maltopyranoside (UDM). The solution was applied to a monoS column (GE Healthcare); sE trimer was eluted with a 2 M NaCl step gradient and further purified by size-exclusion chromatography on Superdex 200 (GE Healthcare). Protein was dialyzed extensively using a 50-KDa molecular-weight cutoff membrane (Spectrapor).
Synthetic analogs of 1662G07 were prepared as described in Text S1 [33].
All screens were preformed at the NSRB at Harvard Medical School. Binding experiments were carried out in Corning, low-volume 384 well microplates and analyzed in a PerkinElmer EnVisions instrument (excitation wavelength, 485 nm; emission wavelength, 535 nm). sE trimer was added to each well at 0.150 µM in 30 µL of TAN buffer. 0.1 µL of compound was transferred to each well and incubated at room temperature for 1 hr before the addition of DV2419–447 at a final concentration of 20 nM. After a 3-hour incubation, plates were read and fluorescence polarization measurements recorded. The original hit described in this paper came from the Maybridge 5 screening library at the NSRB at Harvard Medical School.
This assay was essentially as described previously [24], [34]. Liposomes were made with added trypsin at 10 mg/ml. Unencapsulated trypsin was removed by passage of the suspension through a Superdex 200 gel filtration column (GE Healthcare). Small molecules and peptide prepared as DMSO stocks were diluted into 50 µL of TAN buffer in the presence of purified virions. Reactions were incubated at 37°C for 15 mins, before the addition of trypsin-loaded liposomes, acidified with MES pretitrated to reach a final pH of 5.5, and incubated at 37°C for 15 mins. Reactions were neutralized to pH 8.0 with 1 M TEA. Trypsin digestion proceeded for 1 hr at 37°C. Aliquots of the reaction were resuspended in SDS-loading buffer with 2 mM PMSF, incubated for 20 mins at 100°C, and analyzed by SDS-PAGE followed by immunoblotting with an anti-dengue core and anti-E antibody.
Experiments were performed in duplicate on a Biacore 3000 instrument. DI/DII protein was immobilized to a CM5 biosensor chip per manufacturer's instructions. All experiments were carried out at 25 C in HBS-EP buffer (10 mM HEPES, 150 mM NaCl 3 mM EDTA and 0.005% (vol/vol) P20 surfactant). Sensorgrams were obtained by passing over small molecules diluted in HBS-EP buffer at specified concentrations at a flow rate of 50 µL/min with a 2 minute association phase and 10 minute dissociation phase. The sensor surface was not regenerated between experiments. Identical injections over blank lanes without protein were used and subtracted from the data to account for background and nonspecific interactions with the biosensor chip.
C6/36 cells were maintained in L-15 medium supplemented with 10% fetal bovine serum penicillin and streptomycin (Invitrogen). For viral plaque assays, BHK-21 cells were seeded (5×104cells/well) in 24-well, treated tissue-culture plates in α-MEM supplemented pen/strep antibiotics, and 5% Fetal Bovine Serum (FBS). Cells were plated <12 hrs before use and stored at 37°C with 5% CO2.
Dengue virus serotype 2 New Guinea Clone (NGC) was adsorbed to confluent layers of C6/36 cells for 1 hr at 25°C with rocking every 15 mins. L-15 medium (Mediatech) was added, and cells were incubated at 25°C until syncytium formation was observed. The supernatant was clarified by centrifugation at 1600 RPM at 4°C and stored at −80°C.
BHK-21 cells were seeded as described above. Aliquots from infections were diluted in 10 fold dilutions in Earle's balanced salt solution (EBSS), and 100 µl of each dilution were added to cells. Plates were incubated for 1 hr at 37°C and rocked every 15 mins. Unadsorbed virus was removed by washing with 1 ml PBS, after which 1 ml of α-MEM supplemented with 2% carboxymethylcellulose (CMC), pen/strep antibiotics, HEPES and 2% FBS, was added to each well and incubated at 37°C for 4 days. The CMC overlay was aspirated, and cells were washed 2× with 1 mL PBS and stained with crystal violet.
Virus supernatant was diluted in EBSS to a stock concentration that would allow for infection at MOI of 1, based on 50,000 seeded cells. Small molecules (or carrier) were added to the inoculum as indicated for each experiment. Cells were infected for 1 hr at 37°C with gentle rocking every 15 mins. Virus (or virus∶small-molecule mixtures) were washed from cells with 1 mL of PBS and overlay medium (α-MEM supplemented with HEPES, pen/strep antibiotics and 2% FBS) added. Plates were incubated at 37°C for 24 hrs. Aliquots of the supernatant were withdrawn and stored at −80°C.
BHK-21 cells were seeded at a density of 15,000 cells in a 96 well format. Compounds or vehicle were serially diluted in EBSS and 100 µl were transferred to each well. Plates were incubated at 37°C for 1 hr, media was aspirated and cells were washed 2× with 200 µl of PBS. 200 µl of α-MEM supplemented with pen/strep antibiotics, and 2% FBS was added and incubated for 24 hrs at 37°C. 20 uL of alamarBlue (Invitrogen) was added directly to each well and incubate for 2 hrs and read for absorbance at 570 nm.
|
10.1371/journal.pntd.0006416 | Zika virus infection in the Veterans Health Administration (VHA), 2015-2016 | Zika virus (ZIKV) is an important flavivirus infection. Although ZIKV infection is rarely fatal, risk for severe disease in adults is not well described. Our objective was to describe the spectrum of illness in U.S. Veterans with ZIKV infection.
Case series study including patients with laboratory-confirmed or presumed positive ZIKV infection in all Veterans Health Administration (VHA) medical centers. Adjusted odds ratios of clinical variables associated with hospitalization and neurologic complications was performed.
Of 1,538 patients tested between 12/2015-10/2016 and observed through 3/2017, 736 (48%) were RT-PCR or confirmed IgM positive; 655 (89%) were male, and 683 (93%) from VA Caribbean Healthcare System (VACHCS). Ninety-four (13%) were hospitalized, 91 (12%) in the VACHCS. Nineteen (3%) died after ZIKV infection. Hospitalization was associated with increased Charlson co-morbidity index (adjusted odds ratio [OR] 1.2; 95% confidence interval [CI], 1.1–1.3), underlying connective tissue disease (OR, 29.5; CI, 3.6–244.7), congestive heart failure (OR, 6; CI, 2–18.5), dementia (OR, 3.6; CI, 1.1–11.2), neurologic symptom presentation (OR, 3.9; CI, 1.7–9.2), leukocytosis (OR, 11.8; CI, 4.5–31), thrombocytopenia (OR, 7.8; CI, 3.3–18.6), acute kidney injury (OR, 28.9; CI, 5.8–145.1), or using glucocorticoids within 30 days of testing (OR, 13.3; CI 1.3–133). Patients presenting with rash were less likely to be hospitalized (OR, 0.29; CI, 0.13–0.66). Risk for neurologic complications increased with hospitalization (OR, 5.9; CI 2.9–12.2), cerebrovascular disease (OR 4.9; CI 1.7–14.4), and dementia (OR 2.8; CI 1.2–6.6).
Older Veterans with multiple comorbidities or presenting with neurologic symptoms were at increased risk for hospitalization and neurological complications after ZIKV infection.
| Zika virus (ZIKV) infection has become an important flavivirus infection that affected over a half of a million people in the Western Hemisphere by the end of 2016. Here we show risk factors for hospitalizations and neurologic complications in a US Veteran population. Over 700 Veterans with confirmed or presumed positive ZIKV were included. Our study showed that older Veterans with multiple comorbidities and those presenting with neurologic symptoms were more likely to be hospitalized, while if a patient presented with a rash they were less likely to be hospitalized. Neurologic complications were more likely in those hospitalized or those with a prior history of a cerebrovascular disease or dementia. Better understanding of those patients most at risk for severe disease can help providers when evaluating and treating patients with ZIKV infection.
| Zika virus (ZIKV) is a flavivirus transmitted primarily by Aedes species mosquitoes. Since the first reported primate ZIKV infection in 1947, sporadic human cases have occurred in Africa and Asia, followed by outbreaks in Micronesia and French Polynesia, culminating in widespread infection in the Americas in 2015–2016 [1–4]. In May 2015, locally transmitted infection in the Western Hemisphere was first reported in Brazil; the predominant strain was related to the Asian genotype [5]. ZIKV disseminated among this largely immunologically naïve population, where the World Health Organization estimates >534,000 confirmed or suspect cases, involving the majority of Western Hemisphere countries by the end of 2016 [6]. As of September 2017, 5,431 cases have been reported in the continental United States (U.S.), of which 5,155 were travel-associated, 225 were locally acquired mosquito-borne cases, and 36,644 cases were reported in Puerto Rico and the U.S. Virgin Islands [7].
ZIKV infection is often asymptomatic and usually self-limited, with most symptoms resolving in 7–10 days [4]. Patients typically present with rash, arthralgia, conjunctivitis, or fever [1, 3, 4]. More serious complications include congenital syndrome (microcephaly and other fetal abnormalities), Guillain-Barré syndrome (GBS) and other neurological disorders [8–23]. ZIKV is detectable for approximately 1 week in blood and 2 weeks in urine [24–26].
The Veterans Health Administration (VHA) has health care facilities throughout the U.S. and territories. We perform ongoing surveillance for emerging pathogens, and reported the first ZIKV case in Puerto Rico in December 2015 [27]. Since 46% of all U.S. Veterans and 62% of Veterans in Puerto Rico are aged ≥65 years and have significant comorbidities, they could be at higher risk for severe ZIKV infection compared to those in the general U.S. population exposed to the virus (i.e., returning travelers and those living in areas with local ZIKV transmission) [28–30]. Herein, we describe characteristics of ZIKV-infected Veterans and investigate risk factors for hospitalization and neurological complications.
ZIKV testing and surveillance were conducted as part of VHA operations and public health activities. As such, the VHA Office of Research Oversight considers public health investigations as operational and not research in VHA [31]. Since only ZIKV-positive cases required reporting to public health and U.S. federal agency authorities, negative cases were not reviewed. Patient data were anonymized after data abstraction for analyses.
We identified patients from all VHA facilities with ZIKV test results for specimens collected between December 1, 2015–October 31, 2016, utilizing VHA national data sources. Additional case finding was performed by querying inpatient and outpatient encounter data for Zika-specific International Classification of Diseases, Clinical Modification, 10th Revision (ICD-10-CM) code A92.5 and from VHA facility communications with VA leadership.
Testing for ZIKV was performed at the VHA’s Public Health Reference Laboratory (PHRL), and public health, federal or commercial laboratories. Testing and confirmation of ZIKV infection in our patient population is summarized in Fig 1. Initially, PHRL utilized a ZIKV reverse transcriptase PCR (RT-PCR) assay described previously [32]. Following U.S. Food and Drug Administration approval in March 2016, the Centers for Disease Control and Prevention (CDC) Trioplex RT-PCR assay for ZIKV, Dengue virus (DENV) and chikungunya virus (CHIKV) (in serum, whole blood, urine and spinal fluid) and ZIKV MAC IgM enzyme-linked immunosorbent assay (ELISA) (Anti-Zika Virus IgM Human MAC-ELISA kit, CDC) (for serum only) were used according to manufacturer’s recommendations [33, 34]. DENV (DENV Detect, InBios, Seattle, WA) and CHIKV (Abcam, Cambridge, MA) serum IgM ELISA assays were also performed per manufacturer’s recommendations. Testing methods performed by non-VA laboratories were unable to be confirmed. Samples with presumptive positive, equivocal or inconclusive ZIKV IgM results with an RT-PCR result that was either negative or not performed were sent to CDC for confirmatory testing using PRNT for DENV and ZIKV IgM [25]. Coinfection was defined as CHIKV and ZIKV-positive RT-PCR or IgM assays. In addition, patients with positive DENV RT-PCR and ZIKV RT-PCR assays would be considered coinfected. Potential cross-reaction of tests was defined as positive DENV IgM or RT-PCR and ZIKV IgM with plaque reduction neutralization test (PRNT) positive for DENV and ZIKV IgM results within 30 days of testing. Patients who were ZIKV RT-PCR and DENV IgM positive without PRNT were unable to be categorized as coinfected or cross-reactive. Since laboratory testing was performed based on clinician orders, not all assays were performed in all patients. In addition, CDC-recommended testing strategies changed over the course of 2016 [25, 26, 35].
For all patients with available ZIKV-positive diagnostic test results, we extracted demographics, clinical symptoms during acute illness and travel history from clinical notes, laboratory results (leucocyte, lymphocyte and platelet count; creatinine, alanine [ALT] and aspartate [AST] aminotransferases), concomitant medications (including 3-hydroxy-3-methylglutaryl-coenzyme (HMG-CoA) reductase inhibitors, antidiabetic, nonsteroidal anti-inflammatory drugs (NSAID), antidementia, oral glucocorticoids, antineoplastics, antivirals (for HIV-1 treatment), immunosuppressants, and intravenous immunoglobulin), hospitalizations, and outcomes from VHA electronic health records (EHR). Comorbidities were identified by extraction of discharge and encounter ICD-10-CM codes and provider notes. Age-adjusted Charlson co-morbidity indices (CCI) were calculated as a measure of a patient’s health status [36]). Laboratory cutoffs were as follows: leukopenia (<4,500 white blood cells [WBC]/μL); leukocytosis (>11,000 WBC/μL); lymphopenia (<1,000 lymphocytes/μL); and thrombocytopenia (<155,000 platelets/μL). Acute kidney injury (AKI) was calculated based on serum creatinine levels that were collected at presentation and the most recent serum creatinine prior to ZIKV infection (used as baseline). AKI was categorized into stage I (1.5–1.9 times baseline), stage II (2.0–2.9 times baseline), and stage III (≥3.0 times baseline) [37]. Abnormal hepatic function was determined by elevated transaminasemia 1–1.9 times upper limit of normal (ULN) (AST 34, ALT 40), 2–2.9 times ULN, and ≥3 times ULN [38]. Neurologic complications were determined by review of encounter ICD-10-CM codes after ZIKV diagnosis through 3/31/2017.
Positive cases were categorized as laboratory-confirmed which was defined as a patient with detectable ZIKV RNA by RT-PCR in serum or urine or a patient with positive ZIKV IgM ELISA result and confirmatory PRNT positive for ZIKV IgM only. A presumed positive case was a patient with a positive serum ZIKV IgM and negative DENV IgM result or not tested and a PRNT result positive for ZIKV and DENV IgM. Characteristics of patients with ZIKV infection diagnosed in VA Caribbean Health Care System (VACHCS) were compared to elsewhere in the U.S. In addition, factors associated with 1) hospitalization and 2) timing of diagnosis (in relation to infection) was assessed. Patients with a positive RT-PCR result for ZIKV (regardless of their IgM laboratory test results) were assumed to have been diagnosed earlier during their infection than patients with only an IgM-positive result for ZIKV (i.e., early vs. late where the latter was used as the referent group).
Student’s t-test and χ2 test were used to estimate associations between continuous and categorical variables, respectively. Logistic regression was used to estimate crude and adjusted odds ratios (OR) and 95% confidence intervals (CI) for factors associated with hospitalization and timing of diagnosis. For all clinical and medication-related data, “no” and “unknown” responses were combined as “no” and served as the referent group for all logistic regression models. A multi-stage backwards model building approach was used to develop a parsimonious main effects model (S1 Text and S1 Fig). That is, age group (in 10-year age categories), age-adjusted CCI, and the individual comorbidities used in the CCI were included in stage I. Non-significant comorbidities were removed from the model. Clinical findings, laboratory findings, and medications prescribed as an outpatient prior to ZIKV infection were added and subsequently removed (per non-significance) in stages II, III, and IV, respectively. Age group and age-adjusted CCI remained in the model, regardless of statistical significance until the completion of stage IV. The Kaplan-Meier log-rank test was used to estimate differences in length of stay. Non-parametric tests (e.g., Wilcoxon, Mann-Whitney) were used for non-normally distributed data (e.g., age among those who died). Death among patients with laboratory-confirmed ZIKV infection was analyzed in separate age-adjusted models. An alpha of 0.05 was used to determine statistical significance. Analyses were performed by using SAS 9.4 (SAS Institute, Inc., Cary, North Carolina).
We identified 1,538 VHA patients with ZIKV test results during December 2015– October 2016 (Fig 1). PHRL performed 1,424 (93%) of these tests and the remainder were performed at non-VA laboratories. Seven hundred thirty-six (48%) patients were RT-PCR-positive or serum IgM presumed positive confirmed with PRNT. Of these, 585 patients were laboratory-confirmed by RT-PCR (n = 569) or positive IgM with PRNT positive for ZIKV IgM only (n = 16). Per CDC guidelines and since there was a lack of active dengue cases seen by PCR or dengue specific IgM testing, the remaining 151 patients were presumed positive for ZIKV as their ZIKV IgM PRNT was positive for both ZIKV and DENV [24]. Demographic and clinical factors are summarized in Table 1; cumulatively, there were 655 (89%) male patients, with the majority (93%) of patients diagnosed at VACHCS, and the remaining 7% of patients diagnosed at 24 other VHA medical centers. Documented travel for those diagnosed in the continental U.S. included non-Puerto Rico Caribbean (18), Puerto Rico (16), Central America (9), South America (2), Indonesia (1), and Senegal (1). Six patients had exposure only in Florida. Mean age of all patients was 58.8 years (range 20–99). Patients from VACHCS versus returning travelers with ZIKV infection were older (mean age 60 versus 47 years; p< 0.001).
Four hundred seventy-four of 736 (65%) patients presented to an emergency department. Most common documented symptoms in patients with ZIKV infection were arthralgia/myalgia 92%, rash 90%, conjunctivitis 75%, and reported fever 66% (Table 2). Documented fever and myalgia/arthralgia or rash was reported for 378 (51%) patients and subjective fever and rash for 315 (43%) patients.
The distribution of laboratory findings during their ZIKV illness is shown in Table 2. Among ZIKV-positive patients, and of those who had hematology and chemistry testing performed, at their nadir, 37% had leukopenia (median, 3,700 WBC/μL; range, 800–4,400), 30% had lymphopenia (median, 740 lymphocytes/μL; range, 0–990), and 25% had thrombocytopenia (median, 126,500 platelets/μL; range, 17,000–149,000). Eleven percent had leukocytosis (median, 15,000 WBC/μL; range, 11,000–39,200). Twenty-five (5%) patients had acute kidney disease and 121 (32%) patients had elevated serum transaminases.
Concomitant use of HMG-CoA reductase inhibitors was the most frequently observed medication class (232 [32%]), followed by antidiabetics (124 [17%]), NSAID (83 [11%]), antidementia (40 [5%]), glucocorticoids (11 [2%]), antineoplastics (8 [1%]), antivirals (5 [0.7%]), and immunosuppressants (3 [0.4%]). No patients received intravenous immunoglobulin.
At VACHCS, 91 (12%) of 683 patients were hospitalized with median acute care length of stay (LOS) of 6 days (range 1–214 days), including 20 (3%) who were admitted to intensive care (ICU) with median LOS of 4 days (range 1–30 days); at VHA hospitals elsewhere in the U.S., 3 of 53 (6%) returning travelers with known hospitalization status were hospitalized with median LOS of 4 days (range 1–6 days) but none were admitted to an ICU. The length of stay between these two groups was not significantly different (p = 0.13). The average age among hospitalized patients was higher among the 91 patients at VACHCS than the three returning travelers in the continental U.S. (75 vs. 60 years, p<0.001). Crude measures of association with hospitalization are shown in Table 3.
Adjusted ORs, 95% CI and p-values controlling for all significantly associated factors are presented in Table 4. The odds of hospitalization significantly increased with CCI, connective tissue disease, congestive heart failure, dementia, neurologic symptoms, GBS, leukocytosis, thrombocytopenia, AKI, and glucocorticoid steroid use within 30 days of ZIKV testing. Patients presenting with a rash were less likely to be hospitalized. In additional adjusted analyses reported in Table 4, only having rash, conjunctivitis, leukopenia or lymphopenia at presentation were significantly associated with a positive RT-PCR test.
Forty-six (6%) patients with ZIKV infection (37 confirmed, 9 presumed positive) also had neurologic complications as summarized in Table 1. Five patients had cerebrospinal fluid (CSF) tested for ZIKV, all of whom had at least one of the identified neurologic complications. One patient with altered mental status, meningitis and viral encephalitis was positive for ZIKV by RT-PCR in CSF and serum. CSF findings for this patient were consistent with a viral etiology, demonstrating mild pleocytosis (WBC 12/cm3, 88% polymorphonuclear leukocytes, 12% lymphocytes) and normal CSF protein level (30.2 mg/dL). As shown in Table 4, neurologic complications were significantly more likely in patients with a prior history of cerebrovascular disease (CVD) and dementia as well as those who had been hospitalized.
Of 81 women with a positive ZIKV test, 50 were of childbearing age (18–52 years old) and four were pregnant at the time of infection. Two of these patients were from VACHCS and two were identified as returning travelers. Three patients delivered their babies (further details on the outcome of the babies is unknown) and one patient miscarried at 9.5 weeks.
Fourteen patients were positive for DENV IgM and ZIKV RT-PCR alone, of whom one was hospitalized. These 14 were unable to be categorized as coinfection or cross-reaction as they did not have PRNT performed. Three additional patients were positive for all three viruses (CHIKV IgM, DENV IgM and ZIKV RT-PCR), of whom one was hospitalized. Fifty-six patients (8%), all diagnosed at VACHCS, were positive for ZIKV (RT-PCR [n = 43] or IgM [n = 13]) and CHIKV IgM coinfection, of whom nine (16%) were hospitalized. In adjusted analysis, age was significantly associated with coinfection and arthralgia/myalgia was significantly less common in these patients. There was no increased risk of hospitalization or neurologic complications associated with coinfection.
Nineteen (3%) patients died post-ZIKV infection, all of whom presented to VACHCS with ZIKV related symptoms of which 16 were hospitalized (Table 5). Fourteen of 19 had viremia at presentation (Table 5). The mean age of ZIKV patients who died was 82 years (range, 50–99 years), compared to 59 years (range, 20–98 years) for those at VACHCS who survived (p<0.001). The mean time from ZIKV testing until death was 39 days (range 3–104 days). Eighteen (95%) had at least one CCI condition. Thus, it was difficult to determine whether ZIKV infection contributed to death or not.
Our study is the first to characterize U.S. Veterans with ZIKV infection. Testing varied based on test availability, provider preference, or presenting symptoms. The majority received a diagnosis in Puerto Rico, although 53 were returning travelers or had locally acquired infection elsewhere in the U.S. Among returning travelers, three were hospitalized, whereas in Puerto Rico, where patients were older and had more comorbidities, approximately 13% of patients with ZIKV infection were hospitalized, of whom 3% were admitted to ICU, and 3% died post ZIKV infection. Although we cannot directly link the deaths with ZIKV, the number of deaths was higher among VHA patients compared with a report from Puerto Rico in December 2016 that described only 5 deaths identified by surveillance on the island [39]. CCI was associated with increased risk for hospitalization which was possibly related to a lower threshold for hospitalization in those with significant chronic illness. After adjusting for CCI, connective tissue disease, CHF, dementia as well as presenting with neurologic symptoms, leukocytosis, thrombocytopenia, AKI or being prescribed glucocorticoids 30 days prior to ZIKV diagnosis was associated with increased risk for hospitalization. However, presenting with a rash made hospitalization less likely and no patients receiving a NSAID were hospitalized.
Hospitalization and deaths are reported to be uncommon in ZIKV infection [3, 40–44]. During the 2007 ZIKV outbreak in Micronesia, among 49 confirmed and 59 probable cases, patients presented with typical symptoms described here, but none were hospitalized and none died [3]. Although patients in that study were on average >10 years younger and fewer had comorbidities than U.S. Veterans. In Brazil, among 119 ZIKV confirmed patients only one hospitalization and no deaths were reported [41]. Hospitalizations and death (<1%) were noted in Puerto Rico from November 2015–July 2016 [40]. In our Veteran population, 3% died after a ZIKV diagnosis and 13% were hospitalized which is higher than other ZIKV studies and may be related to Veterans increased comorbidities [30].
Several studies have documented coinfection with ZIKV and CHIKV [45–47]. A prior study identified patients from Nicaragua with positive CHIKV and ZIKV [47]. Since cross-reaction is unlikely between these viruses, these patients were noted to have coinfection. In their study, 16/263 (6%) ZIKV-positive patients were noted to have coinfection with CHIKV and ZIKV [47]. In the Nicaraguan cohort, patients with coinfection trended toward more hospitalization and had similar symptoms to those monoinfected [47]. In our study, 56/736 (8%) patients were identified as being positive for ZIKV and CHIKV IgM (with or without positive DENV IgM). No patients were identified with ZIKV and CHIKV or DENV by RT-PCR. In patients with coinfection, there was no increased risk in hospitalization or neurologic complications but there was an increased risk of coinfection with advanced age. Symptoms were similar between groups except there was a decrease in documented arthralgia/myalgia in coinfected patients.
ZIKV has been documented to have congenital as well as neurologic complications [8–23]. Forty-six patients in our cohort were also noted to have neurologic complications after ZIKV infection. While these neurologic complications are quite broad, they identify potential complications post-ZIKV infection. Prior history of CVD and dementia as well as being hospitalized with ZIKV increased risk of neurologic complication. It was difficult to confirm whether these other neurologic complications were the result of ZIKV infection. Since neonatal and pediatric care was not provided by VA, the status of the infants exposed to ZIKV is unknown.
There are several limitations to our retrospective study. Cases not tested or with results not documented in VA’s EHR could not be identified; asymptomatic and mild cases were unlikely to have testing performed; early dated cases were not tested for ZIKV IgM as it was not available at the time of clinical testing, so some cases that were RT-PCR negative may have been missed; samples from VACHCS prior to December 2015 that were tested for DENV and CHIKV were not tested for ZIKV; samples from facilities in the continental U.S. were only tested for those ordered by the provider and complete testing may not have been ordered by provider depending on timing of symptoms and possible exposure. Some health departments had strict testing criteria and submitted Veteran samples may have been rejected or not tested. Across the VA system, and specifically at VACHCS, testing was not restricted, particularly for testing performed in VA, as there was an ongoing outbreak. We were unable to determine if deaths or neurologic complications were directly related to ZIKV infection. No ICD-10-CM diagnosis code was available for ZIKV until 10/1/2016 for additional case identification purposes. Medications not obtained within the VA could not be identified. Patients receiving care outside of the VA were unable to be reviewed for neurologic complications. Only Veterans who presented to VHA facilities and had appropriate diagnostic testing completed were included. Since asymptomatic individuals were unlikely to be tested for ZIKV, overall burden of disease was unable to be determined. Given the nature of the investigation, primary focus was placed on ZIKV-positive patients. Sample sizes among certain subgroups limited inferences from statistical analyses. Veterans represent a unique group of patients who tend to have increased age and comorbidities compared to the general population [30]. Among returning travelers, many of whom presented for care in the U.S. during the convalescent period, when diagnosis is dependent upon serology, some diagnoses could have been missed as ZIKV IgM typically declines after several weeks to months [26].
Clinicians practicing in areas with ZIKV transmission should be aware that ZIKV infection among elderly patients and patients with comorbidities, including connective tissue disease, dementia and CHF, those on glucocorticoids, and those presenting with neurologic symptoms, leukocytosis, AKI, and thrombocytopenia may have more severe disease. In addition, patients hospitalized and those with prior history of CVD and dementia were more likely to have neurologic complications.
Larger studies are required to determine risks associated with atypical complications, intensive care utilization and death associated with ZIKV infection; and whether prevention strategies or closer monitoring for those at greatest risk for such complications after ZIKV infection should be targeted.
|
10.1371/journal.ppat.1004734 | Dermal Neutrophil, Macrophage and Dendritic Cell Responses to Yersinia pestis Transmitted by Fleas | Yersinia pestis, the causative agent of plague, is typically transmitted by the bite of an infected flea. Many aspects of mammalian innate immune response early after Y. pestis infection remain poorly understood. A previous study by our lab showed that neutrophils are the most prominent cell type recruited to the injection site after intradermal needle inoculation of Y. pestis, suggesting that neutrophil interactions with Y. pestis may be important in bubonic plague pathogenesis. In the present study, we developed new tools allowing for intravital microscopy of Y. pestis in the dermis of an infected mouse after transmission by its natural route of infection, the bite of an infected flea. We found that uninfected flea bites typically induced minimal neutrophil recruitment. The magnitude of neutrophil response to flea-transmitted Y. pestis varied considerably and appeared to correspond to the number of bacteria deposited at the bite site. Macrophages migrated towards flea bite sites and interacted with small numbers of flea-transmitted bacteria. Consistent with a previous study, we observed minimal interaction between Y. pestis and dendritic cells; however, dendritic cells did consistently migrate towards flea bite sites containing Y. pestis. Interestingly, we often recovered viable Y. pestis from the draining lymph node (dLN) 1 h after flea feeding, indicating that the migration of bacteria from the dermis to the dLN may be more rapid than previously reported. Overall, the innate cellular host responses to flea-transmitted Y. pestis differed from and were more variable than responses to needle-inoculated bacteria. This work highlights the importance of studying the interactions between fleas, Y. pestis and the mammalian host to gain a better understanding of the early events in plague pathogenesis.
| Flea-borne transmission is central to the natural history of the plague bacillus Yersinia pestis, and infection within the context of flea feeding may affect the pathogenesis of bubonic plague. We analyzed the mammalian host response to Y. pestis in the skin immediately after transmission by its natural vector, the rat flea Xenopsylla cheopis, to observe differences relative to the response to needle-inoculated bacteria. Our results show that uninfected flea bites induce minimal inflammation, but flea-transmitted Y. pestis cause the recruitment of neutrophils roughly in proportion to the number of bacteria deposited in the skin. We observed interactions of flea-transmitted bacteria with macrophages, a cell type much more permissive than neutrophils for survival and growth of Y. pestis. We found that dendritic cells, important sentinel antigen presenting cells, were recruited to, but had minimal interaction with, flea-transmitted bacteria. Additionally, we found that Y. pestis could disseminate from the flea bite site to the draining lymph node and spleen as early as 1 h after flea feeding, significantly earlier than has been previously reported. This study reveals important differences between needle-inoculated and flea-transmitted Y. pestis in the immediate host response to infection and improves our understanding of the early host-bacterium interactions in plague pathogenesis.
| Bubonic plague is the most common form of plague in humans and is the result of transmission of Yersinia pestis into the dermis via the bite of an infected flea. The bacteria survive in the skin and eventually disseminate to the dLN where they replicate to high numbers forming an enlarged lymph node termed a bubo. The cellular architecture of this bubo eventually becomes compromised resulting in hematogenous spread of the bacteria followed rapidly by death of the host. Fleas can also deposit bacteria directly into the bloodstream of a mammalian host resulting in primary septicemic plague that may constitute as many as one third of human cases [1,2].
Y. pestis evolved from its closest relative Y. pseudotuberculosis, approximately 1500 to 6400 years ago [3]. An essential step in evolution from an orally acquired pathogen that causes mild gastroenteritis to a highly pathogenic, flea-transmitted pathogen was aquistion of the ability to form a biofilm in the flea [4]. This biofilm blocks the proventriculus, a valve structure between the esophagus and midgut of the flea, and interferes with the flea’s ability to take a blood meal [5]. Blocked fleas make repeated attempts to feed until they eventually succumb to starvation or dehydration. Lorange et al. studied the vector efficiency of blocked rat fleas and found that less than half of the fleas transmitted Y. pestis while attempting to feed. For the fleas that did transmit, as many as 4000 CFU were detected, but the median number transmitted was 82 CFU [6].
The exact events that occur in the dermis immediately after deposition of Y. pestis by a flea remain enigmatic. Macrophages are considered permissive for Y. pestis survival whereas neutrophils are much more bactericidal toward the organism [7]; however, up to 10% of Y. pestis may survive after phagocytosis by neutrophils [8,9]. Y. pestis has been observed within macrophages and neutrophils early after intraperitoneal infection [10], but it is unclear if an intracellular phase is important in bubonic plague pathogenesis.
The most important virulence factor of Y. pestis is the pCD1 plasmid-encoded type III secretion system (T3SS). The T3SS effector proteins are preferentially translocated into phagocytes in vivo [11] where they disrupt multiple signaling pathways in phagocytes resulting in cellular paralysis, necrosis or apoptosis [12]. Genes encoding the T3SS are induced by growth at 37°C, but minimally expressed in the flea midgut [13, 14]. Y. pestis also produces a proteinaceous antiphagocytic capsule called F1. Similar to the T3SS, the F1 capsule is poorly expressed in the flea and induced by growth at 37°C [14,15]. Thus, there is likely a period immediately after deposition of the bacteria in the dermis until the T3SS apparatus, its secreted effectors and F1 capsule can be expressed when the Y. pestis is vulnerable to phagocytes.
Neutrophils are highly phagocytic innate immune cells that ingest and destroy invading bacteria. We have previously used intravital microscopy to examine Y. pestis-host cell interactions in vivo [16]. We found that large numbers of neutrophils are recruited to the infection site within 2–3h after i.d. injection of Y. pestis. Interestingly, recruited neutrophils rapidly associated with bacteria and many trafficked Y. pestis away from the injection site. In contrast, dendritic cells (DC), potent antigen presenting cells, were not recruited to the injection site and showed minimal interaction with bacteria [16].
Many previous studies of the early events following Y. pestis infection, including our own, have used intradermal needle inoculation to model bubonic plague transmission. However, needle inoculation differs from the natural route of transmission, the bite of an infected flea, in a number of ways. The flea mouthparts that are inserted into the skin during feeding are at least an order of magnitude smaller in diameter than the 30 gauge needle used for i.d. injections. Flea saliva also contains a number of molecules whose homologs in other blood feeding arthropods affect innate immunity [17]. Additionally, Y. pestis isolated from flea midguts display a markedly different phenotype when compared to in vitro cultured bacteria, including increased expression of biofilm extracellular matrix components and antiphagocytic factors [14]. Thus, we hypothesized that transmission by the natural plague vector might alter the numbers or composition of innate immune cells recruited to the site of infection and their interactions with bacteria. The low transmission efficiency of the flea vector makes quantitative assessment of cellular interactions difficult and led us to develop intravital microscopy methods to study these interactions in vivo. The goal of the present study was to characterize the Y. pestis-host cell interactions that occur in the dermis early after transmission of bacteria by the rat flea Xenopsylla cheopis.
Before we could characterize the responses to flea-transmitted Y. pestis in vivo, we needed to develop methods for reliably identifying flea bite sites on the mouse ear and characterize the neutrophil response to uninfected flea bites. To examine the gross effects of flea feeding on mouse skin, we constructed a simple clamp-on feeding chamber that would allow fleas to feed on a mouse ear (S1 Fig). The chamber containing fleas was placed on the ear for 50 min and dissecting microscope images of the ear were captured before and after feeding. Overall, the most noticeable effect of uninfected flea feeding was marked vasodilation in the ear (Fig. 1A). Occasionally, a flea bite would result in a small, discrete erythematous spot at the bite site, but more often there was no obvious visible indicator of where the fleas had fed. Similar results were seen after feeding of uninfected and infected fleas (S2 Fig). The absence of any consistent localized gross pathology after flea feeding made it difficult to reliably identify flea bite sites. Fortunately, intraperitoneal injection of mice with the non-membrane permeant fluorescent DNA stain Sytox Blue prior to flea feeding resulted in the staining of cells damaged as the fleas inserted their mouthparts into the skin. Foci of Sytox Blue stained nuclei of damaged cells can be seen in areas where fleas have fed (Fig. 1B). To confirm that these areas are flea bites, we injected mice i.v. with the vascular dye Q655 prior to flea exposure. Damage to blood vessels during flea feeding caused localized vascular dye leakage resulting in bright Q655 staining surrounding the vessel. These bright Q655 areas corresponded to areas of Sytox Blue staining (Fig. 1C, D left panel). As further confirmation of our ability to identify flea bite sites, we anesthetized fleas while their mouthparts were embedded in the mouse skin and used microscissors to cut the mouthparts off above the skin. The highly autofluorescent nature of the flea exoskeleton allowed for confocal imaging of the mouthparts, which were found embedded in an area containing both Sytox Blue and Q655 staining (Fig. 1C). Thus, the Sytox Blue method is a reliable way of identifying flea bite sites on mouse skin.
To characterize the neutrophil response to individual uninfected flea bites, we fed fleas on LysM-eGFP transgenic mice that express high levels of eGFP in neutrophils and lower levels in macrophages in the skin [18]. The total neutrophil recruitment to the flea bite sites was evaluated and assigned a numerical score from 0 (no recruitment of neutrophils) to 4 (influx of large numbers of neutrophil resulting in a aggregated mass of eGFPbright cells at the bite site). We found that uninfected flea feeding recruited remarkably few eGFPbright neutrophils to the flea bite, despite the cellular damage and vascular leakage present at the bite site (identified by Sytox Blue or Q655 staining, respectively) (Fig. 1D, S1 Video). Neutrophil recruitment scores for four independent experiments ranged from 0 to 2 with a mean of 0.9 +/-0.4 (Fig. 2). In Fig. 1D multiple flea bites can be seen in the micrograph by Sytox Blue and Q655 staining. Neutrophils appear to be much more heavily recruited to the flea bite near the center of the field. Interestingly, we observed the mobilization and migration of eGFPdim cells towards the bite site over the course of the experiment (S1 Video). Similar eGFPdim cell movement towards uninfected flea bites was observed in four independent experiments. These eGFPdim cells have been characterized as F4/80+, CD11b+ macrophages in the dermis of the LysM-eGFP mouse [18]. Thus, uninfected flea bites result in very little neutrophil recruitment and resident macrophages appear to migrate towards flea bite sites.
Y. pestis is typically transmitted by fleas in which the bacteria have established a biofilm that blocks the proventriculus. These blocked fleas are unable to take a normal blood meal and make repeated unsuccessful attempts to feed, partially withdrawing their mouthparts and reprobing. We hypothesized that this might result in more damage to the skin and consequently increased neutrophil recruitment in comparison to uninfected flea bites. To test this, blocked fleas infected with a T3SS deficient strain of Y. pestis expressing the fluorescent protein mCherry were allowed to feed on LysM-eGFP mice for approximately 50 min. Again, the Sytox Blue reagent was used to identify flea bite sites. We found obvious flea bites where bacteria could not be detected in >50% of the experiments involving feeding of blocked fleas, which is in agreement with what is known about flea transmission efficiency [6]. These bites were imaged to determine the neutrophil response to blocked flea bites without the influence of bacteria at the bite site. We observed a highly variable neutrophil response to blocked flea bites (Fig. 3). The responses ranged from recruitment of very few neutrophils (Fig. 3A, S2 Video), similar to what is seen with uninfected fleas, to an influx of large numbers of neutrophils to the flea bite site (Fig. 3B, S3 Video). When the neutrophil recruitment in nine independent experiments was scored, the results ranged from scores of 0 to 4 with a average score of 2 +/- 0.4 (Fig. 2). The skin of mice fed on by blocked fleas had more foci of Sytox Blue staining than skin fed on by uninfected fleas and these foci often appeared in clusters, presumably due to a flea making repeated attempts to feed in the same location. The numbers of neutrophils recruited did not appear to correlate with the amount of Sytox Blue staining at the bite site (Fig. 3). Thus, the neutrophil response to blocked flea bites is much more variable than the response to uninfected flea bites.
Similar to observations of uninfected flea bites, movement and migration of eGFPdim macrophages towards blocked flea bites was common, occurring in seven out of nine independent experiments. In the two experiments where macrophage migration was not seen, large numbers of eGFPbright neutrophils were recruited to the bite site, which may have obscured observation of the macrophage movement.
To characterize the neutrophil response to Y. pestis transmitted via the bite of an infected flea, blocked fleas infected with Y. pestis pMcherry were fed on a LysM-eGFP mice, flea bites were identified by Sytox Blue staining, and bite sites that contained mCherry+ bacteria were imaged. Consistent with previous studies [6], fleas transmitted a highly variable number of bacteria into the skin (determined qualitatively by image analysis). We show three example experiments representing responses to low (roughly ten or fewer), moderate (roughly hundreds) and high (roughly thousands) numbers of transmitted bacteria as determined by visual inspection of the bite site (Figs. 4, S4, S5 and S6 Video). For the purpose of comparison, images from the 0 h and 4 h timepoints after needle inoculation of Y. pestis (~1000 CFU) or a sterile 30 gauge needle stick alone are also shown in Fig. 4. In seven independent experiments where bacteria could be seen at flea bites, the neutrophil recruitment scores ranged from 0 to 4 with an average of 2.8 +/- 0.4 (Fig. 2). Overall, the number of neutrophils recruited to bite sites containing bacteria was higher than for uninfected flea bites or blocked flea bites without transmission. Furthermore, neutrophil recruitment appeared to correlate with the number of bacteria present at the bite site (Fig. 2).
Interestingly, even when large numbers of neutrophils were recruited to the bite site and associated with bacteria, very little translocation of the bacteria was observed. When bacteria were observed moving, they were frequently (observed in four out of seven experiments where bacteria were present at the flea bite site) associated with eGFPdim cells, which are likely macrophages (Fig. 5, S7 Video). Movement of bacteria in association with eGFPbright neutrophils was a rare event, observed in only 1 of the 7 experiments. This is in contrast to what is observed after needle inoculation of bacteria into the dermis, where many bacteria are trafficked away from the injection site in association with neutrophils [16].
Because the neutrophil recruitment to needle-inoculated Y. pestis is so robust, the presence of large numbers of eGFPbright cells may obscure bacteria-eGFPdim macrophage interactions. To address this possibility, we treated Lys-eGFP mice with anti-GR1, an antibody that efficiently depletes neutrophils and, to a lesser extent, inflammatory monocytes, thus permitting the visualization of the macrophage response to Y. pestis in the near absence of neutrophils. We observed movement of eGFPdim macrophages towards the injection site (S8 Video), similar to what is seen in response to flea bites (S1–S5 Video). However, in contrast to what was observed after flea-transmission (S7 Video), in four independent experiments with needle-inoculated Y. pestis we did not observe movement of bacteria in association with eGFPdim cells, suggesting that flea-transmitted Y. pestis may be more likely than needle-inoculated bacteria to interact with macrophages in vivo.
Dendritic cells are antigen presenting cells that reside in peripheral tissues and migrate into the lymphatics after contact with pathogens [19]. To determine whether or not DC interact with Y. pestis after flea-borne transmission, we used a transgenic mouse expressing yellow fluorescent protein (YFP) under control of the itgax promoter [20]. Itgax encodes a component of CD11c, a molecule widely used to identify DCs.
The bite sites of uninfected fleas, blocked fleas that did not transmit bacteria, and blocked fleas that had deposited Y. pestis in the dermis were imaged for at least 4 hours post-feeding (Fig. 6A). In response to uninfected flea bites DCs appear to randomly move through the dermis (Fig. 6A, S9 Video), similar to what is observed in a naïve mouse ear (Fig. 6A, S10 Video). Consistent with what was observed after needle inoculation of these mice [16], we did not observe any notable interaction between DCs and flea-transmitted bacteria (Fig. 6A, S11 Video). Interestingly, while there was no net influx of a large number of DCs like that seen with neutrophils, the cells that were present appeared to migrate towards the flea bite sites that contained Y. pestis. A similar phenomenon was seen at some blocked flea bites where no bacteria were deposited in the dermis, but was much more variable (Fig. 6A, S12 Video). To quantify this cellular movement, image analysis software was used to track the migration of these cells over the course of the experiment. Cell tracking and displacement are shown in the bottom panels of Fig. 6A. The direction of displacement of each cell track was scored as being “toward”, “away” from or “neutral” relative to the flea bite site (Fig. 6B). Additionally, the average number of cell tracks with displacement >30 μm was determined for each experiment (Fig. 6C). We conclude that DCs migrate towards flea-transmitted Y. pestis or blocked flea bites, but not to uninfected flea bites in the dermis and that there was overall more displacement of DCs when bacteria were present at the bite site.
Fleas are considered to be primarily capillary feeders; they probe the skin with their mouthparts until a blood vessel is cannulated and a blood meal is siphoned directly from the vessel [21, 22]. Blocked fleas can deposit Y. pestis in the extravascular dermal tissue or, less frequently, directly into the lumen of a blood vessel [2]. To determine the numbers of bacteria transmitted by fleas in our experiments and the tissue localization of flea-transmitted Y. pestis early after infection, we collected ear dermis, dLN and spleen tissue samples from mice after completion of the intravital microscopy experiments described above (approximately 5 h after termination of flea feeding). Tissues were triturated and plated to determine the number of colony forming units (CFU) present. The results of each independent experiment consisting of an individual mouse are shown in S1 Table. The number of blocked fleas that fed on each mouse varied from one to seven. We recovered no CFUs from 22% of the mice tested despite many of these mice being fed upon by as many as four blocked fleas. Among mice that had detectable bacteria in the dermis after flea exposure, the number of dermal CFUs ranged from 5 to 3660 with a median of 237.5 CFUs.
The number of CFUs cultured from the spleen served as an indicator of transmission of bacteria directly into the bloodstream. Among the 28 mice that had detectable CFUs in any of the tissues tested, 23 (82%) had anywhere from 1 to 4000 CFUs/spleen. Two mice had bacteria in their spleens, but no bacteria were detected in their dermis or dLN, indicating that fleas had deposited bacteria directly into the lumen of blood vessel during the feeding attempt. Despite harvesting the tissues at the early time point of ~5 h post-feeding and the use of a highly attenuated strain of Y. pestis, we found a surprisingly high number of bacteria present in the dLN. The numbers ranged from 15 to 1000 CFUs with a median of 270 CFUs/LN. Thus, dissemination of bacteria from the dermis to draining lymph node can occur within 5 h of flea feeding.
The above experiments were done in conjunction with the intravital microscopy studies, thus the mice were exposed to variable numbers of blocked fleas, received variable numbers of flea bites, and were assayed ~5 h after flea feeding. To quantify transmission by individual fleas, we performed experiments where mice were exposed to 1 to 3 blocked fleas placed on one or both ears in an effort to consistently obtain mice that had been fed on by only 1 blocked flea. Mice were euthanized 1 h after termination of flea exposure and their ear, dLN and spleen tissues assayed for CFU. Each time a mouse ear had been fed on by at least one blocked flea it was recorded as a feeding event. In total, we exposed a total of 25 mice and recorded 31 feeding events. Of these events, 14 (45.2%) resulted in transmission of bacteria into at least one of the tissues assayed and these are depicted in Fig. 7. Nine (29%) of the feeding events resulted in deposition of Y. pestis into the dermis. We detected bacteria in the spleen of 9 (29%) mice 1 h after removal of fleas, suggesting that bacteria were introduced directly into the bloodstream during flea feeding. Bacteria were detected in the dLN after 6 (19.4%) individual feeding events. The presence of bacteria in the dLN at this early time point indicates that some bacteria disseminate to the dLN within 2 h after introduction into the dermis.
The degree to which flea transmission influences the pathogenesis of bubonic plague or the innate immune response to infection is unknown. Here we characterize the very early neutrophil, macrophage and dendritic cell recruitment to flea bites and flea-transmitted Y. pestis. We developed a method for reliably and accurately identifying flea bite sites in mice using the DNA stain Sytox Blue. Using this method we imaged the host cellular response to uninfected flea bites and found minimal recruitment of neutrophils to the bite site. This was surprising in light of previous work showing a rapid neutrophil response to tissue damage [18, 23]. Specifically, a study by Peters et al. showed robust neutrophil recruitment to uninfected sand fly bites on the same LysM-eGFP mouse strain used in our study [18]. Sand flies are “pool feeders” in that they feed by wounding the dermal microvasculature with serrated mouthparts and siphoning blood from a hemorrhagic pool formed within the wound. In contrast, fleas are considered “capillary feeders” that use their small mouthparts to cannulate a dermal blood vessel, with apparently little damage to the cells at the bite site. This may result in less inflammation and neutrophil recruitment than is seen at sand fly bite sites. Additionally, flea saliva contains a variety of components homologous or analogous to salivary proteins in other blood feeding arthropods that are known to be anti-inflammatory [17]. Both of these factors may be responsible for the low numbers of neutrophils recruited to uninfected flea bites.
Interestingly, we observed the mobilization and migration of eGFPdim cells towards the flea bite site. In the LysM-eGFP transgenic mouse used in this study, these eGFPdim cells in the dermis are largely F4/80+ tissue resident macrophages [18]. We did not observe movement of these cells towards needle inoculation sites in our previous study [16], but the large numbers of eGFPbright neutrophils recruited to tissue damage done by the needle made it difficult to see the dim macrophages.
When we examined blocked flea bites in LysM-eGFP mice where bacteria had been deposited at the bite site, we observed more neutrophil recruitment relative to uninfected flea bites and the neutrophil numbers appeared to correlate with the amount of bacteria at the site. This suggests that the neutrophils were recruited to the bacteria and not the bite itself. It also shows that any suppressive effect that flea saliva may have on neutrophil recruitment could not override the response to bacteria in the dermis. Despite the presence of a large number of neutrophils, we observed very little movement of Y. pestis at the bite site. This is in contrast to our previous study showing many injected bacteria being trafficked away from the injection site in association with eGFPbright neutrophils [16].
Y. pestis isolated from flea midguts are more resistant to phagocytosis by macrophages and neutrophils than broth-cultured bacteria due to upregulation of a family of insecticidal-like toxin complex proteins in the flea [14, 24]. The two-component regulatory system PhoP-PhoQ, important for the resistance of Y. pestis to stressors experienced in the mammalian host such as low pH, osmotic or oxidative stress, or antimicrobial peptides is also upregulated in the flea relative to in vitro broth-cultured bacteria [14, 25]. Additionally, Y. pestis forms a biofilm in the flea midgut as result of increased production of a polysaccharide extracellular matrix (ECM) in this environment. The effects of this ECM on mammalian host response are unknown, but structurally similar ECM produced by Staphylococci protects against innate immune effectors [26]. Thus, the phenotype of flea-derived Y. pestis differs considerably from broth-grown bacteria in ways that may influence pathogenesis and innate host response. Further work will be needed to evaluate the interactions of flea-grown Y. pestis with innate immune cells in vivo.
Interestingly, when bacterial movement at the flea bite site was observed, many of these bacteria were associated with eGFPdim macrophages. In each case the macrophages did not transport bacteria completely away from the injection site, but remained in the field of view for the duration of the experiment (Fig. 5, S7 Video). Again, this is in contrast to experiments with needle-inoculated bacteria where most of the Y. pestis movement was in association with neutrophils that transported bacteria completely out of the field of view [16]. However, the large number of eGFPbright neutrophils present may have obscured the rare eGFPdim events in these experiments. To address this, we needle-inoculated PMN-depleted LysM-eGFP mice with Y. pestis expressing dsRed and imaged them by confocal. While eGFPdim macrophages were recruited to the injection site, we observed minimal movement of bacteria in association with these cells. These results suggest that flea-transmitted bacteria may preferentially interact with macrophages over neutrophils. This would have implications for Y. pestis pathogenesis, as macrophages are much more permissive for Y. pestis survival and growth than neutrophils [7].
Imaging of blocked flea bites in CD11c-YFP mice revealed minimal interactions between YFP+ cells and flea-transmitted Y. pestis at the bite site. It is important to note that YFP expression in these mice is not limited exclusively to DCs, nor does every subset of DC in the dermis express YFP. It is more accurate to classify the YFP+ dermal cells in our experiments as antigen-presenting mononuclear phagocytes [27]; however, for simplicity, we refer to these YFP+ cells as DCs. While we did not observe a massive influx of DCs, they did appear to mobilize and migrate specifically towards blocked flea bites containing bacteria. The consequences of this migration of DCs towards flea-transmitted bacteria are unknown. DCs do not appear to associate with bacteria at the bite site in the timeframe studied, but it remains possible that they would show more association with bacteria later after infection. These results are consistent with our previous study showing minimal interaction between needle inoculated Y. pestis and DC early after infection [16].
Uninfected flea bites did not recruit DCs and migration towards blocked flea bites where bacteria were not detected was variable. It is plausible that some bacterial components, such as lipopolysaccharide, could have been introduced into the bite site by blocked fleas even if no whole bacteria were transmitted. These bacterial components could act as pathogen associated molecular patterns (PAMPs) that directly or indirectly stimulate recruitment of innate immune cells [28]. It is also possible that a very small number of bacteria might have been transmitted, but were undetectable by microscopy. These factors may explain the variability in cellular response we observed. Additionally, blocked fleas are unable to take a blood meal and tend to probe the skin in repeated unsuccessful feeding attempts. The additional tissue damage from this probing could explain the increased cellular recruitment to blocked compared to uninfected flea bites, although the amount of Sytox Blue staining at the bite site did not appear to correlate with neutrophil recruitment.
The CFU assays of the dermis, dLN, and spleen early after flea feeding yielded highly variable results, consistent with previous studies on the regurgitative transmission mechanism of X. cheopis [6]. It was not uncommon to find several hundred or even thousands of bacteria in the spleen after flea feeding. This is most likely due to regurgitation of bacteria directly into the bloodstream during the blocked flea’s attempt to feed as has been previously described [2]. Interestingly, several animals had hundreds or more CFU in the dLN at ~5 h post-feeding. This prompted us to look 1 h post-feeding where we found some animals with a small number of CFU in the dLN. Overall, these data suggest that very rapid dissemination to the spleen and dLN is a common occurrence after flea transmission of Y. pestis. The data also suggest that a small number of flea-transmitted bacteria may move so rapidly into the lymphatics that they bypass any significant interactions with phagocytes at the bite site. The ultimate fate of these early LN disseminators is unknown.
Despite the historical significance of Y. pestis and the importance of fleas in the plague transmission cycle, the early events in the skin after deposition of bacteria via blocked flea bite are poorly understood. Here we have characterized the innate cellular recruitment to uninfected and infected flea bites in vivo. We also gathered quantitative data on the numbers and tissue distribution of Y. pestis transmitted by fleas. Our results show a much greater neutrophil response to flea-transmitted Y. pestis than to uninfected flea bites. We also observed migration of resident tissue macrophages towards uninfected and blocked flea bite sites and their association with flea-transmitted Y. pestis. Similar migration of dendritic cells towards infected, but not uninfected, flea bites was observed. Interestingly, we found Y. pestis in the dLN by 1 h after flea exposure, suggesting that initial dissemination of bacteria to the LN occurs more quickly than was previously appreciated [29, 30]. In support of this, Gonzalez et al. recently reported that needle-injected Y. pestis could be found in the dLNs of some mice as early as 10 min post-infection [31]. Future work will be aimed at determining the fate of these early disseminators and their importance in bubonic plague pathogenesis.
Xenopsylla cheopis fleas were infected with Y. pestis pMcherry (strain KIM6+ [virulence plasmid negative, pigmentation locus positive] transformed with a pMcherry plasmid [Clontech]) using a previously described artificial feeding system [5]. Fleas were monitored for blockage by microscopic examination for up to six weeks post-infection. Blockage was diagnosed by the presence of fresh blood in the flea esophagus, but not the midgut, immediately after feeding.
C57BL/6J LysM-eGFP knock-in mice were originally created by T. Graf [32] (Albert Einstein University, Bronx, NY) and were bred by Taconic Laboratories under a contract with NIAID. C57BL/6J (stock number 000664) and CD11c-YFP (stock number 008829, originally described Lindquist et al. [20]) mice were purchased from The Jackson Laboratory (Bar Harbor, ME). Ten- To 20-week-old female mice were used in all experiments. All mice were maintained at the Rocky Mountain Laboratories animal care facility under specific-pathogen-free conditions.
For experiments involving PMN-depletion, mice were injected i.p. with 250 μg anti-GR1 antibody (clone RB6–8C5, BioXCell, West Lebanon, NH) 24 h and 4 h prior to infection with ~1000 CFU of dsRed-expressing Y. pestis, as described in [16]. This treatment results in >95% depletion of Ly6G+ neutrophils. The Y. pestis strain expressing dsRed was used instead of the mCherry-expressing strain in this experiment to be consistent with a previous study of the response to needle-inoculated Y. pestis, and due to a higher level of fluorescent protein expression in broth culture.
Mice were anesthetized by subcutaneous injection of a ketamine/xylazine mixture and secured on a heating pad to maintain body temperature. Where indicated, mice were injected with 250 μM Sytox Blue (Life Technologies) in 150 μL PBS i.p. and, in some cases, 60 μl of Qtracker655 (2 μM stock, Life Technologies) in 150 μl PBS i.v. 10 to15 min prior to flea exposure. Fleas were immobilized by incubation on ice and placed in a custom-made feeding chamber consisting of a 200 μl PCR tube and a foam padded plastic clamp (S1 Fig). This chamber was then clamped onto the ear of the mouse and the fleas allowed to warm to room temperature. The fleas were in contact with the mouse for 10 min for uninfected fleas or 50 min for blocked fleas. The mouse and feeding chamber were then placed in a jar containing isoflurane for approximately 30 sec to anesthetize the fleas. The chamber was then removed from the ear and the fleas collected and microscopically examined to determine if fresh blood was present in their digestive tract indicating feeding. In some cases, a model SMZ1500 dissecting microscope (Nikon, Tokyo, Japan) equipped with a model DP72 color camera (Olympus, Center Valley, PA) was used to capture images of mouse ears before and after exposure to fleas.
The ears of LysM-eGFP or CD11c-YFP mice were imaged by confocal microscopy as previously described [16]. Briefly, mice were anesthetized with an isoflurane-O2 mixture provide by nose cone and their ears mounted to a coverslip on the stage of a Zeiss LSM 510 Meta confocal microscope (Zeiss, Oberkochen, Germany) equipped with an incubated chamber set to 30°C. Z stacks were acquired with a 20x objective at 2 min intervals for the indicated duration and the images obtained were processed using Imaris 6.3.1 software (Bitplane, South Windsor, CT). All supplemental video files are shown at the same magnification with the exception of S7 Video which has been digitally zoomed using the Imaris software.
Neutrophil recruitment was scored by assessment of total neutrophil accumulation observed over the duration of videos of Lys-eGFP mice fed upon by uninfected or blocked fleas. Each video was scored by 3 lab members on a scale from 0 to 4 in whole number increments, with 0 representing essentially no net recruitment of neutrophils and 4 representing massive accumulation of neutrophils forming a large aggregate at the bite site. The results are shown as the mean and standard error of the mean (SEM).
Tracking of YFP+ cells in time series of CD11c-YFP mice was accomplished using the tracking function within the Imaris 6.3.1 software package. Once the cellular movement had been tracked, we used the software to determine the direction of net displacement of each cell. Limiting further analysis to YFP+ cells with a net displacement of at least 30 μm over the course of the experiment, we scored each cell displacement event as being towards (displacement within a 45° angle in the direction of the bite site), away (displacement within a 45° angle in the opposite direction of the bite site), or neutral (all remaining displacement events) relative to the flea bite site. For mice that had not been fed on by fleas, a spot at the center of field of view was arbitrarily chosen to represent the flea bite site.
Ear, draining cervical lymph node and spleen tissues were collected from mice after euthanasia. Ears were separated with forceps into ventral and dorsal halves. Tissues were placed in Lysing Matrix H bead tubes (MP Biomedicals) containing 500 μl of PBS and disrupted with a Fastprep 120 (Thermo Savant). The numbers of Y. pestis pMcherry CFU in the tissue samples were determined by dilution and plating on blood agar plates containing 100 μg/ml carbenicillin.
All animal studies were performed under protocols adhering to guidelines established by the Public Health Service Policy on Humane Care and Use of Laboratory Animals. The protocols were reviewed and approved by the Rocky Mountain Laboratories Animal Care and Use Committee (AALAS unit number 000462, PHS-OLAW number A-4149–01).
For experiments measuring neutrophil recruitment scores, data were analyzed using a Kruskal-Wallis nonparametric test followed by a Dunn’s multiple comparison test. For experiments determining the direction of DC migration, data were analyzed using two-way ANOVA with Tukey’s multiple comparison post-test. For experiments measuring total DC displacement, data were analyzed using one-way ANOVA with Holm-Sidak’s multiple comparisons post-test.
|
10.1371/journal.pcbi.1005931 | Predicting pathogenicity behavior in Escherichia coli population through a state dependent model and TRS profiling | The Binary State Speciation and Extinction (BiSSE) model is a branching process based model that allows the diversification rates to be controlled by a binary trait. We develop a general approach, based on the BiSSE model, for predicting pathogenicity in bacterial populations from microsatellites profiling data. A comprehensive approach for predicting pathogenicity in E. coli populations is proposed using the state-dependent branching process model combined with microsatellites TRS-PCR profiling. Additionally, we have evaluated the possibility of using the BiSSE model for estimating parameters from genetic data. We analyzed a real dataset (from 251 E. coli strains) and confirmed previous biological observations demonstrating a prevalence of some virulence traits in specific bacterial sub-groups. The method may be used to predict pathogenicity of other bacterial taxa.
| An important challenge in Computational Biology is the analysis of genetic molecular data through sophisticated computer science and mathematical methods that are implemented by interdisciplinary research groups. The resulting comprehensive approach, based on the BiSSE model and microsatellites profiling (TRS-PCR), can be used to predict pathogenicity behavior in bacterial taxa. As proof of concept, we applied the procedure to real clinical data sets of genetic information obtained from a unique collection of bacterial populations (251 strains). Our results showed that a state-dependent model was able to predict pathogenicity behavior of E. coli population. Furthermore, we confirmed previous biological observations indicating a prevalence of some virulence genetic traits in bacteria.
| The diverse species of E. coli display a large repertoire of genetic traits—pathogenicity factors, allowing the colonization of the human host. Depending on the occupied niche, individual virulence factors (VFs) are favored, allowing the survival of the pathogen. However, pathogenicity factors are maintained only when they favor the development of the pathogen, during the colonization of the host [1]. Otherwise they are eliminated when not required, as they are costly traits, or when their presence (expression) promotes detection by the host’s immune system [2,3].
Microsatellites, stretches of DNA consisting of repeated short segments of nucleotides (sequence motifs) are commonly found in bacterial genomes. A special class of microsatellites, that of trinucleotide repeat sequences (TRS motifs), is genetically unstable and this instability depends mainly on the length and number of copies of the repeated motif [4,5]. In the case of bacterial genomes a TRS rarely exceeds 10 copies and is therefore relatively stably transmitted in subsequent generations. The number of such loci (with the number of repetitions n ≥ 3) varies depending on the species of the microorganism and is, for example 1667 for S. aureus JH1, 2568 for E. coli CFT073 and 4201 in the case of M. tuberculosis. Amplification of DNA regions located between the TRS motifs allows one to obtain band patterns specific to the genus, species or a bacterial strain [6–8]. In the case of E. coli, CGG- and / or GTG-PCR patterns are correlated with their phylogenetic membership and also group strains having similar sets of VFs [9,10]. Therefore, the question is whether the observed phenomenon of clustering is only a reflection of the genetic status quo, or can it also be helpful in predicting directions of pathogenicity development in the E. coli population. Such a hypothesis was verified by employing the binary-state speciation and extinction model—BiSSE [11] with an appropriate probabilistic interpretation (see S1 Appendix).
BiSSE is a theoretical model, which was introduced into the phylogenetic community by Maddison et al. [11]. Apart from special subcases, see e.g. [12], the likelihood is not analytically tractable but can be obtained numerically by solving an ODE system (as in the diversitree R package [13–15]). Since its introduction a number of generalizations have been implemented such as quantitative state speciation and extinction (QuaSSE, [14]) where the speciation and extinction rates depend on an arbitrary (even continuous) suite of traits, or Hidden State Speciation and Extinction model (HiSSE, [16]). Even though the likelihood function is not analytically tractable one can deduce large sample properties of the model by studying branching processes on generalized state spaces. In particular Janson [17] provides results characterizing the limit behavior (almost sure convergence and central limit theorems).
In this paper we apply the BiSSE model to estimate parameters from a collection of 251 strains from the clinical isolates of E. coli. We present an application of microsatellites, specifically TRS microsatellites, as pathogenicity markers and we analyze E. coli strains using the BiSSE model.
A collection of 128 clinical E. coli strains (set U) was gathered between June 2005 and September 2006 from the urine of patients in various wards of the Military Teaching Hospital No. 2, Medical University of Lodz, Poland. The second collection (set K) composed of 123 isolated from children with diarrhea in the Lodz region (Poland) and were obtained from the Medical Laboratory SYNEVO in Lodz, Poland. Isolates were collected from January 2009 to May 2010. Genomic DNA isolation and purification was performed with the use of a GenElute Bacterial Genomic DNA Kit (Sigma-Aldrich, St. Louis, MO). The quantity and purity of each genomic DNA sample was determined spectrophotometrically at 260 nm (BioPhotometer, Eppendorf, Germany). The DNA samples were diluted to 20 ng/μl and then used. The possession of virulence genes, typical for uropathogenic (UPEC) and intestinal E. coli (IPEC) was determined by multiplex-PCR, according to procedures described elsewhere [9,10,18–22]. Detailed characteristics of the collection of strains are presented in Table 1 and Table 2.
A collection of 251 genomic DNA samples were isolated from E. coli strains and TRS-PCR profiling using GTG and CGG primers was performed. Two TRS-PCR reactions were performed for each strain using primers containing GTG and CGG repeats respectively, according to procedures described elsewhere [9,10]. The PCR products, 8 μl of 50μl, were resolved by electrophoresis on 1.6% agarose gels (15×15 cm, 4 mm thick) in 1×Tris-acetate-EDTA (TAE) buffer, 2.5 V cm-1, until the dye (bromophenol blue) migrated 6 cm from the top of the gel. Such stringent conditions for the electrophoretic separations allow for carrying out trustworthy analyses. The DNA products for all of the primers ranged from 0.1 kbp to 2.5 kbp. The gels were stained with ethidium bromide (1 μg ml-1), visualized on a UV-transilluminator, and photographed (Fc8800, Alphainnotech). Subsequently, gels were optimized according to recommendations provided by BioNumerics version 5.00 software (Applied Maths, Belgium) and normalized with regard to a 100 bp Plus DNA size marker (Fermentas, Thermo Scientific Waltham, MA, USA). The CGG-PCR and GTG-PCR band profiles for each strain from the collection were obtained and respective dendrograms were constructed using the BioNumerics software (Pearson correlation, optimization 1%, position tolerance 1%). Finally, the average similarity Neighbor Joining dendrogram based on the two trees was assembled. The results are shown in Fig 1. Such dendrogram and virulence information were subsequently analyzed by our wrapper, around make.bisse() and find.mle() functions, R script.
In this study we used the BiSSE model [11] for binary states with four rate parameters. BiSSE models (Fig 2) the evolution of a binary trait (two possible states 0 and 1) and allows for estimation of the speciation (λ0, λ1), extinction (we assumed μ0 = μ1 = 0), and transition between states (q01, q10) rates. Knowledge of these rates sheds light on whether the trait controls diversification rates or not. In our case the trait levels correspond to non-pathogenic (0) and pathogenic (1). The transition rate from 0 to 1 is q01 and from 1 to 0 is q10. If the species is in state 0, then it has speciation rate λ0 and in state 1 speciation rate λ1.
Notice that in our setting we do not assume any extinction events, i.e. the extinction rates are set to 0, while in the general BiSSE model they can be non-zero. We may concisely describe the model as follows. Let N0(t) be the number of 0 strains at time t and N1(t) the number of 1 strains. Of course N(t) = N0(t)+N1(t) is the total number of strains present in the system at time t. We assume that at time 0, at the root of the tree there is one strain alive, N(0) = 1. We will estimate the root state, i.e. whether our data supports N0(0) = 1 or N1(0) = 1.
Immediately with the introduction of the BiSSE model there was concern about its power, i.e. its ability to distinguish between competing hypotheses of symmetric versus asymmetric models (given pairs of parameters equal versus not equal) [23]. Simulation studies indicated that a minimal sample size should be about 300 [23]. However, these investigations were done under the full six parameter BiSSE model. Later investigations (e.g. [24,25]) indicate that some questions can be analyzed based on much smaller samples. If some parameters are set to 0 then the power can increase dramatically and give sensible results with 100 species [24]. Asymmetric speciation rates can be detected with as few as 45 contemporary tip species [25,26]. In our setting the extinction rates are fixed at 0. Since these parameters are the most difficult to estimate [24,25], the consideration of a restricted sub-model should improve the situation. Quoting [24] p. 391, "… there are also many reasons for guarded optimism." especially as, quoting [25] "… low power should tend to reduce our ability to detect differences between parameters, rather than exacerbate them".
We wrote a wrapper script around the make.bisse() and find.mle() functions of the diversitree R package [13,14] that does model selection and then calculates the limit behavior of the model. We demonstrate the application of the BiSSE model to estimate parameters from genetic traits (see scheme of research hypothesis) and to illustrate this approach we estimate parameters from a collection of clinical E. coli strains. We used the diversitree R package to estimate four parameters (λ0, λ1, q01, q10) from the dendrograms. We considered various models: (λ0, λ1, q01, q10), (λ0, λ1, q01 = q10) and (λ0 = λ1, q01 = q10). This particular functionality is actually available through the diversitree::constrain() function. However, our wrapper function is more general and allows the user to specify an arbitrary parametrization of BiSSE's parameters. In particular we do not have the restrictions "Terms that appear on the right hand side of an expression may not be constrained in another expression, and no term may be constrained twice." (from diversitree::constrain()’s help). Our wrapper function should be useful to researchers as it seems that biological studies can require restricted BiSSE setups (e.g. [24,25]).
Model selection was done using AICc [27]. Assessment of model fit was done by comparing the observed fractions of pathogenic strains to the composite parameter P1 (see Section Probability of maintaining the VF in E. coli strains) in Table 3. Furthermore, in Table 3 we can see that the Taylor expansion approximation (see Section Probability of maintaining the VF in E. coli strains) of P1 corresponds well to the theoretical and observed proportions. The estimation of the four parameters was based on the provided phylogeny and observed states. From the estimated parameters we extracted, using an R script, the almost sure limits of the proportion of the VF in E. coli strains (see S1 Appendix).
All calculations were done in R on the multicore computational server of the Department of Mathematics Uppsala University (R 3.2.5 for Ubuntu 12.04.5 LTS on a 1.4GHz. AMD Opteron Proc. 6274). We ran the computation on 4 cores and the whole analysis took about 3 days.
Source code and sample data freely available for download at https://github.com/BISSE-TRS/ppbEcoli, distributed under the GNU GPLv3 license.
In this work, we ask whether, with the disposal of dendrograms based on TRS profiles and the BiSSE method, it is possible to predict the maintenance of particular VF features in a population. A diagram summarizing our work is presented in research hypothesis, Fig 3.
In our study we take the viewpoint that strains are genetically variable but do not go extinct in a population. Extinction is a principle of evolution, but this phenomenon is attributed to species. In our case we do not have classical extinction of species present. Rather, we observe that with time the bacterial genetic pool of strains becomes more diverse. Hence, we focus on the no extinction model (μ0 = μ1 = 0) [28,29]. Even though BiSSE is known to have low power for samples less than 300, Maliska et al. [25] indicated the asymmetries in speciation rates can be detected with as few as 45 species. Hence, their estimates are a primary focus of our understanding of VF dynamics in E. coli. Here, we demonstrated differences in rates of speciation depending on the absence (λ0) or presence (λ1) of the given trait of virulence. Fig 4A shows results obtained for intestinal E. coli strains and Fig 4B shows results for strains isolated from urine. In the case of strains isolated from stool samples a higher rate of propagation can be observed for those not possessing cnf1, hly1, papC, sfa, tsh and usp genes. It is not surprising given that such virulence factors (except tsh) typically occur in uropathogens [1,30–32]. On the other hand, possession of iron and iutA resulted in much higher rate of propagation. In the case of strains isolated from urine most of the virulence factors had a stimulating effect on dissemination of strains except for astA and tsh genes. One could expect this, as urine is not naturally inhabited by microorganisms and therefore, numerous virulence factors facilitate colonization.
Here we studied rates of mutation in pathogenic (q01) and non-pathogenic (q10) directions for strains isolated from stool (Fig 5A) and urine (Fig 5B). Interestingly, in both cases when differences were pronounced the q10 transition was preferred. This is consistent with the fact that maintenance of a VF is energetically costly for microorganisms and additionally, lack of the virulence factor allows for “hiding” from the host’s immunological defense system. Furthermore, highly virulent strains may sensitize individuals allowing for recurrent infections caused by these less virulent strains [1,33].
Based on the estimated BiSSE rates it was possible to estimate the long term proportions (P1: = v1/(v0+v1), see S1 Appendix, Thm. 2.2) of the VF features in the populations. The results are shown in Fig 6. Among analyzed VF features the following traits had a higher than 50% chance for being maintained in an E. coli population–cnf1, fimG, fyuA, hly1, iroN, iutA, sat, sfa and usp. The vast majority of these traits exhibited pathogenicity maintenance in the strains isolated from urine. This seems to be justified by the fact that the VFs mentioned above are necessary for the colonization of the urinary tract in humans i.e. adhesins (fimG, sfa), toxins (hly1, cnf1, sat), iron uptake system (fyuA, iutA, iroN) and bacteriocin (usp) [1,31,32,34]. These VFs, however, are not necessary for the development of intestinal pathogens. If the non-pathogenic strains speciate faster than the pathogenic ones (i.e. λ0>λ 1), then a Taylor expansion of P1 points to a very simple formula for it: q01/λ 0 (provided this is lesser than 1). In Table 3 we also included this simplified calculation.
In this paper we presented a comprehensive approach for predicting pathogenicity in a population based on a state dependent model and TRS-PCR profiling. Additionally, this paper shows that it is possible to apply this approach to real laboratory genetic data–from 251 E. coli strains. Our first research goal was to infer dendrograms for the E. coli population. This required the gathering of a unique collection of bacterial populations (251 strains) and a detailed laboratory genetic analysis, including CGG- and GTG-PCR profiling as well as the identification of pathogenic traits. Next, we applied the BiSSE model to such a collection of genetic data. Any BiSSE analysis of biological data runs the risk of low power and one should be careful with drawing conclusions. However, in our case there is place for "guarded optimism” as we restrict our model by excluding the two parameters most difficult to estimate (the extinction rates). We used AICc to distinguish between competing models and remembering that "… low power should tend to reduce our ability to detect differences between parameters, rather than exacerbate them" [25] we notice that most VFs have equal transition rates. The exceptions to this are hly1 (in U), iroN (in K) iutA (in U), papC (both U and K), sat (in U), sfa (in U), usp (in U), cnf1 (in U). In all of these cases we have q10>q01, i.e. the loss of pathogenicity is favoured. Furthermore we can see that asymmetry (loss of pathogenicity) is preferred in the urine environment. Such a behavior was previously observed by others [1,33].
Our computational results confirmed previous biological observations demonstrating a prevalence of some virulence traits in specific bacterial sub-groups [21,30]. The necessity of harboring some VFs in E. coli pathogens was indicated. For example, UPEC strains exist within the intestinal tract of humans but possess specific factors (adhesins, toxins, siderophores and bacteriocins) that permit their successful transition from the intestines to the urine tract. These VFs are encoded by genes located at the selected regions of chromosomal DNA, plasmids and/or transposons, named pathogenicity islands (PAIs). PAIs are flexible genetic elements, holding the mobility sequences, which are transferred horizontally between the bacterial cells [2,3]. This phenomenon is significant for bacterial population evolution/diversity. Additionally, it allows for VFs’ synergy during the process of pathogenicity. For example, iroN and sfa are located on PAI III in E. coli 536 and hly and cnf1 are encoded by PAI II in E. coli J96. It may suggest that these VF pairs will be co-transmitted. However, in our study only 35,8% of strains harboring iroN encodes also sfa and 86,6% of strains encoded for both hly and cnf1. In the latter case however, we need to keep in mind that alpha-haemolysin gene cluster is present also on plasmids and the other PAIs, some of which do not encode the CNF-1 [35]. In addition, one needs to remember that not always the PAIs are transmitted completely and due to recombination errors, some sets of features may not be lost or acquired jointly [28].
As mentioned above, our research has been conducted using 251 E. coli strains that included two collections–from urine and stool samples. These collections were not equal in terms of their virulence factors repertoires (Table 1) therefore, it would be interesting to extend this research to more strains harboring numerous intestinal VFs. Since the population studied was divided into two collections isolated from two different environments one could also consider the GeoSSE model to capture potential differences between the urine and stool environments. However, on the one hand the sample size of 251 is probably too small for such a complex model (10 parameters, even with extinction set to 0). On the other hand the BiSSE model is a submodel of GeoSSE. GeoSSE has separate BiSSE models in each environment and then transition rates between the environments. Hence, as we analyze the two environments separately ignoring the transitions, the use of GeoSSE would probably result in noticing finer details in the data, i.e. studying it with a tool that has a higher resolution-the interaction between the environments. BiSSE allows us to observe more general properties inside each environment. These are already consistent with biological intuition.
Additionally, this paper presents a method of estimating the probability of persistence of the VF in E. coli strains. Noteworthy, this is a comprehensive approach and it may be used to predict pathogenicity of other bacterial taxa. We believe that our developed software should be useful for biologists that want to use restricted BiSSE models or who want to parametrize the parameters. Our wrapper function seems to be flexible enough for such purposes.
The binary state dependent model and TRS-profiling appear to be useful tools for predicting persistence of pathogenicity in an E. coli population.
|
10.1371/journal.pntd.0000466 | Prevalence of Buruli Ulcer in Akonolinga Health District, Cameroon: Results of a Cross Sectional Survey | Buruli ulcer (BU) is a chronic, indolent necrotizing disease of the skin and underlying tissues caused by Mycobacterium ulcerans, which may result in functional incapacity. In 2002, Médecins Sans Frontières (MSF) opened a BU programme in Akonolinga Hospital, Cameroon, offering antibiotic treatment, surgery and general medical care. Six hundred patients have been treated in the project to date. However, due to the nature of the disease and its stigmatization, determining the exact prevalence and burden of disease is difficult and current estimates may not reflect the magnitude of the problem. The objectives of this survey were to estimate the prevalence of BU in the health district of Akonolinga, describe the geographic extension of the highly endemic area within the health district, and determine the programme coverage and its geographical distribution.
We conducted a cross-sectional population survey using centric systematic area sampling (CSAS). A 15×15 km grid (quadrats of 225 km2) was overlaid on a map of Akonolinga district with its position chosen to maximize the area covered by the survey. Quadrats were selected if more than 50% of the quadrat was inside of the health district. The chiefdom located closest to the centre of each quadrat was selected and Buruli cases were identified using an active case finding strategy (the sensitivity of the strategy was estimated by capture-recapture). WHO-case definitions were used for nodules, plaque, ulcer, oedema and sequelae. Out of a total population of 103,000 inhabitants, 26,679 were surveyed within the twenty quadrats. Sensitivity of the case finding strategy was estimated to be 84% (95%CI 54–97%). The overall prevalence was 0.47% (n = 105) for all cases including sequelae and 0.25% (n = 56) for active stages of the disease. Five quadrats had a high prevalence of >0.6% to 0.9%, 5 a prevalence >0.3% to 0.6% and 10 quadrats <0.3%. The quadrats with the high prevalence were situated along the rivers Nyong and Mfoumou. Overall coverage of the project was 18% (12–27%) for all cases and 16% (9–18%) for active cases, but was limited to the quadrats neighbouring Akonolinga Hospital.
Prevalence was highest in the area neighbouring the Nyong River. Coverage was limited to the area close to the hospital and efforts have to be made to increase access to care in the high prevalence areas. Use of the CSAS method was particularly useful for project planning and to identify priority areas of intervention. An added benefit of the method is that the survey procedure incorporated an awareness campaign, providing information about the disease and treatment to the population.
| As long as there is no strategy to prevent Buruli ulcer, the early detection and treatment of cases remains the most promising control strategy. Buruli ulcer is most common in remote rural areas where people have little contact with health structures. Information on the number of existing cases in the population and where they go to seek treatment is important for project planning and evaluation. Health structure based surveillance systems cannot provide this information, and previous prevalence surveys did not provide information on spatial distribution and coverage. We did a survey using centric systematic area sampling in a Health District in Cameroon to estimate prevalence and project coverage. We found the method was easy to use and very useful for project planning. It identified priority areas with relatively high prevalence and low coverage and provided an estimate of the number of existing cases in the population of the health district. The active case finding component of the method used served as an awareness campaign and was an integrated part of the project, creating a network of health delegates trained on Buruli ulcer.
| Buruli ulcer (BU) is a neglected tropical disease caused by Mycobacterium ulcerans, belonging to the same family of organisms causing tuberculosis and leprosy. Awareness about the public health importance of the disease was raised in 1998 by the World Health Organisation (WHO) initiative [1].
BU affects predominantly children between 5 and 15 years. The clinical lesions of BU generally start as a painless subcutaneous nodule that secondarily ulcerates, presenting characteristic undermined edges. M. ulcerans produces a toxin, mycolactone, which destroys the skin and the subcutaneous tissues, induces necrosis and ulcerations. Ulcers are chronic, indolent and mainly located on the legs and arms. Some patients develop osteomyelitis and joint lesions. Natural evolution of the disease may lead to spontaneous healing but in the absence of early detection and appropriate treatment, the disease can extend, disseminate and leave functional incapacity [2]. Clinical diagnosis for the ulcerative form is straightforward for trained medical staff, although more difficult for the nodules, plaque and oedematous forms [3].
Based on some observational studies, the WHO recently recommended the use of the combination of Rifampicin/Streptomycin for BU treatment [4]. However, surgery remains important for BU treatment. In the early stages of infection, surgery is curative and highly cost effective, since it requires a simple excision followed by an immediate closure. In the disease's later stages, wide excisions, including healthy tissues, are needed to stop the infection and prevent recurrence or relapse at the same site. This is followed by skin grafting and requires long hospital stays [5]. As long as the mode of transmission is not understood, and in the absence of an effective vaccine, control strategies promoting early detection and treatment have achieved the best results in limiting morbidity and costs associated with the disease [6].
Although BU has been reported in 30 countries in Africa, Asia and the Western Pacific [7], determining the exact prevalence and burden of disease is difficult and current estimates may not reflect the magnitude of the problem. These difficulties include un-diagnosed cases due to fears of stigmatization, little knowledge of the disease among both the population and health workers and the variability in clinical presentation of the disease. Further, BU occurs primarily in remote rural areas where the population may have limited access to health care and the disease is not notifiable in many countries [8]. Prevalence estimates are needed for appropriate resource allocation and to plan control strategies.
In Cameroon, BU cases have been reported in 6 provinces, Adamaoua, Central, South, South-East, East and Extreme North. A national survey identified Akonolinga as a health district of high prevalence [9]. BU endemic areas are located along the Nyong River (Ayos and Akonolinga health districts) with an estimated prevalence of 0.44% in 2001 [10]. Recently, several risk factors were identified including swamp wading, wearing shorts, lower-body clothing while farming, living near cocoa plantation or wood and using adhesive bandages when hurt [11]. In 2002, Médecins sans Frontières (MSF) opened a BU project in Akonolinga District, one of the 135 health districts of Cameroon, in collaboration with the local and national health authorities of Cameroon. The project was set up in Akonolinga Hospital, with a passive case detection strategy. To date, 600 BU patients have been treated in the MSF project, which offers antibiotic treatment, surgery and general medical care. Most patients present late to the Akonolinga Hospital, presenting mainly with ulcerative lesion (about 80%) and advanced stages of BU. A study conducted in 2004 in the district described stigmatisation of BU patients and reported that traditional healers were the first source of treatment [12].
In March 2007 we conducted a cross-sectional survey to: 1) estimate the prevalence of BU in the target population of the project; 2) to estimate the proportion of BU cases visiting the MSF project at least once (coverage); and 3) to estimate the proportion of patients visiting another service provider such as a traditional healer or peripheral health centre at least once (health seeking behaviour). We also aimed to describe the spatial distribution of the prevalence as well as that of health seeking behaviour to help target the most affected areas and to address access problems for certain communities.
Akonolinga health-district is a 1-hour drive from Yaoundé in the department of Nyong and Mfoumou. The health district is at its longest distance approximately 70 km east to west, and 100 km north to south. It has a surface of approximately 4500 km2. The district hospital is in Akonolinga, situated in the geographic centre of the district. The Akonolinga health-district has a total population of 103,000 inhabitants.
We performed a cross-sectional survey using centric systematic area sampling (CSAS). CSAS has been used successfully in past research in malnutrition and other low prevalence diseases [13]. It is particularly well suited to situations where the disease is visible, of low prevalence and where geographic distribution of prevalence and program coverage is of interest.
A 15×15 km grid (quadrats of 225 km2) was overlaid on a map of Akonolinga district with its position chosen to maximize the area covered by the survey. Quadrats were selected if more than 50% of the quadrat was inside of the health district, resulting in 20 quadrats identified.
Our sampling unit was the chèferie (chiefdom), the lowest administrative unit in Cameroon. The chèferie located closest to the centre of each quadrat was selected. Chèferies may be comprised of one to several villages. If the total population of the selected chèferie was below 1000 persons, the next closest chèferie was also included to obtain our required sample size as discussed below. Population information for the selected chèferies was obtained from the chief of each chèferie and crosschecked with the Chief's Office, Department of Nyong and Mfoumou and the Department of Development that compared the figures with the 2005 census. All inhabitants of the chèferies were invited to participate in the study.
To estimate an expected prevalence of all forms of BU of 0.6% with 0.1% precision, our required sample size was18,742 inhabitants. Using the same prevalence estimate, and assuming 50% program coverage with 10% precision, our required sample size was 83 cases, corresponding to a population of 13,900 persons.
We used the WHO case-definition of BU [2], limiting the definition of an active case to nodule, plaque, oedema and ulcer (Table 1). We did not include the WHO papula stages because of the very low specificity of the clinical signs and considering that this clinical form is quite rare in West Africa. Sequelae were defined as having a history of BU and complications resulting directly from the lesion (e.g., restricted limb movement, amputation, organ loss). Disfiguring stellar scars not associated with disabilities were not considered as sequelae.
Our secondary endpoints concerned program coverage, specifically attending the BU hospital, the Ministry of Health (MOH) Health Centre or a traditional healer for BU treatment. We defined “covered” as having visited the service at least once to seek treatment for the presenting BU lesion.
BU cases were identified using a combined active case finding strategy. In a preparation phase, all selected chèferies were visited. A meeting was held in Akonolinga with the chiefs of all the selected villages, to explain the survey. Health delegates from the MOH network received two days of training in survey procedures.
In each village, the survey started with a meeting, to explain the objectives of the survey and the clinical signs of BU. Villagers were informed that, on a specified day, a medical team would come and screen every suspect case of BU at a central location. Traditional healers were contacted and informed and asked to send their patients to this central screening location. They were assured that there would be no attempt to take patients away from them. Special attention was paid to ensure that key informants (women leaders, traditional healers, village leaders) understood the objectives of the survey and the different clinical forms of BU, making every effort to use non-medical terms. During the meeting, the villagers were also informed about the second part of the active case finding strategy which consisted of house to house visits by health delegates, identifying suspect cases in the household and informing them personally about the central screening. The chief introduced the health delegates identified for this task to the community.
Health delegates had at least one week between the village meeting and the day of the central screening to perform the house to house visits. They provided information about the survey and identified suspected BU cases. They discussed the fear of stigmatisation with suspected cases, and arranged individual meetings with the medical team for suspected patients who did not want to come to the central meeting point, or arranged for transport for disabled patients. They asked for oral consent of suspected cases identified for possible inclusion in the survey. They also collected information on patients who were living in a household but at the time of the survey were admitted at Akonolinga Hospital. These patients were interviewed at the hospital.
A team comprised of one doctor/nurse experienced in BU, one medical assistant and one interviewer performed the consultations and interviews at the central screening location. The lesion was inspected, measured and categorized according to clinical criteria using the clinical case definitions. A short standardized questionnaire was administered, inquiring when the first symptoms started and where the patients were seeking care. When relevant, patients were asked why they did not go to Akonolinga Hospital. Regular field visits were made to supervise the patient interviews and questionnaire procedures.
For the statistical analyses, CSAS was treated as a random sample [14]. Prevalence (P) was computed as (detected cases/population)*(100/sensitivity (%) of the case finding strategy)*100. Coverage was calculated as the number of detected cases who had visited the healthcare provider/total number of detected cases. Capture/recapture was used to estimate the sensitivity of the case finding strategy. The estimated total number of patients (N) was N = {[(M+1)(C+1)]/(R+1)}−1 [15] with M representing the cases detected by central location screening, C the cases detected by the combined active case finding strategy, and R cases that were detected by both strategies. Since it is impossible to have fractions of cases, the estimated value for N was rounded up to the nearest whole number [16],[17]. Sensitivity of the combined case finding strategy was computed as S = C/N.
Answers to open questions in the questionnaire about health care seeking behaviour were noted word-for-word and coded in a content analysis. Resulting codes were grouped in categories.
Authorization to conduct the survey was granted by the Ministry of Health, Department of Research. Approval from the National Ethical Committee (012/CNE/MP/07) and from the Ethical Committee of MSF was obtained. Informed written consent was asked from all ulcer patients who participated in the survey.
Out of a total population of 103,000 inhabitants, 26,679 lived in the sampled chèferies. A total of 105 BU cases were identified. The age of the cases ranged from 2 to 75 years. The median age for active cases was 15.5 years (Interquartile range (IQR) 11–34 years) and 16 years for patients with sequelae (IQR 13–25). The female/male ratiobwas 0.8 (46/59) for all cases. Of the 105 cases identified, 49 (46.7%) presented with a sequelae and 56 (53.3%) were active cases. A total of 93 cases (88.6%) presented with one lesion and 12 cases (11.4%) with two lesions. The location of the first lesion was predominantly on the legs, 68 cases(64.8%).
Of the 56 active cases, the major or first lesion was an ulcer in 48 cases(85.7%), an oedema in 4 cases(7.1%) and a nodule in 4 cases (7.1%). The median diameter of ulcers was of 4 cm (IQR 2–7 cm) meaning that half of the ulcer cases would have been classified as category 1 or early lesions according to the new WHO categories. The median delay since beginning of the first BU symptoms for active cases was 12 weeks (IQR 3–30).
The overall prevalence of all BU cases was 4.7/1000 (95%CI: 4.1–7.3/1000) and the prevalence for active BU cases was estimated as 2.5/1000 (95%CI: 2.2–3.9/1000). Prevalence estimates per quadrat were categorized as low, middle or high. For active cases, categories were: 0 to 2 cases/1000; >2/1000 to 4 /1000 and >4/1000 to 6/1000 (Figure 1). For all BU cases, categories were: 0 to 3 cases/1000; >3/1000 to 6 /1000; and >6/1000 to 9/1000 inhabitants (Figure 2). The spatial distribution of prevalence estimates per quadrat showed that quadrats with high estimates were predominantly situated along the Nyong and Mfoumou rivers (Figures 1 and 2).
Of the 105 cases identified, 19 cases had visited Akonolinga Hospital resulting in an estimated coverage for the MSF BU project of 18% (95%CI 11–27%). A total of 23 patients visited the health centre at least once, leading to an estimated coverage of the peripheral MoH health centres of 22% (95%CI 15–31%). A high number of patients (77/105) consulted a traditional healer at least once, yielding a coverage of 73% (95%CI 64–81%).
Coverage estimates per quadrat were classified in 5 categories. Coverage for all BU cases for the MSF project was above 60% in the quadrat including the Akonolinga Hospital; Health centre coverage was above 60% in 3 quadrats; and above 60% in 17 quadrats for traditional healers coverage (Table 2). The same trend in coverage was seen for active BU cases. Coverage of the MSF project was 16% (95%CI 9–28%) and of the peripheral MOH Health centres was 23% (95%CI 14–36%), while coverage of traditional practitioners was 61% (95%CI 48–73%).
The geographic distributions of the coverage of the three health care providers (Hospital, health centres and traditional healers) are shown in Figure 3. The quadrate including the hospital had the highest hospital coverage and the lowest traditional healer coverage.
We visited three chèferies (2700 inhabitants) in the health district that were not part of the survey sample to estimate the sensitivity of the case finding strategy. We found 8 cases by central location screening (M), 11 cases by combined active case finding strategy (C), and 7 cases were found by both strategies (R). The total estimated number of cases was 13. The sensitivity of the active case finding method was 84.6% (95% CI: 53.7–97.3%).
Out of 86 patients who did not present to Akonolinga Hospital, 79 answered the question on motive for non-attendance. In the content analysis, a total of 87 reasons were coded. Twenty-five (31.6%) patients answered that they did not have enough money; 14 (17.7%) that the hospital was too far; 22 (27.8%) mentioned a lack of information; 13 were not aware of the services offered at the hospital; 9 didn't know it was free of charge; and 12 (15.2%) said that they did not want surgery.
Prevalence and mapping of Buruli ulcer is one of the priority areas identified by the research subgroup at the 5th WHO Advisory Group Meeting on Buruli Ulcer held in March 2002. Studies, like the one reported here, were noted as most likely to provide immediate direct benefit to Buruli ulcer patients in the medium term. The use of CSAS allowed for the identification of areas with relatively high prevalence and low coverage. An added benefit of the method is that the survey procedure also served as an awareness campaign, providing information about the disease and treatment to the population.
As previously described (4, 8), cases identified were predominantly younger than 15 years and more often male than female. This underlines once again the importance of BU as a potentially severe disabling disease that occurs at a young age. Most of the active cases (85.7%) identified presented with an ulcerative form of the disease, corresponding to what is seen at admission to the hospital (81.6%).
The overall prevalence was 0.47% (n = 105) for all BU cases including sequelae and 0.25% (n = 56) for active stages of the disease in accordance with a survey conducted previously in the region (8). Because of the lack of comparable survey in other regions it is difficult to compare our results to other estimates [18], but they corresponds to the estimated prevalence of tuberculosis in 2006 in Cameroon [19] and are slightly higher than the 6000 cases detected in a national survey in Ghana in 1999 [8]. Quadrats with higher prevalence were situated along the Nyong and Mfoumou Rivers confirming reports that cases are often found near slow moving water.
The quadrat of the area of the hospital had a high prevalence of BU, but the two other high prevalence areas were at a distance of 25 to 40 km from the hospital in the southwest of the health district.
Overall coverage of the MSF project was disappointing with 18% (12–27%) for all cases and 16% (9–18%) for active cases and was limited to quadrats neighbouring Akonolinga hospital. If we combine the geographical distribution of prevalence and coverage, we can identify the southwest of the health district as a priority area for intervention with high prevalence and low project coverage estimates. The coverage of the MOH health centres (22%) was slightly higher than the coverage of the hospital. It is important to note that this was also the case in some of the areas furthest from hospital and indicates the importance of decentralization and a functioning referral system. The high proportion of patients having visited the traditional healers, particularly in remote areas, underlines the importance of reviewing possibilities to integrate traditional healers in the project approach.
Buruli ulcer represents a financial burden for the patients and the health structures [20] Reasons given for not choosing the hospital as a health care provider were mainly financial and distance to the hospital. Lack of information about the existence of the project was also mentioned. Information on free treatment and decentralization towards health centres of diagnosis, antibiotic treatment and daily dressing, could remove barriers and rapidly improve coverage, especially when first targeting areas with high prevalence and low coverage.
A total of 15% of patients who had not gone to the hospital said it was because they did not want surgery. Providing information on the multidisciplinary combination of BU treatments (antibiotics, dressing, surgery, physiotherapy and nutrition) and giving patients the possibility to make an informed choice on which part of the treatment to accept might reduce fears and improve collaboration with traditional healers.
The quadrat including Akonolinga town presented a high prevalence and a high coverage. Since this quadrat had presumably a high population density, the overall prevalence and coverage for the health district might be slightly underestimated [11]. Population estimates of the sampled chèferies were obtained from the chiefs of the villages and confirmed by the prefecture and did not vary largely among the quadrats. Although all chiefs and all health delegates completed the same training, trust among health delegates, chiefs and villagers is a social reality that might influence the sensitivity of case finding and might vary largely. This will remain a weakness of active case finding strategies that are based on existing social networks.
In conclusion, this method was easy to use. It provided estimates of overall prevalence and coverage and identified high prevalence and low coverage areas for intervention. In addition the survey can be considered an information and awareness campaign in itself that also allowed to create a network of health delegates trained on Buruli ulcer that might refer patients in future.
|
10.1371/journal.pbio.1000371 | Most “Dark Matter” Transcripts Are Associated With Known Genes | A series of reports over the last few years have indicated that a much larger portion of the mammalian genome is transcribed than can be accounted for by currently annotated genes, but the quantity and nature of these additional transcripts remains unclear. Here, we have used data from single- and paired-end RNA-Seq and tiling arrays to assess the quantity and composition of transcripts in PolyA+ RNA from human and mouse tissues. Relative to tiling arrays, RNA-Seq identifies many fewer transcribed regions (“seqfrags”) outside known exons and ncRNAs. Most nonexonic seqfrags are in introns, raising the possibility that they are fragments of pre-mRNAs. The chromosomal locations of the majority of intergenic seqfrags in RNA-Seq data are near known genes, consistent with alternative cleavage and polyadenylation site usage, promoter- and terminator-associated transcripts, or new alternative exons; indeed, reads that bridge splice sites identified 4,544 new exons, affecting 3,554 genes. Most of the remaining seqfrags correspond to either single reads that display characteristics of random sampling from a low-level background or several thousand small transcripts (median length = 111 bp) present at higher levels, which also tend to display sequence conservation and originate from regions with open chromatin. We conclude that, while there are bona fide new intergenic transcripts, their number and abundance is generally low in comparison to known exons, and the genome is not as pervasively transcribed as previously reported.
| The human genome was sequenced a decade ago, but its exact gene composition remains a subject of debate. The number of protein-coding genes is much lower than initially expected, and the number of distinct transcripts is much larger than the number of protein-coding genes. Moreover, the proportion of the genome that is transcribed in any given cell type remains an open question: results from “tiling” microarray analyses suggest that transcription is pervasive and that most of the genome is transcribed, whereas new deep sequencing-based methods suggest that most transcripts originate from known genes. We have addressed this discrepancy by comparing samples from the same tissues using both technologies. Our analyses indicate that RNA sequencing appears more reliable for transcripts with low expression levels, that most transcripts correspond to known genes or are near known genes, and that many transcripts may represent new exons or aberrant products of the transcription process. We also identify several thousand small transcripts that map outside known genes; their sequences are often conserved and are often encoded in regions of open chromatin. We propose that most of these transcripts may be by-products of the activity of enhancers, which associate with promoters as part of their role as long-range gene regulatory sites. Overall, however, we find that most of the genome is not appreciably transcribed.
| In recent years established views of transcription have been challenged by the observation that a much larger portion of the human and mouse genomes is transcribed than can be accounted for by currently annotated coding and noncoding genes. The bulk of these findings have come from experiments using “tiling” microarrays with probes that cover the non-repetitive genome at regular intervals [1]–[9], or from sequencing efforts of full-length cDNA libraries enriched for rare transcripts [10],[11]. Additionally, capped analysis of gene expression (CAGE) in human and mouse show that a significant number of sequenced 5′ tags map to intergenic regions [12]. Estimates of the proportion of transcripts that map to locations separate from known exons range from 47% to 80% and are distributed approximately equally between introns and intergenic regions. Dubbed transcriptional “dark matter” [13], the “hidden” transcriptome [1], or transcripts of unknown function (TUFs) [4],[14], the exact nature of much of this additional transcription is unclear, but it has been presumed to comprise a combination of novel protein coding transcripts, extensions of existing transcripts, noncoding RNAs (ncRNAs), antisense transcripts, and biological or experimental background. Determining the relative contributions of each of these potential sources is important for understanding the nature and possible biological function of transcriptional dark matter.
Homology searches for transcripts mapping outside known annotation boundaries [10], as well as cDNA sequencing efforts, indicate that it is still possible to find new exons of protein coding genes [10],[15],[16]. The genomic positions of TUFs are also biased towards known transcripts [8], suggesting that at least a portion may represent extensions of current gene annotations. Nevertheless, the majority of dark matter transcripts is thought to be noncoding [2],[4],[5],[10]. Previous efforts to characterize dark matter transcripts have revealed the existence of thousands of ncRNAs with evidence for tissue-specific expression [17],[18], as well as over a thousand large intervening noncoding RNAs (lincRNAs) originating from intergenic regions bearing chromatin marks associated with transcription [19]. Other studies have reported new classes of ncRNAs, such as those that cluster close to the transcription start sites (TSSs) of protein coding genes [20]–[24]. These promoter-associated RNAs (pasRNAs) typically initiate in the nucleosome free regions that mark a TSS, with transcription occurring in both directions. Finally, results from the ENCODE pilot project have suggested a highly interleaved structure of the human transcriptome, with an estimate that as much as 93% of the human genome may give rise to primary transcripts [9]. Though this estimate was based on a combination of sources that included rapid amplification of cDNA ends coupled to detection on tiling arrays (RACE-tiling), manually curated GENCODE annotations, and paired-end sequencing of long cDNAs (GIS-PET), it was dominated by the results of RACE-tiling experiments that alone found 80% genome coverage, compared to 64.6% and 66.4% for GENCODE annotations and GIS-PET, respectively.
The fact that most TUFs do not appear to be under evolutionary selective pressure [25] has prompted suggestions that at least some of the transcriptional dark matter may constitute “leaky” background transcription [9],[26]. Consistent with this notion, many of the intergenic and intronic transcripts are detected at low levels, close to the detection limit of qPCR or Northern blots [13]. Presumably as a consequence, validation rates for unannotated transcribed regions detected in tiling array experiments have varied between 25% and 70% [1],[5],[27], and a comparison [13] of human chromosome 22 data from three major tiling array studies done on different platforms [1],[3],[27] also revealed little overlap of expressed probes, with 89% of overlapping positive probes mapping to exons or introns of known transcripts. While this low overlap may be due to differences in the samples analyzed [4], there is also evidence that some dark matter transcripts may be due to experimental artifacts. For example, a reassessment of the analysis parameters used in the tiling array study by Kampa et al. [2] revealed a similar number of transcribed fragments in real and randomized microarray data [28]. These issues make it difficult to assess the level of false positives in tiling array experiments.
Transcriptome sequencing (RNA-Seq) has emerged as a new technology that does not suffer from many of the limitations of array platforms such as cross-hybridization [29]. The technique has a wide dynamic range spanning at least four to five orders of magnitude [30],[31] and allows accurate quantitation of expression levels, as determined by experiments using externally spiked-in RNA controls and quantitative PCR [30]. These characteristics make RNA-Seq suitable to accurately assess the relative proportion of sequence from the known versus the dark matter transcriptome. Comparisons between studies of eukaryotic transcriptomes have shown that the estimated proportion of transcriptional dark matter reported in RNA-Seq studies is consistently lower than estimates from tiling arrays [32]. Although most RNA-Seq studies to date have focused on polyadenylated (PolyA+) RNA, which would be enriched for coding transcripts, this cannot fully account for the differences, as most tiling array studies show nearly the same degree of nonexonic transcription for PolyA+ as for total RNA sources [1]–[9]. Indeed, it was reported that even in the most mature form of PolyA+ RNA isolated from the cytosol, approximately half of the transcribed sequence does not correspond to known exons [5]. Moreover, RNA-Seq data from Arabidobsis rRNA-depleted total RNA samples contained a relatively small proportion (3.5%) of intergenic reads [33]. These results may not be characteristic of the larger and more complex human and mouse transcriptomes, but they do present an example in which the proportion of dark matter transcripts is relatively low in a more heterogeneous RNA pool. Other studies, in contrast, reported a higher proportion of nonexonic reads in yeast [34] and for total RNA in human [35], leaving unresolved the question of the quantity and character of dark matter transcripts.
To investigate the extent and nature of transcriptional dark matter, we have analyzed a diverse set of human and mouse tissues and cell lines using tiling microarrays and RNA-Seq. A meta-analysis of single- and paired-end read RNA-Seq data reveals that the proportion of transcripts originating from intergenic and intronic regions is much lower than identified by whole-genome tiling arrays, which appear to suffer from high false-positive rates for transcripts expressed at low levels. The majority of RNA-Seq reads that map to intergenic regions either display a high degree of correlation with neighboring genes or are associated with more than 10,000 potential novel exonic fragments we identified in human and mouse. A genome-wide analysis of “de novo” splice junctions in human samples further revealed 2,789 previously uncharacterized transcript fragments that have no overlap with exons of known gene annotations, 1,259 of which map to intergenic regions. We also find 4,544 additional exons for annotated transcripts, 723 of which extend transcripts at the 5′ end and include likely alternative promoters. The novel exons from spliced transcripts are supported by EST data, are generally more conserved, and derive from coding as well as noncoding transcripts. We conclude that analysis of data from tiling arrays leads to vast overestimates of the proportion of transcriptional dark matter. However, the mammalian transcriptome does contain thousands of unannotated transcripts, exons, promoters, and termination sites. Intriguingly, there is a strong overlap of short intergenic transcripts with DNase I hypersensitive sites, suggesting that they may be the equivalent of pasRNAs for distant enhancers.
We directly compared the accuracy of tiling arrays and RNA-Seq in identifying known transcribed regions from polyadenylated (PolyA+) RNA. To avoid potential genomic abnormalities of cell lines we mainly focused on transcriptome data from tissue sources. For microarray expression profiling, we used Affymetrix whole-genome tiling arrays at a 35 bp resolution for four human and four mouse tissues. In addition, we generated RNA-Seq data for cDNA fragments from human whole brain tissue (multiple donors) and a mixture of cell lines, which were sequenced at both ends on an Illumina genome analyzer to an average depth of 23 M paired 50 nt reads per sample. To match coverage across a wider variety of tissues, we supplemented the paired-end RNA-Seq data with publicly available 32 nt single-end PolyA+ selected datasets, sequenced to an average depth of 22 M reads for 8 human tissues from single donors [16]. RNA-Seq data for mouse were obtained from Mortazavi et al. [36] and consisted of 25 nt single-end data for PolyA+ RNA from three tissues, sequenced to an average depth of 73 M reads. The resulting combined dataset contained tissue-matched RNA-Seq and tiling array data for 4 human and 3 mouse tissues. For our analyses, we only considered RNA-Seq reads that could be unequivocally mapped to unique positions in the genome. This avoided erroneous identification of transcribed regions and facilitated comparisons to data obtained from tiling arrays, which were designed for the non-repetitive part of the genome. Overall the total number of uniquely mapped reads numbered 185.6 M and 79.8 M for the human and mouse genomes, respectively (see Table S1 for a breakdown per tissue). Since the arrays contained only perfect-match probes, the raw intensity data were normalized against a genomic DNA reference to correct for any bias in probe sequence composition (Materials and Methods).
We compared the performance of tiling arrays and RNA-Seq for human total brain tissue, since it had the highest combined sequence coverage of any tissue used in this study (50.2 M uniquely mapped reads from three independent samples, corresponding to 2.1 Gb of sequencing data). Figures 1A and 1B show the relation between the fraction of detected transcript fragments on tiling arrays (transfrags) or in RNA-Seq data (seqfrags) that overlap known RefSeq exons (i.e., precision) and the total fraction of exons recovered (i.e., recall). Tiling array transfrags were identified by selecting consecutive probes that scored above a range of intensity thresholds, with additional limits on the minimum length of each transfrag (minrun) and the maximum gap between probes meeting the threshold (maxgap). The analysis was performed directly on the normalized intensity data, or after applying additional median smoothing across neighboring probes in the genome within a sliding window, to reduce intensity variability. Seqfrags were defined as consecutively transcribed regions in the uniquely mapped RNA-Seq data, and performance was evaluated over a range of thresholds set on the minimum number of reads per seqfrag. We find that RNA-Seq offers superior precision in identifying RefSeq exons compared to tiling arrays, while achieving a high level of recall (Figure 1A, 1B). This difference remains apparent even over a broad range of parameter settings typically used to identify transcribed regions in tiling array data. These observations do not directly demonstrate that tiling arrays have a higher false-positive rate, as a lower precision would also be expected if the majority of the genome were transcribed: the difference between platforms could also reflect a lack of sensitivity to detect unannotated transcripts expressed at lower levels in RNA-Seq data, due to insufficient sequencing depth. If this were the case, however, we would expect that the precision-recall curves for RNA-Seq data would look progressively more similar to those of the tiling arrays with increasing read counts. Instead, when we examined the effect of varying sequencing depth by sampling smaller subsets of reads from the combined human brain RNA-Seq datasets we found that increased sequencing improves recall without a loss in precision (Figure 1B). Thus, the discrepancy with tiling arrays increases rather than decreases with greater sequencing depths.
We also directly compared RNA-Seq read coverage with tiling array measurements at the same genomic location. Figure 1C shows a direct comparison between the number of reads and the normalized probe signal intensity. Consistent with the precision-recall curves that show that high precision in tiling array experiments is only achieved at the most stringent intensity thresholds (Figure 1A), we find that the agreement between sequencing data and array intensities data is poor for all but the most highly transcribed regions. Indeed, the normalized intensity distribution for tiling array probes overlapping transcribed regions in RNA-Seq data with single-read coverage is essentially random (Figure 1D), consistent with previous observations that the correlation between RNA-Seq data and tiling arrays is poor for transcripts expressed at low levels [29],[36]. We do note, however, that the tiling arrays and RNA-Seq data generally agree on the location of the greatest transcript mass (Figure 1C, red line). The increased precision of RNA-Seq is presumably due to reduced ambiguity in detecting transcripts at lower expression levels, relative to microarrays, in which signal from cross-hybridization increasingly contributes to false-positive detection at low expression levels. It is thus conceivable that the proportion of dark matter transcripts based on tiling array experiments is considerably overestimated. Given the improved performance of RNA-Seq over tiling arrays, we therefore focused on RNA-Seq data to revisit the nature of dark matter transcripts.
To assess the proportion of unique sequence-mapping reads accounted for by dark matter transcripts in RNA-Seq data, we compared the mapped sequencing data to the combined set of known gene annotations from the three major genome databases (UCSC, NCBI, and ENSEMBL, together referred to here as “annotated” or “known” genes). When considering uniquely mapped reads in all human and mouse samples, the vast majority of reads (88%) originate from exonic regions of known genes (Figure 2A). These figures are consistent with previously reported fractions of exonic reads of between 75% and 96% for unique reads [16],[33],[36]–[38], including those of the original studies from which some of the RNA-Seq data in this study were derived. When including introns, as much as 92%–93% of all reads can be accounted for by annotated gene regions. A further 4%–5% of reads map to unannotated genomic regions that can be aligned to spliced ESTs and mRNAs from high-throughput cDNA sequencing efforts, and only 2.2%–2.5% of reads cannot be explained by any of the aforementioned categories. The proportions of mapped reads are consistent between tissues and cell lines and independent of read sequence length (Table S1). Altogether, dark matter transcripts only account for a small proportion of PolyA+ transcripts.
While annotated exons can explain the majority of reads, they make up a much smaller proportion of the total transcribed area of the genome: 22.3% in human and 50.6% in mouse (Figure 2B). Nevertheless, complete annotated gene structures in both organisms still account for ∼75% of the total transcribed area. The apparent discrepancy in transcribed intronic versus exonic area in human versus mouse is directly related to the combined increased sequencing depth for the human samples (Table S2). This is illustrated in Figure 2C, which shows the relationship between the amount of sequence coverage in the combined PolyA+ RNA-Seq data from human brain samples and the transcribed area. While the exonic transcribed area levels off quickly at around 500 Mb of RNA-Seq coverage, intergenic and intronic areas keep increasing at roughly constant rates. When we extrapolate from the observed relationship between the amount of mapped sequence data and genomic area covered (Figure 2D), we find that given sufficient sequencing depth the whole genome may appear as transcripts. However, the fact that such pervasive transcription would only be detected at sequencing depths more than two orders of magnitude above current levels suggests that these transcripts may largely be attributed to biological and/or technical background. Indeed, the vast majority of intergenic and intronic seqfrags have very low sequence coverage (Figure 2E, 2F), exemplified by the fact that 70% (human) to 80% (mouse) of the transcribed area in these regions is detected by a single RNA-Seq read in only one sample, much of which is consistent with random placement (see below).
The low coverage and ubiquitous character of the intronic seqfrags suggests that they may represent random sampling from partially processed or unprocessed RNAs. We also note that 4.5% of all mapped (non-unique) human RNA-Seq reads correspond to rRNAs and sn(o)RNAs, suggesting that the PolyA+ selection did not fully exclude RNAs that are not polyadenylated. Alternatively, some of these transcripts may be polyadenylated under normal conditions, or they could correspond to degradation intermediates [39]. We note that, as the number of reads increases, the amount of transcribed area in intergenic regions increases at a much lower rate than in intronic regions (Figure 2C), even though intergenic regions make up a larger proportion of the human genome (1.7 Gb compared to 1.3 Gb for introns), further supporting the notion of random sampling of introns. In the complete set of uniquely mapped human brain RNA-Seq data, intergenic reads appear 3.8-fold less often than reads in intronic regions. In contrast, the cumulative read coverage is much higher for mRNA and EST exons than it is for either introns or intergenic regions (Figure 2E, 2F), indicating that many mRNAs and ESTs likely constitute valid transcripts that are not currently annotated in the three major genome databases. In summary, even though the genome may be randomly transcribed at very low levels, the vast majority of sequence reads in PolyA+ samples corresponds to known genes and transcripts, arguing against widespread transcription to the extent reported previously.
We next sought to gain further insight into the nature of dark matter seqfrags, focusing mainly on intergenic regions to avoid possible interference from unprocessed RNAs in introns. Potential sources of seqfrags in intergenic regions include 5′ and 3′ extensions of known genes, aberrant termination products, pasRNAs, and novel genes. We therefore began our characterization of intergenic seqfrags by examining their relationship to neighboring genes. In both human and mouse PolyA+ RNA-Seq data, we observed that the average read density in intergenic regions is dramatically higher near the starts and ends of annotated genes (Figure 3A) and can extend up to a distance of ∼10 kb from both the transcription start and ends. We also observed bias towards genes in our tiling array analysis (unpublished data), as did a previous analysis using tiling arrays [8], but this study found the bias to be equal between 5′ and 3′ ends. In RNA-Seq data, the effect is stronger at the 3′ compared to the 5′ end of genes. Most transcripts at 3′ ends are consistent with alternative cleavage and polyadenylation (APA) site usage and unannotated UTR extensions of genes [16] or 3′ associated RNAs [12], rather than new exons, since in our splicing analysis (see below) we found very few instances of 3′ intergenic seqfrags linked to new 3′ exons (unpublished data). The increased number of transcripts at the 3′ end of genes is consistent with observations that RNA polymerase II can remain associated with DNA for up to 2 kb following the annotated ends of known mRNAs [40].
To determine the strand of origin of the positionally biased intergenic transcripts and to assess whether this bias was limited to PolyA+ RNA, we examined additional available sequencing-based transcriptome datasets. These included strand-specific RNA-Seq data from human rRNA-depleted whole brain and universal reference RNA [35], as well as from mouse brain PolyA+ [41] and rRNA-depleted total RNA (NCBI short read archive, SRX012528). We also incorporated data from CAGE-tag [12] and Paired-End diTag (GIS-PET) sequencing studies [42], which specifically targeted transcript ends. In all these datasets we find that most reads originate from known exons (Table S3), and among intergenic reads we find the same striking increase in read frequency in intergenic regions proximal to genes (Figure 3B, Figure S1, Table S3) as in PolyA+ samples. The enrichment of CAGE tags is consistent with peaks found at both the 5′ and 3′ ends of genes [12], and the majority of transcripts at the 3′ end of genes are in a sense orientation relative to the neighboring genes (Figure 3B). While CAGE tags are also enriched at 3′ ends of genes in the same orientation, the effect is less pronounced compared to RNA-Seq reads, suggesting that a significant number of transcripts in these regions result from alternative termination of protein-coding genes. Transcripts in intergenic regions flanking TSSs are approximately equally distributed between the sense and antisense strand (Figures 3B, S1A, and S1B), consistent with divergent transcription from promoter regions [12],[20]–[24], as well as unannotated 5′ transcript ends.
To examine the relationship between genes and gene-associated transcripts in greater detail, we next determined whether the increased sequence coverage of seqfrags in intergenic regions flanking genes correlated with the coverage of genic trancripts across the 11 human PolyA+ RNA-Seq samples (the same analysis could not be done for the mouse data, as the number of available samples was too low to reliably estimate correlations). To this end, we first identified intergenic seqfrags by merging overlapping RNA-Seq reads from all human samples and then determined the sequence coverage for seqfrags and genes in each sample. Figure 3C shows that the correlation in coverage between intergenic seqfrags and neighboring genes is much higher than it is for randomly selected genes, indicating that expression in intergenic regions is positively associated with that of the flanking genes. This effect is strongest up to a distance of 10 kb from the gene but persists to a lesser degree over larger distances (Figure 3D). After setting a threshold of p<0.05, based on how often the correlation coefficient between a given seqfrag and neighboring gene was expected to occur at random (Materials and Methods), we find a significantly increased correlation with intergenic seqfrags for 2,970 annotated genes, 934 of which remain after multiple testing correction (Table 1). Consistent with the increased read frequency at 3′ ends of genes, the number of genes with correlated intergenic seqfrags at the 3′ end is 3-fold greater than at 5′ ends of genes (Table 1). Many of the correlated seqfrags at 3′ ends are directly adjacent to the annotated genes (see Figure 3E for a representative example), adding further support to our hypothesis that many of these transcripts are linked in their expression. Additionally, we found a small number of extensions at larger distances, which are consistent with unannotated novel 3′ and 5′ exons (unpublished data and see below).
The total number of genes with correlated 5′ and 3′ intergenic seqfrags is likely underestimated in our analysis, as a minimum number of sequence reads in each sample are needed to calculate a correlation coefficient. Many transcribed intergenic regions detected at very low coverage had to be excluded from the correlation analysis, even though these low coverage regions are clearly enriched in regions flanking known genes (Figure S2). Consequently, some positional bias is still observed after removing the regions identified in this analysis (unpublished data), and correlated transcription in regions flanking genes is likely far more widespread. This is particularly relevant because while the 10 kb flanking regions make up only ∼18% of the total intergenic area, they account for as much as 78% of the intergenic reads in human and mouse PolyA+ RNA. The same trend holds true for CAGE and GIS-PET datasets, as well as RNA-Seq datasets from rRNA-depleted human total RNA (Table S3). Although gene-flanking regions in rRNA-depleted mouse brain total RNA accounted for only 30.7% of intergenic reads, further inspection revealed that most of the reads outside these regions were linked to a small number of seqfrags (21) with excessive read counts (>10,000) confined to a small area (5 kb). This strongly suggests that there are a very small number of unannotated specific transcripts expressed at high levels, and after excluding these outliers, 71.1% of intergenic reads are found near genes (Table S3). The majority of intergenic dark matter transcripts are therefore linked to annotated protein-coding genes, either as extended transcripts or separate noncoding transcripts such as pasRNAs.
Even when combining RNA-Seq data from all human or mouse tissues, read coverage in intergenic regions is very low (Figure 2B, 2C). To determine whether intergenic seqfrags are the result of low-level random background initiation, or whether they instead derive from a limited set of unannotated transcripts, we investigated the RNA-Seq read distribution in these regions. If the low-coverage intergenic seqfrags are indeed due to a uniform level of background initiation, reads should be spread evenly and the number of reads per kb of intergenic sequence should follow a random (Poisson) distribution. Given the observed transcriptional bias in regions flanking genes, we only considered intergenic regions that were at least 10 kb away from annotated genes (corresponding to ∼82% of all intergenic sequence). These trimmed regions account for 0.8% of the total number of reads in the human PolyA+ RNA-Seq data (1.64% for mouse), with an average coverage that is 9.4-fold lower than in intronic regions (3.3-fold for mouse). We find a clear departure from a random distribution in the trimmed intergenic regions of both species (Figure 4A, 4B), including several thousand loci with greater than 20 reads, which should not occur under our null hypothesis. We also independently assessed seqfrags that are supported by only a single RNA-Seq read in one tissue (“singletons”), which account for ∼70% of transcribed area in the trimmed intergenic regions in the human and mouse genomes. The distribution of singleton seqfrags is much closer to the random distribution (Figure 4D, 4E), although some deviation still persists for these low-coverage regions. To exclude that our observations are due to an inherent bias in cDNA library amplification or sequencing, e.g., due to GC content, we repeated the same analysis for an equal number of genomic DNA-Seq reads from HeLa cells [43] or a pool of human sperm DNA from four donors [44]. Both of these datasets were similarly generated on an Illumina genome analyzer and closely follow a random distribution (Figure S3). Taken together, these results indicate that while most reads >10 kb away from annotated genes are placed in a way that resembles random distribution across the genome, some have a non-random character, including several thousand regions with high read coverage that may be derived from unannotated novel transcripts.
To estimate the proportion of intergenic regions transcribed above background levels, we selected all 1 kb regions with a significantly higher read count compared to the random distribution (p<0.05) for all reads, or singleton reads only. At the lower thresholds based on singleton read frequencies, 3.0% (39.1 Mb) and 0.9% (11.4 Mb) of trimmed intergenic regions contain transcripts in the human and mouse genomes, respectively, decreasing to 1.2% (15.8 Mb) and 0.42% (5.25 Mb) at the more stringent thresholds. The increased area in the human compared to the mouse genome is consistent with the broader range of tissues assayed by RNA-Seq. The fraction of trimmed intergenic regions with significantly increased read counts is higher in human total RNA compared to PolyA+ RNA (Figure 4C, 4F): 4.1% (53.9 Mb) or 2.5% (32.8 Mb) at the lower and higher stringency levels, respectively. Considering that the total RNA sequence data was derived from a smaller sample set, this suggests that there are additional unprocessed and/or noncoding transcripts in intergenic regions not detected in PolyA+ RNA.
We also applied an additional threshold to identify putative novel exonic regions in the trimmed intergenic areas, selecting for seqfrags with a PolyA+ RNA-Seq read count greater than or equal to that of the top 5% of seqfrags detected in known introns (6 reads for human and 4 for mouse). At these thresholds we find 16,268 potentially “exonic” seqfrags in human (spanning 2.5 Mb) and 11,533 in mouse (spanning 0.66 Mb), which account for 56.9% and 87.4% of the reads in the trimmed intergenic regions in each organism, respectively. The area covered by the putative exonic seqfrags is 3.8% of the total area covered by seqfrags overlapping known exons in the human genome and 1.4% for the mouse genome. The putative exonic seqfrags tend to be well conserved at the sequence level compared to a random selection of intergenic sequences (Figure 5A, 5B), as judged by PhastCons conservation score based on multiple alignments among 18–22 mammalian genomes. This is significant, considering that the overall conservation for intergenic and intronic reads is close to random (Figure 5C, 5D). Taken together, our results show that a limited number of conserved novel exonic seqfrags can explain the majority of intergenic transcript mass detected in PolyA+ RNA, with a small proportion of low-level transcripts over a broad area that may be due to random initiation events.
We next attempted to identify novel transcript structures by detecting splice junctions between transcribed regions in the genome using Tophat [45]. Tophat uses a two-stage approach that first aligns unspliced RNA-Seq reads to the genome to identify transcribed areas, which are then examined in the second stage to identify junction sequences spanning all possible 5′ and 3′ combinations of these regions, using the reads that could not be mapped in the first stage. The main advantage of this approach is that it does not require a predefined set of annotated exons and it can therefore identify splicing between unannotated regions of the genome. Moreover, as the analysis takes the canonical splice junction donor and acceptor sites (GT-AG) into account, it is possible to determine the strand of origin for each junction, despite the fact that the PolyA+ RNA-Seq data used in this study were not generated in a strand-specific manner. We restricted our analysis to human samples, since we found the reads in the mouse dataset to be too short to reliably detect junction sequences.
Overall, we found 160,516 unique splice junctions in the 11 PolyA+ human RNA-Seq samples, 151,708 (94.5%) of which can be classified as “known,” meaning that they span any two exons within a single annotated transcript (Table S4). The remaining 8,808 novel junctions involved a single known exon or spanned two unannotated regions in the genome. In total, we could detect 57.8% of all exons in the combined set of gene annotations by at least one junction. Only 300 junctions bridged exons between transcripts, and almost all mapped to tandem-repeated regions in the genome (Table S5). Considering the high degree of sequence similarity between the repeated regions, some of these are presumably due to mapping inaccuracies. A significant proportion of bridging junctions (47%, 25% with confirmed deletions) also overlap regions with validated copy number variations (CNVs) that are common in the general population [46], suggesting that others may result from gene fusions following deletion events. These findings further argue against pervasive transcription to the extent reported in previous studies.
We assessed the false positive rate in the detected junctions by randomizing the sequences of potential splice junction reads and determined it to be 0.054% for paired-end reads and 2.7% for single-end reads (see Materials and Methods). The higher accuracy for paired-end reads demonstrates the considerable advantage of using longer reads to accurately assess splice junctions. Indeed, we found that the shorter 32 mer reads are particularly sensitive to false positive detections due to the presence of low-complexity regions and PolyA/T repeats, and we therefore applied additional filtering steps to exclude the affected junctions (see Materials and Methods for details). The longer read lengths of the paired-end compared to single-end RNA-Seq samples, combined with a 4-fold increase in sequencing depth, also resulted in a more than 3-fold higher splice junction detection rate.
The fact that short RNA-Seq reads typically cover only a single junction between exons makes it difficult to determine which combinations of alternative splice junctions correspond to transcripts observed in vivo. We therefore instead focused on identifying transcriptional units (TUs) that represent the aggregate assembly of all connected splice junctions. Thus, a completely reconstructed TU for an annotated gene will comprise the full complement of exonic regions, though these may be used in different configurations in alternatively spliced transcripts. Splice junctions were considered connected if they were directly adjacent to each other on the same strand, arranged in a head-to-tail configuration, and (i) the “facing” junction ends overlapped, or (ii) the complete region between facing splice junctions was transcribed, or (iii) facing junctions were within a distance of 200 bp (i.e., the approximate average exon size).
The vast majority of TUs we identified (91.2%) overlap with at least one exon of an annotated gene (Table 2), and 92.1% of exons in these TUs overlap known gene annotations (Table 3). We also detected 3,451 unannotated internal exons in 2,720 genes, as well as 723 and 370 unannotated 5′ and 3′ exons, affecting 544 and 290 genes, respectively. Among the TUs that are not connected to known gene annotations (i.e., independent TUs), 1,259 map to intergenic regions, the majority of which (82.6%) consist of a single junction. Only a minor fraction of independent TUs (4.8% of the total number of TUs) overlap genic regions on the sense or antisense strand. As it is possible that additional rare splice junctions are not detected in our analysis, some independent TUs overlapping genes in the sense direction may yet turn out to be connected to the gene they overlap. The majority of novel exons in the reconstructed TUs overlap with exons from the UCSC mRNA and spliced EST tracks (Table 3), providing further evidence that they are derived from true splicing events. A small number (73) further overlap exons predicted by Wang et al. [16], which were derived from an analysis of splice junctions associated with computationally predicted exons. Taken together, our findings confirm that the vast majority of spliced transcripts in PolyA+ RNA are linked to known gene annotations and argue against widespread interleaved transcription of protein-coding genes in the human genome. The full set of TUs and junctions has been made available on our supplementary website (http://hugheslab.ccbr.utoronto.ca/supplementary-data/hm_transcriptome/).
To further characterize the 4,544 novel exons connected to existing transcripts, as well as the 2,789 novel independent TUs (i.e., multi-exon transcripts), we assessed their expression levels, degree of conservation, and coding potential. As expected, novel exons detected as part of TUs that overlap annotated transcripts show evidence of increased conservation compared to randomly positioned exons (Figure 6A). Consistent with our analysis, a significant proportion of these exons overlap with Exoniphy predictions of evolutionary conserved protein-coding exons [47], most notably for novel 3′ (20.5%) and 5′ exons (18.9%) (Table S6A). The degree of overlap was significantly higher compared to random selections from intergenic regions (p<0.0001). In contrast, we observed little overlap with conserved RNA secondary structures as predicted by the Evofold [48] and RNAz algorithms [49] (Table S6A). We further examined whether the novel 5′ exons overlapped regions of open chromatin that typically mark regulatory regions [50]–[52] and which can be identified using digital DNase I hypersensitivity assays [53]. To this end, we used publicly available genome-wide data on DNase I hypersensitivity hotspots generated by the UW ENCODE group for 11 cell lines [54]. Consistent with their expected association with promoter regions, we found that the majority of novel 5′ exons overlapped the complete set of DNase I hypersensitivity zones identified by the HotSpot algorithm [53] in all 11 cell lines, as well as a more restricted set that only included hotspots found in both replicates for 8 cell lines (p<0.0001) (Table S6A).
Most of the novel exons are expressed at lower levels compared to the other exons of the gene they are linked to, which suggests that they derive from low-frequency alternative splicing events in the tissues we examined (Figure 6B). Indeed, we find direct evidence of alternative splicing for 2,526 (73%) of the novel internal exons and 2,370 of these (94%) are overlapped by junctions that bypass the novel exon. For novel exons at the 5′ and 3′ termini there is direct evidence for alternative splicing for 310 (43%) and 144 (39%), respectively. Among these are 145 cases of clear alternative promoter usage, where we find splice junctions between internal exons and the annotated promoter, as well as alternative junctions that link to a more distal promoter (Table S7). Figure 7A shows an example of one such alternative promoter for the SLC41A1 gene, encoding a solute carrier family protein.
In contrast to many of the transcribed fragments reported in tiling array studies, we find evidence for higher overall conservation for exons in independent TUs in intergenic regions, and those overlapping genes on the sense or antisense strand (Figure 6A). We assessed the coding potential of the independent TUs using a support vector machine classifier that incorporates quality measures of the available open reading frames (ORF) and blastx results [55]. Larger independent TUs with three or more exons show a general tendency to be coding: 60.8% in the case of intergenic TUs, and 70.8% and 41% for TUs overlapping genes on the sense and antisense strand, respectively. An example of a coding transcript with a translated ORF that has high sequence similarity to the elongation factor TU GTP binding domain is shown in Figure 7B. Some of the other translated TUs with clear similarities to existing proteins have stop codon mutations within the ORF, indicating that they could be pseudogenes.
None of the smaller intergenic TUs (containing only a single splice junction) were classified as coding. We note, however, that it is challenging to reliably detect the coding potential of small transcript fragments, and some of the TU fragments may in fact be part of larger coding transcripts. Indeed, when we extended the independent TUs by incorporating seqfrags overlapping the flanking junction sequences in the detected TUs, the proportion of potential coding transcripts increased to 8.3% for TUs overlapping gene regions on the antisense strand and to ∼17% for TUs overlapping genic regions on the sense strand and intergenic TUs. Moreover, we find a significant overlap with Exoniphy predictions of coding exons, ranging between 10.5% for intergenic TUs and 21% for antisense TUs (Table S6B). Further investigation will be required to characterize these smaller TUs.
Even among the larger intergenic TUs with three or more exons, there is a subset of 116 transcripts that appear to be noncoding and are thus potential human lincRNAs, one example of which is shown in Figure 7C. The fact that we could not perform a comprehensive splice junction analysis in the mouse RNA-Seq data precludes us from making a detailed comparison with the previously identified mouse lincRNAs [19], however we do find a significant overlap between 95 of the mouse intergenic seqfrags with a read count above background and 30 of the lincRNA regions (Table S8A,B). The observation that there is little overlap (0%–1%) between reconstructed TUs and Evofold and RNAz predictions (Table S6B) suggests that most transcripts identified here do not fold into conserved RNA structures. In summary, our results reveal novel alternatively spliced exons and promoters in the human genome that are used at relatively low frequencies, as well as new lincRNA candidates.
Only a small proportion (3.6%) of the 16,268 human intergenic seqfrags we identified with a read count above background were found to be part of TUs, which was surprising given that we could identify splice junctions for the majority of seqfrags in annotated exons. The lack of junctions connecting intergenic seqfrags cannot simply be explained by a reduced detection rate due to lower read counts compared to exonic seqfrags, as the proportion of intergenic seqfrags with detected junctions is consistently lower even at high coverage levels (Figure 8A). We therefore conclude that the majority of intergenic seqfrags are derived from unspliced single-exon transcripts. However, the remaining 15,646 human seqfrags that are not part of TUs are often spaced closely together, suggesting that they may be part of a single transcript, or are processed individually from larger precursor transcripts. Indeed, in many cases the intervening sequence between consecutive seqfrags is classified as transcribed when allowing reads mapping to multiple positions in the genome (see, for example, Figure 8B). When we group neighboring seqfrags with a maximum gap of 500 bp, 8,536 seqfrag clusters remain in human (7,976 of which show no evidence of splicing) and 5,506 in mouse.
We used the support vector machine classifier and Exoniphy predictions of coding exons, described above, to examine the coding potential of the unspliced intergenic seqfrags. Only 1.4% and 3.5% of human and mouse intergenic seqfrags with a read count above background overlap Exoniphy predictions, respectively (Table S8A,B). Moreover, out of the top 5% largest human intergenic seqfrags, ranging in size between 0.4 and 3.8 kb, only 12% were classified as coding. Taken together, these observations strongly suggest that the majority of the small intergenic seqfrags we identified are noncoding. As in the case of intergenic TUs, these transcripts also display little overlap with Evofold and RNAz regions.
The most striking property of the unspliced seqfrags is their strong association with open chromatin: 6,407 out of the 15,646 (40.9%) human intergenic seqfrags overlap with DNase I hypersensitivity hotspots identified in one of the 11 cell lines that were assayed, 3.4-fold more than would be expected by chance (Table S8A). Figure 8C shows a clear enrichment in tags from hypersensitive sites for RA-differentiated SK-N-SH neuroblastoma cells across the full length of brain-expressed seqfrags. Moreover, the typical size of the unspliced seqfrags (median 111 bp) is smaller than that of the DNase I-hypersensitive regions (median 248 bp), and unlike coding transcripts and other ncRNAs, many of the seqfrags appear to be contained entirely within the DNase I-hypersensitive regions. We expect that the true number of seqfrags associated with DNase I hypersensitive regions may be larger, considering that the cell lines assayed only account for a small selection of the cell types represented in the tissues and cell types assayed by RNA-Seq. Thus, these analyses reveal the existence of thousands of small intergenic transcripts associated with open chromatin.
In contrast to earlier studies based on oligonucleotide tiling array analysis of RNA [1]–[9], GIS-PET [9], and RACE-tiling arrays [9], but consistent with other RNA-Seq studies [16],[33],[36]–[38], we find that the proportion of dark matter transcripts among polyadenylated RNA from a large variety of different tissue types is small. Our comparison between tiling arrays and RNA-Seq data from the same tissues indicates that tiling arrays are ill-suited to accurately detect transcripts expressed at low levels. The major fraction of nonexonic transcripts in RNA-Seq data is associated with known genes and includes thousands of new alternative exons and hundreds of alternative promoters. However, we do not find evidence for widespread interleaved transcripts as previously described [9]; virtually all exon-exon junctions detected correspond to junctions within the same gene. Aside from new exons, most of the transcripts that are within or proximal to known genes can be explained as pasRNAs or terminator-associated RNAs, pre-mRNA fragments, or by alternative cleavage and polyadenylation site usage. The relatively small fraction of seqfrags that are not associated with known genes corresponds strongly to DNase I-hypersensitive regions. Altogether, we propose that most of the dark matter transcriptome may result from the process of transcribing known genes. Pervasive transcription of intergenic regions as described in previous studies occurs at a significantly reduced level and is of a random character.
The intergenic regions that are transcribed above background consist of a mix of both coding and noncoding transcripts. In contrast to the extensive intergenic transcription reported in tiling array studies, we found relatively few transcripts in these regions (16,268 seqfrags expressed above background levels in human and 11,533 in mouse). These numbers may be smaller, as some adjacent seqfrags may be parts of a single transcript that contain regions with sequence mapping ambiguities, or they may be larger as more tissues and cell types are surveyed.
The fact that non-exonic transcripts do not overlap with Evofold or RNAz regions argues against widespread roles as structural RNA. The most compelling support that these transcripts may have an independent function comes from the fact that they overlap with DNase I hypersensitive regions and that, unlike the many transcripts found by tiling array studies and from deep sequencing of subtracted cDNA libraries [11], the transcripts found by RNA-Seq show a significantly higher degree of conservation between species. We note, however, that these same two properties are consistent with low-level transcription from enhancers. Indeed, in yeast, it is known that placement of a strong activating transcription factor binding site in random regions of the genome results in the formation of a promoter [56]. Thus, single-exon intergenic seqfrags may represent the analog of pasRNAs for enhancers.
Our findings are based primarily on analysis of PolyA+ enriched RNA; however, our conclusions are corroborated by CAGE tags, GIS-PET, and RNA-Seq analysis of rRNA-depleted total RNA. Similar conclusions to ours were also reached in an independent RNA-Seq analysis of rRNA-depleted human total RNA (G. Schroth, pers. communication). It does not appear as if additional sequencing would substantially alter our conclusions, since coverage bias towards known exons increases with the number of reads. Moreover, while RNA-Seq analysis of PolyA+ RNA biases against very long and very short RNAs, this would not be expected to affect our ability to detect the widespread and pervasive transcription reported previously. Nonetheless, analysis of further tissues and cell types would be expected to identify additional intergenic ncRNA seqfrags that are more abundant but expressed in rare or specialized cell types. It is also likely that total RNA harbors additional transcripts not seen in PolyA+ enriched RNA and that are not evident in current total RNA-Seq analyses due to limitations in read counts.
A major remaining question is the possible function of the novel intergenic transcripts, if any. Undoubtedly, there are many functional ncRNAs remaining to be characterized [57]. However, we and others have emphasized that expression, conservation, and even localization and physical interactions of these RNAs do not constitute direct evidence for function [32]. Promoters and terminators are known to produce transcripts that appear to be associated primarily with the mechanics of gene expression and do not have known independent functions. To be conservative, a null hypothesis should perhaps be that novel transcripts—particularly those that are small and low-abundance—are a by-product rather than an independent functional unit [58]. Searching for phenotypes caused by genetic perturbation may be the most useful approach to disproving the null hypothesis.
Total and PolyA+ samples for tiling array hybridizations from pooled human and mouse heart, liver, testis, and whole brain tissues were obtained from Clontech (Table S9). All human RNA samples were derived from tissues of individuals that suffered sudden death. The human whole brain PolyA+ RNA used for paired-end sequencing came from a Microarray Quality Control (MAQC) sample (Ambion) that consisted of a mixture of RNA from 23 Caucasian males. The PolyA+ selected universal human reference sample (Stratagene) consisted of pooled RNA from 10 human cell lines (Adenocarcinoma, mammary gland; Hepatoblastoma, liver; Adenocarcinoma, cervix; Embryonal carcinoma, testis; Glioblastoma, brain; Melanoma; Liposarcoma; Histiocytic Lymphoma, hystocyte; Lymphoblastic leukemia, T lymphoblast; Plasmacytoma, B lymphocyte).
All RNA samples were DNase treated with 10 units of DNase I (Fermentas) per 50 ug of RNA prior to cDNA synthesis and purified with RNeasy spin columns (Qiagen) using a modified protocol that retains small RNAs <200 nt. Double stranded cDNA synthesis was done as previously described in Kapranov et al. [5]. Briefly, 9 ug of total RNA was reverse transcribed in a reaction that contained 1,800 units of SuperScript II enzyme (Invitrogen) and 83.3 ng of random hexamers and Oligo(dT) primers per ug of RNA. The cDNA was then used for second strand synthesis, after which the double-stranded cDNA (ds-cDNA) was purified using PCR purification columns (Qiagen) in combination with the nucleotide cleanup kit protocols. Following fragmentation and biotin labeling, 7 ug of ds-dDNA was hybridized per array.
The Affymetrix Human and mouse tiling arrays version 2.0R were originally designed for the NCBI genome assemblies v34 and v33, respectively, and were remapped to more recent genome builds (v36 for human and v37 for mouse) using BLAT [59], not allowing for any mismatches in the alignments. A small number of probes mapping to multiple locations in the genome were assigned a position that would conserve probe order relative to the original array design. In cases where this was not possible, the position on the same chromosome nearest to the original probe location was selected, or a match was randomly selected if none could be found on the same chromosome. In total, 99.5% of probe sequences could be remapped to the new mouse genome assembly, and for the human arrays this number was close to 100%. Updated bpmap files are available on request.
Arrays were scanned using an Affymetrix GeneChip scanner 3000 and raw probe intensities were obtained using the Affymetrix GeneChip Operating Software. Each array was quantile normalized against a reference genomic DNA hybridization using the Affymetrix Tiling Array Software v1.1 to obtain intensities corrected for probe sequence bias (Figure S4). The probe intensity data were further smoothed by calculating the pseudomedian of genomic DNA-normalized intensity values of probes that lie within a genomic sliding window around each probe [5]. The size of the sliding window was determined by the bandwidth parameter (BW) as follows: (2× BW) +1. Transcribed regions (transfrags) in tiling array data were selected as previously described [5], by joining positive probes together using three parameters: (i) an intensity threshold to select positive probes, (ii) the maximal distance (MAXGAP) that two neighboring positive probes can be separated by, and (iii) the minimal transfrag length (MINRUN). A range of BW, MAXGAP, and MINRUN parameter combinations were applied and used to assess precision and recall of exons in known transcripts.
Libraries for paired-end sequencing were prepared according to the manufacturer protocols. After selecting for cDNA fragments with a size distribution around 200 bp, 50 bp on both ends were sequenced in an Illumina Genome analyzer II. Single-end RNA-Seq data with a read length of 32 nt for PolyA+ RNA for 8 human tissues from individual donors (Adipose, Brain (2×), Colon, Heart, Liver, Lymph Node, Skeletal Muscle and Testis) were obtained from a previous study by Wang et al. [16]. Twenty-five mer single-end read data for PolyA+ RNA from three mouse tissues (Brain, Liver, Skeletal muscle) were taken from Mortazavi et al. [36]. Both literature datasets were produced following similar protocols that included a fragmentation step followed by a size selection for fragments of ∼200 bp and sequencing on an Illumina Genome analyzer. All paired-end RNA-Seq data are available on our supplementary website (http://hugheslab.ccbr.utoronto.ca/supplementary-data/hm_transcriptome/).
Single-end read RNA-Seq data were mapped to the NCBI human and mouse genome assemblies v36 and v37, respectively, using Seqmap v1.0.10 [60]. Several parameter settings were tested, and the maximum number of uniquely mapped reads (best unique hit) was obtained by restricting the read length to the first 25 bases and allowing for only one mismatch (Figure S5). These settings were subsequently used for all single-end read mappings. Paired-end reads from human brain and UHR samples were split and independently mapped using bowtie [61], selecting only the unique best hits from alignments that had a maximum of two mismatches in the seed sequence (first 28 bases) and an overall sum of mismatch phred quality scores no greater than 70. Single-end reads or tags from strand-specific datasets (Table S3, Figure S1) were also mapped using bowtie [61] to maintain strand information.
For the overlap analysis with known gene annotations, we combined the following tracks from the University of California Santa Cruz (UCSC) genome browser: UCSC known genes, Refseq genes, ENSEMBL genes, RNA genes, miRNAs, and snoRNAs (February 2009). In addition, mRNA and spliced EST tracks were obtained from the same source (September 2009) for a secondary mapping of seqfrags or sequence reads that did not match known gene annotations. Non-redundant sets of genes, mRNAs, and spliced ESTs were prepared by merging overlapping features, where the resulting exonic regions were defined as the union of exons in the source annotations and introns as the intervening regions between merged exons. For the calculation of the proportion of reads accounted for by each annotation category, reads were considered exonic if they partially or fully overlapped a merged exon, and intronic or intergenic if they were fully contained in these respective regions. The proportion of transcribed area was calculated by intersecting the genomic coordinates of continuously transcribed genomic regions (seqfrags) and the various genome annotation categories.
PhastCons [62] conservation scores for the human and mouse genomes were obtained from the UCSC website and were based on multi-species alignments of 18 (hg18-phastCons18way) and 20 (mm9-phastCons20way) placental mammals, respectively. Conservation scores were assigned to each seqfrag by taking the maximum PhastCons score in the genomic region covered by the seqfrag. For comparison purposes, a background score was determined for each seqfrag in the same manner, after reassigning seqfrags to random positions in the genome or within intergenic regions.
To calculate Pearson correlation coefficients between the expression levels of transcribed intergenic regions and the closest neighboring genes, overlapping mapped reads from all 11 human RNA-Seq samples were first merged into seqfrags. For each seqfrag, the nearest neighboring known transcript with read coverage in at least five RNA-Seq samples was then selected from the full set of transcripts in the UCSC known gene, Refseq gene, ENSEMBL gene, RNA gene, miRNA, and snoRNA tracks. In case multiple transcripts were found at the same distance (e.g., alternatively spliced transcripts), the transcript was selected that maximized the number of available data points for correlation analysis. The transcript expression levels in each tissue were defined as the median read coverage per kb of exon sequence and further adjusted for the difference in sequence coverage between RNA-Seq samples. Read coverage for intergenic seqfrags was determined analogously. Pearson correlation coefficients between transcript and seqfrag expression levels were only calculated if the read coverage for both the seqfrag and transcript were above zero in at least five of eleven samples, and all other intergenic seqfrags were removed from the analysis. The significance of the correlations was determined by comparing seqfrag expression levels to those of 1,000 randomly selected genes that met the same cutoff criteria. Nominal p values were defined as the proportion of random permutations where the correlation coefficient exceeded the observed correlation with the closest neighboring gene. Nominal p values were further adjusted for multiple testing by applying a Benjamini-Hochberg FDR correction [63] using the multtest R package from Bioconductor [64].
To determine whether RNA-Seq reads that map outside genes follow a random (Poisson) distribution, intergenic regions were divided into 1 kb segments and the total number of reads and the number of singleton reads in each segment was counted. Regions flanking genes up to a distance of 10 kb were excluded, as reads in these regions are more frequent, and correlated with known genes. For comparison, a random distribution was derived by sampling an equal number of uniquely mapped random reads with the same size distribution as the mapped RNA-Seq reads. To avoid a potential bias from the paired-end reads, we only mapped one of the reads in a pair. Comparisons between random and observed distributions were visualized in rootograms, which plot the square root of the number of segments as a function of the number of reads in each segment, allowing for a better assessment of differences at the tail of the frequency distribution.
Analysis of novel splice junctions was performed using Tophat [45], which uses a detection method outlined in Figure S6A. Briefly, Tophat searches for splice junctions by first mapping RNA-Seq reads to the genome to identify “islands” of expression, which are equivalent to seqfrags. In contrast to the mapping of unspliced RNA-Seq reads described above, Tophat allows multiple genomic matches for each read (up to a maximum of 40 copies) during this mapping step. Each expression island is then considered a potential exon and used to build a set of potential splice junctions, taking into account the canonical splice donor and acceptor sites (GT-AG) within each island and a small flanking region of 45 bp. Subsequently, each possible pairing of neighboring junction sequences up to a specified distance (determined by the maximum allowed intron size) is compared to the set of “missing” RNA-Seq reads that could not be matched to the genome in the first mapping step to identify sequences that span junctions. Islands with high coverage are also examined for internal junctions, to account for the possibility that the intervening intronic region between two highly expressed exons is fully transcribed at lower coverage. Paired-end reads were analyzed with Tophat version 1.0.10, which features improvements in splice junction detection specific to paired sequencing data by taking the distance between read pairs into account. In contrast, single-end read data were analyzed using Tophat version 0.8.3, as we found that this version offered greatly improved sensitivity for shorter unpaired reads.
Splice junctions in paired-end read data were mapped allowing for a maximum intron size of 500 kb, which is sufficient to encompass 99.99% of all introns and 99% of all transcripts in the complete set of annotated transcripts described above. The minimum required read match size at each junction end (i.e., anchor size) was set to 8 nt. Finally, the minimum isoform fraction was set to 0.15 to suppress junctions that were supported by too few alignments relative to the junction exons. The isoform fraction was calculated as S/D, where S is the number of reads supporting each junction and D is the average coverage of the junction exon with the highest coverage [45]. Splice junctions in single-end read data were mapped using the full 32 mer read length, rather than the shorter 25 mer reads used for the mapping of unspliced reads. The Tophat parameter settings for single-end reads were the same as for the paired-end reads, with the exception that the minimum anchor size was set to 11 and the intron size was set to 20 kb (sufficient to bridge the length of 93.98% of all annotated introns and 53% of all transcripts). The adjusted parameters for single-end read data increased precision due to the shorter read lengths (Figure S7), at the expense of a somewhat reduced ability to detect long-range splice junctions. Finally, junctions with identical sequence that mapped to more than one genomic location in both the single- or paired-end RNA-Seq data were dropped from the analysis. Alternative splicing events were defined as junctions that shared the same start position with another junction but ended at a different position, or vice versa.
In order to estimate the proportion of false positives in the splice junction prediction, we adjusted Tophat to use a modified set of reads in the splice junction detection step. The initial read mapping stage to identify islands of expression was unchanged, but the sequence for the set of “missing” reads used in the second stage to detect splice junctions between islands was reversed (Figure S6A), resulting in a scrambled set of potential junction sequences with very similar sequence properties, in particular for low-complexity and repetitive regions. In addition, the pairing of reads in the paired-end dataset was randomized. With the modified sets of “missing” reads, 62 junctions were detected in the brain and 60 in UHR sample, corresponding to an estimated false positive rate of 0.054% for paired-end read samples at the selected analysis thresholds.
At 7.3%, false positive rates for single-end reads were significantly higher, consistent with the shorter read lengths. Further examination of junction sequences revealed an over-representation of PolyA and PolyT repeats in junction sequences of single- compared to paired-end read samples (Figure S8). We believe that enrichment of these repeats is due to a bias in mapping short reads sequenced from PolyA tails, and additional filtering steps were therefore applied to exclude junctions with a PolyA/T repeat size larger than 5. Moreover, any junction found to contain more than 20% low-complexity regions as assessed by the DUST algorithm (http://compbio.dfci.harvard.edu/tgi/software/) and repeatmasker (http://www.repeatmasker.org) was discarded. After applying these filters to the real and randomized junction set, the false positive rate for detection of splice junctions in the single-end read set was reduced to 2.7%.
TUs were defined as described in the main text. Facing splice junctions arranged in a head-to-tail fashion were first assembled into tissue-specific TUs if (i) splice junction ends overlapped (ii) the complete region between facing splice junctions was transcribed or (iii) if facing splice junctions were within a distance of 200 bp (same range as the average exon size) (Figure S6B). TUs were then combined across tissues where TUs with at least one overlapping exon were merged to create a non-redundant set. Exons were detected either partially, with junctions on only one side (e.g., 5′ and 3′ terminal exons), or completely, with supporting junctions defining boundaries on both sides.
The coding potential of novel transcript fragments was assessed using a support vector machine classifier [55] that assesses the protein-coding potential based on several sequence features that incorporates quality assessments of the predicted ORF as well as BLASTX comparisons with the NCBI non-redundant protein database.
Statistical significance for overlaps between genomic feature sets (i.e., Exoniphy predicted coding exons [47], RNAz [49], and EvoFold [48] conserved RNA structures, DNase I hypersensitivity sites generated by the UW ENCODE group [54], and enhancer sets [65],[66]) and exons in transcript units or significant seqfrags in trimmed intergenic regions was calculated by permutation analysis. In each permutation round, seqfrags or TU exons were assigned random positions within intergenic regions (for novel 5′ and 3′ exons connected to annotated genes), trimmed intergenic regions (for seqfrags in intergenic regions at least 10 kb away from genes), or introns (for novel internal exons for annotated genes, as well as exons in independent sense and antisense TUs). p values were defined as the proportion of times that an overlap count greater than or equal to the number of observed overlaps was found in 10,000 permutations. Coordinates of genomic feature sets were obtained from the UCSC genome browser or the original publications and mapped to the hg18 genome build using the UCSC LiftOver tool when needed.
Affymetrix tiling array data are available at GEO (record GSE19289).
|
10.1371/journal.pntd.0007124 | Field evaluation of a locally produced rapid diagnostic test for early detection of cholera in Bangladesh | Cholera remains a substantial health burden in Asia and Africa particularly in resource poor settings. The standard procedures to identify the etiological organism V. cholerae are isolation from microbiological culture from stool as well as Polymerase Chain Reaction (PCR). Both the processes are highly lab oriented, labor extensive, time consuming, and expensive. In an effort to control for outbreaks and epidemics; an effective, convenient, quick and relatively less expensive detection method is imperative, without compromising the sensitivity and specificity that exists at present. The objective of this component of the study was to evaluate the effectiveness of a locally produced rapid diagnostic test (RDT) for cholera diagnosis.
In Bangladesh, nationwide cholera surveillance is ongoing in 22 hospitals covering all 8 divisions of the country since June, 2016. In the surveillance, stool samples have been collected from patients presenting to hospitals with acute watery diarrhea. Crystal VCTM (Span diagnostics, India) and Cholkit (locally produced RDT) have been used to detect V. cholerae from stool samples. Samples have also been sent to the main laboratory at icddr,b where the culture based isolation is routinely performed. All the tests were carried out for both direct and enriched stool samples. RDT sensitivity and specificity were calculated using stool culture as the gold standard.
A total of 7720 samples were tested. Among these, 5865 samples were solely tested with Crystal VC and 1355 samples with Cholkit whereas 381 samples were tested with both the RDTs. In comparison with culture, direct testing with Crystal VC showed a sensitivity of 72% (95% CI: 50.6% to 87.9%) and specificity of 86.8% (95% CI: 82.8% to 90.1%). After enrichment the sensitivity and specificity was 68% (95% CI: 46.5% to 85.1%) and 97.5% (95% CI: 95.3% to 98.8%) respectively. The direct Cholkit test showed sensitivity of 76% (95% CI: 54.9% to 90.6%) and specificity of 90.2% (95% CI: 86.6% to 93.1%).
This evaluation has demonstrated that the sensitivity and specificity of Cholkit is similar to the commercially available test, Crystal VC when used in field settings for detecting V. cholerae from stool specimens. The findings from this study suggest that the Cholkit could be a possible alternative for cholera endemic regions where V. cholerae O1 is the major causative organism causing cholera.
| Cholera still remains a burning public health issue in the developing world. Microbiological culture is the gold standard method for cholera diagnosis. However, it requires well equipped laboratories and 24–72 hours’ time for the isolation of pathogens, which may not be feasible for hard to reach areas and during epidemics or seasonal outbreaks. For the effective control of disease outbreaks, detection methods should be both quick and easy without sacrificing specificity and sensitivity. Rapid diagnostic test for cholera could be a potential alternative for early detection of the disease. Addressing this issue in our study, we report the performance of a rapid diagnostic test (RDT), Cholkit for the diagnosis of cholera cases using stool obtained in field settings and the assessment of its performance with those of microbial culture and Crystal VC assay, a commercially available dipstick.
| Even with the development of modern established treatments and preventative measures, cholera still remains a major health burden in low-income countries with limited resources, particularly in the developing world. Cholera is a water-borne infectious disease which can be characterized by life-threatening secretory diarrhea, often accompanied by numerous voluminous watery stools and vomiting [1]. Clinical consequences range from asymptomatic to severe disease with massive watery diarrhea which may become fatal if untreated [2]. Globally, an estimated 1.3 billion people are at risk of cholera where India and Bangladesh jointly constitute the largest share of population at risk. In Bangladesh, according to estimations, at least 66 million people are at risk of cholera with an estimated 109,052 cholera cases annually [3]. While many infections can result in only mild symptoms, at least 300,000 severe cases occur annually which are severe enough requiring hospitalisation [4]. The causative agent of cholera at present is Vibrio cholerae O1, a Gram-negative pathogen. To date, more than 200 V. cholerae serogroups have been identified where most serogroups are non-pathogenic. Only isolates of serogroup O1 (consisting of two biotypes known as ‘classical’ and ‘El Tor’ and the serotypes Ogawa and Inaba) and O139 have been reported to be pathogenic and cause cholera epidemics and pandemics. However, in the last decade no epidemics due to V. cholerae O139 have been reported and only sporadic clinical cases have been observed [5].
Stool culture remains the reference method for laboratory surveillance of cholera though the sensitivity of direct stool culture is not 100% and depends on the concentration of V. cholerae (106–107 CFU) in stool specimens [6–10]. Moreover, due to limited facilities in peripheral and field sites, diagnosis is a major hindrance for early detection of cholera in endemic regions or during a cholera epidemic. The routine culture also costs approximately 6–8 USD/case [8] and the procedure involves isolation of the bacteria, routine microbiological and biochemical analyses which is lengthy and requires about 24–72 hours. Additionally, microbiological facilities are often not feasible in remote locations and transport to the closest sufficiently equipped laboratory may add further costs. Various molecular-based techniques have been developed including PCR for the rapid detection of virulence and regulatory genes [11]. Although the specificity of PCR method is relatively high, it requires expensive equipment and technical expertise which may is very often not available in small laboratories or field settings.
For the effective control of disease outbreaks, diagnostic methods should be both quick and easy without sacrificing specificity and sensitivity of detection. RDT for cholera could be a potential alternative with advantages such as it is rapid, requires minimum training, easy to use and interpret, can be stored at ambient temperature, reasonably priced and can be deployed widely for early confirmation of cholera outbreaks. One of the most recent cholera RDTs currently available in the market is Crystal VC™ (Span Diagnostics Ltd, Surat, India), a dipstick assay initially developed by the Institut Pasteur which is now being produced by Span Diagnostics (Surat, Guzarat, India). The assay relies on the detection of the lipopolysaccharide (LPS) antigen of both V. cholerae O1 and O139 serogroups by monoclonal antibodies based on a one-step vertical-flow immunochromatography principle. Crystal VC has shown sensitivity ranging from 94–100%, and a specificity range of 84–100% [9, 12–14]. However, the O1 and O139 together in Crystal VC lead to lower specificity. Recently, another RDT named ‘Cholkit’ has been developed by our group. Cholkit is a lateral flow immunochromatography test for the qualitative determination of LPS antigen of only Vibrio cholerae O1 serogroup using monoclonal antibody specific to V. cholerae O1 [15]. The objective of this study was to evaluate the performance of the RDT Cholkit and compare the performance with Crystal VC assay, a commercially available RDT designed to detect V. cholerae O1 and O139.
The study protocol was approved by the Research Review Committee (RRC) and Ethical Review Committee (ERC) at the icddr,b. Informed written consent was taken from adult patients, or guardians on behalf of children.
Since 2016, icddr,b has been running a nationwide enteric disease surveillance in collaboration with Institute of Epidemiology Disease Control & Research (IEDCR) under the Government of Bangladesh (GoB). The surveillance is being conducted in different districts comprising of 22 sentinel sites (health facilities), covering all 8 divisions across Bangladesh (Fig 1). Stool samples were obtained from individuals seeking treatment with complaints of acute watery diarrhea. A diarrheal visit was defined as a patient (age > 2 months) attending hospital with 3 or more loose or liquid stools in last 24 hours or less than 3 loose/ liquid stools causing dehydration; or at least one bloody loose stool in last 24 hours, as well as (age < 2 months) changed stool habit from usual pattern in terms of frequency (more than usual number of purging) or nature of stool (more water than fecal matter).
Patients presented with acute watery diarrhea were requested to provide a stool sample. Freshly collected stool samples were immediately used for the direct dipstick assay at the sentinel sites. Fecal specimens were concurrently enriched overnight at 37°C in alkaline peptone water (APW) (1% peptone, 1% NaCl; pH-8.5) and dipstick assays were performed on the following day. For culture, stool samples were placed into the Cary Blair transport medium and transported to the icddr,b laboratory fortnightly by maintaining the cold chain (2−80 C). Initially all stool specimens (n = 381) were tested with both Crystal VC and Cholkit simultaneously. After observing similar performance of two RDTs, the kits were separately provided in different field sites.
Conventional stool culture was carried out by streaking stool directly on selective TTGA (taurocholate-tellurite gelatin agar) plates, and plates were incubated overnight at 37°C. Enrichment was performed in APW overnight at 37°C, followed by plating on TTGA to isolate V. cholerae. Colonies morphologically consistent with V. cholerae were tested for agglutination reaction with monoclonal antibodies specific to V. cholerae serovar O1 (Ogawa or Inaba) and O139.
Clinical and sociodemographic data were collected as per the original protocol requirement. Data were checked and then entered into the visual studio version 10.0 (Texas, USA). After completing data entry, data were transferred into the SQL server 2008. Data consistency was checked using SQL query. The primary endpoint was the assessment of the performance of the RDT using microbiological stool culture result as the gold standard for comparison. Sensitivity (true-positive or TP rate) was defined as the probability that patients with laboratory-confirmed cholera had a positive RDT. Specificity (true-negative or TN rate) was defined as the probability that patients with no laboratory-confirmed cholera had a negative RDT. The positive predictive value (PPV) was the probability that patients with a positive RDT had V. cholerae isolated from stool culture. The negative predictive value (NPV) was the probability that patients with a negative RDT had no V. cholerae isolated from a stool culture. Proportion test statistics was used for calculating p-values to distinguish the difference between two RDT kits in terms of sensitivity and specificity.
Statistical analyses were conducted using STATA version 13 (USA). Sensitivity and specificity were determined based on the comparison of Cholkit and Crystal VC results with the lab culture test and presented as percentages. Along with the percentages of sensitivity and specificity, 95% Clopper-Pearson confidence intervals (CIs) were as estimated for better predictions.
From 22 sentinel surveillance sites, a total of 7220 patients who presented with acute watery diarrhea were recruited into the study and analyzed to evaluate the performance of two RDT Kits (Fig 2). Among them 50% were from <5 years age group and 5% from 5–17 years old and the rest 45% from those who were older. Mean age of the participants was 18.75 years, and 55% were male. Among them, 381 stool samples (both direct and enriched stool) were tested by using both Cholkit and Crystal VC at the field sites, and the performance was compared with microbiological culture.
Amongst 381 stools, V. cholerae was isolated from 25 (6.6%) samples by culture. Positivity by Crystal VC with direct and enriched sample was 65/381 (17.1%) and 26/381 (6.8%), respectively, whereas Cholkit with direct and enriched sample was positive for 54/381 (14.2%) and 37/381 (9.7%) respectively (Table 1). Crystal VC on direct stool showed a sensitivity of 72.0% (95% CI: 50.6% to 87.9%), specificity of 86.8% (95% CI: 82.8% to 90.1%) and after enrichment the sensitivity and specificity were 68% (95% CI: 46.5% to 85.1%) and 97.5% (95% CI: 95.3% to 98.8%) respectively. Negative predictive values (NPV) of Crystal VC were similar; however, we found different positive predictive values (PPV) 27.7% and 65.4% on direct and enriched stool respectively. Test results on direct sample of Cholkit revealed a sensitivity of 76.0% (95% CI: 54.9% to 90.6%) and specificity 90.2% (95% CI: 86.6% to 93.1%) while enrichment revealed 64% (95% CI: 42.5% to 82.0%) and 94.1% (95% CI: 91.1% to 96.3%) respectively. The sensitivity and specificity of the RDTs using either direct or enrichment methods, were not found to be different (p>0.05). The PPVs of Cholkit was 35.2% and 43.2% on fresh and enriched samples, whereas NPVs were similar (Table 2).
A total of 5,865 direct stool samples and a subset of 614 enriched stools were tested with Crystal VC. On the other hand, 1,355 direct stools and a subset of 424 enriched samples were tested with Cholkit (Table 3). The sensitivity and specificity of Cholkit with direct stool was 79.4% (95% CI: 62.1% to 91.3%), and 87.4% (95% CI: 85.5% to 89.1%) respectively, while the sensitivity and specificity of Cholkit with enriched stool was 66.7% (95% CI: 47.2% to 82.7%), and 94.4% (95% CI: 91.7% to 96.5%), respectively. PPVs were 13.9% on direct stool and 47.6% on enriched sample, whereas NPVs showed similar result on both. In contrast, sensitivity and specificity of Crystal VC with direct stool was 72.2% (95% CI: 64.6% to 78.9%) and 77.1% (95% CI: 75.9% to 78.2%) respectively, while the sensitivity and specificity of Crystal VC was respectively 68.3% (95% CI: 51.9% to 81.9%) and 90.8% (95% CI: 88.1% to 92.9%) with enriched stool. The results of NPVs are almost similar, while PPVs are 8.2% on fresh stool and 34.6% on enriched sample (Table 4).
Rapid and accurate diagnosis of cholera at the earliest stages of an epidemic is a key feature to assist in early management of cholera outbreaks. Thus, there is a pressing need for simple and inexpensive RDT to correctly identify patients with cholera. Till date many RDTs for early detection of cholera have been evaluated [16]. Although, the sensitivity and specificity of these tests were substantial, all may not be suitable for use in the field settings. Crystal VC has been the most widely used as cholera RDT till date. Although Crystal VC is well regarded for higher sensitivity, the presence of O139 in the kit has been reported to lead to lower specificity. Recently our group has developed Cholkit RDT which has showed similar sensitivity but improved specificity compared to Crystal VC in the laboratory settings (Sayeed 2018). This study was conducted to evaluate Cholkit in field settings and compare its performance with Crystal VC.
We tested both RDTs with diarrheal stools obtained from our ongoing cholera surveillance studies. Initially, we tested both RDTs simultaneously in the field sites. Analysis with 381 diarrheal stools showed similar sensitivity and specificity. Thereafter, the kits were distributed separately in the field sites. In the field settings, Cholkit showed similar sensitivity and specificity as Crystal VC. Crystal VC detect both V. cholerae O1 and V. cholerae O139. In contrast, the newly developed RDT, Cholkit only detects V. cholerae O1. V. cholerae O1 is responsible for the majority of cholera outbreaks worldwide while V. cholerae O139 is confined to Southeast Asia and has not been involved or reported in outbreaks for more than a decade [5]. The O139 serogroup was recognized first in Bangladesh in 1992 and in nearby Southeast Asian countries [17, 18]. Since then there was only one reported outbreak with O139 serogroup in Dhaka, Bangladesh in 2002 [19]. Since then the O139 serogroup has appeared sporadically in clinical and environmental samples in Bangladesh. However, no small or large scale outbreak has been reported due to O139 [5, 20]. During epidemics where V. cholerae O1 is the responsible strain, in endemic areas, or in surveillance studies where V. cholerae O1 is the only prevalent strain; testing cholera stool with Crystal VC may create misleading interpretation and ambiguity in calculating the specificity due to the false positive result of V. cholerae O139 and consequently decrease its specificity. In accordance with our observation, evaluation of Crystal VC in field settings conducted by Ley B et al also reported similar false positive V. cholerae O139 [21] where the authors speculated that field workers may often over-interpret faint test lines as positive for V. cholerae O139. Moreover, we also cannot exclude the possibility of false positive O139 results from stool if V. cholerae O1 concentration is high in stool. Considering that, Cholkit might overcome this limitation because it does not have a test band for V. cholerae O139 and can be a suitable RDT as an alternative to Crystal VC during the predominant V. cholerae O1 era. However local production of RDT will also reduce the cost and its accessibility in Bangladesh.
This study has a number of limitations. First, the study did not assess whether the RDT results were affected by the level of skill of the technician and previous intake of antibiotics or intravenous fluids. Compared to the laboratory validation of Cholkit conducted by Sayeed et al [15], we observed lower sensitivity and specificity for both RDTs in the field settings. Previously, Kalluri et al have assessed the impact of the technician’s qualification on the performance of Crystal VC [22]. The reported RDT sensitivities of 94% and 93% when carried out by laboratory technicians and field workers respectively, were similar, but RDT specificity was higher when performed by the technicians (76% versus 67%) [13]. Harris et al and Mukherjee et al have also reported similar observations [8, 23]. Second, while confirmation of V. cholerae isolates was performed at the icddr,b laboratory, culture-negative stool samples were not validated further. In particular, we did not perform PCR testing on our RDT-positive, culture-negative samples. Bhuiyan et al [12] reported five stool samples by multiplex PCR that were O1 dipstick positive but culture-negative and found that all five were negative by PCR, indicating that the five dipstick-positive results were false positives. Thus, we cannot entirely exclude the possibility of false negativity by stool culture. Third, Alam et al pointed out that the dipstick may detect non-culturable forms of V. cholerae that have transformed into a coccoid form due to unfavorable intra-host conditions, such as antibiotic treatment prior to testing [24]. Lastly, we were not able to perform the cost analysis since Cholkit is not yet available commercially. However, according to the manufacturer, the cost of the locally produced RDT may be less compared to the commercially available RDT Crystal VC, since only it is based on a single monoclonal antibody.
Early diagnosis of cholera in an outbreak and endemic settings is of substantial public health importance. Moreover, rapid and correct detection of cholera cases in the initial stages of an outbreak is critical for containment of the infection. The RDT Cholkit has shown comparable performance to Crystal VC in the field settings. In resource -limited settings where culture facility is not readily available, Cholkit has good utility and can potentially be used as an early warning tool for cholera outbreaks in the field.
Our study demonstrated that the RDT Cholkit, locally developed in Bangladesh is comparable to Crystal VC in terms of sensitivity and specificity and can be used for monitoring cholera hotspots and epidemics. The kit will also be relatively cheaper than the commercially available RDT in the market. The Cholkit has only monoclonal antibody that detects V. cholerae O1. In a cholera endemic region like Bangladesh where V. cholerae O1 is the only prevalent strain, it is more efficient to have a test available for the O1 serogroup only. The study demonstrates the feasibility of using RDTs for monitoring cholera in resource poor as well as in hard to reach areas. This analysis also demonstrates the presence of cholera hotspots in different parts of Bangladesh in the surveillance carried out in 8 divisions of the country. However, confirmation of the RDT tests with bacteriological culture was also carried out to further strengthen and confirm the results. The information obtained from this study will be useful for planning preventive measures for eliminating cholera in Bangladesh which is an agenda for the global road map of ending cholera by 2030. In conclusion, our data shows that cholera RDTs will be helpful in predicting the population based incidence of cholera in the country and this information can also be utilized in other countries endemic or having epidemic potentials.
|
10.1371/journal.pcbi.1004935 | Molecular Dynamics of "Fuzzy" Transcriptional Activator-Coactivator Interactions | Transcriptional activation domains (ADs) are generally thought to be intrinsically unstructured, but capable of adopting limited secondary structure upon interaction with a coactivator surface. The indeterminate nature of this interface made it hitherto difficult to study structure/function relationships of such contacts. Here we used atomistic accelerated molecular dynamics (aMD) simulations to study the conformational changes of the GCN4 AD and variants thereof, either free in solution, or bound to the GAL11 coactivator surface. We show that the AD-coactivator interactions are highly dynamic while obeying distinct rules. The data provide insights into the constant and variable aspects of orientation of ADs relative to the coactivator, changes in secondary structure and energetic contributions stabilizing the various conformers at different time points. We also demonstrate that a prediction of α-helical propensity correlates directly with the experimentally measured transactivation potential of a large set of mutagenized ADs. The link between α-helical propensity and the stimulatory activity of ADs has fundamental practical and theoretical implications concerning the recruitment of ADs to coactivators.
| The regulated transcription of eukaryotic genes is governed by gene-specific transcription factors that contain activation domains to stimulate the expression of nearby genes. Activation domains are unable to take up a defined three-dimensional conformation. Nevertheless, as we demonstrate in our study, molecular dynamics simulations reveal that the key docking point of such domains (centered around several large hydrophobic amino acid sidechains) folds into fluctuating α-helical conformations. Analysis of published data shows that this tendency of adopting such local structures correlates directly with stimulation activity. We also investigate the interaction of these structurally unstable domains with a coactivator interaction partner. Computational simulations are ideally suited for analysing the rapidly changing, "fuzzy" interactions occurring between these protein partners. We gained new insights into the competitive nature of the key hydrophobic sidechains in binding to a pocket on the coactivator surface and documented for the first time the rapidly changing movements of an activation domain during these interactions.
| Control of gene expression plays a crucial role throughout all three evolutionary domains of life, allowing cells to establish cellular identity, adapt to environmental challenges and prevent diseases caused by misregulation of transcription [1]. The expression of the genome is controlled predominantly by a network of gene-specific transcription factors (GSTFs) that, after binding to target sites on DNA, regulate the rate of expression of nearby genes. GSTFs performing as transcriptional activators usually contain one or multiple activation domains (ADs; [2]) that orchestrate localized remodelling of the chromatin structure, enhanced recruitment of components of the basal transcriptional machinery on the core promoter and/or stimulate promoter escape and subsequent elongation events [3–6]. These activities typically require binding of the ADs to coactivators that integrate and convey activation signals to other components of the transcriptional machinery [6,7]. The Mediator complex surrounding the basal transcriptional machinery during transcription initiation [8–11] contains coactivators that have been shown experimentally to interact with ADs to regulate gene-specific transcription (Fig 1; [11–13]).
While more than 50 common structural motifs have been described for the DNA-binding domains, the available knowledge concerning the structure and function of ADs is comparatively limited [14]. The first ADs described almost three decades ago were shown to be both necessary and sufficient to confer the transcriptional stimulatory properties [2,15]. From a structural perspective, ADs are often characterized by their unusual primary amino acid sequence abundant in acidic amino acids, glutamine or proline residues [14–17]. The enrichment for such amino acids is thought to discourage the formation of higher order structures and thus results in an intrinsically disordered structure ("acid blobs and negative noodles" or "polypeptide lasso" structures [18–20]). In turn, the intrinsic disorder allows ADs to interact in a highly adaptable manner with a range of coactivators, culminating in a synergistic regulation of the basal transcriptional machinery by one or multiple activators (Fig 1; [21,22]). The affinity of AD-coactivator binding is reasonably high (low micro- to high nanomolar range [12,21,23]) and results in interactions lasting for several milliseconds. NMR-studies provided structural insights into a various aspects of AD-coactivator complexes (TFIID/Taf40-VP16 [24]; TFIIH/Tfb1-VP16 (PDB#2K2U [23]); NcoA1-STAT6 (PDB#1OJ5 [25]); MDM2-p53 (PDB#1YCQ [26]); CBP-CREB (PDB#1KDX [27]; MED25/VP16 (PDB#2XNF [12] and 2KY6 [13]; GAL11-GCN4 (PDB#2LPB [11]). Site-directed mutagenesis and structural studies have shown that evolutionarily highly conserved bulky hydrophobic residues within ADs play a key structural role in mediating interactions with coactivators (Fig 2A and S1 Text, [16,23,24,26,28]). When bound to coactivators, ADs form a "fuzzy" family of stochastically related structures (Fig 2D, [29–31]).
Many of the yet unanswered questions regarding AD-coactivator interactions are challenging to address experimentally, especially those concerning the dynamic range of AD conformations over time, key interaction points on coactivator surfaces, the energetics of such interactions and the structures of ADs prior to binding coactivators. Computational approaches are highly effective to model such systems on the atomic level, to study their behavior and gain new mechanistic insights that consolidate present knowledge and guide future experimental work. Here we describe the results obtained from a series of long, fully atomistic molecular dynamics simulations focusing on the experimentally well-characterized GCN4-GAL11 system from Saccharomyces cerevisiae. Accelerated molecular dynamics (aMD) methods [33,34] provide powerful tools for investigating the binding of the ADs to their coactivator targets, as well as for studying the structural properties of ADs in isolation. We describe the structural interplay of AD-coactivator complexes and explore an extensive experimental data set based on synthetic AD variants to demonstrate a high degree of correlation between the α-helix propensity, degree of "fuzziness" and the transactivation potential.
The yeast transcriptional activator GCN4 contains two tandemly arranged ADs (Fig 2A) that stimulate the expression of more than 70 "downstream" genes. The GCN4 ADs achieve this task by targeting a variety of components of the basal transcriptional machinery, including the coactivator GAL11 (also known as MED15) within the mediator complex [21]). GAL11 contains three structurally independent AD-binding domains ("Activator-Binding Domains" ["ABDs"]; Fig 2A). For one of these, ABD-1, a high-resolution structure shows a stable α-helical structure that includes a groove for interactions with ADs (PDB#2LPB; Fig 2B–2D; [11]). NOE and spin-labeling data of GAL11/ABD-1 complexed with the central AD of GCN4 (GCN4-cAD) were used to create several models illustrating the diversity of interaction between this coactivator and the AD. The bound cAD models contain a short helical stretch (encompassing GCN4 residues 116 to 124) that includes three large hydrophobic residues (W120, L123 and F124) highly conserved during evolution (Fig 2A and S1 Text). The coactivator GAL11/ABD-1 interaction surface displays three computationally detectable "hot spots" ("Pocket #1", "Pocket #2" and "Pocket #3") [32] that are distinguished by their concave topology and potentially become occupied by these particular GCN4 hydrophobic residues (Fig 2C).
We subjected PDB#2LPB-model 1 to extensive aMD simulations to gain deeper insight into various structural aspects, such as variation in AD secondary structure, orientation relative to the coactivator surface and energetic changes underpinning the conformational changes. Simulations were carried out as four independent replica runs with different initial Boltzmann distributions of particle velocities. Each simulation lasted for one microsecond, but the results reflect a period around two or three orders of magnitude longer due to the acceleration protocol used (that is, hundreds of microsecond- to millisecond-range; Table 1).
We were initially curious to see whether the aMD simulations would recreate the different binding states previously proposed by Brzovic et al. [11]. We used distance measurements between GCN4-W120 or GCN4-F124 relative to GAL11/ABD1-A126 (which forms the floor of Pocket #1; Fig 2B) to monitor pocket occupancy. The measurements show that the two key hydrophobic residues, in full agreement with the NMR-based models [11], behave in a switch-like manner and bind to GAL11/ABD1 in the three major binding states via a series of intermediate conformations (Fig 3). At various stages, the GAL11-ABD1 pocket is occupied by the sidechains of either GCN-4/W120 (Fig 3A) or F124 (Fig 3C), respectively. On several occasions, we observe a double occupancy (Fig 3B). A molecular movie illustrates a full time course of the dynamic change, including the changeover between W120 and F124 (S1 Movie). In addition to pocket occupancy state, the NMR-based models also postulate that the GCN4-cAD helical portion takes up several different orientations relative to GAL11-ABD1. Angular measurements of vectors characterizing the GCN4-cAD helix relative to GAL11-ABD1 α-helix 4 (Fig 4) correspond to orientations directly comparable to the previously described ones, but also suggest the presence of additional states representing transitional conformations. Because W120 and F124 act as pivot points in a comparable manner, the various pocket occupancy states and helix orientations observed do not appear to show any significant correlation.
Because we started the aMD simulations from just one of the 13 different models proposed previously, we wondered to what extent he observed motions of the GCN4-cAD on GAL11-ABD1 reflected the conformational space defined by the twelve remaining models. In principle, any extensive simulation of a single member of a family of structural conformers should reveal conformations that encompass the conformations of the majority of the other family members, as these structures are expected to interconvert freely during simulation. Plots of the phase space of the combined trajectories along three coordinates (helical angle; distances of the two key hydrophobic residues (W120 and F124) relative to Pocket#1) demonstrate that approximately 87% of the model coordinates are within highly populated regions (Fig 5). We conclude that the choice of 2LPB-model#1 as the starting structure for all four aMD simulations did not result in an unusually biased sampling of conformational space.
Flexibility and structural adaptability of the cAD thus enables a highly dynamic interplay that accommodates several different combinations of pocket occupancy and helical orientation. This raises intriguing questions regarding the energetics of such a variable interaction. We calculated the molecular mechanics per-residue decomposition of free binding energy (ΔGBinding) measurements along the trajectories in one-nanosecond intervals using the Molecular Mechanics Generalized Bourne Surface Area (MM-GBSA) method [35]. The van der Waals decomposition data of the GCN4-cAD confirms the dominating contribution of GCN4-W120 and F124 in binding to GAL11-ABD1 (Fig 6A; electrostatic interactions play a mostly invariant role in the GAL11-ABD1/GCN4 cAD interaction: S1 Fig). Despite the major conformational changes of the GCN4 cAD relative to the coactivator surface, the energetic contributions of GCN4-W120 and F124 interactions remain relatively steady throughout all four independent simulations. A more detailed study of these interactions reveal the varying role of at least five residues within the GAL11-ABD1 Pocket #1 in mediating these contacts (Fig 7). Two hydrophobic residues, GAL11-M173 or Y220, interact with GCN4-F124 alternatively, depending on whether the F124 sidechain is located within Pocket #1 (Figs 3A and 7A), or has moved out of it and is replaced by GCN4-W120 (Figs 3C, 7B and 7C). While GCN4-F124 occupies Pocket #1, W120 makes favorable hydrophobic contacts with the sidechains of GAL11-K217 and K221 (Fig 7A–7C). The formation of alternative—but energetically equivalent—contacts thus underpins several alternative modes of binding that are conformationally quite different from each other. Another conserved residue, GAL4-L123 (Fig 2A), provides notable van der Waals contributions, mostly in conjunction with F124, and occasionally substitutes for F124 in a reversible manner (particularly obvious in simulation GAL11-ABD1/GCN4-cAD _aMD_no1; Fig 6A).
This analysis also identifies several additional residues (GCN4-M107, F108, Y110, L113, I128, and V130) as making significant additions to ΔGBinding, but in a distinctly non-systematic manner. The residues are nodes in a structurally highly flexible network that facilitates short-lived interactions, but do not a follow recurrent pattern due to substantial and unstable conformational changes in the GCN4-cAD. To exemplify the role of these residues, we investigated the structural interactions of GCN4-M107 in more detail. In simulation GAL11-ABD1/GCN4-cAD _aMD_no2, this residue is seen as providing a substantial van der Waals contribution lasting throughout most of the second half of the simulation (timeframe 1,400–2000 ns aMD; Fig 6A). During this time, GCN4-M107 interacts predominantly with two leucine residues, L169 and L227, which are located on two different α-helices of GAL11 (helix 1 and 4, respectively), but are spatially close to each other and interact with each other via hydrophobic interactions in the folded GAL11 structure. GCN4-M107 interacts with either residue on its own, or even with both leucines by bridging them (Fig 6B and S2 Movie).
Altogether, we interpret these findings the following way: GCN4-W120 /GCN4-F124 anchor GCN4-cAD to the GAL 11 surface, but provide no preferential stabilization of conformation and/or orientation of the cAD relative to the coactivator surface. Additional hydrophobic residues located on GCN4 contribute notable, but temporary contacts (estimated by molecular mechanics measurements to contribute only between -3 to -6 kcal.mol-1 each towards ΔGBinding). The fuzzy interaction thus results from a variable combination between two relatively strong contributors (-8 to -10 kcal.mol-1 each) that are regularly supported by a host of additional minor contributors subjected to continuous change. This molecular interaction pattern predicts that the binding free energy of the GCN4 cAD to GAL11-ABD1 is far from constant and subject to constant fluctuations. The MM-GBSA estimates support the idea of substantial variation in binding affinity (over a three-fold range on the micro/millisecond time scale [S2 Fig]). The half-life of the GCN4-GAL11 interaction is estimated to be in the low millisecond range [11], which supports this interpretation. Although the constant change in affinity may result in a reasonable average affinity (and may include very high affinity states), it will also drop in affinity with statistical regularity to levels that facilitate immediate dissociation. The aMD simulations, by allowing coactivator-AD domains to be monitored over longer timer frames, convey this message much clearer than the limited number of currently existing static snap-shots of models demonstrating alternative conformations [11].
Having examined the molecular dynamics of the GAL11-ABD1/GCN4-cAD fuzzy complex, we turned our attention to the intrinsic structural properties of these two interaction partners. Specifically, we wanted to investigate to what extent binding of GCN4 affects the structure of GAL11 and, more importantly, how extensively the GCN4-cAD is structured on its own. We started by monitoring the formation/maintenance of secondary structure of the GAL11-ABD1/GCN4-cAD complex simulations described above. The analysis showed formation of a stable helical structure ("Helix #1") typically encompassing GCN4 residues S117 to D125 (Fig 8A). Simulation GAL11-ABD1/GCN4-cAD _aMD_no1 exceptionally shows a mixture of 310-helix and α-helix during the first 750 ns of simulation before settling into a α-helical pattern, whereas all other three simulations (including the final stages of GAL11-ABD1/GCN4-cAD _aMD_no1) display extensive and stable α-helices throughout the entire time course. These results are essentially in agreement with the 13 different models presented in PDB#2LPB [11], although the simulations suggest that the C-terminal border of Helix#1 routinely extends one residue further than previously proposed to include position D125. In addition to Helix #1, the occasional presence of another N-terminally located structure ("Helix #2") is evident. Helix #2 is less stable in GAL11-ABD1/GCN4-cAD _aMD_no1, no2 and no3 and either takes up a partial 310-helix conformation (GAL11-ABD1/GCN4-cAD _aMD_no1 and no3), or disappears eventually. In GAL11-ABD1/GCN4-cAD _aMD_no4, Helix #1 and #2 fuse into a single contiguous α-helix (spanning from M107 to N126 at its borders) that remains intact until the end of the simulation. Helix #2 includes the hydrophobic residues (GCN4-M107, F108, Y110, L113, I128, and V130) identified above as making occasional energetically favorable contributions to ΔGBinding.
We next asked to what extent the observed α-helical propensity of the GCN4-cAD was encoded within its primary structure. ADs are intrinsically disordered and are often thought to only adopt significant secondary structure upon binding to a coactivator target [11,29]. This model is, however, controversial. Whereas some NMR and circular dichroism studies of several isolated ADs claim an absence of significant secondary structure elements [11,36], other investigations suggest the presence of a significant fraction of transient α-helices [37–39] or β-sheets [40] in the unbound state of various ADs. In order to eliminate any structural "memory" from the starting structure, we constructed a model of the GCN4-cAD as a completely unfolded polypeptide from its primary amino acid sequence. After aMD simulation, any conformational changes—including the formation of secondary structure elements -will therefore solely be determined by the intrinsic properties of the polypeptide sequence itself. After a short implicit solvation minimization step to fold up the structure in a more compact random coil, we set up four independent microsecond aMD simulations under identical conditions as used previously for the aMD simulations of the GAL11-GCN4 complex (Table 1). Such simulations sample folding pathways and, especially relevant for disordered structures, reveal shifts in equilibria between short-lived conformations. The formation of α-helices occurs on the nanosecond-microsecond time scale [41] and is therefore well within the scope of the chosen simulation parameters.
An investigation of secondary structure elements formed in the GCN4-cAD aMD simulations reveals an unexpectedly high degree of spontaneously formed α-helices (Fig 8B). The formation of α-helices is especially favored in the central portion of the GCN4-cAD that contains the bulky hydrophobic residues that have been experimentally identified as critically important for the transactivation function, as well as making significant contributions to the free energy of binding to coactivators (Fig 6A). Although traces of β-sheet can be seen in GCN4_aMD_no4 (Fig 8B), these structures appear short-lived and do not support the conclusions reached by a previous study [40]. We conclude that the GCN4-cAD has intrinsic potential to form α-helical elements, even in absence of a coactivator, making it likely that these spontaneously preformed secondary structure elements represent key structural features required for coactivator interaction and binding specificity. The absence of substantial random coil elements in the region surrounding GCN4-W120, L123 and F124 allows us to postulate further that the cAD engages most likely with the coactivator with the necessary α-helices already locally preformed prior to first contact.
Although expected to have a less substantial effect, we also attempted to quantitate the effect of AD binding on the conformation of a coactivator. Consequently, we set up four independent one-microsecond aMD simulations of the GAL11-ABD1 in the absence of the GCN4-cAD (Table 1). A comparison of root mean square fluctuation (RMSF) measurements in simulations GAL11-ABD1 _aMD_no1 to no3 in the bound and unbound state shows that ABD1 becomes structurally more restricted upon cAD binding. Especially ABD1 residues involved in either pocket formation or binding of the cAD helix become less mobile (S3 Fig). GAL11-ABD1 _aMD_no4 undergoes a more substantial conformational change that includes a concerted movement of helices 1 and 2 and alters the ABD1 interaction surface. The original pocket for binding GCN4-W120 or F124 is no longer present, suggesting that this conformation of ABD1 may not be able to bind GCN4. The altered surface, however, develops new pockets, that may potentially offer alternative binding sites for other activators.
Up to now, we have focused our attention on a naturally occurring AD/coactivator complex that has been shown to be physiologically relevant [11,21]. Extensive mutagenesis experiments have revealed the existence of a cryptic AD within the primary amino acid sequence of GCN4. This "cAD-like" activation domain, encompassing GCN4 residues 81–100, partially matches the structural criteria for an AD, but does not display a detectable transactivation potential (Table 2; [42]).
Substitutions of hydrophobic residues within the cAD-like motif improve its activity and make it as potent as the GCN4-cAD. This modified cAD-like domain has proven an excellent testing ground for studying the transactivation potential of an array of directly comparable structures created by high-throughput site-directed mutagenesis [42]. We included two examples in our analyses and will refer from here onwards to the members of this collection as "cAD-like xx" (where xx stands for the transactivation potential that the sequence confers). For example, cAD-like07 refers to a 'weak' cAD-like variant that is capable of stimulating ARG3 induction ~7-fold (which is equivalent to the activation potential of the GCN4-cAD). On the other hand, cAD-like96 identifies a strong transactivator motif capable of stimulating ARG3 induction ~96-fold [42]. The system thus offers ideal conditions for further elucidation of the functional necessities of an exceptionally potent AD and its interactions with coactivators.
We set up in silico folding aMD simulations for the cAD-like07 and cAD-like96 motifs under identical conditions used earlier for the GCN4-cAD (Table 1). Taking into account that we previously observed significant α-helical propensity in the isolated and de novo folded GCN4-cAD (Fig 8B), one of the first questions we asked was whether such a propensity could also be detected in the cAD-like variants (while embedded within the same primary sequence context as used in the experimental work). The only ordered secondary structures formed under these conditions are α-helices, albeit with a noticeable difference in effectiveness. In the case of cAD-like07, contiguous α-helical regions are present fleetingly throughout most of the simulation period, but these fluctuate considerably in length and position relative to the underlying primary amino acid sequence (Fig 9A). In some instances all α-helical structures were lost, but restored in a fully reversible manner shortly afterwards. We conclude that cAD-like07 displays a notable tendency towards α-helical conformations, but these structures undergo a constant equilibrium between conformations of different α-helical content and are consequently unable to adopt a higher-order structure with a degree of stability exceeding the nanosecond range. In contrast, within the first 200 ns of aMD simulation the cAD-like96 variant adopts an extensive α-helical conformation that stably propagates afterwards and encompasses the three key hydrophobic residues (W94, L97 and F98) that mediate coactivator contact. The substitutions distinguishing cAD-like96 from cAD-like07 are four tryptophan residues (Table 2; W93, 95, 96 and 99; Fig 9B). Tryptophan is the strongest known helix conformer in short helices [43] and therefore the extensive helicity in cAD-like96 observed in the aMD simulations is in excellent agreement with expectations.
The detected differences in secondary structure content and stability between cAD-like07 and cAD-like96 strongly suggest that pronounced α-helical propensity constitutes a key factor in determining the transactivation potential of an AD, even in absence of additional conformational changes induced by binding to the coactivator surface. We tested this concept further by investigating whether there was a correlation between α-helical propensity predicted by standard bioinformatics tools and the observed effectiveness in mediating transactivation in vivo. A plot of predicted α-helical propensity [44] of 24 different cAD-like variants [42] against experimentally measured transcriptional simulation provides previously undocumented evidence for a strong correlation between these two variables (Fig 9C). The results show that this approach allows a direct prediction of transactivation potentials of cAD-like variants with 95% confidence using only primary amino acid sequence information.
The extensive, stable α-helicity, combined with the presence of additional bulky hydrophobic residues next and between residues W94, L97 and F98 raises some intriguing questions regarding the interaction of cAD-like96 with the GAL11-ABD1 coactivator module. As there is no structural data available for this system, we created a starting structure by in silico substitutions of the orthologous cAD residues in the GAL11-ABD1/GCN4-cAD NMR model (PDB#2LPB-model 1). Subsequently, four independent aMD simulations were carried out using the conditions described earlier. Just as expected from the results of the simulations of cAD-like96 on its own (Fig 9B), the cAD-like96 adopts a continuously stable α-helical conformation that includes positions W94, L97 and F98 throughout all four aMD simulations (Fig 9D).
Because the cAD motif is surrounded by several additional tryptophan residues, Warfield et al. suggested that these tryptophans might be able to occupy the pocket in a similar manner to the original cAD motif key residues and contribute to increased binding efficiency [42]. A molecular mechanics decomposition of the van der Waals forces of the aMD trajectories of GAL11-ABD1/cAD-like96 (Fig 10) reveals interesting similarities and differences to the previously shown GAL11-ABD1/GCN4-cAD results (Fig 6A). First, the main ΔGBinding contributions are once again centered on two regions (W94 and L97/F98), in addition to N-terminal contacts (L84, P87, L89) that provide fleeting contributions reminiscent of the pattern found for the GCN4-cAD (Fig 6A and 6B; note that these additional contacts, compared to GCN4-cAD [Fig 8A], are not in an α-helical conformation [Fig 9D]). It is noticeable, however, that the main contributors in cAD-like96 play a less distinct, broader role; in GAL11-ABD1/cAD-like96_aMD_no1 and no4, L97 makes a major contribution, but is distinctly supported by the flanking residues W96 and F98. Such a more diffuse energetic contribution is also observable near the W94 position. In GAL11-ABD1/cAD-like96_aMD_no3, W95 makes the dominant van der Waals contribution instead of W94 (a state that is briefly and reversibly explored in aMD2 at ~1,900 nanoseconds; Fig 10). A situation where W94 and W95 simultaneously occupy the ABD-1 pocket is not observed. As the helix would have to be distorted for these two residues to gain access to the pocket, it is unlikely for this confirmation to occur. In GAL11-ABD1/cAD-like96_aMD_no4, both W93 and W94 contribute apparently equally and create a stable configuration that remains essentially unchanged throughout one microsecond of aMD simulation conditions.
After identification of the possible binding states, angular measurements of the helical domain of cAD-like 96 and ABD1 α-helix 4 were performed to analyse the relative orientations of these structures relative to each other. The measurements show that the main orientations observed for the cAD-like 96 helix range typically between ~60° and ~120° (Fig 11). In comparison, the GCN4-cAD helix adopts a significantly wider range of orientations (Fig 4). Consequently, even though rotations are observable for both GAL11-ABD1/GCN4-cAD and GAL11-ABD1/cAD-like96 simulations, the maximal rotation performed by the helical cAD-like 96 domain is only 60° compared to ~180° observed for the GCN4-cAD helix. The orientations also last significantly longer and do not follow the frequent and abrupt changes observed for GCN4-cAD. We conclude that overall the binding of cAD-like96 to GAL11-ABD1 is conformationally significantly more restricted and therefore reduced in "fuzziness". The increased degree of α-helicity, redundancy of hydrophobic contacts and reduced conformational freedom documented in the aMD simulations provide a quantitative base for understanding the high transactivation potential displayed by cAD-like96.
Synthetic Biology aims at a quantitative knowledge of molecular structure/function relationships in order to provide the tools required for reshaping the properties of living organisms in a preconceived manner. A high-level of understanding of the processes underlying gene expression mechanisms will, without doubt, be required to achieve such a goal [45]. While decades of laboratory-based investigations have identified the key players of the transcriptional machinery and revealed how they network with each other, our insights are still mostly limited to a qualitative understanding at this stage.
Molecular dynamics simulations offer the ability to model accurately the dynamic interplay of proteins with high precision. While complex systems consisting of tens- to hundreds of thousands of atoms can be studied effectively with currently available high-performance computing hardware, simulations lasting for microseconds and beyond are still challenging. Enhanced sampling methods, such as accelerated molecular dynamics (aMD) effectively extend the range by two- or three orders of magnitude into the millisecond range [33,34]. This state-of-the art technology opens the door towards a better understanding of functional interactions involving the formation of "fuzzy" complex involving intrinsically disordered molecular partners. Computational approaches are of particular relevance in an area that defies conventional experimental approaches due to the high degree of structural flexibility, the rich diversity and the short duration (half-lives typically in the millisecond range) of such interactions. Enhanced sampling method MD simulations are ideally suited for such situations and offer genuine opportunities for gaining quantitative insights into this highly relevant, but still poorly understood field.
Activation domains (ADs) are intrinsically disordered structures that have mostly defied a detailed understanding of structure/function relationships. In this study, we employed molecular dynamics simulations to model an extensively studied experimental system, the interaction between the activator GCN4 with the coactivator GAL11. The experimental identification of in vivo coactivator targets [21], availability of high-quality structural models [11], combined with an extensive collection of functionally characterized mutants and artificial AD variants [11,42], makes GCN4 an ideal model system for a more thorough understanding of the fundamental aspects of transcriptional activation in eukaryotic systems. The GAL11-ABD1/GCN4-cAD complex was simulated stably in multiple aMD simulations. The molecular behavior observed likely represents motions lasting hundreds of microseconds due to the acceleration parameters employed [34]. None of these simulations has yet resulted in dissociation of GCN4-cAD from GAL11-ABD1. Estimates of ΔGBinding revealed, however, substantial fluctuations, that highlight the intrinsic instability of the complex and support the idea that the half-life of this complex is only in the millisecond range [11]. Longer aMD simulations under the conditions described, possibly employing reduced affinity mutants (such as GCN4-W120A or F124A; [11]), may eventually include a complete dissociation event. Similarly, long aMD simulations may reveal a real-time association of a free AD to a coactivator at a level of detail comparable to the binding of drugs to protein target sites [46].
Simulations of complexes between GAL11-ABD1 and GCN4-cAD or cADlike96 allowed us to observe a rich pattern of conformational changes that have thus far not been documented through any other experimental or theoretical approach. Using a single model as a starting structure (model 1) we confirmed essentially all aspects of the 13 different models included in the PDB structure (PDB#2LPB) and obtained an representative selection of intermediate structures (Fig 5) that can be viewed as a molecular movie (S1 Movie). The extensive collection of structures enabled us to gain insights into the constant and variable aspects of orientation of ADs relative to the GAL11-ABD1 (Figs 3, 4 and 11), changes in secondary structures (Figs 8 and 9), and key energetic contributions stabilizing the various conformers at different time points (Figs 6 and 11). Mutagenesis studies identified the cAD motif as W, L and F with spacing of i, i+3 and i+4. MM-GBSA calculations reinforced this result by demonstrating that the W and F residues are in both the cAD and cAD-like96 two major contributing residues towards binding ΔGBinding.
We also observed that the GCN4-cAD adopts α-helicity during aMD simulation in the complete absence of any interaction partner. The aMD simulations, starting from a polypeptide chain devoid of any secondary structure, demonstrate the formation of extensive α-helical conformations surrounding and including the conserved hydrophobic residues in a highly reproducible manner (Figs 8B, 9A and 9B). Although the presence and extent of these helices fluctuate on the hundreds of nanoseconds or microsecond time scale, the key hydrophobic residues are frequently (~65% for cAD-like07), often (~80% for GCN4-cAD) or essentially constantly (cAD-like96) arranged within an α-helical conformation, which could facilitate interactions with the coactivator surface, especially during the recruitment stage. The currently most widely—although not universally—accepted model ([37–39,47]) is based on the concept that ADs are extensively unstructured and only take up significant proportion of α-helical structures after interactions with the coactivator surface. These concepts are predominantly based on NMR-measurements showing only narrowly dispersed resonances in the 1H-N-dimension of various ADs, suggesting an absence of significant secondary structure elements. NMR-studies of poorly structured small domains remain, however, challenging. The energy barrier between disordered state and local α-helix conformations can be as low as 1.0–1.5 kcal/mol [48]. Differences between the intracellular environment and experimental conditions (low pH/low ionic concentrations, absence of divalent ions, effects of terminal flexibility, absence of local hydrophobic packing [49]) are likely to influence the formation and dynamics of small, unstable secondary structure elements. Apart from such experimental parameters, overlapping resonance effects in the spectra of flexible peptides, in conjunction with time-averaging phenomena can result in an underestimation of local structures in NMR experiments [50,51]. In the present case it is, however, likely that minor inaccuracies in the simulation parameters used in this study resulted in an overestimation of the stability of local α-helices. Although the Amber14SB force-field is generally not prone to overstabilize α-helical structures (in contrast to Amber ff03, CHARMM27 [52,53]), the results obtained here are in conflict with experimental data that show only 8–10% helical character in the free GCN4-cAD [11]. While the simulations correctly identify the regions displaying high α-helical propensity within the primary sequence of GCN4-cAD, it is evident that no conclusions should be drawn—as with aMD simulations in general—regarding the kinetic aspects of the data. Our conclusions regarding a recently published experimental data set based on the extensive mutagenesis of the cAD-like domain are, however, based on a different type of analysis and therefore not affected by such kinetic considerations [42]. Although not pointed out by the authors of this study, we observed a distinct correlation (r² = 0.89 for the first linear regression) between the published transactivation potential of 24 cAD-like variants and their theoretically predicted α-helicity (Fig 9C). This difference in α-helicity also emerges very clearly form the simulations of the three different ADs (GCN4-cAD, cAD-like07 and cAD-like-96) described here, both as free structures folded de novo, as well as complexed with GAL11-ABD1 (Figs 8 and 9). The α-helicity of a cAD-like variant can be determined using a quantitative helix-coil transition model [44], so that it should be relatively straightforward to design new cAD-like variants with predictable transactivation potentials.
The concept, that ADs either contain—or have a strong natural conformational bias towards taking up transient secondary structure elements—is not new. Experimental investigations of several other ADs has shown that they contain a significant fraction of transient α-helices in their unbound state (p53 [37]; pKID [38]; ACTR [39]; ERM [47]). We therefore conclude that eukaryotic ADs of widely different origin and specificity may contain pre-structured α-helical domains with low energy barriers of folding detectable by aMD de novo folding methods. Although sequence motifs with a high α-helical propensity can be clearly identified in unbound ADs in simulations, we note that there are distinct changes to the length and positions of these helices after binding to the coactivator. In the GCN4-cAD (compare Fig 8A with 8B), as well as in the cAD-like96 (compare Fig 9B and 9D), the boundaries tend to become more strictly defined (especially at the N-terminus) once bound to GAL11-ABD1, suggesting that coactivator-binding imposes a quantifiable degree of structural ordering on the AD. In addition, in the case of GCN4-cAD we detect reproducibly the formation of a previously unrecognized N-terminal secondary helix (Helix#2; Fig 8A) that spatially organizes up to four large hydrophobic residues (GCN4-M107, F108, Y110, and L113) into a structure that makes energetically significant van der Waals interactions contributions towards coactivator binding (Fig 6A and 6B and S2 Movie).
The degree of structural orderliness affects multiple key parameters [54]. A tight coupling between folding and binding may enhance equilibrium distinctions for interactions with different targets. A high degree of preformed structure may prove kinetically (dis)advantageous for binding to various target sites [55], so that the α-helical content of an unbound AD is likely to exert a significant influence on the rate of binding to various available interaction partners, even before more stable contacts are established through subsequent refolding/realigning events. The variable presence of α-helical modules (and location relative to the underlying primary sequence containing the conserved hydrophobic residues) may therefore encode a high degree of selectivity regarding to the binding of the various components of the transcriptional machinery [21].
The molecular mechanisms that GSTFs employ to regulate the expression of their genes are still poorly understood. There is some evidence that the binding of activation domains to basal transcription factors and coactivator complexes induces major conformational changes that could allosterically transmit signals to other components of the transcriptional machinery [56–58]. This hypothesis suggests that transient interactions of activation domains with their targets could trigger the transition between long-lived alternative coactivator conformations. Such mechanisms are, however, exceedingly difficult to study using biochemical or computational tools. In the study reported here, we have found no evidence for any significant conformational change induced in the GAL11-ABD1 structure as a direct consequence of GCN4-AD binding. Even if such changes were observed, it would still be unclear whether (and how) such an alternative conformation could be allosterically transmitted to the remainder of the GAL11 subunit (and beyond) because currently our structural knowledge of GAL11 is restricted to the ABD1 domain. An alternative—and not necessarily conflicting—view is that ADs exert most of their functions through stabilizing the assembly or position of other functional components of the transcriptional machinery, such as the basal transcriptional initiation complex. Eukaryotic promoters are potentially regulated through dozens of GSTFs bound at nearby enhancer modules, so that a multitude of energetically weak, short-lived interactions between ADs and a variety of targets could provide a significant stabilization effect through synergistic action. The short interaction half-lives and multi-target specificity of the structurally disordered ADs may under such circumstances provide the flexibility to respond to rapidly changing regulatory requirements, or provide the possibility of some components, such as RNA polymerases to "break free" of these complexes after transcription initiation. Our work documents a positive correlation between α-helicity and transactivation potential, suggesting that the overall effectiveness of AD-binding to their targets can be directly controlled through changes in α-helical propensity during evolution. Such changes may, however, have to be counterbalanced with a need for a degree of intrinsic structural disorder to sustain the ability of ADs to interact with multiple target sites. Modelling approaches, including additional AD-coactivator targets, or studying the effect of AD-interactions with larger complexes, offer great opportunities to gain further insights into the dynamical processes of coactivator-activator interactions and open numerous theoretical and applied avenues for the future. Such strategies will most likely be part of synthetic biology approaches that aim at designing artificial transcription factors with a precisely controlled range of specificity and transactivation potential in eukaryotic systems.
All structures were prepared to the same specifications to maximize comparability between the simulations. For the ABD1/cAD complex simulation both polypeptide chains of Model 1 of the GCN4-GAL11 complex (PDB#2LPB) were capped (acetyl and N-methylamide groups added to the N- and C-termini, respectively) using Yasara Structure [59]. For the ABD1/cAD-like96 simulation the GCN4-cAD structure (PDB 2LPB-Model 1) was mutagenized in silico with Yasara Structure [59] to create the cAD-like96 sequence. The coordinates were prepared for simulation in LEaP (AmberTools 14/15) with the Amber 14SB forcefield [60], neutralized and solvated in a TIP3P [61] solvent box with a minimum distance of 15 Å between solute and border. The final ionic concentration within the water box was adjusted to a final concentration of 150 mM NaCl. Capped structures of the GCN4-cAD, GCN4 cAD-like07 and cAD-like96 were built de novo from their primary amino acid in LEaP, and prefolded using 10 ns of GB implicit MD before solvating them under the same conditions described above.
The solvated models were minimized, heated to 300K and relaxed before performing a conventional MD (cMD) production run for 100 ns at a target pressure of one atmosphere to obtain values for the total potential and dihedral energy values (NPT). Simulations were carried out using the pmemd.cuda (Amber 14) applying the hybrid single/double/fixed precision model (SPFP) GPU support [62,63] using 2 fs time steps with a 10 Å cut off under control of a Langevin thermostat [64] and the SHAKE algorithm to restrain hydrogens [65]. Long-range electrostatic interactions were calculated using the Particle Mesh Ewald approximation [66]. The average total potential energies and the average dihedral energies were obtained from the cMD simulations and utilised to calculate the thresholds for dual boost aMD using an α-value of 0.2. All aMD simulations were performed with a target temperature of 300 Kelvin, and a target pressure of one atmosphere (101.325 kPa). Temperature was controlled by the Andersen temperature-coupling scheme and the pressure was controlled by the isotropic position scaling protocol applied in AMBER. Four independent 1000 ns aMD simulations were run for each structure. Details of simulations performed are summarized in Table 1. The structural models and trajectory data are available as supporting data (S1 and S2 Data Sets)
Mapping of interaction hotspots was performed using the FTMAP algorithm (http://ftmap.bu.edu [67]). Trajectory visualisation, secondary structure analysis (based on STRIDE; [68], imaging and file conversion was performed with VMD v.1.9.2 [69]. CPPTRAJ from AmberTools 15 was utilised for distance and angle measurement [70]. Bio3D was implemented for RMSD, RMSF and principal component analysis [71,72]. Visualisation of the analytical data was performed with CRAN [73]. The MM-GBSA estimation of binding free energies was performed employing the Amber forcefield ff99 [74] using the MMPBSA.py script [75]. Residue-specific decomposition was based on adding the 1–4 non-bonded interaction energies (1–4 EEL and 1–4 VDW) to the internal potential terms.
The cAD sequences were acetylated at the N-terminus and amidated at the C-terminus before predicting their α-helical properties at the residue level at pH 7.0, 150 mM NaCl and 300K [44].
|
10.1371/journal.pntd.0000222 | Interactions between Natural Populations of Human and Rodent Schistosomes in the Lake Victoria Region of Kenya: A Molecular Epidemiological Approach | Schistosoma mansoni exists in a complex environmental milieu that may select for significant evolutionary changes in this species. In Kenya, the sympatric distribution of S. mansoni with S. rodhaini potentially influences the epidemiology, ecology, and evolutionary biology of both species, because they infect the same species of snail and mammalian hosts and are capable of hybridization.
Over a 2-year period, using a molecular epidemiological approach, we examined spatial and temporal distributions, and the overlap of these schistosomes within snails, in natural settings in Kenya. Both species had spatially and temporally patchy distributions, although S. mansoni was eight times more common than S. rodhaini. Both species were overdispersed within snails, and most snails (85.2% for S. mansoni and 91.7% for S. rodhaini) only harbored one schistosome genotype. Over time, half of snails infected with multiple genotypes showed a replacement pattern in which an initially dominant genotype was less represented in later replicates. The other half showed a consistent pattern over time; however, the ratio of each genotype was skewed. Profiles of circadian emergence of cercariae revealed that S. rodhaini emerges throughout the 24-hour cycle, with peak emergence before sunrise and sometimes immediately after sunset, which differs from previous reports of a single nocturnal peak immediately after sunset. Peak emergence for S. mansoni cercariae occurred as light became most intense and overlapped temporally with S. rodhaini. Comparison of schistosome communities within snails against a null model indicated that the community was structured and that coinfections were more common than expected by chance. In mixed infections, cercarial emergence over 24 hours remained similar to single species infections, again with S. rodhaini and S. mansoni cercarial emergence profiles overlapping substantially.
The data from this study indicate a lack of obvious spatial or temporal isolating mechanisms to prevent hybridization, raising the intriguing question of how the two species retain their separate identities.
| One of the world's most prevalent neglected diseases is schistosomiasis, which infects approximately 200 million people worldwide. Schistosoma mansoni is transmitted to humans by skin penetration by free-living larvae that develop in freshwater snails. The origin of this species is East Africa, where it coexists with its sister species, S. rodhaini. Interactions between these species potentially influence their epidemiology, ecology, and evolutionary biology, because they infect the same species of hosts and can hybridize. Over two years, we examined their distribution in Kenya to determine their degree of overlap geographically, within snail hosts, and in the water column as infective stages. Both species were spatially and temporally patchy, although S. mansoni was eight times more common than S. rodhaini. Both species overlap in the time of day they were present in the water column, which increases the potential for the species to coinfect the same host and interbreed. Peak infective time for S. mansoni was midday and dawn and dusk for S. rodhaini. Three snails were coinfected, which was more common than expected by chance. These findings indicate a lack of obvious isolating mechanisms to prevent hybridization, raising the intriguing question of how the two species retain separate identities.
| One of the world's most prevalent neglected diseases is schistosomiasis, which is caused by flatworms of the genus Schistosoma. It is estimated that 200 million people world wide are infected [1]. Schistosomiasis is notable for its chronic nature, for being difficult to control on a sustained basis, and for the limited options currently available for control [2]. Schistosoma mansoni is the most widespread and best known of the human-infecting schistosomes. It is a genetically diverse parasite with complex epidemiology, particularly in East Africa, which is also its hypothesized place of origin [3].
Epidemiological studies of S. mansoni understandably often focus on human infections [4], but due to the longevity of schistosome infections in the human host and to the high vagility of humans, studies of humans alone make it difficult to detect when and where transmission actually occurs. By examining snails, the obligatory hosts for the larval stages of schistosomes, we can gain a much needed perspective, one that allows the determination of where human-infective cercariae are actually being produced, and thus identifies likely sites of active transmission. Also, during the molluscan phase of the schistosome life cycle, schistosome sporocysts may encounter other individuals of the same or a related schistosome species, or of unrelated species of digenetic trematodes (see [5] for an overview of some of the possible interactions), potentially influencing the dynamics of transmission. Molecular epidemiological investigations have shown that S. mansoni infections tend to be overdispersed (aggregated in a small proportion of host individuals) in their molluscan hosts, with some snails harboring as many as 9 distinct parasite genotypes [6],[7]. Such patterns could be the result of differing levels of susceptibility, acquired immunity [7],[8], microhabitat variation of snails and miracidia, and/or competitive interactions within the snail, and as they may influence transmission of infection to humans, should be further investigated.
In western Kenya, where our studies were undertaken, S. mansoni is likely to encounter and interact with its sister species, S. rodhaini. This species is typically considered a parasite of rodents although it has been reported from wild felids, canids, and even humans, although this latter observation has not been confirmed with molecular techniques [9]–[14]. Evidence from experimental infections of baboons suggests S. rodhaini cannot infect these primates unless they are coinfected with S. mansoni [15]. Although S. mansoni is primarily a parasite of humans and secondarily other primates, rodents can serve as reservoir hosts, including in East Africa [16]. In some locations such as Guadeloupe, rodents are the exclusive definitive host for S. mansoni [17]. Overlap of both schistosome species in the same individual rodent host was reported by Schwetz [18] who found eggs of both species in rodents the Democratic Republic of the Congo, although he considered the eggs shaped like those of S. mansoni to be a different variety of this species. Both schistosome species infect the same species of Biomphalaria snails and past reports indicate that they can infect the same individual snail host [19]; therefore they potentially influence each other in terms of infection patterns, development, and cercarial release patterns. Also, these two species hybridize readily in the laboratory [20]–[22] and a natural hybrid has been found from a snail in the Lake Victoria region [23].
Hybridization is an important epidemiological concern because hybrids could directly infect humans or lead to gene introgression between the species, which both could alter their biology and capacity to cause pathology. However, in the face of possible hybridization and definitive and intermediate host overlap, these two species are apparently able to maintain their identity [23], which unless contact is very recent, suggests the presence of isolating mechanisms including ecological, geographical, or temporal isolation. Théron and Combes [24] hypothesized that the time of day of cercarial emergence of each species could serve as an isolating mechanism since at different times of the day, different host species would be utilizing aquatic habitats. Most schistosome cercariae emerge from their snail hosts following a predictable circadian pattern [25]–[27], one that is genetically controlled [28]. Schistosoma mansoni cercariae are diurnal and are typically released during daylight hours, but populations vary concerning their exact time of emergence ([17] and references therein). Previous studies have shown that S. rodhaini is nocturnal and emerges after dark between 18:00–22:00 hours [27],[29]. These emergence times correspond to times when their putative hosts are present in the water and available for infection, humans during the day and rodents at night. However, schistosome cercariae remain active and infective in the water column for up to 9 hours in an experimental setting [30]. This longevity creates the potential for overlap in actual transmission times, even if the cercariae emerge at different times.
Using schistosome specimens derived from field collections of snails over a two year period in the Lake Victoria region of Kenya, and applying molecular techniques to these specimens, we addressed several questions concerning the epidemiology of S. mansoni and S. rodhaini, and investigated potential ecological, spatial, and temporal isolating mechanisms: 1. Do S. mansoni and S. rodhaini co-occur spatially and temporally and how prevalent are they? 2. Does either species outnumber the other in terms of number of snails infected and number of cercariae produced per snail? 3. How common are hybrids in snails? 4. How are both species distributed within their snail hosts in terms of abundance (number of genotypes per snail), and how does this correspond to the number of cercariae produced? 5. Can snails become coinfected with both species and is there any evidence the two species co-occur more or less often than expected by chance? 6. Do these species overlap on a microtemporal scale, or is there overlap in the circadian pattern of cercarial emergence for each species? 7. How are these patterns influenced when snails are coinfected with multiple multilocus genotypes or species?
Snails were collected at various sites in western Kenya in the Lake Victoria Basin (Table 1). Snails were isolated in individual wells of tissue culture plates in aged tap water for 24–48 hours and examined for shedding cercariae. Infected snails were given an individual identification number and their cercariae were used to infect mice (Swiss Albino, male and female, 6–7 weeks old), in most cases two mice per infected snail. Infections were performed via skin penetration of the abdomen while the mice were anesthetized with sodium pentobarbital. Infection doses of 10 to 200 cercariae were used depending on the number released by the snail.
Infected snails were subjected to 24 hour cercarial release profiles every 4–7 days after collection for as long as they survived. Profiles were created by counting the number of cercariae released every hour for 24 hours as the snails were moved hourly between wells of 24 well tissue culture plates, each well with 1 mL of aged tap water. Snails were kept under natural lighting (not direct sunlight) in Kenya in a laboratory with east facing windows. Additional replicates were performed in a laboratory with west facing windows and the peak emergence times did not change. Cercariae were either counted directly using a stereomicroscope if few were released, or a subsample was counted by mixing the well with a pipette, removing a subsample of 200 µL, and counting them on a gridded plate after staining with iodine. The final count was then multiplied by 5 to estimate the number in 1 mL. To determine if snails were shedding multiple genotypes or multiple species at different time intervals, cercariae were pooled into 4 time intervals (3:00–9:00, 9:00–15:00, 15:00–21:00, and 21:00–3:00) and used to infect 1–2 mice per time interval. Recovery of adult worms from mice 7 weeks post-exposure was accomplished by perfusion [31]. Gender of the worms was determined by examining adult morphology and was generally obvious with a few exceptions of infections with immature worms, which were scored as unknown. Adult worms were stored in 95% ethanol at 4°C until further use. The methodology described above has been fully approved for the use of animals by the University of New Mexico Institutional Animal Care and Use Committee (Protocol #07UNM003) and Board of Animal Care and Use of the Kenya Medical Institute.
Adult worms recovered from the mice were subsampled so that at least 16 individuals from every snail during each time interval were assayed if available. Snails did not shed during all time intervals and not all infections yielded at least 16 worms. The HotSHOT [32] method was used to prepare genomic DNA of the worms for PCR. To determine the number of genotypes of cercariae that were released from a snail, 7 previously published microsatellite loci [33],[34] were amplified in 1 multiplexed PCR reaction, the P17 panel, as described by Steinauer et al [35]. PCR products were genotyped using an ABI3100 automated sequencer (Applied Biosystems) and scored with GeneMapper® v. 4.0 (Applied Biosystems) software. All genotype calls were verified manually. Individuals with the same genotypes at all 7 loci that emerged from the same snail were considered to be clones descended from a single miracidium and are referred to as a multilocus genotype, although the probability that identical individuals arose from sexual reproduction was also calculated with GENCLONE 1.1 [36]. Part of the 16S and 12S genes (16S-12S) of the mitochondrial DNA from each multilocus genotype was amplified and sequenced using the method of Morgan et al. [23]. Sequences were submitted to GenBank Data Libraries (Accession numbers EU513397-EU513598)
Both the 16S-12S data and microsatellite data were used for species identification. Reference individuals from laboratory reared specimens and also field collected specimens of S. mansoni from Kenya, Egypt, and Brazil were used to establish species level differences with the markers. The 16S-12S data was aligned along with reference sequences from GenBank (S. mansoni: AY446260 and AY446261 (Madagascar); AY446262 and AY446263 (Kenya); AY446259 (Ghana), AF531310 (Tanzania); and S. rodhaini: AF531309, AY446265, and AY446264 (Kenya). The total dataset included the following number of specimens for each species: S. mansoni, 190; S. rodhaini, 24; S. haematobium, 1; S. bovis, 2. Sequences were aligned with ClustalX [37] using a gap opening penalty of 15 and extension penalty of 0.2. Identical sequences were identified using Sequencher 4.6 (Genecodes) and redundant sequences were removed from the alignment. Phylogenetic analyses using the minimum evolution optimality criterion was performed on the data using the model of evolution selected by the likelihood ratio test implemented in MODELTEST 3.0 [38]. Tree searches were done heuristically using PAUP* 4.0b10 [39] with tree bisection reconnection (TBR) branch swapping on initial trees that were obtained by random stepwise addition of taxa, replicated 100 times. Node support for the node separating S. mansoni and S. rodhaini was assessed by bootstrap analysis [40] using the faststep option with 10,000 pseudoreplicates. Species identification was based on clustering with reference sequences from GenBank. Genetic divergence was calculated using MEGA version 2.1 [41]. Within clade divergences and net between clade divergences were calculated using uncorrected p-distances, which is the proportion of sites that differ between two taxa. For the microsatellite data, a population assignment test was performed with GenAlEx [42] using the “leave one out method” to assess whether the microsatellite markers agreed with the 16S-12S data and could differentiate the species using the 7 microsatellite loci. The loci were also compared by eye to determine which were able to differentiate the species.
Prevalence, or percentage of infected snails, of schistosomes and of each schistosome species was calculated for each collection and also pooled across collections by site (Table 2). A proportion of infections (33%) could not be identified to species because the snails never released enough cercariae to infect mice, the mice did not become infected by the cercariae, or the mice died before worms could be recovered. Therefore, estimated prevalence values were also calculated by apportioning the total prevalence value to each species based on their proportion in the known specimens at each site. Both raw prevalence and estimated prevalence values are given in Table 2. To test if prevalence (raw values) and mean intensity (number of genotypes per snail) of infection was positively correlated as noted in previous studies [43], a Pearson's correlation was calculated on the log transformed values using the same software. Also, an analysis of covariance (ANCOVA) that examined the difference in the total number of cercariae released between species and its relationship to snail size was performed. Only snails infected with a single genotype and that shed more than 90 cercariae were used in this analysis. The model included species as a categorical variable and snail size as a covariable as well as the interaction between the terms.
To determine if coinfections in snails were random occurrences or if they were the product of a structured community, the observed parasite communities were compared to a null models of communities based on the observed values of the species' prevalence as described by Lafferty et al. [44]. Expected numbers of coinfected snails were calculated as the product of the number of snails collected and the prevalence (as a proportion) of each parasite species present in the population at the site of interest, during the time of interest (not pooled spatially or temporally). The expected number was compared to the observed number using χ2 goodness of fit tests.
Two-tailed Fisher's Exact tests were used to detect if the proportion of each genotype of cercariae shed from multiply infected snails varied among replicates over time using VassarStats (www.faculty.vassar.edu/lowry/VassarStats.html). Only one snail yielded enough data to examine the three way relationship among genotype, replicate, and time of day (most snails yielded adults mostly from a single time period, 9:00–15:00). This snail was coinfected with both schistosome species, 3 genotypes of S. rodhaini and 1 genotype of S. mansoni. These data were analyzed with a 3-way contingency table with a log-linear analysis for goodness of fit using VassarStats, and the standardized deviates were examined to determine which categories contributed the most to observed significant values.
Alignment and removal of redundant sequences yielded 512 bp for 64 taxa: 61 S. mansoni and one each of S. rodhaini, S. bovis, and S. haematobium. The evolutionary model selected by the likelihood ratio test implemented by MODELTEST 3.0 [38] was the unequal-frequency Kimura 3-parameter model. Phylogenetic analysis yielded 9 trees that did not differ in their groupings of specimens between species (Fig. 1). Within S. mansoni 1.5% sequence divergence was detected; however, no variation was detected in S. rodhaini (24 specimens) or S. bovis (2 specimens). The net between groups genetic distance between S. mansoni and S. rodhaini was 9.3%, which was greater than the distance between S. haematobium and S. bovis (7.6%). A population assignment test using the microsatellite markers yielded 100% assignment of the individuals of S. mansoni and S. rodhaini to their species based on the 16S-12S data (Fig. 2). Two loci were completely non-overlapping between S. mansoni and S. rodhaini (SMD28 and SMD89 from [34]), and one locus (SMD43 from [33]) did not amplify in S. rodhaini. There was no evidence of hybrids based on the mtDNA and microsatellite markers which were concordant in their identification of each individual. Also, no individuals were found to have microsatellite signatures that were indicative of hybrids either in the nonoverlapping loci or the other loci as shown by the population assignment test, which placed the species in relatively tight groups (Fig. 2).
A total of 22,641 snails were collected in the Lake Victoria basin over a 2 year period. Of these snails, 236 (157 B. sudanica and 79 B. pfeifferi) were infected with schistosomes, a prevalence of 1.04%. Not all schistosome infections were identified, but of the 167 that were, 90% were S. mansoni and 8.1% were S. rodhaini, and 1.9% were mixed species infections. Most infections of S. rodhaini occurred in B. sudanica and only one individual of B. pfeifferi was infected with this species, which was a coinfection with S. mansoni. The sex ratio of adults obtained from mice of S. mansoni was male biased (2.36), while that of S. rodhaini was more equivalent (1.11).
Prevalence of schistosome infection varied spatially and ranged from 0.11–3.65% among positive collection sites (Table 2). Prevalence was the highest for both S. mansoni and S. rodhaini at the Car Wash site, which is an area along the shore of Lake Victoria in the city of Kisumu, Kenya, where a population of car washers earns their living by washing vehicles in the lake and is known to be infected with schistosomes [45]. Schistosoma mansoni was more prevalent and widespread than S. rodhaini which was only present at 7 of the 14 collection sites where S. mansoni occurred, and there were no sites where only S. rodhaini occurred. Total prevalence (added over time) of S. rodhaini was not greater than S. mansoni at any one site, but was more prevalent in 7 of the 169 individual collections at Nawa, Nyabera, Usare Beach, and Lwanda. Seasonal patterns of prevalence were not evident, but prevalence for both species was low between November 2004 and March 2005, and increased between September 2005 and March 2006 (Fig. 3).
Examination of the number of genotypes per schistosome species per infected snail included a dataset that consisted only of snails that yielded 8 or more adult worms for DNA analysis and totaled 138 snails. The total number of adults genotyped was 4,777, with a mean of 34.1 per snail (2.5 standard error), range of 8–217, and median of 24 adults per snail. Many snails were sampled over multiple days or shedding intervals that were 4–7 days apart. Snails were sampled over a mean of 2.3 (0.18 standard error) replicates, and ranged between 1 and 6 replicates. For S. mansoni, the 7 loci were adequate to determine that identical individuals were derived from clones and not separate sexual reproduction events. The Psex values (probability that the same multilocus genotype was produced from independent sexual reproduction events) ranged from 1.2×10−27 to 0.000735 for this species. For S. rodhaini, individuals were less diverse and Psex values ranged from 1.2×10−8 to 0.1442; however, this method does not take into account the probability that two individuals that are identical due to sexual reproduction infect the same individual snail host, which is 8.8×10−5 for S. rodhaini. Therefore, it is highly unlikely that we are missing genotypes of either species due to identical individuals in the same snail hosts.
Of the snails that yielded at least 8 adults (128 for S. mansoni and 12 for S. rodhaini, with 2 of these snails coinfected with both species), most harbored only one genotype, but multiple infections of up to 4 genotypes were found (Table 3). A total of 152 genotypes of S. mansoni were found in 128 infected snails and 14 genotypes of S. rodhaini were found in 12 infected snails. There was a significant positive correlation between prevalence and mean intensity (number of genotypes per snail) r2 = 0.264, p<0.05.
Three snails harbored genotypes of both S. rodhaini and S. mansoni, and were found at different sites during the last week of October of 2005 or 2006: Asembo Bay, Nyabera, and Asao. Statistical comparison with null communities indicated that the schistosome communities were structured and coinfections were more common than expected by chance at all three collecting sites, Nyabera (χ2 = 49.3, P<<<0.0001), Asao (χ2 = 140.1, P<<<0.0001), and Asembo Bay (χ2 = 305.4, P<<<0.0001). According to the calculated expected values, one would have to collect 15,692, 40,571, and 94,769 snails at each site, respectively, to find one coinfected snail.
Circadian cercarial emergence profiles were generated based on 226 replicates from 100 snails infected with S. mansoni and 27 replicates from 8 snails infected with S. rodhaini (identified based on mtDNA sequences and microsatellite genotypes). Peak cercarial emergence of S. mansoni occurs between 8:00–13:00 and emergence of S. rodhaini was bimodal with a peak occurring between 5:00 and 8:00 and also 19:00 to 22:00 (Fig. 4). The ANCOVA revealed a significant interaction between parasite species and snail size, making the other effects difficult to interpret because of the uneven slopes (species: F1,127 = 4.702 p = 0.032; size: F1,127 = 0.401 p = 0.528, interaction: F1,127 = 5.087 p = 0.026). Using separate regressions, S. mansoni cercarial abundance has a significant positive relationship with snail size (F1,117 = 9.275 p = 0.003 r2 = 0.073), and S. rodhaini does not (F1,10 = 2.003 p = 0.187), a result that could be an effect of sample size since there were far fewer snails infected with S. rodhaini. A T-test indicated that there was no difference in cercarial production by species (Tdf = 19 = 1.237, p = 0.231).
Snails that were infected with multiple genotypes differed in the ratios of each genotype released and the proportions of each ranged from 50% of each to 95% and 5% of each. A total of 11 snails were examined for independence between the genotypes released and replicates performed over time (typically a week apart). Results indicated significant differences or non-independence between genotype and replicate for 6 of the snails (Table 4). The patterns of five of the six indicated a replacement pattern in which an initially dominant genotype is less represented in later replicates. The remaining snail showed variable proportions over 3 replicates; however, one genotype was always dominant. The five snails with nonsignificant values displayed a more constant pattern of cercarial release in which the proportions of each genotype did not change over time.
For the mixed species infections, limited data were obtained from two of the three snails. For the Asao snail, only 7 worms of 2 female genotypes was recovered, 6 of which were S. mansoni and 1 was S. rodhaini. Interestingly, a single species infection of S. rodhaini was never found at this site. For the Asembo Bay snail, cercariae were collected twice, 28 days apart. In the first collection, 16 adults were genotyped and all were one female genotype of S. mansoni. Unfortunately for the second collection, only 3 adults were recovered: one was a male S. rodhaini and 2 were a male genotype S. mansoni, but a different genotype than released previously. More extensive data was obtained from the Nyabera snail, which shed 1 male genotype of S. mansoni and 3 genotypes of S. rodhaini, one male and two females. Each of 4 replicates of circadian cercarial emergence showed a peak from 8:00–10:00 hours, which corresponds to S. mansoni emergence, and also an earlier morning peak that corresponds to S. rodhaini emergence. Two replicates also showed nocturnal peaks that also correspond to S. rodhaini (Fig. 5). The number of adults obtained from infections of mice with cercariae collected from different time pools of these circadian profiles indicated that S. rodhaini was more common: 94% were of this species, and 43% of these were of the male genotype. The three-way contingency table analysis indicated that all variables, genotype (G), replicate (R), and time of day (T) and their interactions, were significant (G by R: G2 = 17.3, p = 0.0002; G by T G2 = 22.44, p = 0.0002; R by T G2 = 18.84, p<0.0001; G by T by R: G2 = 55.12, p<0.0001). The three largest standardized deviates by more than a value of 1 included the comparison of the S. mansoni genotype between 9:00 and 15:00 hours (3.256), a female S. rodhaini genotype during 3:00 to 9:00 hours (2.111), and the S. mansoni genotype between 21:00–3:00 hours (−2.036). These values indicate that the S. mansoni genotype was more common than expected during 9:00 and 15:00, the peak emergence time for this species (and the only time period that this species was collected) and less common than expected during 21:00–3:00, a time period when this species rarely emerges (Figs. 4–6). Also, one of the female genotypes of S. rodhaini (R3) was more common than expected during the 3:00–9:00 time period, one of the peak emergence times of this species (Fig. 6).
Schistosoma mansoni and S. rodhaini both have spatially and temporally patchy distributions in snails in the Lake Victoria region of Kenya and active infections (those producing cercariae) are characterized by low prevalence of about 1% combined. Although this number may be characterized as low in a relative sense, given the prodigious number of snails supported by Lake Victoria and its environs, this level of infection in snails is responsible for relatively high levels of infection in humans around the lake that can reach up to 80% in school children [46]. Most of the snails were infected with S. mansoni, which was about 8 times more common and more widespread than S. rodhaini. At every site where S. rodhaini was collected, S. mansoni was also collected, but S. mansoni was the sole species collected at 7 of the 14 sites. Also, S. rodhaini was not collected during a large part of the entire sampling period, while S. mansoni was present at some sites during all collection periods. The difference in the abundance and distribution of the species likely is due to differential definitive host use. Schistosoma mansoni primarily infects humans, which generally have larger, less subdivided, and more widespread populations than do rodents, the putative definitive hosts for S. rodhaini. Also, humans, and therefore their worms, are much longer lived than rodents and their worms, and serve as a more stable reservoir that continuously passes eggs and maintains the population. This difference is also reflected in the patterns of genetic diversity in that S. rodhaini showed little variation relative to S. mansoni, even when sample sizes are taken into account, reflecting a small population size for S. rodhaini that potentially has been bottlenecked in the past. Although S. mansoni outnumbered S. rodhaini in terms of numbers of infected snails, there was no difference in the number of cercariae produced by either species per infected snail and this number was not influenced by snail size.
Temporal patterns of prevalence were not obvious in the data, but prevalence varied spatially from 0.11–3.65% at positive sites, with the highest levels of infection occurring at Car Wash site. Although snails are in relative low abundance here due to the less than optimal habitat due to the clearing of vegetation for washing cars, human activity and fecal material are abundant so that the snails that are there are likely to be infected, including with multiple genotypes: 8 of the 21 snails with multiple infections were collected at this site. We also collected at two additional sites that were approximately 210 m and 585 m along the shore from the Car Wash site, Tilapia Beach and Powerhouse. Infection prevalence declined the further the sites were from the Car Wash site, even though snails are much more common at these sites.
Both species were overdispersed in their snail hosts, a pattern that is typical for schistosome populations in snails when prevalence is low [43]. One of the factors that likely leads to the observed pattern is the aggregation of miracidia in microhabitats occupied by particular snails [47] and low probability of contact between miracidia and snails since infection is relatively rare in this system. The fact that mean intensity and prevalence are positively correlated also suggests that probability of encounter plays a large role in determining parasite distribution, or in other words, some snails are “unlucky” and happen to be in the microhabitat where feces are deposited and eggs are hatching. Excess of multiple infections can also be explained by variability in susceptibility of infection of individual snails. Some individuals may be more susceptible or “worm-prone” and are thus likely to acquire multiple genotypes, while other snails are resistant and acquire none. Also, acquired susceptibility of snails could also lead to an excess of multiple infections. In this case, a snail that acquires one genotype becomes more susceptible to additional infections. On the other hand, lack of multiple infections can be explained by probability of encounter, differential compatibility between hosts and parasites, acquired resistance, and competition [8], [48]–[50]. One potential limitation of the methodology used in this study is the possibility of underestimating the number of genotypes that infect a snail. If rare genotypes occur in the sample (in which case they would be difficult to detect by any method) or if certain genotypes are rare due to low infectivity to mice, they may not be detected using our methodology. However, with a minimum sample of 16 worms, and a mean of 34.1 worms sampled per snail, this error likely is low.
The schistosome populations are structured in a way that leads to snail co-species infection more commonly than expected by random infection. Interestingly, two of these snails were also infected with multiple genotypes of one of the species so that the three snails harbored 2, 3, or 4 total genotypes. This result could be explained by the unlucky snail hypothesis mentioned above since microhabitats that are hotspots of transmission for one species could also be a hotspot for the other species. The “Worm-Prone” and the “Acquired Susceptibility” hypotheses mentioned above could also explain this pattern, but would require interspecific facilitation, a phenomenon not unknown in trematode-snail interactions [48]. Experimental infections of snails with one or both species are underway to distinguish among these possibilities. It is also possible that coinfections of definitive hosts play an important role in determining community structure at the snail level because the progeny of both species would be deposited together in the same microhabitat. Our preliminary data from worm burdens of rodents in the region have revealed only one individual that was infected with S. rodhaini, and that individual also was infected with S. mansoni.
Circadian cercarial release cycles were strongly tied to the light/dark cycle in that S. mansoni began to emerge as light intensity increased with the start of the daylight period, and S. rodhaini emerged immediately before and after the daylight period. Peak cercarial emergence of S. mansoni occurred earlier in the 24 hour cycle than most previously studied populations that typically undergo peak emergence when light intensity is the greatest, around noon or later, although this characteristic is known to vary among populations [29],[51],[52]. The bimodal cercarial release pattern of S. rodhaini has not been reported previously, and only twilight emergence was reported from populations from Burundi and Uganda [27],[29]. A possible morning peak of emergence in a Ugandan isolate of S. rodhaini was reported by Fripp [53]; however, his results are unclear because the snails were not monitored over a 24 hour period. In the present study, emergence of S. rodhaini varied among individuals and among replicates of individuals in the number of emergence peaks that occurred. In some cases both peaks occurred, but in others, only one peak occurred. Intraspecific differences in emergence time may correspond to differential definitive host use as this characteristic is likely selected for by the time that definitive hosts are present in the water and available for transmission [17],[54]. Therefore, we suspect that in Kenya S. rodhaini infects a host or group of hosts that are most active in the water just after sunset and right before sunrise.
Three snails were coinfected with both S. mansoni and S. rodhaini, and data from the cercarial emergence profiles of one of these snails indicate that the presence of each species does not influence the other's cercarial release patterns, which is consistent with results from other studies that have examined snails infected with both S. haematobium and S. bovis [55] or with different populations or “strains” of S. mansoni [56]. However, the data from the adults obtained from infections with mice also suggest that S. mansoni emergence is not influenced by coinfection, but S. rodhaini emergence may be because more adults of one genotype of this species were obtained from mice infected with cercariae that emerged between 9:00 and 15:00 hours than adults of S. mansoni. This result is unexpected since this is not the typical emergence time for S. rodhaini. Also, it is anticipated that mechanisms that separate the temporal emergence of each species would evolve particularly if they coinfect the same individual snail host because cercariae released concurrently are likely to infect the same definitive host individuals, thus potentially leading to hybridization. An alternative explanation to the observed results is that the actual number of adults of each species may be biased due to infection success since S. rodhaini may be better adapted to rodents, which are their presumed principal definitive hosts in nature. However, even if the proportions are biased, the data still indicate that the two species are emerging from snails concurrently.
The proportions of genotypes that emerged from snails infected with multiple genotypes varied among circadian cercarial emergence replicates (typically 1 week apart) for about half of the snails examined. Replacement of one predominant genotype by another was the most common pattern detected. It is hypothesized that infection of these snails by the different genotypes occurred sequentially with a large time interval between infections so that one genotype has developed and produces cercariae before the other has developed to the same stage. Possible complete replacement of genotypes was only detected in two snails, but was confounded by small sample sizes of worms and not included in the statistical analyses. An alternative explanation is that since cercarial production occurs in cohorts [57], the genotypes are producing their cohorts asynchronously leading to a pattern that appears to be replacement particularly when only 2 replicates of data are collected. However, in all 7 of the snails where 3 or more replicates were performed, the genotype in majority did not alternate and instead followed a pattern of replacement. The alternative to a replacement pattern was a constant pattern in which the proportions of genotypes did not differ among replicates. This constant pattern may be indicative of infections that were acquired simultaneously and are therefore at the same stage of development within the snail. Interestingly, within these infections the proportions of genotypes were mostly skewed, with the most even ratio being 61:38. This skew suggests that there are other mechanisms besides timing of infection that affect cercarial output possibly including competition between genotypes or variation in compatibility of snail and schistosome genotypes that directly affects cercarial production. These mechanisms are best addressed experimentally to determine the roles of infection timing and competition on genotype “success”, and can be performed to remove the effect of infection bias that may occur when the cercariae are introduced into mice.
Among the factors examined, this study revealed no evidence for ecologically induced isolating mechanisms that prevent S. mansoni and S. rodhaini from encountering one another and hybridizing. These species overlap on a microgeographic scale (individual sites and individual snails) and also temporally both on a seasonal scale and a circadian scale. Even though the emergence peaks of the cercariae do not directly overlap, the cercariae of these two species certainly overlap to some degree since the cercariae remain in the water column and infective for up to 9 hours, and therefore it is difficult to imagine how this would effectively isolate the two species. Also, competition within or among individual snail hosts does not seem to play a large role since coinfections were more common than expected by random infection. If anything, this observation in conjunction with the fact that S. rodhaini was only found in habitats also occupied by S. mansoni, suggests a pattern of co-occurrence as opposed to isolation. The number of cercariae produced per individual snail did not differ between the species; however, if both species share the same host pools, and if there are no strong mating barriers, it is surprising that S. mansoni has not driven S. rodhaini to extinction through hybridization since snails infected with the former species are eight times more common. However, it is possible that our sampling area represents the edge of the range of S. rodhaini and sampling throughout the Rift Valley may reveal larger, more stable populations that disperse to less ideal habitats through movement of snail or mammal hosts. However, the lack of genetic diversity suggests that migration from larger populations is not occurring on a regular basis.
It is unknown how long S. mansoni and S. rodhaini have been in contact in Kenya and if their original divergence was due to sympatric or allopatric speciation. If the latter has occurred and we are witnessing relatively recent secondary contact, then this situation seemingly parallels one occurring in Cameroon in which S. intercalatum is thought to be endangered due to its interactions with S. haematobium and S. mansoni [58]. Decline of S. intercalatum has occurred in recent years (1968-present) and is directly correlated with the introduction of S. haematobium in the region [58]. However, the molecular data suggest S. mansoni and S. rodhaini diverged approximately 2.8 million years ago [3], and it seems likely that they have coexisted in the Lake Victoria basin for a long time. The most likely isolating mechanism separating the two species is the difficulty of S. rodhaini in infecting non-human primates [15] and presumably humans as well, and the preponderance of S. mansoni infections in humans. We have collected both species in the same rodent hosts (unpublished observations) but the relative frequency with which such coinfections occur may be insufficient to break down the genetic differences between the two species, or mate recognition systems may hinder interspecific reproduction when they do encounter each other in a host. What is still lacking is a full understanding of the definitive hosts used by S. rodhaini to propagate itself, whether these hosts are routinely colonized by S. mansoni, and whether the species will hybridize if they encounter each other in the same host. Future monitoring of schistosome populations in Western Kenya and further studies on introgressive hybridization will give further insight on the interactions between these species. |
10.1371/journal.pntd.0000712 | Immunity to Lutzomyia intermedia Saliva Modulates the Inflammatory Environment Induced by Leishmania braziliensis | During blood feeding, sand flies inject Leishmania parasites in the presence of saliva. The types and functions of cells present at the first host-parasite contact are critical to the outcome on infection and sand fly saliva has been shown to play an important role in this setting. Herein, we investigated the in vivo chemotactic effects of Lutzomyia intermedia saliva, the vector of Leishmania braziliensis, combined or not with the parasite.
We tested the initial response induced by Lutzomyia intermedia salivary gland sonicate (SGS) in BALB/c mice employing the air pouch model of inflammation. L. intermedia SGS induced a rapid influx of macrophages and neutrophils. In mice that were pre-sensitized with L. intermedia saliva, injection of SGS was associated with increased neutrophil recruitment and a significant up-regulation of CXCL1, CCL2, CCL4 and TNF-α expression. Surprisingly, in mice that were pre-exposed to SGS, a combination of SGS and L. braziliensis induced a significant migration of neutrophils and an important modulation in cytokine and chemokine expression as shown by decreased CXCL10 expression and increased IL-10 expression.
These results confirm that sand fly saliva modulates the initial host response. More importantly, pre-exposure to L. intermedia saliva significantly modifies the host's response to L. braziliensis, in terms of cellular recruitment and expression of cytokines and chemokines. This particular immune modulation may, in turn, favor parasite multiplication.
| Transmission of Leishmania parasites occurs during blood feeding, when infected female sand flies inject humans with parasites and saliva. Chemokines and cytokines are secreted proteins that regulate the initial immune responses and have the potential of attracting and activating cells. Herein, we studied the expression of such molecules and the cellular recruitment induced by salivary proteins of the Lutzomyia intermedia sand fly. Of note, Lutzomyia intermedia is the main vector of Leishmania braziliensis, a parasite species that causes cutaneous leishmaniasis, a disease associated with the development of destructive skin lesions that can be fatal if left untreated. We observed that L. intermedia salivary proteins induce a potent cellular recruitment and modify the expression profile of chemokines and cytokines in mice. More importantly, in mice previously immunized with L. intermedia saliva, the alteration in the initial inflammatory response was even more pronounced, in terms of the number of cells recruited and in terms of gene expression pattern. These findings indicate that an existing immunity to L. intermedia sand fly induces an important modulation in the initial immune response that may, in turn, promote parasite multiplication, leading to the development of cutaneous leishmaniasis.
| The intracellular protozoan parasites of the Leishmania species are transmitted to vertebrate host through the bites of sand flies. Within the vertebrate host, Leishmania parasites reside in phagocytes and induce a spectrum of diseases ranging from a single self-healing cutaneous lesion to the lethal visceral form. It is currently estimated that leishmaniasis affects two million people per year worldwide [1].
Leishmania braziliensis, the main causative agent of cutaneous leishmaniasis (CL) in Brazil, can be transmitted to the human host by the bite of the sand fly Lutzomyia intermedia. [2], [3]. Several studies have shown that pre-exposure to saliva or to bites from uninfected sand flies results in protection against subsequent infection with Leishmania major [4]–[7], Leishmania. amazonensis [8], and Leishmania chagasi [9]. On the contrary, pre-exposure to Lutzomyia intermedia saliva enhanced infection with L. braziliensis in the mouse model; disease exacerbation was correlated with generation of a Th2 response evidenced by a reduction in the IFN-γ/IL-4 ratio [10]. Importantly, individuals with active CL showed higher humoral immune responses to L. intermedia saliva compared with control subjects, a finding also demonstrated with Old World CL [11] . These data indicate an association between disease and immune response to L. intermedia saliva in humans.
In the case of L. intermedia, the lack of protection observed following pre-exposure to saliva in the murine model may be related to differences in the initial inflammatory response induced by the salivary proteins. Several studies have shown the potential of salivary antigens from Lutzomyia longipalpis, Phlebotomus duboscqi, Phlebotomus papatasi and Phlebotomus ariasi to modulate cell recruitment and production of immune response mediators [12]–[17] however, little is known regarding these effects when using L. intermedia saliva. Our group has previously shown that pre-treatment of human monocytes with L. intermedia followed by L. braziliensis infection led to a significant increase in TNF-α, IL-6, and IL-8 production [18], indicating the ability of L. intermedia saliva to alter the inflammatory milieu. To gain further information regarding the events associated with the initial host response to L. intermedia saliva, we employed the air pouch model of inflammation. This model simulates inoculation of the sand fly in a closed environment and allows for subsequent analysis of inflammatory parameters and mediators induced in vivo by distinct stimuli [19]. Using this model, we showed that saliva from L. longipalpis rapidly induced CCL2 expression and macrophage recruitment, in synergy with L. chagasi parasites, in BALB/c mice [20]. Here we describe the ability of L. intermedia salivary gland sonicate (SGS) to modulate the host immune response in naïve and in SGS-sensitized mice. We have demonstrated that L. intermedia salivary proteins induce neutrophil recruitment and modulate cytokine and chemokine expression. Crucially, a downregulation in CXCL10 paralleled by an increase in IL-10 expression was observed in SGS-sensitized mice stimulated with saliva+L. braziliensis. This correlates with disease exacerbation previously observed in mice immune to L. intermedia SGS and challenged with L. braziliensis [10].
Leishmania braziliensis promastigotes (strain MHOM/BR/01/BA788 [21]) were grown in Schneider medium (Sigma Chemical Corporation, St. Louis, MO, USA) supplemented with 100 U/ml of penicillin, 100 µg/ml of streptomycin, 10% heat-inactivated fetal calf serum (all from Invitrogen, San Diego, CA, USA), and 2% sterile human urine. Stationary-phase promastigotes from second passage culture were used in all experiments.
Female BALB/c mice (6–8 weeks of age) were obtained from CPqGM/FIOCRUZ Animal Facility where they were maintained under pathogen-free conditions. All procedures involving animals were approved by the local Ethics Committee on Animal Care and Utilization (CEUA—CPqGM/FIOCRUZ).
Adult Lutzomyia intermedia sand flies were captured in Corte de Pedra, Bahia, and were used for dissection of salivary glands. Salivary glands were stored in groups of 20 pairs in 20 µl NaCl (150 mM)-Hepes buffer (10 mM; pH7.4) at −70°C. Immediately before use, salivary glands were disrupted by ultrasonication in 1.5-ml conical tubes. Tubes were centrifuged at 10,000×g for two minutes, and the resultant supernatant—salivary gland sonicate (SGS)—was used for the studies. The level of lipopolysaccharide (LPS) contamination of SGS preparations was determined using a commercially available LAL chromogenic kit (QCL-1000; Lonza Biologics, Portsmouth, NH, USA); LPS concentration was <0.1 ng/ml.
BALB/c mice (groups of five to six) were immunized three times with SGS (equivalent to one pair of salivary glands) in 10 µl of PBS in the dermis of the right ear using a 27.5 G needle. Immunizations were performed at two-week intervals. Control mice were injected with PBS. Development of an immune response against L. intermedia saliva was confirmed by ELISA as previously described [10].Immune sera were pooled from SGS-immunized mice and employed in neutralization experiments. Immune mice were employed in air pouch experiments.
Air pouches were raised on the dorsum of anesthetized BALB/c mice (groups of five to six) by injection of 3 ml of air, as described elsewhere [22]. Air pouches were inoculated with either one of the following stimuli: L. intermedia SGS (equivalent to one pair of salivary glands/animal); L. intermedia SGS pre-incubated with a pool of anti-SGS immune sera (SGS+50 µl of immune serum pre-incubated for one hour at 37°C); a pool of anti-SGS immune sera alone; stationary-phase L. braziliensis promastigotes (105 parasites); or L. braziliensis+SGS. Air pouches in control mice were injected with endotoxin-free saline (negative control) or with LPS (Calbiochem, San Diego, CA, USA) (20 µg/ml; positive control). After twelve hours, animals were euthanized and pouches washed with 5 ml of endotoxin-free saline for collection of exudates containing leukocytes. Lavage fluids were washed, and cell pellets were resuspended in saline, stained in Turk's solution, and counted in a Neubauer hemocytometer. Cells were cytoadhered to glass slides using Shandon cytospin2 and stained with hematoxylin and eosin to determine proportions of monocytes/macrophages, neutrophils, lymphocytes, basophils, and eosinophils. Air pouch lining tissue was placed in 5–10 volumes of RNAlater (Ambion Inc., Austin, TX, USA), and samples were stored at −80°C.
Total RNA was extracted from the air pouch lining tissue using the RNeasy Protect Mini Kit (Qiagen, Inc., Santa Clara, CA, USA) according to manufacturer's instructions. The resulting RNA was resuspended in 20 µl diethyl pyrocarbonate (DEPC)-treated water and stored at −80°C until use. cDNA synthesis for detection of cytokine mRNA was performed after reverse transcription (Im Prom-II™ reverse transcription system). Real-time PCR was performed in triplicate on the Abi Prism 7500 (Applied Biosystems, Inc., Fullerton, CA, USA); thermal cycle conditions consisted of a two-minute initial incubation at 50°C followed by ten-minute denaturation at 95°C and 50 cycles at 95°C for 15 seconds and 60°C for one minute each. Each sample and the negative control were analyzed in triplicate for each run. The comparative method was used to analyze gene expression. Chemokine or cytokine cycle threshold (Ct) values were normalized to GAPDH expression as determined by ΔCt = Ct (target gene)−Ct (GAPDH gene). Fold change was determined by 2−ΔΔCt, where ΔΔCt = ΔCt (target)−ΔCt (saline) [23]. The following primers were employed: GAPDH (Forward: 5′-TGTGTCCGTCGTGGATCT GA-3′; Reverse: 5′-CCTGCTTCACCACCTTCTTGA-3′); CCL2 (Forward: 5′-CAGGTC CCTGTCATGCTTCTG-3′; Reverse: 5′-GAGCCAACACGTGGATGCT-3′) ; CCL3 (Forward: 5′-TCTTCTCAGCGCCATATGGA-3′; Reverse: 5′-CGTGGAATCTTCCGG CTGTA-3′); CCL4 (Forward: 5′-TGCTCGTGGCTGCCTTCT-3′; Reverse: 5′-CAGGAA GTGGGAGGGTCAGA-3′); CXCL1: (Forward: 5′-CCGAAGTCATAGCCACACTCAA-3′; Reverse: 5′-AATTTTCTGAACCAAGGGAGCTT-3′); CXCL10: (Forward: 5′-GGACGG TCCGCTGCAA-3′; Reverse: 5′-CCCTATGGCCCTCATTCTCA-3′); IFN-γ (Forward: 5′-CTACACACTGCATCTTGGCTTTG-3′; Reverse: 5′-TGACTGCGTGGCAGTA-3′); TNF-α (Forward: 5′-GGTCCCCAAAGGGATGAGAA-3′; Reverse: 5′-TGAGGGTCT GGGCCATAGAA-3′); and IL-10 (Forward: 5′-CAGCCGGGAAGACAATAACTG-3′; Reverse: 5′-CGCAGCTCTAGGAGCATGTG-3′). Primers were designed using Primer Express Software (Applied Biosystems).
BALB/c mice (n = 5) were intradermally immunized with L. intermedia SGS (equivalent to one pair of salivary glands) or injected with PBS three times in the right ear at two-week intervals. After the third injection, pre-sensitized or control animals were intradermally inoculated with L. intermedia SGS, in the opposite (left) ear dermis. Twenty-four and forty-eight hours after SGS injection, animals were euthanized and the ear was biopsied and stored in 10% neutral buffered formalin. Ears were mounted in paraffin blocks, sectioned at 5-µm intervals, and stained with hematoxylin and eosin for histologic analysis. Paraffin-embedded sections of ears fixed in 10% neutral buffered formalin were used for immunohistochemistry. Myeloperoxidase rabbit anti-mouse (Dako, Carpenteria, CA, USA) was used at 1∶1000 dilution. A secondary biotinylated goat anti-rabbit antibody was used at 1∶500 for 15 minutes (Vector Laboratories, Burlingame, CA, USA) and detected by R.T.U. Vectastin Elite ABC reagent (Vector Laboratories) and DAB chromagen.
Data are presented as the mean with 95%CI. The significance of the results was calculated using nonparametric statistical tests: two-sided Mann-Whitney for comparisons between two groups; Kruskal-Wallis followed by Dunn's multiple comparison test for comparisons between three groups. Analyses were conducted using Prism (GraphPad Software Inc., San Diego, CA, USA); a P-value of <0.05 was considered significant.
We initially studied the cellular recruitment induced by L. intermedia SGS inoculation. Air pouches were induced in BALB/c mice and subsequently probed with different stimuli: endotoxin-free saline; L. intermedia SGS; or LPS. L. intermedia SGS induced a significant increase in leukocyte accumulation in the air pouch compared with saline injection (Figure. 1A). Most cells recruited by inoculation of L. intermedia SGS into air pouches were neutrophils, followed by monocytes (Figure 1B). LPS inoculation was used as a positive control for cell recruitment and, as expected, led to a predominant recruitment of neutrophils (Figure 1B). Moreover, inoculation of L. intermedia SGS did not lead to significant changes in either eosinophil or lymphocyte recruitment.
To confirm that the effect of L. intermedia SGS on leukocyte accumulation within air pouches was specific, we pre-incubated SGS with anti-SGS immune sera obtained from mice immunized with L. intermedia SGS (data not shown, [10] ). Pre-incubation of L. intermedia SGS with anti-SGS immune sera inhibited leukocyte accumulation induced by L. intermedia SGS by 56% (Figure 2A), whereas air-pouch inoculation with immune sera alone led to a cellular recruitment similar to that induced by saline (Figure 2A). Notably, the significant decrease in cellular recruitment following incubation of L. intermedia SGS with antisera was associated with a significant reduction (81%) in the number accumulating neutrophils (Figure 2B). Recruitment of monocytes, lymphocytes, and eosinophils, however, remained unchanged (Figure 2B).
L. intermedia SGS was able to induce a significant increase in leukocyte recruitment in the air-pouch model of inflammation when compared with saline (Figure 1). This effect was particularly powerful on neutrophil migration and was abrogated when SGS was pre-incubated with anti-SGS-specific antiserum (Figure 2B). We then investigated the initial inflammatory response in mice that had been previously immunized with L. intermedia SGS. Air pouches were raised on the back of immune mice, and pouches were stimulated with L. intermedia SGS. Control mice were injected with endotoxin-free PBS. Mice immunized with L. intermedia SGS showed a significant increase in the total number of leukocytes (Figure 3A) accumulating in the air pouch compared with control mice injected with PBS. Surprisingly, this increase was associated with an accumulation of neutrophils (53%) migrating to the air pouch (Figure 3B), whereas migration of monocytes, eosinophils, and lymphocytes remained unaltered in SGS-immunized mice compared with control mice injected with PBS. Because chemokines, together with adhesion molecules, are key controllers of leukocyte migration, we tested for chemokine expression in the pouch lining tissue. CXC-class chemokines act mainly on neutrophils, whereas CC-class chemokines act on a larger group of cells including monocytes, eosinophils, and lymphocytes. Additionally, cytokines have long been recognized as key elements in the host response against Leishmania (reviewed in [24]. As shown in Figure 3C, expression of CXCL1, CCL2, and CCL4 was significantly upregulated in SGS-immunized mice compared with control mice injected with PBS. Moreover, SGS-immune mice also displayed a significant increase in TNF-α expression without significant modulation in expression of IL-10 or IFN-γ (Figure 3D).
We then investigated whether the neutrophil accumulation effect observed in air pouches raised in SGS-immune mice and stimulated with SGS could be replicated in the ear dermis. As shown in Figure 4, ear sections from control mice injected with PBS showed very few inflammatory cells at either 24 or 48 hours after SGS challenge. In contrast, ear sections from SGS-immunized mice displayed, 24 hours after SGS-challenge, numerous polymorphonuclear and few mononuclear cells (Figure 4); at 48 hours, the inflammatory infiltrate was further increased. Presence of neutrophils was confirmed by myeloperoxidase staining and was not observed in control mice injected with PBS.
Because SGS-immune mice displayed enhanced neutrophil recruitment, we investigated whether the presence of L. braziliensis, the parasite transmitted by L. intermedia sand flies, would exert any effect in this outcome. Therefore, air pouches were raised on the back of either naïve or SGS-immunized mice and pouches were stimulated with L. braziliensis (Lb) or L. braziliensis+L. intermedia SGS (Lb+SGS). In naïve mice, we did not detect significant differences in the number of accumulating leukocytes (Figure 5A) or in the recruited cell subsets (Figure 5B) following inoculation with Lb or Lb+SGS (Figure 5B). On the contrary, in SGS-immunized mice, Lb+SGS led to a robust and significant increase in the number of accumulating leukocytes compared with Lb alone (Figure 5C). The increase in the number of leukocytes was due to accumulation of neutrophils in the pouches upon inoculation of Lb+SGS (Figure 5D). There was no significant modulation in the recruitment of monocytes, eosinophils, or lymphocytes in naïve or SGS-immunized mice upon inoculation of Lb or Lb+SGS (Figure 5B and 5D, respectively).
We then investigated the modulation in cytokine and chemokine expression induced by L. braziliensis alone or in the presence of saliva in naïve and in SGS-immunized mice. In naïve mice, pouch stimulation with Lb+SGS induced a significant increase in CXCL10 and CCL2 expression compared with pouch inoculation with Lb alone (Figure 6A). In SGS-immunized mice, chemokine expression was over two-fold higher compared with naïve mice (Figure 6B). More important, pouch inoculation with Lb+SGS led to a different pattern of chemokine expression as indicated by a significant upregulation in expression of CXCL1, CCL3, and CCL4 compared with inoculation of Lb alone (Figure 6B). Of note, in SGS-immunized mice, pouch inoculation with Lb+SGS led to a significant decrease in CXCL10 expression (Figure 6B) as opposed to naïve mice, in which pouch inoculation with Lb+SGS led to upregulation in CXCL10 expression (Figure 6A). Regarding cytokine expression, naïve mice displayed augmented expression of both TNF-α and IL-10 upon pouch inoculation with Lb+SGS (Figure 6C) compared with inoculation with Lb alone. In SGS-immunized mice, stimulation with Lb+SGS led to specific increase in IL-10 expression (Figure 6D). In this same group, inoculation of Lb+SGS was not capable of significantly decreasing expression of IFN-γ and TNF-α (Figure 6D).
Sand flies use saliva to manipulate host homoeostasis, favoring the acquisition of a blood meal. These sand fly salivary molecules modify the skin microenvironment and this, in turn, may favor infection by Leishmania parasites (rev. in [25]). Indeed, we previously observed that L. intermedia SGS-immune mice show a higher disease burden when challenged with L. braziliensis [10]. To gain understanding of the early events associated with inoculation of L. intermedia sand fly saliva, we evaluated leukocyte migration and chemokine/cytokine expression induced in the air-pouch model of inflammation. Importantly, the L. intermedia sand fly is the vector of L. braziliensis [2], [3], the main etiologic agent of cutaneous leishmaniasis.
Injection of L. intermedia SGS into air pouches led to a significant increase in the recruitment of neutrophils and monocytes, corroborating previous findings that both of these cell populations are recruited to the site of saliva inoculation [7], [10], [12], [17], [20], [26]. Indeed, the initial events following saliva inoculation have recently been explored by in vivo live imaging [27]. It was shown that sand fly biting leads to potent neutrophil migration and that these cells are efficiently infected by L. major, indicating that neutrophils may serve as host cells for Leishmania in the early phase of infection, as previously suggested [28], [29]. Differently from L. longipalpis saliva [20], L. intermedia did not lead to accumulation of eosinophils, which are strongly related to mosquito bites and allergies. This distinction in the cellular recruitment induced by L. intermedia vs. L. longipalpis saliva may be explained by variation in the salivary components within sand flies, such as maxadilan, present only in L. longipalpis [30], and hyaluronidase, present in both L. longipalpis and various species within the genus Phlebotomus [31], [32].
Pre-incubation of L. intermedia SGS with specific antisera was able to partially neutralize the leukocyte-recruiting effects of SGS, mainly decreasing the number of accumulating neutrophils, without a significant effect on monocytes. Similarly, Belkaid et al. showed that anti-SGS antibodies could neutralize the ability of P. papatasi SGS to enhance L. major infection in BALB/c mice [5]; however, SGS-immune mice showed an enhanced neutrophil recruitment upon stimulation with SGS in pre-sensitized animals. The actual levels of anti-saliva antibodies into the pouch exudates are unknown and may not be sufficient to neutralize the in vivo effects of the saliva. Another possibility for the in vivo findings is that salivary molecules are able to trigger cytokine/chemokine expression, despite the presence of neutralizing antibodies, leading to enhanced neutrophil recruitment.
Leukocyte recruitment to sites of inflammation is a key event in both innate and adaptive immunity, and chemokines are major players that regulate the sequential steps of leukocyte rolling, firm adherence, and transmigration. In this sense, we tested for CXC-class chemokines, that act mainly on neutrophils, and CC-class chemokines that act on a larger group of cells including monocytes, eosinophils, and lymphocytes. In mice sensitized and stimulated with L. intermedia SGS, we saw increased neutrophil recruitment and significant upregulation in the expression of CXCL1, CCL2, and CCL4. Indeed, CXC chemokines, such as CXCL1, are critical molecules for neutrophil recruitment [33], and CXCL1 is also a dominant chemokine in murine inflammatory responses [34]. CCL2 mediates neutrophil adherence and transmigration, a process dependent on activation of mast cells and release leukotrienes and PAF [35], and CCL4 expression has been associated with a type 1 immune response [36]. Therefore, the enhanced neutrophil chemotaxis in SGS-immunized mice may result from a concomitant upregulation in CXCL1 and CC chemokines (CCL2 and CCL4) and may be further amplified by upregulation in TNF-α, favoring a pro-inflammatory environment as shown by upregulation in CCL4 expression. Indeed, OVA-immunized mice displayed increased neutrophil migration upon antigen stimulation [37]; this effect was dependent on the release of TNF-α, and leukotriene B(4) [38] and mediated by CCL3 [39] . Increased neutrophil recruitment was also observed when SGS immunization was conducted in the ear dermis: SGS challenge led to development of an inflammatory reaction characterized by the presence of numerous neutrophils, confirming previously published results [10]. Similarly, exposure of mice to the bites of uninfected L. longipalpis, the vector of L. chagasi, induced an analogous effect [12]. In addition, it has been shown that PSG, the proteophosphoglycan-rich gel secreted by L. mexicana, also leads to potent neutrophil and macrophage recruitment [40].
In naïve mice, sand fly saliva [4], [5], [41]–[43] and fPPG, a component in PSG [44],favor the initial establishment of Leishmania infection. In naïve mice, pouch stimulation with L. braziliensis+SGS was unable to alter the cellular recruitment induced by L. braziliensis alone (Figure 5A), as opposed to previous studies conducted with L. longipalpis SGS+L. chagasi [20] or with L. major+L. longipalpis SGS [45]; however, pouch stimulation with Lb+SGS induced significant upregulation in the expression of CCL2, CXCL10, TNF-α, and IL-10 (Figure 6A). Accordingly, experimental infection with L. braziliensis leads to increased leukocyte recruitment, CCL2 and CXCL10 expression [46], and production of IL-10 [21]. More recently, increase CXCL10 and IL-10 expression were observed upon infection of human monocytes with L. braziliensis [47]. Therefore, we can suggest that, although presence of sand fly saliva does not add to the cellular recruitment induced by L. braziliensis, salivary antigens modulate the microenvironment, which may favor parasite establishment as previously suggested [48]. Here we were unable to determine parasite load in cellular exudates obtained from stimulated pouches; however, earlier work from our group also showed that pre-treatment of human monocytes with L. intermedia SGS followed by L. braziliensis infection led to a significant increase in TNF-α production without significant augmentation in the parasite load [18].
Pre-exposure to L. longipalpis [9] or P. papatasi saliva [5] or to bites from uninfected P. papatasi [49] results in protection against leishmaniasis; however, pre-exposure to L. intermedia saliva does not generate a protective effect upon a challenge infection with L. braziliensis+L. intermedia SGS [10] although SGS immunized mice do show a significantly lower initial parasite burden after challenge with L. braziliensis+SGS. We hypothesized that this early control in parasite load could be exerted by inflammatory cells (mono and polymorphonuclear cells) that are recruited following stimulation with saliva [10]. Indeed, the results herein show that SGS-immune mice displayed increased leukocyte recruitment, with a marked neutrophil influx (Figure 3) and a similar finding was observed upon inoculation of Lb+SGS (Figure 5). We have recently shown that macrophages and neutrophils collaborate towards L. braziliensis elimination from infected macrophages [50]. Therefore, the current results support our previous hypothesis that an initial inflammatory environment may account for the early control of parasite load in SGS-immunized mice upon challenge with Lb+SGS. This control, however, is limited and L. braziliensis multiplication is later on observed, probably resulting from the pathogen favorable immune response (lower IFN-γ to IL-4 ratio) developed in SGS-immunized mice [10]. Indeed, in the present work, SGS-immunized mice stimulated with Lb+SGS showed decreased CXCL10 expression paralleled with an increased IL-10 expression. Presence of CXCL10 is seen in many Th1-type inflammatory diseases, where it is thought to play an important role in recruiting activated T cells into sites of tissue inflammation [51]. IL-10, on the contrary, is associated with a non-healing L. major infection [52] and L. major persistence [53]. Consequently, lack of CXCL10 and presence of IL-10 may create a de-activating environment, favoring L. braziliensis expansion in the context of SGS-immunized mice.
We cannot exclude that the increased neutrophil recruitment observed in SGS-immunized mice may also be relevant to the “Trojan horse” model, as documented for L. major infection [29], in which parasites within neutrophils are silently transferred to macrophages and successfully establish infection. Indeed, the early influx and persistence of neutrophils after sand fly transmission of L. major appears critical for the development of cutaneous disease [27]. Additionally, L. major internalization delays the neutrophil apoptotic death program and induces CCL4 release, which recruits macrophages to the infection site [29], [54]. Indeed, increased CCL4 expression was observed upon inoculation of Lb+SGS.
Collectively, our data show that in naïve mice, inoculation of L. intermedia saliva plus L. braziliensis modifies the initial inflammatory environment as seen by increased neutrophil recruitment and IL-10 and TNF-α expression. Crucially, in mice sensitized with L. intermedia saliva and stimulated with L. braziliensis, these initial events are further modulated, as seen by a specific decrease in CXCL10 and a persistently increased IL-10 expression. We can speculate that the resulting effects leads to the higher disease burden as previously documented [10]. This study again shows important effects of the L. intermedia sand fly and L. braziliensis interaction. More important, it emphasizes how the immune response to sand fly may exert an under-appreciated role in endemic areas. We are currently characterizing L. intermedia salivary antigens to further identify the components that may induce the effects described here.
|
10.1371/journal.pgen.1007258 | The long-range interaction map of ribosomal DNA arrays | The repeated rDNA array gives rise to the nucleolus, an organelle that is central to cellular processes as varied as stress response, cell cycle regulation, RNA modification, cell metabolism, and genome stability. The rDNA array is also responsible for the production of more than 70% of all cellular RNAs (the ribosomal RNAs). The rRNAs are produced from two sets of loci: the 5S rDNA array resides exclusively on human chromosome 1 while the 45S rDNA arrays reside on the short arm of five human acrocentric chromosomes. These critical genome elements have remained unassembled and have been excluded from all Hi-C analyses to date. Here we built the first high resolution map of 5S and 45S rDNA array contacts with the rest of the genome combining over 15 billion Hi-C reads from several experiments. The data enabled sufficiently high coverage to map rDNA-genome interactions with 1MB resolution and identify rDNA-gene contacts. The map showed that the 5S and 45S arrays display preferential contact at common sites along the genome but are not themselves sufficiently close to yield 5S-45S Hi-C contacts. Ribosomal DNA contacts are enriched in segments of closed, repressed, and late replicating chromatin, as well as CTCF binding sites. Finally, we identified functional categories whose dispersed genes coalesced in proximity to the rDNA arrays or instead avoided proximity with the rDNA arrays. The observations further our understanding of the spatial localization of rDNA arrays and their contribution to the architecture of the cell nucleus.
| The repeated ribosomal DNA (rDNA) array gives rise to the nucleolus, an organelle that is involved in key cellular processes such as stress response, cell cycle regulation, RNA modification, and production of more than 70% of all cellular RNAs (the ribosomal RNAs). This critical genome element has remained unassembled and has been excluded from all Hi-C analyses to date. Here we built the first map of 5S and 45S rDNA contacts with the rest of the genome. The map yielded a number of novel results and challenge the expectation that 5S and 45S arrays are close together in the nucleus. The rDNA arrays share common sites of contact across the genome, are biased towards segments of closed, repressed, and late replicating chromatin, and display greater proximity or avoidance to functionally coherent gene sets. The results further our understanding of the rDNA arrays and their localization in the nuclear environment.
| Ribosomal RNAs (rRNAs) are essential components of the cell, and are encoded in the 5S and 45S ribosomal DNA (rDNA) arrays of higher eukaryotes [1–4]. The 5S rDNA array resides on chromosome 1 and encodes the 5S rRNA, whereas the 45S rDNA array resides on five human acrocentric chromosomes and encodes the 18S, 5.8S, and 28S rRNA components of the ribosome [5–7]. The nucleolus, the first recognized nuclear organelle, is the site of 45S rRNA transcription [1, 2, 4, 8]. The lack of homology between the 5S rDNA and the subunits of the 45S rDNAs arrays reflect deep evolutionary separation. For instance, RNA polymerase I is exclusively dedicated to the transcription of the 45S rRNA, while RNA polymerase III transcribes the 5S rRNAs and tRNAs. The distinct RNA polymerase machineries required for transcription of 5S and 45S subunits are a conserved feature of yeasts, plants, fruit flies, and humans. Furthermore, distance to the nucleolus is thought to be relevant for global gene expression. For instance, proximity to the nucleolus can in some cases promote inactivation of certain RNA polymerase II transcribed genes [9], although the observation has not been systematically tested across the genome. Finally, localization of the 5S array has been documented at the periphery of the nucleolus [9, 10], but also away from the organelle [11], with a substantial fraction of cells showing 5S arrays that are localized elsewhere in the nucleus [10]. Uncovering physical contacts between the rDNA arrays and the rest of the genome can expand our understanding of nuclear architecture, nucleolar structure and function, and the mechanism of concerted copy number variation between 5S and 45S rDNA arrays. However, studies of nuclear architecture have largely excluded analyses of spatial interactions with the 5S and 45S rDNA arrays.
Ligation-capture Hi-C sequencing technology [12–14] enabled a revolution in our understanding of nuclear organization with the identification of hundreds of topologically associated domains (TADs). Human TADs span an average 900 KB each and display remarkable conservation with TADs identified in mice. TADs display, moreover, remarkable structural stability through development and when cells are perturbed in gene knockdown experiments [15, 16]. On the other hand, deep sequencing of nucleoli led to the documentation of nucleoli associated DNA (naDNA) and the identification of nucleolus associated domains (NADs) [17–19]. While NADs display size variation spanning multiple orders of magnitude, they are generally large. NADs covering less than 0.1 MB are relatively rare with most NADs around 1 MB or larger. The domains encompass about 5% of the human genome, are represented in all chromosomes, and are now recognized to be stably associated with nucleoli. Analysis of rDNA interactions with Hi-C might provide a complementary approach to localize the rDNA in the nuclear space possibly informing nucleolar interactions with the genome at a different scale than those afforded by analysis of naDNA.
Here we addressed the landscape of long-range rDNA interactions with 16,482,743 reads identified from a total of >15 billion (15,165,355,427) Hi-C reads in five cell types and two cell lines. The data enabled a map of long-range rDNA interactions at 1MB resolution, and the identification of segments displaying statistically significant differential contact density between cells. The map yielded a number of observations and suggest that the 5S and 45S arrays are not as spatially close as typically expected, yet share significant overlap with common contacts elsewhere. Finally, the data uncovered functionally coherent categories whose dispersed genes either coalesce in proximity to the rDNA arrays or avoid proximity with the rDNA arrays.
We investigated human Hi-C data for two cell lines and five cell types; the two cell lines represent the most replicated human Hi-C datasets to date, yet yielded a relatively small number of rDNA informative reads. For instance, we mined 5,356,990,189 high quality Hi-C reads in LCL to identify 13,528,436 reads with at least one end mapped to the 45S rDNA and 105,147 reads with at least one end mapped to the 5S rDNA (S1 and S2 Tables). Similarly, for K562 cells, we mined 903,837,936 high quality Hi-C reads to identify 1,698,063 reads with at least one end mapped to the 45S rDNA and 47,691 reads with at least one end mapped to the 5S rDNA. This represents a 0.25% and 0.19% recovery rate of 45S rDNA reads in shotgun Hi-C in LCL and K562, respectively. These numbers were substantially larger than the meager 0.002% and 0.005% recovery rate for 5S rDNA reads in LCL and K562, respectively. Similar recovery rates were obtained with the other five cell types studied (Table 1). Overall, we uncovered 16,322,538 reads with at least one end mapped to the 45S rDNA and 160,205 reads with at least one end mapped to the 5S rDNA (Table 1). The mining effort illustrates the challenge in recovering rDNA information in shotgun Hi-C experiments. Nevertheless, the data revealed that rDNA contacts are dispersed across the entire genome, with segments differing in the density of rDNA interaction. The maps also revealed that naDNA and rDNA-contacts are not overlapping domains and likely reflect different attributes of the nucleolus/rDNA (S1 Fig).
Here we partitioned human autosomes (Chr 1 to 22) into 2897 segments of 1MB, 2465 and 2658 of which had no evidence of containing a 5S or 45S pseudogene, respectively. Segments containing an rDNA pseudogene were disproportionately found adjacent to centromeric and telomeric regions, and were excluded from all further analyzes. Unsurprisingly, all 1MB segments across all chromosomes displayed evidence of rDNA contact (S1 and S2 Figs). Moreover, at the 1MB scale, we observed good reproducibility between replicates of a cell line using the same restriction enzyme as well as different restriction enzymes, with consistent results across biological replicates and across cell lines/cell types (S3, S4, S5, S6, S7, S8 and S9 Figs). Fig 1 illustrates the distribution of rDNA contact density for 1MB segments before normalization by sequencing effort. The data shows a 5-10-fold variation in the logarithm of the contact density across segments within a cell type. The mean difference in the average contact density among cells reflects variation in the amount of Hi-C data in each cell type. For 45S rDNA contacts all 1MB segments contained appreciable density of contacts in LCL and K562. However, the ESC and ESC-derived cell types (ESC set) displayed a truncated distribution with many segments that contained very few rDNA contacts (Fig 1A). This was due to the lower number of Hi-C reads for those cells (Table 1). The resolution was much worse for the 5S rDNA arrays (Fig 1C). Therefore, the following analyses focused primarily in the data for LCL and K562 cell lines, with the ESC or ESC-derived cells mostly used for comparisons.
Here we first addressed variation in rDNA contact density across cell lines (LCL vs K562 data collected with the same enzyme and protocol). We found 808 segments of 1MB with significantly different Density of Interactions (DI) of 45S rDNA contacts between LCL and K562 (Fig 2; Fig 3, FDR < 0.05; S3 Table), whereas none is identified among biological replicates of LCLs processed with different enzymes (Fig 3). We observed that 350 DI segments displayed increased density in LCL and 458 segments displayed increased density in K562. Among those 808 DI segments, 302 of them displayed a greater than 2-fold difference in contact density between LCL vs K562. Similarly, nearly half of the 224 segments of 1MB in chromosome 1 showed evidence of DI density between LCL and K562 (Fig 3; chromosome 1: 106 segments significantly different, and 118 non-significant bins), with 97 segments displaying greater contact density in LCL and 9 segments containing greater contact density in K562. Chromosome 1 had the largest number of significantly different DI, followed by Chr 13 (89), Chr 9 (63), and Chr 6 (61). Among the five cell types (ESC related), there were 193 segments of 1MB with significantly different DI with the rDNA (FDR<0.05; S4 Table). Finally, we detected a meager 15 segments with evidence of differential DI between LCL and K562 for the 5S rDNA (FDR<0.05); the small number of differential DI likely reflects the many fewer 5S rDNA reads and thus the much-lowered statistical power of this analysis. Similarly, there was not enough Hi-C data to enable statistical analysis of DI with the 5S rDNA among the five ESC related cell types.
We identified 9,595 and 9,864 genes without evidence of a 5S or 45S rDNA pseudogene, respectively. The remaining genes were excluded from all further analyzes. The data showed a continuous distribution of rDNA-gene contact density for the 45S and 5S rDNA (Fig 1B), with much better resolution for the 45S rDNA than for the 5S rDNA (Table 2, Table 3). As expected, the rDNA-gene contact density was correlated with gene length. We have thus calculated the 45S contact density per gene per nucleotide (“Contacts per gene per nucleotide, CPGN”). This removed the correlations between gene length and 45S rDNA contacts and revealed that CPGN for the 45S rDNA arrays was strongly correlated between LCL and K562 (rho = 0.65; P < 0.001). This correlation was stronger than those between LCL and ESC (rho = 0.27; P < 0.001) or between K562 and ESC (rho = 0.34; P < 0.001). The lower correlations with ESC might partially reflect the lower resolution of the ESC contact map with a substantial fraction of genes showing less than 10 reads with rDNA contacts (Table 2). Indeed, although the overall amount of HI-C data was large, the resolution to ascertain 45S rDNA-gene contacts was only sufficient for LCL and K562, the two biological sources with the largest number of Hi-C reads to date. The issue of low rDNA-gene resolution was particularly evident for the 5S rDNA. Out of 9595 genes analyzed for 45S rDNA arrays, there were 67 and 612 genes with zero 5S contacts in LCL and K562, respectively. For the ESC set, however, there were 1745 genes with zero contacts with the 45S rDNA arrays. Out of 9864 genes analyzed for 5S rDNA arrays, there were 5916 and 7494 genes with zero 5S contacts in LCL and K562, respectively. For the ESC set, we observed that greater than 95% of the genes had zero 5S contacts. The density of 5S rDNA-gene contacts was most strongly correlated with gene length (rho > 0.3, P < 0.001), but calculating the 5S contact density per gene per nucleotide (“Contacts per gene, CPGN”) removed the positive association. Among genes with at least one read showing 5S-rDNA contact in both LCL and K562 we found that CPGN is strongly correlated between LCL and K562 (rho = 0.64, P < 0.001). Evidence for a positive association between the density of 45S rDNA contacts and the density of 5S rDNA contacts is also observed in other partitions of the data, and across genes and 1MB segments in both LCL and K562 (Table 4).
Here we tested for variation in rDNA-gene contact density between LCL and K562. For the 45S array, we observed 731 genes with fold change in interaction density >2 for the LCL vs K562 comparison (experiments with the same enzyme and protocol); 97 genes (FDR < 0.05) displayed significantly different DI after multiple corrections (S10 Fig). For the analyses of 45S rDNA contacts variation among five ESC related cell types, we observed 435 genes with significantly differential density of rDNA contacts (FDR < 0.05). For the 5S array, we observed 954 genes with DI fold change >2 in the LCL vs. K562 comparison. However, none of these genes reached statistical significance, possibly due to the higher variance emerging from the low coverage and thus limited number of 5S contacts in each gene. There was not enough data for statistical analyses of variation in 5S rDNA contact among the five ESC related cell types.
Here we estimated contact densities per base pair in three ways. First, the average contact per base pair across the whole genome was calculated by dividing the total number of mapped rDNA-genome reads by the genome length (3 billion base pairs). The average contact rate is estimated as 4.8 x 10−5 and 3.7 x 10−3 contacts per base pair for the 5S and 45S rDNA, respectively (S5 Table). Hence, for the 45S rDNA each base pair in the genome is expected to have 0.37 mapped reads. Second, the average contact per base pair was estimated after filtering out bins with pseudogenes. Here we divided the total number of rDNA-genome reads within 1MB segments without a pseudogene by the total sequence length in those segments. This yielded an estimated average contact rate of 2.0 x 10−5 and 1.7 x 10−3 contacts per base pair for the 5S and 45S rDNA, respectively. These numbers are comparable with those estimates using all rDNA reads and the whole genome. Third, we estimated the average contact rate per base pair in protein-coding genes by dividing the total number of rDNA-gene reads by the total length of nucleotides within genes, after excluding genes with evidence of containing rDNA pseudogenes. This yielded an average contact rate for genic segments of 2.2 x 10−4 and 0.016 contacts per base pair for the 5S and 45S rDNA, respectively (S5 Table). Collectively, these estimates of contact rate are useful in evaluating regions with putative enrichment or deficit in rDNA contacts.
We examined the relationship between various genomic attributes and the density of rDNA contacts. First, the data showed a significant association between the number of 45S rDNA contacts and the A/B compartments. Specifically, the B compartment of closed chromatin displays an enrichment in rDNA contacts, whereas the A compartment of open chromatin displays a deficit of rDNA contacts (P < 0.01, Chi-square test; Fig 4). In addition, we examined 15 functional annotations; significant enrichments were observed in segments of repressive chromatin, as well as in segments annotated as repetitive or containing insulator regions (P < 0.01, Chi-square test; Fig 4; S11 Fig). Finally, we examined segments of CTCF binding; CTCF is a conserved 11-zinc finger DNA binding protein that regulates chromosome architecture [20]. Using the CTCF database we estimated that CTCF binding segments constitute <7.5% of the human genome. On the other hand, we observed that 37% and 29% of all 45S rDNA-genome reads overlapped a CTCF binding segment in LCL and K562, respectively. These figures are in good agreement with the 35% of all rDNA-genome reads that overlapped a CTCF binding segment in the ESC cell set. These represent a >4-fold enrichment that indicate a significantly higher percentage of 45S rDNA contacts with CTCF binding sites (P < 0.05, one proportion test).
We selected a small set of genes to be examined in greater detail. Specifically, we examined genes that are (i) known to regulate rDNA function or structure and/or (ii) whose expression are associated with rDNA CN variation [21–23]. For instance, the CTCF gene is located on Chr16 and displayed a meager 118 contacts with the 45S rDNA in LCLs, which is significantly lower (P-value < 0.001, one proportion test) than the expected 1198 contacts calculated based on the genome wide average contacts per base pair (1.56%) and the length of the CTCF gene. Thus, the CTCF gene appears to be in repulsion to the rDNA arrays. Similarly, CBX1(Hp1beta), Ubf1, and KDM4B had fewer hits than expected (P < 0.0001 for all of them, one proportion test). Thus, we examined the top 400 genes that are positively and negatively associated with rDNA CN variation in LCL [21]. Collectively, however, these genes were neither enriched nor depleted in rDNA contacts, with a distribution of contacts that is undistinguishable from all other genes in the genome (Fig 5). Nevertheless, nucleolar, mitochondrial, and ribosomal genes were also associated with variation in rDNA array CN [21], and could reveal a distinct pattern. Accordingly, genes that localize to the nucleolus as well as ribosomal genes showed a distribution of contacts that was significantly shifted towards a greater than average number of contacts with the rDNA array in both LCLs and K562 (Fig 5 and Fig 6).
Next, we addressed if the higher density of rDNA contacts in nucleolar, ribosomal, and mitochondrial genes would emerge as significant gene ontology enrichments when genes with a high CPGN are selected. To address the issue, we examined the genes in the top 5% higher number of 5S and 45S contacts after correction for gene length (i.e., CPGN). For 5S rDNA-gene contacts in LCL the cell component category of mitochondrion (GO:0005739) emerged on the top of the list, with 56 candidates (out of 494 genes) localized to the mitochondrion. The association is functionally intriguing and also emerged in the K562 dataset (S6 Table). The same class emerged among the top 5% in the 45S rDNA in LCL, with 63 candidates in the mitochondrion (GO:0005739; adjusted P < 0.05, after correction for multiple testing). The class includes interesting candidates such as seryl-tRNA synthetase 2 (mitochondrial SARS2; ENSG00000104835), tRNA 5-methylaminomethyl-2-thiouridylate methyltransferase (TRMU; ENSG00000100416) and tRNA methyltransferase 1 (TRMT1; ENSG00000104907), Era like 12S mitochondrial rRNA chaperone 1 (ERAL1; ENSG00000132591). In addition, 10 other functionally coherent cell components emerged for 45S rDNA-gene contacts in LCL (S7 Table; adjusted P < 0.05, for all classes in LCL; see S8 and S9 Tables for data on K562 and the ESC set). Four of those categories are highly significant GO terms containing the protein-components of the ribosome (GO:0005840~ribosome, GO:0022625~cytosolic large ribosomal subunit, GO:0015935~small ribosomal subunit, and GO:0022627~cytosolic small ribosomal subunit). Collectively, the data suggest that highly transcribed genes encoding protein constituents of the ribosome are co-localized in proximity to the rDNA arrays (Table 5). In addition, one GO term related to nucleolar function (GO:0005730~nucleolus) also emerged as significantly enriched with 39 genes in the top 5% of genes with higher numbers of 45S rDNA-gene contacts in LCL. Genes in this set include intriguing candidates such as NOP2 nucleolar protein (NOP2; ENSG00000111641), FSHD region gene 1 (FRG1; ENSG00000109536), Sirtuin 6 (SIRT6; ENSG00000077463), and MDM2 (ENSG00000135679).
Among genes in the bottom 5% CPGN in 45S, we observed seven HOX genes dispersed across several chromosomes (HOXA1, HOXA6, HOXA7, and HOXA11 on Chr 7, HOXB5 on Chr 17, HOXC11 on Chr 12, and HOXD13 on Chr 2), three of which showed zero 45S rDNA contacts [HOXA7(Chr 7), HOXC11(Chr 12), and HOXD13(Chr 2)] even in the dense LCL map. This suggests that developmentally regulated Hox genes are rarely localized in proximity to the rDNA arrays. Furthermore, we also found several other developmental genes in the set of 67 genes with zero contacts with 45S rDNA genes, further indicating that developmental genes show “repulsion” from the rDNA genes. Interesting candidates include NK2 homeobox 3 (NKX2-3) on Chr 10, BMP3 on Chr 4, BMP5 and BMP6 on Chr 6, as well as NOTCH1 on Chr 9. Interestingly, the histone cluster 1 H1 family member d (HIST1H1D) on Chr 6 also emerged without a single 45S rDNA contact in the dense 45S map of LCLs. Finally, we confirmed the lack of Hi-C contacts between the 5S and 45S arrays [24]. The segments proximal to the 5S array also displayed depletion in 45S rDNA contacts. The gene RHOU, for instance, is located adjacent to the 5S array and emerged in the bottom 3% of the distribution of 45S rDNA contact density.
Multicopy ribosomal DNA arrays are essential components of the genome. Yet ribosomal DNA arrays are also among the most variable segments of the genome. The arrays have lagged behind with limited assemblies and little understanding of their nuclear localization. Here we report a detailed contact map of spatial interactions between the rDNA arrays and the rest of the genome. Although there are huge amounts of HI-C data, analyses of rDNA contact density for specific regions/genes remained a challenge because rDNA reads constitute a fraction of the Hi-C reads. Thus, we combined multiple Hi-C datasets to identify the subset of reads containing information on rDNA contacts. The effort was computational intensive because the fraction of rDNA reads in shotgun Hi-C is very small. This is particularly evident in the case of the 5S rDNA array: the contact data remained sparse even for LCL, the cell line that has by far the largest amounts of data collected from multiple Hi-C experiments. Nevertheless, we identified consistency of rDNA-gene contacts across different cells (LCL and K562; especially for 45S), which point to replicable spatial interactions. Heatmaps enabled visualization of rDNA contacts along the human genome with statistical analyses pinpointing significant differences in the density of contacts. While the approach can be applied to other multicopy genes as well as single copy genes or regions, we caution that the typical resolution of shotgun Hi-C is not sufficiently high. Indeed, limited resolution was apparent for both the 5S rDNA and 45S rDNA arrays, which required combining multiple datasets to ascertain contacts with genic and non-genic segments of the genome. In summary, the LCL map achieved good resolution for 5S and 45S contacts but the K562 set is quite sparse for 5S contacts, and both 5S and 45S maps are very sparse in the case of the ESC cell and ESC-derived cell types.
Variation in rDNA contact density across genes reflects variation in proximity to the rDNA arrays. The data displayed over 100-fold variation in contact density across genes and revealed several intriguing patterns. First, the compilation enabled us to conduct statistical tests of the differential density of rDNA interactions between LCL and K562. These 45S maps are sufficiently dense, with differences in contact density likely reflecting differences in nuclear organization between these cells. These differences are not surprising since the LCLs are immortalized cells derived from lymphocytes whereas K562 is a myelogenous leukemia. K562 has, moreover, undergone genomic rearrangements [25]. While the data also suggested variation across ESC and ESC-derived cell types, greater coverage for these cells is necessary to draw sufficiently dense contact maps for a more fine-grained and meaningful biological contrast. An intriguing suggestion is that the rDNA/nucleolus represents a keystone in nuclear structure around which the rest of the genome is functionally organized [26] [24]. In this case, rDNA-contact differences between cells are bound to emerge and reflect functional variation.
Second, as a class, the rDNA proximity with genes previously identified as associated with rDNA CN variation across genotypes in human populations is undistinguishable from the background of genes. This indicates that genes impacted by rDNA CN are not spatially close to the rDNA arrays and are not enriched in direct rDNA contacts. This is not an unexpected observation, because the association of gene expression variation with rDNA CN includes hundreds of genes, with only a fraction of them likely to be directly regulated by the rDNA array (i.e. genes associated with rDNA CN are presumably modulated by both direct and indirect effects emerging from the rDNA). While we suggest that changes in nuclear architecture could be one way to for the rDNA to exert regulatory effects, the mechanisms through which rDNA CN directly modulates gene expression are likely varied and the ratio of direct to indirect effects is unknown. Third, we observed that genes encoding proteins that localize to the mitochondria display a disproportionally large number of contacts with the 45S rDNA. Concordantly, genes localized to the mitochondria also emerged as enriched in 5S rDNA contacts. The data suggests that genes localized to the mitochondria might be collectively regulated through aspects of nuclear architecture that are influenced by the rDNA. Noteworthy, connections between the rDNA array and mitochondrial gene expression and function have been uncovered before. In Drosophila, Paredes et al (2011) observed that engineered deletions in the rDNA array preferentially impacted the expression of genes whose protein products localized to the mitochondrion [27]. In humans, Gibbons et al (2014) observed that rDNA CN variation is associated with the expression of genes whose protein product localize to the mitochondrion as well as genes encoding protein components of the mitochondrial ribosome and mitochondrial DNA copy number [21]. Interestingly, in addition to its well-documented role as a structural component of the cytosolic ribosome, the 5S rRNA is also specifically imported into the mitochondria [28, 29].
Fourth, we observed that genes localized to the nucleolus and encoding protein components of the ribosome were significantly enriched for 45S rDNA contacts. The finding points to the specificity of rDNA-genome interactions and suggests that ribosomal gene regulation might be directly influenced by the rDNA array. This pattern of rDNA-gene contacts might partially explain the observation that genes whose expression was correlated with rDNA CN included several candidates encoding the protein components of the ribosome. Indeed, sequence specific inter-chromosomal interactions between the yeast rDNA array and an intergenic segment adjacent to the largest RNA pol I subunit has recently been demonstrated [30]. All in all, our study identified functionally coherent genes and GO categories that are depleted and enriched in direct rDNA contacts. Ribosomal DNA contacted regions for all chromosomes along the human genome suggest a structural component underlying the global regulatory consequence of rDNA CN variation [21]. Finally, we note that as much as 29% of the 45S rDNA reads have both ends mapped in the 45S rDNA. These partially reflect linear proximity along the 45S rDNA unit but could also emerge from looping substructures with contacts between distant units; looping and contact among non-adjacent units has been suggested to facilitate ultra-structural organization of the array and coordinate transcription among rDNA repeat units [6, 7, 24, 31–35].
Concerted copy number variation (cCNV) refers to the correlation in copy number of 5S and 45S rDNA [36]. This co-variation in copy number across genotypes with variable rDNA array size is observed in human lymphoblastoid cells (LCLs) and occurs despite 5S and 45S rDNA residence on different chromosomes and lack of sequence homology between 5S and 45S rDNA subunits. Therefore, physical linkage between loci cannot explain the co-variation. On the other hand, spatial co-localization of the arrays as well as cellular processes of recombination such as those of micro-homology mediated end joining could conceivably contribute to the emergence of cCNV. Our results, however, confirmed a lack of direct 5S-45S contacts in Hi-C, an observation that is in agreement with a previous study [24]. This included a lack of 45S rDNA contacts with genes that are adjacent to the 5S rDNA array. The gene RHOU, for instance, is located next to the 5S and emerged in the bottom 3% of the distribution of 45S rDNA contact density. This indicates that the 5S and 45S rDNA are not in close enough proximity or that large protein complexes prevent the formation of 5S-45S Hi-C reads. The findings support the hypothesis that physical interactions occurring between 5S and 45S rDNA arrays are more restricted than previously anticipated. On the other hand, the denser maps presented here indicate that the 5S and 45S arrays share overlapping contact maps and many regions of the genome display a high density of contacts with both rDNA arrays. For instance, the density of 5S and 45S contacts is strongly correlated across genes and 1MB segments in both LCL and K562 cells. Whether or not this overlapping contact map is relevant for cCNV remains to be determined, but the evidence suggests that the two arrays are not completely independent. Coordination between them is likely to be relevant, with costs and benefits to 5S array proximity with the 45S arrays [24]. All in all, the association between contact density for the 5S and 45S arrays suggest that cCNV might be facilitated by structural proximity. Similarly, rDNA mediated structural changes in the nucleus might partially explain the regulatory consequences of naturally occurring variation in rDNA copy number [21].
From an evolutionary perspective, the co-existence of two clusters of rDNA loci (5S and 45S) might incur costs and benefits compared to rDNA residency on a single location. In some plants and yeasts, the 5S and 45S/35S rDNA subunits are spatially adjacent in the genome [7, 37–41], whereas in Drosophila and mammals, the 5S and 45S arrays reside on different chromosomes. However, the correlated contact maps for the 45S and 5S rDNA arrays suggest that they preferentially anchor at overlapping domains. This might narrow their spatial distances, and could explain why the 5S and 45S arrays can display apparent proximity to one another in a fraction of the cells as observed in cytological preparations. However, the lack of direct Hi-C 5S-45S contacts might suggest a model of competitive exclusion for similar anchoring sites, and predicts that a segment is in close proximity to either the 5S or 45S rDNA at each time. In cases of cytological proximity between the 5S and 45S arrays, large protein complexes might be present and prevent the emergence of direct 5S-45S inter-chromosomal contacts in the scale captured by Hi-C technology. Furthermore, the enrichment of rDNA contacts with ribosomal protein coding genes is surprising and might help explain the association between rDNA CN and the expression of these genes [21]. It suggests a structural component to the regulatory role of the rDNA and raises the possibility that the arrays might exert direct modulation of some genes via changes in nuclear organization. The data suggest that models that exclusively consider proximity to the rDNA arrays/nucleolus as a repressive modifier of gene expression might be overly simplistic. Rather, the distal and proximal association of genes with the rDNA arrays appears functionally motivated, as in the case of developmental genes or ribosomal genes. For instance, ribosomal gene proximity to the rDNA arrays could help facilitate coordinated Pol I, Pol II and Pol III responses. Collectively, these structural rDNA-mediated associations might have partially evolved to mitigate the fitness costs of dosage imbalances among highly expressed RNA and protein components of the translational machinery.
The human 5S rDNA along with flanking regions (chr1: 228,765,135–228,767,255) and the human 45S rDNA (GenBank reference number U13369.1, with modifications) were obtained as recently described [21, 36, 42, 43]. The 45S reference comprises the 18S, 5.8S and 28S rRNA encoding segments, external transcribed sequences (ETS) and internal transcribed segments 1 and 2 (ITS1 and ITS2), as well as a ~32 Kb non-coding intergenic spacer (IGS). Both 5S and 45S segments contain repetitive elements, such as Alu and Line1; all analysis carried out in this study used 5S and 45S sequences masked for these repeats.
Raw Hi-C reads for LCLs and erythroleukemia K562 (K562) cells were downloaded from the Gene Expression Omnibus (GEO) repository with accession number GSE63525 [44]. Biological replicates with more than 1 technical replicate were included for a total of 6,017,877,658 reads in LCL and 1,366,228,845 reads in K562. In addition, raw Hi-C reads for five cell types were obtained from GEO data with SRA Study number SRP033089 [45]. The five cell types comprised the H1 embryonic stem (ES) cells and four differentiated cell-types derived from H1 [Mesendoderm (ME) cells, Mesenchymal stem (MS) cells, Neuronal Progenitor (NP) cells, trophoblast-like (TB) cells] [45, 46]. The number of reads studied and recovery rates for 5S and 45S informative reads was summarized in Table 1.
All data were downloaded in SRA format and converted into FASTQ files by the NCBI SRA Toolkit’s command (fastq-dump). FASTQ files were quality and adapter trimmed with Trim Galore. The trimming criteria required minimal quality score (> 20) and length (>50 bp). Next, we identified Hi-C reads that mapped to the 5S rDNA array or the 45S rDNA array. In this step, both forward and reverse reads were mapped independently to the 5S rDNA and 45S rDNA using Bowtie2 [47]. We used unpaired mapping with ‘very-sensitive’ mode (combinations of parameters: -D 20 -R 3 -N 0 -L 20 -i S, 1, 0.50). The mapping results were sorted and converted into binary format using SAMtools [48] and bed format using BEDTools [49]. We then extracted reads that mapped to the rDNA array and mapped the opposite end to repeat libraries. Reads for which one end mapped to repeats library were excluded. Finally, in order to identify potential confounders due to rDNA pseudogenes, both rDNA references were blasted against the human genome separately. Putative pseudogenes were identified as significant hits (E-value <1 × 10−4) using BLASTN [50, 51]. A segment of 1 MB was excluded from the analysis if an rDNA blast hit is identified within it. Similarly, a gene was removed from analysis if an rDNA blast hit is identified within its boundaries.
To identify spatial variation in genomic contact density along the chromosomes we segmented the human genome GRCh37/hg19 assembly into 3,173 bins of 1MB using BEDTools [49]. Bins with rDNA pseudogenes were excluded. Contact densities were summarized for each bin for each of 5 cell types and 2 cell lines. We calculated the number of Contacts Per Million reads (CPM) to normalize the data and control for different number of reads in each of the seven conditions. This placed all the data in a comparable scale, to enable visualization of contact density along the human genome using heat maps in the 'gplots' R package [52].
The term of “rDNA-gene contact” refers to reads with one end mapped to rDNA arrays and the other end mapped between the first and the last exon of an annotated gene in the human genome. We extracted coordinates of these reads using BEDTools [49] and the Gene Transfer Format (GTF) file: Homo_sapiens.GRCh37.75.gtf from the Ensembl database. GC content and length were also computed for each gene. To normalize contact densities in genes of different length, we computed the number of contacts per gene length in nucleotides (Contacts reads per gene per nucleotide, CPGN). The web based tool DAVID v6.8 [53] was used to investigate gene ontology enrichments for the top 5% of genes with greater CPGN for 5S rDNA or 45S rDNA genes. This corresponds to 494 out of 9864 genes for 5S-gene contacts, and 480 out of 9595 genes for 45S-gene contacts. The “one proportion” test [54] was also applied to address whether the number of mapped reads per base pair within a gene is significantly different from the genome wide average.
We modeled differential contact density per 1MB and per gene using the edgeR package and statistical approaches adapted from RNA-seq analysis [55, 56]. Raw counts for physical contacts with rDNA loci within each bin along the human genome are modeled using generalized linear models (likelihood ratio tests) implemented in the edgeR package [55, 56]. These approaches were recently been used to detect differential interaction density (DIs) in Hi-C data [19, 57, 58]. The models identified statistically significant differences among cell lines/types in rDNA contacts density per MB and within genes. The Benjamini-Hochberg method was used for multiple testing correction [59], and statistical significance was denoted by FDR < 0.05. We applied the method to ascertain significant differences between LCL and K562 data from a single publication. For statistical comparison, we focused specifically on 11 biological replicates for LCL (collected with the Mbol enzyme) contrasted with two biological replicates for K562 (collected with the Mbol enzyme) and two biological replicates for LCL (collected with the DpnII enzyme). Each biological replicate consists of multiple technical replicates. We also evaluated variation among the five ES derived cell types, each with two biological replicates.
We cross-referenced the rDNA contact map with several sources of functional annotation. First, Hi-C studies proposed the partition of the genome into A and B compartments that are widely interpreted as open and closed chromatin, respectively [60]. A/B coordinates were downloaded for LCL cells and 12 cancer types [60]. Second, coordinates of 15 functional regions identified in hESC using ChromHMM [61] were downloaded. Third, information on replication timing along the genome was downloaded from the Replication Domain Database (www.replicationdomain.org). Finally, CTCF binding coordinates were obtained from the CTCFBSDB database [62]. We extracted the coordinates for all the segments in each annotation and addressed its density of rDNA contacts. BEDTools was used to assess the number of mapped reads that overlapped with each annotated segment for each dataset. The percentage of mapped reads was calculated by dividing the number of reads mapped to the segment by the total number of mapped reads. The genome wide average read per base pair was used to compute the expected number of reads in the functional segment. Statistical significance was assessed with Chi-square tests. In addition, we applied the “one proportion” statistical test [54] to address whether the numbers of mapped reads per base pair within a functional segment (e.g., CTCF binding) is significantly different from the genome-wide average per nucleotide contact rate.
|
10.1371/journal.pcbi.1003417 | The Free Energy Landscape of Dimerization of a Membrane Protein, NanC | Membrane proteins are frequently present in crowded environments, which favour lateral association and, on occasions, two-dimensional crystallization. To better understand the non-specific lateral association of a membrane protein we have characterized the free energy landscape for the dimerization of a bacterial outer membrane protein, NanC, in a phospholipid bilayer membrane. NanC is a member of the KdgM-family of bacterial outer membrane proteins and is responsible for sialic acid transport in E. coli. Umbrella sampling and coarse-grained molecular dynamics were employed to calculate the potentials of mean force (PMF) for a variety of restrained relative orientations of two NanC proteins as the separation of their centres of mass was varied. We found the free energy of dimerization for NanC to be in the range of to . Differences in the depths of the PMFs for the various orientations are related to the shape of the proteins. This was quantified by calculating the lipid-inaccessible buried surface area of the proteins in the region around the minimum of each PMF. The depth of the potential well of the PMF was shown to depend approximately linearly on the buried surface area. We were able to resolve local minima in the restrained PMFs that would not be revealed using conventional umbrella sampling. In particular, these features reflected the local organization of the intervening lipids between the two interacting proteins. Through a comparison with the distribution of lipids around a single freely-diffusing NanC, we were able to predict the location of these restrained local minima for the orientational configuration in which they were most pronounced. Our ability to make this prediction highlights the important role that lipid organization plays in the association of two NanCs in a bilayer.
| Cells are surrounded by selectively-permeable bilayer membranes, enabling the cell to control its internal environment. Embedded within these membranes are a variety of membrane proteins, many of which facilitate this environmental control and are integral to numerous metabolic processes. Their location within the membrane and their mutual association are controlled by many factors. We use molecular dynamics simulations to investigate the free energy of association for a pair of relatively simple membrane proteins. By doing so, we are able to characterize the effect that the geometrical properties of the protein have on their mutual association in a bilayer environment, showing that there is a correlation between the buried surface area of two proteins when in contact and the strength of their interaction. We also observe the effect of protein-lipid-protein interactions in this free energy characterization. Such interactions are related to the preferential distribution of lipids around proteins in the membrane.
| Cellular membranes not only separate the contents of a cell from its surroundings, they also play a key role in cell regulation and metabolism. Accounting for approximately a quarter of the coding regions of an organism's genome [1], membrane proteins control the transport of solutes between a cell and its surroundings, facilitate cellular movement, and regulate many aspects of cellular behaviour.
Gram-negative bacteria are surrounded by two membranes separated by a periplasmic layer. The outer membrane lipid bilayer is composed of phospholipids in the inner (i.e. periplasmic) leaflet, and of lipopolysaccharides in the outer leaflet. Within this membrane are many species of outer membrane proteins (OMPs), a class of integral membrane proteins whose secondary structures are almost exclusively [2]. Many of these are porins (OmpC, OmpF, LamB, NanC, for example), through which small (approximately ) molecules can diffuse across the membrane. Porins provide a route for many antibiotics into bacterial cells and are potential vaccine targets [3].
Both in vivo and in vitro, membrane proteins are often present in a crowded environment. Thus, cell membranes generally have a high membrane area fraction (approximately 25% or greater) occupied by proteins [4]. A similar degree of crowding may be found in membranes studied in vitro [5], [6]. Such crowding may result in the clustering of proteins [7]. Whilst the majority of discussion as to the nature of membrane protein cluster formation has focussed on lipid rafts [8], it should be noted that lateral interactions of crowded membrane proteins are a more general property [9]–[12]. In vitro, control of lateral association of a single membrane species in a highly crowded system may be used to induce two-dimensional crystallization [13].
Interactions within a crowded environment may lead to dynamic lateral interactions of membrane proteins, for example, those seen in recent time-resolved AFM studies of OmpF-containing membranes [6]. In studying such interactions, one wishes to distinguish between specific oligomerization of membrane proteins (for example dimerization of transmembrane in glycophorin [14] and of in OMPLA [15]) and less specific interactions. Specific protein interactions are those in which the distributions of orientations of the oligomerized proteins are grouped almost exclusively into very few states (often only a single state). Less specific (or non-specific) interactions are those determined by other effects, such as (local) crowding, rather than purely due the specific interactions between residues on each of the proteins. In less specific oligomerization there may be some orientational dependence, but a more broad distribution of orientations is generally found among the oligomers. Benjamini and Smit found that it was important to determine the effect that non-specific interactions had on the crossing angle for pairs of before investigating the role of any specific interactions between the helices [16]. It is therefore of interest to explore the energy landscape of lateral interactions of a relative ‘featureless’ OMP. NanC (Figure 1A) provides a good example of such a protein, as it is both structurally [17] and functionally [18] monomeric, whilst forming two dimensional crystals in DMPC lipid bilayers [19]. NanC is member of the KdgM-family of bacterial outer membrane proteins, responsible for sialic acid transport in E. coli.
Experimental determination of the free energy of membrane protein dimerization in vitro has been used to characterize their properties in a membrane or membrane-like environment [20]–[25]. Characterization of the free energy landscape for membrane proteins gives us an insight into how the proteins will move and interact within the membrane and allows us to make predictions about their behaviour. There are many published examples of experimentally determined dimerization energies for membrane proteins and peptides [20]–[22], [24], [25], but relatively few for proteins: one important example being the dimerization free energy of the phospholipase OMPLA, which was found to be in the region to [23].
Computer simulations provide a complement to both in vitro and in vivo experiments [26], enabling us to probe the microscopic interaction underlying membrane protein association. Molecular dynamics (MD) simulations have been used to explore a range of membrane proteins [26], in addition to related approaches such as Monte Carlo [27] and Brownian dynamics [28] simulations. In particular, simulations using a coarse-grained approximation [29] have been used to study dimerization of transmembrane domains [30] and of rhodopsin [31]. In the latter case the simulations were also used to characterize large-scale organization of rhodopsin dimers into rows-of-dimers, as seen experimentally in disk membranes.
Many computational studies that explore free energy landscapes use the potential of mean force (PMF) [32] as a convenient description because it enables us to characterize a given reaction or transition as a function of a specific reaction coordinate (or set of coordinates). Not only does this enable us to characterize the free energy landscape as a function of the reaction coordinate (or coordinates), it also provides an opportunity subsequently to parameterize reduced models of complex systems using different simulation paradigms [33].
Calculations of PMFs for the association of membrane proteins have largely focussed on proteins. Thus, dimerization free energy landscapes for transmembrane have been calculated using MD with umbrella sampling [30], [34] and with adaptive biasing force methods [35]; and also using Monte Carlo [27], and dissipative particle dynamics [36]. These have yielded free energies of dimerization in the region of to . To date there has only been one computational study to calculate the association free energy of a membrane protein: the free energy of association for two OmpF trimers was estimated to be in the region of [6].
It has long been suggested that lipids play an important role in the interaction between proteins in a membrane [37], [38]. For example, simulation studies have shown that hydrophobic mismatch may modulate the aggregation of proteins in the membrane [39], [40]. It is therefore important that we capture the effects that lipids have on free energy landscapes if we are to understand membrane protein association in different bilayer environments.
In this paper we develop and apply a method for calculating the free energy of association for rotationally restrained proteins in a lipid bilayer. This allows us to resolve detailed structure in the (one-dimensional) PMFs, which reflect protein-lipid-protein interactions. We apply this method to characterize the association free energy of a coarse-grained model of NanC.
In order to characterize the proteins' free energy of association, we employed MD simulations of a coarse-grained model of NanC in a POPE bilayer to calculate the PMF [29], [41]. Two orthogonal orientations of our coarse-grained NanC are shown in Figures 1B and C. It can be seen from these two views of the coarse-grained protein that, perpendicular to the pore axis, the protein is approximately elliptical in cross section, as it is wider in one direction (Figure 1B) than it is in the other orthogonal direction (Figure 1C). We calculated the PMF for four different orientational configurations of two NanC proteins, as shown in Figures 1D–G. These four configurations were chosen as examples of extremes of the possible contact regimes between the two proteins on association. In each of the configurations the proteins have either a wide (e.g. Figure 1B) or narrow face (e.g. Figure 1C) facing the other protein. These orientational configurations are categorized as: two configurations corresponding to maximal protein contact, where wide faces of both proteins face the other protein (shown in Figures 1D and E); one intermediate configuration, in which a narrow face of one protein faces a wide face of the other protein (shown in Figure 1F); and one configuration corresponding to minimal contact, where narrow faces of both proteins face the other protein (shown in Figure 1G). From these combinations of protein orientations we were able to investigate the differences between the PMFs for the various protein contact regimes.
The PMFs calculated for each of the four rotational combinations are shown in Figure 2. The PMFs were set to zero at an inter-protein separation of 8 nm, where the potentials have become approximately constant. The sampling methods, biasing potentials, rotational restraints and simulation details are given in the Methods section.
We categorized the PMFs in Figure 2 by the depths of their potential well, which resulted in three categories of well depth. The first category contains the PMFs in Figures 2A and B, which both have depths of approximately occurring at inter-protein separations of approximately 3.2 nm. This first category corresponds to the orientational configurations of maximal contact, and , where wide faces of both proteins are brought into contact (shown in Figures 1D and E, respectively). It is interesting to note that the depths of the PMFs for these two parallel and anti-parallel orientational configurations are approximately the same. They are also similar in depth to the orientationally-unrestrained PMF calculated for this coarse-grained NanC system (shown in Figure S2), which has a depth of . This is much greater than the to calculated for the dimerization of OMPLA [23], the only experimental free energy for dimerization of an OMP, but as that was for a protein exhibiting specific oligomerization measured in detergent micelles, we would not expect a good agreement. The next category contains the PMF in Figure 2C, with a potential well depth of approximately occurring at a separation of approximately 3.5 nm. This corresponds to the intermediate orientational configuration in Figure 1F, where a wide face of one protein is brought into contact with a narrow face of the other. The decrease in the depth of the PMF indicates that the configurations with two wide faces in contact are more stable than this intermediate contact configuration, where . The third category contains the PMF shown in Figure 2D, which is the shallowest of the four PMFs with a potential well depth of approximately , occurring at an inter-protein separation of 3.5 nm. This PMF corresponds to the orientational configuration with minimal protein contact, where narrow faces of both proteins are brought into contact (shown in Figure 1G). This configuration is the least stable of the four configurations considered here. The correlation between the depth of the PMFs and the orientational configuration of the proteins suggests that the strength of the interaction may correlate with the overall extent of the resultant protein-protein interface.
As well as the restrained global minima (the global minima for the specific restrained orientations) of the potential wells in the PMFs of Figure 2, there are also multiple local minima, which occur at a variety of centre of mass separations. For example, the PMF in Figure 2B has a restrained global minimum (labelled ) and two higher-energy local minima (labelled and ), which we refer to as restrained metastable states. By fitting quadratic curves to the minima in Figure 2B we calculated their locations as 3.26 nm, 3.62 nm and 4.07 nm for , and , respectively.
The nature of the restrained global () and local ( and ) minima is illustrated by the simulation snapshots shown in Figures 3A–C. These snapshots were taken from the simulation windows used to calculate the PMF in Figure 2B for an orientational configuration of . The snapshot shown in Figure 3A is from the umbrella sampling window in which the proteins were restrained with a separation of 3.3 nm, which is closest to the minimum at 3.26 nm in Figure 2B, labeled . We see that there is one lipid molecule between the two proteins at this restrained global minimum. It should be noted that this is the only lipid in between the two proteins; there is no equivalent lipid on the extracellular side of the membrane (the view from the other side of the membrane is shown in Supporting Information Figure S1), so the restrained global minimum configuration for this orientation has space for one lipid on the periplasmic side of the membrane. A snapshot from the umbrella sampling window with the proteins restrained with a separation of 3.6 nm is shown in Figure 3B, which is the window closest to the minimum at 3.62 nm, labelled in Figure 2B. We can see that there are two lipid molecules between the two proteins in this snapshot. The snapshot in Figure 3C is taken from the umbrella sampling window in which the proteins are restrained with a separation of 4.1 nm, which is the window closest to the minimum at 4.07 nm, labelled in Figure 2B, in which we see that three lipid molecules can occupy the space between the two proteins. These observations suggest that the existence of these restrained metastable states is a result of protein-lipid-protein interactions in this orientationally-restrained system.
To investigate the suggestion that these restrained metastable states were the result of the lipid ordering between the proteins, we calculated the lipid distribution around a freely diffusing NanC in a POPE bilayer. The distribution for a specific coarse-grained particle in the tail of all of the lipid molecules is shown in Figure 3D, where distinct annuli are visible, indicating regions of preferred occupation. We calculated the lipid distribution in a direction that corresponds to the direction of the other protein for the orientational configuration , indicated by the region between the dashed lines in Figure 3D. The average lipid distribution across both leaflets and all coarse-grained lipid particles in this direction is shown in Figure 3E, where again we can see there are preferred distances from the protein at which the lipids are observed. Further details of the averaging calculation are given in the Methods section.
We can use this directional lipid distribution to predict the separations at which the region between two proteins would be optimally packed by the lipid molecules. The alignment process is illustrated in Figures 3F–G and explained in the Methods section. For the minimum labelled in the PMF in Figure 2B, which occurs at a separation of 3.26 nm, we predict an optimal separation of 3.24 nm with one intervening lipid. For the first restrained metastable state labelled in Figure 2B, which occurs at a separation of 3.62 nm, we predict a separation of 3.63 nm with two intervening lipids. For the second restrained metastable state labelled in Figure 2B, which occurs at a separation of 4.07 nm, we predict a separation of 4.02 nm with three intervening lipids. Our predictions for the locations of the restrained metastable states are in close agreement with their location in the PMF. This supports our suggestion that the restrained metastable states observed in the PMFs are due to the protein-lipid-protein effects caused by the distribution of lipids between the two NanC proteins. For the other orientational configurations, the proteins have different faces facing the other protein and will therefore have a different optimal lipid distribution for each face. This may be one reason why the restrained local minima are better defined for and occur at regular intervals.
Such features are not usually observed in PMFs calculated with proteins that are free to rotate (for example, see Figure S2 for an orientationally-unrestrained PMF calculated for this coarse-grained NanC system). In the orientationally-unrestrained case the proteins would be able to rotate to alter the distance between their surfaces, provided they are not perfectly rotationally symmetric, so that the intervening region could be optimally packed with lipids without leaving any voids. However, for a system with rotationally restrained proteins, there is an optimal separation at which multiple lipid molecules can occupy the intervening space between the proteins.
Also observed in each of the PMFs is an energetic barrier, which occurs at an inter-protein separation of approximately 5.5 nm. Extending the arguments made above about the interaction of the two proteins individual lipid distributions, we can see that at distances greater than 2.5 nm in Figure 3E the fluctuations in lipid distribution have decayed to small oscillations around some constant average value, which indicates that these lipids are not as strongly influenced by the protein. From this argument we can think of this barrier as the point at which the lipids whose positions are strongly dependent on each protein begin to interact with one another, that is, there are lipids between the proteins that are affected by both of the proteins. We can think of this as the separation at which the annuli of lipids around each protein overlap with each other.
We wished to formally characterize the dependence of the PMF depth on the orientation of the proteins, which we suggested was related to the extent of protein contact. To do this we calculated the solvent accessible surface area (SASA) of the two proteins as a function of the separation of their centres of mass. For proteins with an approximately elliptical cross-section, we would expect the orientations with greater contact between the proteins to have a larger buried surface area. However, given the significance of the lipid effects that we identified above, it is important to corroborate this. Any features of the combined surfaces of the two proteins that would allow room for a lipid could have a large effect on the free energy.
The SASA is calculated using a spherical probe whose size determines the level of detail in the surface calculated for a specific set of atoms/particles. We used a probe with a radius of 0.47 nm, which is twice the radius of the coarse grained particles (0.235 nm) and should be a reasonable measure for the size of a lipid. We chose this size probe because it is the lipids that are the ‘solvent’ of interest when we bring two proteins together in a membrane. Further details are given in the Methods section.
For each of the four orientational configurations, Figure 4 shows the buried surface area as a function of distance from the minimum of their respective PMFs. We chose this measure since we wanted to remove the effect that the difference in protein radius has on the location of the minimum. In Figure 4 we see that there is a stratification of the buried surface areas in the region around the minima of the PMFs. As with the PMF depths in Figure 2, the buried surface areas can be divided into three categories. The buried surface area is largest for the orientations and , where the two wide protein faces are brought together. The next largest buried surface area around the minimum of the PMF is for the orientation , where one narrow face is brought into contact with one wide face. The smallest buried surface area around the minimum of the PMF is for the orientation , where two narrow faces are brought into contact. The correlation between the depth of the PMF and the buried surface area can be seen in the inset plot in Figure 4, in which these two quantities are plotted. We see that there is a negative correlation between the two quantities. For a protein orientation with a larger buried surface area, the minimum of the PMF is deeper.
We have characterized the free energy landscape of a pair of NanC proteins in a phospholipid bilayer. An interesting feature of these restrained free energy calculations is that certain restrained metastable states, which would usually not be seen, are now resolved. These local minima are associated with the ability of lipids to occupy the space between the two proteins at a given separation. Niemelä et al. [42] found that close to proteins, the lipids in a bilayer have reduced mobility, diffusing with the protein, and it is interactions involving these surrounding lipids upon protein association that we are observing here. Proteins' interactions with lipids have been shown to modulate local lipid formation [43], further demonstrating the important role of interactions involving both proteins and lipids in determining structures observed in the bilayer. PMFs for the dimerization of TM helices [27], [30] and for other more complex proteins, including rhodopsin [44] and OmpF [6] have revealed similar features, suggesting that a role for lipids in the energetics of their interactions may be a general feature of membrane proteins. These features may not affect the kinetics of association, as we do not know if they present metastable barriers to association, but they will affect the dynamics of the system; the NanC proteins will need to negotiate the complex free energy landscape created by these protein-lipid-protein interactions if they are to reach an energetically stable state through oligomerization.
This result may also be compared with studies of membrane protein interactions using more approximate (and hence more general) models and DPD simulations [36]. For example, such studies have suggested that changes in lipids may result in the modulation of mismatch-driven interactions of membrane proteins [45].
We identified a correlation between of the depth of the well in the free energy of association with the buried surface area at the interface of the two proteins. More generally, it has been suggested that oligomer stability of membrane proteins such as glycophorin A [22] and bacteriorhodopsin [46] may be correlated with the buried surface area at the interface. However, studies of the dimeric outer membrane protein OMPLA [15] failed to reveal such a correlation. This may reflect the role of lipids in OMPLA dimerization, confirming the need for detailed energy landscape calculations such as those presented herein.
Features of the methodology used in this work mean that care should be taken when interpreting the results. The treatment of solvents in the coarse-grained model is only approximate, so entropic contributions to solvation and lipidation/delipidation may not be captured as reliably as with a fully atomistic model. The nature of the coarse-grained model also does not enable us to separate out the contributions to the PMF due to energy and entropy, as is sometimes done using atomistic calculations. However, the observations we make here relating to the behaviour of lipids is mostly phenomenological and any quantitative observations are limited to relative comparisons between simulations of the same system. Furthermore, in choosing to look at a highly restrained system where the relative positions and orientations are restrained, we are also looking at the change in free energy along a narrow slice through configuration space. Although this path may be tightly defined, it is only by using such a highly restrained system that we are able to identify some previously unobserved behaviour, specifically the effect of protein-lipid-protein interactions on the free energy of protein dimerization. Such effects would usually be lost when averaging over a larger range of configurations.
The results presented here highlight some of the effects that contribute to the free energy of association for a bacterial outer membrane protein that undergoes non-specific oligomerization in a POPE bilayer. These processes will play a role in many protein-protein interactions, even those with some specific oligomerization modes, although in the latter case the non-specific interactions will likely be masked at close range by the specific interactions. We would expect the protein-lipid-protein interactions to be present in many membrane protein systems, as they seem to be determined by the underlying lipid-protein interactions. The PMF for the association of two OmpF trimers calculated by Casuso et al. [6] had a potential well that was approximately twice as deep as the ones we present here for NanC. However, the OmpF protein is much larger than NanC and the oligomerized proteins would therefore have a correspondingly larger buried surface compared to our NanC system.
Given the conclusions of this study, it will be of great interest to apply similar methods to those presented to calculate orientationally-dependent PMFs for a variety of other membrane proteins. Information obtained from PMFs, such as the orientational dependence of the free energy of association, are necessary for parameterizing yet coarser (i.e. more approximate) models (for example those of Yiannourakou et al. [33]), in order to enable simulation studies of the emergent properties of large, crowded and complex membrane models [47].
Umbrella sampling was used to obtain the PMF for each of the four orientational configurations while varying the inter-protein separation [32]. The umbrella sampling was performed using simulation windows in which one protein was restrained at relative positions with the desired inter-protein separation. We chose this measure as our reaction coordinate because it is a natural choice for characterizing the separation of two proteins and it would also enable the PMFs to be used to parameterize larger scale models, as done by Yiannourakou et al. [33].
To calculate the PMF from these individual biased simulation windows we employed the weighed histogram analysis method (WHAM) [48]. For the WHAM method to produce a converged PMF, we need to ensure that all points along the reaction coordinate are sufficiently sampled. This means that the histograms from the sampling of the reaction coordinate in each simulation window need to overlap with adjacent simulation windows and that these histograms are smooth. The effect of enforcing these requirements is that all points along the reaction coordinate are thoroughly sampled in multiple simulation windows. It is these considerations that determined our positioning of the umbrella sampling windows and the strength of the biasing potentials used in each.
For each orientational configuration, umbrella sampling windows were distributed at positions with varying inter-protein separation along a line connecting the centres of mass of the two proteins. The umbrella potential was applied to the centre of mass of each protein's particles, restraining them at relative positions with the desired centre of mass separation. The sampling windows were distributed at 0.1 nm intervals from an inter-protein separation of 2.8 nm to 8 nm. In each of these simulation windows we applied a harmonic umbrella potential with a force constant of . To improve the overlap of the histograms from adjacent simulation windows in the region of the local minima, which improved the resolution, additional simulation windows were used. These additional simulation windows were distributed at 0.05 nm intervals from a inter-protein separation of 2.8 nm to 4.5 nm, where the proteins were in close proximity. These closely separated windows had a stronger harmonic force constant of , in order that we could better resolve the barriers surrounding the restrained local minima.
The lipidation/delipidation of the protein-protein complex at close range is a slow process. It is important that we adequately sample both the lipidated and delipidated state. To do so, we also performed simulations in which that interface was manually delipidated, with the intervening lipids returned to the bulk of the bilayer, and the system re-equilibrated. Manually delipidated simulations were carried out for the orientational configurations , but were not required for as no persistent lipidated state was observed. Delipidated simulation windows were distributed from an inter-protein separation of 2.8 nm to 3.7 nm for , and to 4.0 nm for , separated by 0.5 nm in all cases and using the stronger position restraint of .
To analyse the free energy of association for specific relative orientations of NanC we had to ensure that we restrained their orientations in each simulation window, as well as their relative positions. This was achieved by applying a rotational potential to the particles of each protein. The rotational potential for each protein acted around a vector in the direction (approximately perpendicular to the plane of the membrane) through the protein's centre of mass. By applying a suitable rotational potential to the proteins, we were able to restrain their rotation without influencing the positional umbrella potential. Kutzner et al. [49] showed that we can virtually eliminate both the radial forces and forces parallel to the axis of rotation by using a restraining potential of the form(1)where is a unit vector parallel to the rotation axis; and are the current and reference positions of the particle, respectively; and are the current and reference positions of the centre of mass of the particles in each protein, respectively; is a rotation matrix, which describes the motion of the potential; is the force constant for the rotational potential, and is a small constant required to avoid a singularity at the axis of rotation.
The application of this potential to the particles of each protein results in a purely rotational force (a torque) about the proteins' centres of mass that acts in an approximately perpendicular direction to the plane of the membrane. This rotational potential is implemented by the enforced rotation feature of the GROMACS MD simulation package, which at each time-step applies an appropriate translational force to each of the restrained particles in order to create the desired torque [49], [50]. The proteins were inserted into the membrane in the desired relative orientations for each configuration and so we wished to restrain their orientations, which was achieved by setting the rotation matrix, , equal to the identity matrix. The reference positions for the particles were taken from the initial protein structures at maximum separation. The force constant was set to , which kept rotational drift below , and the constant was set to . Kutzner et al. showed that values for greater than were shown to give good results for rotating the subdomain of ATPase using the same value of that we have used here [49].
To calculate the PMF of association for two NanC proteins we employed coarse-grained molecular dynamics simulations of a bilayer system. The system consisted of the two coarse-grained proteins embedded in a symmetric bilayer formed from 424 coarse-grained POPE lipid molecules. The bilayer was solvated and counter ions were added to neutralize the system. The coarse-grained force field we implemented was a modified version of the MARTINI forcefield [29], [41], in which approximately four heavy atoms were mapped to each coarse-grained particle. This mapping can be seen between Figure 1A and Figure 1B. All simulations were run using GROMACS v4.6 ScalaLife 2012 (available from http://www.scalalife.eu) [50]. The simulations were performed under conditions of constant temperature (310 K) and pressure (1 bar) using a timestep of 40 fs. We have provided a GROMACS simulation configuration (mdp) file in the Supporting Information (Data S1) for a simulation window in which the protein restrained at relative positions corresponding to a centre of mass separation of 4 nm with a force constant of .
Each of the simulation windows were equilibrated for between and . The production simulations consisted of at least of simulation for each of the 0.1 nm separated simulation windows, where the applied umbrella potential force constant was . The length of simulation was increased if the PMF had not converged sufficiently. The convergence of each PMF was evaluated by comparing the PMFs obtained using non-intersecting subsets of production simulation data (see Figure S3). For each of the 0.05 nm separated simulation windows, where the larger force constant of was applied, of production simulations were performed. We also performed of production simulation for the manually delipidated simulation windows, which were separated by 0.05 nm. To combine the simulation data to obtain the PMFs we used the g_wham program, distributed with GROMACS [51], using a tolerance of .
In an attempt to predict the location of the restrained metastable states we performed a simulation of a single NanC protein freely diffusing in a POPE bilayer. This extended simulation consisted of a single coarse-grained NanC protein model (the same model used for the PMF calculations) embedded in a 25 nm square membrane constructed from coarse-grained POPE molecules. To analyse the lipid distribution around the single protein, we rotated each frame of the trajectory so that the NanC protein was aligned with its position at the start of the simulation. From this aligned trajectory we were able to calculate the position of the lipid particles in relation to the protein for the entire simulation. The particle density was calculated for a 6 nm square region around the NanC protein for each of the particles in the coarse-grained lipid molecules.
To calculate the protein density in a given direction, we calculated a linear projection of this two-dimensional density. For the case of the protein orientation configuration , the direction we are interested in is the same for both of the proteins, as they have the same face oriented toward the other protein. This direction is marked by the dashed lines in Figure 3D. In order to characterize the lipid particle density in this direction, we projected the two-dimensional density onto a series of 4 nm lines emanating from the protein's centre of mass, at regular angular intervals, within the region marked by the dashed lines. The dashed lines represent an angular window of and the individual projection lines were separated by .
To predict the location of the minimum and the local minima of Figure 2B, assuming the lipid behaviour corresponds to that shown in Figure 3D, we aligned the peaks of the mean lipid species plot (where the mean was taken across the linear projections for all coarse-grained lipid particles in both leaflets and is shown in Figure 3E) with those of the same plot overlaid with the x-axis reversed. For the case of a single intervening lipid we aligned the first peak with the first peak of the reversed plot (see Figure 3F). For the case of two intervening lipids we aligned the first peak with the second peak of the overlaid plot (see Figure 3G). Finally for the case of three intervening lipids, we aligned the first peak with the overlaid third peak and the second peak with the overlaid second peak (see Figure 3H).
To obtain the buried surface area of the proteins at various positions along the reaction coordinate, we analysed the surface area of the simulation windows with the higher translational restraining potential, , to enable the analysis of the surface area on a finer scale, using window separations of 0.05 nm instead of 0.1 nm. Using the higher force constant also ensured that the surface area was measured for a conformation that was sampled closer to the centre of the window; with the weaker force constant we would be measuring the surface area for conformations with separations that could differ significantly from the position of the window centre. All of the surface area calculations were carried out using the g_sas tool in GROMACS using of production simulation trajectory.
|
10.1371/journal.pgen.1004033 | Reactivation of Chromosomally Integrated Human Herpesvirus-6 by Telomeric Circle Formation | More than 95% of the human population is infected with human herpesvirus-6 (HHV-6) during early childhood and maintains latent HHV-6 genomes either in an extra-chromosomal form or as a chromosomally integrated HHV-6 (ciHHV-6). In addition, approximately 1% of humans are born with an inheritable form of ciHHV-6 integrated into the telomeres of chromosomes. Immunosuppression and stress conditions can reactivate latent HHV-6 replication, which is associated with clinical complications and even death. We have previously shown that Chlamydia trachomatis infection reactivates ciHHV-6 and induces the formation of extra-chromosomal viral DNA in ciHHV-6 cells. Here, we propose a model and provide experimental evidence for the mechanism of ciHHV-6 reactivation. Infection with Chlamydia induced a transient shortening of telomeric ends, which subsequently led to increased telomeric circle (t-circle) formation and incomplete reconstitution of circular viral genomes containing single viral direct repeat (DR). Correspondingly, short t-circles containing parts of the HHV-6 DR were detected in cells from individuals with genetically inherited ciHHV-6. Furthermore, telomere shortening induced in the absence of Chlamydia infection also caused circularization of ciHHV-6, supporting a t-circle based mechanism for ciHHV-6 reactivation.
| Human herpesviruses (HHVs) can reside in a lifelong non-infectious state displaying limited activity in their host and protected from immune responses. One possible way by which HHV-6 achieves this state is by integrating into the telomeric ends of human chromosomes, which are highly repetitive sequences that protect the ends of chromosomes from damage. Various stress conditions can reactivate latent HHV-6 thus increasing the severity of multiple human disorders. Recently, we have identified Chlamydia infection as a natural cause of latent HHV-6 reactivation. Here, we have sought to elucidate the molecular mechanism of HHV-6 reactivation. HHV-6 efficiently utilizes the well-organized telomere maintenance machinery of the host cell to exit from its inactive state and initiate replication to form new viral DNA. We provide experimental evidence that the shortening of telomeres, as a consequence of interference with telomere maintenance, triggers the release of the integrated virus from the chromosome. Our data provide a mechanistic basis to understand HHV-6 reactivation scenarios, which in light of the high prevalence of HHV-6 infection and the possibility of chromosomal integration of other common viruses like HHV-7 have important medical consequences for several million people worldwide.
| Human herpesvirus 6 (HHV-6) is a ubiquitous pathogen with >90% seroprevalence in healthy adults. Although the process of viral latency is not completely understood, in some cases it is achieved by the integration of the viral genome into telomeric regions of host cell chromosomes (ciHHV-6) [1], and then subsequently vertically transmitted through the germ line [1]–[3]. Approximately 1% of the human population carries genetically inherited HHV-6. After becoming latent, HHV-6 persists in a dormant state with minimal viral transcription or translation in human host cells and without the production of infectious virions and any detectable clinical complications. However, under various physiological conditions, latent HHV-6 is reactivated and forms infectious viral particles (for further details see reviews [4]–[6]). Reactivated HHV-6 has been associated with various human diseases [7]–[10].
The ∼160 kb linear double-stranded genome of both species of HHV-6 (HHV-6A and -6B) is flanked by two distinct regions, ranging from 8 to 13 kb in length [11]–[13], called direct repeats (DR) at the left (DRL) and the right (DRR) ends of the genome, respectively. Multiple passages of HHV-6 in the laboratory have led to the shortening of DR due to the deletion of specific regions in DRL and DRR [11], [12]. Both the DR regions possess two well-defined stretches of telomeric repeats (T1 and T2). These repeat regions contain several copies of the sequence (TTAGGG)n, which are also found at the termini of linear eukaryotic chromosomes. The left end of each DR has short heterogeneous stretches of telomeric repeats (DRL-T1 and DRR-T1) whereas the right end of both the DRs has a single long stretch of homogeneous telomeric repeats (DRL-T2 and DRR-T2). Several other herpesviruses including Marek's disease virus (MDV) and HHV-7 have a similar genome organization containing multiple stretches of telomeric repeats at both ends. Although chromosomal integration of HHV-7 has not been identified so far, it has been suggested that homologous recombination between viral telomeric repeats and human telomeres mediates the integration of HHV-6 [14] and MDV [15] into the host cell genome. The presence of telomeric repeats within the viral genome may have dual functions, required for both the integration (to acquire latency) and its excision (to reactivate from latency) [14], [16]. However, the reactivation machinery and the exact mechanism for HHV-6 reactivation are currently unknown.
We have previously shown that replication of the ciHHV-6 genome is efficiently reactivated in blood cells of patients that are infected with C. trachomatis without the formation of viral particles [17]. In this study, we have used Chlamydia infection as a model to understand the mechanism of ciHHV-6 reactivation. Our results provide strong evidence for the existence of a t-circle based mechanism for the circularization of the integrated viral genome, which is possibly independent of the viral infectious cycle.
The reproducible and strong reactivation of ciHHV-6 replication by Chlamydia suggests that chlamydial infection triggers the exit of the integrated virus from the host genome and the subsequent formation of circular viral DNA, which functions as a template for rolling circle replication. Since previous work has demonstrated that HHV-6 contains telomere-like sequences within its genome and it integrates within the telomeres of eukaryotic chromosomes we investigated the role of these sequences in ciHHV-6 reactivation. We devised a hypothesis for the reactivation mechanism based on the following considerations: (1) since telomeres are dynamic DNA structures, they are subject to reorganization by the telomere maintenance machinery [18]. Therefore, the HHV-6 reactivation may be a consequence of the direct changes of telomeres at chromosomal ends. (2) The reactivated HHV-6 must form a circular DNA to allow rolling circle replication. (3) The circular HHV-6 DNA must maintain at least one complete and reconstituted DR with the packaging signal sequences pac1 and pac2. These sequences are required for the packaging and assembly of intact viral particles. (4) If the telomeric repeats mediate viral reactivation, excised extra-chromosomal HHV-6 should have variable telomeric-sequence lengths. Based on these considerations, we propose the model presented in Figure 1 as a basic mechanism for the formation of complete and reconstituted viral DNA.
HHV-6 integrates into telomeric regions of human chromosomes possibly by homologous recombination [2], [14] and in the process loses the distal end of the viral genome including the DRR- pac2 signal sequence essential for the packaging of viral DNA [2], [16] (Step 1; Figure 1). During the cell cycle, telomeric repeats can be added to the end of chromosomes by the cellular telomerase complex containing the RNA subunit hTR and the catalytic subunit, hTERT. We therefore propose that the overhanging end (step 1, DRL; Figure 1) of ciHHV-6 will progressively shorten during subsequent cell divisions until the host telomerase complex is able to bind to the exposed heterogeneous viral telomeric repeats and add telomeric repeats to stabilize the chromosomal ends. This process should lead to the loss of the pac1 sequence from the DRL of the integrated viral genome (Figure 1). We hypothesize two ways by which the viral genome can subsequently be excised from the chromosome: (i) by homologous recombination within the integrated viral genome leading to circular viral DNA formation possibly mediated by the telomeric circle (t-circle) formation machinery of the host cell or (ii) by telomeric loop (t-loop) formation by branch migration and Holliday junction resolution (See reviews [18]–[20]). T -circles are duplex or single-stranded extra-chromosomal DNA circles formed from telomeric repeat sequences at the ends of human chromosomes and play a key role in maintenance of telomeres [20] whereas t-loops are loop like structures, frequently found at telomeric ends of chromosomes, which stabilize the telomere through the formation of the multiprotein shelterin complex [18]. T-loop formation is frequently observed in many different cell types; whereas single homologous recombination events during alternate lengthening of telomeres is suggested to form circular DNA molecules (t-circles) in the absence of telomerase activity [21], [22]. Since the extra-chromosomal HHV-6 DNA must be in a circular form for rolling circular replication, we considered both the t-circle and/or t-loop formation as possible mechanisms by which ciHHV-6 can form extra-chromosomal circular viral DNA.
There are two major possibilities for the successful reconstitution of the circular viral genome. A homologous recombination event between DRL-T1 and DRR-T1 would lead to the formation of a circular viral genome with a fixed length of DR-T2 and the presence of an incomplete DRL in the human genome of reactivated cells. Alternatively, the left end of the DRL could be first removed by a shorter t-circle formation process (Step 2, Figure 1), resulting in telomeric elongation of the chromosomal end from a homogenous telomeric repeat region of DRL (DRL-T2) (Step 3, Figure 1). A second t-circle formation between DRL-T2 (Step 4, Figure 1) and DRR-T2 would result in the excision of the whole viral genome from the chromosome thereby generating a circular viral genome with a single reconstituted DR (Step 4, Figure 1). The circular viral genome can further undergo rolling circle amplification to form concatemeric viral DNA that can be cleaved to form linear double-stranded viral DNA molecules including two complete DRL and DRR sequences (Step 5, Figure 1).
Reactivation of ciHHV-6 from human telomeres may be initiated by structural changes at the chromosomal ends. To analyze whether chlamydial infection has any impact on chromosomal ends, changes in telomere length were measured by telomere restriction fragment analysis during C. trachomatis infection. This assay determines telomere lengths by digesting DNA with frequently cutting restriction enzymes that do not cleave within telomeric sequences. Interestingly, we observed strong telomere shortening between 24–36 h of Chlamydia infection in different cell types including wild type HeLa (Figure 2A), HeLa229 (Figure S1A) and one of the ciHHV-6A cell lines (HSB-ML) (Figures 2B), which was followed by partial repair of telomeric ends. The presence of any nonspecific DNA degradation during Chlamydia infection was excluded by control hybridizations (Figure S1A). Termination of chlamydial infection by the addition of doxycycline, 24 h post infection, prevented telomere shortening indicating an active involvement of Chlamydia in this process (Figures 2B, S1A). Interestingly, persistent Chlamydia, a viable but non-productive form of Chlamydia induced by addition of penicillin, inhibited telomere repair (Figure 2C), suggesting that both telomere shortening and repair are actively induced by Chlamydia infection. The loss of telomeric sequences and subsequent defective telomere repair were also detected by fluorescent in situ hybridization (FISH) 48 h after Chlamydia infection (Figure 2D), where several single chromatids in Chlamydia infected cells showed weak or no telomeric signal (Figure 2D).
To test if HHV-6 reactivation involves t-circle formation after Chlamydia infection, we applied neutral-neutral 2D DNA electrophoresis. This method enables the discrimination between linear and circular DNA of the same size (Figure S1B) due to their differential migration in agarose gels [23]. This unusual migration behavior of DNA can be further resolved by either increasing the voltage or agarose concentration. Based on this principle, DNA is separated in two different dimensions during 2D DNA electrophoresis, first according to mass and then according to shape. The method has been previously extensively used to study the organization of telomeric DNA [23]–[28]. Interestingly, we detected increased t-circle formation in primary human PBMCs (Figure 2E) as well as in HeLa cells (Figure S1C) after Chlamydia infection. Thus, our data provide strong evidence for changes in telomere length and increased t-circle formation during Chlamydia infection, which can contribute to ciHHV-6 excision from human telomeres and subsequent reactivation.
The presence of well-defined stretches of telomeric repeats within the viral DR could potentially lead to the formation of t-circles of variable sizes thereby generating viral DNA containing telomeric repeats of differing lengths. To investigate the excision of viral genomes as a consequence of t-circle formation, we investigated two different viral reactivation scenarios. We have frequently observed viral integration into host cell chromosomes during productive infection of HSB-2 cells. These cells productively infected with HHV-6A (Figure S2A) contained a fraction of the replicating viruses integrating into the telomeres (Figure S2B) and therefore could serve as source for virus replication and reactivation. Alternatively, we selected HSB-ML cells, which in contrast harbor ciHHV-6A and undergo viral reactivation and formation of extra-chromosomal viral DNA upon infection with Chlamydia (see Materials and Methods for details). To compare the length of tandem arrays of telomeric repeats, extra-chromosomal viral DNA was isolated from these cells using low melting DNA gel electrophoresis and then subjected to telomere restriction fragment analysis. In addition, S1 nuclease digests were used to monitor the amount of single-stranded nucleic acids in the isolated DNA. Differential lengths of telomeric repeats ranging from 0.25 to 2 kb in length were detected during productive HHV-6 infection in HSB-2 cells and in HSB-ML cells after virus reactivation (Figure 3A).
To check for the presence of multiple conformations of extra-chromosomal viral genomes, 2D DNA electrophoresis was performed with total DNA from HHV-6A-infected HSB-2 cells and ciHHV-6 HSB-ML cells reactivated by chlamydial infection. We detected circular HHV-6 DNA (marked with a white arrowhead in Figure 3B) in addition to the linear viral genome (marked with a red arrowhead) in HSB-2 cells, which co-hybridized with two different telomeric probes as well as with the HHV-6 specific probe (Figure 3B). However, HSB-ML cells contained very low amounts of circular DNA upon reactivation by Chlamydia infection. In addition, we observed a distinct band of viral DNA that was smaller in size in both samples from the different cell-lines (Figure 3B, blue arrow head), which may represent a shorter form of circular viral DNA. Full-length circular as well as linear double-stranded HHV-6 DNA was also detected in the total DNA isolated from primary PBMCs from one ciHHV-6 patient (Figure 3C) with ongoing natural viral reactivation at the time of blood sample collection. During further analysis, single-stranded circular HHV-6 DNA was detected in Chlamydia reactivated HSB-ML cells (marked with yellow arrowhead in Figure S2C), whereas this form was hardly detectable in HHV-6A infected HSB-2 cells (Figure S2C). Single-stranded circular viral DNA was clearly longer in size than the average telomeric circle in eukaryotic cells (Figure S2C) and was not present in control DNA samples (Figure S2D). S1 nuclease treatment digested both forms of circular DNA (Figure S2E) thereby confirming the presence of nicked circular double-stranded and/or single-stranded circular viral DNA in HHV-6A-infected cells.
To verify the presence of short t-circles in ciHHV-6A harboring HSB-ML cells, we gel purified DNA bands of approximately 10 kb in size (verified for presence of HHV-6 by southern hybridization) and performed TEM analysis. We observed mostly circular DNAs of varying sizes (Figure 3D) whereas control DNA preparations from uninfected HSB-2 cells (HHV-6A negative) did not contain any circular DNA. Thus, these results support our hypothesis that short t-circles carrying viral DR are formed in ciHHV-6 cells.
The absence of homogeneous stretches of telomeric repeats at DRL-T1 suggested that telomere addition and chromosomal end maintenance might begin from the DRL-T2 region. Therefore, we expected the formation of shorter t-circles between telomeric repeats added to DRL-T1 and DRL-T2 (Step 2, Figure 1). This process should generate a shorter DRL without most of the DRL ORFs. To test this hypothesis, we performed Southern hybridization analysis with total DNA from HSB-2 cells with both, chromosomally integrated HHV-6 and ongoing productive viral infection (Figures S2A, S2B) and from uninfected HSB-2 cells (Figure S3A). Re-hybridization of the same blot with 2 different HHV-6 probes complementary to the two ends of the DRL as well as with a telomere specific probe confirmed the formation of short t-circles in these cells. DNA bands of the same length, detected by 2 different HHV-6 probes, could also originate from head-to-tail fused DNA concatemers. However, the length of head-to-tail fused concatemers should differ from those originating only from short t-circle formation (described in detail in Figures S3B, S3C), supporting the hypothesis that frequent t-circle formation removes parts of the viral DRL leading to the generation of short circular DNA molecules containing viral DR.
T-circle formation between DRL-T1 and DRL-T2 (Step 2, Figure 1) should result in chromosomal ends having overhanging DRL-T2 (Step 3, Figure 1). To test this, we performed Southern hybridization analysis with total DNA from HHV-6 infected and uninfected HSB-2 cells and detected an approximately 950 bp fragment in infected HSB-2 cells with a probe hybridizing outside of the DRL that was not detected by a DR-specific probe (Figures 4A, 4B). Interestingly, this band gave a poor telomeric hybridization signal, indicating the presence of an extremely short telomere at its left end. Similar restriction digestion and Southern hybridization experiment was carried out using total DNA from a haploid chronic myeloid cell line (KBM-7) [29], infected with HHV-6A, which allows productive virus infection (Figure S4A). Viral DNA ending at DRL-T2 as well as different sizes of short t-circles were detected in these cells (Figure 4C). Thus, the data suggest that part of the integrated HHV-6 DNA lacks DRL-T1 and contains a very short telomeric overhang starting at DRL-T2.
We detected distinct bands of short circular viral DNAs in HSB-2 cells with productive viral infection and in the Chlamydia infected HSB-ML cells (Figure 3B, marked with blue arrowhead). These short circular DNAs did not hybridize with a probe located in the U1 region, outside the viral DR, demonstrating that these short t-circles do not contain viral DNA outside the DR. The size of these bands correlate with the shorter DNA fragments from gel purified extra-chromosomal DNA (Figure 3A, marked with red arrowhead) indicating that short circular DNA molecules containing HHV-6 DR (Step 2, Figure 1) are frequently present in these cells and can be detected by various methods. Since ciHHV-6 blood DNA samples cannot be assessed for the presence of short t-circles by Southern analysis due to an insufficient DNA yield, we used inverse PCR to test for the presence of short t-circles in total DNA from freshly isolated ciHHV-6 PBMCs as well as some of the previously described cell lines carrying ciHHV-6 (detailed experimental approach is described in Figure S4). We amplified fragments of different sizes (Figures 4C), which were subsequently confirmed by southern hybridization using multiple probes and in part by sequencing to distinguish between short DRL-T1 and DRL-T2 containing t-circles (see Text S1). Variable lengths of telomeric repeats were detected in all the sequenced products, which did not show any major sequence differences within the viral DR. Our data indicate that shorter t-circle formation from ciHHV-6 DRL is a frequent event since several different sizes of short circular DNA containing partial DRL and its telomeric repeats were detected in all the cell types tested. Minor differences in the length of short t-circles were observed in the total DNA isolated before and after Chlamydia-mediated ciHHV-6 reactivation in HSB-ML cells and in one sample from ciHHV-6 PBMCs (P4) (Figure 4D). We detected several smaller amplification products using the HHV-6 probe 1 (Figure 4D), which was not detected by telomere probes. This results from the generation of incomplete PCR products due to difficulties in amplification of DNA from within highly GC-rich telomeric repeat regions. Thus, these results confirm our hypothesis that short extra-chromosomal t-circle formation is a frequent event and it is not altered by the subsequent reactivation of the viral DNA.
Validation of the final steps of the proposed model for HHV-6 reactivation required the detection of circular viral DNA with a single reconstituted DR (Step 4, Figure 1). We used inverse PCR with an inverse primer pair (table S1) located outside the DR (Figure 5A) to amplify circular viral DNA from several different cell types (Figures 5A, B). Once again, variable lengths of telomeric repeats were observed in extra-chromosomal HHV-6 DNA from HSB-2 cells (Figure 3A). Therefore, we size fractionated the full-length HHV-6A DNA from HSB-2 cells and used it separately for inverse PCR and subsequent Southern hybridization to determine the sequence composition of the different PCR fragments. We amplified circular HHV-6 DNA from HSB-2 cells as well as from one of the ciHHV-6 PBMCs (P4) infected with C. trachomatis (Figure 5B). In addition, PCR amplified DNA bands were gel purified, cloned and sequenced. Sequencing of PCR products confirmed the results of Southern analysis. The results clearly differentiated three distinct groups of DRs within these samples. We identified a fully reconstituted DR ∼9.7 kb in one fraction of total DNA from HSB-2 cells. Two other fractions contained a ∼8 kb smaller DR and one distinct fragment of ∼3.2 kb was detected in all the fractions of viral DNA from HSB-2 cells as well as in the total genomic DNA isolated after Chlamydia mediated reactivation of ciHHV-6 PBMC (P4). Interestingly, these fragments represented reconstituted short DRs, which lacked most of the DR (DR1–DR7) (Figure S5A). We also observed smaller incomplete DRs in HSB-2 cells that were not detected with probes located between DRL-T1 and DRL-T2 or the telomeric probe (Figures S5B, S5C, S6) indicating recombination between DRL-T2 and DRR-T1 facilitated by short telomeric repeats. Sequencing reads of variable sizes of HHV-6 DNA containing a single DR confirmed two possible combinations for t-circle formation, either between DRL-T2 and DRR-T1 or between DRL-T2 and DRR-T2 resulting in reconstituted DRs. We observed strong variation in the size of the reconstituted DR-T1 (Figure S6). Interestingly, circular HHV-6 DNA with incomplete DR were observed in ciHHV-6 PBMC (P4) only after Chlamydia infection, thereby confirming the viral DNA circularization event during chlamydial reactivation of ciHHV-6. Even though circular DNA was present in Chlamydia reactivated HSB-ML cells (Figure 3C), we could not detect any circular viral DNA by inverse PCR in these cells. This may be due to sequence variations in the primer-binding region or formation of a longer DR, which cannot be amplified in these PCR conditions.
Mammalian telomeric TTAGGG repeats bind the key dimeric DNA binding protein telomere repeat binding factor-2 (TRF2), which plays a key role in maintaining telomere integrity [30]. A mutant of TRF2 with a deletion in the N-terminal basic domain (TRF2ΔB) has previously been shown to induce homologous recombination mediated t-loop formation and subsequent telomere loss [25]. To find experimental evidence for t-circle dependent ciHHV-6 reactivation, we over-expressed human TRF2ΔB (Figure 6A) in various cell types. As expected, TRF2ΔB over-expression induced telomere shortening in all the ciHHV-6 cell types tested (Figure 6B) and caused cell death within 5–7 days after lentivirus infection. Since TRF2ΔB is known to induce t-circle formation, we predicted that circularization of ciHHV-6 would be observed in these cells leading to circular extra-chromosomal viral DNA formation. Extra-chromosomal circular viral DNA with a single DR was detected by inverse PCR (as described in Figure 5) and confirmed by Southern analysis using a HHV-6 specific probe (Figure 6C). Control cells that did not undergo t-circle formation did not contain circular viral DNA molecules. In addition, fragments were purified, cloned and sequenced to verify the origin of the DNA. Thus, our results confirmed the involvement of the host cell t-circle formation machinery in ciHHV-6 reactivation.
Although reactivation of latent HHV-6 has implications in the progression of many diseases including MS and CFS [8], [10], [31], the exact mechanism for latent HHV-6 reactivation, including that of ciHHV-6, remains unknown. We recently described Chlamydia infection as a natural trigger to excise the ciHHV-6 genome from the host cell telomere. In this study, we have followed an experimental approach to understand the mechanism of ciHHV-6 reactivation.
We found chromosomal integration of HHV-6A in HSB-2 cells during productive viral infection (Figure S2B). Furthermore, we have previously reported the presence of ciHHV-6 in HeLa cells after HHV-6A infection [32] and other human cell lines have been generated carrying ciHHV-6 [2], [16]. Bearing in mind that about 1% of humans carry genetically inherited HHV-6, we propose that the integration of the HHV-6 genome into human chromosomes is a frequent event. Reactivation of these integrated HHV-6 sequences likely involves their excision from human chromosomes leading to the formation of extra-chromosomal HHV-6 genomes and subsequent replication. HHV-6 contains a potential origin of lytic replication site (OriLyt) within its linear viral genome [33]. However, a linear genome cannot function as a template for viral replication. Therefore, circularization of the genome is required for its continued replication. Although some evidence exists for head-to-tail fusion and circularization of HHV-6 during viral DNA replication [13], [34], this cannot explain the reactivation mechanism of ciHHV-6. For example, head-to-tail end fused viral DNA would maintain the same viral genome size during productive infection with identical lengths of telomeric repeats (DR-T2) within the viral genome. However, we observed viral DNA with varying lengths of telomeric repeats during viral replication (Figures 3A, S5, S6) corroborating similar results published from other laboratories [16], [35], [36]. As previously demonstrated [16], [37], we detected viral DNA with a single reconstituted DR (Figure 5), a DNA configuration that cannot originate from end fusion. Therefore, recombination events involving telomeric repeats within viral genomes are likely necessary for viral DNA circularization and subsequent replication. As HHV-6 integrates within the telomeric repeats located at the ends of human chromosomes, telomeric recombination events may facilitate the excision and circularization of integrated HHV-6.
Since C. trachomatis infection reactivates ciHHV-6 replication [17], we investigated the effect of chlamydial infection on host cell telomere integrity. Cells infected with C. trachomatis experienced drastic telomere shortening (Figures 2A, 2B, S1A) and subsequent repair, which was dependent on the presence of viable and active bacteria, leading to increased t-circle formation. Changes in telomere length are frequently correlated with increased t-circle formation [25], [26], [38], supporting the hypothesis that Chlamydia-mediated telomere alteration initiates circularization of ciHHV-6. In line with this notion, single-stranded viral DNA was detected in Chlamydia-infected ciHHV-6 cells (Figure S2C) which is also consistent with the model since telomeric circle formation frequently leads to formation of single-stranded circular telomeric DNA [24]. In addition, we recapitulated ciHHV-6 circularization and viral genome reconstitution by t-circle formation and telomere shortening independent of Chlamydia infection by modulating telomeric protein complexes (Figure 6C). Telomeres are regulated and maintained by the multiprotein shelterin complex, which includes TRF2 [18]. TRF2 plays a crucial role in preventing non-homologous end joining at the end of functional telomeres through the formation of t-loops, thereby protecting telomeres from potentially harmful deletions [18]. However, the deletion of the N-terminal basic domain of TRF2 (TRF2ΔB) enhances t-loop formation through t-loop homologous recombination [25]. We utilized these properties of TRF2ΔB to induce t-loop formation in ciHHV-6 cells and showed that enhanced t-loop formation led to weak but definite circularization of viral DNA (Figure 6C). Our results thus provide direct evidence for the involvement of the telomere maintenance machinery in viral reactivation.
It is a well-known phenomenon that HHV-6A and -6B produce high amounts of viral DNA during productive infection but the formation of infectious viral particles is inefficient [39]. We propose that the reactivation of ciHHV-6 from human telomeres is an incomplete process because of the high frequency of shorter t-circle formation that only rarely results in the successful reconstitution of a complete viral genome. Frequent loss of viral DNA between DR1–DR6 in laboratory strains of HHV-6 and packaging of incomplete HHV-6 DNA lacking parts of DR (DR1–DR6) has been previously reported [39]. We have also detected loss of either most of the DR (from DR1–DR8) or parts of the DR (between DR1–DR7) in various cell types during ciHHV-6 reactivation (Figures S5, S6). On the basis of these results, we propose that the infectious nature of HHV-6 genome may be independent of the completeness of the reconstituted circular viral genome corroborating with earlier reports showing the presence of incomplete DR in infectious viral particles [39]. Previous observations of the loss of identical lengths of DR from both ends of HHV-6 DNA [39] supports our model of HHV-6 replication from a circular DNA intermediate containing a single DR. Our results indicate a predominant role of DRL-T2 during viral circularization (Figures S5, S6). This reinforces the idea that most of the left part of the viral DRL between pac2 and DRL-T2 is preferentially removed from the integrated viral genome thereby producing overhanging viral DNA ends at DRL-T2 (Step 3, Figure 1), which can subsequently recombine with the telomeric repeats of DRR (both DRR-T1 or DRR-T2) to form a reconstituted circular viral genome. The observation of DRL-T2 overhangs at the end of chromosome with frequent short telomeric repeats (Figures 4B, 4C) is in line with this hypothesis. Recent studies have utilized single telomere length analysis (STELA) assays to show similar short, unstable telomeric repeats at the sites of HHV-6 integration [16], [40]. However, results obtained with STELA assays should be interpreted with caution since extra-chromosomal telomeric circle-encoded linear DNA containing parts of the viral DR may also be amplified by this technique. To our knowledge, this is the first report to show how the telomere maintenance machinery is exploited to reactivate a latent virus after a prolonged non-infectious state. The enormous prevalence of HHV-6 infection and the possibility of chromosomal integration of other common viruses such as HHV-7, suggests that our data can form a basis to understand HHV-6 reactivation and the subsequent medical consequences for several million people worldwide.
For the study of latent ciHHV-6 activation, established ciHHV-6 cell line, HSB-ML (a tetraploid T-cell line derived from HSB-2 cells with 2–5 copies of ciHHV-6), JL-LCL and PL-LCL were kindly provided by the HHV-6 Foundation, USA (www.hhv-6foundation.org/, www.bioworldantibodies.com). Haploid chronic myeloid cell line KBM-7 [29] was a kind gift from Prof. Thijn R. Brummelkamp. Fresh blood samples from 5 individuals with ciHHV-6 were provided by the HHV-6 Foundation, USA. Viral load was re-verified by qPCR, which confirmed the ciHHV-6 status of these cells. Wild type HeLa (ATCC CCL-2), HeLa229 (ATCC CCL-2.1) and HSB-2 [32] were grown in RPMI 1640 media and 5% fetal bovine serum (FBS) at 37°C and 5% CO2. Fresh PBMCs were isolated as described previously [17].
The ciHHV-6 blood samples were collected under written informed consent under IRB# CI001-HHV-6 approved by The Essex Institutional Review Board Committee, USA.
Cells were infected with C. trachomatis at a multiplicity of infection (MOI) of 1–5 as described previously [32].
Total DNA was extracted from whole blood samples using QIAamp DNA Blood Mini Kit (Qiagen, Germany) following the manufacturer's protocol. DNA extraction from all the cultured cells was done using DNAzol (Invitrogen) following manufacturer's protocol. In particular, all the experiments involving HHV-6 DNA analysis were carried out using DNA samples extracted with DNAzol as this method is non-invasive and maintain the genomic DNA in a non-shearing form.
To study HHV-6 reactivation, ciHHV-6 cell lines and fresh blood samples from individuals with ciHHV-6, were infected with C. trachomatis serovar L2 at an MOI of 1–5. Chlamydial infection was monitored by phase contrast microscopy. After 2–3 days of infection, cells were grown in fresh media supplemented with 1 µg/ml of doxycycline, which allowed Chlamydia-infected cells to recover.
For telomere length analysis, total DNA from HeLa229 cells was digested with MspI and HhaI, which do not cut telomeric DNA, for overnight and then separated on a 0.8% agarose gel. These gels were subsequently used for Southern hybridization. As a control, the same DNA samples were digested with HindIII and processed similarly.
For Southern hybridization, agarose gels were incubated in 0.125M HCl for 8–10 min and then in DNA denaturation buffer (1.5M NaCl, 0.5M NaOH) for 30 min. DNA was transferred to Nylon-XL membrane (Amersham Hybond-XL, GE life sciences) by capillary transfer using denaturation buffer for transfer. After transfer, membrane was washed with neutralization buffer (3M NaCl, 0.3M Tri-sodium citrate, 0.5M Tris, pH 8.0) for 15 min and was subsequently pre-incubated in hybridization buffer (GE life sciences, USA). After 1 h of pre-incubation, either random primed probes (GE life sciences, USA) or end labeled probes (Table S1) were added to the hybridization buffer and incubated overnight at 42°C. Membranes were washed and exposed overnight to phosphor storage screens (Fujifilm), which were then scanned by Typhoon 9200 imager (GE Healthcare). PCR amplified Ctr LcrH/SycD gene product of 136 bp was used for random priming and subsequent as probe. For HHV-6A probe 5 and 6 (Figure 4C), PCR products were amplified using primer pairs described in Table S1 and were used for random priming.
Separation of DNA in neutral–neutral 2D gels was performed as previously described [24]. Briefly, 8–10 µg of DNA was digested with appropriate enzymes and was first separated on 0.4% agarose at low voltage in 0.5× TBE, and the gel was stained with 0.3 µg/ml ethidium bromide. The lane was cut and placed on a clean gel support at 90° to the direction of the first electrophoresis, cast with 1.1% agarose containing 0.3 µg/ml ethidium bromide and run in 0.5× TBE. The first dimension was run overnight at 1 V/cm, and the second at 4 V/cm for 4 h, both at room temperature. Southern blot analyses were performed as described above.
Total DNA containing HHV-6 DNA was separated using a 0.6% low melting agarose gel for overnight. After ethidium bromide staining, the desired bands were cut out and eluted using the phenol-chloroform extraction method.
Viral DNA was purified from low melting agarose gel. Prolonged exposure to UV light was prevented in order to avoid any DNA break. For visualizing double-stranded DNA, purified DNA in TE-buffer was mixed with ammonium acetate and cytochrome c (50–200 ng DNA in 2–10 µl TE-buffer, 200 µl 0.2 M ammonium acetate, 0.8–2.0 µl of 1% cytochrome c [in distilled water]) and 100 µl drops were placed on parafilm. After incubation for 20 min at room temperature, the cytochrome c coated DNA was picked up by touching collodion-coated grids to the surface of the drops. Grids were immediately stained with 5×10−5 M uranyl acetate in 90% ethanol or 30 sec, washed for 30 sec in 90% ethanol, air dried and metal coated with platinum/palladium by rotary shadowing under an angle of 5–7%.
100 ng of total genomic DNA was used to amplify short t-circles using a primer pair facing against each other (see table S1) and Phusion high-fidelity master mix with GC buffer (Thermo scientific). The following amplification cycles were used: Initial denaturation at 98°C for 2 minutes, 28 cycles of denaturation at 98°C for 30 seconds, primer annealing at 64°C for 30 seconds and primer extension at 72°C for 7 minutes. Final extension was done at 72°C for 30 minutes. Amplified PCR products were run on 1% agarose gel and were used for Southern hybridization. Amplified PCR products were cloned into TOPO 2.1 vector and sequenced using M13 forward and reverse primers.
100 ng of total genomic DNA was used to amplify circular or concatemeric HHV-6 DNA having a single direct repeat (DR) using a primer pair facing against each other (see table S1) and LA Taq (Takara Biosciences). Long PCR amplification was performed as follows: an initial cycle of denaturation at 92°C for 4 minutes was followed by 10 cycles of denaturation at 92°C for 10 seconds, primer annealing at 64°C for 30 seconds and primer extension at 68°C for 6 min, followed by 20 additional cycles under similar conditions except that the primer extension time was increased for 20 seconds per subsequent cycle. PCR was terminated with a final extension step for 30 minutes at 72°C. Amplified PCR products were run on 1% agarose gels and subjected to Southern hybridization. Amplified PCR products were cloned into TOPO 2.1 vector and sequenced using M13 forward and reverse primers. For inverse PCR in Figure 6C, a different primer pair (For3 and Rev3, see Table S1) was used.
FISH and Co-FISH experiments were performed using the following protocol. Metaphase spreads were prepared after 2–3 hrs of colcemid treatment using standard cell biology techniques. For co-FISH, slides were hybridized with 2 different probes using previously described techniques [41]. For co-FISH with blood cells, PBMCs were stimulated for 72 h with 10 µg/mL PHA and then cultured in RPMI1640 media containing 100 units/mL IL-2 and 10% FCS. A custom designed Alexa-488 tagged PNA oligo probe (Panagene, South Korea) against HHV-6 (Alexa488-OO-GCG TCA TAA TGC TCA ACA-CONH2) was used for FISH analysis using manufacturer's protocol. Alexa488-tagged Tel-G probes and Cy5-tagged Tel-C probe were purchased from Eurogentec, Germany (Cat No. PN-TC055-005, TG055-005). For single chromatid telomere staining (Figure 2D), HeLa cells were incubated with 5-bromo-2′-deoxyuridine (BrdU) and 5-bromo-2′-deoxycytidine (BrdC). Newly synthesized DNA strands were subsequently digested with Exonuclease III.
To validate the origin of extra-chromosomal HHV-6 DNA from the ciHHV-6, we separated extra-chromosomal HHV-6 DNA from the chromosomal DNA by agarose gel electrophoresis. DNA from both the fractions were eluted and used for amplification of viral U94 ORF by PCR. Amplified DNA was cloned into TOPO 2.1 vector and sequenced.
Constructs for TRF2 and TRF2ΔB overexpression [25] were obtained from Addgene, USA. Detailed protocol for lentivirus generation and infection are previous described [42]. Rabbit monoclonal TRF2 antibody was purchased from Abcam, UK (Cat no. ab108997).
|
10.1371/journal.ppat.1000017 | IFN-Lambda (IFN-λ) Is Expressed in a Tissue-Dependent Fashion and Primarily Acts on Epithelial Cells In Vivo | Interferons (IFN) exert antiviral, immunomodulatory and cytostatic activities. IFN-α/β (type I IFN) and IFN-λ (type III IFN) bind distinct receptors, but regulate similar sets of genes and exhibit strikingly similar biological activities. We analyzed to what extent the IFN-α/β and IFN-λ systems overlap in vivo in terms of expression and response. We observed a certain degree of tissue specificity in the production of IFN-λ. In the brain, IFN-α/β was readily produced after infection with various RNA viruses, whereas expression of IFN-λ was low in this organ. In the liver, virus infection induced the expression of both IFN-α/β and IFN-λ genes. Plasmid electrotransfer-mediated in vivo expression of individual IFN genes allowed the tissue and cell specificities of the responses to systemic IFN-α/β and IFN-λ to be compared. The response to IFN-λ correlated with expression of the α subunit of the IFN-λ receptor (IL-28Rα). The IFN-λ response was prominent in the stomach, intestine and lungs, but very low in the central nervous system and spleen. At the cellular level, the response to IFN-λ in kidney and brain was restricted to epithelial cells. In contrast, the response to IFN-α/β was observed in various cell types in these organs, and was most prominent in endothelial cells. Thus, the IFN-λ system probably evolved to specifically protect epithelia. IFN-λ might contribute to the prevention of viral invasion through skin and mucosal surfaces.
| Virus-infected cells can secrete interferons (IFNs), cytokines that induce an infection-resistant state in neighboring cells. IFNs are critical to slow down early multiplication of pathogens in the body. Two IFN families exhibiting strikingly similar properties were described: type I IFNs (or IFN-α/β) and type III IFNs (or IFN-λ). Our work addressed the question of the redundancy of these two IFN systems in vivo. First, we found that the relative expression of IFN-λ over that of IFN-α/β exhibited some extent of tissue specificity and was low in the brain. Next, we used a strategy based on in vivo expression of cloned IFN genes to compare the responses of different tissues to IFN-α and IFN-λ. As was suggested by previous in vitro work, response to IFN-λ appeared to be restricted to epithelial cells, unlike response to IFN-α which occurred in most cell types. Tissues with a high epithelial content such as intestine, skin or lungs were the most responsive to IFN-λ and expressed the higher amounts of IFN-λ receptor. Our data suggest that the IFN-λ system evolved as a specific protection of epithelia and that it might contribute to prevent viral invasion through skin and mucosal surfaces.
| Type I interferon (IFN), also called IFN-α/β, was originally discovered owing to its potent antiviral activity [1]. Type I IFN was later shown to display pleiotropic activities. It modulates innate and acquired immune responses, cell growth and apoptosis [2].
Type I IFN forms a vast multigenic family [3]. Human and mouse genomes carry 13 or 14 genes coding for closely related IFN-α subtypes [4],[5]. In addition, they contain genes coding for IFN-β, IFN-κ [6], IFN-ε/τ [7] and IFN-ω (human) or limitin/IFN-ζ (mouse) [8]. MuIFN-α subtypes share about 90% amino acid sequence identity with each other and approximately 30% sequence identity with other type I IFN subtypes. Some of these IFNs are glycosylated while others are not [4],[5],[9],[10]. In spite of this remarkable variability, all type I IFN subtypes appear to bind the same heterodimeric receptor [11], raising the question of the reason for type I IFN gene multiplicity. Some data suggest that various IFN subtypes might exhibit different affinities for each of the receptor subunits and hence, generate signals that could vary in nature, duration, or intensity. For instance, Jaitin and his collaborators reported that IFN-α/β subtypes differ in their affinity for IFNAR1 and that this receptor subunit is the limiting factor for ternary complex formation [12]. Binding to the IFNAR1 subunit would favor signaling pathways leading to antiproliferative activity whereas binding to the IFNAR2 subunit would favor signaling pathways leading to antiviral responses [13]. Such subtle binding differences could explain the few qualitative differences observed in the activity of different IFN subtypes. Alternatively, the multigenic nature of the IFN family could allow individual IFN subtypes to be expressed in a tissue or in a cell-specific fashion.
Intriguingly, the multigenic type I IFN system cohabits with the seemingly redundant type III IFN system discovered more recently. Type III IFN (also called IFN-λ or IL-28/29) is structurally and genetically close to the members of the IL-10 family of cytokines but displays type I IFN-like activity [14],[15]. In humans, 3 genes code for the 3 members of this new family: IFN-λ1, IFN-λ2 and IFN-λ3. Among these molecules, only HuIFN-λ1 is glycosylated [14],[15]. In the mouse, the IFN-λ1 gene is a pseudogene. IFN-λ2 and IFN-λ3 genes encode glycosylated proteins [16].
IFN-λ expression has been shown to depend on the same triggers (viral infection, TLR ligands) [17],[18] and signal transduction pathways [19]–[21] as those inducing type I IFN expression. Type I and type III IFNs bind unrelated heterodimeric receptors. The type I IFN receptor is made of the ubiquitously expressed IFNAR1 and IFNAR2c subunits [22]. The type III IFN receptor is made of the IL-10Rβ subunit which is widely expressed and shared by other IL-10 related cytokines, and of the IL-28Rα subunit which is specific to IFN-λ and responsible for signal transduction [14]–[16],[23]. Although type I and type III IFN receptors are unrelated, they trigger strikingly similar responses, mostly through the activation of STAT-1 and STAT-2, and, to a lesser extent, of STAT-3 [14], [16], [24]–[26]. Association of phosphorylated STAT-1 and -2 with IRF-9/p48 yields the ISGF3 complex which induces the transcription of hundreds of genes, the so-called “interferon stimulated genes” (ISGs). These ISGs encode proteins such as Mx1, OAS or IFIT, which mediate the antiviral effects of IFN [27]. IFN-α/β and IFN-λ were also reported to activate the MAP kinase pathway through JNK and p38 phosphorylation. ISGs activated by type I and type III IFNs were found to be similar [25],[26]. Accordingly, type III IFN was shown to display antiviral [23],[24],[28],[29], antiproliferative [16],[30], and immunomodulatory properties [31],[32], similar to those of type I IFN.
It has been shown that, in vitro, cell responses to IFN-λ closely depend on the expression of the IL-28Rα receptor subunit [18],[26]. Overexpression of IL-28Rα in non-responding cells restored the response of these cells to IFN-λ [26]. IL-28Rα expression has been detected in primary keratinocytes and colonic cells, but not in splenocytes, fibroblasts and endothelial cells, indicating that the IFN-λ receptor can be expressed in a cell-specific fashion [16],[24]. These data suggest that, in vivo, distinct cells or tissues might be targeted by IFN-α/β and IFN-λ. However, few data are available about production of IFN-λ and about the tissue and cell specificity of the response to this IFN in vivo.
To examine possible tissue specificity of IFN-λ expression, we compared the expression of type I and type III IFNs in the brain and in the liver, using various viral infection models. To compare the responsiveness of different tissues and cells to type I and type III IFNs, we used a strategy based on in vivo expression of cloned IFN genes. We observed some tissue specificity in the production of type III IFN and a clear tissue specificity in the response to type III IFN. At the cellular level, the response to IFN-λ showed a marked specificity for epithelial cells, thus clearly differing from the response to IFN-α.
Currently available in vitro data do not reveal differential expression of type I and type III IFN genes. To test whether some tissue specificity exists in the production of type III versus type I IFN in vivo, we compared IFN-α, IFN-β and IFN-λ expression in the brain and in the liver of mice infected with various RNA viruses: Theiler's virus (TMEV, the neurovirulent strain GDVII or the persistent strain DA1), LACVdelNSs (La Crosse virus mutant lacking the IFN-antagonist protein NSs), Mouse Hepatitis virus (MHV, strain A59) or Lactate dehydrogenase-elevating virus (LDV).
For detection of mouse IFN-λ, we designed new primers that amplify both IFN-λ2 and IFN-λ3 transcripts, but not putative transcripts from the IFN-λ1 pseudogene. For detection of mouse IFN-α, we designed primers that are specific for IFN-α5 (Table 1). This IFN subtype has been shown to be among the most prominently induced IFN-α subtypes in the brain, after both LACV and TMEV infections [33]. Using the RT-PCR-cloning-sequencing strategy used in the former study [33], we observed that IFN-α5 was also among the most prominently expressed IFN-α subtypes (20.4%) in the liver of MHV-infected mice (Figure 1). Thus, IFN-α5 expression appears to be a good marker to follow global IFN-α expression in both infected livers and brains.
We first analyzed IFN production in mice infected intracerebrally (i.c.) with MHV-A59 or intraperitoneally (i.p.) with LDV (Table 2). Following i.c. injection, MHV-A59 can spread within the central nervous system (CNS), by the hematogenous and neuronal routes. The virus can also enter the bloodstream via the disrupted blood-brain-barrier at the inoculation site and reach the liver where it replicates. MHV-A59 strain is known to target a large range of cells including hepatocytes, macrophages (including Kupffer cells and microglial cells), endothelial cells, glial cells and neurons [34]. LDV injected i.p. rapidly infects a population of LDV-permissive macrophages in the mouse [35]. One day post-infection, which corresponds to the peak of viremia, LDV antigen-positive cells have been detected in most organs, including the liver and the leptomeninges of the brain. Subsequently, the virus establishes a persistent infection that is limited by the number of available target cells [36],[37]. Thus, groups of C57BL/6 mice were infected either i.c. with MHV-A59 or i.p. with LDV, since these infection models allow to compare the IFN responses in the brain and the liver of the same animals. Mice infected with MHV-A59 were sacrificed at 72h post infection, when clinical signs of encephalitis were prominent. LDV-infected mice were sacrificed at 24 hours post infection, which corresponds to the peak of viremia and of IFN expression [36],[38].
In mice infected with LDV (Figure 2A), we noticed a striking difference in the relative expression of IFN-λ in the brain and in the liver. IFN-λ mRNA was readily detected in the liver but was hardly detectable in the brain (1 out of 9 mice had detectable amounts of IFN-λ mRNA in the brain). In contrast, IFN-α and IFN-β mRNAs were clearly detected in both the liver and the brain of these mice. The expression of IFN-λ, relative to that of type I IFN was significantly lower in the brain than in the liver (Table 3). In mice infected with MHV-A59 (Figure 2B), the same trend was observed. The differences were less extensive, yet statistically significant (Table 3).
We then examined, in diverse experimental infection conditions (see Table 2), whether the same trend of lower relative expression of IFN-λ in the brain than in the liver existed. IFN production was examined in the brain of mice infected with neurotropic viruses (TMEV-DA1, TMEV-GDVII, LACVdelNSs). At the time point analyzed (Table 2), both TMEV strains inoculated intracerebrally primarily infect neurons, as do LACVdelNSs inoculated intraperitoneally [39]–[41]. In the brain of mice infected with these viruses, IFN-λ expression was either non-detectable (TMEV-DA1) or very low, compared to that of IFN-α or IFN-β (TMEV-GDVII and LACVdelNSs) (Figure 2C, 2D, 2E, 2F). In contrast, in the liver of mice infected i.p. with MHV-A59, although variation existed between experimental groups, IFN-λ expression was close to or higher than that of IFN-α (Figure 2G, 2H). Taken together, our data suggest that IFN-λ expression (relative to that of IFN-α/β) is restricted in the brain as compared to the liver.
We next analyzed whether the response to specific IFNs also exhibited some degree of tissue specificity in vivo. To this end, we chose to compare ISGs expression in peripheral organs and in the CNS, after in vivo expression of cloned IFN genes. IFN was expressed in vivo from expression vectors that were electroinjected in the tibialis anterior muscle [42]. An advantage of this technique over the administration of recombinant IFN is that gene products are expected to carry native post-translational modifications like glycosylation.
We tested the efficacy of the procedure by following plasmid-driven expression of luciferase, using in vivo imaging. As shown in Figure 3, luciferase expression was readily detectable in the tibialis muscle after 2 days, and lasted up to 3 or 4 months after a single plasmid electroinjection. Then, to test whether IFN could be expressed in vivo, in this experimental setting, mice were electroinjected in the tibialis anterior muscle, with plasmid DNA coding for MuIFN-α6T (accession AY220465) or for a mutant of this IFN carrying a glycosylation site (D68N mutation). PCR analysis and sequencing of PCR products confirmed that the IFN subtype expressed in the muscle corresponded to the subtype expressed by the injected plasmid (data not shown). Two days and seven days after electroinjection, both glycosylated and non-glycosylated forms of circulating IFN-α6T, expressed from tibialis muscles, induced the expression of various ISGs (OASl2, Mx1, IRF7 and Ifit1) in the injected muscle but also in liver, spleen and kidney. ISGs were also upregulated, but to a lesser level, in the brain and in the spinal cord (Figures 4 and 5, Tables 4 and 5, and data not shown). When the empty vector was electroinjected, upregulation of ISG expression was detectable in the injected muscle but hardly, if at all, in other tissues. Experiments conducted in IFNAR1-KO mice failed to reveal transcriptional upregulation of ISGs by IFN-α6T (Figure 5, Tables 4 and 5), showing that the induction of ISGs, observed in mice carrying the type I IFN receptor gene, was indeed type-I IFN-dependent.
Thus, electrotransfer of plasmid DNA in vivo allows the expression of circulating IFN which activates ISG expression in the tissues examined. In this experimental setting, no significant difference was detected between the activities of glycosylated and non-glycosylated forms of IFN-α6T.
We used this in vivo expression strategy to compare the tissue specificities of the responses to type I and type III IFNs. Seven days after electrotransfer, we measured, by real-time RT-PCR, ISG expression in the organs of mice that received plasmids coding for either IFN-α6T or IFN-λ3 (Figures 4 and 5 and Tables 4 and 5). Interestingly, although IFN-α6T induced bona fide ISG expression in all organs tested, response to IFN-λ3 exhibited some tissue specificity. In response to IFN-λ3, expression of OASl2 (Figure 4), Mx1 (Figure 5), and IFIT1 (not shown) was close to background in the brain, spinal cord, spleen, liver, and muscle but was detected in the kidney. In different experimental settings (Table 4), induction of OASl2 expression by IFN-λ3 was ≤3.1±0.7 in the brain but ranged from 6±0.9 to 27±5.2 in the kidney. Accordingly, induction of Mx1 expression in mice carrying functional Mx1 alleles was ≤2.3±0.4 in the brain but ranged from 8.4±2.4 to 29±0.6 in the kidney. Experiments performed in IFNAR1-KO mice (Figure 5 and Table 5) indicated that the induction of ISG expression observed with IFN-λ3 did not depend on the activation of the type I IFN system.
In cell lines, the response to type III IFN was shown to be related to the expression level of the α subunit of the IL-28 receptor. Differential expression of IL-28Rα could thus explain the tissue selectivity of IFN-λ responses in vivo. We used real-time RT-PCR to compare the expression levels of IL-28Rα and of the ubiquitously expressed IFNAR1 subunit of the type I IFN receptor, in the brain, liver and kidney. Expression of IFNAR1 and IL-28Rα were influenced neither by IFN-α nor by IFN-λ expression (Figure 6). In the kidney, which showed good responsiveness to type III IFN, IL-28Rα expression was clearly higher than in brain and liver (Figure 6).
In order to identify the cells responding to type I and type III IFNs in vivo, we performed immunohistofluorescence using Mx1 as a marker of the IFN response. On one hand, we studied the IFN response in the kidney which was found to respond readily to both IFN-α and IFN-λ (see Figure 4). On the other hand, we examined the IFN response in the brain. In contrast to the kidney, this organ readily responded to type I IFN but hardly responded to type III IFN.
In the kidney, IFN-α-induced Mx1 expression was widespread (Figures 7C, 7E, 7G, and 8A). Mx1 labeling was prominent in endothelial cells (Figures 7C, 7E, 7G) but Mx1-positive cells also included epithelial cells of the tubules and of the urinary epithelium (not shown). The neighboring adipose tissue was also strongly responsive to IFN-α6T (Figure 8A). In contrast, Mx1 expression in response to IFN-λ was strikingly restricted to epithelial cells (Figures 7D, 7F). Labeling of the urinary epithelium was prominent (Figure 7H) (much stronger than in response to IFN-α expression). Glomeruli were negative (Figure 7D). In the cortex and medulla, only epithelial cells were positive. Adipose tissue showed background-like labeling (Figure 8B).
In the brain, very few cells responded to IFN-λ, as expected from the very low expression of ISGs in this organ. These cells appeared to correspond to rare epithelial cells of the meninges and to cells of the choroid plexus. In the choroid plexus, the comparison between Mx1 expression in response to IFN-α and to IFN-λ was exemplar (Figure 8C–F). IFN-α induced mostly Mx1 expression in the endothelial cells of the vessels comprised between the two monolayers of cuboidal epithelial cells (Figure 8C and 8E). Some epithelial cells were also Mx1-positive. In response to IFN-λ, Mx1 expression was prominent in epithelial cells but absent from endothelial cells (Figure 8D and 8F).
In view of the striking restriction of the IFN-λ response to epithelial cells in the brain and in the kidney, we tested whether the responsiveness of different tissues to IFN-λ would parallel their epithelial nature. Thus, we used real-time RT-PCR to compare, in different tissues, i) ISG induction in response to systemically expressed IFN-λ versus IFN-α (Figure 9A), and ii) IL-28Rα versus IFNAR1 expression (Figure 9B). Response of the tissues to IFN-λ (over IFN-α) nicely paralleled IL-28Rα (over IFNAR1) expression. Interestingly, tissues like stomach, intestine, skin, and lung, which have an important epithelium component showed the highest IFN-λ over IFN-α responsiveness. The small apparent differences seen between relative expressions of IL-28Rα (over that of IFNAR1) in tissues of gastro-intestinal tract and in lungs or skin were not significant. Also, these differences did not appear when considering IL-28Rα expression alone (data not shown). In contrast, nervous tissues and spleen responded very poorly to IFN-λ and expressed small amounts of IL-28Rα. Surprisingly, the liver responded poorly to IFN-λ and expressed low amounts of IL-28Rα, in spite of the epithelial nature of the hepatocytes. In contrast, the response of the heart was surprisingly high.
Many data converge to show that type I IFN can be expressed by virtually all nucleated cells, including some neurons. In contrast, little is known about the specificity of IFN-λ expression. Upregulation of IFN-λ transcription has been shown to depend on the same stimuli, sensors, and signal transduction pathways as those involved in type I IFN production [17]–[21],[28]. IFN-λ expression has been mainly described in vitro, in MD-DCs, pDCs, macrophages, and in numerous lymphoid, myeloid and epithelial cancer cell lines [18],[28]. In these studies, IFN-α/β and IFN-λ have been shown to be expressed simultaneously. In MD-DCs and in pDCs, upon influenza A or Sendai virus infection, IFN-α/β and IFN-λ were expressed at the same order of magnitude and with similar kinetics [43].
Our data show that expression of IFN-λ in the central nervous system is minimal, even under conditions of strong IFN-α and IFN-β expression, as those observed after infection by LACVdelNSs or TMEV-GDVII. In contrast, in the liver, IFN-λ was readily expressed after both LDV and MHV-A59 infections. The difference between relative type III and type I IFN expression levels detected in the liver and in the brain was highly significant in the case of C57BL/6 mice infected i.p. with LDV or infected i.c. with MHV-A59. A similar trend of low relative expression of IFN-λ in the brain was observed with the other infection models (different viruses and different mouse strains). However, our study does not exclude a possible influence of the mouse genetic background in the relative expression of type I and type III IFN genes.
Nevertheless, our results show that some differential tissue specificity exists in the production of type I and type III IFNs. This suggests that the molecular pathways leading to type I and type III IFN gene expression vary either qualitatively (some specific factors required for IFN-λ gene induction) or quantitatively (different thresholds of sensors, signal transduction or transcription factors required for the activation of type I and type III IFNs). The tissue specificity of IFN-λ production observed in this work probably results from cell type specificity. In vitro, IFN-λ was shown to be notably produced by MD-DCs and pDCs [43]. If these cells are also important IFN-λ producers in vivo, the paucity of DCs, in particular of pDCs, in the CNS might be the reason for the low expression of type III IFN in this organ.
In cell lines, IFN-λ responses have been shown to correlate with expression of IL-28Rα. On the basis of IL-28Rα expression and of IFN-λ responsiveness of cell lines and primary cells, it was suggested that IFN-λ could be primarily expressed by cells of epithelial origin. Accordingly, in vivo, IFN-λ proved to be effective against some viruses known to infect epithelial cells such as Herpes simplex virus-2 [17]. Indirect evidence also comes from the fact that Yaba-like disease virus, a virus with tropism for the dermis was found to produce a type III IFN antagonist protein [11]. However, until now, no direct in vivo data identified the cells responding to IFN-λ.
Here, we show, by immunohistochemistry, that the response to IFN-λ involves primarily epithelial cells, at least in the kidney and in the CNS. In the kidney, Mx1 expression in response to IFN-λ was notably prominent in cells of the pluristratifiated urinary epithelium. In contrast, endothelial cells which responded nicely to IFN-α failed to respond to IFN-λ. In the choroid plexus of the brain, response to IFN-α was most prominent in endothelial cells and detectable in cuboidal epithelial cells. In contrast, response to IFN-λ was only detectable in cuboidal epithelial cells. At the tissue level, responsiveness to IFN-λ, as measured by ISG induction, correlated with IL-28Rα over IFNAR1 expression. Again, epithelium-rich tissues such as stomach, intestine, skin or lung were responsive to IFN-λ. It is not clear, however, why the liver was not more responsive and why the heart appeared to be as responsive as the lung.
IFN-λ was reported to share, with type I IFN, immunomodulatory activities. For instance, IFN-λ was found to modulate the Th1/Th2 balance of the immune responses [32]. However, in agreement with previous studies, our data show that neither endothelial cells nor spleen cells, two important components of homing and activation of immune cells, responded detectably to IFN-λ, though the response of a small cell population could easily have been undetected. It will be of interest to identify the target cells that mediate the immunomodulatory function of IFN-λ.
Type I IFN turned out to have much impact on CNS pathologies. On one hand, type I IFN was shown to be instrumental in the resistance of humans and mice to neurotropic viral infections [44],[45]. On the other hand, type I IFN proved to be beneficial against autoimmune disorders like multiple sclerosis [46],[47] and the murine experimental autoimmune encephalitis [48]. IFN-β was shown to decrease the relapse rate and disease activity in relapsing-remitting MS [49]. However, exposure to type I IFN can also cause adverse effects. IFN treatment often triggers flu-like symptoms. When prolonged, for instance in the case of hepatitis C treatment, type I IFN treatment can lead to neurological or neuropsychiatric adverse effects like depression [50],[51].
IFN-λ could represent an interesting alternative to type I IFN. Indeed, IFN-λ appears to activate the same set of genes as type I IFN and most biological functions of type I IFN appear to be shared by type III IFN. We observed that the CNS is both a poor IFN-λ producer and a poor responder to this cytokine. In the CNS, the blood-brain barrier is mostly made of the tight junctions that bridge the endothelial cells and thus prevent the diffusion of metabolites from the blood to the CNS parenchyma. The lack of responsiveness of endothelial cells to circulating IFN-λ could thus explain the global absence of response to IFN-λ in the CNS. In the choroid plexus, however, endothelial cells are fenestrated. In this structure, the blood-brain barrier is formed by the tight junctions occurring between the cuboidal epithelial cells (Figure 8G). Response of these cells to IFN-λ suggests that they express the IFN-λ receptor on their basolateral membrane which is accessible to factors diffusing from the bloodstream. The low responsiveness of the CNS to IFN-λ does not appear to result solely from the combination of the blood-brain barrier and lack of endothelial cell responsiveness. Our RT-PCR data show that expression of the IL-28Rα receptor chain is very low in the entire brain. This suggests that, even in inflammatory conditions (such as in MS or during viral infection) known to affect the integrity of the blood-brain barrier, the CNS would be expected to respond poorly to IFN-λ. This fits with the observation that IFNAR1-KO mice (which have an intact IFN-λ system) exhibit extreme susceptibility to many neurotropic viral infections [45]. It will be of interest to test whether, owing to the low responsiveness of the CNS, IFN-λ would exhibit less toxicity than IFN-α/β. This might be of interest if the effective targets of the IFN treatment are in the periphery and, of course, responsive to IFN-λ.
Although type I and type III IFNs signal through different receptors, these two IFN families share common features. Production of both IFN types can be triggered by the same stimuli and responses of cells to type I and type III IFNs involves the upregulation of the same set of genes. Why these two seemingly redundant systems co-evolved is not fully clear. Previous data based on cell lines and primary cells responsiveness to IFN-λ suggested that a key difference between the type I and type III IFN systems could be the cell specificity of IFN-λ receptor expression [18],[26]. Our work confirms that, in vivo, a major difference between the type I and type III IFN systems is the cell type-restricted nature of responses to IFN-λ. Type III IFN appears to have evolved primarily as a protection of epithelial cells. However, type I IFN also acts on these cells, leaving open the question of redundancy.
On one hand, IFN-λ could be viewed as a leftover of an ancestral antiviral protection system that arose to protect simple organisms. In the evolution, type III IFN-like genes, which occur in the genome of the fish, appear to have preceded type I IFN genes that emerged with the development of birds and tetrapods [52]. Type I IFN would have evolved faster to become the primary antiviral protection system, active in many cell types. In this hypothesis, IFN-λ would only play the role of a back-up system.
On the other hand, the co-existence of two systems with overlapping specificities might have been selected because both systems contribute to the protection against live-threatening and/or widespread pathogenic viruses or microorganisms. Our data suggest that the primary function of IFN-λ would be the protection of epithelial structures. Many viruses use epithelial cells as primary replication sites. These include viruses like poxviruses, herpesviruses and influenza virus that could have had enough impact on species populations to drive some evolution of the genomes. The effect of IFN-λ against vaginal infection by HSV-2 [17], the inverse correlation between rhinovirus-induced IFN-λ expression and viral load in infected volunteers [53], and the antagonistic activity of Yaba-like virus against IFN-λ [11], support an active role for this IFN. IFN-λ is thus also expected to contribute to the defense of respiratory epithelia, against influenza virus. Accordingly, recent findings suggest that IFN-λ contributes to the protection of airways against influenza A virus, through induction of Mx gene expression (Markus Mordstein and Peter Staeheli, unpublished observations). IFN-λ might also be instrumental in the early defense of the intestinal mucosa against very common pathogens such as rotaviruses or possibly against bacteria. Further studies are needed to confirm that, in these tissues, the primary targets of IFN-λ activity are also the epithelial cells, and to evaluate how much protection is added by the IFN-λ system to the very potent IFN-α/β system. The notion that some differential regulation exists in the production of type I and type III IFNs might also broaden the range of the response or accelerate the reactivity of the body to some specific pathogens.
3–4 week-old female FVB/N, 129/Sv, C57BL/6 mice (infection experiments) and 7–8 week-old female or male FVB/N mice (electroinjection experiments) were obtained from Charles River Laboratories or from the animal facility of the Univ. of Louvain, Belgium. Congenic mice carrying a functional Mx1 gene were from the breeding colony of the Univ. of Freiburg, Germany. These mice were BALB.A2G-Mx1 and B6.A2G-Mx1 (designated Mx1/WT) [54] as well as B6.A2G-Mx1 mice lacking a functional type I IFN receptor (designated Mx1/IFNAR1-KO) [55]. Handling of mice and experimental procedures were conducted in accordance with national and institutional guidelines for animal care and use (Agreement ref. UCL/MD/2006/034).
Viruses used in this study were: Theiler's murine encephalomyelitis virus (TMEV) persistent strain DA (DA1 molecular clone), and neurovirulent strain GDVII [40], La Crosse virus deleted from the NSs gene (LACVdelNSs) [56], mouse hepatitis virus, strain A59 (MHV-A59) [57] and lactate dehydrogenase-elevating virus of the Riley strain (LDV) [58].
Intracerebral infections (i.c.) were done by injection of 40 microliters of serum-free medium containing 103 PFU of TMEV(GDVII), 105 PFU of TMEV(DA), or 2×104 TCID50 of MHV-A59. Control mice were injected with 40 microliters of serum-free culture medium. Intraperitoneal (i.p.) infections were performed by injection of 250 microliters of serum-free medium containing 104 PFU of LACVdelNSs, 2×107 ID50 of LDV, or 1 or 2×104 TCID50 of MHV-A59.
Mice were anesthetized and perfused with PBS before organs harvest. RNA was isolated from organs using the technique described by Chomczynski and Sacchi [59] and reverse-transcribed as previously described [60]. Real-time RT-PCR was performed, as described previously [60], using SybrGreen and the iCycler or the MyIQ™ apparatus (Biorad). Standards consisted of 10-fold dilutions of known concentrations of murine genomic DNA, of plasmids carrying the PCR fragment of interest (pCR4-Topo, Invitrogen) or plasmid pcDNA3-IFN-α5 [10] or pEF-IFN-λ3 [16] (kindly provided by S. Kotenko). Primers sequences and PCR conditions used are presented in Table 1. The IFN-subtype specificity of primers for IFN-α5 was confirmed. No PCR product was detected when plasmids encoding the other IFN-α subtypes were used as templates. Moreover, when genomic DNA was used as template, the IFN-α5 gene segment was specifically amplified, as confirmed by sequencing of the PCR products.
The firefly luciferase gene was cloned from pGL3 (Promega) in pcDNA3 (Invitrogen) using HindIII-XbaI restriction sites, to yield pCS41. Plasmid pcDNA3-muIFNα6T [10] was subjected to site directed mutagenesis [61] with oligonucleotide TM439 (5′ GGA GGG TTG CAT TCC AAG CAG CAG A 3′) to generate the Asp to Asn78 mutant (D78N) that carries a N-glycosylation site. The mutated IFN-α6T region was recloned in pcDNA3 and sequenced to make sure that no unexpected mutation occurred during the mutagenesis procedure.
MuIFN-λ3, was cloned from pEF-2-mIFN-λ3 [16] into pcDNA3 (Invitrogen) using Asp718-EcoRI restriction sites. The human IFNGR2 signal sequence and the N-terminal FLAG coding sequences present in pEF-2-mIFN-λ3 were replaced by a sequence encoding the wild-type murine IFN-λ3 signal sequence. To this end, the 3′ complementary primers TM723 (5′ AAA GGT ACC GCC ACC ATG CTC CTC CTG CTG TTG CCT CTG CTG CTG GCC GCA 3′) and TM724 (5′ AAA GGA TCC GCT TGG GTT CTT GCT AGC ACT GCG GCC AGC AGC AGA GGC AA 3′) were used for PCR and the resulting fragment was cloned in the recombinant plasmid using the Asp718 and BamHI restriction sites. The muIFN-λ3 region was sequenced to make sure that no unexpected mutation occurred during PCR and subcloning steps. The plasmid obtained, pcDNA3-IFN-λ3, encodes a wild-type muIFN-λ3 with a wild-type signal sequence. A similar procedure was followed to obtain pcDNA3-IFN-λ2 from pEF-2-mIFN-λ2 [16].
Mice were anesthetized with 200 µl of a mix of Medetomidin hydrochlorid 100 µg/ml (Domitor) and Ketamine 500 µg/ml (Anesketin) given i.m. Before DNA injection, mice were shaved locally, using depilatory cream. 10 µg of endotoxin free plasmid DNA (Qiagen endofree) in 25 µl of PBS were injected in the left and right tibialis anterior muscles of the mice. Electric pulses (80 V per 4 mm, 8 pulses, 20 msec/pulse, pause: 480 msec) were then administered using a Cliniporator system (Cliniporator, IGEA, Carpi, Italy) equipped with 4 mm electrode plates [62]. For all experiments, conductive gel was used to ensure electrical contact with the skin (EKO-GEL, ultrasound transmission gel, Egna, Italy). Mice were then woken up by i.m. injection of 250 µl of Atipamezol 500 µg/ml (Antisedan).
Mice were anesthetized as for DNA electroinjection and given 3 mg of Luciferin (Xenogen) in 100 µl of PBS, intraperitoneally. 10 min after luciferin injection, luciferase activity was monitored in vivo using a CCD camera (IVIS 50, Xenogen) [63]. Mice were then woken up, as described above.
Mice were anesthetized before being euthanized for organs harvest. They were perfused with PBS. Freshly collected brains and kidneys were immersed in buffered formaldehyde 4% for 24h at room temperature and then embedded in paraffin. Tissue sections of 8 µm in thickness were cut, placed on SuperFrost Plus slides, dried at 37°C overnight, and processed by standard methods for immunohistochemistry. Briefly, sections were deparaffinized, permeabilized for 5 min in phosphate-buffered saline (PBS) containing 0.1% Triton X-100, and washed in PBS. Sections were then treated for 90 min at 97°C in sodium citrate buffer 0.01 M - pH 5.8, to unmask antigens. Blocking was performed by incubating sections for 1 hour with normal goat Serum (Sigma) diluted 1/50 in PBS. Then, immunolabeling was done in blocking solution containing the antibodies. Mx1 protein was detected with rabbit polyclonal antibody AP5 [64] that recognizes the C-terminal 16 amino acids of Mx1. It was used at a dilution of 1/150. For immunofluorescent labeling, the secondary antibody (at 1/800) was a goat anti-rabbit antibody coupled to Alexa 488 (Molecular Probes). |
10.1371/journal.pgen.1002648 | A New Role for Translation Initiation Factor 2 in Maintaining Genome Integrity | Escherichia coli translation initiation factor 2 (IF2) performs the unexpected function of promoting transition from recombination to replication during bacteriophage Mu transposition in vitro, leading to initiation by replication restart proteins. This function has suggested a role of IF2 in engaging cellular restart mechanisms and regulating the maintenance of genome integrity. To examine the potential effect of IF2 on restart mechanisms, we characterized its influence on cellular recovery following DNA damage by methyl methanesulfonate (MMS) and UV damage. Mutations that prevent expression of full-length IF2-1 or truncated IF2-2 and IF2-3 isoforms affected cellular growth or recovery following DNA damage differently, influencing different restart mechanisms. A deletion mutant (del1) expressing only IF2-2/3 was severely sensitive to growth in the presence of DNA-damaging agent MMS. Proficient as wild type in repairing DNA lesions and promoting replication restart upon removal of MMS, this mutant was nevertheless unable to sustain cell growth in the presence of MMS; however, growth in MMS could be partly restored by disruption of sulA, which encodes a cell division inhibitor induced during replication fork arrest. Moreover, such characteristics of del1 MMS sensitivity were shared by restart mutant priA300, which encodes a helicase-deficient restart protein. Epistasis analysis indicated that del1 in combination with priA300 had no further effects on cellular recovery from MMS and UV treatment; however, the del2/3 mutation, which allows expression of only IF2-1, synergistically increased UV sensitivity in combination with priA300. The results indicate that full-length IF2, in a function distinct from truncated forms, influences the engagement or activity of restart functions dependent on PriA helicase, allowing cellular growth when a DNA–damaging agent is present.
| Translation Initiation Factor 2 (IF2) is a bacterial protein that plays an essential role in the initiation of protein synthesis. As such, it not only has an important influence on cellular growth but also is subject to regulation in response to physiological conditions such as nutritional deprivation. Biochemical characterization of IF2's function in replicating movable genetic elements has suggested a new role in the maintenance of genome integrity, potentially regulating replication restart. The parasitic elements exploit the cellular replication restart system to duplicate themselves as they transpose to new positions of the chromosome. In this process, IF2 makes way for action of restart proteins, which assemble replication enzymes for initiation of DNA synthesis. For the bacterial cell, the restart system is the means by which it copes with accidents that result in arrest of chromosomal replication, promoting resumption of replication. We present evidence for an IF2 function associated with restart proteins, allowing chromosomal replication in the presence of DNA–damaging agents. As the IF2 function is a highly conserved one found in all organisms, the findings have implications for understanding the maintenance of genome integrity with respect to physiological status, which can be sensed by the translation apparatus.
| Translation Initiation Factor 2 (IF2; for a review, see [1]) is an essential cellular protein that brings mRNA, the 30S ribosome, and the initiator fMet-tRNA together into the 30S initiation complex and then promotes association with the 50S ribosomal unit to form the 70S initiation complex [2]–[4]. We have previously identified it as an essential component for reconstituting bacteriophage Mu replication by transposition in vitro, a process in which IF2 makes way for initiation of DNA synthesis by the cellular restart proteins [5]. This finding raises the question whether IF2 could play an important function in the maintenance of genome integrity by regulating the engagement or activity of restart proteins.
For bacteriophage Mu transposition in vitro [6], IF2 plays a critical part [5] during the transition from strand exchange catalyzed by MuA transposase [7], [8] to the assembly of the replisome by the host replication restart proteins [9] (Figure 1; for a review, see [10]). IF2 binds to Mu DNA only upon disassembly of the oligomeric MuA transpososome that remains tightly bound to Mu ends after strand exchange [11]–13. This process begins as ClpX weakens the transpososome assembly [14]–[16] and is completed by host factors which promote transition to replisome assembly [5], [9], [15], [17]. Strand exchange creates a fork at each Mu end, creating a potential site for initiating Mu DNA replication. However, the Mu forks retain a block to initiation of DNA replication even after transpososome disassembly, and IF2 appears to play a key role in unlocking this complex [5]. Restart proteins are subsequently assembled, beginning with the displacement of the IF2 by PriA helicase. The reaction in vitro specifically requires the E. coli replication restart proteins PriA, PriC, and DnaT but not PriB, indicating that the mode of Mu replication reconstituted in this system is through the PriA-PriC restart system [18], [19]. (The PriA-PriC pathway is one of the two major cellular restart pathways, the other being the PriA-PriB pathway, which requires PriA, PriB, and DnaT [18].) Additionally, only truncated forms of IF2 (IF2-2 and IF2-3; Mr of 79.7 and 78.8 k compared to 97.3 k for full-length IF2-1), synthesized from two internal, in-frame start codons within the infB gene, have been found to be active in this in vitro system.
Indeed, the role of the various IF2 forms in translation remains unclear. Full-length (IF2-1) and truncated (IF2-2/3) forms are present in nearly equimolar amounts under normal growth conditions [20], [21], and IF2-2/3 levels increase with respect to IF2-1 during cold shock [22]. Mutations that prevent expression of IF2-1 or IF2-2/3 elicit cold sensitivity [21]. However, even IF2 with one-third of its residues deleted from the N-terminal end has intact activities in vitro as translation factor and supports cell viability when present in excess [23], [24].
IF2's role in Mu DNA replication by transposition in vitro raises the question whether it can influence or regulate the engagement of cellular restart mechanisms. The apparent function implied by the Mu replication system is that by binding to forked DNA templates, it may promote or regulate the action of restart proteins. IF2's molecular chaperone activity [25] potentially plays a function similar to ClpX, promoting remodeling of the nucleoprotein assembly at the Mu ends for the transition to a new complex [5] or plays a key part in the activation of enzymatic functions necessary for replication restart. Moreover, IF2's major function as translation factor as well as its possible function as a transcriptional activator [26], [27] also indicate its potential to influence restart mechanisms by promoting expression of proteins needed for this process. Indeed, the role of IF2 in Mu replication may be an idiosyncrasy of Mu as a parasite exploiting host proteins to promote its own propagation; alternatively, it may reflect IF2's cellular role in regulating engagement of restart functions, a function that Mu exploits as a parasite.
In this work, we examined whether IF2 function can affect specific pathways for replication restart by perturbing its function with mutations that prevent expression of IF2-1 or IF2-2/3. Only truncated forms of IF2 have been found to be active in the reconstituted Mu replication system by the PriA-PriC pathway [5]. While this result does not necessarily indicate that only the truncated forms of IF2 may be involved in restart mechanisms (the in vitro system may have lacked factors needed to engage IF2-1), it nevertheless suggests functional differences between isoforms that may be examined in vivo.
Here, we demonstrate that the loss of IF2-1 or IF2-2/3 results in different defects in restart mechanisms that cope with DNA damage during cell growth. In particular, the loss of IF2-1 elicits a phenotype that is analogous to a certain restart mutant. No matter the mechanism by which IF2 influences restart mechanisms, the results indicate a new function of IF2 in influencing the engagement of restart mechanisms, the relative levels of IF2 isoforms having the potential to affect the choice or course of the restart mechanism. We discuss the potential for IF2 to regulate maintenance of genome integrity with respect to cell physiology, suggesting a means for coordinating replication, recombination, and repair with translation status.
In the in vitro Mu replication system, binding of IF2-2 can be detected after strand exchange just prior to the binding of the restart protein PriA [5]. Since this is the major basis for suspecting that IF2 may serve a function that affects activity of restart functions, we wished to confirm that IF2 indeed binds at or near Mu ends in vivo when Mu development is induced. Chromatin immunoprecipitation (ChIP) analysis was conducted with extracts of induced lysogens expressing IF2 with an N-terminal S tag (S-IF2) after extensive RNase treatment.
Mu DNA was co-precipitated with S-IF2-1 and S-IF2-2 in induced GTN373 (a thermoinducible Mu lysogen) at 35 min postinduction (Figure 2A, S-IF2-1 and S-IF2-2), using antibody against the S tag. In contrast, relatively little Mu DNA was precipitated with S-IF2-2 upon inducing the isogenic lysogen that has a clpX knockout mutation (Figure 2A, S-IF2-2 ClpX−) and thus cannot support Mu replication [28]. This result parallels findings in vitro that the omission of molecular chaperone ClpX from the reaction system does not permit binding of IF2-2 to Mu DNA and the initiation of Mu replication [5], [15]. As it appeared that Mu ends were being enriched in immunoprecipitations when cells were undergoing Mu replication, we repeated the ChIP with 5-fold less antibody to ascertain whether bound S-IF2-2 in induced GTN373 is concentrated around Mu ends. In the immunoprecipitated samples, the Mu ends sequences were enriched over the center sequences (18 kb from either end) as well as host DNA (Figure 2C). In the control PCR amplification of total DNA, the Mu end and center sequences were amplified to the same extent. Mu PCR products were produced at higher levels than the host thrA PCR product at 35 min postinduction, reflecting the replication of Mu during lytic development. IF2 does have some nonspecific DNA binding activity [27]. Thus, the enrichment of Mu end sequences with respect to Mu center sequences by immunoprecipitation is the best indicator of preferred IF2 binding at or near Mu ends although the enrichment of Mu end sequences with respect to host DNA is also clear in this analysis.
To ensure that the anti-S tag antibody was specifically precipitating Mu DNA bound to S-IF2-2, we compared the co-precipitation of Mu DNA (35 min postinduction) in induced lysogens expressing S-IF2-2 and untagged IF2-2 (Figure 2B). When the IF2-2 had no S tag, no more Mu DNA was captured in the immunoprecipitation than in the no-antibody control.
The results indicate that not only truncated IF2-2/3 but also full-length IF2-1 bind at or near Mu ends upon induction of Mu development, corroborating the role IF2 plays in vitro in promoting initiation of Mu DNA replication by restart proteins. In vitro, IF2 makes way for the binding of PriA [5], which binds to forked DNA structures [29], [30] such as the Mu fork, and PriA subsequently displaces IF2 from Mu DNA. The ChIP analysis by itself can only indicate a preponderance of IF2 binding around the Mu ends and does not rule out the possibility that IF2 binds at nearby sites. Nevertheless, these results together with the role IF2 plays in vitro strongly suggest that there are IF2 molecules bound at the Mu fork during lytic development. The role played by IF2 in Mu replication raises the question whether IF2 function can regulate the engagement or activity of restart functions.
We constructed a series of strains with infB alleles that only allow expression of full-length IF2-1 or the truncated forms IF2-2/3 to examine their effect on restart functions. The infB alleles were introduced into the chromosome where a transposon vector was inserted, and then the natural infB allele was knocked out by introduction of the del(infB)1::tet allele, which precisely deletes the natural cistron for IF2 (Figure 3A–3B). To prevent the expression of IF2-1, we deleted sequences around the translation initiation start site for IF2-1. Sequences from 14 nucleotides upstream of the IF2-1 start codon to 32 nucleotides upstream of the IF2-2 start codon were deleted (Figure 3B); this is known to permit expression of the truncated IF2 forms while eliminating IF2-1 expression [21]. The resulting allele, denoted as infB(del1) to indicate that the deletion prevents expression of IF2-1, supports the synthesis of only IF2-2 and IF2-3. Expression of the truncated IF2 forms were prevented by changing the start codons of IF2-2 and IF2-3, gug to guc (g474c) and aug to acg (t494c); these mutations have previously been shown to eliminate expression of the truncated forms while leaving a functional IF2-1 [21]. We shall refer to this allele as infB(del2/3) to indicate that the mutations prevent expression of IF2-2 and IF2-3 even though del2/3 is not a deletion mutation. The resulting infB del1, del2/3, and wild-type (wt) alleles were introduced into the transposon site as part of the nusA infB operon (<nusAinfB> to signify that this is encoded within the transposon).
The natural infB allele could be readily knocked out by introducing the del(infB)1::tet allele when the operon in the transposon had infB(wt), infB(del1), and infB(del2/3) alleles. (The procedures for verifying deletion of the natural infB allele as illustrated in Figure 3C and for verifying infB alleles by PCR and sequencing will be described under Materials and Methods.) The <infB(del2/3)> and especially the <infB(del1)> strains display some measure of cold sensitivity, growing very slowly at 25°C and below, consistent with previous reports about strains with analogous alleles [21]. We determined that the strain with the single copy <infB(del1)> as sole allele was highly sensitive to MMS whereas the strains with <infB(wt)> and <infB(del2/3)> as sole alleles were not (Figure 4A and Figure S1A).
The results indicate that the del1 mutation causes the inability to grow in the presence of MMS. The question is whether this is due to a general deficiency in repair, recombination, and restart functions, resulting from a generally deficient translation initiation function, or whether there is any specificity of the defect. We should note that the <infB(del1)> strain (ArgA−) was at least moderately proficient in homologous recombination measured by P1 transduction, although the frequency of Arg+ transductants was reduced approximately 5 fold compared to <infB(wt)> and <infB(del2/3)> strains (Figure S1B). While some reduction in homologous recombination frequency may be part of the phenotype of this strain, the reduction seen here is modest compared to the 20–50 fold reduction in P1 transduction demonstrated for the priA knockout strain [31].
To determine whether it is indeed IF2-1 that is needed to maintain MMS-resistance, we complemented the <infB(del1)> allele of strain GTN1156 with the infB(del2/3) allele, harbored as part of a nusA infB operon on a plasmid with a pSC101 replicon, pSPCnusAinfB(del2/3). While the empty plasmid vector could not confer MMS-resistance and homologous recombination proficiency, IF2-1 expressed from pSPCnusAinfB(del2/3) did restore high viability on MMS plates (Figure 4A). In contrast, IF2-2/3 expressed from the plasmid-borne infB(del1) allele only partially restored viability on MMS plates (Figure 4A). While the multicopy infB(del1) allele did increase dramatically the viable count on MMS plates, the colonies grew up very slowly, and the viable count on these plates was still 5–10 fold lower than that of the strain with the multicopy infB(del2/3) allele (Figure 4A). These results illustrate functional differences between IF2-1 and IF2-2/3 in promoting recovery after MMS treatment. They also indicate that IF2-2/3 when expressed from a multicopy vector may compensate for the lack of IF2-1, albeit inefficiently.
To confirm that it was not just the DNA segment deleted in the infB(del1) allele but the full-length IF2-1 protein that was needed for complementation, we introduced IF2 G domain mutations, infB(c1227a) or (c1501a), which result in the IF2 D409E and D501N alterations, respectively, into pSPCnusAinfB(del2/3). The infB(D409E) allele is an example of a viable G mutant that is functional at 37°C [32] whereas infB(D501N) is a recessive allele that is lethal as a single-copy gene [33]. Introduction of pSPCnusAinfB(del2/3,D409E) into GTN1156, but not pSPCnusAinfB(del2/3,D501N), restored MMS-resistance (Figure 4A). IF2-1 must therefore be providing the function needed for viability in MMS. The level of homologous recombination in the <infB(del1)> mutant, examined by P1 transduction, could also be increased by supplying the various infB alleles on the plasmid vector (Figure S1C). Due to the relatively modest effect on homologous recombination, this aspect of the infB(del1) mutant was not further examined.
The <infB(del1)> strain, which produces only IF2-2/3 and has extremely low viability in the presence of 6 mM MMS, attains high viability when complemented with the plasmid-borne infB(del2/3) allele, which restores IF2-1 production (Figure 4A and 4B). This indicates that the multicopy infB(del2/3) allele is dominant over the <infB(del1)> allele. The inability of pSPCnusAinfB(del2/3,D501N) to restore efficient growth of the <infB(del1)> strain in MMS could indicate the inactivation of a necessary function of IF2-1 by the D501N mutation. Alternatively, the <infB(del1)> allele may be dominant negative over the multicopy infB(del2/3,D501N) allele in terms of supporting growth in MMS.
Although the D501N mutation is lethal when present as a single-copy infB allele, this mutation is recessive to the wild-type allele [33]. We therefore tested whether the multicopy infB(del2/3,D501N) allele on the plasmid could support viability by itself. The natural infB allele in strain GTN932 that bears plasmids pSPCnusAinfB(del2/3), pSPCnusAinfB(del2/3,D409E), or pSPCnusAinfB(del2/3, D501N) could readily be knocked out, leaving the infB on the plasmid as the sole allele in the cell. This allowed us to test whether or not IF2-1(D501N), expressed from multicopy <infB(del2/3, D501N)>, is defective in a function that IF2-1 provides but IF2-2/3 fails to perform. Although the strain with the multicopy infB(del2/3,D501N) as the sole allele grew relatively slowly, requiring at least twice the incubation time as the other two strains for growth, it was clearly viable and also retained significant viability on MMS plates, comparable to viability of analogous strains with infB(del2/3) and infB(del2/3,D409E) as sole alleles (Figure 4B). That is, the multicopy infB(del2/3,D501N) is able to support high viability in MMS so long as the <infB(del1)> allele is absent. Introduction of the D501N mutation to the multicopy infB(del2/3) allele thus results in loss of dominance over <infB(del1)>, not in the loss of a function needed to maintain viability in MMS. These results suggest that IF2-2/3, at levels produced from the single-copy <infB(del1)> allele, is performing a function in a way that aggravates problems which the cells encounter during growth in MMS, outcompeting IF2-1(D501N) that is able to carry out the function appropriately to maintain viability. In other words, IF2-2/3 does not necessarily lack the capacity to perform the IF2-1 function. Rather, it appears to carry it out in a way that dramatically reduces viability. That is, the recessive properties of infB(D501N) with respect to the infB(del1) allele, including its ability to support resistance to MMS as the sole multicopy allele, suggest that MMS sensitivity of the <infB(del1)> strain is not simply due to a general deficiency in translation initiation function when only IF2-2/3 is present.
We next determined whether the MMS sensitivity of the <infB(del1)> mutant reflected deficiency in the levels of repair or restart proteins in these mutants. In the analysis described above, MMS resistance was measured by growth of cells on plates containing MMS. By this analysis, cells must not only survive initial exposure to the DNA-damaging agent but also grow into colonies in its presence. We also measured the ability of strains exposed to MMS to recover and grow in the absence of MMS in order to assess their capacity to repair DNA lesions and restart DNA replication. Strains that are defective in genes such as priA, recA, and polA that participate in DNA repair or replication restart are known to be quite sensitive as measured by initial exposure for 15 min in MMS and plating without MMS to determine the number of survivors; the alkA tag mutant, which is defective in a major mechanism for repairing alkylated bases (base excision repair), is also sensitive to MMS by this criteria [34]. The <infB(del1)> mutant was as resistant to MMS as infB(wt) strains by this criteria (Figure 4C), with MMS resistance comparable to strains with natural infB, <infB(wt)>, and <infB(del2/3)> alleles; in contrast the recA938 mutant was highly sensitive by this criteria (Figure 4C). It should be noted that when cells were deficient in both the PriA-PriB and PriA-PriC pathway (deficient in both PriB and PriC), they had very low viability even without MMS treatment (Figure S2A). As restart mutants tend to have very low viability even without MMS, we measured MMS sensitivity of a del(dnaT)759::kan mutant with a dnaC(a491t) suppressor mutation, which greatly increases cell viability. Even with the suppressor mutation, the dnaT knockout strain was significantly more sensitive to the 15-min MMS treatment (Figure S2B) than the <infB(del1)> mutant. These results indicate that levels of repair and restart factors in the <infB(del1)> strain are sufficient for the recovery of DNA replication and cell growth after DNA damage by MMS. However, there was a 1000-fold reduction in viability of the <infB(del1)> mutant on 6 mM MMS plates (Figure 5A). That is, the <infB(del1)> strain is proficient in repairing DNA damage and resuming DNA replication after the 15-min exposure in MMS, but it is severely defective in its ability to sustain growth in MMS. Thus, the <infB(del1)> mutant is not able to cope with the sustained damage to DNA during cell growth. This could indicate that a repair or restart factor, although not deficient, is sufficiently low such that it cannot keep up with constant DNA damage inflicted on MMS plates; alternatively, it is possible that the regulation of repair and restart processes are not appropriate for efficiently supporting DNA replication under these conditions.
Introduction of the sulA::Mud(lac, Ap, B::Tn9) allele greatly restored viability of the <infB(del1)> mutant in MMS (Figure 5A). The sulA gene, which is a component of the SOS system induced by DNA damage, is a cell division inhibitor [35]. In mutants such as the priA null strain, which has a constitutively induced SOS system, the high expression of sulA results in loss of viability, which can be largely restored by sulA mutations [36]. It is important to note that the sulA::Mud(lac, Ap, B::Tn9) allele did not fully restore viability to the <infB(del1)> mutant. Moreover, scorable colonies on MMS plates required incubation for over 36 hours at 37°C whereas infB(wt) colonies readily arose in 16 hours. That is, the sulA mutation did not fully revert <infB(del1)> to the wild-type phenotype.
Interestingly, the priA300 mutant had a phenotype much like the <infB(del1)> mutant, resistant to MMS when exposed to MMS and plated in its absence but highly sensitive when plated on 6 mM MMS plates (cf. Figure 4C with Figure 5A). The priA300 allele encodes for a helicase-deficient PriA that is fully proficient in primosome and replisome assembly by the PriA-PriB pathway [19], [37]. The priA300 mutant has previously been shown to have essentially a wild-type phenotype unless that mutation is combined with mutations affecting other restart functions such as priB; wild-type properties of the priA300 mutant include homologous recombination proficiency and relatively high UV resistance [19], [38]. As with the <infB(de11)> mutant, the sulA::Mud(lac,Ap,B::Tn9) allele could restore viability of the priA300 mutant in MMS (Figure 5A). In addition, priA300 was epistatic with the infB(del1) allele, causing no significant increase in MMS sensitivity (Figure 4C and Figure 5A). In contrast, the <infB(del1)> del(priB)302 combination (GTN1117) was synergistic, reducing viability to 0.010±0.002% on the MMS plates. The priB knockout alone did not have such a severe effect; the <infB(wt)> del(priB)302 strain (GTN1133) had a viability of 43±8% on MMS plates. In addition, knockout of priC did not increase UV sensitivity; the <infB(wt)> del(priC)752::kan strain (GTN1059) had a viability of 84±7% on MMS plates. These results indicate that the PriA-PriC pathway, which requires PriA helicase, is not solely responsible for allowing cell growth in the presence of MMS and that the PriA-PriB pathway most likely makes a significant contribution to mechanisms dependent on PriA helicase as well. We shall further examine the interactions of priA300 and del(priB)302 with the <infB(de11)> and <infB(de12/3)> alleles by UV sensitivity. The epistatic relationship between the infB(del1) and priA300 alleles suggests that the loss of IF2-1 specifically affects the activity or engagement of factors in restart pathways that require PriA helicase.
Despite the high MMS sensitivity of the <infB(del1)> strain, it did not resemble the priA knockout mutant in terms of having constitutively high levels of SOS induction (Figure 5B). Expression from the sulA::lacZ SOS reporter was significantly lower than the strain with wild-type priA and infB and the priA300 strain. The latter strain had moderate basal levels of SOS induction, which was significantly less than that of the priA knockout. Treatment of the wild-type and priA300 strains with 18 mM MMS elicited moderate increases in SOS expression; in contrast, treatment of the <infB(del1)> strain elicited over a 10-fold increase in SOS expression, consistent with the role of SOS induction reducing the strain's viability upon MMS treatment,
Although the <infB(del1)> strain was sensitive to growth in MMS, it was slightly more resistant to UV light than the <infB(wt)> strain (Figure 6A). In fact, the <infB(del2/3)> mutant, which was found to be the most MMS-resistant, was slightly more UV sensitive than the <infB(wt)> strain (Figure 6A). These results do not rule out the possibility that the del1 and del2/3 mutations impair or knock out restart mechanisms engaged after UV irradiation. As there are multiple restart pathways in the cell, the PriA-PriB and PriA-PriC pathways being the two major ones [18], the del1 or del2/3 mutation may predominantly affect only one pathway but not the other. To test this possibility, we examined the effect of the infB alleles in combination with priB or priC knockout alleles.
It is well established that the knockout of priB or priC has little to no effect by itself [39] in contrast to the priA or dnaT knockouts, which affects both major restart pathways and elicits high sensitivity to DNA-damaging agents and low viability [18], [36], [40]. As expected, neither the priB nor priC knockout had any effect on UV sensitivity when introduced into the parent strain (GTN932) used to construct the various <infB> mutants (Figure 6D). While the del(priC)752 allele had absolutely no effect on single-copy <infB(wt)>, <infB(del1)>, and <infB(del2/3)> strains (Figure 6C; cf. with Figure 6A), the del(priB)302 clearly had a synergistic effect with the infB(del1) mutation to elicit relatively high UV sensitivity (Figure 6B). This finding that the priB knockout, but not the priC knockout, is synergistic with the <infB(del1)> allele to increase UV sensitivity indicates that the loss of full-length IF2-1 diminishes the PriA-PriC pathway for recovery after UV irradiation. Introduction of a pBAD24-priB plasmid into the del(priB)302 <infB(del1)> strain (GTN1117), allowing the expression of PriB driven by the PBAD promoter with arabinose as inducer, increased its UV resistance to levels comparable to the del(priB)302 <infB(wt)> strain (GTN1133; Figure 7A), confirming that the deficiency of GTN1117 can be reversed by expressing PriB. This indicates that the activity of repair and restart proteins needed for recovery after UV irradiation in GTN1117, which has the <infB(del1)> allele, is comparable to that in GTN1133, which has the <infB(wt)> allele. Therefore, the increased UV sensitivity of GTN1117 with respect to GTN1133 is most likely due to some type of deficiency in the PriC-dependent pathway. We were unable to measurably increase UV resistance by expressing PriC from pBAD24-priC (Figure 7A). Indeed, PriC in its active form must be present in GTN1117. When the chromosomal priC was knocked out in pBAD24-priC/GTN1117 (GTN1566), expression of PriC from the plasmid vector became essential for viability with or without pre-treatment with MMS (Figure S2A), viability being less than 0.1% in the presence of glucose. In the presence of arabinose, viability of GTN1566 with or without MMS treatment was comparable to the strain with an intact chromosomal priC. That is, active PriC can be expressed from pBAD24-priC or the chromosomal priC gene in the <infB(del1)> genetic background, and supplementation of PriC expression in GTN1117 from the plasmid cannot restore any measure of UV resistance. These results suggest that its relatively high UV sensitivity is not caused by a deficiency in PriC, PriA, and DnaT.
Although the del(priB)302 <infB(del1)> strain (GTN1117) has high UV sensitivity, its ability to recover after a 15-min exposure to MMS was comparable to the wild-type control (Figure 4C). Moreover, Mu plating efficiency on this strain is not dramatically reduced, indicating that the PriA-PriC pathway can promote Mu replication in the absence of IF2-1 (Figure S3), a result consistent with properties of Mu replication in vitro [5]. In general, the Mu plating efficiencies on the various <infB(wt, del1, or del2/3)> strains, whether in the PriB+PriC+, del(priB)302, or del(priC)752::kan genetic backgrounds, were nearly the same. These results indicate that restart proteins needed to promote Mu replication by the PriA-PriC pathway are present at sufficient levels to support lytic development. They also suggest that the defect of the infB(del1) allele is not a deficiency in restart activity needed for recovery but rather in the regulation of restart activity needed to maintain replication in the presence of the DNA-damaging agent.
Although the effect is not as much as in the <infB(del1)> background, the del(priB)302 allele also did significantly increase UV sensitivity when introduced into the <infB(wt)> background (cf. the solid square data points in Figure 6A and 6B) whereas it had essentially no effect in the natural infB(wt) background (Figure 6D). This may reflect a small change in relative or absolute levels of full-length and truncated IF2 when the infB allele is expressed from the transposon site, a change that has no discernible effect unless specific restart mechanisms are inactivated as with the del(priB)302 mutation. Interestingly, in the <infB(wt)> background both the priA300 (Figure 7B) and the del(priB)302 (Figure 6B) allele increased UV sensitivity to the same level. Like the del(priB)302, the priA300 allele is known to have little effect on UV sensitivity [19], and indeed we found essentially no effect of the priA300 allele in the GTN932 background (Figure 6D), which has the natural infB(wt) allele. As we described above, the priA300 and <infB(del1)> alleles both independently elicit sensitivity to growth in MMS, and the two mutations are epistatic for this trait, consistent with a model in which PriA helicase and IF2-1 function in the same pathway to maintain efficient growth in MMS. In the UV sensitivity analysis, the infB(del1) allele was also found to be epistatic with priA300, not being able to elicit further UV sensitivity in the priA300 background (Figure 7B). That is, loss of IF2-1 attenuates pathways dependent on PriA helicase such as the PriA-PriC pathway. In contrast, the infB(del2/3) allele was synergistic with priA300 to increase UV sensitivity (Figure 7B). The results indicate that loss of IF2-2/3 from the infB(del2/3) allele results in deficiency of a restart pathway that is distinct from the IF2-1/PriA helicase-dependent pathway. Mu plating efficiency on the three priA300 strains with the <infB(wt)>, <infB(del1)>, and <infB(del2/3)> were essentially the same, the titer obtained on the latter two strains being greater than 90% of the titer on the priA300 <infB(wt)> strain. Thus, as with the del(priB)302 <infB(del1)> combination, which also synergistically contributes to high UV sensitivity, the priA300 <infB(del2/3)> combination does not lead to an inability to initiate Mu replication by the available host restart machinery.
What is notable about the UV sensitivity analysis is that the combination of priA300 <infB(del2/3)> or del(priB)302 <infB(del1)> mutations does not produce the extremely severe phenotype of the priA300 del(priB)302 combination, which elicits a phenotype analogous to the priA knockout [19]. That is, loss of IF2-1 or IF2-2/3 does not result in the inability to promote replication restart by the respective pathways they influence, but rather the loss of each IF2 isoform affects some mechanism needed to maintain high viability when the restart mechanism is engaged after DNA damage. However, under normal growth conditions or if cells are allowed to recover after MMS treatment or UV irradiation without the presence of DNA damaging agents, there is little effect of knocking out IF2 isoforms, and a mild effect is seen when these mutations are combined with the restart mutation del(priB)302 or priA300, which by itself has little effect under normal growth conditions. We examined the cell morphology of the various infB mutants to examine whether there is an increased incidence of sporadic SOS induction, leading to filamentation [41] of a small fraction of the cells in the population.
The strains with the single del(priB)302 or <infB(del2/3)> mutant had essentially wild-type morphology (Figure S4A), 100% of cells being 0–8 µm in length when at least 40,000 cells were analyzed. Cells with the single <infB(del1)> or priA300 mutation (GTN1114 and GTN1298, respectively) tended to be longer in size, with a higher incidence of moderate sized filaments (examples of moderate filaments are indicated by white arrows). In a sample of 1000 cells, 1% of the cells were in the 8–30 µm range for GTN1114 and GTN1298. Consistent with the relatively low basal levels of SOS expression measured for the <infB(del1)> mutant at the macroscopic level (Figure 5B), the level of its filamentation was quite low compared with that of the priA knockout mutant (Figure S4B), but the moderate filamentation suggests an increased incidence of sporadic SOS induction.
What was notable for the synergistic <infB(del1)> del(priB)302 combination (GTN1117) was that it gave rise to a low but significant frequency of very large filaments greater than 30 µm (Figure S4A). The incidence of filaments over 30 µm in size was found to be 0.13% in a screening of 33,000 total cells, most of these large filaments (0.10% of total cells) being over 50 µm in length. Only one other combination of an infB allele with the priA300, del(priB)302, or wild-type restart functions (Figure S4A) yielded any filaments over 50 µm in 100,000 cells screened. The mutant with the synergistic <infB(del2/3)> priA300 combination (GTN1297) produced filaments greater than 30 µm at a significantly lower frequency of 0.02% in a screening of 100,000 cells, of which only 3 were greater than 50 µm. Filaments in the 30–50 µm range also arose with the single <infB(del1)> or priA300 mutants (GTN1114 and GTN1298, respectively) but with a frequency of no more than one in 40,000 cells. No filaments of greater than 30 µm were detected with the <infB(wt)> (GTN1050), <infB(del2/3)> (GTN1115), del(priB)302 (GTN1133), and the <infB(del1)> priA300 (GTN1323) strains when at least 100,000 cells were examined. The results indicate that the <infB(del1)> del(priB)302 mutant (and, to a lesser extent, the <infB(del2/3)> priA300 mutant) has an increased incidence of very high SOS induction (leading to the formation of giant filaments) in a small fraction of the cell population growing in LB, suggesting a reduced capacity to cope with accidents that might occur during DNA replication for normal cell growth. However, these mutants clearly do not have the characteristics of extensive SOS induction as with a priA knockout strain such as GTN430 (Figure S4B; 3% of cells producing filaments greater than 30 µm in a sample of 4000 cells, 2% greater than 50 µm).
The characteristics of strains such as the <infB(del1)> or priA300 mutant are more akin to a priA knockout strain that has acquired a suppressor mutation in dnaC (Table 1). GTN412, which is a Mucts62 lysogen, can support Mu replication upon thermoinduction to yield a high level of infective centers, has a high level of viability on MMS plates, and has a relatively low level of expression from its SOS reporter gene (dinD::lacZ). Introduction of the priA knockout decreased viability on MMS and formation of Mu infective centers by several orders of magnitude. The presence of a suppressor mutation in dnaC (GTN522) did diminish cell filamentation (Figure S4B; the number of filaments >30 µm are reduced to 0.05% from 3%, measured in a sample of 15,000 cells), reduce the level of SOS induction as indicated by the dinD::lacZ reporter, and restore the ability to form Mu infective centers, but this strain retained the severe sensitivity to growth in the presence of MMS, a central feature of the both the <infB(del1)> and priA300 mutants. This is consistent with the ability of the dnaC suppressor mutation to bypass the requirement for PriA to initiate DNA synthesis at forked DNA structures [38], [42]; however, without PriA the mechanism for promoting replication restart and promoting high viability in the presence of MMS (the IF2-1/PriA helicase-dependent pathway) appears to be compromised. In the same way, a priA300 or <infB(del1)> mutant may be able to promote replication restart by a less preferred pathway, which may permit replication restart to proceed but does not do so in a way that supports high viability during growth in the presence of MMS.
The present work indicates a special relationship between the PriA helicase function and IF2-1 (see Table 2). Both the PriA helicase function and IF2-1 are required to allow cells to grow with maximal viability in the presence of MMS. Nevertheless, neither of these mutants display the severe characteristics of the priA knockout, having UV resistance that is comparable to wild type and being able to recover from MMS treatment with very high viability provided that it can do so in the absence of MMS. The defect of the priA300 mutant, previously shown to have nearly a wild-type phenotype [19], is a surprising new phenotype, being defective in the ability to grow in the presence of MMS but not in its ability to recover from MMS treatment. Even more surprising is the finding that the loss of the IF2-1 function elicits the same phenotype. Another characteristic which indicates that the infB(del1) allele affects some aspect of replication restart is the suppressing effect of knocking out sulA, a mutation that greatly increases viability of both the infB(del1) and priA300 mutant on MMS plates. Moreover, MMS treatment of inf(del1) mutant promotes an especially high level of SOS induction compared to the level promoted in wild type.
A relationship between full-length and truncated IF2 isoforms and replication restart functions is further indicated by UV sensitivity analysis. Both the <infB(del1)> and <infB(del2/3)> exhibit UV resistance comparable to wild type, but the combinations of <infB(del1)> del(priB)302 and <infB(del2/3)> priA300 significantly enhance UV sensitivity. Moreover, the <infB(del1)> del(priB)302 mutant (and, to a lesser extent, the <infB(del2/3)> priA300 mutant) display an increased frequency of sporadic SOS induction, indicated by the increased frequency of very long filaments over 30 µm. Clearly, the general population of these cells do not display the same high level of SOS induction of the priA knockout cells at the macroscopic level. The sporadic nature of filamentation is consistent with the thinking that these cells are mostly proficient in coping with accidents of DNA replication which may arise during normal growth conditions, unlike the priA knockout that copes with such accidents poorly. One would expect that only a small minority of cells would need to cope with a large number of DNA lesions during growth in LB unless a DNA-damaging agent such as MMS is present. The combination of <infB(del1)> del(priB)302 and <infB(del2/3)> priA300 alleles may sufficiently attenuate the major pathways that lead from DNA damage to replication restart, thus manifesting a modest but significant increase in sensitivity to UV irradiation.
The epistatic relationship between the priA300 and infB(del1) alleles revealed by both UV sensitivity and viability on MMS plates indicates that IF2-1 and PriA helicase function in common pathways as proposed in Figure 8A. This includes mechanisms in both the PriA-PriB and PriA-PriC pathway, for neither the priB or priC knockout has the severe effect of priA300 for growth on MMS plates. What remains of the major restart pathways when PriA helicase is inactive are mechanisms in the PriA-PriB pathway that can operate in the priA300 background [19]. Thus, the effect of the infB(del2/3) allele in this genetic background (increased UV sensitivity and increased incidence of sporadic cell filamentation) suggests that IF2-2/3 plays a role in this pathway. However, we have yet to find a phenotype for the infB(del2/3) allele alone, comparable to MMS sensitivity of the infB(del1) mutant, and whether IF2-2/3 is a key participant in PriA helicase-independent restart mechanisms (Figure 8A) remains be determined.
Finally, the characteristics of the <infB(del1)> and priA300 mutants and especially the infB(del1) del(priB)302 double mutant are more like the priA knockout with a suppressor mutation in dnaC rather than the priA knockout with no suppressor. The <infB(del1)> and priA300 mutants, like the priA knockout with suppressor, do not exhibit the extreme sensitivity to UV irradiation, the massive cell filamentation, and the inability to support Mu replication that is characteristic of the priA knockout. Nevertheless, all of these mutants are not able to grow efficiently on media containing 6 mM MMS, their viability on MMS plates being approximately 0.1% or less. For the priA knockout, the dnaC suppressor allows replication restart to proceed, but the bypass of the restart proteins compromises maintenance of high cell viability when DNA replication proceeds during relatively high rates of DNA damage. Similarly, replication restart mechanisms can still operate in the <infB(del1)> mutant, and the lack of IF2-1 may bypass the preferred pathway that maintains high cell viability during growth in the presence of MMS. As IF2-1 and IF2-2/3 share 726 common residues, IF2-1 having 157–164 extra residues at the N-terminal end, it is quite conceivable that IF2-2/3 can replace IF2-1 in the IF2-1/PriA helicase-dependent pathway, allowing replication restart to proceed but lacking the function need to maintain high cell viability. The ability to grow under conditions that damage DNA at elevated levels could provide cells with the selective advantage that conserves the function of restart proteins despite the fact that suppressor mutations can bypass the need for these proteins. For example, the fact that the helicase motif of PriA is highly conserved among diverse bacteria [43] has been puzzling in light of the fact that its inactivation by the priA300 mutation seemed to have little effect on the cell phenotype, but the ability of cells with active PriA helicase to grow under conditions that damage DNA at a relatively high rate would indeed be a selective advantage that would conserve this motif.
The phenotype of the <infB(del1)> mutant raises the question of what IF2-1 could be doing to influence cellular recovery after DNA damage by a PriA helicase-dependent pathway. First, IF2-1 and IF2-2/3 could have different preferences for mRNAs such that IF2-1 specifically promotes the translation of factors needed to support this pathway. Such a mechanism would be novel as such a role of the various IF2 isoforms in promoting differential gene expression has yet to be described. Second, IF2 may act as a transcription factor and the various IF2 isoforms may have different activity in this regard such that IF2-1 is specifically needed to regulate expression of genes needed for PriA helicase-dependent pathways. The finding that IF2 can selectively promote transcription of rRNA by RNA polymerase in vitro [26] and the identification of a region in the carboxy terminal region of IF2 with nonspecific DNA binding activity [27] have prompted the proposal that IF2 has activity influencing transcription. Third, IF2 has been shown to have molecular chaperone activity [25]. The IF2 isoforms may ensure that specific factors in their respective pathways are active when required. We have previously proposed a role of IF2 as a chaperone performing a function much like ClpX (Figure 1B–1C) where IF2 binds to a Mu end and prepares the DNA template for assembly of restart proteins, a process beginning with displacement of IF2 from DNA by PriA helicase. The analysis of this present work cannot definitively establish that any one of these possibilities is the basis for IF2's influence on cellular restart mechanisms; however, we favor the third mechanism in which IF2 acts as molecular chaperone, based on the role of IF2 in bacteriophage Mu replication in vitro [5], [17], the phenotype of the infB(del1) mutant, and the relationship of this allele with priA300.
A key question regarding the function of IF2-1 is, why does its loss lead to a severe decrease in viability during growth on MMS despite the fact that the cell remains proficient for supporting replication restart? We suspect that the loss of the preferred IF2 isoform for a restart mechanism, loss of PriA helicase activity, or the complete loss of PriA in the presence of a dnaC suppressor results in the inability to fine-tune the progression of restart pathways, a level of regulation that becomes essential when cells must grow under conditions that damage DNA at a high rate. If we speculate that the role of IF2 in Mu replication in vitro is applicable for cellular restart mechanisms, we can illustrate the type of regulation that IF2 might exert (Figure 8B).
An important difference between IF2-1 and the truncated forms IF2-2/3 for the assembly of restart proteins at stalled forks may be the mechanism by which they respond to a hypothetical go-ahead signal for restarting DNA replication. When DNA damage is accumulating at a relatively high rate, a mechanism that regulates restart by preventing re-establishment of the replication fork until the template is relatively free of DNA damage may ensure efficient DNA replication in the presence of a DNA-damaging agent. For example, restarting DNA replication before the DNA is relatively free of lesions will only result in the stalling of the fork again, causing delay in establishing a productive replication fork and thus inducing a high level of SOS response that may become toxic.
These considerations are reminiscent of the findings of Flores et al. [44], who determined that priA300 greatly diminishes viability of the holDG10 mutant. The holD gene encodes the Psi unit of DNA polymerase III holoenzyme, and the mutant Psi causes frequent replication fork stalling. That is, the effect of priA300 becomes discernible only when the rate of replication fork stalling becomes high. As noted by Flores et al. [44], the deficiency in PriA helicase caused by the priA300 mutation may lead to the inability to promote duplex opening on the DNA substrate for DnaB helicase loading and replisome assembly [29]; alternatively, another function of PriA besides the helicase could be inactivated by the priA300 mutation, leading to the inability to cope with frequent fork arrest in the holDG10 mutant. The PriA function needed to sustain high viability of the holDG10 mutant may be related to the pathway in which both IF2-1 and PriA helicase play a role. When cells must grow in the presence of MMS, the action of PriA helicase to displace IF2-1 may play a critical function to ensure maximal cell viability, or conceivably, the inactivation or attenuation of another function by priA300 may prevent what we call the IF2-1/PriA helicase pathway from operating optimally. This example underscores the possibility that PriA helicase as well as IF2-1 play multiple roles for replication restart, some of which may be part of their mutual participation in the IF2-1/PriA helicase pathway and some of which may not. PriA helicase may play important roles in duplex opening for DnaB loading as well as displacement of IF2 to initiate replication restart, but only the latter may be essential for the IF2-1/PriA helicase pathway.
The role of IF2 isoforms in influencing replication restart mechanisms has important implications for how replication restart and the maintenance of genome stability may be regulated with respect to cell physiology. As a translation factor, IF2 has a strong influence on cell growth and progression through the cell cycle while responding to cellular signals such as the alarmone (p)ppGpp [45], which is an indicator of nutritional deprivation. Depending upon the physiological status, how replication restart is carried out can be critical in determining cell viability, and IF2 may respond to cellular signals to determine the conditions for restart. The IF2 function in translation is a highly conserved one found in all living cells [46], [47]. Its role in influencing pathways for maintaining genome integrity prompts the question whether this general function has been conserved in other organisms to play some function in coordinating replication, recombination, and repair functions with respect to growth conditions.
All experimental analysis was conducted with derivatives of GTN932 (Hfr del(gpt-lac)5; see Table S1), an E. coli K-12 strain that is a derivative of PK191 [48]. We have conducted PCR and sequencing analysis to verify that this line of strains have wild-type relA, not the relA1 allele [49] as sometimes reported for PK191 strains. The del(priB)302 and priC303::kan alleles from JC19272 [39], priA2::kan from PN104 [36], del(priC)752::kan from JW0456-1, del(dnaT)759::kan dnaC(a491t) from JW4336-2, and del(argA)743::kan from JW2786-1 [50] were introduced into bacterial strains by P1vir transduction as previously described [39]. Inheritance of del(priB)752::kan was screened by PCR analysis with primers PriBupper and PriBlower (Table S2). The priA300 was introduced by P1 transduction, first transferring the metB1 allele by selecting for the closely linked btuB3191::Tn10 from CAG5052; the priA300 was then transferred from SS97 by selecting for Met+ transductants (tetracycline-sensitive transductants were chosen) [19], which were screened by PCR amplification with primers PriA-Nseq and PriA-Cseq and sequenced with revPriA820 primer. The sulA::Mud(lac,Ap,B::Tn9) from SS97 [18] or dinD1::Mud1(lac,Ap) from PN104 [36] was introduced into strains by P1 transduction and selection on ampicillin plates; transductants were screened for disruption of the sulA or dinD genes with primers sulAupper and sulAlower or dinDupper and dinDlower, respectively. The clpX::kan strain was constructed as previously described [15].
The del(infB)1::tet allele was constructed by first integrating a single copy nusAinfB operon into a random site of the host chromosome as part of the EZ-Tn5 transposon. The natural infB cistron was precisely excised and replaced with a tetR cistron from pACYC184 [51], using recombineering methods [52] to generate the del(infB)::tet allele. As recombination events at the natural infB site were very difficult to isolate, we created a PCR template to generate the del(infB)::tet allele, with approximately 1-kb of DNA from upstream and downstream of infB to flank the tet cistron. This template on the pGEM-Teasy vector (Promega) was amplified using PfuUltra High Fidelity DNA polymerase (Stratagene) using the primers nusLower and rbfUp2, and the PCR product was used to transform heat-induced DY330 flgJ::<nusAinfB-kan>.
The various flgJ::<nusAinfB-cat> alleles were constructed by introducing infB mutations into the nusAinfB operon harbored on the EZ-Tn5 transposon. The transposon was from the pMOD-6<KAN-2/MCS> purchased from Epicentre, and it was introduced into DY330 as a transpososome according to the instructions of the manufacturer. The transposon was determined to be integrated in the flgJ gene by a single primer PCR and sequencing method [53]. For introduction of various infB alleles at the transposon site, the transposon was modified by recombineering [52]. Heat-induced DY330<KAN-2/MCS> was transformed with a PCR product made by amplifying the cat gene of pACYC184 with primers DelMOD6Cat and lowerKanCat (see Table S2). The resulting strain DY330<del(kan)::cat)>, which is chloramphenicol-resistant and kanamycin sensitive, serves as the strain for introducing various alleles at this site.
PCR products for introducing the nusA infB operon at the transposon were made using pMOD-6<KAN-2/MCS> constructs as template. The nusA infB operon, amplified from the E. coli chromosome using PfuUltra High Fidelity with primers argRmetYp2 and IF2BamHI, was cloned between the SphI and XbaI site of pMOD-6<KAN-2/MCS> (promoter side of the operon is proximal to the SphI site). Various infB mutations were introduced into the resulting plasmid. The operon was then amplified using primers lowerMod6Tn and antiSqRP, and the PCR product was used to transform heat-induced DY330<del(kan)::cat)>, selecting transformed cells on LB plates containing 25 µg/ml kanamycin and screening for chloramphenicol sensitivity. To construct versions of these flgJ::<nusAinfB> alleles that encode chloramphenicol rather than kanamycin resistance, heat-induced DY330 flgJ::<nusAinfBkan> strains were transformed with PCR products made by amplifying the cat gene of pACYC184 with primers upperKanCat and lowerKanCat. This inactivates the kan gene while leaving intact the nusAinfB contained within transposons. The resulting constructs were always verified by sequencing as described below.
We could readily knock out the natural infB allele of a strain with the <nusAinfB(wt, del1 or del2/3)> cassette by introducing the del(infB)::TetR allele. As the expression of tetracycline resistance was relatively feeble from this allele, introduction of the knockout was most conveniently done by co-transduction with the closely linked argG; Arg+ transductants of a del(argG)781::kan recipient strain co-inherited the del(infB)::TetR allele at a frequency greater than 80%, provided that an infB allele which supports cell viability was provided from another site. Even when the second infB function was supplied by pSPCnusAinfB(del2/3,D501N), greater than 80% of the Arg+ transductants coinherited del(infB)1::tet allele, indicating that the multicopy infB(del2/3,D501N) can maintain cell viability (the presence of the D501N mutation in the sole infB allele was verified by sequencing). When the second nusAinfB operon was present on the chromosome, it was introduced into the transposon inserted in flgJ. The various flgJ::<nusAinfB-cat> alleles were constructed by recombineering methods in DY330 as described above and transferred to other strains by P1vir transduction. The nusAinfB operon contained within the transposon includes all three ArgR binding sites (see Figure 3A) and extends to the stop codon for infB.
As the nusAinfB operon in the transposon lacks downstream genes such as rbf in the natural operon, the infB alleles at the natural site and the transposon in flgJ can be separately amplified for DNA sequencing (Figure 3A–3C; primers p1 and p2 for the natural site and p1 and p4 for the transposon site). Thus, the presence of infB at the natural site could readily be detected by primers (p1 and p2) annealing to sites flanking infB to yield a 4.7-kb band (Figure 3C, lanes 1, 3, and 9), confirmed by 2.8-kb band yielded by one primer (p3) annealing within infB and one (p2) downstream of the gene (lanes 2, 4, and 9). (See the list of primers in Table S2.) Knockout of the natural infB, in contrast, could be detected with the formation of a 3.2-kb band with primers p1 and p2 (lanes 5 and 7) and no bands (lanes 6 and 8; cf. with lanes 2, 4, and 10) with p3 and p2.
We found this to be the best method for constructing strains with various single-copy infB alleles, for the replacement of the wild-type infB allele at the natural site proved to be very difficult. As constructed strains were suspected to be potential restart mutants, their dnaC allele was sequenced to determine whether any suppressor mutations have accumulated there [39]. None of the mutants we isolated had as severe a phenotype as the priA null mutant, and no suppressor mutations in dnaC were detected.
All pSPCnusAinfB plasmids with various infB alleles were constructed using the pBAD43 plasmid vector (a gift from Dr. Jonathan Beckwith, Harvard Medical School) [54]. This plasmid is a relatively low copy plasmid, having a pSC101 plasmid origin and conferring spectinomycin resistance. The nusAinfB operon, amplified by PCR using primers p1nusAinfB and IF2BamHI (see Table S2) and PfuUltra High Fidelity DNA polymerase, was inserted into the NsiI-BamHI site of the pBAD43 vector. The ara and PBAD sequences required for arabinose-based gene expression by this plasmid were deleted by digestion with NsiI-BamHI and replaced with the nusAinfB operon, which begins downstream of the metYp2 promoter, including the last 5 nucleotides of the Fis binding site and ending with the stop codon for infB. As a vector control for the pSPCnusAinfB plasmids, pBAD43 was used.
Construction of pBAD24 plasmids [55] that express IF2-1, IF2-2, and S-tagged IF2-2 (S-IF2-2) has been described previously [5]. The plasmid for expressing S-IF2-1 was similarly constructed by amplifying the infB gene using primers Stag-IF2-1 and IF2BamHI, which introduce the S-tag coding sequence. The coding sequence was ligated into the NdeI-BamHI site of a pBAD24 vector whose NcoI site has been modified to an NdeI site. The priB and priC genes were cloned into pBAD24, amplifying these genes using the NdeI-priB/PstI-priB and NdeI-priC/PstI-priC oligonucleotides and ligating into the NdeI/PstI site of the pBAD24 vector.
Site-specific mutagenesis was carried out using the QuikChange Lightning Multi-Site-Directed Mutagenesis Kit purchased from Stratagene, using primers listed for this purpose in Table S2. The infB(del1) deletion was generated by amplifying the nusAinfB operon harbored on a plasmid vector with 5′-phosphorylated primers delIF2-1UP and delIF2-LOW (see Table S2), with PfuUltra High Fidelity and circularizing the linear PCR product with T4 DNA ligase. All mutations were verified by sequencing.
ChIP analysis was conducted by modification of previously published procedures [56], [57]. The major change was the incubation of cell lysate with 50 µg/ml RNase A at 37°C for 30 min just before the immunoprecipitation step. Additional details are described in Protocol S1.
Sensitivity of strains to MMS was measured both by direct plating on LB plates containing 6 mM MMS and by 15 min exposure to 0–18 mM MMS, the latter based on the procedure by Nowosielska et al. [34]. β-galactosidase activity was measured according to the procedure of Miller [58]. Mu was plated on LB plates at 37°C with 10 mM magnesium sulfate on a background of indicator cultures. Mu infective centers from thermoinducible lysogens were plated on a background GTN932 indicator at 42°C. Mucts62 lysogens were grown at 30°C. Cultures of priA2::kan strains were maintained in Davis minimal medium (Difco) containing glucose, thiamine, proline, and histidine, and the viable count was determined on plates containing the same media.
All results from measuring MMS and UV sensitivity, homologous recombination proficiency, enzyme assays, and Mu plating efficiency are indicated with error expressed as the standard deviation from the mean (at least three independent experiments; the number of independent experiments is indicated). See Protocol S1 for additional details.
|
10.1371/journal.pgen.1004378 | Genome-Wide Nucleosome Positioning Is Orchestrated by Genomic Regions Associated with DNase I Hypersensitivity in Rice | Nucleosome positioning dictates the DNA accessibility for regulatory proteins, and thus is critical for gene expression and regulation. It has been well documented that only a subset of nucleosomes are reproducibly positioned in eukaryotic genomes. The most prominent example of phased nucleosomes is the context of genes, where phased nucleosomes flank the transcriptional starts sites (TSSs). It is unclear, however, what factors determine nucleosome positioning in regions that are not close to genes. We mapped both nucleosome positioning and DNase I hypersensitive site (DHS) datasets across the rice genome. We discovered that DHSs located in a variety of contexts, both genic and intergenic, were flanked by strongly phased nucleosome arrays. Phased nucleosomes were also found to flank DHSs in the human genome. Our results suggest the barrier model may represent a general feature of nucleosome organization in eukaryote genomes. Specifically, regions bound with regulatory proteins, including intergenic regions, can serve as barriers that organize phased nucleosome arrays on both sides. Our results also suggest that rice DHSs often span a single, phased nucleosome, similar to the H2A.Z-containing nucleosomes observed in DHSs in the human genome.
| The fundamental unit of chromatin is the nucleosome, which consists of 147 bp of DNA wrapped around a histone octamer containing four core histones (H3, H4, H2A, and H2B). Nucleosome positioning in the genome affects the DNA accessibility for regulatory proteins, and thus is critical for gene expression and regulation. Genomic regions associated with regulatory proteins are associated with a pronounced sensitivity to DNase I digestion, and are thus called DNase I hypersensitive sites (DHSs). It is well known that only a subset of nucleosomes are reproducibly positioned in eukaryotic genomes. However, it is less clear what factors determine genome-wide nucleosome positioning, especially in intergenic regions. We mapped both nucleosome positioning and DHS datasets across the rice genome. We discovered that DHSs located in a variety of contexts, both genic and intergenic, were flanked by strongly phased nucleosome arrays. We confirmed the same association of DHSs with phased nucleosomes in the human genome. We conclude that genomic loci associated with a diverse set of regulatory proteins are major determinants of nucleosome phasing, and this is true in both genic and intergenic regions.
| The fundamental unit of chromatin is the nucleosome, which consists of 147 bp of DNA wrapped around a histone octamer containing four core histones (H3, H4, H2A, and H2B) [1]. Since the DNA has to bend sharply around the surface of the histone octamer, flexible or intrinsically curved sequences are favorable for nucleosome formation [2]. In contrast, poly(dA:dT) stretches, which are intrinsically stiff, have been shown to be unfavorable for nucleosome formation and are more enriched in linker sequences [3]–[5]. The intrinsic properties of poly(dA:dT) are also important for nucleosome depeltion, promoter accessibility and transcriptional activity [6]. In vitro nucleosome assembly studies in yeast (Saccharomyces cerevisiae) and Caenorhabditis elegans have confirmed the DNA sequence preferences in nucleosome formation [7], [8]. However, nucleosome organization in vivo is determined by several factors that can override the sequence preferences, including gene transcription, action of nucleosome remodeling complexes, and presence of histone variants and histone modifications [2], [6]. In fact, a sequence preference-based model could only explain ∼50% of the in vivo nucleosome positions in S. cerevisiae [9]. Similarly, only 20% of the human genome is occupied by preferentially positioned nucleosomes [5]. It is important to take such numbers with caution, however, as the calculations are affected by the sequencing methodology and the cell/tissue types used in analysis [10].
Relationships between nucleosome organization and gene expression have been well demonstrated in several model eukaryotes. Phased nucleosome arrays have been observed on both sides of the promoters of active genes [5], [8], [11]–[15]. The promoter itself was traditionally considered to be nucleosome free or depleted, producing what is often called a “nucleosome-free region” (NFR). The first nucleosome downstream and upstream of the promoter are named +1 and −1 nucleosomes, respectively. Nucleosomes after the +1 or before the −1 nucleosome become progressively less phased. Nucleosome positioning in the human genome appears to correlate with the levels of Pol II in the promoter region: better phasing is observed with higher levels of Pol II and less phasing with lower levels of Pol II [13]. So far, the majority of the nucleosome organization studies have been focused on genomic regions associated with transcription. It is unclear, however, what factors determine nucleosome positioning in intergenic regions.
Rice (Oryza sativa) has been used as model species for plant genome research. The rice genome is relatively small (∼400 Mb) and is one of the best sequenced genomes in higher eukaryotes [16]. Various genome-wide genomic and epigenomic datasets have been developed in rice [17]–[22]. Thus, rice provides an excellent model system for nucleosome positioning studies. We generated genome-wide nucleosome positioning data in rice. We mapped both nucleosome positioning and DNase I hypersensitive site (DHS) datasets in the rice genome. We discovered that DHSs associated with different genomic regions, including promoters, genes, and intergenic regions, were all flanked by strongly phased nucleosome arrays. Our results support the barrier model for nucleosome organization. The DHSs, which are likely bound to regulatory proteins, can serve as the barriers to organize phased nucleosome arrays on both sides. Thus, genome-wide nucleosome positioning appears to be orchestrated by genomic regions associated with regulatory proteins.
DHSs are markers of regulatory DNA and span all classes of cis-regulatory elements, including promoters, enhancers, insulators, silencers and locus control regions [23]. We applied a strategy of mapping both nucleosome positioning and DHS datasets to examine whether nucleosome positioning is associated with all cis-regulatory elements across the rice genome. All datasets used in the analysis were developed using rice leaf tissue at the same developmental stage (see Materials and Methods). Rice chromatin was digested by micrococcal nuclease (MNase) into mono-nucleosome size. Mono-nucleosomal DNA was isolated and sequenced (MNase-seq) using Illumina sequencing platforms. We obtained a total of 38 million (M) single-end reads from our first MNase-seq experiment and mapped ∼26 M to unique positions in the rice genome. We also conducted pair-end sequencing of an independent MNase-seq library, obtained 274 M paired-end reads, and mapped ∼231 M read pairs to unique positions in the rice genome.
We previously identified a total of 97,975 DHSs (leaf tissue) in the rice genome [24]. We grouped these DHSs into five categories based on their locations in the genome: 13,272 in proximal promoters (within 200 bp upstream of a TSS), 13,607 in distal promoters (200–1000 bp upstream of a TSS), 25,922 within genes, 4,249 in downstream regions of genes (within 200 bp downstream of the end of transcription), and the remaining 41,602 in intergenic regions. We then aligned both DNase-seq and MNase-seq reads to the rice genome. Strikingly, we observed peaks of read alignments oscillating from both sides of DHSs, indicating the presence of regularly spaced, phased nucleosomes. This phenomenon was evident both in forward and reverse oriented reads (represented by positions of their 5' ends) and in both single-end reads (Figure 1) and paired-end reads (Figure S1). The highest amplitudes of the oscillations were immediately adjacent to boundaries of the DHSs, suggesting that the nucleosomes close to the DHSs were more phased than those far from the DHSs. Phased nucleosomes were not observed in regions flanking randomly selected genomic regions (Figure 1F). The pattern of phased nucleosome arrays surrounding the DHSs is highly similar to the phased nucleosomes surrounding the promoters of active genes reported in model animal species [5], [11], [13].
We also examined nucleosome phasing surrounding TSSs in the rice genome independently of DHSs. Clearly-phased nucleosomes were detected downstream of TSSs of expressed genes (Figure 2A), but not downstream of TSSs of non-expressed genes (Figure 2B), similar to the patterns observed in human and yeast genomes [5], [11], [13]. However, phased nucleosomes were not detected upstream of TSSs of expressed genes (Figure 2A), although phased nucleosomes were detected on both sides of the promoter DHSs (Figures 1A, 1B). In contrast, phased nucleosomes were observed on both sides of TSSs in human and yeast genomes [5], [11], [13].
We noticed that the average lengths of most DHSs in different genomic regions, except for those located in proximal promoters, were similar in the rice genome, with ∼50% DHSs in the size of 35–150 bp. In contrast, the lengths of DHSs in proximal promoters were more variable, including ∼79% DHSs >150 bp (Figure 2C). We suspected that the variable lengths of the DHSs in proximal promoters may mask the detection of nucleosome phasing in front of TSSs. We sorted the DHSs in proximal promoters based on lengths and examined the nucleosome positioning of all active genes associated with these DHSs. Phased nucleosomes were observed on both upstream and downstream of the TSSs of these genes (Figure 2D), which confirmed our prediction.
We wanted to examine if phased nucleosomes are associated with the binding sites of specific rice transcription factors. IDEAL PLANT ARCHITECTURE1 (IPA1), a member of the SPL transcription factor family, is a key regulator in determining plant architecture and enhancing grain yield in rice [25]. A genome-wide IPA1-binding site map has recently been developed using ChIP-seq method and shoot apices tissue from 4-week-old rice seedling [26]. We found that 87.8% of the IPA1-binding sites (5,298 of 6,032) are associated with DHSs, despite of the fact that the DHS data was developed from 2-week-old seedling tissue [24]. An IPA1-binding site was considered to be flanked by phased nucleosome if the ±50 bp regions of the site overlap with a phased nucleosome. Under this criteria, 33.2% (1,757 of 5,298) of the IPA1-binding sites were flanked by phased nucleosomes (see an example in Figure 3), which is significantly higher than the frequency observed from 5,298 randomly selected regions (24.3%, binomial test, p<0.001). In addition, 5,197 and 2,898 of the IPA1-binding sites contain the IPA1-binding motif, GTAC, and another over-represented motif, TGGGC[C/T], respectively [26]. We found that 33.1% of the GTAC-containing sites and 36.2% of the TGGGC[C/T]-containing sites were flanked by phased nucleosomes under the same criteria.
Mapping of both DNase-seq and MNase-seq datasets revealed peaked MNase-seq reads from both forward and reverse strands on both sides of DHSs (Figures 1A–1D). These results suggest that the DHS regions, although highly sensitive to DNase I cleavage, may span a structure that is more inhibitory to MNase digestion than the DHS-flanking regions. The most likely candidate for this predicted structure is a phased nucleosome within each DHS. This predicted nucleosome partially overlapped with the TSSs in proximal promoters (Figure 1A). We named this predicted nucleosome as “-1 nucleosome” because of its location in front of the TSS. The mapping results and our prediction are in agreement with a recent report that active promoters and other regulatory regions in the human genome are not nucleosome free, but are enriched with special nucleosomes containing both of the widely conserved histone variants H3.3 and H2A.Z [27]. These regions were previously considered as “nucleosome free” because nucleosomes carrying both H3.3 and H2A.Z are unusually unstable under the conditions that were commonly used for nucleosome preparation [27], [28]. This instability is believed to facilitate the access of transcription factors and regulatory proteins [27]. Nucleosome formation in promoters was detected during the activation of the zygotic genome of zebrafish [29].
The DHSs in intergenic regions were associated with a unique nucleosomal positioning pattern. The intergenic DHSs lacked the forward MNase-seq peak and the reverse MNase-seq peak, respectively, on the two sides of the DHSs (Figure 1E), suggesting that either these DHSs lack nucleosomes or the nucleosomes are poorly phased. Thus, intergenic DHSs are likely more dynamic with nucleosome occupation, which could mask the identification of a positioned nucleosome. Intergenic DHSs are highly enriched with enhancers in mammalian species [23], [30]. Thus, many of these regions may be associated with regulatory proteins in a cell type-specific manner, which would also mask the identification of positioned nucleosomes in datasets generated from tissues with mixed cell types, such as leaf. We previously demonstrated that rice DHSs generally lack histone modification marks associated with histone H3. However, intergenic DHSs were uniquely enriched with H3K27me3, suggesting a dynamic nucleosome occupation in these regions [24].
Since the DHSs in proximal promoters were more variable in lengths (Figure 2C), we further investigated the positions of the -1 nucleosomes relative to the DHSs with different lengths. We divided the DHSs into five different groups based on their lengths (320–480 bp, 200–320 bp, 140–200 bp, 80–140 bp, and 20–80 bp, respectively). DHSs within the same group were aligned by their 5' ends. All DHSs with a length >140 bp showed a similar nucleosomal positioning pattern (Figures 4A, 4B, 4C). These DHSs appeared to span a single, phased nucleosome, although the DNA length of the DHSs in 320–480 bp is close to two nucleosomes, which may reflect nucleosomes with longer linkers, or nucleosomes tightly associated with other regulatory proteins. These results indicate that the -1 nucleosome in these promoters can accommodate variable amounts of DNA, perhaps reflecting the existence of diverse proteins that interact tightly with the -1 nucleosome or with promoter DNA.
The sizes of 2,495 DHSs (out of 11,718) in proximal promoters were <140 bp, which is shorter than the sequences required to wrap a single nucleosome. These DHSs did not appear to span a nucleosome, but appeared to be enriched in the 3′ portion of the -1 nucleosome (Figure 4D) or were located between the -1 and +1 nucleosome (Figure 4E). Thus, the small DHSs tend to be located in the linker regions. The levels of DNase I sensitivity within these small DHSs were clearly lower than those of the DHSs >140 bp (Figure 4).
We observed a superposition between the forward and reverse MNase-seq reads in genic and promoter regions, which indicates very little or no space between 5' ends of forward and reverse oriented reads (Figures 1A–1C). However, a clear shift between the forward and reverse reads was observed in intergenic regions (Figure 1E). We wondered if this shift was caused by longer linkers that connect the phased intergenic nucleosomes (Figure S2). We investigated the lengths of linkers between phased nucleosomes associated with different genomic regions. We used paired MNase-seq reads and employed 1-bp resolution to calculate the distribution of forward and reverse MNase-seq reads rather than using the 20-bp windows that we used for the other analyses. We measured the distance between maxima of adjacent peaks from reverse to forward strand, respectively, to estimate the length of the linkers between two adjacent nucleosomes. Assuming a constant nucleosome core DNA length of 147 bp, the average length of linkers between two phased nucleosomes in intergenic regions was 35.3 bp, which was significantly longer than the average lengths of linkers between two adjacent nucleosomes within genes (8.1 bp) and in proximal promoters (8.5 bp) (Figure 5A, p<0.005, Kolmogorov–Smirnov test). We also calculated linker lengths in the human genome using human MNase-seq data [13], and found a similar pattern as in rice: the linker length in intergenic regions in the human genome was 38.7 bp, compared to only ∼11.5 bp and 10.1 bp, respectively, for the linkers in proximal promoters and genic regions (Figure 5A).
A weakness of the above method of calculating linker length is that it is influenced by the severity of MNase digestion as MNase can either digest into the nucleosome core DNA or fail to completely digest the linker DNA. Thus, we used an alternative method to estimate the linker lengths in different genomic regions in rice. Since the position of the nucleosome center (dyad), which can be identified as the middle position of each paired-end read, is not affected by different levels of MNase digestion, we can calculate the spacing of between two adjacent nucleosomes using the midway point between paired MNase-seq reads rather than 5' ends. We found that the average spacing between two nucleosomes adjacent to intergenic DHSs was ∼191 bp (Figure 5B), which is significantly longer than the spacing between nucleosomes adjacent to DHSs in proximal promoters (175 bp) and genes (176 bp). The average spacing of nucleosomes associated with various histone modification marks was recently reported in human CD4+ T cells [5]. The average spacing of nucleosomes associated with H3K4me1 and H3K27ac, both euchromatin marks, are 178 bp and 179 bp, respectively. In contrast, the average spacing of nucleosomes associated with H3K9me3 and H3K27me3, both heterochromatin marks, are 205 bp [5]. Thus, linkers of nucleosomes in heterochromatin are significantly longer than the linkers of nucleosomes in euchromatin. These results are in agreement with the linker length difference in genic and intergenic regions observed in both rice and human genomes (Figure 5).
We exploited the genomic datasets from the human genome to examine a similar association of DHSs with nucleosome positioning. Human CD4+ T cell line has been extensively used in epigenomics profiling, including histone modifications [31], nucleosome positioning [13], and DHS mapping [32]. We found that the relationship between DHSs and nucleosome positioning using datasets from the CD4+ T cell line was highly similar to the patterns observed in rice. The DHSs in proximal promoters (Figure 6A), genes (Figure 6B), and intergenic regions (Figure 6C) were flanked by phased nucleosomes. Interestingly, a similar shift between the forward and reverse MNase-seq reads was also observed in intergenic regions (Figure 6C).
Since H2A.Z-associated nucleosomes were found in regions that were previously thought to be nucleosome free, we investigated if DHSs in the human genome span H2A.Z-associated nucleosomes. Mapping of H2A.Z ChIP-seq dataset [31] together with DHS data [32] revealed a phased H2A.Z-associated nucleosome within DHSs in proximal promoters and genic regions in the human genome (Figures 6A, 6B). The intergenic DHSs tended to locate between two phased H2A.Z nucleosomes (Figure 6C). These results suggest that human DHSs span a phased H2A.Z nucleosome, which is also supported by previous data that a single H2A.Z nucleosome can be mapped within CTCF-binding sites in low-salt condition in the human genome [27]. The positions of the H2A.Z nucleosomes within human DHSs are highly similar to the implicated nucleosome within rice DHSs. Thus, we predict that the implicated nucleosome associated with rice DHSs likely contains H2A.Z, which serve as ‘place holders’ to facilitate binding of tanscription factors. The instability and dynamic replacement by regulatory proteins of these nucleosomes result in the DHSs in these genomic regions.
Genome-wide nucleosome positioning maps have been generated in several eukaryotes, including yeast [9], [11], [33]–[35], Drosophila melanogaster [12], C. elegans [36], humans [5], [10], [13], and Arabidopsis thaliana [37]. It has been well documented that only a subset of nucleosomes are phased in any genome. Most consistently, active genes form highly phased nucleosomes flanking the TSSs, which led to the suggestion that transcription may promote nucleosome organization [8], [38]. Proper function of the adenosine triphosphate (ATP)-dependent chromatin remodeling enzymes was recently found to be key for nucleosome positioning in yeast [39]–[41] and mammalian species [42]. It also suggests that transcription or the transcription initiation complexes do not play a direct role in nucleosome phasing surrounding TSSs [40], which is also supported by the fact that genes with poised Pol II in the human genome exhibited a similar pattern of nucleosome phasing to the expressed genes [13].
A barrier model was proposed to explain genome-wide nucleosome positioning [3], [43]. Nucleosomes can be organized passively at regular intervals surrounding a barrier. The barrier model can be used to explain the phased nucleosome arrays surrounding TSSs in that each TSS indirectly dictates a phased position for the next adjacent nucleosome. Whatever factors that determine spacing of nucleosomes in that context would then force the subsequent nucleosome to also be phased, and so on until an array of phased nucleosomes is formed. A barrier can only enforce its effect within a limited distance, resulting in the decay of nucleosome phasing away from the barrier. The effect of the barriers appear to be bidirectional since phased nucleosome arrays are formed on both sides of the TSSs.
Gaffney et al. (2012) recently mapped nucleosomes surrounding the binding sites of 35 different transcription factors in human lymphoblastoid cell lines. Strongly positioned nucleosome arrays were found to flank the binding sites, including those at least 1 kb away from a known TSS [10]. Phased nucleosome arrays were observed around the binding sites of other regulatory proteins, such as the mammalian insulator protein CTCF [5], [44] and repressor protein NRSF/REST [5]. Hughes et al. (2012) recently studied nucleosome positioning of S. cerevisiae strains containing large genomic regions from other yeast species [15]. Nucleosome-depleted regions (NDRs) fortuitously arose in coding regions of the foreign genomic sequences. Interestingly, these NDRs are associated with binding of TFIIB, an essential component of the RNA polymerase II core transcriptional machinery, and were flanked by phased nucleosomes [15]. These results are all in favor of the barrier model because the binding of a regulatory protein to both promoters and non-promoter regions can create a barrier for nucleosome organization. The regulatory proteins reported to be involved in nucleosome positioning include nucleosome remodelers and transcription factors, including activators, components of the preinitiation complex and elongating Pol II [6].
We demonstrate that DHSs in the rice genome are flanked by phased nucleosome arrays on both sides (Figure 1), which is highly similar to the nucleosome arrays flanking TSSs. Phased nucleosome arrays were associated with DHSs located in different genomic regions, including those inside of genes and intergenic regions. A similar association of DHSs with phased nucleosomes was also observed in the human genome (Figure 6). It has been well documented in different eukaryotes that DHSs represent regions associated with various regulatory proteins. For example, the binding patterns of 21 developmental regulators in Drosophila were quantitatively correlated with DNA accessibility in chromatin that can be measured by the DNase I sensitivity [45]. More strikingly, 94.4% of a combined 1,108,081 binding sites from all human ENCODE transcription factors fall within DHSs [23]. Similarly, we previously found that ∼90% of the binding sites of two of the best characterized transcription factors in A. thaliana, APETALA1 and SEPALLATA3, were covered by DHSs [46]. Thus, the association of DHSs with phased nucleosome arrays shows that the barrier model can be extended to an entire genome: any genomic region associated with regulatory proteins can serve as a barrier for nucleosome organization, and these regions can be either directly associated with transcription, such as promoters, or indirectly associated with transcription, such as the insulators. This model would also predict different nucleosome positioning profiles in different organs/tissues and in different developmental stages due to differential binding of regulatory proteins.
A DHS-based barrier can be permanent, such as the promoters associated with constitutively expressed genes, or be temporarily, such as binding sites of transcription factors associated with tissue- or organ-specific gene expression. Regulatory proteins can bind DNA tightly or loosely (or dynamically, with transient nucleosome formation in the same region), which may result in “hard” barriers or “soft” barriers. Hard barriers will result in well positioned and well phased nucleosome arrays; whereas soft barriers may result in “fuzzy” and less phased nucleosome arrays. In Drosophila, the binding sites of transcription factors that are flanked with strongly positioned nucleosome arrays were more sensitive to DNase I digestion and have more pronounced DNase I footprints [10]. These results support that the levels of transcription factor occupancy at the binding site determine the levels of positioning of the flanking nucleosome arrays, thus, the level of “hardness” of the barrier.
In summary, we demonstrate that DHSs located across the rice genome are flanked by strongly phased nucleosome arrays. We confirmed the same phenomenon in the human genome by analyzing publically available datasets. Our results support the barrier model for nucleosome organization as a general feature of eukaryote genomes. We propose that genome-wide nucleosome positioning in the eukaryotic genomes is orchestrated by genomic regions associated with regulatory proteins.
Rice cultivar “Nipponbare” seeds were germinated at room temperature for three days. Germinated seeds were then sowed in soil to continue to grow in the greenhouse. The seedlings continued to grow for two weeks under 12 hrs day/night cycles and 32°C/27°C corresponding to day and night, respectively. The seedlings were then harvested for nuclei isolation, the same growing stage/condition used for developing DNase-seq and RNA-seq datasets previously [24]. The nuclei were then digested with a series of concentrations of micrococcal nuclease (MNase). The MNase-digested DNA was separated using 2% agarose gel containing ethidium bromide and visualized under UV light. Nuclei were digested into ∼80% nucleosome monomers and ∼20% dimers. The mono-nucleosomal DNA was then excised from the gel and purified using a gel purification kit (Qiagen, 28006). The purified DNA was used for MNase-seq library development, including end blunting, adding “A” base to the blunt DNA fragments, ligating “A” tailed DNA fragments with either single-end adapter or pair-end adapter, and enriching ligated DNA fragments by PCR. The final, amplified DNA was purified and sequenced with 36 bp SR (single reads) or PE (paired end) using Illumina sequencing platforms.
We mapped the MNase-seq reads to the rice genome (TIGR release 5) using MAQ software [47] with default parameters (except 1-bp mismatch allowed). Only the reads aligning to a unique position in the rice genome were used for further analysis. DNase-seq and RNA-seq dataset were generated from our previous work [24]. Methods for mapping DNase-seq and RNA-seq reads were described previously [24]. We used the same methods to analyze datasets from human CD4+ T cell line, including DNase-seq dataset [32], MNase-seq dataset [13], and H2A.Z ChIP-seq [31]. All sequence reads from human CD4+ T cell line were aligned to human genome build 37 of NCBI using MAQ software using default parameters (except 1-bp mismatch allowed). We used F-seq [48] with 200-bp bandwidth parameter to identify rice DHSs. To control the FDR of the identified DHSs, we generated 10 random datasets each containing the same number of sequence reads as our DNase-seq dataset. The FDR was calculated as ratio of DHSs identified from random datasets to DHSs identified from the DNase-seq dataset. We controlled the FDR<0.05. We used the same method and parameters as Boyle et al. [32] to identify the DHSs in human CD4+ T cell line. We employed nucleR [49] to predict phased nucleosomes based on pair-end MNase-seq data using nonparametric method. We removed all fragments >200 bp (distance between the paired reads) and trimmed the fragments to the middle 40 bp to remark the position of dyad. The dyad positions were transformed by Fast Fourier Transform to show distribution of nucleosomes in Figure 3 and to identify the phased nucleosomes. The programs for data processing and statistical test were written in Perl or R (http://www.r-project.org/).
MNase-seq data has been deposited to NCBI under accession number GSE53027.
|
10.1371/journal.pntd.0006545 | Toxoplasmosis seroprevalence in rheumatoid arthritis patients: A systematic review and meta-analysis | Toxoplasmosis is a cosmopolitan infection caused by an intracellular obligatory protozoan, Toxoplasma gondii. Infection to this parasite in immunocompetent patients is usually asymptomatic, but today it is believed that the infection can be a risk factor for a variety of diseases, including rheumatoid arthritis (RA). RA is an autoimmune disease and the most common type of inflammatory arthritis that is a major cause of disability. The aim of this systematic review and meta-analysis was to address the association between RA and toxoplasmosis in light of the available research.
Based on the keywords, a systematic search of eight databases was conducted to retrieve the relevant English-language articles. Then, the studies were screened based on the inclusion and exclusion criteria. The random effect model was used to calculate the odds ratio (OR) using forest plot with 95% confidence interval (CI).
Overall, 4168 Individual, extracted from 9 articles were included for systematic review evaluation, with 1369 RA patients (46% positive toxoplasmosis) and 2799 individuals as controls (21% positive toxoplasmosis). Then, eight articles (10 datasets) were used for meta-analysis (1244 rheumatoid arthritis patients and 2799 controls). By random effect model, the combined OR was 3.30 (95% CI: 2.05 to 5.30) with P < 0.0001.
Although toxoplasmosis could be considered as a potential risk factor for rheumatoid arthritis, more and better quality studies are needed to determine the effect of T. gondii infection on induction or exacerbation of RA. Our study was registered at the International Prospective Register of Systematic Reviews (PROSPERO; code: CRD42017069384).
| Toxoplasma gondii is an intracellular obligatory protozoan, which causes toxoplasmosis. T. gondii infection in immunocompetent individuals is mostly asymptomatic, but it may be reactivated as a result of immune disorders inducing serious complications. Rheumatoid arthritis (RA), as a complex autoimmune disease, is a major cause of significant and progressive disability, articular complications, and premature death. Studies confirmed an interaction between infections and environmental factors as the potential risk or protective factors determining the development of autoimmune diseases. In this study, we investigated the association between toxoplasmosis and RA.
| Toxoplasmosis is a parasitic disease with worldwide distribution caused by obligate intracellular coccidian protozoan Toxoplasma gondii (T. gondii) [1]. It is estimated that one-third of the world’s population are infected with this parasite in both developed and developing countries [2, 3]. Humans can be infected with the parasite through different routes, including consumption of raw or undercooked meat containing tissue cysts of the parasite, ingestion of sporulated oocysts from contaminated water and food, and vertical transmission during pregnancy through the placenta to the fetus [4].
T. gondii remains in the infected host tissues perpetually [5]. Most immunocompetent individuals, if infected with this parasite, are asymptomatic or show minor symptoms [6]. The most common symptom of toxoplasmosis in humans is lymphadenopathy that may be associated with fever, sore throat, muscle pain, fatigue, and headache [4]. In congenitally infected and immunocompromised patients, this disease is more likely to bring about severe complications [7]. Myocarditis and polymyositis have been reported in immunocompetent individuals with acute toxoplasmosis [8]. Furthermore, toxoplasmosis may cause polyarthritis in the hand and knee joints [9]. Polytenosynovitis (inflammation of a tendon sheath) caused by T. gondii has also been reported [10].
Rheumatoid arthritis (RA) is a common autoimmune disease, which is a major cause of inflammation of the joints and the principal cause of disability that affects 0.5–1% of the population [11, 12]. The disease presents with swollen joints, production of autoantibodies (rheumatoid factor), and systemic effects [13].
In recent years, the role of infectious agents, especially bacteria and viruses, has been identified in the pathogenesis of autoimmune diseases, while the role of parasitic infections due to their vague effects on host immunity has not been well-investigated. Experimental evidence may support the protective effect of specific parasitic infections in the susceptibility to autoimmunity [14]. Some geoepidemiological studies showed that host genetic susceptibility interacts with lifestyle and environmental factors, such as socioeconomic status, dietary habits, environmental pollutants, and ultraviolet radiation exposure; further, infections increase the risk of developing autoimmunity [15]. On the other hand, infectious diseases may contribute to the development of autoimmune diseases through molecular mimicry and epitope spreading [16]. Therefore, the aim of this systematic review and meta-analysis was to provide an updated review of data about the relationship between toxoplasmosis and RA.
This study was performed according to the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) and its checklist [17]. Individuals with RA, along with a control group, were surveyed. To begin, we searched scientific databases for all the articles on the association between toxoplasmosis and RA published up to the first of January 2018. These keywords were used alone or in combination: “Toxoplasma gondii”, “toxoplasmosis”, “seroprevalence”, “prevalence”, “rheumatoid arthritis”, “rheumatoid factor”, “meta-analysis”, and “systematic review”.
A literature review was carried out using English databases including “PubMed”, “Google Scholar”, “Science Direct”, “Scopus”, “Web of Science”, “EMBASE”, “CINAHL”, and “ProQuest”. The systematic search of articles was conducted from March 4 to December 31, 2017 by two researchers independently. Also, for completing the checklist, we investigated all the references lists of the selected articles manually. In this study, only English-language articles were analyzed; furthermore, unpublished studies were not evaluated.
After completing the search, the selected articles were reviewed by the two researchers independently. All the duplicate and irrelevant studies were excluded after reviewing the title, abstract, and full text of the articles. Moreover, to prevent reprint bias, the results of the articles were carefully investigated and duplicates were omitted.
In order to assess the quality of reporting of the studies, standard Strengthening the Reporting of Observational Studies in Epidemiology checklist (STROBE) was used [18]. S1 Checklist represents the quality score of different eligible studies. This checklist included items assessing the study methodology, study type, study population, sample size, sample collection methods, statistical tests, and presentations. In our study, articles were evaluated based on STROBE assessment (low quality: less than 16.5, moderate quality: 16.6–25.5, and high quality: 25.6–34). The articles we entered in our meta-analysis had acceptable quality.
Abstracts and full texts were assessed independently by the two researchers using a piloted form. The final decisions about the eligibility or exclusion of studies were made separately. Disagreements were resolved with provision for arbitration from a third reviewer. Following the removal of duplicate entries, articles were evaluated according to the following criteria: (1) cohort or case-control studies about the relationship between toxoplasmosis as an exposure and rheumatoid arthritis as a disease, (2) the studies conducted only on humans, (3) the presence of case and control groups, (4) the studies where toxoplasmosis was diagnosed by detecting IgG and/or IgM antibodies against T. gondii in individuals with definitive diagnosis of RA, and (5) the studies providing details on the seroprevalence rate of toxoplasmosis and RA.
The exclusion criteria comprised: (1) studies that were only descriptive, (2) studies that only presented the final result and did not provide the raw data, (3) articles that were not available in English language, and (4) the studies conducted on animals.
Articles were carefully studied and the following data were extracted: first author, year of publication, the number of patients and controls, the number and percentage of the positive and negative cases of serum IgG and IgM in patients and controls, as well as information about age and gender and laboratory results. In studies where two different populations were studied, data were extracted separately.
The meta-analysis was executed with the Stats Direct statistical software (http://statsdirect.com). For displaying the heterogeneity between studies, χ2-based Cochrane test (Q) and I2 index were applied [19]. Due to significant heterogeneity between the studies, a random effect model was used to combine the results of the studies. Forest plot was used to indicate the prevalence of toxoplasmosis in each study and to determine pooled estimate prevalence in the studies. Odds ratios (ORs) and 95% confidence intervals (CI) were used for estimating the risk of T. gondii infection (the significance of P<0.05). OR > 1 indicates the positive effect of Toxoplasma on RA and an OR < 1 shows that toxoplasmosis has a protective effect against RA. Publication bias was examined by funnel plots and the statistical significance was assessed by the Egger test [20]. Also, it was performed a sensitivity analysis to identify probably effect of each article on the overall results by excluding them using Stata version 14 (Stata Corp, College Station, TX, USA).
The study protocol (CRD42017069384) was registered on the website of the International Prospective Register of Systematic Reviews (PROSPERO) [21].
Our preliminary search of eight databases yielded 8234 papers. After a primary screening of the titles of the articles based on keywords, 124 studies were extracted. Sixty-five articles were also excluded from the study due to duplication. In the next step, by screening the abstracts of the articles and based on the inclusion/exclusion criteria, 43 other articles were excluded. After reading the full text of the articles, 10 other papers were omitted, and three studies were added to the collection after reviewing the references. After the final review of the articles, nine eligible studies [14, 16, 22–28] were identified for systematic review. Another study was excluded due to the absence of a healthy control group [16]. Finally, eight of these nine articles [14, 22–28] were entered into the meta-analysis with respect to the inclusion/exclusion criteria (Fig 1). The studied articles were published between 2007 and 2017. We identified 11 datasets from the nine articles that met the inclusion criteria, eight of which were case-control, two cross-sectional, and one were cohort studies (Table 1). The surveys were conducted in Latin America [14], Europe [14, 16], Egypt [24, 25], Iraq [22, 23, 27], Czech and Slovak [26], and China [28].
Our meta-analysis was performed among 4168 people including 1369 RA patients and 2799 controls. In all the studies, blood samples were collected from patients and controls. To identify anti-Toxoplasma antibodies (IgG and IgM) in those studies, ELISA [22–24, 26–28], CFT [26], EIA [25], chemiluminescence [16], and BioPlex 2200 system [14] were used (Table 1).
Except for Fischer et al. [16], who only evaluated IgG, other authors surveyed both IgG and IgM antibodies. However, only three studies had reported a titer of antibodies [23–25], and others had described the percentage of positive antibodies in patients and controls. Finally, all the studies analyzed the relationship between toxoplasmosis and RA with respect to the percentage of seropositive and seronegative individuals (patients and controls).
As shown in Fig 2, the prevalence of toxoplasmosis in RA patients in these studies varied from 25% to 77% with an overall seroprevalence of 46% (95% CI [37; 56]). However, the total prevalence of this disease in the control subjects entered in these studies was 21% (95% CI [14; 28]), which varied from 0% to 48% in various studies (Fig 3).
According to Fig 4, the odds of toxoplasmosis in RA patients are 3.30 times compared to that of controls with 95% CI: 2.05 to 5.30 and P < 0.0001. Nonetheless, the heterogeneity analysis of the effect size of arthritis (Q = 32.77, P = 0.0001, I2 = 72.5%) showed a relatively high heterogeneity in our meta-analysis.
Begg and Egger tests were used to evaluate publication bias. Negligible publication bias was observed using both Begg test (P = 0.0286) and Egger test (P = 0.0446) in the included studies. The results of sensitivity analysis showed that the impact of each study on meta-analysis was not significant on overall estimates (Fig 5).
Toxoplasma gondii is an important opportunistic parasite infecting one-third of the world’s human population and it is considered a silent threat [29]. Though a clear relationship between toxoplasmosis and autoimmune diseases, including RA, has not yet been well documented, a higher prevalence of anti-T. gondii antibodies was reported in patients with rheumatoid arthritis [16]. Thus, we designed this systematic review and meta-analysis to explore the possible association between Toxoplasma infection and RA, an autoimmune disease causing pain and disability [30].
Although few studies were included in our meta-analysis, our findings showed that the prevalence of toxoplasmosis in the control group was 21%, which is almost in agreement with the results obtained by Dubey and Beattie [31]. This seroprevalence was significantly different from the prevalence in RA patients (46%).
According to Table 2, the lowest OR was reported in Latin America and the highest OR in Europe. The difference in ORs can be attributed to the significant difference in host response and virulence of parasitic strains [32]. In addition, we found high heterogeneity in the relationship between RA and T. gondii infection in this systematic review. The high heterogeneity index is suggestive of potential variation, which could be due to difference in genetic potential of humans, which is affected by lifestyle and environmental factors such as dietary habits, environmental pollution, exposure to ultraviolet radiation, various types of infections, and socioeconomic status [33].
Our findings suggest that T. gondii may trigger a pathologic process in individuals, which can ultimately lead to RA. This finding has been reported in other autoimmune diseases such as diabetes mellitus [34, 35], lupus erythematosus [36], and autoimmune thyroid diseases [37]. The higher prevalence of T. gondii in people with chronic diseases can be explained by the following reasons: 1) toxoplasmosis can contribute to the progression of chronic diseases and 2) treatment of these diseases with immunosuppressive drugs increases the susceptibility of patients to infections, including toxoplasmosis [6]. Recent treatments for RA patients with anti-tumor necrosis factor-α (TNF-α), which leads to brain toxoplasmosis, are indicative of this issue [27, 38]. On the other hand, some toll-like receptors (TLRs) have been identified in mammals, for which some pathogens act as ligands, and as a result of binding between the TLRs and pathogens different types of immune responses can be induced. Based on reference, T. gondii may be used as ligands for TLRs, which can induce inflammatory response [39].
Also, studies show that T. gondii increases the expression of interleukin 17 (IL-17) in patients [40], and since this cytokine is involved in the pathogenesis of many autoimmune diseases, including RA [41], a significant relationship between toxoplasmosis and RA can be explained.
RA patients have autoantibodies and rheumatoid factors in their blood [42]. In two studies, these disease activity markers were found to have a significant relationship with toxoplasmosis, especially in high titers [24, 25]. This indicates that T. gondii can induce or exacerbate arthritis symptoms [43–45].
Because in the studied articles the relationship of age and sex with the prevalence of toxoplasmosis was not evaluated, we avoided the meta-analysis of these risk factors. In addition, diversity in the quality of studies and methods of measuring antibodies limited the interpretation and analysis of these items. These two issues were the important limitations of our meta-analysis.
Despite the significant relationship found between T. gondii infection and RA in this systematic review and meta-analysis study, further studies are needed on the following grounds: 1) the limited sample sizes in the articles, 2) difference in the quality of the reports, 3) diverse methods of measuring anti-parasitic antibodies, and 4) lack of evaluation of various risk factors such as age and gender.
One of the most important achievements of our study is that although T. gondii infection affects about one-third of the world’s population and possibly causes and exacerbates the symptoms of RA, only few studies have addressed this subject. These studies were conducted only in Latin America, Europe, and few regions of Asia and Africa. Accordingly, further studies are needed to achieve accurate results from other parts of the world. Also, further studies will be necessary to clarify the pathogenesis of T. gondii in humans to understand whether T. gondii is a cofactor in the development of autoimmune diseases.
|
10.1371/journal.pgen.1008054 | Antero-posterior ectoderm patterning by canonical Wnt signaling during ascidian development | Wnt/β-catenin signaling is an ancient pathway in metazoans and controls various developmental processes, in particular the establishment and patterning of the embryonic primary axis. In vertebrates, a graded Wnt activity from posterior to anterior endows cells with positional information in the central nervous system. Recent studies in hemichordates support a conserved role for Wnt/β-catenin in ectoderm antero-posterior patterning at the base of the deuterostomes. Ascidians are marine invertebrates and the closest relatives of vertebrates. By combining gain- and loss-of-function approaches, we have determined the role of Wnt/β-catenin in patterning the three ectoderm derivatives of the ascidian Ciona intestinalis, central nervous system, peripheral nervous system and epidermis. Activating Wnt/β-catenin signaling from gastrulation led to a dramatic transformation of the ectoderm with a loss of anterior identities and a reciprocal anterior extension of posterior identities, consistent with studies in other metazoans. Surprisingly, inhibiting Wnt signaling did not produce a reciprocal anteriorization of the embryo with a loss of more posterior identities like in vertebrates and hemichordate. Epidermis patterning was overall unchanged. Only the identity of two discrete regions of the central nervous system, the anteriormost and the posteriormost regions, were under the control of Wnt. Finally, the caudal peripheral nervous system, while being initially Wnt dependent, formed normally. Our results show that the Ciona embryonic ectoderm responds to Wnt activation in a manner that is compatible with the proposed function for this pathway at the base of the deuterostomes. However, possibly because of its fast and divergent mode of development that includes extensive use of maternal determinants, the overall antero-posterior patterning of the Ciona ectoderm is Wnt independent, and Wnt/β-catenin signaling controls the formation of some sub-domains. Our results thus indicate that there has likely been a drift in the developmental systems controlling ectoderm patterning in the lineage leading to ascidians.
| The Wnt/β-catenin pathway is a system of cell-cell communication. It has an ancient origin in animals and plays multiple roles during embryogenesis and adult life. In particular, it is involved in determining, in the vertebrate embryo, the identity of the different parts of the body and their relative positions along the antero-posterior axis. We have investigated in an ascidian (or sea squirt) species, a marine invertebrate that is closely related to vertebrates, whether this pathway had a similar role. Like in vertebrates, activating Wnt/β-catenin led to a posteriorization of the embryo with a loss of anterior structures. By contrast, unlike vertebrates, ascidian embryos formed rather normally following Wnt/β-catenin inactivation. Since hemichordates (or acorn worms), earlier divergent invertebrates, use Wnt/β-catenin in a manner comparable to vertebrates, it is in the ascidian lineage that changes have occurred. Consequently, ascidians build an antero-posterior axis, very similarly organized to that of vertebrates, but in a different way.
| Ascidians belong to the tunicates, the sister group of the vertebrates. Together with the cephalochordate (amphioxus) and vertebrate phyla they form the super-phylum of chordates whose specific body plan includes a notochord and a dorsal hollow neural tube during embryonic life. Comparative developmental studies between these three phyla is essential for elaborating evolutionary scenarios explaining the emergence of chordates and their diversification [1–4]. Ascidians are particularly puzzling organisms since they took a significantly different evolutionary path from other chordates resulting in divergent morphological, embryological and genomic features. Their development is fast and stereotyped with very few cells; and ascidian genomes have undergone compaction and extensive rearrangements when compared to vertebrates and amphioxus. This raises the question of whether developmental mechanisms controlling typical chordate structure formation are conserved between ascidians and other chordates. In particular, ascidian embryos are the emblematic examples for the concept of mosaic development. However, it is well known that cell-cell communication is involved in cell fate determination, yet possibly at only short distances (i.e. neighboring cells) [5–7].
Wnt signaling, one of the pathways present in animals, allows cells to communicate through the secretion of the Wnt ligands that bind to their cognate Frizzled receptors. It is involved in a wide range of biological processes during embryogenesis and adult homeostasis [8,9]. The canonical Wnt pathway (cWnt) that uses the protein β-catenin as a central mediator to control target gene transcription is extensively involved in axis formation during the development of many metazoans [10,11]. Three discrete developmental processes contribute to antero-posterior (AP) axis formation in bilaterians: germ layer specification, AP patterning and posterior growth. During cleavage/blastula stages, nuclear accumulation of β-catenin in the vegetal hemisphere specifies endomesoderm in several phyla (nemerteans, echinoderms, hemichordates and ascidians) [12–15]. A similar function for the specification of the endoderm at the oral pole of cnidarians suggests that this constitutes an ancient function at the base of metazoans [16,17]. In vertebrates, a posterior to anterior gradient of activity provides cells with positional information and patterns the central nervous system (CNS) [18–21]. A recent study in hemichordates demonstrated that this function for cWnt is conserved at the base of the deuterostomes [22]. Finally, in both insects and vertebrates, Wnt signaling controls body elongation during posterior growth [23].
Posterior growth does not exist in ascidians since the embryo elongation at the improperly named tailbud stages occurs through cell division and rearrangement without proper addition of new tissue from a growth zone [24]. Embryonic axes are determined very early and can be identified in the fertilized egg before first cleavage [5]. cWnt participates in endomesoderm formation along the animal-vegetal axis [13,25]. The AP axis is orthogonal and determined following ooplasmic movements that localize asymmetric cleavage determinants to the posterior. A consequence for AP patterning is that anterior (so called a-line) and posterior (b-line) ectoderm precursors have intrinsically different potentials in response to neural induction as soon as they arise at the 8-cell stage [26]. Interestingly, not only the CNS but also the epidermis is patterned along the AP axis; and this patterning also involves signals from vegetal tissues [27]. The transcription factor FoxA-a is the anterior determinant that establishes the early a-line versus b-line potentials [28,29]. Since direct transcriptional FoxA-a targets are Wnt antagonists–a-line expressed Sfrp1/5 and Ror genes–there is a potential role for Wnt signaling in ectoderm AP patterning. Moreover, the AP identity of the two sensory pigment cells within the CNS is controlled by Wnt signaling [30]. However, a global function for Wnt in ectoderm AP patterning has not been investigated; and this is the topic of the present manuscript.
Sequencing and annotation of the ascidian Ciona robusta (formerly known as C. intestinalis type A) has revealed a complement for Wnt signaling compatible with a functional pathway [31–34]. A recent phylogenetic analysis has shown that the ten Wnt ligands found in the Ciona genome correspond to 10 out of the 13 families present at the base of chordates, with the loss of Wnt1, Wnt4 and Wnt8 [35]. Their spatio-temporal expression has been described throughout embryogenesis for eight of them, but only a few show a restricted pattern (Wnt3, Wnt5, Wnt6 and Wnt7) [35–39]. In particular, they do not display a staggered expression in the posterior of the embryo as observed for many metazoans including vertebrates and amphioxus (reviewed in [10], [35]). The only possible similarity would be the expression of the four above mentioned ligands in caudal muscle at cleavage/gastrula stages and the epidermal expression of Wnt5 at the posterior tip of the forming tail. At the opposite pole of the embryo, the Wnt antagonists, Sfrp1/5 and Ror genes, are expressed in the anterior ectoderm as described above [28,29,36,37]. The C. robusta genome contains five Frizzled receptors [33]. The expression pattern is known for three of them (Frizzled1/2/7 and Frizzled5/8 are maternally and ubiquitously expressed; Frizzled4 is expressed in the ectoderm from the 16-cell stage and later in various discrete regions), but does not allow us to predict where and when Wnt signaling is active [36].
In the present study, we have combined ectopic activation and down-regulation of the cWnt pathway to assess the effects on AP patterning of the C. intestinalis embryonic ectoderm. Activating cWnt from gastrulation leads to a loss of anterior ectoderm that is converted in posterior ectoderm. By contrast, inhibiting cWnt has varying effects depending on the ectoderm derivatives. Epidermis AP patterning is unchanged. The CNS is largely unaffected, except for its anterior and posterior ends, suggesting a function of cWnt in refining a global AP pattern that is defined by other means. Finally, the early definition of the caudal peripheral nervous system (PNS) requires cWnt signaling but redundant mechanisms allow proper differentiation of this tissue. Consequently, while the Ciona ectoderm displays a sensitivity to cWnt activation that is compatible with the expected function for cWnt at the base of deuterostomes, cWnt is only marginally required for ectoderm AP patterning.
LiCl or small molecule inhibitors of Gsk3β have been previously used in ascidian embryos to activate the cWnt pathway [13,25,40]. We have further developed such treatments using two distinct inhibitors, 1-azakenpaullone and BIO [41,42]. These inhibitors were tested at two doses (5 and 10 μM for 1-azakenpaullone; 1 and 2.5 μM for BIO). The results presented here correspond to the highest dose for each molecule, conditions leading to fully penetrant and identical phenotypes for both inhibitors. As expected, early treatments led to ectopic endoderm formation as revealed by staining for endoderm specific endogenous alkaline phosphatase activity (S1 Fig). Treatments starting at the 32-cell stage or later produced embryos with a dramatically abnormal morphology but without ectopic endoderm, allowing us to determine effects on ectoderm patterning without interfering with germ layer formation. A previous report has suggested that activating the cWnt pathway interferes with epidermal sensory neuron (ESN) formation along the AP axis in the tail [40]. We reproduced the reported results: a loss of anterior ventral ESN formation (revealed by the expression of Etr at late tailbud stages) when the treatment was initiated at early neurula stages (stage 14) and an absence of effect when the treatment was initiated at initial tailbud stages (stage 17) (Fig 1D–1G). However, when the treatment was initiated at the onset of gastrulation (stage 10), we observed ectopic ESNs located in the ventral trunk midline (Fig 1B and 1C). Caudal ESNs are known to arise from a neurogenic territory characterized by the expression of Klf1/2/4 [43]. The presence of ectopic ESNs in the trunk upon cWnt activation was accompanied by the ectopic expression of Klf1/2/4 in the ventral trunk midline (Fig 1I–1L), suggesting that these ectopic ESNs arise from an ectopic neurogenic territory. Interestingly, Klf1/2/4 ectopic expression was also observed in treatments starting at stage 14 while Etr expression was repressed. To further investigate the apparent posteriorization of the ectoderm following cWnt activation, we determined the expression of Zf115, a gene with a marked restricted expression in the tail epidermis [44]. Zf115 was ectopically expressed in the entire trunk epidermis for both of our early treatments (Fig 1N–1Q) but not for the latest treatment (S2 Fig). To further delineate the sensitivity of the ectoderm to cWnt activation, we performed 30 min treatments at various developmental stages and assessed the expression of both Etr and Zf115. Ectopic expression of Etr in the ventral trunk was observed when such short treatments were performed during gastrulation (stages 10 to 13), but not later (S2 Fig). The loss of anterior ventral tail Etr expression described above was not observed in the pulse treatments suggesting a longer exposure time may be required. Ectopic Zf115 expression in the trunk was observed for all pulse treatments with a reducing effect as the treatment was delayed: treatment at the onset of gastrulation led to an ectopic expression in the entire trunk while later treatments led to an extension limited to the posterior trunk (S2 Fig).
Above results suggest that cWnt activation converts trunk ectoderm into tail ectoderm with a maximum sensitivity during early gastrulation.
In this section, we will determine what are the effects of activating cWnt from gastrulation on all three ectoderm derivatives: the epidermis, the peripheral and the central nervous system. We have thus examined by in situ hybridization the expression of a panel of AP markers for the ectoderm at early tailbud stages (stage 19/21) following 1-azakenpaullone or BIO treatment from initial gastrula (stage 10). Both drugs led to similar effects (Figs 2 and S3). Interestingly, while we observed a dose response on the morphology of embryos treated with 1-azakenpaullone, the effect on the AP markers examined remained consistent for all doses (S4 Fig). We did not observe a graded effect similar to what was observed when the treatment was staggered over timed intervals (S2 Fig).
To verify the specificity towards Wnt/β-catenin signaling of the above results, we overexpressed Wnt5, a ligand normally restricted to the posterior ectoderm [36], throughout the ectoderm using DNA electroporation. This led to ectopic expression of the tail midline markers Msxb, Klf1/2/4 and Nkx-C in the ventral trunk epidermis (S5 Fig). However, the embryo morphology was severely affected, making gene expression analysis tedious. We turned to overexpression of ΔN-β-catenin, a version of β-catenin that is deleted from the N-terminal domain (containing Gsk3β phosphorylation sites) and that behaves as a dominant active form [40]. We could reproduce the results obtained using Gsk3 inhibitor treatments: ectopic expression of Six1/2, Six3/6, Msxb and Klf1/2/4 (Figs 3B, 3D and 3L and S5), and repression of the epidermal expression of Hox1, Islet and Ror-a (Fig 3F, 3H and 3J). The CNS expression of Hox1 was not affected since we targeted the ectoderm using the promoter of the Fucosyl transferase gene [40]; CNS Hox1 positive cells originate from vegetal lineages and do not express this gene. These observations strengthen our findings that AP patterning defects result from direct action of Wnt/β-catenin signaling.
In previous results, we observed that activating cWnt led to the formation of an ectopic neurogenic territory in the ventral trunk epidermis. It is known that Bmp signaling is required to specify the tail ventral midline and that Bmp signaling is active throughout the ventral epidermis, both in the trunk and the tail [43,46]. We thus expressed Noggin, a secreted Bmp antagonist, together with ΔN-β-catenin. As predicted, ectopic expression of Klf1/2/4 and Msxb was suppressed (Figs 4E and S6). As previously reported, when Bmp signaling was activated, Klf1/2/4 and Msxb were ectopically expressed throughout the tail epidermis (Figs 4C and S6) [43,46]. Activation of cWnt signaling in addition to Bmp led to ectopic activation of both genes in the trunk epidermis as well (Figs 4F and S6), suggesting that combining cWnt and Bmp signals is sufficient to launch the tail ventral neurogenic program (Fig 4G).
We have used the overexpression of two different proteins to block Wnt signaling. TcfΔC is a dominant negative form of the transcription factor Tcf that normally regulates transcription, together with β-catenin, downstream of the binding of a Wnt ligand to a Frizzled receptor. It contains a C-terminal deletion that eliminates the DNA binding domain of Tcf and has been previously used in Ciona to inhibit β-catenin nuclear activity during endomesoderm formation [25,47]. Sfrp1/5 is a naturally secreted antagonist of Wnt signaling that acts by sequestering Wnt ligands and thus preventing them from binding to Frizzled receptors [48]. Both molecules were overexpressed in the entire ectoderm using the promoters of the Friend of gata (Fog) or Fucosyl transferase (Ft) genes [40,49]. The following combinations produced the strongest phenotypes and were used in subsequent experiments: pFog>TcfΔC and pFt>Sfrp1/5.
We first determined the efficiency of our constructs by testing their ability to counteract the effect of Gsk3 inhibitor treatment. Overexpression of TcfΔC was sufficient to suppress the ectopic activation of both Six3/6 and Klf1/2/4 triggered by 1-azakenpaullone treatment (Fig 5A–5D). We could not perform the same assay for Sfrp1/5 since it acts upstream of Gsk3 inhibitors in the cWnt pathway. However, its overexpression had similar effects as TcfΔC overexpression did.
Given the robust phenotypes on epidermal expression following cWnt activation, we expected a strong reciprocal effect for Wnt inhibition: loss of posterior markers and posterior extension of anterior marker expression. We were surprised to see that epidermal expression of Islet, Ror-a, Hox1 and Cdx was unchanged (Fig 5E, 5F, 5I and 5J), with possibly a weak reduction in levels of expression in the most affected embryos as depicted for Cdx on Fig 5J. The only clear difference we could detect was a repression of the epidermal expression of Hox12 using both constructs (Fig 5Kii and 5Kiii), but we did not detect a concomitant posterior extension of Cdx into the tail tip (Fig 5J). Consequently, with the exception of Hox12 and possibly the tail tip, epidermis AP patterning is largely unchanged following Wnt signaling inhibition.
We have determined, at tailbud stages, the expression of CNS genes whose expression was modified following cWnt activation: Six1/2, Six3/6, Hox1 and Hox12. Both Six1/2 and Six3/6 were robustly repressed by pFog>TcfΔC, but only slightly downregulated by pFt>Sfrp1/5 in a minority of embryos (Fig 5G and 5H). This suggests that both genes could be positively regulated by cWnt signaling although in a ligand independent manner. Anterior tail nerve cord Hox1 expression was unchanged (Fig 5I). Expression of Hox12 in the posterior of the tail nerve cord was downregulated by pFt>Sfrp1/5 but unaffected by pFog>TcfΔC (Fig 5K). This difference possibly stems from the embryonic origin of the Hox12 expressing cells that may be of vegetal origin (A-line). Consequently, they do not express the promoters used and as such, do not express the transgenes. Since Sfrp1/5 is a secreted molecule it can prevent these cells from receiving Wnt signals.
In summary, CNS patterning is regulated by Wnt signaling at only the anteriormost and the posteriormost regions of the axis.
The above results prompted us to test the effects of Wnt signaling modulations on the early formation of the CNS at the neural plate stage (stages 13/14) (Fig 6). Etr, whose expression is found in the CNS precursors (rows I to IV according to [50]) and in the palp forming region at the medial anterior neural plate border (rows V and VI; Fig 6Ai and 6D), displayed a loss of expression in this latter territory following cWnt activation (Fig 6Aii). This corroborates previous results obtained at later stages with the markers Ror-a, Otx and Islet. However, the loss of this marker does not correspond to a conversion into more posterior neural tissue since the expression of Six3/6, which is immediately posterior to the anterior neural plate border, was unchanged (row IV; Fig 6B). Moreover, Ap2-like2 expression was also unchanged and did not extend posteriorly (Fig 6C) suggesting that a conversion into epidermis had not occurred. When Wnt signaling was inhibited, Etr was ectopically expressed laterally in the anterior neural plate border (Fig 6Aiii and 6Aiv). These observations suggest that Wnt signaling regulates medio-lateral patterning of the anterior neural plate border. Importantly, as development proceeds, the medial part of the anterior neural plate border stained by Etr will form the very anterior palps region while the lateral part, Etr negative, will form the region immediately posterior containing anterior apical trunk ESNs (aATENs) (Fig 6D and 6E) [51,52].
We next tested the requirement of Wnt signaling for tail PNS formation. In a reciprocal manner to their activation following cWnt activation (S5 Fig), the genes Msxb, Klf1/2/4 and Nkx-C were strongly downregulated following either TcfΔC or Sfrp1/5 overexpression (Figs 7C, 7D and S7). Of these, Msxb was the most affected gene and displayed a complete loss of expression by in situ hybridization for the strongest phenotypes (Figs 7C, 7D and S8). We next assessed ESN formation using Etr as a marker. To avoid confusion with CNS staining, we only scored ventral ESNs. The number and location of ESNs is stochastically determined and so varies from embryo to embryo [43]. For the control embryos electroporated with pFog>Venus, we counted 6.8 ESNs on average (n = 44 embryos). The numbers for the experimental embryos were as follows: 6.8 ESNs for pFt>Sfrp1/5 (n = 46) and 5.8 for pFog>TcfΔC (n = 42). This suggests that Wnt signaling is not essential for tail PNS formation. A possible explanation comes from the observation of Achaete-scute a-like2, a transcription factor expressed in the tail neurogenic midlines [45]. Contrary to the other three genes examined, Achaete-scute a-like2 expression was unchanged following either activation or inhibition of Wnt signaling (Fig 7E–7H). Achaete-scute a-like2 could thus compensate the downregulation of other transcription factors and allow tail PNS formation when Wnt signaling is blocked.
The impacts of modulating cWnt activity on the AP pattern of the ectoderm are summarized in Fig 8. Activating cWnt using either pharmacological Gsk3 inhibitors or overexpression of a constitutively active β-catenin led to very dramatic modifications of the ectoderm AP axis. By analyzing the expression of markers that delineate broad AP domains of the epidermis and the neurogenic epidermis of the PNS (the anterior palp forming region and the caudal midlines), we observed a loss of anterior identity (Otx, Ror-a, Islet and Etr in the palp region; FoxF in the trunk; Bmp2/4, Smad6/7, Nkx-A and Nk4 in the ventral trunk; Figs 1–3 and S2–S4) and a concomitant anterior extension of posterior identity (Cdx, Zf115, Msxb, Klf1/2/4 and Nkx-C; Figs 1–3 and S2–S5), suggesting that the trunk epidermis was respecified as tail epidermis. In the CNS, the situation is less extreme: posteriorization was observed in the anteriormost region (ectopic expression anteriorly for Six1/2 and Six3/6; Figs 2 and 3 and S3 and S4) and within the tail nerve cord (loss of Hox1 in the anterior tail nerve cord and anterior ectopic expression of Hox12; Figs 2 and S3 and S4). Inhibition of cWnt led to downregulation of Msxb, Klf1/2/4 and Nkx-C but not of Achaete-scute a-like2 in the tail neurogenic midlines (Figs 7 and S7), and only to downregulation of Hox12 in the epidermis (Fig 5). In the CNS, the anteriormost and posteriormost regions were affected as revealed by the repression of Six1/2, Six3/6 and Hox12 (Fig 5). The above patterning defects triggered by cWnt are also possibly at play in the other germ layers, endoderm and mesoderm (Figs 2 and S3 and S4), as observed in some vertebrates [53,54].
Data from cWnt activation together with the expression of Wnt5 and Wnt6 posteriorly and of Wnt antagonists (Sfrp1/5, Ror-a and Ror-b) anteriorly fit with a global view of graded Wnt activity from posterior to anterior. These data are in agreement with the proposed ancient role for cWnt signaling in patterning the AP axis during early embryonic development, at least at the base of the deuterostomes [22]. In particular, the ascidian epidermis that contains neurogenic domains (forming ESNs) is highly regionalized along the AP axis. While we are not aware of similar organization in vertebrates, with the exception of specific regions such as the amphibian cement gland, similarities might be drawn with the hemichordate neurogenic ectoderm whose AP pattern is regulated by Wnt signaling [22].
Our results from cWnt activation are at first glance similar to what has been observed in other metazoans–repression of anterior identities and promotion of more posterior identities. We would like to discuss these observations by combining cWnt inhibition data and by restricting our comparisons to deuterostomes (Fig 9).
The marked difference between activation (dramatic posteriorization phenotypes) and inhibition (discrete and limited phenotypes) of cWnt signaling was rather puzzling. An obvious explanation could be the incomplete inhibition of the pathway. TcfΔC has been previously used to inhibit endomesoderm formation in Ciona [25] and we have shown that it can suppress the action of the Gsk3 inhibitor 1-azakenpaullone (Fig 5); and Sfrp1/5 led to similar effects in the experiments presented here. The activation data that we have presented (Figs 1 and S2) show that the ectoderm is responsive to cWnt signaling for a prolonged period of time during gastrulation and neurulation, and cWnt signaling might consequently be ongoing during this time window (only around 4hrs in Ciona developing at 18°C [61]). We thus tested various combinations of the two ectodermal drivers (pFog and pFt), a strong ubiquitous driver (pEf1α [62,63]) as well as combinations of TcfΔC and Sfrp1/5 (S8 Fig). This did not lead to a dramatically stronger repression of the genes that we have tested. While we cannot rule out that cWnt is active in the ectoderm before our earliest driver (pFog: 16-cell stage), we conclude that a partial inhibition of cWnt is not the most likely explanation for the modest phenotypes that we have observed.
In addition to Wnt signaling, several pathways (Fgf, retinoic acid, Shh) participate in patterning the CNS of the vertebrate embryo [56,64–66]. In Ciona, while retinoic acid regulates AP identity of both the CNS and the epidermis at the level of the anterior tail, Fgf regulates tail tip identity of the epidermis but also CNS patterning at various places (tail tip, posterior sensory vesicle, pigment cells, anterior neural plate border) [24,40,67–70]. Fgf signaling is thus likely to interact with cWnt and may act as a redundant signal that compensates for the loss of Wnt signaling in our experiments; we aim at testing their respective functions in future experiments.
A major outcome of cWnt signaling is the regulation of gene expression and transcriptional reporters containing Tcf binding sites have been used as proxies to determine cWnt activity. We have used a reporter previously described in Ciona [40] and found the same global conclusions: reporter activity could be detected in endomesoderm derivatives and in the neurogenic tail ventral midline (S1 Table). We also detected activity in the posterior dorsal midline. In addition, our quantification of reporter activity showed that while endomesodermal activity was detected in a large majority of the embryos, epidermal activity was found, at best, in around 10% of the embryos. Furthermore, this reporter was not active in the CNS regions where we functionally uncovered a role for Wnt signaling. This suggests that cWnt activity in the ectoderm may be very low or possibly at levels undetectable by the reporter used, or that this reporter may not be a faithful readout of cWnt in the ectoderm.
Finally, a major explanation for the modest roles of cWnt in ectoderm AP patterning is likely to stem from the mosaic development of ascidians. In particular, it is well known that the binary AP difference in the ectoderm occurs as early as the 8-cell stage between the trunk and the tail ectoderm precursors, and that FoxA-a acts as an anterior determinant [26,28,71,72]. cWnt might thus be involved, possibly together or redundantly with other signals, in refining this basic pattern. For example, both Wnt5 and Wnt6 are expressed posteriorly and could participate in the definition of the posteriormost CNS and caudal PNS. Wnt6 is also expressed transiently in the anterior neural plate border similarly to Six3/6 and could play a role in the patterning of this region of the embryo [36,37]. Further combinatorial and targeted experiments will be required to definitively determine the precise function of Wnt signaling in ectoderm patterning.
While cWnt is not essential for caudal PNS formation, we have uncovered two distinct functions. First, cWnt appears to interact with Bmp signaling to position within the embryo the ventral neurogenic midline, by regulating the expression of the gene Msxb (S6 Fig). This is not the only mechanism involved since the expression of Achaete-scute a-like2, another early midline gene, is Wnt independent. It would be interesting to uncover and compare the mechanisms that initiate the transcription of both genes in the tail ventral ectoderm through the study of their cis-regulatory DNAs. The timed activation of cWnt allowed us to uncover a later function for cWnt that is independent of the posteriorization; cWnt repressed ESN formation (Fig 1). It is well known that Notch signaling regulates the number of ESNs that form in the caudal midlines and launches a proneural transcriptional cascade [43,73–75]. It will thus be important to determine whether cWnt interacts with this GRN and at which level.
Ripe adults of Ciona intestinalis (formerly referred to Ciona intestinalis type B [32]) were provided by the Centre de Ressources Biologiques Marines in Roscoff (EMBRC-France). Embryo obtention and electroporation were performed as described [72]: 50 μg of each plasmid DNA were used in a 350 μl electroporation volume placed in a 4 mm cuvette and a single pulse of 25V for 32 ms was applied using an ECM830 electroporator (BTX, Harvard Bioscience). Stock solutions of 1-azakenpaullone (191500, Calbiochem, Merck) and BIO (361550, Calbiochem, Merck) were prepared at 10 mM in DMSO. Dilutions were made in sea water just before use at the concentration indicated in the text. Embryo staging and neural plate description were performed according to [50,61].
We have used several previously reported constructs: pFog>Noggin, pFog>Admp and pFog>Venus [43], pFt>ΔN-β-catenin and p12xTcf>nlsLacZ [40]. The other constructs were generated using dedicated Gateway vectors [76]. The activity of the following promoters has been previously described: pFog (pan-ectodermal from the 16-cell stage) [49], pFt (pan-ectodermal from the 64-cell stage) [40] and pEf1α (ubiquitous from early gastrula stages) [62,63]. While the first two were available in Gateway vectors, the last one was introduced following PCR amplification (Forward primer: AAAAAGCAGGCTTTGCTTTACCATCGCGTGACG, reverse primer: AGAAAGCTGGGTTTTGGAAGGTTGGGGTTAACC) using pSPCiEF1α>Cas9 [77] as a template. We have used entry clones containing the coding sequence of ΔN-β-catenin (generated by a BP reaction from pFt>ΔN-β-catenin [40]), TcfΔC (generated by PCR from pRN3-TcfΔC [25]. Forward primer: AAAAAGCAGGCTCAGAAAAAATGCCTCAGTTAAACTCGGA, reverse primer: AGAAAGCTGGGTTCATGGCCGACTTGGTTTG), Sfrp1/5 (generated by RT-PCR from initial tailbud stages C. robusta RNA. Forward primer: CAGAAAAAATGGGATCGTGGATAAAAGGA, reverse primer: TTATCTCCCAGCAGAACCAGTG) and Wnt5 [78] (clone cien109569).
Whole mount in situ hybridization and X-gal staining (detection of β-galactosidase activity following p12xTcf>nlsLacZ electroporation) were performed as previously described [26,79]. Dig-labeled probes were synthesized from C. robusta clones described in previous publications [26,80,81], obtained from cDNA libraries [78,82] or generated by cloning RT-PCR products (from initial tailbud stages embryonic RNA) into pGEM-T Easy (Promega) (S2 Table). Effects on gene expression were analyzed for each marker on 15–40 embryos for inhibitor treatments and 40–70 electroporated embryos (the number of independent experiments is indicated in the figure legend). Embryos treated with DMSO or electroporated with pFog>Venus were used as controls.
Colorimetric detection of endogenous alkaline phosphatase activity was adapted from [13]: embryos were fixed 10 min at room temperature in sea water containing 5% formaldehyde, washed twice 10 min in TMNTw (100 mM NaCl, 50 mM MgCl2, 100 mM Tris pH 9.5, 0.1% Tween20) and stained in TMNTw containing 3.3 μl/ml of NBT (50 mg/ml) and 1.75 μl/ml of BCIP (50 mg/ml).
All pictures were taken from embryos in PBTw using a Zeiss Discovery V20 dissecting scope equipped with an AxioCam ERc5s digital camera. Image panels and figures were constructed with Adobe Photoshop and Adobe Illustrator.
The genes described in this study are represented by the following gene models in the KH2012 C. robusta assembly: genes whose expression has been analyzed by in situ hybridization (see S2 Table), Fog (KH.C10.574), Ft (KH.C11.299), Ef1α (KH.C14.52), Noggin (KH.C12.562), Admp (KH.C2.421), β-catenin (KH.C9.53) Tcf (KH.C6.71), Sfrp1/5 (KH.L171.5), and Wnt5 (KH.L152.45).
|
10.1371/journal.pgen.1005654 | SLO BK Potassium Channels Couple Gap Junctions to Inhibition of Calcium Signaling in Olfactory Neuron Diversification | The C. elegans AWC olfactory neuron pair communicates to specify asymmetric subtypes AWCOFF and AWCON in a stochastic manner. Intercellular communication between AWC and other neurons in a transient NSY-5 gap junction network antagonizes voltage-activated calcium channels, UNC-2 (CaV2) and EGL-19 (CaV1), in the AWCON cell, but how calcium signaling is downregulated by NSY-5 is only partly understood. Here, we show that voltage- and calcium-activated SLO BK potassium channels mediate gap junction signaling to inhibit calcium pathways for asymmetric AWC differentiation. Activation of vertebrate SLO-1 channels causes transient membrane hyperpolarization, which makes it an important negative feedback system for calcium entry through voltage-activated calcium channels. Consistent with the physiological roles of SLO-1, our genetic results suggest that slo-1 BK channels act downstream of NSY-5 gap junctions to inhibit calcium channel-mediated signaling in the specification of AWCON. We also show for the first time that slo-2 BK channels are important for AWC asymmetry and act redundantly with slo-1 to inhibit calcium signaling. In addition, nsy-5-dependent asymmetric expression of slo-1 and slo-2 in the AWCON neuron is necessary and sufficient for AWC asymmetry. SLO-1 and SLO-2 localize close to UNC-2 and EGL-19 in AWC, suggesting a role of possible functional coupling between SLO BK channels and voltage-activated calcium channels in AWC asymmetry. Furthermore, slo-1 and slo-2 regulate the localization of synaptic markers, UNC-2 and RAB-3, in AWC neurons to control AWC asymmetry. We also identify the requirement of bkip-1, which encodes a previously identified auxiliary subunit of SLO-1, for slo-1 and slo-2 function in AWC asymmetry. Together, these results provide an unprecedented molecular link between gap junctions and calcium pathways for terminal differentiation of olfactory neurons.
| Cell type diversity is important for the nervous system to function properly. Asymmetric differentiation of neurons along the left-right axis is one way to achieve diversity; however, the molecular mechanisms used to establish neuronal asymmetry are only partly understood. In the nematode C. elegans, the AWC sensory neuron pair displays stochastic asymmetric identities. Communication between neurons, including the two AWC neurons, through gap junctions inhibits calcium channels in one AWC neuron, resulting in two distinct AWC identities. How gap junctions repress calcium channels in one AWC is not well understood. We show that voltage- and calcium-activated potassium channels provide a molecular link between gap junctions and calcium channels to establish AWC neuronal asymmetry. We show that potassium channels are asymmetrically expressed in AWC neurons, which is dependent on gap junctions. We also find that potassium channels localize close to calcium channels in AWC, suggesting they may functionally couple to establish AWC asymmetry. In addition, our results show that potassium channels regulate the localization of synaptic markers in AWC for asymmetry. Furthermore, we identify an auxiliary subunit of the potassium channels required for their function in establishing AWC asymmetry. These results shed light on mechanisms used to diversify neuronal cell types.
| The nervous system generates a tremendous diversity of cell types that enable formation of functional neural circuits for information processing and mediating behaviors. Cellular diversity is especially important in the developing sensory system as it allows animals to detect different cues in the environment. However, the molecular mechanisms that generate neuronal diversification are only partly understood. One way to generate cellular diversity in the nervous system is to specify different fates and functions of individual cell types across the left-right axis. Left-right asymmetry in the nervous system is present throughout the animal kingdom [1–3]. For example, anatomical and functional asymmetries in the human nervous system have been described, such as the greater size of the planum temporale in the left hemisphere, and the localization of language to the left hemisphere of the brain [4]. Defects in brain asymmetry have been correlated with various neurological diseases such as dyslexia and schizophrenia [5]. In the C. elegans nervous system, two pairs of head sensory neurons display molecular and functional asymmetries: the ASE taste neurons and the AWC olfactory neurons [6–9].
The left and right AWC olfactory neurons appear symmetric at the anatomical and morphological level. However, the two AWC neurons differentiate asymmetrically into two distinct subtypes, one default AWCOFF and one induced AWCON, at both molecular and functional levels in late embryogenesis [10–12]. The AWCON subtype expresses the G-protein coupled receptor (GPCR) gene str-2 and functions to detect the odorant butanone [11,12]. The AWCOFF subtype expresses the GPCR gene srsx-3 and functions to sense the odorant 2,3-pentanedione [12,13]. AWC asymmetry is stochastic, such that the AWCON subtype is induced on the left side of the animal in 50% of the population and on the right side of the animal in the other 50% [11]. AWC asymmetry is maintained throughout the life of an animal [11,14,15].
The default AWCOFF subtype is specified by a calcium-activated protein kinase pathway. In this pathway, calcium entry through voltage-gated calcium channels (the pore-forming α1 subunits UNC-2/N-type or EGL-19/L-type and the regulatory α2δ subunit UNC-36) activates a kinase cascade that consists of UNC-43 calcium/calmodulin dependent protein kinase (CaMKII), the TIR-1 (Sarm1) adaptor protein, NSY-1 MAP kinase kinase kinase (MAPKKK), and SEK-1 MAPKK [10,11,16,17]. TIR-1 assembles a calcium-signaling complex containing UNC-43 (CaMKII) and NSY-1 (MAPKKK) at postsynaptic sites in the AWC axons, in a manner dependent on microtubules and the kinesin motor protein UNC-104, to promote the AWCOFF subtype [10,18]. Intercellular calcium signaling through a transient embryonic neural network, formed between AWC and other neurons via the NSY-5 gap junction protein innexin, coordinates precise AWC asymmetry [19]. In addition, NSY-5 and the NSY-4 claudin-like protein function in parallel to antagonize calcium signaling through mir-71-mediated downregulation of tir-1 expression in the AWCON subtype [20–22]. However, the mechanism by which NSY-5 gap junctions and NSY-4 claudin suppress unc-2/unc-36 and egl-19/unc-36 calcium signaling to induce the AWCON subtype is only beginning to be understood.
The ky389 and ky399 alleles were identified from a forward genetic screen for mutants with two AWCON neurons (2AWCON phenotype) [11]. The ky389 and ky399 mutations were revealed as gain-of-function (gf) alleles of slo-1 in a study demonstrating a central role of slo-1 in behavioral response to ethanol [23]. slo-1 encodes a conserved voltage- and calcium-activated large conductance BK potassium channel [24,25]. Activation of SLO-1 (Slo1) channels causes hyperpolarization of the cell membrane, thereby reducing cellular excitability and limiting calcium entry through voltage-gated calcium channels [26]. The 2AWCON phenotype of slo-1(gf) mutants suggests a sufficient role of slo-1(gf) in promoting AWCON. However, the effect of slo-1 loss-of-function mutations on AWC asymmetry and the mechanism by which slo-1 functions to control AWC asymmetry remained unaddressed. Here we demonstrate that both slo-1 and slo-2 BK channels are necessary for the establishment of AWC asymmetry. We show that slo-1 and slo-2 act redundantly downstream of nsy-5 (innexin gap junction protein) and in parallel with nsy-4 (claudin) to antagonize the function of unc-2 and egl-19 (voltage-gated calcium channels) in the induced AWCON subtype. Asymmetric expression of slo-1 and slo-2 in the AWCON neuron, which is dependent on NSY-5 and NSY-4, is necessary and sufficient for AWC asymmetry. In addition, SLO-1 and SLO-2 BK channels localize close to UNC-2 and EGL-19 voltage-gated calcium channels, suggesting that SLO channels may inhibit calcium channels through functional coupling and negative feedback. Our results also suggest that slo-1 and slo-2 may regulate AWC communication to control AWC asymmetry through modulating UNC-2 synaptic puncta and synaptic vesicle clustering. Thus, our study identifies an unprecedented role of SLO BK potassium channels in mediating transient gap junction signaling for inhibition of a calcium channel-activated kinase cascade in terminal differentiation of olfactory neurons.
Wild-type animals have one default AWCOFF neuron, expressing the GPCR gene srsx-3, and one induced AWCON neuron, expressing the GPCR gene str-2 (Fig 1Ai and 1B). Both slo-1(ky389gf) and slo-1(ky399gf) mutations resulted in expression of the AWCON marker str-2 in two AWC neurons (2AWCON phenotype) (Fig 1Aii, 1B and 1C), as shown previously [11]. Both slo-1(ky389gf)/+ and slo-1(ky399gf)/+ heterozygous animals displayed a less penetrant 2AWCON phenotype (Fig 1B), confirming their characterization as dominant gain-of-function mutants. We set out to further characterize the AWC phenotypes of slo-1(gf) mutants. We found that the AWCOFF marker srsx-3 was not expressed in either of the AWC neurons of slo-1(gf) mutants (Fig 1Aii and 1B), consistent with the 2AWCON phenotype of the mutants (Fig 1Aii and 1B). In addition, overexpression of slo-1(T1001Igf) and slo-1(E350Kgf), containing ky389gf and ky399gf mutations, respectively, in AWC in a wild-type background also caused a strong 2AWCON phenotype (73%-75%) (Fig 1B). However, overexpression of wild-type slo-1 only caused a very weak 2AWCON phenotype (1%) when injected at the same concentration as the slo-1(T1001Igf) and slo-1(E350Kgf) transgenes (Fig 1B). Our results are consistent with the previous electrophysiological study suggesting that slo-1(ky389gf) and slo-1(ky399gf) mutations result in increased SLO-1 channel activity in dopaminergic neurons [23].
Although slo-1(gf) mutants caused a strong 2AWCON phenotype, we found that loss-of-function (lf) mutations in slo-1 did not display any defects in AWC asymmetry (Fig 1B and 1C). This suggests that slo-1 may function redundantly with other genes to establish AWC asymmetry. Since slo-2 encodes the only other calcium-activated SLO-like potassium channel in C. elegans and its expression overlaps with slo-1 [24,27,28], we hypothesized that slo-1 and slo-2 may function redundantly to control AWC asymmetry. Similar to slo-1(lf) mutants, slo-2(lf) mutants did not exhibit abnormalities in AWC asymmetry (Fig 1B and 1D). However, slo-1(eg142lf); slo-2(ok2214lf) double mutants had a complete penetrance of two AWCOFF neurons (2AWCOFF phenotype): the expression of the AWCON marker str-2 was lost and the AWCOFF marker srsx-3 was expressed in both AWC neurons (Fig 1Aiii and 1B). Together, these results suggest that slo-1 and slo-2 have essential and redundant roles in promoting the AWCON subtype.
To determine whether slo-1 and slo-2 affect general AWC fate, we examined the expression of two general AWC markers, the guanylyl cyclase gene odr-1 and the homeodomain protein encoding gene ceh-36, both of which are expressed in both AWC neurons in wild-type animals. Both slo-1(ky399gf) and slo-1(eg142lf); slo-2(ok2214lf) double mutants displayed normal expression of odr-1 and ceh-36 (Fig 1B), suggesting that general AWC identity is not affected by the mutations.
The 2AWCON phenotype of slo-1(gf) mutants and the 2AWCOFF phenotype of slo-1(lf); slo-2(lf) double mutants (Figs 1Aii, 1Aiii, 1B and 2A) indicate that the two BK potassium channels function to promote the induced AWCON subtype. To shed light on how slo-1 and slo-2 promote AWCON, we investigated where they are located within the AWC asymmetry pathway by generating double and triple mutants of slo-1, slo-2, and other genes previously implicated in AWC asymmetry (Fig 2A).
The slo-1(gf) 2AWCON mutants were crossed with 2AWCOFF mutants including nsy-5/innexin(lf), nsy-4/claudin(lf), unc-43/CaMKII(gf), and tir-1/Sarm1(gf) [11,18,20,21]. The 2AWCOFF phenotype of nsy-5(lf) mutants was completely suppressed by slo-1(gf) mutants (Fig 2A), suggesting that slo-1 acts downstream of nsy-5 gap junctions to specify AWCON (Fig 2B). nsy-4(lf); slo-1(gf) double mutants had mixed 2AWCON and 2AWCOFF phenotypes (Fig 2A), suggesting that slo-1 acts in parallel with nsy-4 claudin in promoting AWCON (Fig 2B). Furthermore, the 2AWCON phenotype of slo-1(gf) was nearly completely suppressed by unc-43(gf) and tir-1(gf) mutations (Fig 2A); the 2AWCOFF phenotype of slo-1(lf); slo-2(lf) double mutants was almost completely suppressed by the unc-43(lf) mutants (Fig 2A), which is consistent with the previous notion that slo-1 acts upstream of unc-43 (CaMKII) [16] (Fig 2B).
Since our genetic results put slo-1 and slo-2 (BK potassium channels) at a position similar to unc-2/unc-36 and egl-19/unc-36 (voltage-gated calcium channels) in the AWC asymmetry pathway, we examined the genetic interaction of slo-1 and slo-2 with unc-36. unc-36(e251lf) mutants have a strong 2AWCON phenotype [11](Fig 2A) and were crossed into the slo-1(lf); slo-2(lf) 2AWCOFF mutants. We found that the 2AWCON phenotype of unc-36(lf) and the 2AWCOFF phenotype of slo-1(lf); slo-2(lf) were significantly mutually suppressed in unc-36(lf); slo-1(lf); slo-2(lf) triple mutants (Fig 2A). These genetic analyses suggest antagonistic and parallel functions of BK potassium channels (slo-1 and slo-2) and voltage-gated calcium channels (unc-36) in AWC asymmetry (Fig 2B). unc-2 and egl-19, both of which encode α1 subunits of voltage-activated calcium channels, were shown to have partially redundant functions in AWC asymmetry [13]. unc-2(lj1lf) mutants had a mixed 2AWCON and 2 AWCOFF phenotype, while a reduction of function (rf) allele of egl-19 did not display any AWC asymmetry defects. However, egl-19(rf); unc-2(lf) double mutants caused a strong 2AWCON phenotype reminiscent of unc-36(lf) mutants [13] (Fig 2A), which supports partially redundant functions of egl-19 and unc-2 in AWC asymmetry. To test whether unc-2 and egl-19 interact with slo-1 and slo-2 to establish AWC asymmetry, we determined genetic relationship between unc-2, egl-19, slo-1, and slo-2. The 2AWCOFF phenotype of slo-1(lf); slo-2(lf) was only slightly suppressed or not suppressed in slo-1(lf); slo-2(lf) unc-2(lf) or egl-19(rf); slo-1(lf); slo-2(lf) mutants, respectively (Fig 2A). These results suggest that unc-2(lf) or egl-19(rf) alone are not sufficient to suppress the slo-1(lf); slo-2(lf) 2AWCOFF phenotype. However, egl-19(rf); slo-1(lf); slo-2(lf) unc-2(lf) quadruple mutants displayed a high 2AWCON phenotype, which resembled egl-19(rf); unc-2(lf) mutants (Fig 2A). This result suggests that unc-2 and egl-19 act redundantly to antagonize slo-1 and slo-2 function to promote the AWCOFF identity.
Taken together, these genetic results suggest that slo-1 and slo-2 (BK potassium channels) act downstream of nsy-5 (innexin) and in parallel with nsy-4 (claudin) to antagonize the function of unc-2/unc-36 and egl-19/unc-36 (voltage-gated calcium channels) to induce AWCON (Fig 2B).
Both slo-1 and slo-2 are widely expressed in neurons and muscles [25,28]. To determine if slo-1 and slo-2 are expressed in AWC neurons, we crossed odr-1p::TagRFP (expressed in both AWC neurons) with slo-1p::GFP and slo-2p::GFP transgenic strains. We found that GFP expressed from slo-1p and slo-2p was colocalized with the AWC marker odr-1p::TagRFP (Fig 3A and 3C), suggesting that slo-1 and slo-2 are expressed in AWC.
To determine if slo-1 and slo-2 are expressed asymmetrically in AWC neurons, we compared their respective expression level in AWC left (AWCL) and AWC right (AWCR). Although slo-1 and slo-2 are expressed in both AWC neurons in the majority of wild-type animals, both slo-1 and slo-2 are asymmetrically expressed in AWCL or AWCR in a stochastic manner (Fig 3B and 3D, AWCL>AWCR versus AWCL<AWCR are indistinguishable). Random asymmetry of slo-1 and slo-2 expression in AWC neurons is consistent with the stochastic nature of AWC asymmetry. In contrast, nsy-5(ky634lf) and nsy-4(ky627lf) mutants exhibited a significant increase in the percentage of animals that expressed slo-1 and slo-2 symmetrically in AWCL and AWCR (Fig 3B and 3D, AWCL = AWCR). This data suggests that nsy-5 (innexin) and nsy-4 (claudin) are required for the asymmetric expression of slo-1 and slo-2 in AWC neurons, and is consistent with our genetic analysis demonstrating that nsy-5 acts in parallel with nsy-4 to promote AWCON through slo-1 and slo-2 (Fig 2B). As a control, we examined the asymmetric expression of slo-1 and slo-2 in unc-43(n498gf)/CaMKII mutants, which cause a 2AWCOFF phenotype, similar to that caused by nsy-5(ky634lf) and nsy-4(ky627lf). We found that the asymmetric expression of both slo-1 and slo-2 was unaffected by the unc-43(n498gf) mutation (Fig 3B and 3D). This suggests that unc-43 (CaMKII) does not regulate the expression of slo-1 and slo-2, and is consistent with our genetic results, which place slo-1 and slo-2 upstream of unc-43/CaMKII. The result also supports that the effect of nsy-4 and nsy-5 loss-of-function mutations on asymmetric expression of slo-1 and slo-2 was not due to the 2AWCOFF phenotype.
We also compared expression level of slo-1 and slo-2 in AWCON and AWCOFF, and found that slo-1 and slo-2 are expressed predominantly in the AWCON cell (Fig 3E–3H). These results are consistent with the hypothesis that slo-1 and slo-2 promote AWCON in a cell-autonomous manner.
To determine the site of slo-1 and slo-2 function in promoting AWCON, we performed genetic mosaic analysis in slo-1(lf); slo-2(lf) mutants containing an integrated AWCON marker (str-2p::GFP) transgene and the extrachromosomal array odr-3p::slo-1(overexpressor (OE)); odr-1p::DsRed or odr-3p::slo-2(OE); odr-1p::DsRed. Both odr-3p::slo-1(OE) and odr-3p::slo-2(OE) transgenes rescued the 2AWCOFF phenotype in slo-1(lf); slo-2(lf) mutants and also caused a slight 2AWCON overexpression phenotype (Fig 4A and 4C). Since extrachromosomal transgenes are unstable and can be randomly lost at each cell division, the co-injected marker odr-1p::DsRed (normally expressed in both AWC) was used to indicate the presence of the slo-1(OE) or slo-2(OE) array in AWC. Specifically, we determined if retention of the slo-1(OE) or slo-2(OE) array in only a single AWC cell causes a bias of AWCON choice in that cell when the mosaic animals exhibited a wild-type 1AWCON/1AWCOFF phenotype. We found that the slo-1(OE); slo-2(lf) AWC became AWCON and the slo-1(lf); slo-2(lf) AWC became AWCOFF in the majority of mosaic animals in which slo-1 was expressed only in a single AWC neuron (Fig 4B). Similarly, the slo-1(lf); slo-2(OE) AWC became AWCON and the slo-1(lf); slo-2(lf) AWC became AWCOFF when mosaic animals expressed slo-2 only in a single AWC neuron (Fig 4D). Together, these results support that slo-1 and slo-2 act cell autonomously to specify AWCON. We did observe a very small percentage of mosaic animals in which the slo-1(lf); slo-2(lf) AWC became AWCON (Fig 4B and 4D). This suggests that although slo-1 and slo-2 have a largely cell-autonomous role in promoting the AWCON fate, they may also have a nonautonomous role. This is similar to other genes in the AWC asymmetry pathway, such as nsy-5 and nsy-4 which display both autonomous and nonautonomous roles in AWC asymmetry [20,21].
Mosaic analysis was also performed in transgenic lines in which slo-1(T1001Igf), containing the ky389gf mutation, was overexpressed in a wild-type background, resulting in a strong 2AWCON phenotype (Fig 4E). When the transgene was retained in only one of the two AWC cells, the slo-1(gf) cell became AWCON and the wild-type cell became AWCOFF in the majority of mosaic animals (Fig 4F). This result is consistent with a largely cell-autonomous function of slo-1 in promoting AWCON, and also suggests that the AWC with slo-1(gf) activity may become hyperpolarized, allowing the cell to reduce calcium influx and take on the AWCON subtype.
SLO-1 and SLO-2 have overlapping expression patterns and have been suggested to potentially form heteromeric channels [24,28,29]. In addition, it has been shown that BK channels and N-type voltage-gated calcium channels localize in close proximity to achieve functional coupling of these channels [30]. To determine if SLO-1, SLO-2, UNC-2 (N/P/Q-type calcium channels), and EGL-19 (L-type calcium channels) localize in close proximity in AWC, we generated single copy transgenes expressing functional translational reporters driven by the AWC odr-3 promoter using Mos1-mediated single copy insertion [31–33]. The tagged proteins expressed in these transgenes were functional in rescuing respective mutant phenotypes [34](S1 Fig, Materials and Methods).
These single copy insertion transgenes showed that GFP::UNC-2, SLO-1::TagRFP, SLO-1::GFP, SLO-2::TagRFP, and GFP::EGL-19 were mainly localized on the plasma membrane of AWC cell bodies and also displayed a punctate pattern along AWC axons (Fig 5 and S2 Fig), similar to the previously shown localization pattern of GFP::UNC-2 in AWC [34]. Since these channels were localized throughout the plasma membrane of the AWC cell body and had distinct punctate patterns in AWC axons, we focused on analyzing their localization in relation to each other in AWC axons. We found that both SLO-1::TagRFP and SLO-2::TagRFP were localized adjacent to GFP::UNC-2 and GFP::EGL-19 in AWC axons (Fig 5A and 5B, S2A and S2B Fig). In addition, SLO-2::TagRFP is located close to SLO-1::GFP in AWC axons (Fig 5C). The Coloc 2 plugin in Fiji was used to quantify colocalization of these proteins in AWC axons using three different algorithms (Pearson’s correlation coefficient, Spearman’s rank correlation coefficient, and Li’s ICQ). Each of the algorithms displayed positive correlation indices (Fig 5D and S2C Fig). This further supports that UNC-2 and EGL-19 localize close to SLO-1 and SLO-2, and that SLO-1 and SLO-2 are localized in close proximity as well.
These results support the notion that BK potassium channels (SLO-1 and SLO-2) and voltage-gated calcium channels (UNC-2 and EGL-19) may function in close proximity for rapid activation of SLO-1 and SLO-2 channels by locally increased calcium levels near UNC-2 and EGL-19 calcium channels.
It has been shown that communication between the pair of AWC neurons via chemical synapses in axons is important for induction of the AWCON subtype [11]. Our genetic data suggests that slo-1 and slo-2 are required for the specification of the induced AWCON subtype. In addition, SLO-1 and SLO-2 displayed distinct punctate localization patterns in AWC axons. Thus, we examined whether slo-1 and slo-2 regulate localization of synaptic markers in AWC neurons. To do so, we generated Mos1-mediated single copy insertion transgenes expressing fluorescently tagged synaptic markers, GFP::UNC-2 and YFP::RAB-3, driven by the AWC odr-3 promoter (Figs 5A, 5B and 6). UNC-2 is localized to presynaptic active zones and RAB-3 is a synaptic vesicle marker [34].
In wild type, GFP::UNC-2 was localized in the AWC axon and cell body, and YFP::RAB-3 was mainly localized in a punctate pattern in the AWC axon as shown previously [34] (Fig 6). In slo-1(ky399gf) animals, the intensity of GFP::UNC-2 or YFP::RAB-3 was not significantly affected in the AWC axon and cell body (Fig 6). However, slo-1(eg142lf); slo-2(ok2214lf) mutants displayed significant reduction in intensity of GFP::UNC-2 and YFP::RAB-3 in the AWC axon and cell body (Fig 6). These results suggest that slo-1 and slo-2 are required for localization and/or stability of synaptic markers, UNC-2 and RAB-3, in AWC neurons, which may contribute to the 2AWCOFF phenotype caused by the slo-1(eg142lf); slo-2(ok2214lf) mutations. Our genetic mosaic analysis suggests a minor role of nonautonomous function of slo-1 and slo-2 in establishing AWC asymmetry (Fig 4), which is consistent with a possible role of slo-1 and slo-2 in regulating synaptic communication of AWC neurons. In addition, autofluorescence of the gut found in wild-type animals was visibly decreased in slo-1(eg142); slo-2(ok2214lf) mutants (S3A Fig), suggesting that the SLO channels are required for gut autofluorescence.
As a control, the intensity of GFP expressed from the transgene odr-3p::GFP was analyzed in wild-type and mutant backgrounds, and no significant effect was observed in slo-1(ky399gf) and slo-1(eg142lf); slo-2(ok2214lf) mutants (S3B Fig). This result rules out the possibility that the activity of the odr-3 promoter is regulated by slo-1 and slo-2, and also supports the notion that the effect of slo-1(lf); slo-2(lf) mutations on UNC-2 and RAB-3 is mainly at the subcellular localization level. It is also possible that slo-1(lf); slo-2(lf) mutations may affect unc-2 and rab-3 at post-transcriptional levels, such as translation efficiency, mRNA and/or protein stability.
Previous studies have shown that slo-1(lf) or slo-2(lf) mutations result in increased neurotransmitter release at the neuromuscular junction in the ventral nerve cord [25,35]. However, a recent study showed that UNC-2 localization is not affected at the presynaptic terminals of neuromuscular junctions in slo-1(lf) mutants [36]. Thus, previous findings did not demonstrate a correlation between increased neurotransmitter release and increased localization of UNC-2 or RAB-3 at presynaptic sites of the neuromuscular junction in slo-1(lf) or slo-2(lf) mutants. To examine whether the localization of UNC-2 and RAB-3 is affected in ventral cord motor neurons in slo-1(lf); slo-2(lf) mutants, we quantified the intensity of GFP::UNC-2 and RAB-3::mCherry driven by the unc-25 promoter, which is expressed in ventral cord motor neurons [34]. We examined the axons located anterior to VD5 and DD3 neurons in wild type and slo-1(lf); slo-2(lf) mutants at the L4 stage, but no significant difference was observed (S4 Fig). This suggests that slo-1 and slo-2 do not play an apparent role in the localization of these presynaptic markers in the ventral nerve cord. The different effects of slo-1 and slo-2 mutations on the localization of synaptic markers in AWC neurons and ventral cord motor neurons suggest that slo-1 and slo-2 take on a different function in AWC neurons than in the ventral cord motor neurons.
Although no apparent effect of slo-1(lf); slo-2(lf) mutations on the localization of UNC-2 and RAB-3 was observed in ventral cord motor neurons, the effect of slo-1 and slo-2 mutations on locomotion was performed by analyzing the wavelength and wave width of body wave tracks of wild type, slo-1(lf), slo-2(lf), and slo-1(lf); slo-2(lf) animals. We found that the wavelength of the worm track was not affected in the mutants, however the wave width was significantly increased in slo-1(lf), slo-2(lf), and slo-1(lf); slo-2(lf) mutants (S5 Fig). These results suggest that slo-1 and slo-2 are required for normal locomotion.
Previous studies have identified several modulators of SLO-1 activity in muscles using forward genetic screens. Since genes may interact in similar pathways in different tissues, we chose these candidate genes to determine whether they may also modulate SLO-1 activity in AWC neurons. bkip-1 mutants were identified from a screen for suppressors of the lethargic phenotype of slo-1(gf) mutants. BKIP-1 (BK channel Interacting Protein), a single pass membrane protein, functions as an auxiliary subunit of SLO-1 to assist in regulating neurotransmitter release and regulate the surface expression of the channel [37]. Similar to bkip-1, ctn-1 (α-catulin), identified from two independent screens for suppressors of the slo-1(gf) lethargic phenotype, also regulates surface localization of SLO-1 in both muscles and ventral nerve cord motor neurons [38,39]. In addition, components of the dystrophin-associated protein complex (DAPC), including dys-1 (dystrophin), dyb-1 (dystrobrevin), stn-1 (syntrophin), and dyc-1 (C-terminal PDZ-domain ligand of nNOS), control the localization of SLO-1 in muscles but not in neurons [40,41]. Furthermore, islo-1, encoding a transmembrane protein, functions as an adaptor protein that links the DAPC to SLO-1 for SLO-1 localization in muscles [40].
To determine whether bkip-1, ctn-1, dys-1, and islo-1 play a role in AWC asymmetry, we first examined expression of the AWCON marker str-2p::GFP in their respective loss-of-function mutants, but did not see any defects in AWC asymmetry (Fig 7A). We then determined whether bkip-1(lf), ctn-1(lf), dys-1(lf), and islo-1(lf) mutants suppress the slo-1(gf) 2AWCON phenotype in AWC asymmetry by performing double mutant analysis. We found that dys-1(cx18); slo-1(ky399gf), ctn-1(eg116); slo-1(ky399gf), and islo-1(eg978); slo-1(ky399gf) all displayed the same 2AWCON phenotype as slo-1(ky399gf) animals (Fig 7A). This suggests that dys-1, ctn-1, and islo-1 are not required for slo-1 function in AWC asymmetry. However, bkip-1(zw2) completely suppressed the 2AWCON phenotype of both slo-1(ky389gf) and slo-1(ky399gf) mutants to wild type (Fig 7A), indicating that bkip-1 is required for slo-1 function in promoting AWCON. As shown by our results, slo-1(lf) and slo-2(lf) single mutants did not display AWC asymmetry defects (Figs 1B and 7A). However, both bkip-1(lf); slo-1(lf) and bkip-1(lf); slo-2(lf) displayed a 2AWCOFF phenotype (Fig 7A), supporting a role of bkip-1 in both slo-2 and slo-1 function, respectively. However, the 2AWCOFF phenotype of bkip-1(lf); slo-1(lf) and bkip-1(lf); slo-2(lf) was not 100% as seen in slo-1(lf); slo-2(lf) double mutants (Fig 7A), suggesting that bkip-1 is not the only factor required for slo-1 and slo-2 function in AWC asymmetry. We also determined whether slo-2 is required for slo-1 function by crossing slo-2(lf) mutants into both slo-1(ky389gf) and slo-1(ky399gf) alleles. We found that slo-2(lf) did not suppress the slo-1(gf) 2AWCON phenotype (Fig 7A), suggesting that slo-2 is not required for slo-1 function in AWC asymmetry.
Previous work demonstrates a role of bkip-1 in regulating the surface expression of SLO-1 in muscle dense bodies and the nerve ring [37]. We therefore determined whether bkip-1 affects SLO-1 localization in AWC neurons by examining a functional SLO-1::GFP translational reporter driven by the AWC odr-3 promoter in wild type and bkip-1(zw2) mutants (Fig 7B). We found that in bkip-1(zw2) mutants, SLO-1::GFP intensity was significantly reduced in AWC axons (Fig 7B and 7C) but is not significantly affected in cell bodies (Fig 7D). This suggests that bkip-1 is required for appropriate localization of SLO-1 in AWC axons but not in the AWC cell body. Consistent with the result suggesting that bkip-1 is not the only factor required for slo-1 function in AWC asymmetry (Fig 7A), this result also suggests that slo-1 activity could be required in both the AWC axons (dependent on bkip-1) and cell bodies (independent of bkip-1). We also examined whether bkip-1(zw2) mutants display altered the localization of SLO-2::GFP in AWC axons, but did not find a significant effect (S6 Fig). This result suggests that slo-2 may require bkip-1 in a manner independent of appropriate localization. bkip-1 may be required for appropriate slo-2 expression levels, or BKIP-1 may physically interact with SLO-2.
Together, our results showed that bkip-1 is the only one of the known modulators of slo-1 activity in muscles to be also required for slo-1 and slo-2 function in AWC asymmetry. Thus, our results suggest that slo-1 and slo-2 need a different set of regulators for their function in AWC asymmetry.
The voltage-dependent activation of SLO-1 and SLO-2 channels is modulated by calcium (for SLO-1 and SLO-2) and chloride (for SLO-2) [24,26]. To determine whether any chloride channels or other voltage-gated potassium channels might be involved in establishing left-right AWC asymmetry, we examined AWC asymmetry in mutants of selective channels that have been shown to be expressed in the nervous system (WormBase). Although the majority of mutants examined did not display an AWC asymmetry defect (S7A Fig), a gain of function mutation in unc-103 (ERG voltage-gated potassium channel) resulted in a slight 2AWCON phenotype (S7B Fig). In addition, a gain of function mutation in egl-2 (EAG voltage-gated potassium channel) caused a high penetrance of the 2AWCON phenotype (S7B Fig), as previously shown [11]. We found that the egl-2(n693gf) mutation suppressed the 2AWCOFF phenotype observed in slo-1(eg142); slo-2(ok2214) double mutants, nsy-5(ky634lf), unc-43(n498gf), and tir-1(ky648gf) single mutants (S7B Fig). This suggests that egl-2 may function downstream of these genes to promote the AWCON fate. Alternatively, it is possible that egl-2 may function at the same level as slo-1 and slo-2; and that the production of 2 AWCON neurons by egl-2(gf) in the slo-1(eg142); slo-2(ok2214) mutants is because egl-2(gf) is sufficient to cause enough membrane hyperpolarization to induce AWCON even in the absence of slo-1 and slo-2. Like slo-1(lf) and slo-2(lf) mutants, loss-of-function mutations in egl-2 did not cause a significant effect on AWC asymmetry nor did slo-1(lf); egl-2(lf) double mutants (S7B Fig), suggesting that egl-2 may act redundantly with other factor(s) in promoting AWCON.
Here we identify an essential role of SLO BK potassium channels in asymmetric differentiation of one pair of olfactory neurons. Our findings reveal a functional link between gap junctions and SLO channels in inhibition of voltage-gated calcium channels for diversification of olfactory neurons. To the best of our knowledge, stochastic AWC asymmetry is the first system in which SLO channels are implicated in terminal neuron differentiation, stochastic cell fate determination, and left-right patterning.
Our results suggest antagonistic and parallel functions of BK potassium channels (SLO-1 and SLO-2) and voltage-gated calcium channels (UNC-2/UNC-36 and EGL-19/UNC-36) downstream of NSY-5 gap junctions in AWC asymmetry. UNC-2/UNC-36 and EGL-19/UNC-36 activate a CaMKII-MAP kinase cascade to specify the default AWCOFF subtype, while SLO-1 and SLO-2 inhibit the calcium channel-activated kinase cascade to promote the induced AWCON subtype (Fig 8). Calcium and voltage are potential signals that mediate intercellular communication between the two AWC neurons and other neurons in the NSY-5 gap junction network to coordinate stochastic AWC asymmetry [19,20]. In addition, both SLO BK channels and voltage-gated calcium channels generate voltage and calcium signals, and are subject to calcium- and voltage-dependent activation and inactivation [26,42]. The regulatory loop between gap junctions, SLO BK channels, and voltage-gated calcium channels can potentially generate sustained differences in calcium-regulated signaling outputs between the two AWC cells through positive and negative feedback mechanisms, leading to asymmetric differentiation of AWC cells. This extends the previous model of NSY-5 function in AWC asymmetry by identifying SLO BK channels as the mediators of transient gap junction signaling for antagonizing voltage-gated calcium channel pathways.
Signaling via NSY-5 gap junctions may lead to transcriptional regulation of slo-1 and slo-2 in order to ensure that these genes are expressed asymmetrically in the AWC neurons. Studies have shown that connexin gap junction proteins are capable of regulating gene expression. For example, gap junction communication mediated by Cx43 is required for ERK phosphorylation of the transcription factor Sp1, which in turn leads to appropriate expression of an osteoclastin transcriptional element [43]. It has been suggested that gap junctions may allow diffusion of second messengers such as calcium and cyclic nucleotides, which subsequently can influence gene transcription [43,44]. It has also been suggested that C-terminal tails of connexins may bind to particular proteins, which can then contribute to regulating gene expression [44]. It is possible that NSY-5 gap junctions use similar mechanisms to regulate slo-1 and slo-2 gene expression.
SLO-1 is 55% identical to its mouse orthologue Slo1 and SLO-2 is 41% identical to its mammalian orthologue Slack, while SLO-1 is only 18% identical to its nematode paralogue SLO-2 along the entire channel peptide [28]. SLO-1/Slo1 and SLO-2/Slack have overlapping expression patterns and may form heteromeric channels [24,28,29]. However, functional relationships between SLO-1/Slo1 and SLO-2/Slack have not yet been demonstrated in any biological contexts. Our results show that SLO-1 localizes in close proximity to SLO-2 in AWC neurons. In addition, our results suggest that slo-1 and slo-2 have complete functional redundancy in AWC asymmetry, since loss-of-function mutations in either gene alone did not cause any defects in AWC asymmetry while slo-1(lf); slo-2(lf) double mutants displayed a complete penetrance of the 2AWCOFF phenotype. Functional redundancy between SLO-1/Slo1 and SLO-2/Slack may represent one of the general mechanisms for their roles in other systems.
The voltage range of activation of BK channels is modulated by different intracellular factors including calcium (for SLO-1, Slo1, and SLO-2), chloride (for SLO-2 and Slack), sodium (for Slack), pH, and phosphorylation [24,26]. None of the mutants of chloride channels we examined displayed any AWC asymmetry defects. In addition, although SLO-2 shares a complete redundant function with SLO-1 in AWC asymmetry, it has not been shown that the activation of SLO-1 channels is sensitive to chloride. These findings suggest that SLO-2’s redundant role with SLO-1 in establishing AWC asymmetry may be more dependent on sensitivity to calcium than to chloride.
Calcium-activated BK channels and voltage-gated calcium channels have been shown to localize in close proximity to ensure selective and rapid activation of BK channels by a local increase in cytosolic calcium level [30]. The sensitivity of vertebrate Slo1 channels to calcium provides an important negative feedback for calcium entry in many cell types. For example, activation of Slo1 channels causes transient membrane hyperpolarization, which limits calcium entry through voltage-gated calcium channels to control the burst of calcium action potentials in cerebellar Purkinje cells and to regulate synaptic transmission in presynaptic terminals [26]. Our genetic results and findings that SLO-1 and SLO-2 localize close to UNC-2 and EGL-19 voltage-gated calcium channels are consistent with the physiological roles of vertebrate Slo1 channels in inhibiting voltage-gated calcium channels through functional coupling and negative feedback. By analogy to functional coupling between Slo1 and voltage-gated calcium channels in vertebrates, SLO-1 and SLO-2 may couple with UNC-2/UNC-36 and EGL-19/UNC-36 to generate oscillation of cytosolic calcium and voltage signals to coordinate stochastic AWC asymmetry through a feedback loop. In this hypothetical feedback loop, an increase in voltage triggers voltage-gated calcium channels to open, leading to an increase in intracellular calcium levels. High calcium levels allow the coupled calcium-activated BK channels to open, resulting in a decrease in voltage. The decreased voltage causes the voltage-gated calcium channels to close, leading to a decrease in intracellular free calcium levels and the subsequent closure of calcium-activated BK channels and an increase in voltage. This would initiate another cycle of calcium and voltage oscillation. Previous studies identified two forms of intercellular communication important for AWC asymmetry: one is mediated by NSY-5 gap junctions between the cell body of AWC and other neurons in a network [19,20]; the other is by synaptic connection between two AWC axons [10,11]. Since SLO-1 and SLO-2 are localized in proximity to UNC-2 and EGL-19 at the AWC axons, functional coupling between BK channels (SLO-1 and SLO-2) and voltage-activated calcium channels (UNC-2 and EGL-19) may occur at AWC axons.
Our study has revealed that SLO-1 and SLO-2 have different functions and interacting partners in AWC olfactory neurons than in ventral cord motor neurons. Our genetic analysis suggests that BK potassium channels (SLO-1 and SLO-2) act to antagonize calcium channels (UNC-2/UNC-36 and EGL-19/UNC-36) to promote the AWCON identity. A recent report suggests that in M4 motor neurons, UNC-2 and UNC-36 function to activate SLO-1, which in turn antagonizes the EGL-19 calcium channel to inhibit synaptic transmission at the M4 neuromuscular junction [45]. A recent study showed that UNC-2 localization is not affected at the presynaptic terminals of neuromuscular junctions in slo-1(lf) mutants [36]. However, our results suggest that slo-1 and slo-2 are required for appropriate localization or stability of presynaptic markers UNC-2 and RAB-3 in AWC axons. We also show that SLO-1 and SLO-2 localize in close proximity to both UNC-2 and EGL-19 calcium channels in AWC neurons, in contrast to a report that SLO-2 exclusively couples with EGL-19 but not with UNC-2 in ventral cord motor neurons [35].
BK channels are ubiquitously expressed and have a staggering repertoire of functions in different tissues. To achieve functional diversity, BK channels, which assemble as tetramers of pore-forming α-subunits, can form complexes with various auxiliary β-subunits. For example, the β1 subunit changes gating and calcium sensitivity of Slo1 α subunits, and β2 subunits promote fast inactivation of Slo1 channels [26]. In addition, functional diversity of Slo1 channels can be achieved by alternative splicing, posttranslational modifications, and heteromultimer formation [26]. In C. elegans, several modulators have been identified for surface expression and activity of SLO-1 channels in muscles and neurons [37–41]. Our results show that the auxiliary subunit BKIP-1 is the only previously identified modulator of SLO-1 to be required for SLO-1 and SLO-2 function in asymmetric AWC differentiation. AWC asymmetry may provide an effective model system to identify novel modulators of SLO BK channels in vivo due to the ease of unbiased forward genetic screens in identifying biologically relevant genes and robust phenotypic readouts of SLO channel activity.
Wild type is strain N2, C. elegans variety Bristol. Strains were maintained by standard methods [46]. Mutants used were as follows: nsy-5(ky634) I [20], dys-1(cx18) I [47], ctn-1(eg116) I [38], avr-14(ad1302) I [48], bkip-1(zw2) II [37], clh-1(qa900) II, clh-1(qa901) II, clh-2(ok636) II, clh-3(ok763) II, unc-36(e251) III [46], tir-1(ky648gf) III [18], nsy-4(ky627) IV [21], unc-103(e1597gf) III, egl-19(n582) IV [49], islo-1(eg978) IV [40], unc-43(n498gf) IV [50], unc-43(n1186) IV, slo-1(ky399gf) V [11], slo-1(ky389gf) V [11], slo-1(eg142) V [40], slo-1(js118) V [25], slo-1(js379) V [25], egl-2(n693gf) V, egl-2(n693n904) V, exp-2(sa26ad1426) V, shw-3(ok1884) V, clh-6(ok791) V, avr-15(ad1051) V, slo-2(ok2214) X (C. elegans knockout consortium), slo-2(nf100) X [51], egl-36(n728) X, egl-36(n728n398) X, and clh-4(ok1162) X, unc-2(lj1) X [52].
Integrated transgenes used include kyIs140 [str-2p::GFP; lin-15(+)] I [11], vyIs76 [ceh-36p::myrTagRFP; ofm-1p::DsRed] I, vyIs58 [odr-1p::TagRFP] I, vyIs56 [odr-1p::TagRFP] III, vyIs68 [str-2p::TagRFP; srsx-3p::GFP] II [7], vySi8 [odr-3p::slo-2c::TagRFP; unc-119(+)] II, vySi18 [odr-3p::GFP::unc-2; unc-119(+)] II, vySi23 [odr-3p::slo-1a::TagRFP; unc-119(+)] II, vySi38 [odr-3p::slo-1a::GFP; unc-119(+)] II, vySi39 [odr-3p::YFP::rab-3; unc-119(+)] II, vySi58 [odr-3p::slo-2::TagRFP; unc-119(+)] IV, vyIs51 [str-2p::2xnlsTagRFP; ofm-1p::DsRed] V [18], vyIs74 [ceh-36p::myrTagRFP; ofm-1p::DsRed] V, otIs264 [ceh-36p::TagRFP] [53], kyIs479 [unc-25p::GFP::unc-2; unc-25::mCherry::rab-3; odr-1p::mCherry] [34], vyTi2 [odr-3p::GFP::egl-19], and vyTi4 [odr-3p::GFP::egl-19]. Transgenes maintained as extrachromosomal arrays include vyEx842, 843 [nsy-5p::slo-1 (7.5 ng/μl); odr-1p::DsRed (15 ng/μl); ofm-1p::DsRed (30 ng/μl)], vyEx1573, 1574, 1575, 1576 [nsy-5p::slo-1(T1001Igf) (7.5 ng/μl); odr-1p::DsRed (15 ng/μl); ofm-1p::DsRed (30 ng/μl)], vyEx822, 823 [nsy-5p::slo-1(E350Kgf) (7.5 ng/μl); odr-1p::DsRed (15 ng/μl); ofm-1p::DsRed (30 ng/μl)], vyEx1539, 1540 [slo-1p::GFP (15 ng/μl); ofm-1::DsRed (30 ng/μl)], vyEx1684 [slo-1p::2xnlsGFP (5 ng/μl)], vyEx1701 [slo-1p::2xnlsGFP (2 ng/μl; pRF4(rol-6(su1006) (50 ng/μl)], sEx10749 [slo-2p::GFP; pCeh361], vyEx1122, 1151 [odr-3p::slo-1 (30 ng/μl); odr-1p::DsRed (15 ng/μl); ofm-1p::DsRed (30 ng/μl)], vyEx1682 [ceh-36p::myrTagRFP (5 ng/μl)], vyEx1239 [odr-3p::slo-2c (30 ng/μl); odr-1p::DsRed (15 ng/μl); ofm-1p::DsRed (30 ng/μl)], vyEx1572 [odr-3p::slo-2d (30 ng/μl); odr-1p::DsRed (15 ng/μl); ofm-1p::DsRed (30 ng/μl)], vyEx1418 [odr-3p::slo-1::GFP (20 ng/μl); ofm-1p::DsRed (30 ng/μl], vyEx1393, 1367 [slo-1p::slo-1 (15 ng/μl); odr-1p::DsRed (15 ng/μl); ofm-1p::DsRed (30 ng/μl)], vyEx1266 [slo-1p::slo-1::GFP (7.5 ng/μl); ofm-1p::DsRed (30 ng/μl)], vyEx1594 [odr-3p::slo-1::TagRFP (30 ng/μl); ofm-1p::DsRed (30 ng/μl)], vyEx1325, 1326 [odr-3p::slo-2c::TagRFP (30 ng/μl); elt-2p::GFP (5 ngl/μl)], and vyEx611 [odr-3p::GFP (7.5 ng/μl); elt-2p::CFP (7.5 ng/μl)].
To make nsy-5p::slo-1, a 3420 bp fragment of full-length slo-1a cDNA was amplified from snb-1p::slo-1a (pBK3-1) [25] and cloned into a vector containing a 5556 bp of nsy-5 promoter [20]. nsy-5p::slo-1(T1001Igf) and nsy-5p::slo-1(E350Kgf) were generated by site directed mutagenesis of nsy-5p::slo-1 using a QuikChange II XL Site-Directed Mutagenesis Kit (Stratagene). slo-1p::GFP was made by replacing slo-1a::GFP in the slo-1p::slo-1a::GFP vector containing a 5239 bp of slo-1 promoter [25] with GFP. slo-1p::slo-1 was generated by subcloning the 3420 bp of slo-1a coding region into the vector containing a 5239 bp of slo-1 promoter. odr-3p::slo-1 was made by subcloning the 3420 bp of slo-1a coding region into a vector containing the odr-3 promoter (Roayaie et al. 1998). odr-3p::slo-2c and odr-3p::slo-2d were generated by cloning 3261 bp of slo-2c and 3351 bp of slo-2d, respectively, into the odr-3p vector.
odr-3p::slo-1::GFP was made by subcloning the slo-1a::GFP translation fusion from slo-1p::slo-1a::GFP [25] into the odr-3p vector. To make odr-3p::slo-1::TagRFP, TagRFP was inserted into the slo-1a cDNA at a location corresponding to a region of the protein between S8 and S9, the same insertion site as GFP in slo-1p::slo-1a::GFP [25], using fusion PCR. odr-3p::slo-2c::TagRFP was made by inserting TagRFP into the slo-2c cDNA at a location corresponding to a region of the protein between the last 16th and 15th amino acids from the C-terminus, the same insertion site as GFP in a slo-2::GFP translation fusion construct [28], using fusion PCR. For Mos1-mediated single copy insertion (MosSCI) of these translational reporter transgenes, a pAB1 construct [14], derived from a pCFJ151 MosSCI insertion vector for integration on chromosome II [31], was modified to generate pAB1.1 that includes a new set of restriction enzyme sites. odr-3p::slo-1::TagRFP, odr-3p::slo-2::TagRFP, and odr-3p::YFP::rab-3 fragments were subcloned into pAB1.1 to generate pAB1.1::odr-3p::slo-1::TagRFP, pAB1.1::odr-3p::slo-2::TagRFP, and pAB1.1::odr-3p::YFP::rab-3 respectively. To make pAB1.1::odr-3p::GFP::unc-2, partially overlapped fragments of linearized pAB1.1 vector backbone as well as odr-3p::GFP and GFP::unc-2, both of which were PCR amplified from odr-3p::GFP::unc-2 [34], were assembled and ligated using Gibson Assembly (New England Biolabs; Ipsiwich, MA). pAB1.1::odr-3p::slo-1::GFP was made by Gibson Assembly of 3 partially overlapped fragments of pAB1.1::odr-3p vector backbone linearized from pAB1.1::odr-3p::slo-2::TagRFP, odr-3p::slo-1::GFP, and unc-54 3’UTR. odr-3p::slo-2::TagRFP fragment was cloned into the pCFJ356 MosSCI insertion vector for integration on chromosome IV [33] to generate pCFJ356::odr-3p::slo-2::TagRFP. odr-3p::GFP::egl-19 miniMos construct was generated by replacing snt-1p::HALO in the snt-1p::HALO::egl-19 miniMos construct (pSAM354), containing a section of egl-19 gDNA (exons 5–9 and introns in between the exons) sandwiched between two stretches of cDNA (exons 1–4 and 10–17), with an odr-3p::GFP fragment from odr-3p::GFP::unc-2 [33]. We found that a set-18p::GFP::egl-19 transgenic array rescued the locomotory phenotypes of the egl-19(n582) hypomorph mutant, supporting that GFP::EGL-19 translational reporter is functional.
Transgenic strains were generated by injecting DNA constructs into the syncytial gonad of adult worms (P0) as previously described [54]. F1 worms expressing fluorescent transgenes were picked and cloned (1 worm per plate). The F1 clones that have F2 progeny containing fluorescent transgenes were selected as transgenic lines and analyzed.
MosSCI lines were generated using the direct insertion protocol as previously described [31,33]. Briefly, pAB1.1::odr-3p::GFP::unc-2 (22 ng/μl), pAB1.1::odr-3p::slo-2::TagRFP (71 ng/μl), pAB1.1::odr-3p::slo-1::TagRFP (107 ng/μl), pAB1.1::odr-3p::slo-1::GFP (43 ng/μl), or pAB1.1::odr-3p::YFP::rab-3 (26 ng/μl) was injected along with hsp16.4p::peel-1 (10 ng/μl), eft-3p::mos-1 (50 ng/μl), rab-3p::mCherry (10 ng/μl), myo-2p::mCherry (2.5 ng/μl), and myo-3p::mCherry (5 ng/μl) into ~100 EG4322 (ttTi5605 II; unc-19(ed3) III) worms cultured at 15 or 20°C. pCFJ356::odr-3p::slo-2::TagRFP (34 ng/μl) was injected along with hsp16.4p::peel-1 (10 ng/μl), eft-3p::mos-1 (50 ng/μl), rab-3p::mCherry (10 ng/μl), myo-2p::mCherry (2.5 ng/μl), and myo-3p::mCherry (5 ng/μl) into ~100 EG6703 (unc-19(ed3) III; cxTi10816 IV) worms cultured at 15 or 20°C. Three injected worms were picked to one plate and cultured at 25°C until starvation (~1 week). The starved worms were heat shocked at 34°C for two hours to activate the negative selection marker PEEL-1, which kills animals carrying extrachromosomal arrays. After recovery at 25°C for four hours, worms that were rescued for the unc-119 phenotype and lacked the three mCherry co-injection markers were cloned out from separate plates. The presence of single copy inserts was verified by PCR.
miniMos integration was done as previously described [55]. Briefly, odr-3p::GFP::egl-19 miniMos construct containing a hygromycin resistance cassette (17.5 ng/μl) was injected along with hsp16.4p::peel-1 (10 ng/μl), eft-3p::mos-1 (50 ng/μl), rab-3p::mCherry (10 ng/μl), myo-2p::mCherry (2.5 ng/μl), and myo-3p::mCherry (10 ng/μl) into ~100 vySi8 II; unc-119(ed3) III and vySi23 II; unc-119(ed3) III worms cultured at 15–20°C. Three injected animals were picked per plate and cultured at 25°C. Three days after injection, hygromysin was added directly onto the plate to a final concentration of 0.25 mg/ml and cultured further until starvation (~1 week). The starved worms were heat shocked at 34°C for two hours to kills animals carrying extrachromosomal arrays. After recovery at 25°C for four hours, worms that survived and lacked the three mCherry co-injection markers were cloned out to determine homozygosity.
Transgenic strains expressing fluorescent markers or fluorescently tagged proteins were mounted onto 2% agarose pads and anesthetized with 5mM sodium azide (Sigma) or 7.5mM levamisole (Sigma). Z-stack images were acquired at room temperature (20–22°C) using Zeiss Axio Imager Z1 or M2 microscopes, each of which is equipped with a motorized focus drive, a Zeiss objective EC Plan-Neofluar 40x/1.30 Oil DIC M27, a Piston GFP bandpass filter set (41025, Chroma Technology), a TRITC filter set (41002c, Chroma Technology), and a Zeiss AxioCam CCD digital camera (MRm for Z1 and 506 mono for M2) driven by the Zeiss imaging software (AxioVision for Z1 and ZEN for M2). For comparison of fluorescence intensity, all animals in each set of experiments were subjected to the same exposure time. Fluorescence intensity was measured with AxioVision or ZEN imaging software or NIH ImageJ image processing software. Since the background autofluorescence varies between some genetic backgrounds (S3A Fig), fluorescence intensity of reporter transgenes was subtracted by background fluorescence intensity to obtain corrected fluorescence intensity. Images shown in Figs 1A, 3A, 3C, 3E, 3G, 5A, 5B, 5C, 6A, 6B and 7B, as well as S2A, S2B, S3A, S3B, S4A and S6A Figs were processed with Adobe Photoshop; the same degree of brightness and contrast adjustment was applied to all images in each set of experiments for comparison of fluorescence intensity (Figs 6A, 6B and 7B, as well as S3A, S3B and S6A Figs).
Genetic mosaic analysis was performed with various unstable extrachromosomal transgenic arrays in either wild type or slo-1(eg142lf); slo-2(ok2214lf) mutants. Three different experiments were performed to determine the sites of slo-1 and slo-2 function in AWC asymmetry. odr-3p::slo-1 was injected into slo-1(lf); slo-2(lf) mutants to determine whether slo-1 acts cell autonomously or nonautonomously to rescue the 2AWCOFF mutant phenotype. A similar experiment was performed using the odr-3p::slo-2 extrachromosomal array. In the third experiment, nsy-5p::slo-1(T1001Igf) was injected into wild-type animals. In all three experiments, the odr-1p::DsRed marker (expressed in AWC and AWB) was included in the injection mix to serve as an indicator for presence or absence of the extrachromosomal transgene in AWC. The AWCON and AWCOFF neurons were determined using expression of a stable integrated str-2p::GFP (AWCON marker) transgene. Transgenic strains were passed for minimum of six generations to allow the transgenes to stabilize before scoring for mosaic animals.
Colocalization was quantified using the Coloc 2 plugin (http://fiji.sc/Coloc_2) in Fiji [56]. Three different algorithms were used: Pearson’s correlation coefficient, Spearman’s rank correlation coefficient, and Li’s ICQ. For each colocalization class, images of at least three animals were used for quantification. Positive values of each coefficient indicate positive correlation, values close to zero indicate no correlation, and negative values indicate anti-correlation. Pearson's correlation coefficient ranges from -1 to +1; Spearman’s rank correlation coefficient ranges from -1 to +1; Li's ICQ value ranges from -0.5 to +0.5
Locomotion analysis was performed on L4 animals in wild type, slo-1(eg142), slo-2(ok2214), and slo-1(eg142); slo2(ok2214) animals. Single animals of each genotype were placed on a bacterial lawn and allowed to make tracks. The worm tracks as well as individual worms were imaged and analyzed in ImageJ [57]. All animals were placed on the same batch of NGM plates seeded with the same batch of HB101 and were imaged on the same day. Wavelength was measured as the distance between wave peaks, and at least 3 wavelengths were measured and averaged per animal. The wavelength was normalized by the body length of the animal. Wave width was measured as the distance from the peak to the trough of the worm wave. At least 3 wave widths were measured and averaged per animal. The wave width was normalized by the body length of the animal.
|
10.1371/journal.pntd.0003498 | Prospective Study in a Porcine Model of Sarcoptes scabiei Indicates the Association of Th2 and Th17 Pathways with the Clinical Severity of Scabies | Understanding of scabies immunopathology has been hampered by the inability to undertake longitudinal studies in humans. Pigs are a useful animal model for scabies, and show clinical and immunologic changes similar to those in humans. Crusted scabies can be readily established in pigs by treatment with the glucocorticoid dexamethasone (Dex).
Prospective study of 24 pigs in four groups: a) Scabies+/Dex+, b) Scabies+/Dex-, c) Scabies-/Dex+ and d) Scabies-/Dex-. Clinical symptoms were monitored. Histological profiling and transcriptional analysis of skin biopsies was undertaken to compare changes in cell infiltrates and representative cytokines. A range of clinical responses to Sarcoptes scabiei were observed in Dex treated and non-immunosuppressed pigs. An association was confirmed between disease severity and transcription of the Th2 cytokines IL-4 and IL-13, and up-regulation of the Th17 cytokines IL-17 and IL-23 in pigs with crusted scabies. Immunohistochemistry revealed marked infiltration of lymphocytes and mast cells, and strong staining for IL-17.
While an allergic Th2 type response to scabies has been previously described, these results suggest that IL-17 related pathways may also contribute to immunopathology of crusted scabies. This may lead to new strategies to protect vulnerable subjects from contracting recurrent crusted scabies.
| Scabies is a neglected tropical skin disease caused by the tiny parasitic mite Sarcoptes scabiei. Scabies is common in developing countries, and scabies outbreaks also occur in institutional settings worldwide. Scabies often underlies secondary bacterial skin infection and resulting complications, and is thus associated with considerable morbidity. Crusted scabies is a an extremely severe and debilitating clinical form of the disease, but host immune responses leading to the development of crusted or ordinary scabies are poorly understood. This is largely due to limited access to clinical samples, and the difficulty in monitoring the progression of infestation in human patients. We have overcome this challenge by using a pig model of scabies infestation, since pigs and humans with scabies display clinical and immunological similarities. In this study, we undertook longitudinal analysis of clinical, histological and molecular immunological changes in pigs experimentally infected with scabies. We confirmed that disease severity was associated with a pronounced allergic, Th2 immune response, as previously reported. In a novel finding, we showed that the Th17 associated cytokines interleukin-17 and interleukin-23 were also associated with the development of crusted scabies. This may lead to new immunotherapeutic strategies to protect vulnerable subjects from contracting recurrent crusted scabies.
| Sarcoptes scabiei infestation is associated with considerable global morbidity [1]. The disease is prevalent in overcrowded living conditions, with the highest disease burdens seen in young children [2]. The link between scabies, secondary bacterial infection and sequleae such as post-streptococcal glomerulonephritis [3] has resulted in efforts to reduce the prevalence of scabies in endemic communities.
Ordinary scabies manifests as a localised or general rash with low mite burden (<20 mites). Crusted (Norwegian) scabies is a less common but debilitating form, with proliferation of mites, hyperkeratosis, and risk of serious secondary infection. Crusted scabies requires aggressive treatment, and recrudescence and reinfestation are common [4].
Factors underlying the development of crusted scabies include iatrogenic immunosuppression and other immunosuppressive conditions such as HIV, HTLV-I and systemic lupus erythematosus [5–7]. The disease has also been described in those with no immune deficit [7–9], and reasons for crusted scabies development in this cohort are unknown. Limited humoral and cellular studies conducted to date suggest that crusted scabies is associated with a non-protective allergic T helper (Th) 2 response [10–12], but these are confounded by difficulties in assessing clinical severity [13], and the fact that patients present at an advanced stage of infestation.
Prospective studies are necessary to gain meaningful insights into immune responses driving crusted scabies. Scabies is associated with delayed onset of symptoms (4–6 weeks) in primary infestation, and several studies show that S. scabiei is capable of down regulating cytokine expression, likely suppressing early immune responses to allow mites to establish [14–18]. However, these studies were mostly in-vitro, utilising mite extracts and cultured cells or skin equivalents. We have recently developed a porcine model to investigate aspects of scabies immunology [19–21]. Pigs are a natural host of S. scabiei var suis, developing similar clinical manifestations to humans, including crusted and ordinary scabies. In this study we conducted transcriptional analysis of representative Th1, Th2, and Th17 pathway cytokines in the skin of infected pigs at several time points post infestation, and assessed skin biopsies with different clinical phenotypes by immunohistochemistry for inflammatory markers.
Animal ethics approval was obtained from the QIMR Berghofer Medical Research Institute (Approval 1266) and the Queensland Department of Agriculture, Forestry and Fisheries (Approval SA 2009/07/294). Animals were handled in accordance with good animal practice as defined by the Australian code of practice for the care and use of animals for scientific purposes and the Australian National Health and Medical Research Council’s Animal Code of Practice.
Details regarding trial design have been described elsewhere [20]. The study involved 24 female piglets in four treatment groups (n = 6 per group). Group A: treated daily with 0.25mg/kg oral Dexamethasone (Dex) and ears infested with approximately 2,000 S. scabiei var suis mites. Group B: infested with approximately 2,000 mites. Group C: treated daily with 0.25mg/kg Dex (Dex only control). Group D: No Dex or mite infestation (negative control). While the infested and non-infested groups were kept isolated from each other, the allocation of individual pigs to pens was random, meaning that Dex and non-Dex pigs were housed together. Skin lesions were scored weekly on a 1–8 scale (1 = minimal change, >4 = development of crusts, 8 = extensive crusting. Skin scrapings were collected from a 2cm2 ear region of each pig fortnightly to approximate mite burden, as described previously [19]. Mite burden was graded as follows: – = no mites, + <20 mites/scrape, ++ = >20–100 mites/scrape, +++ = >100 mites/scrape. Two adjacent 3mm skin punch biopsies were collected from the ears of all pigs at week 0, 4, 8, and 12 post-infestation. At this size biopsies healed rapidly with minimal scarring. For infested pigs, biopsies were taken directly from lesional areas where scabies infestation was apparent. Biopsies were full skin thickness, including hyperkeratotic areas (if apparent), epidermis, dermis, and underlying ear cartilage. One biopsy was stored in RNA Later reagent (Life Technologies) and kept at-80°C. The second was collected into 10% neutral buffered formalin, fixed for 24 hours, transferred to 70% ethanol, and kept at 4°C.
Ten pigs were selected for analysis- two from groups A, C and D, and four from Group B. These represented different clinical phenotypes—crusted scabies, ordinary scabies, and non-infested, based on clinical presentation and mite burden. The four pigs in Group B included two pigs with crusted scabies (designated Group B+ in subsequent results). Serial sections (4–7μM) were cut from paraffin embedded biopsies, dewaxed and stained with hemotoxylin and eosin. Slides were examined for cellular, structural and vascular changes (Table 1) and each parameter allocated a score of 0–5 where 0 = minimal change and 5 = extensive change. The inspecting pathologist was blinded to the allocated group.
Immunohistochemistry was undertaken to characterise the cellular infiltrate, as well as for IL17 cytokine staining. T cell numbers were assessed by staining with anti-CD3 antibody. Dewaxed sections were incubated with high pH antigen retrieval solution (Dako pH 9.0) and blocked with purified casein (Medical background sniper, Biocare). Rabbit anti-human CD3 antibody (Biocare), previously established to cross-react with pig tissue, was diluted 1:275 and incubated overnight at room temperature. Sections were washed 3 times for 5 minutes in phosphate buffered saline. Anti-rabbit HRP secondary antibody (Vector labs) was applied for 30 minutes, washed as above, and the HRP substrate Novared (Vector labs) applied and developed for 5 minutes. For mast cell visualisation, sections were stained with toluidine blue. Analysis was performed on CD3 and toluidine blue stained sections by scanning slides with an Aperio XT scanner. Positively labelled cells were counted in 10 fields at 20 X magnification. Cell concentration (cell/mm2) was calculated for each field by dividing the count total by the area of the field (0.234mm2) and the value for each of the 10 fields averaged.
For IL-17 detection, dewaxed sections were blocked for endogenous peroxidases with 1.0% H2O2, 0.1% sodium azide for 10 minutes. Sections were incubated in citrate pH 6.0 antigen retrieval buffer at 97°C, and blocked in 4% skim milk powder, followed by purified casein (Medical background sniper, Biocare) plus 10% normal goat serum and 1.0% BSA. Polyclonal rabbit anti-human IL-17 (Abcam, 1mg/mL), diluted 1:100, was applied for 2 hours at room temperature. This antibody was derived from a synthetic 19 amino acid peptide of human IL-17A, and based on sequence conservation was predicted to be cross reactive with its pig homologue. Sections were washed 3 times for 5 minutes in tris buffered saline. Anti-rabbit-HRP secondary antibody (Mach2, Biocare) was applied for 45 minutes, sections washed as above and developed with 3,3'-diaminobenzidine (DAB) with H2O2 as substrate for 5 to 10 minutes. Staining intensity was assessed qualitatively on a scale of 0–4 by a dermatologist blinded to the allocated group.
Biopsies stored in RNA Later were thawed and homogenised in 600μL TRIzol reagent (Life Technologies) using the Tissue Lyser II homogeniser (Qiagen). Phase separation with TRIzol was undertaken according to the manufacturers’ protocol. The aqueous phase was column purified as per the manufacturers’ protocol (PureLink RNA mini-kit, Life Technologies), including DNAse digestion. RNA was eluted in RNAse free dH20 and stored at-80°C. RNA quantity and integrity was assessed using the Nanodrop ND2000 spectrophotometer (Nanodrop Technologies) and Agilent Bioanalyzer RNA 6000 nano-kit (Agilent Technologies). One μg of purified total RNA was reverse transcribed to cDNA using the QuantiTect reverse transcription kit (Qiagen). The cDNA was diluted 1:4 in dH20 and stored at-20°C.
Primers were designed using Primer3 software (http://frodo.wi.mit.edu/) (Table 2). Hypoxanthine phosphoribosyl transferase 1 (HPRT1) was selected as a reference gene, as this gene is proposed to be stable under different environmental conditions [22]. Gradient PCR to test optimal annealing temperatures was performed using control skin cDNA. PCR products were purified (Roche), cloned (pGEM-T, Promega) and sequenced (Big Dye 3.1, Applied Biosystems). Sequence identity was confirmed using BLASTx (http://blast.ncbi.nlm.nih.gov/).
To assess amplification efficiency, plasmids containing the gene of interest were linearised and serially diluted. qPCR was done using the QuantiTect SYBR green PCR kit (Qiagen). Reactions contained 1 X SYBR green master mix, 0.4μM primers, 1μL diluted plasmid DNA and dH20 to total volume of 10μL. Reactions were cycled in the Rotor Gene 6000 real-time cycler (Qiagen). Cycling conditions were: initial denaturation 95°C, 15 min, followed by 40 cycles of 94°C, 15 s; 56°C, 30 s; 72°C, 30 s; with data acquisition at 76°C, 20 s. Standard curves, melting temperature and efficiency calculations were produced using the Rotor Gene software.
qPCR was run on the cDNA samples for the gene of interest in parallel with HPRT1, allowing for normalisation. A no-RT control containing RNA as template was used to confirm that co-amplification of genomic DNA was not occurring. Each PCR also included a no template control. Reactions were performed in duplicate. Individual reaction mixtures were as above, except that 2μL cDNA was used as template.
To measure transcriptional differences between treatment groups relative to the untreated, uninfested control group (Group D), the ΔΔCt formula was used, corrected for PCR efficiency [23]. Significance of differences between groups was assessed using unpaired T tests at each time-point using GraphPad Prism version 5.0 (GraphPad Software, Inc.).
Details of the clinical phenotypes observed in the trial are presented elsewhere [20]. All pigs in Group A developed crusted mange (skin score >4) from weeks 8–24 post-infestation (Fig 1A). Two pigs in group B also developed crusted mange in the absence of Dex immunosuppression (designated as group B+ in subsequent results). The remaining pigs in group B developed an acute reaction, with lesion severity 1–4, peaking at weeks 8–12 before declining (Fig 1B). Mite counts were associated with lesion scores, with positive scrapings obtained from 4/6 pigs in group A, and 3/6 pigs in group B in week 4. From week 8, differences between the groups became more apparent, with most pigs in group A having heavy mite infestations. In group B, 2 pigs developed heavy infestations (Group B+) while 4 pigs had low-moderate infestations. Pigs in the non-infested groups did not develop skin lesions nor have detectable mites at any time (Table 3).
General histopathology. Major epidermal changes characteristic of severe crusted S. scabiei infestation included acanthosis, rete peg hypertrophy and para-hyperkeratosis (Fig 2A). Other changes included apoptosis / necrosis /erosion, microabscesses and transudation (Table 4). At the dermal level pathology included edema, vasculitis, and infiltrates of granulocytes and monocytes. The level of pathology was associated with clinical severity, with the greatest changes observed in Groups A & B+. Group A had fewer histological changes at 4 weeks, but more dramatic change at 8 and 12 weeks. Histological changes in group B pigs that clinically had self-limiting infestation peaked at week 4 and were reduced at week 8 and 12. Minor changes were observed in one pig in group C (thickening, ortho-hyperkeratosis, minor mononuclear infiltrate). No histological changes were apparent in the group D, Table 4).
CD3 immunolabeling. Pigs in Groups A & B had increased T cell infiltrates relative to non-infested pigs in groups C & D as ascertained by CD3 immunolabeling (Fig 2C, 3A). Positive cells aggregated in a perivascular pattern in the papillary and reticular dermis and in the stratum basale and stratum spinosum of the epidermis. This increase was most marked in pigs in Group A and B+. Maximal infiltration was observed in pigs in group A at weeks 8 and 12, in Group B+ at weeks 4 and 8, and in Group B at week 4 (Fig 3A).
Mast cell staining. Mast cell numbers in crusted pigs were increased relative to non-infested controls at weeks 8 and 12 (Fig 2D, 3B). Positively stained cells were perivascular in the papillary and reticular dermis. No stained cells were present in the epidermis. Mast cell numbers in pigs with ordinary scabies did not change dramatically over the course of infestation, but were slightly elevated relative to other groups at week 4.
IL-17 immunolabeling. IL-17 staining l was moderate to intense in pigs with crusted scabies at weeks 8 and 12 (Group A, B+), while low to moderate in pigs with ordinary scabies (Group B) and minimal in non-infested pigs (Table 4, Fig 4). Where positive, IL-17 labeling was widespread and generally dispersed and located in dermal cells, stratum basale and stratum spinosum, as well as within vessels. There was also a strong signal in keratinocytes (Fig 4). No signal was observed with the isotype control antibody (Fig 4D).
Scabies was associated with significant changes to several cytokines measured, including transforming growth factor β (TGF-β), interleukin (IL)-2, IL-4, IL-13, IL-17 and IL-23 (Fig 5). No significant changes to interferon γ (IFNγ), IL-5, IL-6 or IL-10 were detected. In the following section, while fold changes in transcription are noted, the Group B+ P-values are not reported due to the low numbers of pigs in this group limiting meaningful statistical interpretation. When comparing pigs with crusted scabies (Groups A and B+) to those with ordinary scabies (Group B) an increased magnitude of IL-13, IL-17 and IL-23 responses was observed from 4 weeks. IL-13 was increased both in Group A and B+ by 13-fold at 4 weeks (Group A p = 0.05), in Group A (15-fold, p = 0.002) and B+ (4-fold) at 8 weeks, and in Group A (16-fold, p = 0.009) and B+ (8-fold) at 12 weeks. By contrast in Group B pigs with ordinary scabies elevation of IL-13 was only observed at week 12 (8-fold, p = 0.03). We saw upregulation in the Th17 cytokines, IL-17 and IL-23 only in pigs with crusted scabies. IL-17 was significantly upregulated at all time points, most strongly at week 8 (Group A 41-fold, p = 0.009, Group B+ 37-fold). Upregulation of IL-23 was observed at all time points, with a 30-fold increase observed in Group A pigs from week 4 (p = 0.03).
Transcription of IL-4 was increased in all infected pigs at all time points, with the exception of group B+ at 4 weeks. The greatest elevation of IL-4 was in Group A pigs at 8 weeks (24-fold, p = 0.002). IL-2 levels also increased in all infected pigs from week 4, but the change only became significant after week 8. Similarlary TGB-β was modestly but signficantly upregulated in infected pigs at all time points, with the exception of Group A at week 4.
There were no significant differences between non-infected Dex +ve and Dex—ve pigs (Groups C & D), suggesting that the Dex had little impact on baseline levels of these cytokines in the skin.
Comparison of immune responses in scabies been confounded by the limited availability of clinical samples and standardisation problems related to differences in disease presentation and the existence of co-morbidities [24]. Animal models offer the ability to correlate clinical phenotype with immune parameters and to report the temporal development of immune responses. We observed that phenotypic differences between crusted and ordinary scabies in a porcine model were associated with differences in both the timing and magnitude of cytokine responses and histological changes.
Scabies became clinically apparent in infested pigs from week 4, with mite numbers correlated with the appearance of clinical lesions. As skin scrapings can have poor diagnostic sensitivity due to low mite numbers, and are difficult to perform in large herds or community studies, clinical appearance is more useful as a proxy measure of infestation level [25]. For example, while several pigs in Group B had negative skin scrapings at weeks 4 and 8, clinical scores indicated they were still infested. While most infested pigs had similar mite counts at week 4, pigs that developed crusted scabies had substantially increased mite numbers from week 8, while those with ordinary scabies maintained low or moderate numbers of mites. Notably, two pigs from group B developed crusted scabies in the absence of immunosuppression. While acknowledging the small number and consequent limited interpretation of results for Group B+, we elected to compare these pigs as a separate “subgroup”, as the development of crusted scabies in the absence of Dex immunosuppression is of interest. These clinical observations reflect what is well documented in the literature- while the majority of pigs with sarcoptic mange develop an ‘acute’ manifestation with clinical peak of around 8 weeks before a decline in skin lesions and mite numbers, indicative of a self limiting infestation, some pigs develop chronic hyperkeratotic mange akin to crusted scabies in humans. This reinforces the value of the porcine model to explore protective versus pathologic immune responses in scabies, and further studies by our group have focused on the further study of different clinical phenotypes in pigs not receiving Dex treatment [21].
As histological analysis of scabies lesions has been reported in the literature previously for both pigs and humans, we did not intend to undertake comprehensive histological comparisons in this study, but rather obtain a representative “snapshot” to link our clinical and molecular observations in different clinical phenotypes of scabies. Being mindful of the limited numbers of pigs examined, histopathology generally mirrored clinical observations. An exception was that pigs in Group A had delayed inflammatory responses at week 4 relative to Group B. As the pigs in Group B+ with crusted scabies also had inflammatory changes at week 4, these differences may be more attributable to Dex supressing early inflammatory responses rather than differences between crusted and ordinary scabies. From week 8 pathologic changes between the clinical phenotypes were more apparent, which was also reflected in CD3+ T cell numbers. Increased T lymphocytes in scabies lesions have also been reported in humans [26] and other animals [26–29]. Ongoing work has shown that the CD3+ T cell infiltrate in pigs with crusted scabies is comprised largely of γδ T cells and CD8+ T cells [21]. Although γδ T cells have not yet been examined in human scabies, CD8+ tropism has been observed in crusted scabies [30], while increased CD4+ cell infiltrates were associated with protective immunity in canine mange [28].
Crusted scabies was associated with increased mast cell numbers, most notably at week 12 post infestation. Mast cell numbers remained steady throughout the study in pigs with ordinary scabies. The presence of mast cells is consistent with previous findings [27,29,31,32]. The presence of mast cells, often with accompanying eosinophilia, is reflective of the allergic and immediate hypersensitivity component of the scabies immune response, particularly upon secondary exposure [31]. The role of mast cells and related high IgE levels in protective versus pathologic responses to scabies is yet to be resolved [33].
As well as general T cell proliferation and inflammatory markers such as IL-2 and TGF-β, crusted scabies was associated with a pronounced Th2 response. This was most evident with IL-13, and to a lesser extent, IL-4, whereas IL-5 was not signifigantly elevated. These findings are in accordance with cross-sectional studies on human patients [10], where peripheral blood mononuclear cells (PBMCs) from crusted scabies patients secreted more IL-5 and IL-13, and reduced IFNγ in response to stimulation with S. scabiei antigens [10]. While we did not see any transcriptional changes in IFNγ in the present study, this may be related to the timing of infestation, local versus peripheral responses, or primary versus secondary infestation. For example Lalli et al [34] found that while primary exposure to S. scabiei in mice was associated with an IL-4 response, secondary exposure following immunization was IFNγ oriented. Other studies by our group have shown increased CD4+ IFNγ+ T cells at one week post infestation in PBMCs from mange infested pigs [21].
This is the first study to measure temporal changes in cytokine levels in scabies infested skin. In studies undertaken on clinical patients, little information was available regarding duration of current infection and a key question was if elevated Th2 responses precede, or are simply a consequence of, the extreme antigen burden in crusted scabies [35]. Our studies show that Th2 elevation, particulary of IL-13, occured prior to the development of high mite burdens and before major clinical or histological differences between groups became evident.
The observation of increased IL-17 in the skin by immunohistochemistry and qPCR supports our recent findings of increased CD3+ IL-17+ cells in crusted scabies as determined by intracellular cytokine staining [21]. In this study, increased IL-17 was observed at week 15 post-infestation. Here, we show that transcriptional increases of IL-17 begin from as early as week 4 post-infestation. Again, this was prior to the development of strong clinical or inflammatory changes in the skin, suggesting that the IL-17 increase is associated with a dysregulated response rather than just a consequence of a changed inflammatory skin milieu.
IL-17 is a proinflammatory cytokine implicated with a number of allergic and inflammatory diseases. Traditionally associated with CD4+ T cells (Th17), IL-17 is also secreted by other innate and adaptive immune cells in the skin, including CD8+ T cells, γδ T cells, and mast cells [36]. While γδ cells are likely a major source of IL-17 in crusted scabies [21], the contribution of CD8+ and mast cells to local IL-17 production is still to be investigated. Regardless of the cell type, it is accepted that functional maturation and IL-17 secretion is promoted by increases in IL-23, secreted by dendritic cells, macrophages and keratinocytes, in the presence of TGB-β and IL-6 [36]. These are all present in scabies infested skin, supporting an IL-17 environment. Importantly, IL-23 was only increased in crusted scabies, potentially promoting the subsequent high levels of IL-17. Using human skin equivalents, Morgan and colleagues [15] demonstrated that S. scabiei promotes up-regulation of IL-23 from 48 hours post infestation.
It is suggested that increases in IL-17 could be the result of a dysregulated regulatory T (Treg)/Th17 balance, or due to a deficit in IL-10 [21]. While mite extracts are capable of inducing IL-10 secretion in human PBMCs [37], reduced IL-10 was observed in PBMCs isolated from crusted scabies patients relative to ordinary scabies [10]. In the current study there were no observable differences in IL-10 between crusted and ordinary scabies. A limitation was that other markers of Treg function were not examined. A role for IL-10 regulation of IL-17 is supported by studies in leishmaniasis, where blockade of IL-10 resulted in increased IL-17 and exacerbation of skin pathology [38]. Increased IL-4 is also reported to suppress IL-10, exacerbating syptoms of Th2 mediated atopic dermatitis [39].
An important consideration is the potential impact of Dex on the immune parameters investigated. It is accepted that the effects of Dex are pleotropic, with dose, timing and experimental system appearing to play a role. These preliminary findings need to be supported by larger studies with non-immunosupressed pigs with crusted scabies. While the utilisation of Dex to induce the clinical phenotype of crusted scabies somewhat confounds interpretation of the immunologic parameters measured in this study, the data obtained is still informative. Firstly, comparing immune responses in the crusted scabies phenotype in the presence and absnece of immunosupression assists in refining a common immunopathology, regardless of causation. Secondly, crusted scabies in humans frequently arises from corticosteroid use, so an understanding of immune responses and potential implications for immunotherapy under these conditions are of interest. Thirdly, the effects of Dex on specific aspects of the immune system remain poorly defined in both humans and animals, so this study adds value at a general level. In our study, pigs were maintained on a relatively low dose of Dex (0.25mg/kg), with others reporting that porcine immunne funtion was resistant to higher doses (2mg/kg) [40]. Despite the low dose, there are several factors whereby Dex may be conducive to the development of crusted scabies. Dex may promote Th2 bias, with increased IL-4 and decreased IFNγ [41–43]. Other studies report Dex inhibition of Th2 responses [44,45] but again, these differences may be in part explained by the concentration used, with low doses stimulating, and high doses inhibiting IL-4 [46]. Of particular relevance, populations of double positive Th2/Th17 cells secreting IL-4 and IL-17 have been identified in severe asthma, and these cells were insensitive to Dex [47]. Dex treatment may decrease FoxP3+ CD4+ T cells [48], possibly causing further amplification of Th2 and Th17 pathways and promoting the development of crusted scabies.
This study contributes to the limited knowledge regarding the immunopathogenesis of crusted scabies, with a theme for involvement of Th2 and Th17 related cytokines now emerging, although numbers of pigs and human patients studied remains small. It is now important to gain more detailed insights into pathways of immune dysregulation in crusted scabies, particularly the contribution of regulatory T cells. Longitudinal studies are also needed earlier in infestation, prior to the development of clinical symptoms. Finally, studies where pigs are treated, then reinfested, would be of value to compare primary versus secondary immune responses to S. scabiei.
|
10.1371/journal.pgen.1007891 | Single-molecule dynamics and genome-wide transcriptomics reveal that NF-kB (p65)-DNA binding times can be decoupled from transcriptional activation | Transcription factors (TFs) regulate gene expression in both prokaryotes and eukaryotes by recognizing and binding to specific DNA promoter sequences. In higher eukaryotes, it remains unclear how the duration of TF binding to DNA relates to downstream transcriptional output. Here, we address this question for the transcriptional activator NF-κB (p65), by live-cell single molecule imaging of TF-DNA binding kinetics and genome-wide quantification of p65-mediated transcription. We used mutants of p65, perturbing either the DNA binding domain (DBD) or the protein-protein transactivation domain (TAD). We found that p65-DNA binding time was predominantly determined by its DBD and directly correlated with its transcriptional output as long as the TAD is intact. Surprisingly, mutation or deletion of the TAD did not modify p65-DNA binding stability, suggesting that the p65 TAD generally contributes neither to the assembly of an “enhanceosome,” nor to the active removal of p65 from putative specific binding sites. However, TAD removal did reduce p65-mediated transcriptional activation, indicating that protein-protein interactions act to translate the long-lived p65-DNA binding into productive transcription.
| To control the rate of transcription of genes, both eukaryotes and prokaryotes express specialized proteins, transcription factors (TF), that bind promoter sequences to mark them for the transcriptional machinery including DNA polymerase II. TFs are often multi-subunit proteins containing a DNA-binding domain (DBD) as well as a protein-protein interaction interface. It was suggested that the duration of a TF-DNA binding event 1) depends on these two subunits and 2) dictates the outcome, i.e. the amount of mRNA produced from an activated gene. We set out to investigate these hypotheses using the transcriptional activator NF-κB (p65) as well as mutants affecting one of its functional subunits. Using a combination of live-cell microscopy and RNA sequencing, we show that p65 DNA-binding time indeed correlates with the transcriptional output, but that this relation depends on, and hence can be uncoupled by altering, the protein-protein interaction capacity. Our results suggest that, while p65 DNA binding times are dominated by the DBD, transcriptional output relies upon functional protein-protein interaction subunit.
| Transcription factors (TFs) are fundamental regulatory components of transcription in both prokaryotes and eukaryotes, which can activate or repress the expression of specific genes. The NF-κB family of TFs, universal among nearly all animal cell types, is involved in many signaling pathways and when dysregulated can contribute to several pathologies, including cancer and inflammatory diseases [1]. This is exemplified by the RELA (v-rel reticuloendotheliosis viral oncogene homolog A), or p65 TF, which is implicated in regulating the activation of ~150 genes involved in wide-ranging functions from immune response to metabolism [1]. In its most prevalent form, p65 forms a stable heterodimer with p50 in the cytoplasm [2] (Fig 1A). Upon stimulation, the activated heterodimer translocates into the nucleus [3]. The heterodimer interacts with target DNA regulatory elements through a conserved Rel homology region (RHR) [4–6]. Following DNA binding, p65-mediated transcriptional activation is controlled by two trans-activation domains (TADs), TAD1 and TAD2 [7]. Co-regulators of transcription are recruited at the promoters of target genes via protein-protein interactions mediated by TAD1 and TAD2, eventually leading to the recruitment of RNA polymerase II (RNA pol-II) and subsequent activation of gene expression [8]. Deletion of one or both TADs has been shown to heavily impair p65-dependent transcriptional activation, suggesting a dominant-negative effect of such truncation mutants [7].
There is increasing evidence from biochemistry and live-cell single molecule imaging that in general the duration of binding events of TFs to responsive elements (RE) correlates with transcriptional activity [9–11]. However, as in the case of p65, it remains largely unknown whether protein-protein interactions mediated by p65 TADs can stabilize DNA binding, and thus lead to higher transcriptional activity. Such a stabilization would be expected according to the model of the interferon-β1 enhanceosome (reviewed in [12]). The cooperative binding of eight transcription factors, including NFκB, to the interferon-beta enhancer sequence was shown to form a composite surface for the recognition of the entire enhancer region [13]. Biochemically, the enhanceosome complex was shown to be very stable in vitro [14], a result that was challenged by a subsequent study in which the duration of p65-DNA binding was measured at genetically engineered arrays in cells and found to be very transient [15], an observation incompatible with the formation of stable complexes. Alternatively, perhaps protein-protein interactions mediated by the p65 TADs could destabilize its binding to DNA, and TAD-mediated transactivation would instead be responsible for actively displacing p65 from chromatin. As this live-cell study was performed on artificial arrays of p65 binding sites, the role of TAD-mediated protein-protein interactions for p65-binding stability in the genomic context still remains unexplored.
Here, we combined single-molecule live-cell imaging and genome-wide transcriptomics of wild-type p65 (p65-WT), TAD truncation mutants and p65 DNA-binding affinity mutants to elucidate the role of these domains in the stability of p65 binding and on downstream transcriptional activity. We established point mutants to the DNA-binding domain to modulate p65-DNA binding affinity. We found that the lifetime of binding events of p65 mutants to chromatin in living cells correlated with their reported in vitro binding affinities and genome-wide transcriptional activity. We next examined the effects of TAD deletion mutants. We found that these mutants had DNA-binding kinetics comparable to p65-WT. However, whole transcriptome profiling revealed that TAD truncated forms of p65 did have impaired transactivation capability, suggesting that TAD-mediated protein-protein interactions serve the role of translating longer-lived p65-DNA binding into transcriptionally productive events.
We carried out a p65 DNA-binding kinetics and genome-wide transcriptomic study using a carboxy-terminal Halo-Tag [16] fusion construct of the human p65 (p65-Halo) (Fig 1B). To fluorescently label p65, HeLa cells were transiently transfected with p65-Halo and incubated with Halo-JF549 [17] (Fig 1B). In a large majority of transfected cells (~90%), the labeled p65-Halo was enriched in the cytosol and excluded from the nucleus (Fig 1B, panel “-TNFα”). After 30 minutes of stimulation with TNFα, p65-Halo translocated from the cytosol into the nucleus in ~73% of the cells (Fig 1B, panel “+TNFα”), showing that it was responsive to TNFα treatment.
We also tested the p65-Halo fusion protein for its ability to transactivate two well-known p65 target genes, NFKBIA and Ccl2, either in the presence or absence of TNFα stimulation (S1 Fig). Ectopically expressed p65-Halo upregulated the expression of both genes above their endogenous levels in non-stimulated cells (p < 0.05). Furthermore, upon TNFα stimulation, NFKBIA showed a further significantly (p < 0.05) increased level of expression, whereas Ccl2 upregulation was not significantly different (p > 0.05) from non-stimulated cells. This is likely due to two synergistic factors, that is, the already high expression levels of Ccl2 in the presence of overexpressed p65, and the slower activation rate of Ccl2 as compared to NFKBIA (attributed to the fact that the former requires chromatin remodeling whereas the latter does not [18,19]).
We further verified the interaction of p65-Halo with its consensus DNA sequence (S2 Fig) by using an electrophoretic mobility shift assay (EMSA). JF549-labeled p65-Halo was purified (Materials and Methods) and incubated with Atto647N-labeled consensus oligonucleotide before electrophoretic separation under non-denaturing conditions (Materials and Methods). Increasing concentrations of p65-Halo enhanced the shifted fraction of labeled oligonucleotide, confirming the ability of the fusion protein to bind in vitro to its specific consensus sequence (S2 Fig).
We performed 2D single-molecule tracking (SMT) of individual, JF549-labeled p65-Halo molecules in the nucleus of live HeLa cells after stimulation with TNFα. We excited the sample with a highly inclined and laminated laser illumination (HILO) to minimize background fluorescence from out-of-focus p65 molecules [20] (Materials and Methods). Further, using stroboscopic laser excitation (tint = 5 ms; tgap = 95 ms; power ~1 kW cm−2), we could minimize photobleaching, allowing us to record seconds-long trajectories from both static and mobile p65 molecules (Fig 2A). To selectively identify p65 molecules bound to chromatin, we used the histone subunit H2B fused to Halo tag as an “immobile” control to define an upper threshold for the displacement rmax between two consecutive frames. We found that ~99% of H2B displacements were below rmax = 435 nm (Fig 2A and S3 Fig). Each p65 frame-to-frame displacement satisfying r < rmax was further required to last at least 10 frames to minimize the probability that slowly diffusing molecules would affect the calculated tb [21]. The binding time tb of each DNA-bound p65 single molecule was then directly measured as the number of frames the fluorescence stayed “on” until disappearance. We found that tb of DNA-bound p65 molecules could not be described by a single-exponential decay model (S4 Fig). To find the model that could best describe the survival probability distributions for p65-WT, we compared single and double exponential fits using an F-test (see Methods). We found that a bi-exponential decay model was most likely correct and in good agreement with our photobleaching-corrected data (Fig 2C, S4 Fig, S5 Fig), with lifetimes of tbfast=0.53±0.01s and tbslow=4.13±0.11s, for the short- and long-lived populations of p65-Halo wild-type molecules (p65-WT), respectively (Fig 2D). Here, short- and long-lived populations corresponded to ~95.7% and ~4.3% of p65-WT DNA-bound molecules (Fig 2D).
To modulate the affinity of p65 for DNA, we performed single-point mutagenesis within the p65 DNA-binding domain (DBD) (Fig 2B) [22]. Of the prototypical NFkB heterodimer, both subunits p65 and p50 have a Rel homology domain (RHD) that was shown to be responsible for DNA binding [23]. However, only the modification (acetylation) of p65 was shown to have a regulatory effect on DNA binding and gene activation [22]. We have therefore focused our mutational analysis only on p65. We generated two DNA binding affinity mutants, p65-KKAA-Halo and p65-KKRR-Halo, corresponding to lower (relative KD = 0.1) or higher (relative KD = 3) in vitro binding affinities as compared to p65-WT [22]. We expressed each mutant in HeLa cells, controlling for expression levels (S6 Fig.), and initially performed fluorescence recovery after photobleaching (FRAP, S7 Fig). As expected, the lower DNA-binding affinity mutant p65-KKAA showed faster recovery as compared to p65-WT (t1/2KKAA=0.56±0.04s;t1/2WT=1.3±0.1s) while p65-KKRR displayed comparable recovery dynamics (t1/2KKRR=1.4±0.1s) to p65-WT.
Following SMT, we again compared exponential functions of different degree to find the model that could best describe the survival probability distributions. Although all 1-CDF distributions were best fit with a bi-exponential model, for p65-KKAA we observed the lowest F significance together with the highest error associated to the long binding time (S5 Fig). Consistent with the estimated recovery rates from FRAP, p65-KKAA showed a dominant (99%) short-lived binding time (tbfast=0.43±0.01s) with only a minor (1%) long-lived binding time (tbslow=9.39±2.49s). The p65-KKRR variant displayed two components with a high fraction of long-lived binding events (~22.5%) associated with longer binding times (tbslow=12.8±1.2s). To provide quantitative estimates of the fraction of p65 molecules involved in binding, we repeated the SMT at faster frame rates (tint = 5 ms; tgap = 15 ms) and analyzed the resulting tracks by fitting the distribution of displacements between consecutive frames (Δt = 20 ms) using a three-component diffusive model [24] (Eq 2). The fraction of p65 molecules corresponding to the slowest diffusing component matched the diffusivity coefficient of the histone subunit H2B (~0.04 μm2 s-1) and was identified as the bound fraction (BF) of p65 molecules (S8 Fig.). We noted that p65-KKRR displayed a significantly higher BF (BFKKRR~30%) than p65-KKAA (BFKKAA~4%) and p65-WT (BFwt−p65~21%; S5 Fig) which well explains the higher immobile fractions in FRAP recovery curves (S7 Fig.).
The transcriptional activation potential of p65-Halo mutants was estimated by measuring transcriptome-wide gene expression levels (RNA-seq; see Methods). RNA-seq analysis allowed us to identify differentially expressed genes by comparing stimulated (+TNFα) and non-stimulated cells (-TNFα). Using a false-discovery rate (FDR) lower than 0.1 (Materials and Methods), a total of 1080 genes were scored as differentially expressed (Fig 3A). Of these, we selected only genes directly bound by p65 on the basis of deposited ChIPseq data (ENCODE database). Among the remaining 215 genes, we identified 45 well-characterized p65 targets [25], including FAS, IL23A and TRAF1. The relative fold-changes (FC) of expression of the 215 p65-target genes were then computed for each generated mutant by normalizing against the gene expression levels observed in non-transfected cells (NT) (Fig 3A). This analysis was complemented by determining the z-score of gene expression levels and visualized using a heat-map (Fig 3B). Results obtained with both approaches identified p65-KKAA as a loss-of-function and p65-KKRR as a gain-of-function mutant (Fig 3B–3D).
To investigate the role of protein-protein interactions on the p65 DNA binding time and downstream transcriptional activation, we generated two additional truncation mutants, lacking one (p65-ΔTA1) or both TADs (p65-ΔTAD; Fig 4A). We performed SMT and RNA-seq on these TAD mutants to retrieve binding kinetics (Fig 4B and 4C) and genome-wide transcriptional activation potentials (Fig 4D). An additional truncation construct of p65 lacking the entire DNA-binding domain (p65-ΔDNA) was used as a negative control (Fig 4A). We again evaluated the most likely fitting function and found that all constructs (p65-ΔTA1, p65-ΔTAD and p65-ΔDNA) were best described with a bi-exponential model (S5 Fig). Notably, although a double exponential model was indicated for p65-ΔDNA, similar to p65-KKAA, the double exponential model led to a dominant (~99%) fraction with a short binding time (tbfast=0.46±0.01s) as well as a minor (1%) fraction with a long binding time and a large associated error (tbslow=13.59±2.8s). Both TAD mutants showed fractions of long-binding events (~4%) comparable to those observed with p65-WT (Fig 2C and 2D). Moreover, the durations of such binding events were similar to those found for p65-WT, ~3–6 s (Fig 4C). However, both p65-ΔTA1 and p65-ΔTAD scored as loss-of-function mutants (that is, z-score ~ 0 and median log2FC < 0; Fig 3B and 3C) as their overexpression in HeLa cells led to significantly lower levels of overall target gene transcription. Thus, despite comparable tb, truncation of TAD domains significantly impaired transcriptional activation (Fig 3B and 3C, Fig 4D). Nevertheless, transcriptional activation potentials of both p65 deletion mutants scored higher than the control construct p65-ΔDNA, indicating that a residual transcriptional activation potential was retained in TAD truncation mutants. To test the spontaneous transactivation of our p65 variants, we performed qPCR experiments for a subset of seven target genes testing their relative gene induction compared to non-transfected control cells. We tested the genes NFKBIA and CCL2 (see S1 Fig) as well as five additional well-characterized p65 target genes from our RNA-seq hit list (S9 Fig). In the absence of TNFα, we observe very similar transcript quantities as compared to non-transfected cells (S9A Fig), indicating low spontaneous gene induction as a result of over-expression. Only for EBI3 and TRAF1 did we observe a stronger spontaneous activity in cells transfected with p65-WT and p65-KKRR. However, upon TNFα stimulation, the transcription of all seven genes was enhanced in cells transfected with p65-wt (S9B Fig).
The revealed relationship between transcriptional activation potential, in vitro p65-DNA affinity, and the duration of long-lasting DNA binding events, tbslow, is summarized in Fig 5A. Note that the p65-ΔDNA mutant included in this plot was assigned an arbitrarily low relative KD ~ 10−5. Also, due to their small fraction and the high fitting uncertainty of the long-lived DNA binding time, for p65-KKAA and p65-ΔDNA we plotted tbfast instead. We observed that the median log2FC ratio of p65 DNA-binding affinity mutants correlated with tbslow. This linear dependence is recapitulated when considering correlations between the in vitro binding affinity (relative KD), and tbslow. Notably, the p65-KKAA mutant appeared similar to p65-ΔDNA both in terms of median log2FC ratio and tbslow.
Considering the impairment of protein-protein interactions induced in p65-ΔTA1 and p65-ΔTAD mutants, we observed that both variants displayed tbslow values surprisingly similar to p65-WT (~4 s). Both truncations provoked significantly lower median log2FC ratios as compared to p65-WT, thus scoring much lower transcriptional activation potentials. However, it should be noted that deletion of one or both TADs did not fully abolish p65 mutants’ capability to trigger transcriptional activation, as demonstrated by comparing the median log2FC ratios of p65 truncation mutants and p65-ΔDNA, considered here as a negative control of both DNA-binding and transcriptional activation.
The relationship between TF DNA binding time and downstream transcriptional activation is fundamental to understanding the mechanism of gene expression and its regulation. In this work, we investigated two fundamental questions: (i) How does the binding time change when TF-DNA or TF-co-regulator interactions are modified or abrogated? (ii) What are the downstream effects on gene activation?
We found that the in vitro DNA affinity of p65 is well-reflected in the DNA binding time measured in live cells for the two affinity mutants KKAA and KKRR. The global transcription levels as measured by RNA-seq further suggest higher transcriptional activation by p65-KKRR (gain of function) while p65-KKAA expression led to a globally strongly reduced gene activation (loss of function). We note that p65-KKAA was able to activate a small subset of genes when compared to p65-WT. While it is possible that p65-KKAA binds these genes with higher affinity, this appears to be an unlikely scenario. For both p65-KKAA and p65-ΔDNA, we found only a minor fraction of long-lasting events which were associated with a high fitting uncertainty. Indeed, also from a structural point of view, neither K122 nor K123 are involved in site-specific DNA contacts [23], but instead provide positive charge strengthening DNA-backbone association. Thus, it is not likely that specific sequences would result in stronger binding of p65-KKAA, which is supported by previous data [22]. Additional evidence comes from p65-ΔDNA, where a small subset of genes also shows upregulation in our RNA-seq screen. This cannot be explained by a shifted DNA affinity since this mutant lacks the entire DNA binding region, suggesting that alternative mechanisms lead to an upregulation of those genes.
We found that the contribution of protein-protein interactions to p65-DNA binding stability was negligible at the genome-wide scale, since p65 truncation mutants lacking the TADs show binding times tb and bound fractions BFs comparable to the wild-type form of p65, (p65-WT). The model previously described for the interferon-beta (IFN-β) locus assigns a prominent function to protein-protein interactions for the stability of the protein complex–the enhanceosome—formed by p65 and associated TFs. According to this model, preventing p65 from interacting with co-regulators of the transcriptional machinery should result in shorter interactions between p65 and target chromatin binding sites. On the contrary, our results indicate that p65 tb remains largely unaffected by the absence of one or both p65-TADs. We note that heterodimers of p65 and endogenous p50 likely maintain some degree of protein-protein interactions. Nevertheless, we found that removing TADs from p65 significantly affects gene transactivation, even if it is not completely abolished. Although our qPCR results (S9 Fig.) indicate gene-specific induction by our p65 mutants, the overall trends agree with those obtained by RNA-seq. qPCR recapitulates the higher induction by p65-KKRR and reduced induction by p65-KKAA when compared to p65-WT. We suggest that any follow-up study of our RNA-seq results should be validated by qPCR to identify potential gene-specific effects not captured in RNA-seq.
A possible interpretation of our findings assigns to TADs the general role of translating stable p65-DNA binding interactions into productive transcriptional events. Importantly, the negligible impact of TAD-mediated protein-protein interactions on p65-DNA binding stabilization, we probed at the genomic-scale, does not rule out the validity of the enhanceosome model described for the single IFN-β locus. The specific promoter architecture likely determines the stabilizing contribution of protein-protein interactions at a locus-specific scale, whereas any genomic scale-recorded readout may average these differences out.
Other studies have also challenged the enhanceosome model. One alternative proposes that protein-protein interactions between p65 and downstream components of the transcriptional machinery were proposed to actively evict p65 from chromatin, showing that the nature of the stabilizing contribution may depend on the specific components recruited to a promoter [15]. However, our findings do not recapitulate these experimental results, since p65-WT molecules displayed similar tb as those recorded for both ΔTAD-mutants. The difference between our results and these previous measurements is again one of genetic context: we used native genes as opposed to arrays of multiple p65 binding sites stably integrated into the genome. Thus, our results more directly address the question of the effect of protein-protein interactions at the genomic scale.
In addition to the activation mechanism relying on the stabilization of protein-protein interactions, p65 is capable of triggering transcriptional activation more indirectly [9,11]. According to this indirect model of p65-dependent gene activation, p65 can act as a “pioneer TF” that promotes chromatin opening, making adjacent regulatory elements accessible to secondary TFs. Previous results have shown that DNA binding of other TFs depends on p65 and its active TAD, suggesting it might play a pioneering role [26]. Supporting evidence comes also from a recent large bioinformatics study [27]. Using computational methods, the authors conclude that a substantial amount of NFκB DNA binding occurs outside of pre-accessible chromatin at time scales that are very similar to other “pioneer” TFs such as PU.1. Following binding, transcriptional activation may be elicited, although the exact mechanism of RNA Pol-II recruitment at these sites has not yet been elucidated (Fig 5B). A first, important consequence of this model is that the removal of TADs is not predicted to affect the stability of p65 binding to target regulatory elements, since p65 can still undergo DNA binding through its unaltered DBD. This insight constitutes the main achievement of the present work, as demonstrated above. A second, more subtle implication of the model concerns the detectable levels of transcription when p65 lacks TADs. As previously shown [28], truncation mutants can still trigger transcriptional activation at specific loci. Notably, this is consistent with our results, since we detected residual gene expression levels when either p65-ΔTA1 or p65-ΔTAD1/2 were overexpressed in our cells.
An additional intriguing finding of the present study concerns the positive correlation between p65 binding time, DNA-binding affinity and transcriptome-wide gene expression. This result recapitulates predictions of the “clutch model” [10], extrapolated from biochemical evidence. According to this model, longer TF binding times should yield higher expression levels of target genes, while the fraction of long-lived events was shown to be rather low. A similar fraction of long-lived binding events was previously observed for other TFs (5.5% for p53 and 8.8% for GR) where they were additionally demonstrated to be linked to transcriptional activation [29]. The available data on p65 DNA occupancy and transcriptional activation [30,31] are consistent with the model that long-lived binding events regulate transcription for NFkB as well.
Interestingly, the “clutch model” was recently challenged by two key studies, but found to hold for both the transcriptional activator p53 [11] and artificial repressor-like effectors [9]. Although the present study was carried out by overexpressing recombinant p65 constructs in HeLa cells, our experimental results recapitulate the expected trend of transcriptional levels [22,28]. Future studies may consider genome editing approaches to avoid overexpression and remove contributions from endogenous p65.
A fundamental aspect of the present study is that we combined SMT with RNA-seq to inspect how p65 binding kinetics correlate with the regulation of gene expression at a genomic scale. Our approach sheds new light on the mechanistic role of p65 trans-activation domains in regulating p65-DNA binding kinetics and the relative transcriptional outcome. We gained also new insights on how p65-DNA binding affinity may tune gene expression, underpinning an emerging model of transcriptional regulation in higher eukaryotes.
Human HeLa cells were cultured in full-supplemented DMEM (high-glucose DMEM, Gibco; 10% vol/vol fetal bovine serum, FBS, Gibco; 1% vol/vol of penicillin/streptomycin mix, Gibco and 1 mM L-glutamine, Gibco). For regular HeLa subculturing, a subcultivation volumetric ratio of 1:5–1:7 was used every 24–48 hours, respectively. HeLa cells were transiently transfected with Lipofectamine 3000 (ThermoFischer Scientific). Cells were seeded in 6-well plates at a density of ~2.0*105 cells/well about 16–20 hours before transfection was performed in antibiotic-free, full-supplemented DMEM. 7.5 μL of Lipofectamine 3000 and 5.0 μg of plasmid DNA were then diluted each in 125 μL of room-temperature OptiMEM (Gibco) in two distinct 1.5 mL Eppendorf tubes. Diluted DNA was supplemented with 10 μL (2 μL/μg of DNA) of P3000 reagent, mixed, and added to diluted Lipofectamine 3000 reagent. Complexes were incubated 15 minutes at RT and evenly distributed on 2.0 mL of fresh, full-supplemented DMEM medium without antibiotics. For microscopic imaging, 10–12 hours later, cells were labelled by adding 0.1–0.5 nM JF549 (L. Lavis, Janelia) in phenol-red free DMEM (LifeTechnologies) supplemented with 10% vol/vol FBS for 30 minutes at 37°C/5% CO2. HeLa cells were then washed 3 times for 20 minutes with phenol-red free complete DMEM to remove excess fluorophore. The mammalian expression vector encoding the Halo- and FLAG-tagged, wild-type human p65 (pCI-neo-p65-Halo-FLAG) was originally obtained from Promega and described in [16]. Point mutants (KKAA, KKRR) and deletion mutants (ΔDNA, ΔTAD and ΔTA1) were generated by mutagenesis directly from pCI-neo-p65-Halo-FLAG. The QuickChange Site-Directed Mutagenesis kit (Stratagene) was used to make point mutations within the wild-type p65 coding sequence and generate KKAA, KKRR. To generate ΔDNA and ΔTAD deletion mutants, an overlap extension PCR protocol was used.
qRT-PCR was used to functionally validate the p65 construct encoded in the pCI-neo-p65-Halo-FLAG expression vector. HeLa cells were seeded in 6-well plates and either transfected or not with pCI-neo-p65-Halo-FLAG plasmid using Lipofectamine 3000. 24 hours later, cells were quickly rinsed with pre-warmed, sterile PBS before performing serum-starvation for 4 hours. Cells were then either treated or not with 20 ng/mL human TNF-α (Sigma) for 30 minutes at 37°C/5% CO2. After stimulation, cells were quickly rinsed twice in ice-cold PBS and total RNA was extracted (RNAeasy Mini kit; Qiagen). Briefly, cells were directly lysed in wells using 350 μL RLT buffer supplemented with β-mercaptoethanol (β-MeOH). Lysates were combined with an equal volume of 70% vol/vol ethanol diluted in DEPC-water and loaded into provided silica mini-columns. After processing samples with 700 μL RW1 buffer and twice with 500 μL of RPE buffer, total RNA was eluted in 30 μL of nuclease-free water. Samples were stored at -80°C until use. RNA samples were retro-transcribed with the SuperScript II Reverse Transcriptase (SuperScript II RT; ThermoFischer Scientific). 250 ng of random primers (RPs; ThermoFischer Scientific) were combined with 1 μg of total extracted RNA from the previous step and 1 μL of dNTPs mix (10 mM each; ThermoFischer Scientific) in 0.2 mL sterile plastic PCR tubes. Reactions were incubated 5 minutes at 65°C and after a brief centrifugation, each sample was added with 4 μL of 5X First-Strand Buffer, 2 μL of 0.1 M Di-thio-threithol (DTT) and 1 μL of RNasin (Promega). Tubes were incubated at 25°C for 2 minutes and 1 μL (200 units) of SuperScript II RT was added. Tubes were allowed to incubate at 25°C for 10 minutes and then at 42°C for 50 minutes. Samples were stored at -20°C until use. Quantitative analysis of p65-target genes NFKBIA and Ccl2 was performed with the support of the Gene Expression Core Facility of the EPFL. Briefly, an automatic pipetting system (Hamilton) was used to combine retro-transcribed cDNA templates with primers specific for target genes NFKBIA (FW: 5’-ATGTCAATGCTCAGGAGCCC-3’, RV: GACATCAGCCCCACACTTCA-3’ and Ccl2 (GeneCopoeia) and four additional housekeeping genes (β-glucoronidase, gusB; β-actin, actB; Eukaryotic elongation factor 1-alpha, eEF-1α; and TATA-binding protein, tbp) in a 384 wells-plate. Three technical replicates were measured for each biological condition. For each qPCR reaction, 3.5 μL of forward and reverse primers premix (200 nM final concentration) were mixed with 1.5 μL of cDNA template diluted 1:5 and 5 μL of SYBR Green 2X Master Mix (Applied Biosystems). 384-wells plates were briefly centrifuged and sealed before performing real-time quantitative PCR with an ABI Prism 7900 Real-time PCR machine (Applied Biosystems). To interpret the data, the threshold cycle (Ct) values obtained with the SDS software (Applied Biosystems) were imported into qBase, a Visual Basic Excel based script for the management and automated analysis of qPCR data for further analysis [32]. Ct values were transformed to normalized relative quantities (NRQs) assuming a gene-amplification efficiency of 2 (i.e. equivalent to 100%). This application for Microsoft Excel allows gene expression quantification relying on multiple reference housekeeping genes. NRQs values were reported as averages of three biological replicates ± standard-error of the mean (SEM).
Single-molecule acquisitions to determine p65 binding kinetics were conducted on an Olympus IX81 inverted microscope equipped with a 100x oil-immersion objective lens (Olympus, N.A. = 1.49) and with an air-stream stage incubator (Okolab UNO, Stage Mad City Labs Z2000) that kept cell samples at 37°C and 5% CO2. The setup for single-molecule microscopy was based on an inclined illumination (HILO) scheme to reduce the background signal originated from out-of-focus molecules [20] and arranged as previously described [21]. Specimen was mounted on a piezoelectric stage enabling selection of the focal plane without modifying the position of the objective. Such a configuration allowed us to adjust the focal plane so that to lie approximately in a middle section of the cell nucleus. Single-molecule stacks (300 frames/stack; 128 x 128 pixels; 18.56 x 18.56 μm2) were acquired by strobing the excitation 561 nm laser (Qioptiq iFlex Mustang). Specifically, to record the binding time (tb) of p65 variants, the EM-CCD camera (Evolve 512; Photometrics) and the laser were synchronized by means of a pulse generator in order to avoid photobleaching when the camera shutter was closed, using an integration time (tint) of 5 ms and a gap time (tgap) of 95 ms (referred in the following as “slow movies”). Single-step displacement analysis (ssd; see next section) and bound-fraction (BF) were computed out of “fast movies”, where tint = 5 ms and tgap = 15 ms. An irradiation intensity of ~1 kW/cm2 was used for both settings. Image stacks were collected using μManager open source microscopy software, setting the EM-CCD camera electronic multiplier (EM) gain to 300 AU.
Movies collected for each p65 construct were analyzed using a Matlab routine (MatlabTrack_v5.03) described in [21]. Individual frames were processed with a band-pass filter using a lower threshold of 1 pixel (equivalent to 145 nm) and a higher threshold of 5 pixels both to smooth the diffraction limited spots corresponding to single molecules and suppress pixel noise. Localization of fluorescent peaks was carried out by using a dedicated algorithm implemented within MatlabTrack_v5.03 using an intensity threshold of 500–700 AU, visually adjusted according to the noise level of the movie. These threshold values allowed us to discard dim peak intensities putatively corresponding to out-of-focus molecules. Tracking was performed by using MatlabTrack_v5.03 which implemented the Matlab version of the Crocker and Grier algorithm [33].
We analyzed “slow movies” (tgap = 95 ms) to estimate the distribution of p65 residence times: to this scope we allowed for a maximum displacement between consecutive frames of 5 pixels (725 nm) to selectively identify slowly moving or immobile molecules. To account for blinking of the fluorophore we allowed an arbitrary gap-length of 3 frames. We discarded tracks shorter than 2 frames. A more stringent selection of putative p65 binding events was performed by using an additional filter implemented in MatlabTrack_v5.03. Specifically, we retained only molecules displacing shorter than 3 pixels (435 nm) and longer than 10 frames (1 s) as previously described [21] by comparison with immobile H2B molecules (S3 Fig). This allows to discard slowly mobile p65 molecules that might otherwise be erroneously interpreted as bound.
Trajectory were calculated out of individual movies collected for each p65 mutant (10–12 movies per condition) using MatlabTrack_v5.03. The duration of each track was assumed to be equal to the time the molecule stays bound while unbleached, i.e. the binding time, tb = 1/koff, where koff corresponds to the kinetic dissociation rate of each detected single molecule. Each distribution corresponding to the different p65 variants was normalized against the trajectory length distribution of H2B to account for photobleaching. To assess the photobleaching kinetics, we first measured and computed the cumulative histogram (i.e. the 1- cumulative distribution function plot, 1-CDF) of an immobile nuclear control protein—H2B-Halo—using the same experimental conditions as for p65. H2B-Halo has a markedly slower decay rate describing the effective photobleaching rate kbl of our experimental system (S10 Fig).
For all further measurements of p65-Halo, we normalized the 1-CDF plots against the 1-CDF histogram obtained for H2B to account for photobleaching (S10 Fig).
This yielded the photobleaching-corrected probability of having a molecule still bound after a time t, which we call the survival probability. The main assumptions for our correction approach are (1) that H2B-Halo and p65-Halo are affected by the same photobleaching rate and that (2) this rate is slower than the expected DNA-binding kinetics. In order to resolve the fast and slow kinetic components kfast and kslow, the 1-cumulative distribution function (1-CDF) histogram for different p65 variants was calculated based on the binding time of each individual track. Values corresponding to calculated 1-CDF distributions were then exported in OriginPro (OriginLab) and fitted according to either a mono (y=A1*exp(−xt1))- or bi-exponential (y=A1*exp(−xt1)+ A2*exp(−xt2)) decay function. Goodness of fitting was evaluated using an F-test as implemented in OriginPro2018b (function ‘compare models). The bound fraction (BF) was calculated by performing the single-step displacement (ssd) analysis as described in [21] using “fast movies” collected with tgap = 15 ms. To determine kon*, we measured the fraction of DNA-bound molecules (i.e. the bound fraction, BF) using single-molecule tracking as described previously (Mazza et al., Nucleic Acids Res. 40(15) 2012). A fast image acquisition (tint = 20 ms, no interval) here allows to capture mobile as well as stationary molecules, which are tracked from one frame to the next, while their displacement was quantified. A histogram of this single-step displacement can then be described using a three component model and compared with an immobile control protein (chromatin-bound histone subunit H2B) (S3 Fig). Briefly, the probability density distribution p(r) of displacing a distance between r and r+Δr in the time Δt between two consecutive frames in our single-molecule movies of Halo-tagged p65 was fit by a n-component diffusion model:
p(r)Δr=rΔr∑i=1nfi2DiΔtexp(−r24DiΔt)
Eq 2
where Di are the diffusion coefficients for each of the species and fi are the fractions of molecules with diffusion coefficient Di, with ∑i=1nfi=1. We found that a three-component diffusion model provided adequate fitting of the experimental data and the slowest diffusion component D1 matched the average diffusion coefficient measured for H2B-Halo (~0.04 μm2s-1). This allows to retrieve the fraction of bound molecules (BF). kon* was the calculated using:
BF=kon*kon*+koff
Eq 3
where koff can be retrieved from single-molecule tracking experiments.
The RNA-seq experiment was run through the genomic technologies facility of the University of Lausanne and the bioinformatics and biostatistics core facility of the EPFL. Purity-filtered reads were adapted and quality trimmed with Cutadapt (v. 1.3, [34]) and filtered for low complexity with seq_crumbs (v. 0.1.8). Reads were aligned against Homo sapiens v. GRCh38 genome using STAR (v. 2.4.2a, [35]). The number of read counts per gene locus was summarized with htseq-count (v. 0.6.1, [36]) using H. sapiens v. GRCh38 Ensembl 82 gene annotation. Quality of the RNA-seq data alignment was assessed using RSeQC (v. 2.3.7, [37]). Reads were also aligned to the H. sapiens v. GRCh38 Ensembl 82 transcriptome using STAR (v. 2.4.2a, [35]) and the estimation of the isoforms abundance was computed using RSEM (v. 1.2.19, [38]). Statistical analysis was performed for protein-coding genes and long non-coding genes genes in R (R version 3.1.2). Genes with low counts were filtered out according to the rule of 1 count per million (cpm) in at least 1 sample. Library sizes were scaled using TMM normalization (EdgeR v 3.8.5; [39]) and log-transformed with limma voom function (R version 3.22.4; [40]). Differential expression was computed with limma [41] by fitting data into a linear model, extracting the contrasts for all pairwise comparisons of transfected vs Non-transfected (NT). A moderated F-test was applied and the adjusted p-value computed by the Benjamini-Hochberg method, controlling for false discovery rate (FDR). Genes displaying a FDR < 0.1 were selected as differentially expressed. Data analysis performed as described above identified 1080 differentially expressed genes at FDR 10%.
However, the transcriptional activity of the different p65-Halo variants was assessed using only a subset of 215 differentially expressed genes selected as direct binding targets of p65 on the basis of the ENCODE ChIPseq deposited information (Dr. Jacques Rougemont; Bioinformatics and biostatistics core facility, EPFL). To calculate the intersection between differentially expressed genes based on RNA-seq analysis and the ENCODE ChIPseq database, we first identified find the differentially expressed gene (RNA-seq) coordinates using genrep4humans.py assembly hg19. Then we defined regions of interest on the forward strand from Gene_Start-2000 to Gene_End and on the reverse strand from Gene_Start to Gene_End+2000. We use bedtools intersect to get the intersection of the peaks and the regions promoter+gene. We find at least one peak in 19.9% of the differentially expressed genes: 215 / 1080.
We used venn_mpl.py from pybedtools to plot a Venn diagram of genomic regions. We plotted the intersection between the promoter + gene region and promoter regions of the differentially expressed genes and the ChIPseq Peak regions using the hg19 assembly. We performed the NFKB1_REL_RELA.P2 motif search (obtained from the Swissregulon Database at http://swissregulon.unibas.ch/fcgi/wm?wm=NFKB1_REL_RELA.p2&org=hg19) in the promoter regions of the selected genes (on the forward strand Gene_Start+/-2000 and on the reverse strand Gene_End+/-2000). In order to retrieve the FASTA sequence, we used bbcfutils, a collection of tools used at the Bioinformatics & Biostatistics Core Facility, EPFL, Lausanne, Switzerland. More precisely, the genrep4humans.py script gives the gene coordinates with the assembly hg38, (https://github.com/bbcf/bbcfutils/blob/master/Python/genrep4humans.py).
Average expression levels of genes scoring at least one ChIPseq peak in either the promoter or coding sequence were visually represented in a single heat map calculated in MatLab using the dendrogram function. Z-score defines how many standard deviations a value is away from the population mean:
z=(x−μ)σ
Eq 4
In the dendrogram shown in Fig 3B, the values are normalized per row (i.e. mean = 0, SD = 1). The heat map displays the expression levels of the 215 genes identified from the ENCODE ChIP-seq database obtained for each p65 mutant and corrected for the NT sample. For each p65 mutant, we plotted the relative expression (log2 fold-change; log2FC) of the wild-type p65-transfected condition on x-axis and mutant p65-transfected condition on y-axis, both compared to the wild-type non-transfected condition (NT). We computed the correlation coefficient (r) and the median log2 FC over all genes. The later could then be used to classify p65 mutants as either loss- or gain-of function.
At 12–16 hours post-transfection, Hela cells were labeled with 5 μM OregonGreen Halo Ligand (Promega) for 30 minutes at 37°C/5% CO2. Cells were then extensively washed in pre-warmed phenol red-free DMEM so that to be sure to have eliminated the majority of unbound fluorescent ligand. FRAP experiments have been carried out with the Leica SP8 confocal fluorescence microscope using an oil-immersion PLAN-APOCHROMAT 60X objective. Fluorescence was excited with an Argon laser set at 80% of its total power. Pre- and post-bleach images (256x256 pixels) were acquired with a pinhole aperture set to 2 Airy-units using bidirectional scanning mode for faster acquisition. A total of 50 pre-bleach and 500 post-bleach frames were collected at 0.2% of AOTF and a zoom factor of 8 that resulted in a final pixel size of 180 nm. The delay time between successive frames was 69 ms. Bleaching was obtained using the 488 nm Argon laser set at maximum power and the zoom-in option implemented in the TCS SP8 FRAP module (one bleaching frame only).
Regions of interest (ROIs) corresponding to the photobleached area, the whole nucleus and the background region were manually segmented in Fiji (https://imagej.nih.gov/ij/) for each recorded cell. FRAP curves were then calculated and normalized using FRAPAnalyser 2.0 (http://actinsim.uni.lu/). Double exponential fitting was performed according to:
FRAP(t)=I1∙(1−e−tτ1)+I2∙(1−e−tτ2)
Eq 5
which was implemented in the FRAPAnalyser 2.0. τ1 and τ2 are the time-constants corresponding to the fast and the slow component, respectively, calculated as τ1=t12fast/ln2 and τ2=t12slow/ln2, being t12fast and t12slow the half-times of recovery of the fast and slow fractions.
The global half-time of recovery t1/2global corresponds to the 50% of fluorescence signal recovery. The value of FRAP(t12global) can be computed as:
FRAP(t12global)=FRAP(∞)−FRAP(0)2
Eq 6
Given that: FRAP(t→∞)=I1+I22 and FRAP(t→0)=0,FRAP(t12global)=I12+I22. Therefore, when t=t1/2global we have that:
I1∙(0.5−e−t12globalτ1)+I2∙(0.5−e−t12globalτ2)=0
Eq 7
To compute t1/2global, the equation reported above is solved numerically in R using the function uniroot (see S1 Table).
Suspension-adapted HEK293 cells were routinely maintained in serum-free ExCell 293 medium (SAFC Biosciences, St. Louis, MO) with 4 mM glutamine as described [42] in a shaking ISF-4-W incubator (Kühner AG, Birsfelden, Switzerland) at 37°C in the presence of 5% CO2 at the Protein Expression Core Facility of the EPFL (in collaboration with Dr. D. Hacker). HEK293 cells were transfected with pCI-neo-p65-Halo-FLAG as described in [43]. 24 hours post-transfected cells (~109) were harvested by centrifugation (500×g, 5 minutes at RT) in two 50-mL Falcon tubes. Supernatant was filtered and maintained in the cell incubator at 37°C with 5% CO2 to perform TNF-α stimulation. Cell pellets were pooled together in a single 50-mL Falcon tube and resuspended in 40 mL of the original pre-equilibrated cell culture medium supplemented with 20 ng/mL of human TNF-α (Sigma). The cell suspension was incubated for 30 minutes at 37°C in an orbital shaker (180 rpm). Stimulated HEK293 cells were then pelleted (2000×g for 5 minutes at 4°C) and resuspended in ice-cold PBS added with phosphatase (Sigma) and protease inhibitors (COMPLETE; Roche). Cell washing with supplemented-PBS was repeated once and the cell pellet was frozen at -80°C for 1 hour to help protein releasing from cells due to a facilitated plasma-membrane rupture out of a freeze-thaw cycle. Thawed cells were added with ~5 packed cell volume (pcv) of PBS with 1% vol/vol phosphatase inhibitors (Sigma) and 1 mM DTT and pelleted at 1’800×g for 5 minutes at 4°C. Cell pellet was resuspended in ~3 pcv of hypotonic buffer (10 mM HEPES, pH 7.9 at 4°C; 1.5 mM MgCl2; 10 mM KCl; immediately before use add protease inhibitors (1 COMPLETE table/10 mL of buffer) and 1 mM DTT) and incubated on ice for 10 minutes. Cells were then homogenized with a glass, ice-cold 15-ml Dounce homogenizer (pestle B, 28 strokes on ice; Kimble Chase). This step disrupts the majority of cells membranes but keeps nuclei intact. Nuclei were then pelleted (3’300×g for 15 minutes at 4°C) and resuspended in 1 packed-nuclear volume (pnv) of low-salt buffer (20 mM HEPES, pH 7.9 at 4°C; 25% glycerol; 1.5 mM MgCl2; 0.2 mM EDTA; supplement with protease inhibitors and 1 mM DTT prior to use). Nuclei were then dispersed thoroughly with the 15-mL Dounce homogenizer (pestle B) while adding 5 M NaCl dropwise up to a final concentration of 420 mM to allow chromatin-bound proteins to be extracted from nuclei. Nuclear lysates were incubated for 30 minutes at 4°C on a rotating wheel and ultracentrifuged (100’000×g for 1 hour at 4°C). The supernatant was then collected and diluted with one volume of hypotonic buffer supplemented with 1 mM DTT, 20% vol/vol glycerol, 0.2% NP-40 alternative and phosphatase/protease inhibitors.
The recombinant p65-Halo-FLAG protein was purified from nuclear crude extracts by performing a pull-down with anti-FLAG M2 magnetic beads (Sigma). 2.0 mL of beads were washed three times in 5 mL of equilibration buffer (10 mM HEPES, pH 7.9 at 4°C; 10 mM KCl; 1.5 mM MgCl2; 200 mM NaCl; 0.1% vol/vol NP-40 alternative; 10% vol/vol glycerol) and collected through the magnet. The nuclear crude extract (~6 mL) was then added to beads together with 14 μM of JF549 fluorescent Halo-ligand (from Dr. L. Lavis) and incubated ON at 4°C on a rotating wheel. After extensive washing in elution buffer (10 mM HEPES, pH 7.9 at 4°C; 200 mM NaCl; 0.1% vol/vol NP-40 alternative; 1 mM DTT; 1 mM EDTA; 10% vol/vol glycerol freshly supplemented with protease and phosphatase inhibitors) to remove unbound proteins and excess fluorophore, beads were incubated in elution buffer supplemented with 100 μg/mL of FLAG peptide (Sigma) for 1 hour at 4°C on a rotating wheel. The supernatant was then collected and stored at 4°C until use. The elution step was performed three times and supernatants were pooled together and concentrated in Centricon 10 kDa MWCO centrifuge filters at 5000 × g for ~2 hours at 4°C.
The eluted p65-Halo-FLAG protein concentration was determined by performing denaturing sodium dodecyl sulphate-polyacrylamide gel electrophoresis (SDS-PAGE) against known amounts of bovine serum albumin (BSA) standards, followed by Coomassie staining (SimplyBlue SafeStain; Thermo Fischer Scientific). Variable volumes of eluted p65-Halo-FLAG (0.5 μL, 5 μL and 10 μL) and BSA standards (0.2 μg, 0.5 μg, 1.0 μg, 1.5 μg, 3.0 μg and 4.0 μg) were denatured in 1x Laemmli Sample Buffer (Alfa Aesar) and boiled for 5 minutes at 95°C. Samples (20 μL final volume) were separated using 12% SDS-PAGE prepared from stock 37.5:1 polyacrylamide:bis-acrylamide solution (Fischer Scientific) and run at 120 Volts for ~1 hour in Tris-Glycine running buffer (25 mM TrisCl; 250 mM glycine; 0.1% SDS) using a MiniProtean System (Biorad). For Coomassie staining, the minigel was rinsed three times with ~100 mL deionized water and ~20 mL of blue stain were added and incubated ON. The minigel was destained 2 hours with 100 mL of water. The final protein concentration was ~47 ng/μL (~470 nM) as estimated from densitometry (ImageJ).
For Western Blot, samples preparation and electrophoresis were performed as described above. Proteins were ON-transferred to a nitrocellulose membrane (Protran Hybond ECL; GE Healthcare) at 4°C in Towbin transfer buffer (25 mM TrisCl; 192 mM glycine, pH 8.3; 20% methanol and 0.1% SDS) at 100 Volts using the MiniProtean transfer cassette (Biorad). Membranes were then blocked in non-fat dry milk (5% w/vol; Biorad) for 1 hour at RT and probed with mouse monoclonal IgG1 anti-p50 antibody (1:200; Santa Cruz) ON at 4°C in TBST (20 mM TrisCl pH 7.5; 150 mM NaCl; 0.1% Tween.20) supplemented with 5% w/vol non-fat dry milk. Filters were then washed 3 times for 15 minutes each with TBST and probed with a sheep anti-mouse, peroxidase-labelled antibody (Amersham) for 45 minutes at RT in TBST supplemented with 5% w/vol non-fat dry milk. Membranes were washed 3 times for 15 minutes with TBST and developed with ECL Plus system (Thermo Scientific). Chemiluminescence detection was carried out with a gel fluorescence scanner (ChemiDoc; Biorad). Notably, p65-Halo-FLAG was detected directly through the fluorescence emitted from the covalently-bound JF549 and, therefore, did not need to be probed with a specific antibody.
Synthetic HPLC-purified sense and anti-sense oligo probes encoding the consensus binding sequence of p65 were purchased from Microsynth (Microsynth AG, Switzerland; Sense-p65_κB: 5’-AGTTGAGGGGACTTTCCCAGGC-3’; Anti-sense-p65_κB: 5’-GCCTGGGAAAGTCCCCTCAACT-3’). An Atto647N dye was attached to the 5’ of the sense-strand to visualize DNA by fluorescence detection. To make double-stranded DNA probes, a pair of sense and anti-sense oligos were mixed at 50 μM each in annealing buffer (10 mM TrisCl, pH 8.0; 1 mM EDTA and 50 mM NaCl), then annealed in a PCR machine with the following program: 95°C (3 minutes), 0.1°C/second drop to 55°C, 55°C (60 minutes), 0.1°C/second drop to 25°C as previously reported [44]. In EMSA, 0.5 μM fluorescent double-stranded DNA probe (dsDNA) was mixed with 0.1, 0.3 and 0.6 μg of purified, JF549-labeled p65-Halo-FLAG in 20 μL binding buffer (25 mM HEPES, pH 7.6; 0.1 mM EDTA; 12.5 mM MgCl2; 100 mM KCl; 0.01% NP-40 alternative and 10% glycerol) and incubated at 4°C for 60 minutes and then at RT for 10 minutes before loading into 1% w/vol agarose (Sigma) prepared in 1x TBE buffer (45 mM Tris-borate; 1 mM EDTA) and prerun at 120 Volts, 4°C for 30 minutes. Samples were run by electrophoresis at 150 Volts for ~1 hour at 4°C in 0.5x TBE buffer. Fluorescence signals were scanned with a Chemidoc imaging system (Biorad).
|
10.1371/journal.ppat.1003843 | Recovery of an Antiviral Antibody Response following Attrition Caused by Unrelated Infection | The homeostatic mechanisms that regulate the maintenance of immunological memory to the multiple pathogen encounters over time are unknown. We found that a single malaria episode caused significant dysregulation of pre-established Influenza A virus-specific long-lived plasma cells (LLPCs) resulting in the loss of Influenza A virus-specific Abs and increased susceptibility to Influenza A virus re-infection. This loss of LLPCs involved an FcγRIIB-dependent mechanism, leading to their apoptosis. However, given enough time following malaria, the LLPC pool and humoral immunity to Influenza A virus were eventually restored. Supporting a role for continuous conversion of Influenza A virus-specific B into LLPCs in the restoration of Influenza A virus immunity, B cell depletion experiments also demonstrated a similar requirement for the long-term maintenance of serum Influenza A virus-specific Abs in an intact LLPC compartment. These findings show that, in addition to their established role in the anamnestic response to reinfection, the B cell pool continues to be a major contributor to the maintenance of long-term humoral immunity following primary Influenza A virus infection, and to the recovery from attrition following heterologous infection. These data have implications for understanding the longevity of protective efficacy of vaccinations in countries where continuous infections are endemic.
| Antibody responses to infectious pathogens are critical in host survival, recovery and protection from reinfection; they also correlate with the success of vaccination. It is currently thought that antibody serum titers are maintained at protective levels over long periods of time by specialized long-lived antibody-secreting plasma cells residing in the bone marrow. Indeed, antibodies against the original virus can still be found in survivors of the 1918 Spanish Flu, more than 90 years ago. However, it is also becoming clear that subsequent infection with heterologous pathogens may cause attrition of previously established immunological memory, in order to accommodate new lymphocyte specificities in the finite space of the host. This phenomenon is seemingly at odds with long-term maintenance of immunological memory. We also show that a single episode of malaria, caused by infection by Plasmodium chabaudi, leads to the loss of preexisting plasma cells, serum antibodies and protective immunity against Influenza A virus. However, Influenza A virus-specific immunity does eventually recover in these animals with the replenishment of plasma cells by B cells over the course of several weeks. Thus, the reported mechanism reconciles attrition of immunological memory by heterologous infection and long-term stability, and places B cells, instead of their descendant plasma cells, at the center of humoral memory.
| Infection or vaccination usually induces high levels of antigen-specific antibodies (Abs) in the systemic circulation and mucosal surfaces. These Abs can be maintained for long periods of time in the absence of re-infection, despite the relatively short half-life of serum immunoglobulins, which is measured in weeks [1]. For example, virus-neutralizing Abs have been detected in humans over 90 years after Influenza A virus infection [2] and in mice over 250 days after lymphocytic choriomeningitis virus (LCMV) infection [3]. The establishment of these long-term Ab responses relies on the maintenance of antigen-specific memory B cells (MBCs) and long-lived plasma cells (LLPCs). MBCs and LLPCs occupy distinct anatomical locations in the spleen and bone marrow, respectively, which are thought to be of finite size and under homeostatic control [4]. One consequence of such regulation is that new antigenic challenges, particularly with complex pathogens that generate large populations of MBCs and LLPCs, would affect the maintenance of Ab responses to previously encountered antigens.
Infection with the malaria parasite, Plasmodium, has long been known to induce a strong B cell response, giving rise to large numbers of LLPCs [5], hypergammaglobulinemia in humans [6] and in experimental models [7], [8], and perturbations of splenic and bone marrow microarchitecture [7], [9]. Ab responses, MBCs or LLPCs specific for antigens administered prior to or during an experimental blood-stage malaria infection can be delayed, and/or reduced in magnitude and avidity [8], [10], [11]. Similar observations were made after Trypanosoma brucei infection of mice, which caused a reduction in pre-established MBCs and LLPCs and an increase in susceptibility to heterologous infection [12].
The mechanisms by which subsequent infections may cause the attrition of pre-existing heterologous MBCs and LLPCs are not entirely understood. Apoptosis of pre-existing parasite-specific and unrelated MBCs and LLPCs has been described in non-lethal rodent Plasmodium strain P. yoelii [10]. Immune complexes cross-linking of the inhibitory receptor FcγRIIB on the surface of LLPCs have been shown to induce apoptosis of LLPCs in the bone marrow that were induced by protein immunization [13]. However, it is currently unclear whether or not similar mechanisms underlie loss of pre-established humoral immunity following protein immunization or parasitic infection.
Loss of pre-existing heterologous humoral immunity following parasitic infections has been documented extensively [10]–[12], [14] and is seemingly at odds with long-term maintenance of antiviral Abs [15]. Therefore we examined more closely both the kinetics and potential mechanisms of humoral memory attrition. Here, we investigated whether the blood stages of the malaria parasite would affect pre-established humoral immunity to Influenza A virus. We established a mouse model of sequential infection with Influenza A/Puerto Rico/8/34 (PR8) and the rodent malaria parasite Plasmodium chabaudi chabaudi (AS). We found that sequential infection of PR8-immune mice with P. chabaudi resulted in the loss of pre-established serum PR8-specific Abs and LLPCs in the bone marrow, and this rendered mice more susceptible to PR8 challenge. Moreover, during P. chabaudi infection, LLPCs underwent apoptosis in the bone marrow, through an FcγRIIB-dependent mechanism. However, the loss of pre-established humoral immunity was temporary, as antiviral serum Abs and LLPC numbers did eventually return to levels observed before the P. chabaudi infection. Importantly, B cells were essential for the maintenance of long-lived serum Ab titers to PR8, as B cell depletion in PR8-immune mice resulted in the eventual loss, without recovery, of LLPCs and antiviral serum Abs. These results confirm the detrimental effect of parasitic infection on the LLPC pool and serum titers of antiviral antibodies, which is eventually restored by further LLPC generation, thus reconciling humoral memory attrition by subsequent infection and long-term stability.
BALB/c mice were first infected with PR8 and the kinetics of Ab induction, specific serum Ab concentrations and specific plasma cells and MBCs were quantified at various time points after infection. Intranasal PR8 infection resulted in a gradual increase in serum HA-specific IgG (Fig. 1A), which plateaued at a median of approximately 100 µg/ml 80–100 days post infection, and remained stable for a further 100 days. By contrast, HA-specific IgM increased within the first 14 days, but did not increase further over the 200-day period of the experiment (Figure 1A). Serum neutralizing Ab (nAb) titers measured by a modified viral neutralization test [16] followed a very similar kinetic to HA-specific IgG Ab response measured by ELISA, peaking approximately 80 day after infection and remaining stable for up to 200 days post-infection (Figure 1B)
To determine whether the established anti-PR8 humoral response would be affected by a malaria infection, BALB/c mice were infected with 105 P. chabaudi pE 105 or 150 days after intranasal inoculation of PR8 (Figure 1C), when the PR8 HA-specific IgG response was stable. Infection with P. chabaudi at both time points caused a significant reduction in HA-specific IgG within 21 days after the P. chabaudi infection (Figure 1D–E). Importantly the loss of HA-specific IgG Abs was accompanied by a substantial decrease in titers of PR8 neutralizing Abs (Figure 1F), and consequently there was also a significant loss of anti-viral immunity (Figure 1G), as shown by the increased viral titers on day 3 upon re-challenge of these PR8-P. chabaudi infected mice with PR8 42 days after P. chabaudi infection. Although the cellular immune response to Influenza A virus rechallenge can be highly protective, it is typically delayed in comparison with the immediate protected afforded by pre-existing Abs [17]. Therefore, susceptibility to PR8 re-challenge at this early time-point would be indicative of the loss of PRR-specific humoral immunity after P. chabaudi infection.
The loss of PR8-specific Abs was not due to a reduction in half-life of IgG, as neither acute nor chronic P. chabaudi infection induced increased clearance of IgG (Figure S2). We also established that there was little cross-reactivity of Abs induced by each infection (Figure S1A–C), and Abs induced by infection with P. chabaudi alone were not able to neutralize PR8 in vitro (Figure S1D).
Therefore, a P. chabaudi infection induced loss of pre-established PR8-specific Abs, which was unrelated to homeostatic regulation of immunoglobulin concentrations.
After infection or immunization, serum Ab levels are thought to be maintained by long-lived plasma cells (LLPC) in the bone marrow [3]. Therefore, we investigated whether the reduction of HA-specific Abs, and thus reduced immunity to re-infection with PR8, following a P. chabaudi infection could be due to the loss of LLPC.
First, BALB/c mice were infected with P. chabaudi and the absolute number of bone marrow (BM) cells from femur pairs was determined for up to 80 days after P. chabaudi infection. In addition, parasitaemia was monitored throughout acute P. chabaudi infection. Bone marrow cellularity was reduced by day 8 following the P. chabaudi infection (Figure 2A), coinciding with the peak of parasitaemia, and then recovered on day 20 as the parasitaemia dropped, and remaining stable for up to 80 days post-infection.
We determined the number of HA-specific Ab-secreting cells (ASC) in the bone marrow as a measure of pre-established LLPC. After a primary PR8 infection, HA-specific ASCs accumulated in the bone marrow and reached a stable number by approximately day 50, and remained at this level for up to 250 days (Figure 2B). However, when PR8-immune mice were infected with P. chabaudi 150 days later, there was a significant reduction in the number of HA-specific ASC within 21 and for up to 42 days of P. chabaudi infection (Figure 2C and Figure 2D). There was a distinct loss of HA-specific ASCs relative to the total IgG ASC compartment in the bone marrow, which remained unchanged (Figure 2E) Therefore, infection with P. chabaudi resulted in the rapid and sustained loss of pre-established HA-specific ASC in the bone marrow.
The loss of HA-specific ASC from bone marrow during acute P. chabaudi infection could be due to dislocation by competition with migratory plasmablasts, as previously suggested during a secondary immunization of human subjects with tetanus toxoid [18], or by apoptosis of LLPC, as previously described during infection with non-lethal P. yoelii [10] and after immunization with a immunogenic cocktail of antigens [13].
Newly generated plasma cells (CXCR4+ CXCR5− CD19− MHCII+) were transiently detected in the blood between days 8 and 12 of a P. chabaudi infection (Figure S3C), and migratory plasmablasts (B220+ CD138+) were present in the bone marrow from day 10 onwards (Figure S3A and B). Despite the influx of these plasmablasts, we did not detect any increase in B220− CD138+ LLPCs (Figure S3C) or HA-specific ASC (Figure S3D) above background levels in the blood on days 8 or 10 of P. chabaudi infection, indicating that it was unlikely that pre-established HA-specific ASC were competitively dislocated from bone marrow by migratory plasmablasts, at least not in sufficient frequencies for detection by flow cytometry or HA-specific ELISpot in the blood. By contrast, there was a significant but transient increase in the numbers of Annexin V+ apoptotic bone-marrow cells and bone-marrow LLPCs 4 days after a P. chabaudi infection (Figure 3A–B).
FcγRIIB expressed on LLPCs has been previously implicated in homeostatic regulation of the LLPC niche, in a cell-intrinsic manner, during immune responses by inducing apoptosis of LLPC after ligation by elevated concentrations of immune complexes [13]. Since we observed dramatically elevated levels of total serum IgG during acute and chronic P. chabaudi infection (Figure S4B), we investigated whether apoptosis of LLPC during P. chabaudi infection could be mediated via immune complex ligation of FcγRs.
Bone marrow LLPCs induced by PR8 infection expressed FcγRIIB (Figure 4A). Using C56BL/6 mice lacking either FcγRIIB (FcγIIB−/−) or FcγRI,II, and IIIa (FcγRI,II,III−/−), we asked whether HA-specific Abs and HA-specific ASCs were maintained after a P. chabaudi infection in the absence of these Fcγ receptors. We established that there was a similar loss of HA-specific antibody and ASC in PR8/P. chabaudi infected wild-type C56BL/6 mice compared with BALB/c mice (data not shown), and that the course of infection of P. chabaudi were similar in FcγRI,II,III−/− and FcγRIIB−/− mice compared to those of wild-type C56BL/6 (Figure S5A–C) and [19].
Despite similar peak parasitaemias, there was no loss of either total bone marrow cellularity (Figure 4B) or HA-specific ASCs (Figure 4C) on day 8 of P. chabaudi infection in FcγRI,II,III−/− mice compared with the significant loss in C56BL/6 mice. In line with this, FcγRI,II,III−/− mice retained their pre-established levels of HA-specific serum IgG for up to 56 days after P. chabaudi infection (Figure 4D), whereas there was a significant drop in Ab levels in wild-type C56BL/6 mice at this time. Interestingly, although there was no reduction in total bone marrow cellularity in FcγRI,II,III−/− mice after a P. chabaudi infection, there was a significant loss of total bone marrow cellularity in FcγRIIB−/− mice infected with P. chabaudi (Figure 4E), suggesting that the loss of LLPC and the loss of other cells may be via engagement of different Fc receptors. Importantly, there was no loss of HA-specific ASCs in FcγRIIB−/− mice (Figure 4F), strongly implicating FcγRIIB as the crucial factor in the maintenance of pre-established LLPCs. In line with this, whilst a substantial fraction of total bone marrow cells and bone marrow LLPC in C56BL/6 mice infected with P. chabaudi was apoptotic and expressed Annexin V, there was no change in Annexin V expression levels in infected FcγRIIB−/− mice at all (Figures 4G and 4H). Ligation of the FcγRIIB, therefore, is an important mechanism of loss of pre-established HA-specific bone marrow ASCs and serum Abs during P. chabaudi infection.
To determine the longer-term effects of this P. chabaudi infection on the humoral immune response to PR8, we monitored HA-specific IgG in plasma and numbers of HA-specific ASCs in bone marrow for up to 100 days after P. chabaudi infection. To our surprise, we observed a gradual return of HA-specific IgG by day 42 of P. chabaudi infection to the previously established levels prior to the P. chabaudi infection (Figure 5A). Similarly, the numbers of HA-specific ASCs returned to pre-established numbers 84 days after P. chabaudi infection, and thereafter remained at this level for up to 110 days after P. chabaudi infection (Figure 5B), suggesting that HA-specific ASCs are being replenished in the bone marrow, despite no re-infection with PR8. This increase in plasma cells follows a similar kinetic to the restitution of HA-specific Abs in serum.
Memory B cells (MBCs) can differentiate into Ab-secreting plasma cells on re-challenge with the specific antigen through the B cell receptor or stimulation of Toll-like receptors (TLRs) [20], [21]. Bystander CD4+ T cell help in vitro can also stimulate non-specific MBCs to differentiate into PCs, possibly because of an increased availability and upregulation of co-stimulatory molecules and production of Th2 cytokines [21], [22]. We hypothesized that the differentiation of MBCs or other B cell subsets into plasma cells was responsible for the eventual replenishment of bone marrow HA-specific ASCs and thus of serum HA-specific IgG concentrations to pre-established levels before the P. chabaudi infection.
To determine whether B cells in general and MBCs in particular can contribute to the maintenance of HA-specific serum IgG after infection with PR8, we selectively depleted B cells, but not LLPCs in vivo in PR8-immune hCD20 transgenic mice using the monoclonal anti-hCD20 Ab, 2H7 [23], [24]. This depletion was highly specific for B cells as determined by the surface markers sIgD, CD19 and CD21 (Figure S6C–D). A two-week course of 2 mg/wk treatment with the mAb (Figure S6B) depleted more than 90% of all B cells in spleen, peripheral blood and lymph nodes, and more than 50% of B cells in bone marrow of PR8-immune hCD20tg mice, but not in their hCD20tg-negative littermates (Figure S6D). In line with previous studies [25], [26], treatment with anti-hCD20 mAb did not affect total numbers of CD138+ B220− LLPC in the spleen and bone marrow in hCD20tg mice and hCD20tg-negative littermates either in immediately after treatment or 150 days after treatment (Figure S6E). Similarly, treatment with anti-hCD20 mAb did not deplete pre-existing GC B cells in the spleen (Figure S6F). Instead, numbers of splenic GC B cells were transiently elevated in both hCD20tg and hCD20tg-negative mice immediately after anti-hCD20 mAb administration (Figure S6F).
The majority of MBCs are thought to reside in the spleen. We therefore quantified HA-specific MBC cells in the spleen with ELIspot as a representative measure of total depletion efficacy. More than 90% of HA-specific MBCs were depleted rapidly from the spleen and this loss remained significant even 150 days after depletion (Figure 6A), indicating significant long-term depletion of the vast majority of pre-established HA-specific MBCs. Treatment with anti-hCD20 mAb had no effect on the mean of HA-specific MBCs in hCD20tg-negative littermate controls (Figure 6A).
Contrasting the rapid and sustained reduction in HA-specific MBCs, the total MBC niche appeared to be rapidly depleted, but filled up to pre-existing numbers by day 42 (data not shown), presumably with MBC of irrelevant specificities.
In contrast to HA-specific MBCs (Figure 6A), HA-specific ASCs in bone marrow were not immediately depleted by anti-hCD20 mAb treatment (Figure 6B and C). However, we observed a significant decrease in the number of HA-specific ASCs from 70 days post-depletion onwards in hCD20tg mice, but not in hCD20tg-negative littermates (Figure 6B and C). In contrast, there was no change in total IgG ASCs in hCD20tg mice, even 150 days post-depletion (Figure 6D). The eventual loss of HA-specific ASCs, but not of total IgG ASCs (Figure 6E), suggested that this loss was not a non-specific effect of 2H7 treatment, but perhaps a consequence of the long-term depletion of the HA-specific B cell compartment, which is therefore unable to replenish the HA-specific ASC niche.
Finally, we observed a significant loss of HA-specific IgG only from 70 days post-depletion and for up to 150 days post-depletion (Figure 6F), without any change in total serum IgG in hCD20tg mice (data not shown). This eventual loss of HA-specific ASC and HA-specific serum IgG at later time-points after depletion of HA-specific B cells demonstrates the importance of a complete HA-specific B cell compartment to maintain numbers of HA-specific ASC and HA-specific IgG in mice. The kinetics of the loss of HA-specific ASC and HA-specific Ab were very similar. From these data (Figure 6C and F), in the absence of a HA-specific B cell compartment, the half-life of HA-specific ASC was calculated to be approximately 72 days, whilst the half-life of serum HA-specific Abs is approximately 86 days.
The requirement for HA-specific B cells in maintaining long-term HA-specific ASCs and Abs strongly suggests that HA-specific B cells are very likely to be responsible for the eventual return of HA-specific ASCs, and therefore serum HA-specific Ab at late time points of P. chabaudi infection.
We have used a mouse model to examine the requirements for maintenance of long-term humoral immunity to Influenza A virus, a virus that induces life-long humoral immunity in humans [2]. We show that serum levels of Influenza A virus-neutralizing Abs are maintained by continuous conversion of Influenza A virus-specific B cells into antibody-secreting LLPCs under steady-state conditions. A single malaria episode significantly dysregulates this maintenance of Influenza A virus-neutralizing Abs: a P. chabaudi infection, initiated after the B cell and antibody responses to Influenza A virus reach a stable plateau, results in the loss of pre-established serum Abs and plasma cells specific to Influenza A virus and in increased susceptibility to Influenza A virus re-infection. The loss of LLPC and the reduction in Abs is mediated via an FcγRIIB-dependent mechanism resulting in their apoptosis. However, continuous conversion of Influenza A virus-specific B cells into LLPCs following P. chabaudi infection eventually replenishes the LLPC pool and restores humoral immunity to Influenza A virus, highlighting that this arm of adaptive immunity can withstand considerable homeostatic disruption.
Homeostatic regulation of LLPC occurs as they compete for space in survival niches, supported by intrinsic and extrinsic survival resources [27]. The size of the LLPC niche in the bone marrow is finite [4], [28], [29] and has to accommodate LLPCs with specificities against different infections over time [4]. LLPCs may therefore be lost from their niches by competition from new migrating plasmablasts generated by heterologous infections [4], [18]. During each new immune response, some of the pre-established LLPCs are assumed to be removed [4], [30]. If the numbers of newly migrating plasma cells are sufficiently large to reduce the number of established antigen-specific LLPCs to below the threshold required to sustain enough specific serum Abs to neutralize re-infection, the host may effectively lose protective immunity to that pathogen. Loss of Influenza A virus-specific ASCs and concomitant loss of protective immunity to Influenza A virus re-infection following P. chabaudi infection, in the face of a total IgG ASC pool which did not change during P. chabaudi infection, strongly suggests that such a mechanism operates in this context. Although our analysis took place on day 3 post-rechallenge, which was heavily reliant on pre-existing Abs for protection [17], we cannot formally exclude the potential contribution of cellular immunity. However, unlike the finite LLPC compartment, it is thought that the memory CD8+ T cell compartment is expandable [31] and therefore perhaps not as susceptible to stochastically-determined attrition as the LLPC pool is.
Several mechanisms have been put forward to explain loss of LLPCs from bone marrow as a result of subsequent infection or immunization. As suggested for developing B cells in Trypanosoma brucei infections of mice [32], loss of expression of CXCL12 on LLPCs, which is required for their retention in BM [33], could result in displacement from the BM and subsequent cell death. Although we did not investigate levels of CXCL12 expression on LLPCs, we found no evidence of displacement of LLPCs or HA-specific ASCs into the blood, suggesting that this may not explain the loss of ASCs in this P. chabaudi infection. Rather, our data support the idea that plasma cells undergo apoptosis in situ in the bone marrow, in agreement with previous studies in P. yoelii infections, in which caspase-3-dependent apoptosis of MBCs and LLPCs was evident [10].
Apoptosis of LLPCs following injection of an immunogenic cocktail of antigens has been suggested to result from binding of immune complexes to the inhibitory FcγRIIB expressed on B cells and plasma cells [13]. A cell-intrinsic role for FcγRIIB in LLPC apoptosis was demonstrated in adoptive transfer studies of wild-type and FcγRIIB-deficient immune splenocytes, which were then differentiated into LLPCs upon secondary challenge [13]. Early studies using P. chabaudi-infected mice also implicated immune complexes of lipoproteins and IgG in inhibiting Ab secretion from ASCs in vitro [14]. Here, we show that loss of Influenza A virus-specific LLPCs is also dependent on FcγRIIB, as mice lacking this receptor did not lose pre-established HA-specific ASCs and their LLPCs did not undergo apoptosis following a P. chabaudi infection. Although, a cell-intrinsic role of FcγRIIB in LLPC loss, as previously demonstrated [13], was not investigated in this study, it is likely that the observed loss of LPPCs is brought about by FcγRIIB-dependent apoptosis, triggered by extensive hypergammaglobulinemia and the generation of immune complexes that occur during acute P. chabaudi infection [8], [34].
Surprisingly, in this study we found that Influenza A virus-specific Abs eventually returned to the levels observed before the P. chabaudi infections, suggesting that specific LLPCs were being replenished despite the fact that the Influenza A virus infection had been eliminated several months previously. One explanation for this is that Influenza A virus-specific B cells have been reactivated, differentiating into PCs and repopulating the Influenza A virus-specific LLPC niche. Indeed, we could show that B cell depletion in Influenza A virus-immune hCD20-transgenic mice resulted in specific depletion of Influenza A virus-specific B cells and eventual loss of LLPCs and Influenza A virus-neutralizing Abs. In contrast, the total IgG ASC population in the bone marrow, the major LLPC reservoir, was not affected by anti-hCD20 mediated depletion, indicating that the eventual loss of HA-specific ASCs and HA-specific IgG was due to a depleted HA-specific B cell pool. Hence our data strongly support a mechanism whereby Influenza A virus-specific B cells contribute to the continuous replenishment of Influenza A virus-specific bone marrow plasma cells and serum Abs.
Generation of Influenza A virus-specific ASCs may also originate either from the recruitment of new naïve B cells into a chronic response, an ongoing low-level GC reaction or reactivation of MBCs. However, the GC B cell population was unaffected during B cell depletion in hCD20 transgenic mice, and the naïve B cell population, as well as the total MBC population was relatively quickly restored following cessation of anti-hCD20 treatment. These two populations could in principle restore Influenza A virus-specific ASCs. However, Influenza A virus-specific serum Abs showed very little recovery following anti-hCD20 treatment, as did the Influenza A virus-specific MBC population. This observation strongly suggests that the continuous seeding of the LLPC pool and replenishment following P. chabaudi infection are mediated by MBCs.
MBCs have a variety of properties, which enable them to maintain serum Abs even in the absence of re-infection. MBCs can survive independently of antigen stimulation and in the absence of mitosis [35], are intrinsically programmed for faster signaling and self-renewal [36], and have been documented to re-circulate for up to 90 years after the last known infection [2], [37]. Furthermore, they are unique from other B cell subsets and LLPCs in their independence of the cytokines BAFF and APRIL for their survival [38] and have their own specialized niches like the spleen [39], although the properties of these niches are not well characterized.
While MBCs do not spontaneously differentiate into ASCs, MBCs have a higher propensity to differentiate into PCs than naïve B cells upon activation [21], [40]. MBCs differentiate into plasma cells upon antigenic stimulation and they have the potential to react to a wider range of pathogenic epitopes than the Abs produced by LLPCs, due to their lower-affinity, more polyreactive B cell receptors (BCRs) [41], [42], meaning that both homologous antigen and cross-reactive stimulation can stimulate the MBC B cell receptor. In addition, human and mouse MBCs can differentiate in vitro into plasma cells upon non-BCR-mediated, non-specific Toll-like receptor (TLR) stimulation [22], [40]. Therefore there are a number of ways in which MBCs can maintain serum Abs. Over time, MBCs can continually differentiate into ASCs whenever the host encounters homologous re-infection, cross-reactive heterologous infections, or in any inflammatory context with TLR ligands and bystander T cell help, and thus frequently boost serum Ab titers, and/or replenish the LLPC niche [22].
Although antigen-independent MBC conversion to LLPC has been suggested by studies in humans and in vitro [22], evidence that non-specific TLR or cytokine-mediated reactivation of MBC occurs in vivo in mice is currently lacking [40]. In contrast, mouse studies have demonstrated that persistence of viral antigens following Influenza A virus infection can be very long, spanning weeks or months [43]. It is therefore likely that reactivation of Influenza A virus-specific MBCs and continuous generation of LLPCs is a consequence of antigen retention. Further supporting this notion, MBCs have been shown to reconstitute humoral immunity to cytomegalovirus upon adoptive transfer into re-challenged, but not into antigen-free recipients [44], [45], and MBC reactivation in the lung of Influenza A virus re-challenged mice requires the presence of intact viral particles [46]. Therefore, both antigen retention and continuous generation of virus-specific ASCs may be seen as part of a robust mechanism to ensure both long-term maintenance of Influenza A virus-neutralizing serum Abs and recovery following episodes of attrition.
These reports indicate that there is a strong biological basis for the importance of MBCs, not just in the anamnestic response, but also in the general maintenance of long-lived serum Abs in the absence of re-infection. However these findings are seemingly at odds with the notion that serum Abs are maintained only by LLPCs, which was inferred by previous B cell depletion studies in mice [25], [26]. The latter studies suggested that the LLPCs generated by protein immunization and acute infection with LCMV (Armstrong) [3] are capable of surviving and maintaining pre-established serum Ab titers for long periods of time, despite MBC depletion by monoclonal Abs or irradiation. Our findings suggest that maintenance of long-term humoral immunity exclusively by LLPCs, in the absence of input from B cells, might not be a universal feature of all viral infections. The degree of reliance of the LLPC pool and consequently of protective immunity on continuous conversion of B cells into LLPCs is likely dependent on the size of the virus-specific LLPC pool. A large LLPC pool will require less B cell input before its size is reduced below the critical threshold for protection. In contrast, a small LLPC pool will not be able to resist attrition without continuous B cell input.
Our data have implications for the longevity of protective efficacy of vaccinations in malaria-endemic countries. Field data in humans on the impact of malaria infection on pre-existing immunity are relatively limited. The efficacy of childhood vaccination is reduced in malaria-endemic countries such as Nigeria or The Gambia as compared to non-endemic areas [47], [48], although the precise underlying reasons are not clear. Furthermore, there are documented outbreaks of infectious diseases, such as polio, despite a high level of vaccination coverage in the same areas [49]. However, as the majority of vaccine efficacy studies are carried out in very young infants and children, they might be confounded by a number of factors, such as the immaturity of the immune system [50] and high pre-existing titers of maternal Abs, which inhibit the development of MBCs and LLPCs. It has been documented that a co-infection with P. falciparum (i.e. detectable parasitaemia) suppresses the development of vaccine-induced immune responses [47], [51], although there are also reports that overall vaccine-induced immunity has not been affected in malaria-endemic countries [52]–[55]. Similar studies in farm animals have shown that infection with African trypanosomes significantly reduced the efficacy of several commercial vaccines [56]–[60]; however almost all these studies were done with vaccinations given during the time of Trypanosome infection and so do not provide answers to whether a parasite infection caused a loss of pre-established immunity. There is a dearth of investigations into the longevity of pre-established immunity in older age groups living in or moving into malaria-endemic countries where the parasite may have the ability to abrogate pre-established immunity and render the host susceptible to secondary infection. This would be extremely relevant in the context of multiple vaccination programs in malaria-endemic countries.
All animal experiments were approved by the ethical committee of the NIMR, and conducted according to local guidelines and UK Home Office regulations under the Animals Scientific Procedures Act 1986 (ASPA) and the authority of Project Licenses PPL 80/2236 and PPL 80/2358.
Inbred C56BL/6 and BALB/c mice were originally obtained from the Jackson Laboratory (Bar Harbor, ME) and subsequently bred and maintained in a specific pathogen-free (SPF) unit at the MRC National Institute for Medical Research (NIMR, London) animal facilities for over 30 years. hCD20tg BALB/c [24], Rag2−/− BALB/c [61], FcγRIIB-deficient (FcγRIIB−/−) C56BL/6 mice [62] and FcγRI-, II- and III-deficient (FcγRI,II,III−/−) C56BL/6 mice [63] were backcrossed for at least 7 generations onto the NIMR inbred mice and used with age-matched BALB/c and C56BL/6 controls. Experiments were performed using 8–15 week old female mice of each strain. Animal experiments were performed in accordance with the UK National guidelines (Scientific Procedures) Act 1986 under license approved by the British Home Office and the NIMR Institute Ethical Review Panel.
For induction of non-lethal Influenza A virus infection, mice were infected by instillation of the A/Puerto Rico/8/34 strain of (H1N1) Influenza A virus (PR8) into their nasal cavities without anesthesia. For sub-lethal infection, mice were given light inhalation anesthesia with isoflorane before intranasal instillation with PR8 and allowed to recover. Infection with P. chabaudi chabaudi (AS) (P. chabaudi) was initiated 105 or 150 days after PR8 infection by intraperitoneal (i.p.) injection of 105 iRBCs. Parasitaemia was determined by examination of Giemsa-stained thin blood smears. Chloroquine (CQ) for injection was prepared fresh from chloroquine diphosphate salt (Sigma) for drug-mediated elimination of parasites [8]. The treatment protocol was 10 daily i. p. injections of 40 mg/kg of CQ dissolved in sterile 0.9% saline. The efficacy of the drug treatment was verified by sub-inoculation of blood from the drug treated mice into immune compromised RAG2−/− recipient mice and analyses of thin blood films from the RAG2−/− recipient mice for the next 10 days. The mAb recognizing hCD20, 2H7 [24], was used for B cell depletion. 2H7 was purified from hybridoma culture supernatants and was endotoxin tested with Pyrotell (Associates of Cape Cod) and found to be present at a level of 0.3–0.6 EU/1 ml. Mice were injected i. p. with 2 mg/week of 2H7 in sterile 0.9% saline for 2 weeks.
P. chabaudi (AS) parasites were cloned and maintained at the NIMR, London [64], from an original isolate provided by Professor David Walliker (University of Edinburgh). Cryopreserved parasite stabilates were used for initiating infections in the animals in the manner previously described [65]. Briefly, stabilates were thawed from liquid nitrogen, diluted 1∶1 with 0.9% saline and injected i. p. into BALB/c mice. These parasites were passaged up to four times in mice by i. p. injection of 106, 105 or 104 iRBC per mouse diluted in 100 µl of Kreb's glucose saline. The number of iRBC was calculated by determining the percentage of parasitaemia on thin blood films using 20% Giemsa stain (VWR) and assuming a RBC density of 2.5×109/ml in peripheral venous blood. Experimental mice were infected using iRBC taken from one of the passage mice before the peak of parasitaemia. Each experimental mouse received an i. p. injection of 105 iRBC diluted in 100 µl of Kreb's glucose saline.
Bromelain-digested PR8 haemagglutinin (HA) was prepared as previously described [66]. P. chabaudi-iRBC were solubilized in Triton X-100 and SDS buffer in the presence of protease inhibitors [67]. Proteins were resolved by electrophoresis through NuPAGE 4–12% acrylamide gels in MES buffer (Invitrogen) under reducing conditions. Markers were the broad-range pre-stained protein standard Seeblue2 (Invitrogen). Proteins were transferred onto Hybond C extra nitrocellulose membrane (Amersham Pharmacia), as described previously [68]. Specific proteins were detected using (i) unpurified sera from influenza immune mice or (ii) unpurified sera from P. chabaudi infected mice, followed by Alexa 680-conjugated goat anti-mouse IgG (1∶15,000) (Licor Biosciences) and revealed by scanning the membranes with the Odyssey scanner (Licor Biosciences) using 680EX nm/700EM nm filter settings. Plasma from uninfected mice was used as a negative control.
For surface staining, 5×106 erythrolyzed and washed cells in 50 µl of FACS buffer (PBS, 2% FCS, 0.05% NaN3) were stained with prepared antibody multi-mixes (obtained from eBioscience, BD Pharmingen or BioLegend) in the presence of Fc receptor blocking antibody (clone 24G.2) to prevent non-specific binding of Fc receptors. Cells were incubated at 4°C for 40 minutes. Stained cells were washed two times with 200 µl of FACS buffer. For biotinylated antibodies, streptavidin-conjugated fluorochromes were added at a dilution of 1/100 and further incubated for 10 minutes at room temperature and then washed 3 times with FACS buffer. Cells were acquired using the CyAn ADP flow cytometer (Beckman Coulter) within 2 hours. Data were analyzed using FlowJo (Tree Star, Inc). Flowjo was used for graphical representation. For Annexin V staining, after surface staining was completed, cells were washed 2 times and resuspended in Annexin V Binding Buffer (BioLegend) at a concentration of 1×106 cells/ml. 100 µl of the cell suspension was filtered through a 0.2 µm filter into polypropylene FACS tubes (BD) and 5 µl of Annexin V-Pacific Blue (BioLegend) was added. Cells were incubated for 15 minutes at room temperature in the dark. 400 µl of Annexin V Binding Buffer was added to the tube just prior to analysis.
Serum titers of PR8 neutralizing antibodies were measured as previously described [16]. Sera were collected at indicated time points after PR8 infection, heat-inactivated for 10 minutes at 56°C, and tested using a modified Madin-Darby canine kidney (MDCK)-based assay. Serial dilutions of the sera were added to monolayers of MDCK cells in 96-well plates, which subsequently were infected with a 95% tissue culture-infective dose of PR8. MDCK cell viability was measured with an Alamar blue-based assay 3 days after infection. Cultures were pulsed with Alamar blue for 1 to 2 h, and fluorescence was measured with a fluorescence plate reader (TECAN Safire2).
RNA extraction was performed according to the RNeasy mini kit protocol following the manufacturer's protocol (Qiagen cat: 74106). Total RNA was extracted from whole lung tissues using TRI reagent (Sigma-Aldrich) and subsequently was used for cDNA synthesis with the Omniscript reverse transcription (RT) kit (Qiagen). RNA (1 ng) was used as the template, and cDNA synthesis was initiated by a mixture of 1 µM random hexamers and 1 µM of a primer specific to a highly conserved region of the IAV matrix gene, as previously described [69], [70]. The following primers were used for the amplification of target transcripts: Hprt forward (5′-TTGTATACCTAATCATTATGCCGAG-3′) and reverse (5′-CATCTCGAGCAAGTCTTTCA-3′), IAV matrix forward (5′-AAGACCAATCCTGTCACCTCTGA-3′) and reverse (5′-CAAAGCGTCTACGCTGCAGTCC-3′). Reaction mixtures were incubated at 37°C for 1 h and terminated by incubating the mixture at 90°C for 5 minutes. Expression of mRNA was determined by quantitative reverse transcription-PCR (qRT-PCR) using a DNA master SYBR green I kit (Roche) and the ABI Prism 7000 detection system (Applied Biosystems). The primers used for the amplification of target transcripts are in the Supplemental Methods [70]. Samples were analyzed in duplicate. The housekeeping gene Hprt was used to normalize the critical threshold values for the genes of interest. Levels of IAV matrix mRNA are plotted as arbitrary units relative to Hprt mRNA levels.
HA-specific IgG or IgM serum antibodies at time points after infection with PR8 and P. chabaudi were quantified by ELISA using HA as coating antigen. Results are expressed as µg/ml using an anti-mouse IgG or IgM ELISA (Southern Biotech) and purified Ig (Sigma) as a standard to quantify the amounts of the different isotypes. Amounts of P. chabaudi-specific serum antibodies after P. chabaudi infection were quantified by ELISA using parasite lysate as a coating antigen using hyperimmune serum from multiply-infected mice as a standard. The amounts of P. chabaudi-specific Abs were expressed as arbitrary units (AU) as described [71].
The protocol for determination of the half-life of IgG in mice was adapted from a method previously described [1]. Mice were injected i.p. with 200 µg of anti-TNP mouse IgG2a monoclonal antibodies (Hy1.2; a kind gift of Dr. H–U Weltzien, Max-Planck-Institute for Immunobiology, Freiburg, DE) which was purified from hybridoma culture supernatants as described above for 2H7. Hy1.2 Abs were administered 24 hours or 60 days after infection with 105 iRBC, and into uninfected controls. The concentration of HY1.2 mIgG2a Abs in serum were quantified by ELISA coating with 25 ng/well of TNP-BSA (Biosearch Technologies; diluted in PBS) using purified mouse IgG2a (Sigma) as a standard. Linear regression was used to find the relationship between the logarithm of serum antibody concentration and time since injection. Antibody half-life was then determined using the equation t1/2 = (ln 2)/k where k is the decay constant given by the slope of the best fitting linear function.
HA-specific plasma cells were quantified by a direct ex vivo ELISpot assay as previously described for other antigens [5]. 96-well Multi-screen HA Nitrocellulose filtration plates (Millipore) were coated with 50 µl of 10 µg/ml bromelain-digested PR8 HA diluted in PBS. As a positive control for total IgG secreting cells, some wells on each plate were coated with goat anti-mouse IgG (Invitrogen). The plates were incubated at 4°C overnight, washed twice in PBS, then blocked with 200 µl complete Iscove's medium for 1 h at room temperature. The plates were then washed twice with PBS and cell suspensions added at the following numbers: 1×106, 5×105, 2.5×105 and 1.25×105 per well in 200 µl complete Iscove's medium. The plates were incubated at 37°C, 7% CO2 for 5 h, then washed four times in PBS and four times with PBS with 0.1% Tween (PBS-T). 100 µl of goat anti-mouse IgG biotin conjugated antibody (Invitrogen) diluted 1∶1000 in PBS-T containing 1% FCS was added and the plates incubated overnight at 4°C. Plates were washed four times with PBS-T, and 100 µl of streptavidin-alkaline-phosphatase (BD Pharmingen) diluted 1∶8000 in PBS-T containing 1% FCS added and incubated for 1 h in the dark at room temperature, followed by four washes with PBS-T and four washes with PBS. Detection was carried out by adding 100 µl of BCIP/NBT substrate (BioFX) and incubating in the dark until blue spots appeared. The reaction was stopped by thorough washing with cold tap water and air-dried. Plates were analyzed using the ImmunoSpot reader (CTL).
HA-specific memory B cells were quantified using a limiting dilution ELISpot technique adapted from [72]. Briefly, 21 replicates of two-fold dilutions of cell suspensions of spleen starting with 1×106 cells per well were made on flat-bottomed 96-well plates (Costar) and cultured for 6 days in 200 µl/well complete IMDM containing a ‘stimulant mastermix’ of 0.4 µg R595 lipopolysaccharide (Alexis Biochemicals), 106 irradiated (1,200 rad) naive splenocytes and 20 µl Concanavalin A supernatant [73] per well. After 6 days, cells were washed in complete IMDM containing 1% FCS, harvested and transferred to pre-coated 96-well Multi-screen Nitrocellulose filtration plates. An ex-vivo ELISpot assay for HA-specific and total IgG plasma cell detection performed as described above. Frequencies were determined from the zero-order term of the Poisson distribution, using the Microsoft Excel Trendline option, a straight line of best fit was plotted and values were accepted when r2 values were greater than 0.7.
Statistical analysis was performed using GraphPad Prism 5 software. Continuous data, determined to be approximately normally distributed according to the Kolmogorov-Smirnov test (P>0.10), were analyzed using two-sided unpaired Student's t test. Where data were not normally distributed, data were analyzed using non-parametric Mann Whitney test. P values≤0.05 were considered significant.
|
10.1371/journal.ppat.1000448 | Fitness of Escherichia coli during Urinary Tract Infection Requires Gluconeogenesis and the TCA Cycle | Microbial pathogenesis studies traditionally encompass dissection of virulence properties such as the bacterium's ability to elaborate toxins, adhere to and invade host cells, cause tissue damage, or otherwise disrupt normal host immune and cellular functions. In contrast, bacterial metabolism during infection has only been recently appreciated to contribute to persistence as much as their virulence properties. In this study, we used comparative proteomics to investigate the expression of uropathogenic Escherichia coli (UPEC) cytoplasmic proteins during growth in the urinary tract environment and systematic disruption of central metabolic pathways to better understand bacterial metabolism during infection. Using two-dimensional fluorescence difference in gel electrophoresis (2D-DIGE) and tandem mass spectrometry, it was found that UPEC differentially expresses 84 cytoplasmic proteins between growth in LB medium and growth in human urine (P<0.005). Proteins induced during growth in urine included those involved in the import of short peptides and enzymes required for the transport and catabolism of sialic acid, gluconate, and the pentose sugars xylose and arabinose. Proteins required for the biosynthesis of arginine and serine along with the enzyme agmatinase that is used to produce the polyamine putrescine were also up-regulated in urine. To complement these data, we constructed mutants in these genes and created mutants defective in each central metabolic pathway and tested the relative fitness of these UPEC mutants in vivo in an infection model. Import of peptides, gluconeogenesis, and the tricarboxylic acid cycle are required for E. coli fitness during urinary tract infection while glycolysis, both the non-oxidative and oxidative branches of the pentose phosphate pathway, and the Entner-Doudoroff pathway were dispensable in vivo. These findings suggest that peptides and amino acids are the primary carbon source for E. coli during infection of the urinary tract. Because anaplerosis, or using central pathways to replenish metabolic intermediates, is required for UPEC fitness in vivo, we propose that central metabolic pathways of bacteria could be considered critical components of virulence for pathogenic microbes.
| Bacteria that cause infections often have genes known as virulence factors that are required for bacteria to cause disease. Studying virulence factors such as toxins, adhesins, and secretion and iron-acquisition systems is a fundamental part of understanding infectious disease mechanisms. In contrast, little is known about the contribution of bacterial metabolism to infectious disease. This study shows that E. coli, which cause most urinary tract infections, utilize peptides as a preferred carbon source in vivo and requires some, but not all, of the central metabolic pathways to infect the urinary tract. Specifically, pathways that can be used to replenish metabolites, known as anaplerotic reactions, are important for uropathogenic E. coli infections. These findings help explain how metabolism can contribute to the ability of bacteria to cause a common infection.
| Traditional studies of bacterial pathogenesis have focused on pathogen-specific virulence properties including toxins, adhesins, secretion, and iron acquisition systems, and mechanisms to avoid the innate and adaptive immune response. Examining bacterial metabolism during the course of an infection is also critical to further our understanding of pathogenesis and identifying potential targets for new antimicrobial agents. Infectious diseases represent a serious threat to global health because many bacteria that cause disease in humans such as Staphylococcus aureus, Mycobacterium tuberculosis, and E. coli are steadily developing resistance to many of the available treatments [1]–[3]. Since the introduction of antibiotics in the last century, the emergence of bacteria that resist these compounds has rapidly outpaced the discovery and development of new antimicrobial agents [4]. The need to understand bacterial physiology during infection of the host is critical for the development of new antimicrobials or antibiotics that will reduce their burden upon human health.
Among common infections, urinary tract infections (UTI) are the most frequently diagnosed urologic disease. The majority of UTIs are caused by E. coli and these uropathogenic E. coli (UPEC) infections place a significant financial burden on the healthcare system by generating annual costs in excess of two billion dollars [5],[6]. Because UTIs are a significant healthcare burden and E. coli is one of the best studied model organisms for studying metabolism, these traits can be exploited to understand and identify metabolic pathways that are required for the growth of the bacterium during infection of the host.
Despite being arguably the most studied organism, E. coli metabolism during colonization of the intestine has only recently been explored [7],[8]. Commensal E. coli acquires nutrients from intestinal mucus, a complex mixture of glycoconjugates, and subsequently expresses genes involved in the catabolism of N-acetylglucosamine, sialic acid, glucosamine, gluconate, arabinose and fucose [8],[9]. E. coli mutants in the Entner-Doudoroff and glycolytic central metabolic pathways have diminished colonization levels reflecting the importance of sugar acid catabolism [8]. These findings suggest that commensal E. coli uses multiple limiting sugars for growth in the intestine [8]. Together, this developing body of evidence supports the assertion that E. coli grows in the intestine using simple sugars released by the breakdown of complex polysaccharides by anaerobes [9],[10].
Much less is known about the metabolism of enteric pathogens during colonization of the gastrointestinal tract. Enterohemorrhagic E. coli (EHEC) O157∶H7 requires similar carbon metabolic pathways as do commensal strains, however, mutations in pathways that utilize galactose, hexuronates, mannose, and ribose resulted in colonization defects only for EHEC [9]. It was also found that multiple mutations in a single EHEC strain had an additive effect on colonization levels suggesting that this pathogen depends on the simultaneous metabolism of up to six sugars to support the colonization of the intestine [9]. When faced with limiting sugars due to consumption by other colonizing bacteria, EHEC may switch from glycolytic to gluconeogenic substrates to sustain growth in the intestine [11]. Synthesis and degradation of glycogen, an endogenous glucose polymer, plays an important role for EHEC and pathogenic Salmonella during colonization of the mouse intestine presumably by functioning as an internal carbon source during nutrient limitation [12]–[14]. Although it is not known which external carbon sources are used by S. enterica serovar Typhimurium during colonization it has been demonstrated that full virulence requires the conversion of succinate to fumarate in the tricarboxylic acid (TCA) cycle [15],[16]. These studies have contributed much to the understanding of the in vivo metabolic requirements of EHEC colonization; however, these studies were done in an animal model that is not suitable for studying pathogenesis because these animals do not exhibit signs of EHEC infection [9],[11],[13].
In contrast to the nutritionally diverse intestine, the urinary tract is a high-osmolarity, moderately oxygenated, iron-limited environment that contains mostly amino acids and small peptides [17],[18]. The available studies on UPEC metabolism during UTI has revealed that the ability to catabolize the amino acid D-serine in urine, which not only supports UPEC growth, appears important as a signaling mechanism to trigger virulence gene expression [19],[20]. Metabolism of nucleobases has been demonstrated to play a role for UPEC colonization of the urinary tract; signature-tagged mutagenesis screening identified a mutant in the dihydroorotate dehydrogenase gene pyrD that was outcompeted by wild-type UPEC in vivo [21] and in a separate transposon screen a gene involved in guanine biosynthesis, guaA, was identified and found to be attenuated during experimental UTI [22].
To better understand bacterial metabolism during infection, we used a combination of comparative proteomics and systematic disruption of central metabolism to identify pathways that are required for UPEC fitness in vivo. By examining the expression of UPEC cytoplasmic proteins during growth in human urine, we confirmed that E. coli is scavenging amino acids and peptides and found that disruption of peptide import in UPEC significantly compromised fitness during infection. Consistent with the notion that peptides are a key in vivo carbon source for UPEC, only mutations ablating gluconeogenesis and the TCA cycle demonstrated reduced fitness in vivo during experimental UTI. These findings represent the first study of pathogenic E. coli central metabolism in an infection model and further our understanding of the role of metabolism in bacterial pathogenesis.
Culturing UPEC in human urine partially mimics the urinary tract environment and has proven to be a useful tool to identify bacterial genes and proteins involved in UTI [18], [22]–[24]. Because it is well established that urine is iron-limited and our previous studies clearly demonstrated that the majority of differentially expressed genes and proteins are involved in iron acquisition [18],[23], we determined the protein expression profile of E. coli CFT073 during growth in human urine and compared that with bacterial cells cultured in iron-limited LB medium to unmask proteins involved in processes other than iron metabolism. Using this strategy and 2D-DIGE it was possible to visualize 700 cytoplasmic protein spots, 84 of which were differentially expressed (P<0.05) between urine and iron-limited LB medium (Fig. 1). Of these, 56 were more highly expressed in human urine (green) than in iron-limited LB medium, while 28 demonstrated greater expression in iron-limited LB medium (red) than in urine (Fig. 1).
Proteins induced in human urine with >2-fold differences from expression levels in iron-limited LB medium were identified by tandem mass spectroscopy (Table 1). The results indicate that E. coli growing in urine are expressing proteins involved in the catabolism of pentose sugars; XylA (xylose isomerase), AraF (high-affinity arabinose-binding protein), and the non-oxidative pentose phosphate pathway enzyme TalA (transaldolase) were induced 5.25-, 2.02-, and 5.66-fold, respectively (P<0.001) (Table 1). Other proteins that were induced are the involved in metabolism of the sugar acids gluconate (UxuA, mannonate dehydratase), gluconolactone (YbhE, 6-phosphogluconolactonase), sialic acid (NanA, N-acetylneuraminate lyase), and fructose (FruB, fructose-specific IIA/FPr PTS system component). Multiple isoforms of the periplasmic dipeptide and oligopeptide substrate-binding proteins DppA and OppA were also induced (>2-fold, P<0.009) in urine confirming the notion that amino acids and small peptides are being acquired from this milieu (Table 1). Proteins involved in amino acid metabolism were also identified and include SerA (D-3-phosphoglycerate dehydrogenase) that is involved in serine biosynthesis and two enzymes in the arginine biosynthesis pathway, ArgG (argininosuccinate dehydrogenase) and SpeB (agmatinase) (Table 1). As expected, none of the proteins identified were involved in iron uptake or metabolism, although DppA has been reported to bind heme albeit with less affinity than dipeptide substrates [25].
Notably, there was an increase in abundance for two central metabolism enzymes, TalA, as mentioned above, and TpiA that was increased 4.58-fold (P<0.0001) in urine (Table 1). TalA, a non-oxidative pentose phosphate pathway enzyme, converts sedoheptulose-7-phosphate and glyceraldehyde-3-phosphate to erythrose-4-phosphate and fructose-6-phosphate. Due to the transfer of the glycolytic intermediate glyceraldehyde-3-phosphate by TalA, this enzyme is an important link between the pentose phosphate pathway and glycolysis [26]. TpiA is a glycolytic enzyme that catalyzes the reversible isomerization of glyceraldehyde-3-phosphate and dihydroxyacetone phosphate [27]. The induction of TalA and TpiA suggested that the coupling of the pentose phosphate pathway and glycolysis or gluconeogenesis via the transfer and isomerization of glyceraldehyde-3-phosphate may be an important route of carbon flux through these central pathways during the bacterium's growth in human urine.
To determine whether some proteins identified by 2D-DIGE are required for UPEC fitness during UTI, CFT073 mutants were constructed in the genes: talA, xylA, tpiA, serA, speB, uxuA, nanA, argG, araF, dppA, and oppA. For these studies, an experimental competition between each mutant strain and wild-type parental CFT073 was performed. Wild-type UPEC and the mutant strain were prepared in a 1∶1 ratio and transurethrally inoculated into the bladders of mice. The number of mutant (kanamycin-resistant) and wild-type (kanamycin-sensitive) bacteria recovered from the bladder and kidneys was determined by plating the tissue homogenates for CFU on both LB agar and LB agar containing kanamycin. Mutants containing defects in genes that affect fitness in vivo are out-competed by the wild-type strain when inoculated into the same animal. This was determined by comparing the ratio of colony forming units (CFU) of bacteria recovered from the infection to the ratio of bacteria contained within the inoculum to obtain a competitive index (CI). A CI>1 indicates the wild-type out-competes the mutant strain and a CI<1 indicates the wild-type is out-competed by the mutant. In these series of experimental infections, only mutants defective in peptide transport (ΔdppA and ΔoppA) were dramatically out-competed by wild-type UPEC in vivo, CI>50, P<0.005 for the bladder (Table 2). One additional mutant, ΔtpiA, that functions in both glycolysis and gluconeogenesis, was out-competed by wild-type in the kidneys at 48 hpi, CI = 2.54, P = 0.0206 (Table 2).
Despite the lack of attenuation in vivo for the many of the mutants, these results reveal a number of important findings. The agmatinase mutant ΔspeB out-competed wild-type in the bladder at 48 hpi, CI = 0.14, P = 0.0122 (Table 2). Agmatinase is part of arginine metabolism and catalyzes the formation of the polyamine putrescine and urea from agmatine and H2O. This suggests that accumulation of agmatine or reduced production of urea and putrescine by the mutant may provide a modest advantage over wild-type UPEC during infection of the bladder. CFT073 ΔargG was unable to grow in MOPS defined medium unless supplemented with 10 mM arginine (Fig. 2A), validating the expected auxotrophic phenotype. Similarly, the ΔserA serine auxotroph required supplementation with either 10 mM serine or glycine in MOPS, D-serine was unable to rescue the in vitro growth defect (Fig. 2B). Lack of arginine or serine biosynthesis had little effect upon the ability of UPEC to grow logarithmically in human urine, although the ΔargG mutant consistently entered stationary phase at a lower cell density, with an O.D.600 of 0.45±0.04 compared to 0.59±0.03 for wild-type (P = 0.051) (Fig. 2C). When tested for in vivo fitness, neither the ΔargG nor ΔserA strain were significantly out-competed by wild-type UPEC at 48 hpi (Fig. 2D, 2E, and Table 2). Additionally, there was no preference for serine over arginine or vice versa for UPEC colonization at 48 hpi. When the auxotrophic strains were co-inoculated into the same mice both mutants were recovered at similar levels (Fig. 2F). These data clearly demonstrate that there are sufficient concentrations of arginine, serine and/or glycine in the urinary tract to support growth of these auxotrophic strains.
As mentioned, deletion of the genes encoding periplasmic peptide substrate-binding proteins, dppA and oppA, had the greatest impact on UPEC fitness in vivo of the CFT073 mutants in genes whose products were induced during growth in human urine (Table 2). The dipeptide transport mutant, ΔdppA, failed to maintain colonization in the bladder at 48 hpi, 11/11 bladders had undetectable levels (<200 CFU/g) for this mutant, while wild-type levels from the same bladders reached a median of 104 CFU/g (P = 0.0020) (Fig. 3A). Because these mice had low levels of recoverable UPEC from the kidneys it was not possible to determine the contribution of dipeptide transport for kidney colonization. Import of oligopeptides via the OppA substrate-binding protein is also required for UPEC fitness in vivo. CFT073 ΔoppA was out-competed nearly 500∶1 wild-type∶mutant in the bladder (Table 2) with a 3-log reduction in the median CFU/g from bladder tissue at 48 hpi (P = 0.0047) (Fig. 3B). In these co-challenge infections, wild-type UPEC colonized 10/16 (62%) of kidneys, while ΔoppA was detectable in 4/16 (25%) of kidneys at 48 hpi. The ratio of wild-type∶mutant recovered from the kidneys at this time point was 156∶1 (Table 2) where wild-type UPEC had 3-logs greater CFU/g than ΔoppA (P = 0.0420) (Fig. 3B). Together, the in vivo fitness defect for CFT073 harboring a deletion of either dppA or oppA suggests that peptides may be an important carbon source for UPEC during urinary tract infection.
Previously, we have shown that the low copy pGEN plasmid is maintained in CFT073 in the absence of antibiotic pressure for up to 48 h [28]. Using this ampicillin resistant plasmid system, we cloned the entire dppA gene including 200 bp upstream from the predicted start site of translation and introduced the resulting construct, pGEN-dppA, into the CFT073 ΔdppA strain. To determine if it was possible to complement the ΔdppA defect in vivo, co-challenge infections were performed as described and modified to enumerate bacteria in tissue homogenates by plating on agar containing ampicillin (wild-type CFT073 harboring pGEN) or ampicillin and kanamycin (CFT073 ΔdppA containing pGEN or pGEN-dppA). The ΔdppA mutant containing empty vector (pGEN-) demonstrated the expected fitness defect in bladder colonization when co-inoculated with wild-type CFT073 (pGEN-) (P = 0.0002) while ΔdppA containing a wild-type copy of dppA (pGEN-dppA) restored colonization to wild-type levels in the bladder at 48 hpi (Fig. 3C). Although both mutant (pGEN-) and wild-type (pGEN-) demonstrated poor colonization in the kidneys of these animals, complementation of ΔdppA (pGEN-dppA) resulted in a 2-log increase in median kidney CFU/g at 48 hpi (Fig. 3D).
The requirement for peptide transport for UPEC fitness during infection implicates peptides as an important carbon source in vivo. This predicts that certain central metabolism pathways that operate during catabolism of amino acids or peptides may be more important for in vivo growth of UPEC than pathways that function primarily to catabolize sugars. To test the role of central metabolic pathways during an actual infection mutants were constructed in UPEC strain CFT073 to produce defects in glycolysis (pgi, phosphoglucose isomerase and tpiA, triosephosphate isomerase) [29], the Entner-Doudoroff pathway (edd, 6-phosphogluconate dehydratase) [10], the oxidative branch (gnd, 6-phosphogluconate dehydrogenase) and the non-oxidative branch (talA, transaldolase) of the pentose phosphate pathway [26], gluconeogenesis (pckA, phosphoenolpyruvate carboxykinase) [30], and the TCA cycle (sdhB, succinate dehydrogenase) [31]. The in vitro growth of these central metabolism mutants were examined and compared to wild-type UPEC during culture in human urine, LB medium, and MOPS defined medium containing 0.02% glucose. All of the central metabolism mutants produced similar logarithmic growth as wild-type when cultured in human urine (Fig. 4A) and LB medium (data not shown) under defined inoculation conditions. As expected, only mutants with defects in glycolysis demonstrated diminished growth in MOPS medium containing glucose as the sole carbon source (Fig. 4B). The Δpgi strain produced an extended lag phase of 5.5±1.1 h compared with wild-type (P = 0.001) and ΔtpiA failed to reach exponential phase after 18 h (Fig. 4B). These data and the indistinguishable growth of the glycolysis mutants from wild-type in urine supported the proteomics data and indicated that UPEC growing in urine utilizes carbon sources other than glucose.
To determine the role for central metabolism during E. coli infection of the urinary tract, the ascending model of murine UTI was used as described above to measure the impact that a lesion in central metabolism has upon the relative fitness of the strain in vivo. Mutants with defects in glycolysis had levels of colonization in the bladder at 48 hpi similar to wild-type (P>0.400) (Fig. 5A and 5B). In the kidneys, Δpgi CFU/g were comparable to wild-type (Fig. 5A), while ΔtpiA demonstrated a 10-fold reduction in the median CFU/g (P = 0.0206) (Fig. 5B). The pentose phosphate pathway mutants, Δgnd (Fig. 5C) and ΔtalA (Table 2), were not significantly out-competed by wild-type in vivo. The mutant with a defect in the Entner-Doudoroff pathway (Δedd) also was not impaired in the ability to infect both the bladder and kidneys as indicated by its similar colonization to wild-type at 48 hpi (Fig. 5D). UPEC in vivo fitness was significantly reduced in the TCA cycle mutant ΔsdhB, this mutation resulted in a 50-fold reduction in median CFU/g in the bladder (P = 0.0134) and a 1.5-log decrease in kidney CFU at 48 hpi (P = 0.0400) (Fig. 5E). This defect in the TCA cycle impacted fitness to a greater extent in the bladder, where 11/15 (73%) of mice had undetectable levels of mutant bacteria, than in the kidneys where 6/15 (40%) mice had undetectable counts (Fig. 5E). The gluconeogenesis mutant, ΔpckA had a 2-log reduction in median CFU/g in both the bladder (P = 0.0005) and kidneys (P = 0.0322) and half of the mice (7/14) displayed undetectable levels of ΔpckA at 48 hpi (Fig. 5F).
To verify that this mutation is non-polar as expected and the defect in colonization is not due to a secondary mutation, in vivo complementation experiments were conducted. The ΔpckA mutant with the pGEN empty vector demonstrated a 2-log reduction in CFU/g at 48 hpi (P = 0.0039) in the bladder when co-inoculated into mice with wild-type UPEC containing pGEN (Fig. 6). When CFT073 ΔpckA (pGEN-pckA) were co-inoculated with CFT073 (pGEN-) there was no significant difference in bladder CFU/g at 48 hpi between the strains (Fig. 6). Thus, by re-introducing the pckA gene into the mutant it was possible to complement the ΔpckA defect in bladder colonization at 48 hpi.
The in vitro growth and in vivo fitness for the UPEC central metabolism mutants is summarized in Table 3. As expected, only mutations in glycolysis had a negative effect on growth in defined medium with glucose. Only gluconeogenesis or TCA cycle mutants demonstrated reduced persistence at 48 hpi in both the bladder and kidneys (Table 3). Non-oxidative and oxidative pentose phosphate pathway and Entner-Doudoroff pathway mutants did not demonstrate any colonization defect and of the glycolytic mutants only the triosephosphate isomerase deletion had a measurable defect in the kidneys but not in the bladder (Table 3). Together, the fitness defect for the peptide transport mutants and these data indicate UPEC could be using amino acids as the primary carbon source during infection. Surprisingly, there was no correlation between the ability of the central metabolism mutants to grow in human urine ex vivo and grow in the urinary tract in vivo.
Bacterial pathogenesis traditionally involves studying virulence traits involved in the production of toxins and effectors, iron acquisition, adherence, invasion, and immune system avoidance. Although many paradigms exist that describe mechanisms of pathogenesis, the contribution of microbial metabolism to bacterial virulence during an infection is less understood. Much work has been done studying E. coli as model organism for characterizing individual central metabolism pathways and enzymes [10], [27], [32]–[38]. We have shown here that central metabolism studies in E. coli can be extended to investigate the contribution of central pathways to bacterial pathogenesis using a virulent uropathogenic E. coli strain and a well-established animal model of UTI. It is known that commensal E. coli require the Entner-Doudoroff pathway and glycolysis for colonization in vivo; while the TCA cycle, pentose phosphate pathway, and gluconeogenesis are dispensable in the intestine [8]. In contrast, we have shown that during E. coli infection of the urinary tract, the pathways required for commensal colonization are dispensable while the TCA cycle and gluconeogenesis are necessary for UPEC fitness in vivo. Adaptation to distinct host environments has been previously shown to involve shared traits between commensal and pathogenic strains [39],[40]. Because commensal E. coli are an important natural component of the intestine one concern faced when developing antimicrobials that target pathogenic strains is how to avoid eradicating commensal bacteria. Thus, these findings highlight important differences between commensal and pathogenic E. coli that could be exploited for the development of antimicrobials that target these pathways in this pathogen during infections that may not affect commensal strains. Interestingly, in addition to UPEC, gluconeogenesis is required for virulence in microbes that represent an array of pathogenic lifestyles, from intracellular bacteria and parasites [41],[42], plant-pathogenic [43], and intestinal pathogens [16]; suggesting that anaplerosis may be a common mechanism of microbial pathogenesis.
This study comprehensively examines the role of pathogenic E. coli central metabolism in a disease model and provides insight not only into UPEC metabolism in vivo but also information regarding the nutrients available to support the growth of E. coli within the urinary tract. The proteomics experiments did reveal that UPEC growing in human urine induces expression of multiple isoforms of both dipeptide- and oligopeptide-binding proteins, both of which were found to be required for UPEC to effectively colonize the urinary tract. This indicates that these bacteria actively import short peptides in urine and this function may indicate that peptides are an important carbon source in vivo. Consistent with this, only bacteria with defects in peptide transport, gluconeogenesis, or the TCA cycle demonstrated a significant reduction in fitness in vivo in both the bladder and kidneys. These findings suggest a model that describes the biochemistry of E. coli during UTI. For optimal growth during infection, short peptides are taken up by UPEC and degraded into amino acids that are catabolized and used in a series of anaplerotic reactions that replenish TCA cycle intermediates and generate gluconeogenesis substrates (Fig. 7).
Certain glycolytic steps are irreversible and the reverse gluconeogenic reaction is performed by an enzyme specific for gluconeogenesis. Carbon flux through glycolysis and gluconeogenesis must be carefully controlled by the cell to avoid a futile cycle of carbon metabolism [44]. Allosteric regulation of enzymes that catalyze irreversible reactions in these pathways and catabolite repression are mechanisms used to avoid the futile cycle [45],[46]. A gluconeogenic-specific enzyme subject to allosteric regulation is phophoenolpyruvate carboxykinase that converts oxaloacetate to phosphoenolpyruvate [47]. Deletion of the gene pckA that encodes this enzyme resulted in a significant reduction in UPEC fitness in vivo. Because bacteria prevent glycolysis and gluconeogenesis from occurring simultaneously and deletion of pckA reduced fitness in vivo, we reason that carbon flux through gluconeogenesis during UPEC infection may be an important indication of amino acid catabolism in vivo.
It is not surprising that, in addition to gluconeogenesis, the TCA cycle is also required for UPEC fitness in vivo. These two pathways are connected and collectively described as “filling in” or anaplerotic reactions. The TCA cycle is necessary to provide substrates for gluconeogenesis when cells use amino acids as a carbon source. Gluconeogenic amino acids can be degraded to oxaloacetate or to pyruvate that can be converted to acetyl-CoA and enter the TCA cycle [47]. Oxaloacetate, a TCA cycle intermediate, is converted to phophoenolpyruvate during gluconeogenesis by PckA as described above. A mutation in the TCA cycle enzyme succinate dehydrogenase, sdhB, results in a UPEC strain that has reduced fitness in vivo. This finding suggests that UPEC are growing aerobically in the urinary tract because succinate dehydrogenase is replaced by fumarate reductase during anaerobic growth and therefore, future work could confirm if the reductive TCA cycle is not operating during UPEC infection. The requirement for peptide import and the TCA cycle for UPEC fitness during infection is consistent with the hypothesis that acetyl-CoA production from the degradation of amino acids could be occurring in vivo as has been shown by another group [48].
Interestingly, with the exception of peptide-transport proteins, up-regulation of protein expression in urine ex vivo did not correlate with functional importance in vivo. This could be due to the fact that many central metabolism genes are constitutively expressed and that human urine only partially mimics the complex lifestyle of UPEC during UTI [49]. The absence of host cells and the immune response during growth in urine ex vivo could in part account for this discrepancy. It also remains possible that mutants that lack growth defects in urine but demonstrate reduced fitness in vivo could represent genes or metabolic pathways that are required for intracellular phases of growth during cystitis [50].
Despite these disadvantages, up-regulation of both DppA and OppA expression was seen in urine and loss of either dppA or oppA was found to negatively impact UPEC colonization in vivo. Induction of dppA has been reported in a hypervirulent UPEC strain that has a lacks a functional D-serine deaminase gene (dsdA) [51]. Deletion of dppA in this mutant strain resulted in a loss of the hypervirulent phenotype in vivo and significantly reduced its ability to colonize the urinary tract in competition with wild-type [51]. Surprisingly, in contrast to our findings, this group found that mutation of dppA alone had no effect on UPEC fitness in vivo [51]. Due to lack of complementation, it is unclear from that work why loss of dppA dramatically attenuated a hypervirulent strain but had no effect on wild-type. Despite this inconsistency in that work, the importance of peptide transport for UPEC fitness in vivo is supported by the findings that loss of either dppA or oppA significantly reduced colonization of the urinary tract and that the reduced bacterial colonization in the ΔdppA strain can be restored to wild-type levels by complementing the mutant with a wild-type dppA gene.
In summary, defects in the both branches of the pentose phosphate pathway, the Entner-Doudoroff pathway, and glycolysis had limited or no impact on UPEC fitness in vivo. On the other hand, the TCA cycle- and gluconeogenesis-defective strains demonstrate significant fitness reductions during UTI. The utilization of short peptides and amino acids as a carbon source during bacterial infection of the urinary tract is supported by the observation that UPEC mutants defective in peptide import have reduced fitness in vivo while auxotrophic strains do not. Together, these findings provide compelling evidence to support the notion that catabolism of amino acids to form TCA cycle intermediates and gluconeogenic substrates is important for the ability of UPEC to infect the urinary tract efficiently. This shows that anaplerotic and central metabolism pathways are required for UPEC fitness in vivo and suggest microbial metabolism should be considered important for bacterial pathogenesis.
Strains were derived from E. coli strain CFT073, a prototypic UPEC strain isolated from the blood and urine of a patient with acute pyelonephritis [52]; its genome has been sequenced and fully annotated [53]. Isolated colonies were used to inoculate overnight Luria-Bertani (LB) cultures. Bacteria from overnight cultures were collected by centrifugation, washed with sterile PBS, and 106 CFU were used to inoculate pre-warmed LB or human urine. To mimic iron-limitation in urine, LB containing 10 mM deferoxamine mesylate (Sigma) was used as a growth medium for comparative proteomics. For human urine cultures, mid-stream urine was collected into sterile sample containers from 8–10 male and female donors, pooled, and sterilized by vacuum filtration through a 0.22 µm pore filter. MOPS defined medium containing 0.2% glucose [54] with and without 10 mM L-arginine, L-serine, glycine, aspartatic acid, or D-serine (Sigma) was also used to test growth of mutant strains. Growth curves were established in triplicate using a Bioscreen bioanalyzer in 0.4 ml volumes; OD600 was recorded every 15 min. All cultures were incubated at 37°C; LB overnight and MOPS cultures were incubated with aeration; urine cultures were incubated statically. For preparation of proteins, UPEC isolate CFT073 was grown statically to exponential phase (OD600 = 0.25) in pre-warmed LB or human urine at 37°C in 5×100 ml cultures for each growth medium.
Bacteria were harvested from 500 ml of culture by centrifugation (10,000× g, 30 min, 4°C) and lysed in a French pressure cell at 20,000 psi. Harvested cells were washed and resuspended in 10 ml of 10 mM HEPES, pH 7.0 containing 100 U of Benzonase (Sigma). Following two passes through the chilled pressure cell, lysates were centrifuged (7500× g, 10 min, 4°C) to remove unbroken cells and supernatants were ultracentrifuged (120,000× g, 1 h, 4°C) to remove membranes and insoluble material. Soluble proteins were quantified using the 2D Quant Kit (GE Healthcare) following the manufacturer's protocol and either used immediately in DIGE-labeling procedures or stored at −80°C.
For fluorescence difference in gel electrophoresis (2D-DIGE) [55], bacterial proteins were minimally labeled with cyanine-derived fluors (CyDyes) containing an NHS ester-reactive group as recommended by the manufacturer (GE Healthcare). To determine quantitative differences within the UPEC soluble proteome during growth in human urine, cytoplasmic proteins prepared from human urine cultures were labeled with Cy3, from LB broth with Cy5, and a pooled internal standard representing equal amounts of both urine and LB preparations with Cy2 as described previously [23]. Briefly, 50 µg of protein was incubated with 400 pmol CyDye for 30 min and the reaction was stopped by added 10 mM lysine. Following labeling, samples labeled with each CyDye were pooled (150 µg total protein), mixed with an equal volume of 2× DIGE sample buffer; 7 M urea, 2 M thiourea, 10 mM tributylphosphine (TBP) (Sigma), 2× biolytes 3–10 (Bio-Rad), 2% ASB-14 and incubated on ice for 10 min. For rehydration, samples were brought to 0.35 ml with 1× DIGE rehydration buffer (7 M urea, 2 M thiourea, 5 mM TBP, 1× biolytes 3–10, 1% ASB-14) and used to passively rehydrate pH 4–7 IPG strips (Bio-Rad) overnight at room temperature. Rehydrated IPG strips were equilibrated and subjected to isoelectric focusing for 50,000 V·h and second dimension SDS-PAGE on 10% gels within low fluorescence glass plates (Jule Biotechnologies, Inc.) and were run at a constant current of 55 mA at 4°C for 4 hr. Following SDS-PAGE, image acquisition and pixel intensity was obtained using a Typhoon scanner (GE Healthcare) and differential in-gel analysis and biological analysis of variance were performed using the DeCyder 6.5 software suite (GE Healthcare). Using this software, the normalized spot volume ratios from Cy3 or Cy5 labeled spots were quantified relative to the Cy2-labeled internal standard from the same gel. The Cy2-labeled standard was then used to standardize and compare normalized volume ratios between the Cy3 and Cy5 labeled proteins between gels representing three independent experiments to generate statistical confidence for abundance changes using student's t-test and ANOVA. To identify the proteins, 500 µg of cytoplasmic proteins were focused as described above and spots of interest were excised from a colloidal Coomassie-stained 2D SDS-PAGE gel and subjected to enzymatic digestion with trypsin. Mass spectra were acquired on an Applied Biosystems 4700 Proteomics Analyzer (TOF/TOF). MS spectra were acquired from 800–3500 Da and the eight most intense peaks in each MS spectrum were selected for MS/MS analysis. Peptide identifications were obtained using GPS Explorer (v3.0, Applied Biosystems), which utilizes the MASCOT search engine. Each MS/MS spectrum was searched against NCBInr. Tryptic digestion and tandem mass spectrometry were performed at the University of Michigan Proteome Consortium.
Deletion mutants were generated using the lambda red recombinase system [56]. Primers homologous to sequences within the 5′ and 3′ ends of the target genes were designed and used to replace target genes with a nonpolar kanamycin resistance cassette derived from the template plasmid pKD4 [56]. Kanamycin (25 µg/ml) was used for selection of all mutant strains. Gene deletions begin with the start codon and end with the stop codon for each gene. To determine whether the kanamycin resistance cassette recombined within the target gene site, primers that flank the target gene sequence were designed and used for PCR. After amplification, each PCR product was compared to wild-type PCR product and in cases where size-differences are negligible; PCR products were digested with the restriction enzyme EagI (New England Biolabs). Both the PCR products and restriction digests were visualized on a 0.8% agarose gel stained with ethidium bromide. For in vivo complementation, the dppA and pckA genes were amplified from CFT073 genomic DNA using Easy-A high-fidelity polymerase (Stratagene) and independently cloned into pGEN-MCS [28],[57] using appropriate restriction enzymes. The sequences of pGEN-dppA and pGEN-pckA were verified by DNA sequence analysis prior to electroporation into CFT073 ΔdppA or ΔpckA mutant strains.
Six-to eight-week-old female CBA/J mice (20 to 22 g; Jackson Laboratories) were anesthetized with ketamine/xylazine and inoculated transurethrally over a 30 sec period with a 50 µl bacterial suspension per mouse using a sterile polyethylene catheter (I.D. 0.28 mm×O.D. 0.61 mm) connected to an infusion pump (Harvard Apparatus). To measure relative fitness, overnight LB cultures for CFT073 and the mutant strain were collected by centrifugation and resuspended in sterile PBS, mixed 1∶1 and adjusted to deliver 2×108 CFU per mouse. Dilutions of each inoculum were spiral plated onto LB with and without kanamycin using an Autoplate 4000 (Spiral Biotech) to determine the input CFU/mL. After 48 hpi, mice were sacrificed by overdose with isoflurane and the bladder and kidneys were aseptically removed, weighed, and homogenized in sterile culture tubes containing 3 ml of PBS using an OMNI mechanical homogenizer (OMNI International). Appropriate dilutions of the homogenized tissue were then spiral plated onto duplicate LB plates with and without kanamycin to determine the output CFU/g of tissue. Plate counts obtained on kanamycin were subtracted from those on plates lacking antibiotic to determine the number of wild-type bacteria. Competitive indices were calculated by dividing the ratio of wild-type to mutant at 48 hpi by the ratio of wild-type to mutant input CFU/mL. Groups of 5 mice per co-challenge were used to determine defects in fitness, when a defect was apparent the co-challenge was repeated two more times with groups of 5 mice. Statistically significant differences in colonization (P-value<0.05) were determined using a two-tailed Wilcoxon matched pairs test. All animal protocols were approved by the University Committee on Use and Care of Animals at the University of Michigan Medical School.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.