Matches in SemOpenAlex for { <https://semopenalex.org/work/W4317617488> ?p ?o ?g. }
Showing items 1 to 47 of
47
with 100 items per page.
- W4317617488 abstract "Article Figures and data Abstract Editor's evaluation Introduction Results Discussion Methods Data availability References Decision letter Author response Article and author information Metrics Abstract The regulatory and effector functions of T cells are initiated by the binding of their cell-surface T cell receptor (TCR) to peptides presented by major histocompatibility complex (MHC) proteins on other cells. The specificity of TCR:peptide-MHC interactions, thus, underlies nearly all adaptive immune responses. Despite intense interest, generalizable predictive models of TCR:peptide-MHC specificity remain out of reach; two key barriers are the diversity of TCR recognition modes and the paucity of training data. Inspired by recent breakthroughs in protein structure prediction achieved by deep neural networks, we evaluated structural modeling as a potential avenue for prediction of TCR epitope specificity. We show that a specialized version of the neural network predictor AlphaFold can generate models of TCR:peptide-MHC interactions that can be used to discriminate correct from incorrect peptide epitopes with substantial accuracy. Although much work remains to be done for these predictions to have widespread practical utility, we are optimistic that deep learning-based structural modeling represents a path to generalizable prediction of TCR:peptide-MHC interaction specificity. Editor's evaluation The study provides a significant step forward in the prediction of T cell receptor docking to peptide-major histocompatibility complex ligands using a specialised version of the deep neural network structure prediction program AlphaFold. Progress towards this goal has implications for vaccine development, and cancer immunotherapy and is an intrinsically interesting structural problem due to the variability of the T cell receptor scaffold. https://doi.org/10.7554/eLife.82813.sa0 Decision letter Reviews on Sciety eLife's review process Introduction The specificity of T cell receptors (TCR) for peptides presented by major histocompatibility complex proteins (pMHC) is a critical determinant of adaptive immune responses to pathogens and tumors and of autoimmune disease. A predictive model of TCR:pMHC interactions, capable of mapping between TCR sequences and pMHC targets, could lead to advances in cancer immunotherapy and in the diagnosis and treatment of infectious and autoimmune diseases. Despite recent progress in TCR sequence analysis and modeling (Gielis et al., 2019; Huang et al., 2020; Mayer-Blackwell et al., 2021; Montemurro et al., 2021), a generalizable predictive model of TCR:pMHC interactions remains out of reach: existing predictors can learn to recognize new TCR sequences specific for pMHCs in their training set, but robust generalization to unseen pMHC epitopes has not been convincingly demonstrated (Moris et al., 2021). Two key difficulties are the diversity of TCR:pMHC recognition modes, a consequence of TCR sequence and structural diversity and flexibility in TCR:pMHC docking orientation, and the limited number of experimentally validated TCR:pMHC interaction examples for use in training. We hypothesized that 3D structural modeling might offer a path toward generalizable prediction of TCR:pMHC interactions in the current data-limited regime. At the biophysical level, TCR:pMHC interaction specificity is determined by the structures and flexibilities of the interacting partners. A wealth of structural studies have provided valuable insights into the atomistic determinants of specificity (Rossjohn et al., 2015; Rudolph et al., 2006; Singh et al., 2017). Collectively, these experimentally determined structures define a range of docking geometries that likely covers the majority of unseen interactions; they also provide valuable templates for cutting-edge deep neural network structure prediction methods such as AlphaFold (Jumper et al., 2021) and RoseTTAfold (Baek et al., 2021). These prediction tools feature advanced network architectures with millions of parameters that are trained on structurally characterized proteins and their sequence homologs. Despite being trained on monomeric structures, these approaches can generate state-of-the-art structure predictions for protein complexes, and they have even been used to predict whether or not protein pairs will associate (Humphreys et al., 2021). Here we show that a version of AlphaFold specialized for TCR:pMHC modeling can be used to predict TCR:pMHC binding specificity with some success. Whereas the default AlphaFold version trained to predict protein:protein docking (AlphaFold-Multimer Evans et al., 2021) shows inconsistent performance on TCR:pMHC structures (Yin et al., 2022), our specialized pipeline demonstrates improved accuracy and reduced computational cost. Moreover, this modeling pipeline has significant power to discriminate target peptides from decoy peptides as evaluated on a benchmark of human and mouse MHC class I epitopes. Importantly, success in predicting the correct peptide target correlates with structural accuracy of the models, suggesting that when the pipeline succeeds, it does so by recapitulating key specificity determinants. This work, together with previous studies applying molecular modeling techniques to TCRs (Borrman et al., 2020; Jensen et al., 2019; Lanzarotti et al., 2018; Pierce and Weng, 2013), suggests that structure-based approaches represent a promising path forward for predicting TCR:pMHC interaction specificity. Results Structure prediction We first evaluated the structure prediction performance of a recently released version of AlphaFold (AlphaFold-Multimer Evans et al., 2021) that was specifically trained for protein:protein docking. AlphaFold-Multimer leverages inter-chain residue covariation observed in orthologs of the target proteins to identify amino acid pairs making interface contacts. Given that TCR:pMHC interactions are determined in part by highly variable, non-germline encoded CDR3 regions, it was unclear whether AlphaFold’s strong docking performance on other systems would translate to TCR:pMHC interactions. Indeed, the AlphaFold-Multimer developers noted that it does not perform well on antibody:antigen complexes, which share many features with TCR:pMHC complexes. We tested two versions of AlphaFold-Multimer, one in which the full sequences of the interacting partners are provided as input ('AFM_full': MHC-I or MHC-IIa, beta-2 microglobulin or MHC-IIb, peptide, TCRa, and TCRb variable and constant domains), and one in which only the directly interacting domains are provided as input ('AFM_trim': TCR constant domains, beta-2 microglobulin, and C-terminal MHC domains are removed). Restricting to the core interacting domains speeds the calculations substantially at the risk of introducing decoy docking sites at the location of interfaces with the missing domains. Although both models were capable of generating high-quality predictions on a nonredundant set of 130 TCR:pMHC complexes (as indicated by CDR loop RMSDs at and below ~2 Å; details below), prediction quality was highly variable, and visual inspection revealed that many of the predicted models had displaced peptides and/or TCR:pMHC docking modes that were outside the range observed in native proteins. Additionally, these AlphaFold predictions took multiple hours per target to complete, limiting their throughput. One limitation of AlphaFold-Multimer is that it does not support multi-chain templates (Evans et al., 2021): template information from the database of solved structures can inform the internal conformation of individual chains, but it does not guide the docking of chains into higher order complexes. The constrained nature of the TCR:pMHC binding mode suggests that higher and more consistent prediction accuracy could be obtained by providing additional template information. A challenge when modeling TCR structures is that the V-alpha and V-beta genes largely determine the best structural template, and these genes associate freely rather than in fixed pairings, which means that the optimal structural template for the TCR-alpha chain will often come from a different PDB structure as that for the TCR-beta chain. Additionally, the TCR:pMHC docking mode varies widely within an overall diagonal binding mode, in a way that is not easily predicted directly from sequence, making it challenging to select an optimal template for the TCR:pMHC relative orientation. Guided by these considerations, we developed an AlphaFold-based TCR docking pipeline that uses hybrid structural templates to provide a broad, native-like sampling of potential docking modes (Figure 1). In this approach, individual chain templates are first selected based on sequence similarity to the target TCR:pMHC (Figure 1A). Hybrid complexes are created from these individual chain templates by using a diverse set of representative docking geometries to orient the TCR chains relative to the pMHC (see Methods). Docking geometries are defined in terms of the 6 degrees of freedom that relate the MHC reference frame to the TCR reference frame, where the MHC and TCR reference frames are defined based on internal pseudo symmetry (Figure 1B and D and Methods). These hybrid complexes are provided as templates to multiple independent AlphaFold simulations, four templates per simulation, with the highest confidence model from the simulations taken as the final prediction (Figure 1C). During benchmarking, templates and docking geometries from structures with similar TCRs or pMHCs to the target are excluded to reduce bias toward the native structure (see Methods; this constraint was not applied to the default AlphaFold-Multimer methods). Given that we are providing template information that constrains the inter-chain docking, we chose not to include additional multiple sequence alignment (MSA) information beyond the target sequence. This greatly speeds the predictions: MSA building is the most time-consuming part of the AlphaFold pipeline, and the neural network inference step is also significantly faster without MSA information. Figure 1 with 1 supplement see all Download asset Open asset Constructing diverse hybrid templates for AlphaFold modeling. (A) Four structural templates for each TCR chain and for the peptide:MHC are identified in the Protein Databank (Berman et al., 2000) by sequence similarity search. (B) TCR:pMHC docking geometry is defined by computing the rigid-body transformation between TCR and pMHC coordinate frames. Coordinate frames are oriented based on internal pseudo symmetry as described in the Methods. (C) Three independent AlphaFold simulations are performed, each with four hybrid templates built from the four sets of single-chain templates oriented relative to one another using one of twelve representative docking geometries chosen to cover a wide range of experimentally determined ternary complexes. (D) TCR coordinate frames from class I pMHC ternary structures and the 12 representative transforms (thicker arrows) are shown in a common coordinate system defined by their corresponding pMHC coordinate frames. We found that the hybrid templates AlphaFold pipeline specialized for TCR:pMHC ('AF_TCR') produces higher quality models than either of the Alphafold-Multimer variants on a benchmark set (Figure 2—figure supplement 1) of 130 TCR:pMHC complexes (Figure 2A, Wilcoxon P<10–7 vs AFM_full and P<10–12 vs AFM_trim on the full set; Figure 2B, P<10–3 for both comparisons on 20 targets without a close homolog in the AlphaFold-Multimer training set; and Figure 2—figure supplement 2 for peptide modeling accuracy). The AF_TCR pipeline also outperforms the state of the art TCRpMHCmodels pipeline (Jensen et al., 2019) for Class I MHC TCR modeling (Figure 2—figure supplement 3A–B), and produces better docking geometries than simply borrowing the geometry from the most sequence-similar template (Figure 2—figure supplement 3C). There was a significant positive correlation between predicted and observed model accuracy (Figure 2C). Figure 2 with 6 supplements see all Download asset Open asset TCR modeling accuracy. (A) Comparison between Alphafold-Multimer with full ('AFM_full') or trimmed ('AFM_trim') input sequences and the hybrid-templates TCR pipeline ('AF_TCR'). CDR RMSD values (y-axis) are computed by superimposing the native and modeled MHC coordinates and comparing the placement of the TCR CDR loops (see Methods). (B) Same as in (A) but for the 20 benchmark targets unrelated to any TCR:pMHC structure deposited before May 2018, the cutoff date for the AlphaFold-Multimer training set. (C) AlphaFold’s predicted aligned error (PAE) measure, evaluated between TCR and pMHC, correlates with CDR RMSD between model and native structure. (D) The docking geometry of the final AlphaFold model improves over the best of the 12 templates in 30% of cases (points above the line y=x). (E) The docking geometry of the final AlphaFold model improves over the median of the 12 templates in 94% of cases (points above the line y=x). (F) Fine-tuning AlphaFold’s parameters on human TCR:pMHC complexes improves prediction of mouse TCR:pMHC complexes. Boxes in A, B, and F show the quartiles of the plotted distributions. Figure 2—source data 1 Structure prediction benchmark. CSV file containing information on the ternary complexes in the structure prediction benchmark. https://cdn.elifesciences.org/articles/82813/elife-82813-fig2-data1-v1.csv Download elife-82813-fig2-data1-v1.csv For each benchmark target, the AlphaFold TCR pipeline is provided with 12 hybrid template complexes whose TCR:pMHC docking modes are taken from 12 diverse ternary structures unrelated to the target. We were curious to know whether the AlphaFold simulation was improving on the docking information present in these template structures. To answer this question, we compared the accuracy of the docking geometry present in the final model to the accuracies of the 12 template structures. Since the 12 templates differ in the sequences and structures of their CDR loops, we developed a distance between TCR:pMHC docking geometries that compares the placement of 'generic' CDR loops ('docking RMSD', see Methods). This docking RMSD measure is correlated with CDR RMSD in comparisons of models to natives (Figure 2—figure supplement 4), but it focuses exclusively on the docking geometry and provides a sequence-independent way of comparing binding modes that emphasizes CDR loop placement. For 30% of the targets, the AlphaFold TCR final model had a lower RMSD than the best template docking geometry (Figure 2D); the final model improved over the median template RMSD for 94% of the targets (Figure 2E). To visualize the overall docking geometry landscape of models and natives, we calculated docking RMSD values between all of the native ternary structures and the AlphaFold-TCR and AlphaFold-Multimer models and transformed this distance matrix into a 2D projection (Figure 2—figure supplement 5) using the UMAP algorithm (McInnes et al., 2018). Inspection of this 2D docking geometry landscape reveals regions that are distant from the native structures and only sampled by the AlphaFold-Multimer models, supporting the view that incorporating template docking geometries helps to constrain predictions to native-like geometries. We analyzed the factors contributing to docking prediction accuracy and found that two dominant factors are the degree to which the docking geometry in the native structure deviates from the consensus binding mode (as captured in a multidimensional Z score, see methods) and the MHC class (class II binding modes were better predicted than class I), with minor contributions from V gene template sequence distance, CDR loop modeling accuracy, peptide modeling accuracy, and TCRalpha/TCRbeta docking accuracy (Figure 2—figure supplement 6). An attractive feature of neural network architectures is the potential to 'fine tune' a general network for improved prediction accuracy in a specific domain. We fine-tuned the AlphaFold parameters in the context of the AlphaFold TCR pipeline on the set of 93 human TCR:pMHC complexes from the benchmarking set and subsequently evaluated the performance of this model on the 37 mouse TCR:pMHC targets. Despite the small size of the TCR:pMHC ternary structure database, the fine-tuned model showed improved performance on the mouse targets (Figure 2F; Wilcoxon p<0.015), which are distinct in the details of their epitope, MHC, and TCR sequences from the human training set, suggesting that the model was able to learn generalizable features of TCR:pMHC interactions. This fine-tuning procedure was facilitated by the fact that the AF2 model requires significantly less memory in the absence of MSA information, making it possible to perform parameter optimization on full TCR:pMHC systems without any residue cropping. Binding specificity prediction Having established that the AlphaFold TCR pipeline can generate more accurate TCR:pMHC models than AlphaFold-Multimer, we evaluated its performance in TCR epitope prediction. The general problem of predicting, de novo, which peptide:MHCs a given TCR recognizes is likely to be very difficult due to the diversity of TCR:pMHC recognition modes, the polyspecificity of individual TCRs, and the paucity of available training data (Moris et al., 2021). Here we consider instead the simpler problem of selecting the correct target peptide from a small set of candidates. This might correspond to a real-world scenario in which we know the source antigen from which the unknown peptide epitope is taken, or we have a positive hit in a T cell stimulation assay that implicates a pool of peptides rather than a unique epitope. For benchmarking, we focus on peptide-MHC epitopes for which a repertoire of cognate TCRs has been identified. This allows us to evaluate the sensitivity of the predictions to small changes in TCR sequence. It also lets us investigate a scenario in which we are given not one TCR, but a set of TCRs that are all predicted to recognize the same epitope, and we consider the extent to which this helps to constrain the target epitope. With improved single-cell technologies for paired TCR sequencing, and improved methods for identifying TCR sequence convergence, we hypothesize that this will become an increasingly common scenario. We selected a set of 8 Class I peptide:MHC systems (Table 1) for which a repertoire of paired epitope-specific TCRs and a solved ternary structure were available. These systems include one human (A*0201) and one mouse (H2-Db) MHC allele, each with 9- and 10-residue peptides. TCR repertoires containing more than 50 unique TCR sequences were subsampled to a set of 50 TCRs using an algorithm that removed redundancy while concentrating on the more densely sampled regions of TCR space (see Methods). For each MHC/peptide length combination, we used the NetMHCpan-4.1 (Reynisson et al., 2020) method to select 9 decoy peptides with binding scores in the range of the true peptide binders. We additionally selected 50 irrelevant TCRs at random from human and mouse CD8 T cell datasets made available by 10 X Genomics (these TCRs were used to correct for pMHC-intrinsic effects; see below and Methods). Table 1 Binding specificity benchmark. OrganismMHCPeptide lengthPeptide sequenceAntigenhumanHLA-A*02:019GILGFVFTLFlu M1humanHLA-A*02:019GLCTLVAMLEBV BMLF1humanHLA-A*02:019NLVPMVATVCMV pp65humanHLA-A*02:019YLQPRTFLLSARS-CoV-2 SpikehumanHLA-A*02:0110ELAGIGILTVhuman MART-1humanHLA-A*02:0110KLVALGINAVHCV POLGmouseH2-Db9ASNENMETMFlu NPmouseH2-Db10SSLENFRAYVFlu PA We used the AlphaFold TCR pipeline to generate docked complexes and associated interface accuracy estimates for pairings of each TCR with its true pMHC epitope and with 9 decoy peptides of the same length (Figure 3A). This produces, for each of the eight pMHCs, an Nx10 matrix of predicted interface accuracies (Figure 3B, left panel), where N is the number of TCRs specific for the given pMHC. To generate a single number representing the estimated interface accuracy of a complex, we summed the residue-residue predicted aligned error (PAE) for all TCR:pMHC residue pairs. These raw accuracy estimates showed significant TCR- and pMHC-intrinsic effects (Figure 3B). Certain TCRs had consistently higher or lower than average predicted interface accuracies due to features such as longer CDR3 loops or usage of V genes without a close structural template. We saw similar, albeit weaker, trends across different peptide:MHC complexes, perhaps due to AlphaFold’s confidence in the MHC-bound structure of the peptide. TCR-intrinsic factors do not change the relative order of candidate peptides, but they make comparisons of binding predictions across TCRs difficult; pMHC effects have the potential to change the rank ordering of candidate peptide epitopes. Since we are interested here in evaluating the compatibility between TCR and pMHC and not, e.g., ranking peptides by their affinity for MHC, we corrected for these TCR- and pMHC-intrinsic effects to generate an array of TCR:pMHC binding scores intended to be comparable across different pMHCs and TCRs (Figure 3B, middle panel; lower scores indicate stronger predicted binding, see Methods). Figure 3 Download asset Open asset Structural modeling can sometimes discriminate correct from incorrect TCR:pMHC pairings. (A) For each of the eight peptide:MHC epitopes, we docked multiple cognate TCRs against multiple decoy peptides and the wild type epitope. Here three TCRs and three pMHCs are shown; 9 decoys and up to 50 TCRs were actually modeled. (B) For each candidate TCR:pMHC pairing, the mean AlphaFold predicted aligned error (PAE) for the TCR:pMHC interface was calculated (left) and transformed into a binding score by subtracting out TCR-intrinsic and pMHC-intrinsic factors (middle). These binding scores were averaged to define a repertoire-level binding score for the WT epitope and each of the decoys (bottom). Also calculated was the rank of the WT binding score within the list of all the binding scores for each TCR (right). (C) TCRdist hierarchical clustering tree of the 50 modeled TCRs for the A*02:01 GIL9 epitope, labeled with the TCR sequence information, top-ranked peptide, and rank of the WT peptide, and colored by the rank of the WT peptide. Internal edges, which correspond to multiple ‘leaf’ TCRs, are colored by the rank of the WT peptide after averaging the binding scores over the leaf TCRs. (D) Repertoire binding scores for each of the eight target epitopes and the 9 decoy peptides, with the lowest (most favorable) binding score in each row boxed. (E) Receiver operating characteristic (ROC) curves for discrimination of WT from decoy peptides by binding score. Area under the ROC curve (AUROC) values are given in the legend along with the sequence of the WT peptide. Figure 3—source data 1 Epitope specificity benchmark TCRs. CSV file containing information on the TCR sequences in the epitope decoy discrimination benchmark. https://cdn.elifesciences.org/articles/82813/elife-82813-fig3-data1-v1.csv Download elife-82813-fig3-data1-v1.csv Figure 3—source data 2 Epitope specificity benchmark peptides. CSV file containing information on the peptides in the epitope decoy discrimination benchmark. https://cdn.elifesciences.org/articles/82813/elife-82813-fig3-data2-v1.csv Download elife-82813-fig3-data2-v1.csv We evaluated the accuracy of these binding predictions across the eight pMHC epitopes. First, we calculated the rank of the true peptide epitope amongst the 9 decoy peptides (Figure 3B, right panel) on a per-TCR basis. To visualize how these ranks vary across each pMHC-specific repertoire, we constructed hierarchical clustering trees of the TCR sequences using the TCRdist measure (Dash et al., 2017) and colored them by the rank of the true peptide (Figure 3C and Figure 4). Internal edges, which correspond to multiple ‘leaf’ TCRs, are colored by the rank of the true peptide after averaging the binding scores over the leaf TCRs. Looking across all eight epitopes, we can see, first, that the predictions are not random: on average the correct peptide is ranked more favorably than most of the decoys (i.e. there is more blue than red). For six of the eight epitopes, the correct peptide is ranked first when we average the binding scores of all the TCRs in the repertoire (Figure 3D; Figure 4: the largest branch of the tree is dark blue). It also appears that the epitopes with more sequence-diverse repertoires (A*0201-GLC9 and A*02:01-NLV9) are more challenging to predict: the trees that merge completely at smaller TCRdist values (further to the left) are bluer than the other trees in Figure 4. This can be seen quantitatively by plotting the TCRdiv repertoire sequence diversity measure (Dash et al., 2017) against measures of binding prediction success (Figure 4—figure supplement 1). If we rank the peptides by binding score and compare the recovery of true binder peptides to decoys using receiver operating characteristic (ROC) curves, we can see that some epitopes, such as A*02:01-YLQ9 and A*02:01-ELA10 are predicted very well (by area under the ROC curve, AUROC ≥ 0.96) and some predictions are only slightly better than random (Figure 3E). We find an overall AUROC value of 0.82 when binding and non-binding TCR:pMHC pairs from all epitopes are ranked together. Figure 4 with 1 supplement see all Download asset Open asset Peptide decoy discrimination results for the eight benchmark epitopes. The rank of the wild type peptide relative to the 9 decoys (0=best, 9=worst) is shown in a heatmap and a TCRdist hierarchical clustering tree of the epitope-specific TCRs. Each row of the heatmap corresponds to a single TCR; each column corresponds to one of the 10 modeled peptides, with the wild type peptide on the left. The vertical ordering of the TCRs in the heatmaps and trees is the same. Internal edges of the trees, which correspond to multiple ‘leaf’ TCRs, are colored by the rank of the wild type peptide after averaging the binding scores over the leaf TCRs. We looked to see whether structural modeling accuracy correlated with binding prediction success (Figure 5). Although very few of the specific TCRs being modeled have been structurally characterized, each of the epitopes has at least one solved ternary structure in the protein structure database. For each TCR, we computed docking RMSDs between the TCR:pMHC model in complex with its cognate epitope and the solved ternary structures for that epitope and took the minimum value as a proxy for the accuracy of the predicted binding mode. Figure 5A shows the distribution of these RMSD values across each repertoire. Well-predicted epitopes such as A*02:01-YLQ9 and A*02:01-ELA10 indeed appear to have smaller RMSD values than other repertoires. The mouse pMHC H2Db-ASN9 is an outlier, with an RMSD distribution shifted to very high values. Examination of the three ternary structures for this pMHC revealed that they represent a unique population of TRBV17+ TCRs that is distinct from the consensus repertoire modeled here. Two of the three TCRs bind with a reversed docking orientation (Gras et al., 2016), and the third has a highly displaced binding footprint (Zareie et al., 2021); all three are outliers in a hierarchical clustering tree of Class I TCRs based on docking RMSD (Figure 5—figure supplement 1). If we exclude H2Db:ASN9 and plot docking RMSD to the closest epitope structure versus binding score for the correct peptide, we see that there is a positive correlation (Figure 5B). The TCRs for which the correct peptide is ranked first have a lower RMSD distribution than other TCRs, and this RMSD distribution shifts upward as the rank of the correct peptide declines (Figure 5C). These results suggest that the correct binding predictions are driven at least in part by recovery of native-like structural features (analysis of peptide backbone RMSDs shows a positive, but much weaker, correlation between binding prediction and modeling accuracy: Figure 5—figure supplement 2). Figure 5 with 2 supplements see all Download asset Open asset Success in decoy discrimination correlates with structural modeling accuracy. (A) For each TCR, the structural model in complex with the wild type epitope was compared to all experimentally determined ternary structures for that epitope and the smallest docking RMSD was recorded. The resulting RMSD distributions were smoothed using kernel density estimation and plotted. (B) Scatter plot of docking RMSD to the nearest wild type structure versus the binding score for the wild type peptide. Favorable wild type binding scores correlate with lower RMSD values. (C) Distributions of docking RMSD to the nearest wild type structure (y-axis) as a function of the rank of the wild type peptide (x-axis). When the wild type peptide is ranked first (left violin), the corresponding docking geometries are more similar to those of ternary complexes for that epitope, suggesting higher accuracy. To further investigate the behavior of our modeling approach, we performed an in silico epitope alanine scan of each of the eight pMHC-specific repertoires. We built models and calculated binding scores for each epitope-specific TCR docked to all single-alanine mutants of the native peptide (native alanine residues were mutated to glycine). Binding scores for each TCR and each of the alanine mutants are shown in the heatmaps in Figure 6. Averaging these binding scores over all of the TCRs for each epitope and subtracting the score for the native peptide gives a predicted repertoire-level sensitivity to mutation at each peptide position (Figure 6B). From" @default.
- W4317617488 created "2023-01-21" @default.
- W4317617488 creator A5067744380 @default.
- W4317617488 date "2022-12-09" @default.
- W4317617488 modified "2023-10-01" @default.
- W4317617488 title "Author response: Structure-based prediction of T cell receptor:peptide-MHC interactions" @default.
- W4317617488 doi "https://doi.org/10.7554/elife.82813.sa2" @default.
- W4317617488 hasPublicationYear "2022" @default.
- W4317617488 type Work @default.
- W4317617488 citedByCount "0" @default.
- W4317617488 crossrefType "peer-review" @default.
- W4317617488 hasAuthorship W4317617488A5067744380 @default.
- W4317617488 hasBestOaLocation W43176174881 @default.
- W4317617488 hasConcept C147483822 @default.
- W4317617488 hasConcept C185592680 @default.
- W4317617488 hasConcept C203014093 @default.
- W4317617488 hasConcept C207936829 @default.
- W4317617488 hasConcept C2779281246 @default.
- W4317617488 hasConcept C41008148 @default.
- W4317617488 hasConcept C55493867 @default.
- W4317617488 hasConcept C70721500 @default.
- W4317617488 hasConcept C86803240 @default.
- W4317617488 hasConceptScore W4317617488C147483822 @default.
- W4317617488 hasConceptScore W4317617488C185592680 @default.
- W4317617488 hasConceptScore W4317617488C203014093 @default.
- W4317617488 hasConceptScore W4317617488C207936829 @default.
- W4317617488 hasConceptScore W4317617488C2779281246 @default.
- W4317617488 hasConceptScore W4317617488C41008148 @default.
- W4317617488 hasConceptScore W4317617488C55493867 @default.
- W4317617488 hasConceptScore W4317617488C70721500 @default.
- W4317617488 hasConceptScore W4317617488C86803240 @default.
- W4317617488 hasLocation W43176174881 @default.
- W4317617488 hasOpenAccess W4317617488 @default.
- W4317617488 hasPrimaryLocation W43176174881 @default.
- W4317617488 hasRelatedWork W1598424666 @default.
- W4317617488 hasRelatedWork W1966863828 @default.
- W4317617488 hasRelatedWork W2027267610 @default.
- W4317617488 hasRelatedWork W2031336749 @default.
- W4317617488 hasRelatedWork W2130354820 @default.
- W4317617488 hasRelatedWork W2748952813 @default.
- W4317617488 hasRelatedWork W2770101712 @default.
- W4317617488 hasRelatedWork W2899084033 @default.
- W4317617488 hasRelatedWork W2953216564 @default.
- W4317617488 hasRelatedWork W644041026 @default.
- W4317617488 isParatext "false" @default.
- W4317617488 isRetracted "false" @default.
- W4317617488 workType "peer-review" @default.