Peptide binding specificity prediction using fine-tuned protein structure prediction networks

Abstract

Peptide binding proteins play key roles in biology, and predicting their binding specificity is a long-standing challenge. While considerable protein structural information is available, the most successful current methods use sequence information alone, in part because it has been a challenge to model the subtle structural changes accompanying sequence substitutions. Protein structure prediction networks such as AlphaFold model sequence-structure relationships very accurately, and we reasoned that if it were possible to specifically train such networks on binding data, more generalizable models could be created. We show that placing a classifier on top of the AlphaFold network and fine-tuning the combined network parameters for both classification and structure prediction accuracy leads to a model with strong generalizable performance on a wide range of Class I and Class II peptide-MHC interactions that approaches the overall performance of the state-of-the-art NetMHCpan sequence-based method. The peptide-MHC optimized model shows excellent performance in distinguishing binding and non-binding peptides to SH3 and PDZ domains. This ability to generalize well beyond the training set far exceeds that of sequence only models, and should be particularly powerful for systems where less experimental data is available.

Structure-based prediction of T cell receptor:peptide-MHC interactions

“…Rosetta software package (Leaver-Fay et al, 2011) or other molecular modeling tools (Lee et al,2018). Finally, it may be possible to fine-tune AlphaFold parameters directly to discriminate TCR:pMHC binding examples from non-binding examples, as we have recently demonstrated for peptide:MHC interactions (Motmaen et al, 2022). This would allow us to directly leverage the thousands of validated TCR:pMHC interactions within the context of a structurally-informed training procedure.…”

Identifying endogenous peptide receptors by combining structure and transmembrane topology prediction

“…While conceivable, we refrained from developing a dedicated confidence-derived score such as pDockQ for peptide-receptor docking due to the very limited size of the dataset. Also, optimally, given a larger dataset, a classification model to distinguish true receptors could be trained on top of AlphaFold, as was recently done for peptide-MHC binding [24].…”

Peptide-MHC Structure Prediction With Mixed Residue and Atom Graph Neural Network

“…The advent of large high-quality proteomic data has led to an unprecedented opportunity for deep learning-based protein folding methods such as AlphaFold 2 [Jumper et al, 2021], OmegaFold [Wu et al, 2022], and ESMFold [Lin et al, 2022], which have been proven to predict reliable and accurate protein structures. Furthermore, these trained models can be fine-tuned on pMHC structures for other downstream tasks such as binding prediction [Motmaen et al, 2022].…”

“…A post-processing step recovers the full-atom coordinates and constructs the entire pMHC structure. By training on a novel weighted reconstruction loss, the proposed GNN predicted accurate structures with similar performance compared to AlphaFold 2 [Jumper et al, 2021] and pMHC fine-tuned AlphaFold 2 [Motmaen et al, 2022] with only 1.7M learnable parameters, which represents a more than 98% reduction in the number of parameters. We demonstrate that the predicted structures are closer to native structures, both in terms of geometry and biological consistency.…”

“…The metrics of the proposed model together with the baselines AlphaFold 2[Jumper et al, 2021], pMHC fine-tuned AlphaFold 2[Motmaen et al, 2022], and OmegaFold [Wu et al, 2022] are reported…”