The interaction between peptides and major histocompatibility complex (MHC) molecules is pivotal in autoimmunity, pathogen recognition and tumor immunity. Recent advances in cancer immunotherapies demand for more accurate computational prediction of MHC-bound peptides. We address the generalizability challenge of MHC-bound peptide predictions, revealing limitations in current sequence-based approaches. Our structure-based methods leveraging geometric deep learning (GDL) demonstrate promising improvement in generalizability across unseen MHC alleles. Further, we tackle data efficiency by introducing a self-supervised learning approach on structures (3D-SSL). Without being exposed to any binding affinity data, our 3D-SSL outperforms sequence-based methods trained on ~90 times more data points. Finally, we demonstrate the resilience of structure-based GDL methods to biases in binding data on an Hepatitis B virus vaccine immunopeptidomics case study. This proof-of-concept study highlights structure-based methods’ potential to enhance generalizability and data efficiency, with possible implications for data-intensive fields like T-cell receptor specificity predictions.
27 February 2024 in Journal of Open Source Software
We present DeepRank2, a deep learning (DL) framework geared towards making predictions on 3D protein structures for variety of biologically relevant applications. Our software can be used for predicting structural properties in drug design, immunotherapy, or designing novel proteins, among other fields. DeepRank2 allows for transformation and storage of 3D representations of both protein-protein interfaces (PPIs) and protein single-residue variants (SRVs) into either graphs or volumetric grids containing structural and physico-chemical information. These can be used for training neural networks for a variety of patterns of interest, using either our pre-implemented training pipeline for graph neural networks (GNNs) or convolutional neural networks (CNNs) or external pipelines. The entire framework flowchart is visualized in Figure 1. The package is fully open-source, follows the community-endorsed FAIR principles for research software, provides user-friendly APIs, publicly available documentation, and in-depth tutorials.
Summary: Computational simulations like molecular dynamics and docking are providing crucial insights into the dynamics and interaction conformations of proteins, complementing experimental methods for determining protein structures. These methods often generate millions of protein conformations, necessitating highly efficient structure comparison and clustering methods to analyze the results. In this article, we introduce GradPose, a fast and memory-efficient structural superimposition tool for models generated by these large-scale simulations. GradPose uses gradient descent to optimally superimpose structures by optimizing rotation quaternions and can handle insertions and deletions compared to the reference structure. It is capable of superimposing thousands to millions of protein structures on standard hardware and utilizes multiple CPU cores and, if available, CUDA acceleration to further decrease superimposition time. Our results indicate that GradPose generally outperforms traditional methods, with a speed improvement of 2–65 times and memory requirement reduction of 1.7–48 times, with larger protein structures benefiting the most. We observed that traditional methods outperformed GradPose only with very small proteins consisting of ∼20 residues. The prerequisite of GradPose is that residue–residue correspondence is predetermined. With GradPose, we aim to provide a computationally efficient solution to the challenge of efficiently handling the demand for structural alignment in the computational simulation field. Availability and implementation: Source code is freely available at https://github.com/X-lab-3D/GradPose; doi:10.5281/zenodo.7671922.
Background: Analysis of the distribution of amino acid types found at equivalent positions in multiple sequence alignments has found applications in human genetics, protein engineering, drug design, protein structure prediction, and many other fields. These analyses tend to revolve around measures of the distribution of the twenty amino acid types found at evolutionary equivalent positions: the columns in multiple sequence alignments. Commonly used measures are variability, average hydrophobicity, or Shannon entropy. One of these techniques, called entropy–variability analysis, as the name already suggests, reduces the distribution of observed residue types in one column to two numbers: the Shannon entropy and the variability as defined by the number of residue types observed. Results: We applied a deep learning, unsupervised feature extraction method to analyse the multiple sequence alignments of all human proteins. An auto-encoder neural architecture was trained on 27,835 multiple sequence alignments for human proteins to obtain the two features that best describe the seven million variability patterns. These two unsupervised learned features strongly resemble entropy and variability, indicating that these are the projections that retain most information when reducing the dimensionality of the information hidden in columns in multiple sequence alignments.
Identifying tumor peptides that bind patient MHC proteins and elicit immune responses is central to immunotherapy, yet progress remains limited to a handful of well-studied alleles. Structure-based methods generalize better than sequence-only models but are constrained by the high computational cost of 3D modeling. We present SwiftMHC, the fastest structure-based framework for peptide–MHC (pMHC) modeling and binding affinity prediction. SwiftMHC predicts peptide–MHC binding in 0.009 sec/case in batch mode on a single A100 GPU—nearly an order of magnitude faster than leading sequence-based methods such as NetMHCpan 4.1 (0.081 sec/case)—while performing competitively in predictive accuracy. In addition to affinity estimation, SwiftMHC generates all-atom 3D pMHC structures and achieves a median Cα-RMSD of 1.32 Å against X-ray benchmarks, matching or better than the accuracy of state-of-the-art approaches (e.g., AlphaFold with fine-tuning) but running thousands of times faster (excluding disk-writing time). These results demonstrate the power of task-specific AI trained on physics-derived synthetic data to overcome the scarcity of experimental structures. Optimized for HLA-A*02:01 9-mers but readily extensible to other alleles, SwiftMHC enables rapid and accurate identification of peptides distinct from self at T-cell–exposed surfaces. This capability may expand immunotherapy targets, improve the safety of TCR-based therapies, and accelerate the development of next-generation cancer immunotherapies.
Understanding the peptide presentation mechanism of Major Histocompatibility Complex (MHC) is crucial to study the recognition of pathogens, the treatment of autoimmune diseases and in developing cancer immunotherapies. To comprehend these mechanisms and design effective therapies at affordable costs, it is fundamental to be able to precisely predict in silico which peptides can be bound by each MHC. Many computational approaches have been developed to identify MHC-binding peptides based on their sequences, however, they present biases and their performance on less-studied MHC alleles with limited experimental binding data is severely restricted by their sequence-based (SeqB) nature. We reason that structure-based (StrB) methods could learn the physico-chemical and geometrical rules of peptide-MHC (pMHC) complexes conserved among the alleles, thus generalizing better. We designed three supervised StrB geometric deep learning (GDL) methods, showing superior generalization across all StrB methods over two SeqB methods when the test data is distant from the training data. To further explore data efficiency, we designed a novel self-supervised GDL approach trained only on ∼1K x-ray structures, 3D-SSL, which predicts binding affinities without seeing any during training. Finally, we showcase our StrB method on a case study from the HPV vaccine design to show how StrB methods are robust to biases in the binding data. These findings highlight the potentials of structure-based methods in efficiently utilizing limited data, enhancing generalizability. This study bears important implications on data hungry fields, for example, the long-standing T cell receptor (TCR) specificity challenge.