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

Abstract

Neoantigen-targeting vaccines have achieved breakthrough success in cancer immunotherapy by eliciting immune responses against neoantigens, which are proteins uniquely produced by cancer cells. During the immune response, the interactions between peptides and major histocompatibility complexes (MHC) play an important role as peptides must be bound and presented by MHC to be recognised by the immune system. However, only limited experimentally determined peptide-MHC (pMHC) structures are available, and in-silico structure modelling is therefore used for studying their interactions. Current approaches mainly use Monte Carlo sampling and energy minimisation, and are often computationally expensive. On the other hand, the advent of large high-quality proteomic data sets has led to an unprecedented opportunity for deep learning-based methods with pMHC structure prediction becoming feasible with these trained protein folding models. In this work, we present a graph neural network-based model for pMHC structure prediction, which takes an amino acid-level pMHC graph and an atomic-level peptide graph as inputs and predicts the peptide backbone conformation. With a novel weighted reconstruction loss, the trained model achieved a similar accuracy to AlphaFold 2, requiring only 1.7M learnable parameters compared to 93M, representing a more than 98% reduction in the number of required parameters.

Revolutionizing peptide‐based drug discovery: Advances in the post‐AlphaFold era

“…This method slightly outperforms AF multimer in accuracy while remarkably enhancing computational speed, enabling the calculation of 100,000 predictions in just 4 h. Other noteworthy advancements include TFold, which extends the AF network's prowess by incorporating paired templates and multiple sequence alignments (MSA) derived from binding data, alongside a bespoke sequence‐based network 45 . Additionally, an independent study conducted by InstaDeep and BioNTech harnessed a simplified graph neural network (GNN) model with 98% fewer learnable parameters than AF, yet yielded comparable performance 46 …”

“…45 Additionally, an independent study conducted by InstaDeep and BioNTech harnessed a simplified graph neural network (GNN) model with 98% fewer learnable parameters than AF, yet yielded comparable performance. 46 In the post-AF era, a pertinent inquiry arises: Do traditional docking tools retain their relevance for peptide-based discovery? While AF has indeed revolutionized the field, it is not devoid of limitations.…”

LightMHC: A Light Model for pMHC Structure Prediction with Graph Neural Networks

“…Delaunay et al [2022] enhance performance with GNN but confine to 9-meric peptide backbone prediction, facing scalability issues with side-chain recovery. Pre-trained models like AlphaFold 2 [Jumper et al, 2021, Lin et al, 2022] tend to be overparametrised, while smaller specialised models match single-family performance [Delaunay et al, 2022]. Adaptations of AlphaFold 2 focus on pMHC but lack out-of-sample testing [Motmaen et al, 2022, Mikhaylov and Levine, 2023].…”

“…While AlphaFold 2 [Jumper et al, 2021] and ESMFold [Lin et al, 2022] have demonstrated remarkable successes, these models cannot always capture accurately the conformational dynamics of flexible regions and remain computationally expensive, making them unsuitable for high-throughput inference. There also exist application-specific models that are limited to specific protein domains [Abanades et al, 2022, Delaunay et al, 2022] or backbone atoms [Delaunay et al, 2022, Cohen et al, 2022], or require pre-defined templates [Abanades et al, 2022], limiting their applicability to a broader range of immune recognition molecules. In this work, we focus on pMHC structure prediction task and present LightMHC: a light model combining attention-based graph neural networks and convolutional neural networks.…”

Accurate modeling of peptide-MHC structures with AlphaFold

“…An example of an automated user-friendly tool is PANDORA [6], a MODELLER-based pipeline. Since the deep learning revolution in protein structure prediction [7,8], AlphaFold [8,9] (AF) has been applied to pMHC structure and binding prediction [10], and custom neural nets were built specifically for this task [11,12].…”