[Paper Review] Neural message passing for joint paratope-epitope prediction

Alice Del Vecchio, Andreea Deac+2 more|arXiv (Cornell University)|May 31, 2021

vaccines and immunoinformatics approachesBiochemistry, Genetics and Molecular Biology21 references17 citations

TL;DR

This paper proposes EPMP, a novel asymmetric neural message passing framework for joint paratope-epitope prediction that uses distinct architectures—Para-EPMP for the sequential paratope and Epi-EPMP for the structurally complex epitope—achieving state-of-the-art performance on both tasks, with a 75.2% AUC PR on paratope prediction and 27.7% AUC PR on epitope prediction, significantly outperforming prior methods like PECAN.

ABSTRACT

Antibodies are proteins in the immune system which bind to antigens to detect and neutralise them. The binding sites in an antibody-antigen interaction are known as the paratope and epitope, respectively, and the prediction of these regions is key to vaccine and synthetic antibody development. Contrary to prior art, we argue that paratope and epitope predictors require asymmetric treatment, and propose distinct neural message passing architectures that are geared towards the specific aspects of paratope and epitope prediction, respectively. We obtain significant improvements on both tasks, setting the new state-of-the-art and recovering favourable qualitative predictions on antigens of relevance to COVID-19.

Motivation & Objective

• To address the inherent asymmetry between paratope and epitope prediction, where paratopes are sequence-localized and epitopes are structurally dispersed and context-dependent.
• To improve computational prediction of antibody-antigen binding sites to accelerate in silico vaccine and synthetic antibody design.
• To overcome class imbalance in epitope prediction and improve generalization through multi-task learning and attention-based message passing.
• To develop a joint predictor that leverages structural information from AlphaFold2-predicted models while maintaining high accuracy on both tasks.

Proposed method

• Para-EPMP uses dilated convolutional neural networks (à trous CNNs) to exploit the sequential localization of paratopes, combined with graph neural networks (GNNs) to incorporate 3D structural context.
• Epi-EPMP employs a multi-task learning setup with attention-based GNNs (GATs) to model long-range dependencies and contextual cues from the antibody, improving regularization and generalization.
• The model uses a hybrid architecture where the antibody’s paratope is predicted via a sequence-aware module, while the antigen’s epitope is predicted using antibody-conditional message passing.
• A class-weighted binary cross-entropy loss is used for training, with separate learning rates for GCN and non-GCN layers to stabilize training on the highly imbalanced epitope task.
• The framework leverages structural priors from AlphaFold2-predicted protein structures, treating residues as nodes in a contact graph with edges based on spatial proximity (<4.5Å).
• Multi-task learning explicitly regularizes the epitope predictor by sharing representations with the paratope prediction head, improving attention focus on biologically relevant antibody regions.

Experimental results

Research questions

• RQ1 Can a joint paratope-epitope prediction model achieve superior performance by explicitly modeling the asymmetry between the two tasks?
• RQ2 Does incorporating both sequential and structural inductive biases improve paratope prediction accuracy compared to symmetric architectures?
• RQ3 Can multi-task learning with shared attention mechanisms improve generalization and reduce overfitting in epitope prediction, which suffers from high class imbalance?
• RQ4 Can the model produce biologically meaningful predictions on clinically relevant antigens, such as the SARS-CoV-2 RBD?

Key findings

• EPMP achieves a new state-of-the-art AUC PR of 75.2% on paratope prediction, significantly outperforming PECAN’s 70.0%.
• On epitope prediction, EPMP achieves an AUC PR of 27.7%, surpassing PECAN’s 21.2% and other ablation models like Epi-GCN (22.9%) and Epi-MPNN (20.6%).
• The ablation study confirms that the asymmetric architecture is essential: models without the dedicated sequential component (e.g., Para-MPNN) or without multi-task learning perform worse.
• EPMP successfully identifies the correct binding interface on the SARS-CoV-2 RBD using the B38 antibody, with minimal false positives near the true binding site.
• The model maintains strong performance without transfer learning, outperforming PECAN even when the latter uses transfer learning on general protein-protein interactions.
• The use of multi-task learning with explicit paratope supervision improves attention focus in Epi-EPMP, reducing unsupervised learning instability.