Multimodal Graph Adversarial Networks for Bias-Aware Protein-Peptide Binding Prediction

View at publisher Summarise with VIRAT
Authors: Chetany Kalaneya, Gaurav Ganguly, Rohit Kumawat
Cited by: 0
References: 10

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Drug discovery, immunotherapy, and molecular recognition heavily rely on protein-peptide interactions. However, deep learning models have shown significant bias towards common protein variants and, thus, have a weak generalization capability for rare mutations. We introduce MAGAN-PP (Multimodal Graph Adversarial Network for Protein-Peptide binding), a new framework that combines graph neural networks for the 3D structural representation, multimodal variational autoencoders for the fusion of the structural and textual biological annotations, and adversarial debiasing for the maintenance of fairness across variant types. Our method performs almost equally well for common and rare variants (AUC difference: 0.0167), which is an 88.9 % increase in fairness compared to the baseline models. By achieving an F1-score of 0.696 and providing interpretable SHAP-based feature importance, MAGANPP exemplifies that algorithmic fairness principles can be effectively incorporated into biomolecular prediction tasks. This research sets a new standard for bias-aware computational biology, thus, being a step towards fair drug discovery and personalized medicine.

Fields of Study & Subjects

vaccines and immunoinformatics approaches Computational Drug Discovery Methods Machine Learning in Bioinformatics

Publication details

Journal / source
International Conference on Electronics, Materials Engineering & Nano-Technology (IEMENTech)
Pages
1–6
Publication date
6 February 2026
Language
EN