Geometric Deep Learning for Protein Function Prediction: Integrating 3D Structural Inductive Bias with Interaction Graphs

Abstract

The accurate prediction of protein function from raw amino acid sequences and structural data remains a central challenge in computational biology, essential for advancing drug discovery and understanding cellular mechanisms. While sequence-based methods have historically dominated the field due to data abundance, they often fail to capture the functional implications of distant homology where sequence similarity is low but structural conservation is high. With the advent of highly accurate structure prediction systems, the availability of 3D protein structures has exploded, necessitating novel architectures capable of leveraging this geometric data. This paper introduces GeoProtNet, a comprehensive framework that utilizes Geometric Deep Learning (GDL) to predict protein function, specifically Gene Ontology (GO) terms. We propose a hybrid architecture that processes the protein as a geodesic graph on a Riemannian manifold to capture local chemical environments, while simultaneously integrating global context through a higher-order Protein-Protein Interaction (PPI) graph. By enforcing SE(3)-equivariance within the message-passing mechanism, our model ensures robustness against rotational and translational variations inherent in coordinate data. Experimental results demonstrate that GeoProtNet significantly outperforms state-of-the-art sequence-based and structure-based baselines, particularly in the twilight zone of low sequence identity.