EvoGraph: synergizing evolutionary semantics and tail-aware graph contrastive learning for protein–protein interaction prediction

Abstract

Accurate prediction of multilabel protein–protein interactions is pivotal for deciphering molecular mechanisms and accelerating drug discovery. However, existing methods are often hindered by the inherent long-tailed distribution of interaction types, leading to poor generalization on rare but biologically critical categories. To address this challenge, we propose EvoGraph, a novel framework that synergizes deep evolutionary semantics with tail-aware graph representation learning. Specifically, EvoGraph first leverages the pretrained protein language model to establish a semantic foundation, compensating for sparse signals in rare classes. Driven by the critical need to disentangle complex interaction types and mitigate head-class dominance, we integrate two core mechanisms. To resolve the structural heterogeneity of diverse interaction types, we design a multi-view graph neural network encoder that aggregates type-specific topological contexts. Simultaneously, to counter label imbalance, we introduce a tail-aware contrastive learning module, which regularizes the latent space by clustering minority samples, effectively preventing them from being overshadowed by head classes. Finally, these features are synthesized via a hybrid fusion strategy and optimized with an asymmetric loss. Extensive experiments on four benchmark datasets demonstrate that EvoGraph significantly outperforms state-of-the-art baselines, achieving substantial improvements in identifying tail interaction types without compromising overall performance. The implementation of EvoGraph is available at https://github.com/AlphaZYL/EvoGraph.