Learning a generalized graph transformer for protein function prediction in dissimilar sequences.

Abstract

Background: In the face of a growing disparity between high-throughput sequence data and low-throughput experimental studies, the emerging field of deep learning stands as a promising alternative. Generally, many data-driven approaches are capable of facilitating fast and accurate predictions of protein functions. Nevertheless, the inherent statistical nature of deep learning techniques may limit their generalization capabilities when applied to novel nonhomologous proteins that diverge significantly from existing ones.

Results: In this work, we herein propose a novel, generalized approach named Graph Adversarial Learning with Alignment (GALA) for protein function prediction. Our GALA method integrates a graph transformer architecture with an attention pooling module to extract embeddings from both protein sequences and structures, facilitating unified learning of protein representations. Particularly noteworthy, GALA incorporates a domain discriminator conditioned on both learnable representations and predicted probabilities, which undergoes adversarial learning to ensure representation invariance across diverse environments. To optimize the model with abundant label information, we generate label embeddings in the hidden space, explicitly aligning them with protein representations. Benchmarked on datasets derived from the PDB database and Swiss-Prot database, our GALA achieves considerable performance comparable to several state-of-the-art methods. Even more, GALA demonstrates wonderful biological interpretability by identifying significant functional residues associated with Gene Ontology terms through class activation mapping.

Conclusions: GALA, which leverages adversarial learning and label embedding alignment to acquire domain-invariant protein representations, exhibits outstanding generalizability in function prediction for proteins from previously unseen sequence space. By incorporating the structures predicted by AlphaFold2, GALA demonstrates significant potential for function annotation in newly discovered sequences. A detailed implementation of our GALA is available at https://github.com/fuyw-aisw/GALA.