Effective structure-aware protein alignment via residue-level contrastive learning

Protein alignment is indispensable for biological discovery, supporting structure comparison, functional annotation, and evolutionary inference. While structure-based methods are highly effective at detecting structural similarity, their applicability is constrained by the limited availability of experimentally resolved protein structures and high computational cost. Sequence-based approaches using pretrained protein language models (pLMs) provide scalable alternatives, yet supervised methods based on differentiable dynamic programming have not consistently outperformed simpler unsupervised strategies. Here, we present CLAlign, a structure-aware protein alignment framework based on contrastive learning. CLAlign fine-tunes a pretrained pLM to generate structure-aware residue-level embeddings enriched with structural context, without relying on differentiable dynamic programming. It represents the first supervised approach to consistently outperform unsupervised pLM-based methods, and it naturally extends to both sequence- and structure-based alignment by flexibly adopting different protein language model encoders. CLAlign achieves state-of-the-art accuracy on the MALIDUP and MALISAM benchmarks, outperforming existing sequence-based methods by large margins while remaining highly efficient. Moreover, its alignment scores show clear biological interpretability: in remote homology detection on SCOPe, CLAlign performs comparably to structure-based methods such as TM-align while far exceeding all sequence-based baselines. Together, these results establish CLAlign as a simple, extensible, and biologically meaningful framework for protein alignment.