This is the authors' abstract. We don't add key points for this paper.
Detecting structural similarity at the local level between proteins is central to understanding function and evolution, yet most approaches require 3D models. In this work, we show that protein language models (pLMs), solely using sequence data as input, implicitly capture fine-grained structural signals that can be leveraged to identify such similarities. By mean-pooling residue embeddings over sliding windows and comparing them across proteins with cosine similarity, we find diagonal patterns that reflect locally aligned regions, even without sequence identity. Building on this insight, we introduce a framework for detecting locally aligned structural regions directly from sequences, supporting the development of scalable methods for structural annotation and comparison. The effectiveness and scalability of this framework are further demonstrated through a case study identifying SRC homology 3 (SH3) domains within a large-scale PDB subset, where our approach successfully recovered structurally conserved motifs across diverse sequence contexts. Ultimately, this work provides a lightweight alternative to structure-based methods, paving the way for high-throughput structural discovery using sequence data alone.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Ma et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: