RES-LT: Extracting Higher-Order Topological Features from Protein Language Models for Enhanced Per-Residue Classification

Abstract We introduce RES-LT (Residual Local Topology), a novel topological data analysis (TDA) approach that extracts higher-order structural information from transformer-based protein language models. RES-LT utilizes both $${{H}_{0}}$$ and $${{H}_{1}}$$ persistent homology to characterize residue-residue interactions in proteins through attention maps, generating biologically relevant features for per-residue classification. Implemented on the ESM-2 model family, our framework integrates $${{H}_{0}}$$ and $${{H}_{1}}$$ topological features with standard embeddings to create a powerful hybrid representation. Extensive evaluation demonstrates that RES-LT achieves state-of-the-art performance in conservation prediction and significantly outperforms both traditional approaches and comparable transformer-based methods in binding site identification.