Exploiting Topology of Protein Language Model Attention Maps for Token Classification
Keywords: protein language models, protein property prediction, topological data analysis, attention maps, transformers
TL;DR: We propose to use Topological Data Analysis of attention maps in a protein language model for amino-acid classification
Abstract: In this paper, we introduce a method to extract topological features from transformer-based protein language models. Our method leverages the persistent homology of attention maps to generate features for token (per amino-acid) classification tasks and demonstrate its relevance in a biological context. We implement our method on transformer-based protein language models using the family of ESM-2 models. Specifically, we demonstrate that minimum spanning trees, derived from attention matrices, encode structurally significant information about proteins. In our experiments, we combine these topological features with standard embeddings from ESM-2. Our method outperforms traditional approaches and other transformer-based methods with a similar number of parameters in several binding site identification tasks and achieves state-of-the-art performance in conservation prediction tasks. Our results highlight the potential of this hybrid approach in advancing the understanding and prediction of protein functions.
Supplementary Material: zip
Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 11536
Loading