BERTology Meets Biology: Interpreting Attention in Protein Language Models

2020/06/26 by Jesse Vig, Vig, Jesse, Ali Madani +9 · 19 citations
Biochemistry, Genetics and Molecular Biology · Computer Science · #Bioinformatics and Genomic Networks #Biomolecules (q-bio.BM) #Computation and Language (cs.CL) #FOS: Biological sciences #FOS: Computer and information sciences #I.2 #Machine Learning (cs.LG) #Machine Learning in Bioinformatics #Protein Structure and Dynamics #cs.CL #cs.LG #q-bio.BM

paper · pdf · doi:10.48550/arxiv.2006.15222

To appear in ICLR 2021

openalex publication_date 2020/06/26 · arxiv created 2021/03/28 · arxiv updated 2021/03/30 · openalex created_date 2022/07/26 · openalex updated_date 2026/07/28

Abstract

Transformer architectures have proven to learn useful representations for protein classification and generation tasks. However, these representations present challenges in interpretability. In this work, we demonstrate a set of methods for analyzing protein Transformer models through the lens of attention. We show that attention: (1) captures the folding structure of proteins, connecting amino acids that are far apart in the underlying sequence, but spatially close in the three-dimensional structure, (2) targets binding sites, a key functional component of proteins, and (3) focuses on progressively more complex biophysical properties with increasing layer depth. We find this behavior to be consistent across three Transformer architectures (BERT, ALBERT, XLNet) and two distinct protein datasets. We also present a three-dimensional visualization of the interaction between attention and protein structure. Code for visualization and analysis is available at https://github.com/salesforce/provis.

Cited by

• Quantifying the Role of OpenFold Components in Protein Structure Prediction
• Machine learning approaches for interpretable antibody property prediction using structural data
• Education Paradigm Shift To Maintain Human Competitive Advantage Over AI
• Isotropy and Geometry of Pretrained Protein LMs
• Decoding the Grammar of Protein-Protein Interaction Interfaces with Multimodal Representations
• Directed Evolution of Proteins via Bayesian Optimization in Embedding Space
• ProtSAE: Disentangling and Interpreting Protein Language Models via Semantically-Guided Sparse Autoencoders
• Nucleotide dependency analysis of genomic language models detects functional elements
• A Comprehensive Review of Transformer-based language models for Protein Sequence Analysis and Design
• Enhancing Safe and Controllable Protein Generation via Knowledge Preference Optimization
• Automated Neuron Labelling Enables Generative Steering and Interpretability in Protein Language Models
• Toward the Explainability of Protein Language Models
• Squidly: Enzyme Catalytic Residue Prediction Harnessing a Biology-Informed Contrastive Learning Framework
• Transformers and genome language models
• Into the Unknown: From Structure to Disorder in Protein Function Prediction
• Gene finding revisited: improved robustness through structured decoding from learned embeddings
• Reverse Distillation: Consistently Scaling Protein Language Model Representations
• Retrieval and competition: how a protein foundation model starts a protein
• RAG-ESM: Improving pretrained protein language models via sequence retrieval

Related

• AlphaFold reveals but sometimes distorts an organizational principle of protein folding
• Synthesizing Proteins on the Graphics Card. Protein Folding and the Limits of Critical AI Studies
• Masked Language Modeling for Proteins via Linearly Scalable Long-Context Transformers
• Protein Secondary Structure Prediction Using Transformers
• Insights into the inner workings of transformer models for protein function prediction