BERTology Meets Biology: Interpreting Attention in Protein Language Models

By Jesse Vig, Ali Madani, L. Varshney, Caiming Xiong, Richard Socher, Nazneen Rajani. 2020

381 citations recorded by the source corpus.

Transformer architectures have proven to learn useful representations for protein classification and generation tasks. However, these representations present challenges in interpretability. Through the lens of attention, we analyze the inner workings of the Transformer and explore how the model discerns structural and functional properties of proteins. 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 also present a three-dimensional visualization of the interaction between attention and protein structure. Our findings align with known biological processes and provide a tool to aid discovery in protein engineering and synthetic biology. The code for visualization and analysis is available at https://github.com/salesforce/provis.

Subfields: Molecular Biology, Natural Language Processing, Computational Biology & Bioinformatics

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Molecular Biology

In Natural Language Processing

In Computational Biology & Bioinformatics

Citations in the corpus

Rethinking Attention with Performers (2020, methodology): “Furthermore, we analyze the amino-acid similarity matrix estimated from the attention matrices produced by the Performer model, as described in Vig et al. (Vig et al., 2020).”
Transformer protein language models are unsupervised structure learners (2020, background): “In particular Vig et al. (2020) discovered a correlation between self-attention maps and contact patterns, suggesting they could be used for contact prediction.”
Language models enable zero-shot prediction of the effects of mutations on protein function (2021, background): “Vig et al. [34] and Rao et al. [24] found that transformer attention corresponds to known biological properties such as structure and binding sites and can be used to predict contacts.”
Language Models are Open Knowledge Graphs (2020, methodology): “…(Vaswani et al., 2017) based language models (e.g., BERT and GPT-3), substantial recent work focuses on both visualizing and analyzing the attention (Vig, 2019; Jain & Wallace, 2019; Clark et al., 2019; Michel et al., 2019; Vig et al., 2020; Ramsauer et al., 2020; Hendrycks et al., 2020).”
Graph Representation Learning in Biomedicine and Healthcare (2021, positive): “Additionally, an attention mechanism for protein-sequence embeddings generated by the language model BERT (for ‘bidirectional encoder representations from transformers’) has facilitated the interpretability of the predictions of such networks 129,130 .”
The language of proteins: NLP, machine learning & protein sequences (2021, background): “…a model from scratch, especially when labeled data is scarce. how attention-based models work in proteins, and the parts of proteins that attention heads focus on across different tasks (e.g. which residues are most relevant in predicting which AAs interact in a protein’s 3D structure) [105].”
The Nucleotide Transformer: Building and Evaluating Robust Foundation Models for Human Genomics (2024, methodology): “Finally, to understand whether the model encodes for genomic features, we computed the attention percentages for all heads and layers of the Nucleotide Transformer models over the same nine genomic features, following the methodology introduced in previous work [31].”
Deep neural language modeling enables functional protein generation across families (2021, background): “The self-attention mechanism in each layer infers pairwise interaction relationships between all positions in its input sequence 32 .”