CoChat
Sign Up
Cited by 216Protein Structure and DynamicsMonoclonal and Polyclonal Antibodies Researchvaccines and immunoinformatics approaches

The DOI matches its Crossref record; 75 of the 75 references it cites were found in a registry or an index.

Antibody structure prediction using interpretable deep learning

Open access
Ruffolo, Jeffrey A., Sulam, Jeremias, Gray, Jeffrey J.·2022·Patterns
DOI 10.1016/j.patter.2021.100406PMC PMC8848015

Published version of a preprint

Open this paper in CoChat — free accountOpen PDF

Abstract

Abstract Therapeutic antibodies make up a rapidly growing segment of the biologics market. However, rational design of antibodies is hindered by reliance on experimental methods for determining antibody structures. In recent years, deep learning methods have driven significant advances in general protein structure prediction. Here, we present DeepAb, a deep learning method for predicting accurate antibody F V structures from sequence. We evaluate DeepAb on two benchmark sets – one balanced for structural diversity and the other composed of clinical-stage therapeutic antibodies – and find that our method consistently outperforms the leading alternatives. Previous deep learning methods have operated as “black boxes” and offered few insights into their predictions. By introducing a directly interpretable attention mechanism, we show that our network attends to physically important residue pairs. For example, in prediction of one CDR H3 residue conformation, the network attends to proximal aromatics and a key hydrogen bonding interaction that constrain the loop conformation. Finally, we present a novel mutant scoring metric derived from network confidence and show that for a particular antibody, all eight of the top-ranked mutations improve binding affinity. These results suggest that this model will be useful for a broad range of antibody prediction and design tasks. Significance Accurate structure models are critical for understanding the properties of potential therapeutic antibodies. Conventional methods for protein structure determination require significant investments of time and resources and may fail. Although greatly improved, methods for general protein structure prediction still cannot consistently provide the accuracy necessary to understand or design antibodies. We present a deep learning method for antibody structure prediction and demonstrate improvement over alternatives on diverse, therapeutically relevant benchmarks. In addition to its improved accuracy, our method reveals interpretable outputs about specific amino acids and residue interactions that should facilitate design of novel therapeutic antibodies.

Overview

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1016/j.patter.2021.100406

Cites

75 of 75 references verified0 citation errors · 0 not checkable

Statistics

Figures

Graphical abstract
Graphical abstract CC BY-NC-ND · PubMed Central
Figure 1 Diagram of DeepAb method for antibody structure prediction Starting from heavy and light chain sequences, the network predicts a set of inter-residue geometries describing the FV structure. Predictions are used for guided structure realization with Rosetta. Two interpretable components of the network are highlighted: a pre-trained antibody sequence model and output attention mechanisms.
Figure 1 Diagram of DeepAb method for antibody structure prediction Starting from heavy and light chain sequences, the network predicts a set of inter-residue geometries describing the FV structure. Predictions are used for guided structure realization with Rosetta. Two interpretable components of the network are highlighted: a pre-trained antibody sequence model and output attention mechanisms.CC BY-NC-ND · PubMed Central
Figure 2 Comparison of CDR H3 loop structure prediction accuracy (A) Average RMSD of H3 loops predicted by RosettaAntibody-G (RAb), RepertoireBuilder (RB), ABodyBuilder (ABB), and DeepAb on the two benchmarks. Error bars show standard deviations for each method on each benchmark. (B) Average RMSD of H3 loops by length for all benchmark targets. Error bars show standard deviations for loop lengths corresponding to more than one target. (C) Direct comparison of DeepAb and alternative methods H3 loop RMSDs, with diagonal band indicating predictions that were within ±0.25 Å. (D) Comparison of native rituximab H3 loop structure (white, PDB: 3PP3) to predictions from DeepAb (green, 2.1 Å RMSD) and alternative methods (blue, 3.3–4.1 Å RMSD). (E) Comparison of native sonepcizumab H3 loop structure (white, PDB: 3I9G) to predictions from DeepAb (green, 1.8 Å RMSD) and alternative methods (blue, 2.9–3.9 Å RMSD).
Figure 2 Comparison of CDR H3 loop structure prediction accuracy (A) Average RMSD of H3 loops predicted by RosettaAntibody-G (RAb), RepertoireBuilder (RB), ABodyBuilder (ABB), and DeepAb on the two benchmarks. Error bars show standard deviations for each method on each benchmark. (B) Average RMSD of H3 loops by length for all benchmark targets. Error bars show standard deviations for loop lengths corresponding to more than one target. (C) Direct comparison of DeepAb and alternative methods H3 loop RMSDs, with diagonal band indicating predictions that were within ±0.25 Å. (D) Comparison of native rituximab H3 loop structure (white, PDB: 3PP3) to predictions from DeepAb (green, 2.1 Å RMSD) and alternative methods (blue, 3.3–4.1 Å RMSD). (E) Comparison of native sonepcizumab H3 loop structure (white, PDB: 3I9G) to predictions from DeepAb (green, 1.8 Å RMSD) and alternative methods (blue, 2.9–3.9 Å RMSD).CC BY-NC-ND · PubMed Central
Figure 3 Interpretability of model components (A) Diagram of attention mechanism (with attention matrix A and value matrix V) and example H3 loop attention matrix, with attention on other loops indicated. Attention values increase from blue to red. (B) Model attention over FV structure while predicting each CDR loop for an anti-peptide antibody (PDB: 4H0H). (C) Key interactions for H3 loop structure prediction identified by attention. The top five non-H3 attended residues (H32-Y, L32-Y, L49-Y, L55-F, and L91-S) are labeled, as well as an H3 residue participating in a hydrogen bond (H100-S). (D) Two-dimensional t-SNE projection of sequence-averaged LSTM embeddings labeled by source species. (E) Two-dimensional t-SNE projects of LSTM embeddings averaged over CDR1 loop residues labeled by loop structural clusters.
Figure 3 Interpretability of model components (A) Diagram of attention mechanism (with attention matrix A and value matrix V) and example H3 loop attention matrix, with attention on other loops indicated. Attention values increase from blue to red. (B) Model attention over FV structure while predicting each CDR loop for an anti-peptide antibody (PDB: 4H0H). (C) Key interactions for H3 loop structure prediction identified by attention. The top five non-H3 attended residues (H32-Y, L32-Y, L49-Y, L55-F, and L91-S) are labeled, as well as an H3 residue participating in a hydrogen bond (H100-S). (D) Two-dimensional t-SNE projection of sequence-averaged LSTM embeddings labeled by source species. (E) Two-dimensional t-SNE projects of LSTM embeddings averaged over CDR1 loop residues labeled by loop structural clusters.CC BY-NC-ND · PubMed Central
Figure 4 Prediction of mutational effects with DeepAb model (A) Diagram of ΔCCE calculation for model output predictions for an arbitrary residue pair. Plots show the change in probability density of the predicted geometries for the residue pair after making a mutation. (B) Plot of the combined network metric against experimental binding enrichment over wild type, with negative values corresponding to beneficial mutations for both axes. True positive predictions (red) and mutations to wild type cysteines (yellow) are highlighted. (C) Receiver operating characteristic for predicting experimental binding enrichment over wild type with the combined network metric and each component metric. Area under the curve (AUC) values are provided for each metric. (D) Position of true positive predictions on anti-lysozyme FV structure. (E) Positive predictive value for mutants ranked by the combined metric. (F) Comparison of ΔCCE for a designed eight-point variant (D44.1des, red) to sequences with random mutations at the same positions.
Figure 4 Prediction of mutational effects with DeepAb model (A) Diagram of ΔCCE calculation for model output predictions for an arbitrary residue pair. Plots show the change in probability density of the predicted geometries for the residue pair after making a mutation. (B) Plot of the combined network metric against experimental binding enrichment over wild type, with negative values corresponding to beneficial mutations for both axes. True positive predictions (red) and mutations to wild type cysteines (yellow) are highlighted. (C) Receiver operating characteristic for predicting experimental binding enrichment over wild type with the combined network metric and each component metric. Area under the curve (AUC) values are provided for each metric. (D) Position of true positive predictions on anti-lysozyme FV structure. (E) Positive predictive value for mutants ranked by the combined metric. (F) Comparison of ΔCCE for a designed eight-point variant (D44.1des, red) to sequences with random mutations at the same positions.CC BY-NC-ND · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Research works printed but not linked: 1

  1. Linder, 2020, Fast differentiable DNA and protein sequence optimization for molecular design, arXiv

Cited by

Sorted by citations

See all 216

Related

See all

Metadata

Authors
Ruffolo, J. A., Sulam, J., & Gray, J. J.
Affiliations
Johns Hopkins University (US)
Corresponding author
Gray, Jeffrey J.
ORCID
Venue
Patterns
Publisher
Elsevier BV
Published
2022-02
Volume
3
Issue
2
Pages
100406
ISSN
2666-3899 (ISSN-L)
DOI
10.1016/j.patter.2021.100406
PMC
PMC8848015
OpenAlex
W4226108318
Cited by
216
References
75
Citations by year
2026: 26 · 2025: 42 · 2024: 50 · 2023: 58 · 2022: 37
FWCI
11.62
Type
journal article
Language
en
Open access
Licence
Locations
  • Patterns (publishedVersion, cc-by-nc-nd)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by-nc)
  • PubMed Central (publishedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
Funders
  • Maryland Advanced Research Computing Center
  • National Institutes of Health (R01-GM078221, T32-GM008403)
  • AstraZeneca plc
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Protein Structure and Dynamics
  • Life Sciences › Immunology and Microbiology › Immunology › Monoclonal and Polyclonal Antibodies Research
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › vaccines and immunoinformatics approaches
Keywords
antibody structure prediction, therapeutic antibodies, explainable deep learning, attention mechanism, binding affinity, rational antibody design, hydrogen bonding

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review