Use Structural Evolution Online
Commercially Available Structural Evolution No-Code Web Server
Structural Evolution: Unsupervised Evolution of Proteins and Antibody Complexes with a Structure-Informed Language Model
Structural Evolution is a structure-informed protein language model workflow that proposes beneficial amino acid substitutions for proteins and antibody-antigen complexes without any task-specific training data. It was developed by Varun R. Shanker, Theodora U. J. Bruun, Brian L. Hie and Peter S. Kim at Stanford University, Sarafan ChEM-H and the Chan Zuckerberg Biohub, and published in Science in 2024. The method builds on ESM-IF1, an inverse folding model trained on single-chain structures, and extends it to score variants of multichain complexes. Open-source code is provided by the authors on GitHub.
Evolving a protein in the lab is hard because almost all mutations are deleterious or neutral, and mutations accumulated over successive rounds tend to destabilize the structure and reduce evolvability. Sequence-only language models learn general design principles but do not see the three-dimensional structure that determines function, activity and evolvability.
Structural Evolution addresses both problems. The model predicts each amino acid from the preceding sequence and the full backbone coordinates, so sequences it rates as likely are expected to fold into the input backbone. Restricting exploration to high-likelihood sequences acts as a prior for high-fitness variants, without defining an explicit fitness function.
Property | Detail |
|---|---|
Underlying model | ESM-IF1, a structure-informed autoregressive language model |
Input | Backbone coordinates of a protein or complex (PDB or CIF file) and the target chain to mutate |
Training data needed | None task-specific; fully unsupervised |
Search space scored | About 4,300 single-residue substitutions across antibody VH and VL regions |
Publication | Science 385(6704):46-53 (2024) |
How Structure-Informed Language Model Evolution Works
Condition on structure: the model reads the backbone coordinates of the wild-type protein, or of the whole antibody-antigen complex.
Score mutations: every possible single substitution is scored by sequence log likelihood given the backbone.
Pick top candidates: in the antibody campaigns, the top ten predictions at unique residues per chain were tested in round one.
Combine mutations: in round two, the model was used again to select up to five top-scoring combinations of mutations per antibody chain.
Benchmark Results: Zero-Shot Variant Effect Prediction
Without any experimental training data, the approach recovered top fifth percentile mutations in 9 of 10 deep mutational scanning datasets, compared with 2 of 10 for the sequence-only model ESM-1v. For antibodies, it outperformed ESM-1v, AbLang and abYsis across five mutagenesis datasets, and performance was highest when both antibody and antigen structure were supplied. For CR9114 binding to H1, the correlation rose from 0.17 with antibody alone to 0.65 with the full complex.
Antibody Engineering Results Against SARS-CoV-2 Variants
The authors evolved two clinical antibodies, LY-CoV1404 (bebtelovimab) and SA58, by testing only tens of variants. About one-third of the 20 single substitutions per antibody improved neutralization of the original strain, and nearly half of the LY-CoV1404 substitutions improved neutralization of BQ.1.1.
Outcome | Reported result |
|---|---|
Best LY-CoV1404 design, BQ.1.1 neutralization | 25-fold improvement (structure-informed) vs 2-fold (sequence-only ensemble) |
Best SA58 design, BQ.1.1 neutralization | Over 14-fold (structure-informed) vs 4-fold (sequence-only ensemble) |
LY-CoV1404 binding to BQ.1.1 RBD | Up to about 27-fold better apparent affinity |
SA58 binding to XBB.1.5 RBD | Up to 37-fold better apparent affinity |
Round-two multi-mutation designs | All 11 LY-CoV1404 and 5 SA58 designs beat wild type |
The evolved antibodies showed no substantial change in polyspecificity, and the top SA58 design also improved BA.1 neutralization by over 3-fold. Several of the best substitutions lay in framework regions far from the binding interface, which a CDR-focused search could miss.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with Structural Evolution on Tamarind Bio
Antibody affinity and potency maturation: rank single substitutions for therapeutic antibodies using the antibody-antigen complex structure.
Resilience to escape variants: re-score variants against a new antigen structure, as done for BQ.1.1 and XBB.1.5.
General protein engineering: prioritize mutations for enzymes, kinases and other proteins, such as the top predictions for MAPK1 that matched inhibitor-resistance substitutions.
Small, focused libraries: shortlist a few dozen variants for experimental testing instead of exhaustive scans.
How to Use Structural Evolution on Tamarind Bio
Open the tool: log in to tamarind.bio and select Structural Evolution.
Provide a structure: upload a PDB or CIF file of your protein or complex, ideally the antibody bound to its antigen, since the authors found complex structure gave the best predictions.
Choose the chain to mutate: specify the target chain, for example the antibody heavy or light chain, if exposed.
Set parameters: adjust any scoring or output options if exposed.
Run the job: submit and wait for the log-likelihood scoring to finish.
Download results: retrieve the ranked substitutions.
Select candidates: pick top substitutions at unique residues and consider combining beneficial ones in a second round.
Validate experimentally: test variants in the assay that matters for your goal, since the model does not predict a specific property.
Things to Keep in Mind
The model uses backbone coordinates only, so it has limited ability to find mutations that help by changing the backbone of the mature antibody.
ESM-IF1 was trained only on single-chain structures; its multichain use is an extension reported by the authors.
It does not model an explicit function, so predicted variants are likely to fold well but are not guaranteed to improve any given property.
Improved binding was not sufficient for improved neutralization: four recommended mutations improved binding but were neutral or deleterious for neutralization.
Source: Shanker VR, Bruun TUJ, Hie BL, Kim PS. Unsupervised evolution of protein and antibody complexes with a structure-informed language model. Science 385(6704):46-53 (2024), DOI 10.1126/science.adk8946. Code: https://github.com/varun-shanker/structural-evolution.