Efficient evolution of human antibodies from general protein language models and sequence information alone
Brian HieiD, Varun R. ShankeriD, Duo XuiD, Theodora U. J. BruuniD, Payton A. B. WeidenbacheriD, Shaogeng TangiD, Wesley C. H. WuiD, John E. PakiD, Peter S. KimiD
Abstract
Natural evolution must explore a vast landscape of possible sequences for desirable yet rare mutations, suggesting that learning from natural evolutionary strategies could accelerate artificial evolution. Here, we report that deep learning algorithms known as protein language models can evolve human antibodies with high efficiency, despite providing the models with no information about the target antigen, binding specificity, or protein structure, and also requiring no additional task-specific finetuning or supervision. We performed language-model-guided affinity maturation of seven diverse antibodies, screening 20 or fewer variants of each antibody across only two rounds of evolution. Our evolutionary campaigns improved the binding affinities of four clinically relevant antibodies up to 7-fold and three unmatured antibodies up to 160-fold across diverse viral antigens, with many designs also demonstrating improved thermostability and viral neutralization activity. Notably, our algorithm requires only a single wildtype sequence and computes recommended amino acid changes in less than a second. Moreover, the same models that improve antibody binding also guide efficient evolution across diverse protein families and selection pressures, indicating that these results generalize to many natural settings. Contrary to prevailing notions of evolution as difficult and resource-intensive, our results suggest that when constrained to a narrow manifold of evolutionary plausibility, evolution can become much easier, which we refer to as the "efficient manifold hypothesis."
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 458
- FWCI
- 74.96
- Field percentile
- top 1%
- Organism
- Assay
- Research Focus
- Approach
In-depth summary
Claude Opus 5.5 · xhigh effort · 5 Nov 2026 · from the full text of version 1 (11 Apr 2022) · CC BY-NC 4.0
Consensus likelihoods from general, unsupervised protein language models evolve seven human antibodies using only a wildtype sequence and <=20 tested variants, improving binding affinity up to 160-fold.
Laboratory directed evolution and machine-learning-guided antibody affinity maturation typically require high-throughput screens, structural coordinates, or antigen-specific training data. Using a consensus of six general protein language models trained on natural sequence diversity, a single wildtype sequence suffices to recommend affinity-enhancing mutations in under a second for low-throughput screening.
Key results
ShownScreening 20 or fewer language-model-recommended variants per antibody across two rounds improved binding affinity in all four matured clinical antibodies up to 7-fold.
Testing 8-14 single mutants in round 1 and 1-11 combinations in round 2 by biolayer interferometry improved monovalent Fab dissociation constants (Kd) 7-fold for MEDI8852 against influenza HA H7 HK17 (0.21 to 0.03 nM), 3.4-fold for mAb114 against Ebola GP, 5.1-fold for REGN10987 against SARS-CoV-2 Omicron RBD (1.3-fold against Beta S-6P), and 1.7-fold for S309 against Beta S-6P (vs 1.3-fold for sotrovimab). · Abstract; Efficient affinity maturation with general…; Figure legends; Tables
ShownLanguage-model-guided evolution improved binding affinities of all three unmatured human antibodies by 13-fold to 160-fold, including via non-natural evolutionary pathways.
Fab or IgG Kd improved by 160-fold for mAb114 UCA against ebolavirus GP (33-fold when excluding sites mutated in matured mAb114), up to 23-fold for MEDI8852 UCA against Group 2 HA H4 Hubei (2.6-fold against H1 Solomon, 5.4-fold against H7 HK17), and 13-fold for C143 against SARS-CoV-2 Beta S-6P (3.8-fold against Omicron RBD). · Abstract; Efficient affinity maturation with general…; Figure legends; Tables
ShownRecommended variants showed high expression and binding retention, with half of affinity-enhancing mutations occurring in framework regions and ~16% mutating to rare natural residues.
Across 7 antibodies, only 1 of 122 variants failed to express in Expi293F cells, 71-100% of round-1 Fabs retained sub-micromolar binding, and 15-71% improved Kd >=1.1-fold; 36 of 76 recommended substitutions (18 of 32 affinity-enhancing) fell in Kabat framework regions, and 5 of 32 (~16%) used residues with <=10% frequency in UniRef90 or abYsis. · Efficient affinity maturation with general…; Originality of affinity-enhancing substitutions; Figure legends; Tables
ShownAffinity-matured antibody designs maintained or increased thermostability and significantly enhanced pseudovirus neutralization potency.
Twenty-one of 31 affinity-enhancing Fabs had higher melting temperatures (Tm) than wildtype (all 31 kept Tm > 70 C; mAb114 UCA rose from 74.5 C to 82.5 C; S309 reached 72.8 C vs 69.6 C for sotrovimab), and pseudovirus IC50 improved 1.5-fold for mAb114 (Ebola), 2-fold for REGN10987 (Beta), and 32-fold for C143 (Beta; 19-fold for D614G), correlating with Kd change (Spearman r = 0.82, P = 1.9 x 10^-4). · Improved thermostability and neutralization of…; Figure legends
ShownThe same unsupervised six-model consensus significantly enriched high-fitness single-residue substitutions across eight non-antibody protein families in deep mutational scanning benchmarks.
Evaluating small sets (~10^1) of consensus-recommended substitutions against scanning mutagenesis data significantly increased hit rates (hypergeometric P < 0.05; 12% to 40% positive fraction), including from 7% background to 40% for beta-lactamase ampicillin resistance, 7% to 31% for influenza HA viral infectivity, and 3% to 20% for bacterial PafA enzyme kinetics. · Generality across diverse protein families; Natural protein evaluation based on scanning…; Figure legends; Tables
Methods, data and reproducibility
Scores single-amino-acid substitutions in heavy- and light-chain variable regions from a wildtype sequence using a consensus of six pretrained masked protein language models (ESM-1b and five ESM-1v models). Screens 8-14 single mutants in round 1 and 1-11 combinations in round 2 per antibody via biolayer interferometry, DSF, and pseudovirus neutralization, and benchmarks on 8 DMS datasets.
- Seven human IgG1 antibodies (MEDI8852, MEDI8852 UCA, mAb114, mAb114 UCA, S309, REGN10987, C143) and 122 evolved variants expressed in Expi293F cells · human · n = 7 wildtype antibodies; 122 variants (<=20 per antibody across 2 rounds) · DOI:10.5281/zenodo.6415457
- Recombinant viral target antigens (9 influenza A HAs, ebolavirus Mayinga GP delta309-489, and SARS-CoV-2 Wuhan-Hu-1, Beta, and Omicron Spike/RBD) · virus · n = 9 influenza HA strains, 1 Ebola GP, 3 SARS-CoV-2 Spike/RBD variants · AAG40168.1
- UniRef50 (March 2018) and UniRef90 (March 2020) sequence databases used to train ESM-1b and ESM-1v and estimate natural residue frequencies · multiple species · n = ~27M sequences (UniRef50); ~98M sequences (UniRef90, ~800 Ig regions)
- Deep mutational scanning (DMS) datasets from Livesey & Marsh and Markin et al. (PafA) across human, bacterial, and viral proteins · multiple species · n = 8 protein families (>=90% single-substitution coverage)
- Methods & tools
- ESM-1b, ESM-1v, abYsis version 3.4.1, mafft version 7.475, fuzzywuzzy version 0.18.0, neutcurve version 0.5.7, Octet System Data Analysis Software version 9.0.0.15, PR.ThermControl version 2.3.1, i-control version 2.0
- Software & environment
- Python packages neutcurve v0.5.7 and fuzzywuzzy v0.18.0; mafft v7.475; abYsis v3.4.1; Octet System Data Analysis Software v9.0.0.15; PR.ThermControl v2.3.1; i-control v2.0 [S9, S17, S18, S20, S21].
- Experimental conditions
- Computes recommended amino acid substitutions from a single wildtype sequence in less than a second [S1, S2].
- Reproducibility gaps
- Raw data are deposited on Zenodo (DOI:10.5281/zenodo.6415457) without a full http(s) URL in the text; specific CPU/GPU hardware used for ESM inference is not stated [S23].
Limitations (3)
- Because the language models are trained exclusively on natural protein sequences, this unsupervised strategy may be less effective for unnatural selection pressures, de novo designed proteins, or wildtype sequences already at a fitness peak [S7]. · stated by the authors
- In the deep mutational scanning benchmark, increasing likelihood stringency improved high-fitness variant acquisition across all tested proteins except P53 [S28]. · stated by the authors
- For five of the seven antibodies, second-round combinatorial variants were manually prioritized rather than exhaustively tested because all combinations exceeded ~10 variants [S9]. · stated by the authors
Who should read:Read this if you engineer antibodies or enzymes and want to prioritize high-affinity mutations from a single wildtype sequence without high-throughput screens or structures.
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- Nucleotide context models outperform protein language models for predicting antibody affinity maturation
Mackenzie M. Johnson, Kevin Sung, … Frederick A. Matsen IV·PLoS Computational BiologyPeer-reviewed·Dec 1, 2025
- Antibody Humanization via Protein Language Model and Neighbor Retrieval
Honggang Zou, Rongqing Yuan, … Jinbo Xu·bioRxiv·Sep 4, 2023
- Inverse folding of protein complexes with a structure-informed language model enables unsupervised antibody evolution
Varun R. Shanker, Theodora U. J. Bruun, … Peter S. Kim·bioRxiv·Dec 19, 2023
- Antibody optimization enabled by artificial intelligence predictions of binding affinity and naturalness
Sharrol T. Bachas, Goran Rakocevic, … Roberto Spreafico·bioRxiv·Aug 16, 2022
- Rapid protein evolution by few-shot learning with a protein language model
Kaiyi Jiang, Zhaoqing Yan, … Omar O. Abudayyeh·SciencePeer-reviewed·Nov 21, 2024
- Efficient, Few-shot Directed Evolution with Energy Rank Alignment
Sebastian Ibarraran, Shriram Chennakesavalu, … Grant M. Rotskoff·Journal of Chemical Information and ModelingPeer-reviewed·Aug 26, 2026