henhao
Sign in
BioinformaticsbioRxiv
Preprint

Antibody optimization enabled by artificial intelligence predictions of binding affinity and naturalness

Sharrol T. BachasiD, Goran RakoceviciD, David SpenceriD, Anand V. SastryiD, Robel Gemechu HaileiD, John M. SuttoniD, George W. Kasun, Andrew L. StachyraiD, Jahir M. GutierreziD, Edriss YassineiD, Borka Medjo, Vincent BlayiD, Christa Kohnert, Jennifer T. Stanton, Alexander Brown, Nebojša Tijanić, Cailen M. McCloskey, Rebecca Viazzo, Rebecca Consbruck, Hayley Carter, S. G. LEVINEiD, Shaheed A. AbdulhaqqiD, Jacob D. ShauliD, Abigail B. VenturaiD, Randal S. OlsoniD, , Roberto SpreaficoiD

PDFDOI
Structured summary available · Jump to summary ↓

Abstract

Traditional antibody optimization approaches involve screening a small subset of the available sequence space, often resulting in drug candidates with suboptimal binding affinity, developability or immunogenicity. Based on two distinct antibodies, we demonstrate that deep contextual language models trained on high-throughput affinity data can quantitatively predict binding of unseen antibody sequence variants. These variants span a KD range of three orders of magnitude over a large mutational space. Our models reveal strong epistatic effects, which highlight the need for intelligent screening approaches. In addition, we introduce the modeling of "naturalness", a metric that scores antibody variants for similarity to natural immunoglobulins. We show that naturalness is associated with measures of drug developability and immunogenicity, and that it can be optimized alongside binding affinity using a genetic algorithm. This approach promises to accelerate and improve antibody engineering, and may increase the success rate in developing novel antibody and related drug candidates.

02 /Specification & Classification

Click any tag to browse matching papers
Citations
63
OpenAlex Topics

Abstract summary

Verified abstract synthesis

Claude Opus 5.5 · xhigh effort · 18 Oct 2026 · from the abstract

Language models trained on high-throughput affinity data predict KD of unseen antibody variants across three orders of magnitude, and a "naturalness" score can be optimised alongside affinity.

Antibody optimisation normally screens a small slice of sequence space, so candidates can carry weak affinity or poor developability and immunogenicity. Learning affinity directly from large mutational libraries, and adding a score for resemblance to natural immunoglobulins, lets both properties be searched at once.

Key results

  1. SuggestedContextual protein language models trained on high-throughput affinity measurements predict binding of antibody variants they were not trained on, quantitatively rather than as a ranking.

    Demonstrated for two distinct antibodies, over variants covering a KD range of three orders of magnitude across a large mutational space; no correlation, error or held-out split is reported in the abstract. · Abstract

  2. SuggestedBinding across the mutational libraries shows strong epistasis, so mutations cannot be assessed independently.

    Reported as a model finding used to argue for guided rather than random screening; no magnitude, fraction of epistatic pairs or statistics are given. · Abstract

  3. SuggestedA "naturalness" score for similarity to natural immunoglobulins tracks developability and immunogenicity measures and can be co-optimised with affinity by a genetic algorithm.

    Stated as an association with developability and immunogenicity metrics; neither the metrics, the correlation strength, nor experimental results for the co-optimised variants appear in the abstract. · Abstract

Methods, data and reproducibility

Deep contextual language models were trained on high-throughput binding data from mutational libraries of two antibodies to predict KD of unseen variants. A separate naturalness model scores how closely a variant resembles natural immunoglobulins, and a genetic algorithm searches sequence space for variants scoring well on affinity and naturalness together.

  • high-throughput binding affinity measurements over a large mutational space for two antibodies · n = 2 antibodies; variants spanning a KD range of three orders of magnitude
Methods & tools
deep contextual language model, genetic algorithm
Reproducibility gaps
No repository, dataset release, model architecture or library deposition is named in the abstract.
Limitations (2)
  • Only two antibodies underlie the affinity results, with no library size, accuracy metric or held-out evaluation described in the abstract. · noted in this summary
  • Naturalness is linked to developability and immunogenicity by assertion; no effect sizes, and no experimental characterisation of genetic-algorithm outputs, are reported here. · noted in this summary

Who should read:Read this if you engineer antibodies and want sequence models that score affinity and resemblance to natural immunoglobulins jointly instead of screening a narrow library slice.

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

Similarity Map
SelectedSame fieldCross-field
THISAgarwal '25Ashraf '24Furui '25Zou '23Hossain '25Vashchenko '22
81% semantic matchSynthetic Biology
Aditya A. Agarwal, James Harrang, David Noble+12 more·mAbs·Jul 22, 2025· 17 citations