Machine learning optimization of candidate antibody yields highly diverse sub-nanomolar affinity antibody libraries

Lin Li, Esther Gupta, John Spaeth, Leslie Shing, Rafael Jaimes, Emily Engelhart, Randolph Lopez, Rajmonda S. Caceres, Tristan Bepler, Matthew WalshNature Communications 2023 · DOI

TL;DR

Researchers have developed an end-to-end machine learning framework that uses protein language models and Bayesian optimization to design antibody libraries. Their method produced antibodies with nearly 30 times higher binding affinity than traditional directed evolution methods while maintaining significantly higher sequence diversity.

What Problem Does This Paper Address?

Therapeutic antibodies are the backbone of modern medicine, treating everything from cancer to autoimmune diseases. However, finding the "perfect" antibody is a needle-in-a-haystack problem. The sequence space—the total number of ways you can arrange amino acids—is astronomically large.

Traditional methods like directed evolution rely on iterative rounds of random mutation and screening. This process is slow (often taking over a year), labor-intensive, and frequently gets stuck in "local optima," producing antibodies that bind well but have poor physical properties (like being too sticky or unstable). The authors sought to replace this slow, trial-and-error process with a predictive, "in silico" (computational) design cycle that can explore the sequence space more intelligently.

Methods

The authors built a five-step pipeline that combines high-throughput experimentation with state-of-the-art AI:

  1. Data Priming: They started with a "weak" candidate antibody (Ab-14) and generated ~52,000 random variants. They measured the binding affinity of these variants using a high-throughput yeast mating assay.
  2. Language Modeling: They used BERT-style masked language models (similar to the tech behind LLMs) pre-trained on millions of natural protein sequences (Pfam) and human antibody sequences (OAS).
  3. Supervised Fine-tuning: They fine-tuned these models on their specific binding data to create a "sequence-to-affinity" predictor. Crucially, they used Ensemble and Gaussian Process models to quantify uncertainty—essentially teaching the AI to know when it was "guessing."
  4. Bayesian Optimization: Using the AI as a "fitness landscape," they used three sampling algorithms—Hill Climbing, Genetic Algorithms, and Gibbs Sampling—to find the best possible sequences.
  5. Empirical Validation: Finally, they synthesized the AI's top designs and tested them in the lab to see if the predictions held up.

Datasets

  • Pre-training: The Pfam database (general protein families) and the Observed Antibody Space (OAS), which contains hundreds of millions of human antibody sequences.
  • Fine-tuning: A custom-generated dataset of 52,676 binding measurements for variants of the Ab-14 antibody.
  • Validation: Seven newly designed libraries per antibody chain, each containing thousands of unique, AI-optimized sequences.

Key Insights and Ideas

  • Extrapolation is Possible: Standard ML models often struggle to predict outcomes for data far outside their training set. Here, the ensemble models successfully predicted high-affinity binders with up to 23 mutations away from the starting sequence, far beyond the 3-mutation limit of the training data.
  • The Power of Pre-training: Fine-tuning a model that already "understands" the grammar of proteins (via Pfam/OAS) performed significantly better than training a model from scratch.
  • Diversity vs. Potency Trade-off: The choice of sampling algorithm matters. Hill Climbing produced the most potent binders (the "peaks" of the landscape), while Gibbs Sampling produced the most diverse libraries. Diversity is vital because it gives drug developers more "backups" if the top candidate fails later due to toxicity or manufacturing issues.
  • Structure-Free Design: Unlike many computational tools, this method does not require a 3D structure of the target protein, making it applicable to "dark" targets where structural data is unavailable.

Key Results

  • 28.7-fold Improvement: The best antibody designed by the ML model bound to its target nearly 30 times more strongly than the best candidate produced by a traditional directed evolution baseline (PSSM).
  • 99% Success Rate: In their most successful library, 99% of the AI-designed sequences were better than the original starting antibody.
  • Sub-nanomolar Affinity: The team reached ultra-strong binding levels (pico-molar range) in just a single round of ML optimization.
  • Efficiency: The entire cycle—from initial data generation to validated high-affinity leads—takes only a few months, compared to the typical 12-month industry standard.

Limitations and Future Work

While impressive, the method still requires an initial "wet lab" step to generate ~7,000 measurements for the specific target to prime the model. It isn't yet a "zero-shot" predictor that works for any target out of the box.

The authors also noted that while they focused on affinity (binding strength), future iterations could include "multi-objective" optimization. This would allow the AI to simultaneously optimize for affinity, solubility, and low immunogenicity, essentially designing a "clinic-ready" drug in one go. Finally, integrating target-specific information (like the epitope sequence) could further reduce the need for initial experimental data.

Summary updated

Abstract

Therapeutic antibodies are an important and rapidly growing drug modality. However, the design and discovery of early-stage antibody therapeutics remain a time and cost-intensive endeavor. Here we present an end-to-end Bayesian, language model-based method for designing large and diverse libraries of high-affinity single-chain variable fragments (scFvs) that are then empirically measured. In a head-to-head comparison with a directed evolution approach, we show that the best scFv generated from our method represents a 28.7-fold improvement in binding over the best scFv from the directed evolution. Additionally, 99% of designed scFvs in our most successful library are improvements over the initial candidate scFv. By comparing a library's predicted success to actual measurements, we demonstrate our method's ability to explore tradeoffs between library success and diversity. Results of our work highlight the significant impact machine learning models can have on scFv development. We expect our method to be broadly applicable and provide value to other protein engineering tasks.
BibTeX
@inproceedings{li2023machine,
  author = {Lin Li and Esther Gupta and John Spaeth and Leslie Shing and Rafael Jaimes and Emily Engelhart and Randolph Lopez and Rajmonda S. Caceres and Tristan Bepler and Matthew Walsh},
  title = {{Machine learning optimization of candidate antibody yields highly diverse sub-nanomolar affinity antibody libraries}},
  booktitle = {Nature Communications},
  year = {2023},
  doi = {10.1038/s41467-023-39022-2},
}
Figure