Abstract
Therapeutic antibody design is a complex multi-property optimization problem with substantial promise for improvement with the application of machine-learning methods. Towards realizing that promise, we introduce “Lab-in-the-loop,” a new approach that orchestrates state-of-the-art repertoire mining methods, generative machine learning models, multi-task property predictors, active learning ranking and selection, and in vitro experimentation in a semi-autonomous, iterative optimization loop. By automating the design of antibody variants, property prediction, ranking and selection of designs to assay in the lab, and ingestion of in vitro data, we enable an end-to-end approach to developing computationally-informed therapeutic antibody design pipelines. We apply lab-in-the-loop to eleven seed antibodies obtained via animal immunization with four clinically relevant antigen targets: EGFR, IL-6, HER2, and OSM. Over 1,800 unique antibody variants are tested throughout four rounds of iterative optimization identifying 3–100× better binding variants for all targets and 10/11 seeds, with the best binders exceeding 100 pM affinity, demonstrating a process by which end-to-end machine learning can be developed for therapeutic antibody development.
Introduction
Designing a therapeutic antibody is a "multi-property optimization" challenge. An effective antibody must not only bind tightly to its target (affinity) but also be manufacturable (expression), stable (developability), and highly specific to avoid off-target effects. Traditional discovery methods, such as animal immunization or display technologies, are often linear and struggle to balance these competing requirements simultaneously. While machine learning (ML) offers a way to navigate the vast sequence space of proteins, previous efforts have often struggled with "naturalness" or real-world translation. The "Lab-in-the-loop" (LitL) framework addresses these gaps by integrating generative AI with iterative experimental validation.
Methodology
The LitL framework functions as a semi-autonomous "flywheel" that cycles through computational design and experimental testing.
The pipeline consists of four primary components:
- Generative Ensemble: A mix of "unguided" models (e.g., Walk-Jump Sampling, SeqVDM) for broad exploration and "guided" models (e.g., LaMBO-2, DyAb) for targeted optimization.
- Discriminative Oracles: A multi-task framework called Cortex, built on the LBSTER protein language model, predicts expression yield, binding affinity, and non-specificity.
- Active Learning: The system employs the Noisy Expected Hypervolume Improvement (NEHVI) acquisition function to identify designs on the Pareto frontier—balancing trade-offs between multiple goals.
- High-Throughput Wet Lab: Designs are synthesized and tested in 4–6 week cycles, with the resulting data fed back to refine the models.
The system was pre-trained on massive public databases (UniRef50, PDB, OAS) and fine-tuned on internal Genentech datasets. It was evaluated against four clinical targets: EGFR, IL-6, HER2, and OSM.
Results
The LitL approach demonstrated significant improvements across all targets. Over four rounds of optimization, the team tested over 1,800 unique variants.
Key findings include:
- Affinity Gains: 10 out of 11 starting "seed" antibodies saw at least a 3× improvement, with several reaching 10–100× better binding. The best variants achieved sub-nanomolar affinities (~100 pM).
- Iterative Learning: Model accuracy improved significantly with each round of lab data, highlighting the value of high-quality, task-specific feedback loops.
- Structural Integrity: Crystal structures of the designs confirmed that the AI-suggested mutations preserved the original binding modes while optimizing the interface through "rational engineering" behaviors, such as rigidifying CDR loops.
Discussion
While LitL represents a major step toward autonomous drug discovery, it currently relies on high-quality "seeds" from traditional methods rather than full de novo design. Future work aims to incorporate more complex properties like in vivo pharmacokinetics and immunogenicity into the loop. By moving toward a fully integrated "discovery factory," the time and cost of developing life-saving antibody therapies could be drastically reduced.
Summary updated