Probabilistic Steering of Generative Priors for Lab-in-the-Loop Antibody Optimization
Organizations: BigHat Biosciences
Abstract
Therapeutic antibody development is often framed as comprising two discrete stages: initial discovery of novel binders followed by iterative optimization to refine these hits into therapeutic-grade molecules. Models and methods often reflect these stages, with generative structure models applied to discovery and probabilistic black-box optimization methods applied to refinement. Here we propose to unify these stages by smoothly transitioning a generative discovery model into iterative optimization, allowing the handoff between the structure-based prior and supervision from wet lab data to be guided by the posterior uncertainty of a probabilistic surrogate. We show that this yields improvements in optimization efficiency in the low-N regime, as well as improved diversity amongst the optimized molecules.
Figures & tables
| Target | Method | Designs | Best quality | Mean top-10 | parent (%) |
|---|---|---|---|---|---|
| PT | Best-of- | ||||
| FK | |||||
| OberBO | |||||
| LaMBO | |||||
| SARS-CoV-2 | Best-of- | ||||
| FK |
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
| Campaign | Binder format (length) | Antigen | Redesigned (#) | Target |
|---|---|---|---|---|
| VHH–PT | VHH (117 aa) | PT (internal) | 3 CDRs (31) | in |
| scFv–SARS-CoV-2 | scFv (246 aa) | SARS-CoV-2 peptide | CDRs (33) | in |
| Trastuzumab–HER2 | Fv (227 aa) | HER2 subdomain IV | CDR-H3 (10) |
| complex | binder_seq | target_seq | redesign (0-idx) |
|---|---|---|---|
| scfv_sars | EVQLVETGGGLVQPGGSLRLSCAASGFTLNSYGISWVRQAPGKGPEWVSVIYSDGRRTFYGDSVKGRFTISRDTSTNTVYLQMNSLRVEDTAVYYCAKGRAAGTFDSWGQGTLVTVSSGGGGSGGGGSGGGGSDVVMTQSPESLAVSLGERATISCKSSQSVLYESRNKNSVAWYQQKAGQPPKLLIYWASTRESGVPDRFSGSGSGTDFTLTISSLQAEDAAVYYCQQYHRLPLSFGGGTKVEIK | PDVDLGDISGINAS | 156–172; 188–194; 227–235 |
| trastuzumab_her2_h3mut | EVQLVESGGGLVQPGGSLRLSCAASGFNIKDTYIHWVRQAPGKGLEWVARIYPTNGYTRYADSVKGRFTISADTSKNTAYLQMNSLRAEDTAVYYCSRFEGDWFDAMYYWGQGTLVTVSSDIQMTQSPSSLSASVGDRVTITCRASQDVNTAVAWYQQKPGKAPKLLIYSASFLYSGVPSRFSGSRSGTDFTLTISSLQPEDFATYYCQQHYTTPPTFGQGTKVEIK | HPECQPQNGSVTCFGPEADQCVACAHYKDPPFCVARCPSIWKFPDEEGACQPCPIN | 98–107 |
| Edit distance | d1–3 | d4–9 | d10–14 | d15–19 | d20–26 |
|---|---|---|---|---|---|
| 2,475 | 1,765 | 528 | 633 | 500 | |
| MAE | 0.794 | 0.335 | 0.430 | 0.438 | 0.457 |
| Spearman | 0.738 | 0.669 | 0.724 | 0.682 | 0.687 |
| Campaign | Method | % pI | % GRAVY | OASis pct. (mean) |
|---|---|---|---|---|
| PT | BoN | |||
| FK | ||||
| OberBO | ||||
| LaMBO | ||||
| SARS-CoV-2 | BoN | |||
| FK |
Protein Binder Design
Protein binder design creates proteins, antibodies, and peptides that bind selected molecular targets with desired specificity and function.
See all 30 papers