Authors
Dell'uomo, D., Satz, A., Averso, B.
Abstract
Antibody discovery and engineering remain bottlenecked with current experimental screening approaches that are resource-intensive, time-consuming, and offer limited control over critical properties. We present Hyperbind2, a multi-shot learning platform that progressively improves through iterative lab-to-AI feedback cycles, distinguishing it from static zero-shot approaches. Our platform screens antibody candidates using only primary sequence input, inferring structural context internally so no experimental crystal or NMR structure is required at inference. Our approach embeds antibody and antigen sequences into a shared representation space where binding affinity is modeled as a geometric relationship. Hyperbind2 achieves remarkable efficiency, screening up to 100 million candidates within 48 hours, and demonstrates adaptability through direct learning head model training on minimal experimental data (10-20 binders/non-binders) converging to high-affinity antibodies within three digital-experimental cycles. Model training accuracy improved from 65% to 85% across three rounds. Experimental validation achieved 21% success rate (20/96 candidates <100 nM KD) by Round 3. In a pilot study targeting a challenging multi-pass membrane receptor, Hyperbind2 designed and identified 20 high-affinity scFv antibody candidates that were experimentally validated via SPR/BLI, achieving KD [≤] 100 nM through three multi-shot learning cycles of lab-to-AI feedback, with 3 candidates demonstrating sub-10 nM affinities. Using synthetic datasets derived from validated binders, we demonstrate the model's ability to identify high-affinity binders across diverse therapeutic formats including VHHs, scFvs, and full-length IgGs, with preliminary validation for CARs, BiTEs, and bispecifics. HyperBind2 is offered in two complementary offerings: (1) an open-source computational toolset for academic research, distributed with a lightweight contrastive model and documented workflows; and (2) a commercial platform that supplies lab-ready antibody sequences from proprietary antibody sequence libraries, with no computational work required by lab teams (via abtique.com). Hyperbind2 establishes a synergistic digital-experimental workflow where computational design and screening complements ongoing in vitro experimentation, and a continuous lab-to-AI feedback loop improves the model's performance, significantly accelerating the candidates identification, optimizing experimental resource allocation, and reducing experimental burden.
Preprint server:
bioRxiv
The authors list and abstract were imported from bioRxiv on 08 Nov 2025.
Advertisement
Stats
- Recommendations n/a n/a positive of 0 vote(s)
- Views 44
- Comments 0