mBER: Controllable de novo antibody design with million-scale experimental screening

Recent machine learning approaches have achieved high success rates in designing protein binders that demonstrate in vitro binding to their targets. While open models for unconstrained “minibinder” design have shown great experimental promise, methods for designing binders in specific formats, such as antibodies, have lagged behind in experimental success rates. In this work, we present mBER, an open-source protein binder design system capable of designing antibody-format binders with state-of-the-art experimental success rates. mBER builds on the ColabDesign framework, achieving successful antibody design primarily through the inclusion of informative sequence and structure conditioning information. Using mBER, we designed two libraries comprising over 1 million VHH binders against 436 diverse targets. We experimentally screened the two libraries against 145 of these targets, resulting in a dataset of over 100 million binding interactions. We achieved specific and significant design success against 45% of targets. In a filtered set of designs, we detect binding rates to specific epitopes as high as 38%. Through mBER, we demonstrate that format-specific binder design is possible with no additional training of underlying folding and language models. This work represents the largest reported de novo protein design and validation campaign, and one of the first open-source methods to demonstrate double-digit percentage experimental success rates for antibody binder design.