Folding models retooled as antibody screeners: frozen Protenix lifts Origin-1 macro AUROC from 0.669 to 0.757
Synopsis
Antibody screening usually splits specificity, affinity and developability across separate models, and structural information is often compressed into a single confidence score such as ipTM. FoldAbS freezes the open-source folding model Protenix as a shared encoder, trains only lightweight task heads, and supplies specificity supervision with pseudo-negatives built from antigen mismatch, CDRH3 mutation and CDRH3 replacement. With Protenix as backbone, FoldAbS-Px raises macro AUROC on Origin-1 from 0.6690 to 0.7570 and reports the best results on affinity and four developability properties.
Interpretation
The three antibody-screening objectives, whether a candidate binds its antigen specifically, how strongly it binds, and whether it is developable, can now come from one frozen folding-model representation plus three lightweight task heads. Earlier routes either passed predicted complex coordinates to a downstream model that had to relearn interaction features, or ranked candidates directly by native confidence or interface scores such as ipTM and ipSAE, which are not optimized for screening objectives. On the Origin-1 test set, with the same Protenix-v1 backbone, FoldAbS-Px raises macro AUROC from 0.6690 to 0.7570; on PPIFlow, built from de novo designed antibodies, it raises it from 0.5590 to 0.5920.
The specificity head learns to separate binders from non-binders directly, and the affinity head inherits the interaction representation from the trained specificity head before being converted to regression over continuous binding strength. Specificity training lacks experimentally validated negatives, so pseudo-negatives are generated from observed binding pairs by antigen mismatch, local CDRH3 mutation and whole CDRH3 replacement, with positives and negatives sampled at a 1:2 ratio. Specificity beats Protenix-v1 on six of eight antigen systems in Origin-1 and on five of seven targets in PPIFlow; affinity reports the best Spearman rho and MAE on all three test sets, AFF, SKE and HER2.
Developability prediction takes only the antibody side, regressing aggregation, expression, immunogenicity and thermostability from Protenix single, pair and structure outputs. Earlier developability models mostly relied on sequence descriptors or antibody language-model embeddings, and large-scale benchmarking found no single model leading consistently across datasets. Spearman rho is 0.406 for aggregation, 0.348 for expression, 0.367 for immunogenicity and 0.845 for thermostability, each above its strongest comparison baseline.
Perspective
The demonstrated scope is prediction and screening: specificity, affinity and developability are evaluated on public or curated benchmarks, FoldAbS-Px uses Protenix-v1 as a frozen backbone, and the code is public. The next step the authors propose is to integrate the learned screening signals into hallucination-based or generative antibody design pipelines.
The authors note that the work focuses on prediction and screening rather than closed-loop antibody design, and that specificity supervision partly relies on curated pseudo-negatives that may introduce false-negative noise; developability properties differ widely in sample size, and earlier large-scale benchmarking showed such properties are hard to predict consistently across datasets.