Predicting Experimental Success in De Novo Binder Design: A Meta-Analysis of 3,766 Experimentally Characterised Binders

Abstract

Designing high-affinity de novo protein binders has become increasingly tractable, yet in vitro prioritisation continues to depend on heuristics in the absence of systematic analysis. Here, we present a large-scale meta-analysis of 3,766 experimentally tested binders across 15 structurally diverse targets. Using a unified, high-throughput pipeline that re-predicts each binder-target complex with AF2 (initial guess and ColabFold), AF3 and Boltz-1, we extract over 200 structural, energetic and confidence features per design. We show that interface-focused metrics, most notably the AF3-derived interaction prediction Score from Aligned Errors (ipSAE) outperform commonly used scores such as ipAE and ipTM, with a significant 1.4-fold increase in average precision compared to ipAE. We further show that combining these metrics with orthogonal physicochemical interface descriptors, including Rosetta {Delta}G/{Delta}SASA and interface shape complementarity, improves predictive performance. While overall per-formance varies by target, simple linear models trained on a small number of AF3-derived features generalize well across datasets. We propose interpretable, target-agnostic filtering strategies, such as combining AF3 ipSAE_min rankings with structural filters, to improve precision in selecting binders for testing. Finally, we release the complete dataset establishing a community resource to benchmark and accelerate de novo binder discovery.