SimBinder-IF: Structure-Aware Antibody Affinity Optimization via Efficient Preference Learning

Antibody therapeutic efficacy depends on high-affinity target engagement, yet laboratory affinity maturation is slow and costly. Most protein language models (PLMs) lack explicit training for high affinity, and current preference optimization methods introduce computational overhead without clear affinity improvements. Therefore, structure-aware and parameter-efficient approaches for antibody affinity optimization are urgently needed.

We propose SimBinder-IF, a structure-aware antibody optimization model trained by freezing the Evolutionary Scale Modeling inverse folding (ESM-IF) structure encoder and fine-tuning only its decoder via Simple Preference Optimization (SimPO) to prefer stronger binders. In generalization tests across seven held-out complexes (93 477 mutants), SimBinder-IF shows a 22% numerical increase in average Spearman correlation (0.26 to 0.32) compared with vanilla ESM-IF and outperforms ESM-IF on five of seven assays; however, the assay-level paired Wilcoxon test does not reach statistical significance (p = 0.188). For seed-averaged top-ranking enrichment, SimBinder-IF has the highest 20-fold improvement@20 mean (0.453), whereas ESM-IF has the highest 10-fold improvement@10 mean (0.520 vs. 0.487 for SimBinder-IF). In a case study redesigning antibody F045-092 to target the A/California/04/2009 pandemic H1N1 (pdmH1N1) strain-a target the original antibody fails to bind-SimBinder-IF generates variants with markedly lower predicted binding free energy than ESM-IF (mean ΔΔG: -68.56 vs. -49.77 kcal/mol). Notably, SimBinder-IF updates only 18% of the ESM-IF parameters. Regarding generic protein-protein affinity modelling, SimBinder-IF shows class-dependent preservation with trade-offs, improving performance on antibody-antigen (AB/AG) and protease-protein inhibitor (Pr/PI) complexes while yielding slightly lower aggregate mean correlations than ESM-IF and ProteinDPO.

The source code and model weights for SimBinder-IF are available at https://github.com/MSBMI-SAFE/ SimBinder-IF.

The supplementary file contains additional analyses of paratope prediction, statistical confidence, sequence identity, training efficiency, and the ESM-IF architecture.