MSCORE: Multi-Target Protein Design through Compositional Score Combination

Multi-target protein design represents a fundamental challenge in therapeutic development, where proteins must simultaneously satisfy multiple binding requirements such as target selectivity and pharmacokinetic stability. Current computational approaches include sequential optimization (optimizing targets independently then seeking compromises), multi-objective evolutionary algorithms that treat targets as separate objectives, and structure-based methods using physics-based energy functions. However, these approaches fail to encode the compositional relationships between binding objectives and often require extensive multi-target training data that is scarce compared to abundant single-target datasets. We present MSCORE (Multi-target SCOre REcombination), a framework that decomposes multi-target protein design into target-specific sequence optimization followed by score composition functions. Our approach leverages pre-trained protein language models—trained on abundant single-target data—to generate position-wise binding predictions for each target independently, then combines these predictions using mathematical operators that encode desired inter-target relationships. For cooperative binding scenarios requiring simultaneous enhancement of multiple interactions, we employ multiplicative score combination. For competitive selectivity scenarios requiring discrimination between similar targets, we use ratio-based combination. Both operators incorporate position-specific influence parameters that enable fine-grained control over regional target allocation. We demonstrate MSCORE's effectiveness on antibody design tasks addressing two critical therapeutic challenges. In cooperative binding experiments, MSCORE achieves substantial success rates in simultaneously improving binding to both main and secondary targets. For the challenging competitive selectivity task of distinguishing between single-residue variants, MSCORE consistently achieves directional selectivity while preserving high overall binding affinity, significantly outperforming sequential optimization and naive unmasking baselines. This work establishes score composition as a principled paradigm for leveraging single-target models for multi-target molecular design with broad applications in therapeutic development.