Abstract

Designing proteins for real-world applications requires the simultaneous satisfaction of multiple physicochemical properties. Structure-based de novo protein design has become the prominent design paradigm, successfully creating numerous proteins. Property optimization is commonly introduced during the sequence generation stage of protein design, i.e., inverse folding. Existing methods primarily rely on fine-tuning inverse folding models to design sequences with desired characteristics. However, multi-property optimization through fine-tuning demands datasets annotated with multiple properties-resources that remain extremely limited. Consequently, structure-based protein design has not yet achieved joint optimization of multiple properties. Here, we present Discriminator-Guided Inverse Folding (DGIF), a framework that guides the inverse folding model by adjusting its internal history states through an auxiliary discriminator module. The discriminator integrates multiple property predictors, each trained independently on a single-property dataset, thereby enabling multi-property optimization in the absence of datasets annotated with multiple properties. In addition to substantial improvements in key traits like thermostability and solubility, DGIF can generate protein sequences optimized for both properties simultaneously, with the designed proteins shifting markedly toward the Pareto front that represents optimal trade-offs. Experimental results validate the effectiveness of DGIF for multi-property protein design.