Reinforcement Learning-Augmented ProteinMPNN Improve the Binding Affinity of TNFR1-Targeting Minibinders

Minibinders are compact protein molecules that hold great promise as therapeutic agents due to their target-binding specificity, stability, and potential for oral delivery. However, the primary objective ofthe widely used minibinder sequence design method, ProteinMPNN, is structural fidelity rather than the optimization of specific functional properties such as binding affinity. Consequently, the de novo design of high-affinity minibinders still largely relies on extensive wet-lab screening of numerous candidates. Directly designing high-affinity minibinders in silico thus remains a critical challenge for numerous therapeutic applications. Here, we present a computational framework that integrates a reinforcement learning (RL) framework with ProteinMPNN network to directly generate minibinder sequences with improved functional features. We demonstrate the power of this framework by optimizing the activities of the minibinders targeting Tumor Necrosis Factor Receptor 1 (TNFR1). Experimental validation showed that the optimized minibinders OPT1 and OPT7 exhibited 3-fold and 7-fold higher binding affinity, respectively, and 6-fold and 4-fold greater neutralizing activity in cells compared to the original minibinder S1B2 . Our work establishes the framework as a promising tool that augments AI-driven protein design for the de novo development of high-affinity therapeutic minibinders.