Robust and Reliable de novo Protein Design: A Flow-Matching-Based Protein Generative Model Achieves Remarkably High Success Rates

Abstract

Generative models have achieved significant progress in the field of protein design, particularly in the generation of tertiary structures. However, they still face several challenges, such as balancing the designability and diversity of the generated structures, generating specified structures under highly constrained conditions, and the most challenging aspect, direct functional design such as motifs and binders. We present OriginFlow, an efficient protein generative model that combines Stochastic Differential Equation (SDE) models and flow models based on flow matching. The generated structures exhibit state-of-the-art (SOTA) levels of designability, diversity, and novelty. Moreover, this model demonstrates outstanding performance in functional design aspects such as motifs, binders, and symmetric motif scaffolding. We performed binder design for multiple targets, and all results showed excellent affinity performance in AF3 complex predictions and MD experiments. Wet-lab validation was conducted on PD-L1, RBD, and VEGF targets, achieving 90% expression, solubility, and affinity. These results are significantly superior to all current AI binder design algorithms.