PXDesign: Fast, Modular, and Accurate De Novo Design of Protein Binders

PXDesign is an end-to-end protein binder design pipeline validated on six diverse protein targets, achieving nanomolar binder hit rates of 20–73% on five of them, representing a substantial improvement over prior methods such as AlphaProteo [30]. The high experimental success rate of PXDesign stems from a systematic development process addressing the two essential stages of structure-based design: generation and filtering. In this work, we develop and benchmark both diffusion-based generative models and hallucination-based optimization approaches, identifying complementary strengths across tasks. Beyond generation, we systematically analyze confidence-based filtering and ranking strategies from multiple structure predictors, comparing their accuracy, efficiency, and complementarity on datasets spanning de novo binders and mutagenesis. These analyses inform the construction of a robust, high-hit-rate pipeline for protein binder design. To promote reproducibility and broader adoption, we release our filtering and benchmarking frame-work ( https://github.com/bytedance/PXDesignBench ) and provide public access to PXDesign via a webserver ( https://protenix-server.com ). Abstract Figure