PIDiff is a physicochemically enhanced multimodal diffusion framework for protein‑protein complex design. It jointly generates target‑conditioned sequences and structures by coupling diffusion in discrete sequence space with diffusion on the continuous SE(3) structural manifold. The core module PIFormer interleaves geometry‑aware invariant point attention with global sequence self‑attention, capturing the interplay between spatial and sequential information. We curated PIBench, a dataset of diverse target scaffolds sourced from the PDB, and pretrained PIDiff on it. The model is then fine‑tuned for two therapeutically relevant tasks—minibinder design and antibody design. Experiments show that PIDiff consistently outperforms baseline methods in sequence design success rate, structural binding energy, and sequence‑structure co‑designability.
Review