GeoFlow-V2: A Unified Atomic Diffusion Model for Protein Structure Prediction and De Novo Design

Abstract

We introduce GeoFlow-V2, a unified atomic diffusion model that seamlessly integrates structure prediction and de novo protein design across multiple biological modalities, including proteins, nucleic acids (DNA/RNA), and small molecules. The models core innovation lies in its unified architecture that natively handles both structure prediction (when provided with complete input sequences) and generative design (for partially or fully masked inputs) through a shared atomic diffusion process. Through the integration of structure conditioning constraints and the sequence design module, GeoFlow-V2 operates as a fully bidirectional framework, accepting both sequence and structural inputs while generating corresponding outputs. GeoFlow-V2 can also accommodate diverse experimental constraints and prior knowledge, which boosts performance and enables precise control over the folding and design process. Benchmarking against state-of-the-art methods demonstrates GeoFlow-V2s strong performance in both structure prediction and de novo antibody design. We showcase GeoFlow-V2 s versatility through protein design cases conditioned on diverse target modalities. To maximize accessibility, we engineer key functionalities of GeoFlow-V2 and provide a user-friendly web server at prot.design for non-commercial research use.