Emyx Accelerates All-Atom Protein Generation with High Efficiency

Computational enzyme design requires generative models capable of producing proteins with high geometric accuracy and structural diversity to scaffold catalytic residues and ligands effectively. Existing all-atom protein generators often suffer from high computational costs and limited sample diversity, largely due to inheriting complex architectures from structure prediction tasks. This research introduces Emyx, a novel 140-million-parameter conditional flow matching model that rethinks the architecture for generative tasks. Emyx concentrates its capacity within standard transformer blocks, replacing heavy embedding stacks with lightweight conditional representations and sparse connectivity. This design choice significantly reduces the model's complexity and training overhead. Emyx also features an exact reparametrization of the flow matching interpolant into the EDM noise-level framework, bridging training efficiency with state-of-the-art sampling methods without requiring retraining. Despite being a much smaller model, Emyx outperforms leading generators like Proteina-Complexa and RFdiffusion3 across critical benchmarks, including global fold recovery, catalytic geometry accuracy, structural novelty, scaffold diversity, and geometric validity, all while training approximately four times faster.