Emyx: Fast and efficient all-atom protein generation

A 140M-parameter conditional flow matching model named Emyx can generate all-atom protein structures while training in 682 GPU-hours, roughly four times less than the leading RFdiffusion3 model, according to a preprint posted to arXiv on June 12, 2026 [1][2]. The model, described in a paper titled "Emyx: Fast and efficient all-atom protein generation," concentrates its capacity within standard transformer blocks and replaces heavy embedding stacks with lightweight conditional representations and sparse connectivity [2]. The authors argue that the expensive architectures inherited from structure prediction are unnecessary for generative tasks that condition on sparse geometric constraints rather than rich co-evolutionary signals [2]. Emyx outperforms both Proteína-Complexa and RFdiffusion3 on the AME enzyme design benchmark under strict evaluation criteria that require global fold recovery, catalytic geometry accuracy, structural novelty, scaffold diversity, and geometric validity [2]. The team also derived an exact reparametrisation of the flow matching interpolant into the EDM noise-level framework, which bridges flow matching training efficiency with sampling methods designed for diffusion models without requiring retraining [2]. The preprint appeared on arXiv, an open-access repository that hosts electronic preprints across mathematics, physics, computer science, and related fields [6]. As of November 2024, the repository receives approximately 24,000 submissions per month and has surpassed two million total articles [6]. The paper was submitted under the Machine Learning category and is available in both PDF and experimental HTML formats [1]. Computational enzyme design demands models that can scaffold catalytic residues and ligands with both geometric precision and structural variety [2]. Current all-atom generators have been limited by high training costs and constrained sample diversity, problems the Emyx architecture directly addresses through its reduced parameter count and efficient training regime [2]. The model's 682 GPU-hour training requirement represents a significant reduction from the computational budget needed by comparable systems [2].