[P450Diff2: a diffusion model-based method for generating P450 enzyme sequences]

Cytochrome P450 enzymes constitute the largest superfamily of oxidoreductases in nature, playing pivotal roles in drug metabolism, plant secondary metabolism, and biotransformation of environmental pollutants. To generate artificial P450 enzyme sequences with high fidelity and diversity, we propose P450Diff2, a novel diffusion model-based approach for generating P450 enzyme sequences. Built upon the EvoDiff-Seq framework comprising 640 million parameters, P450Diff2 was trained on a comprehensive dataset of 1 041 254 non-redundant P450 protein sequences collected from NCBI, GMind annotations, RNA-Seq assemblies, and metagenomic databases. Evaluation of the generated sequences revealed that P450Diff2 outperformed the previously proposed P450Diffusion model across multiple metrics, including amino acid composition distribution, sequence feature space coverage, sequence similarity profiles, and structural plausibility. Notably, the generated sequences achieved an average pLDDT score of 72.29. Experimental results further demonstrate that 60% of the generated sequences can correctly fold into biologically active P450 enzymes, indicating that the proposed method not only effectively preserves the structural features of natural sequences but also exhibits strong potential for functional sequence generation. By integrating large-scale sequence generation and screening workflows, this approach holds promise for the rapid design of efficient novel enzymes while significantly reducing the time and cost of experimental validation, offering a valuable and scalable paradigm for de novo enzyme engineering.