EvoDiff

EvoDiff is a diffusion framework for controllable protein generation that operates directly in sequence space, trained on evolutionary-scale protein databases. Its two variants, EvoDiff-Seq and EvoDiff-MSA, generate high-fidelity, diverse, and structurally plausible proteins by conditioning on individual sequences or on multiple sequence alignments respectively. Because it works in sequence space, EvoDiff can design intrinsically disordered regions and other functional elements that structure-based generators cannot reach, while still respecting biophysical constraints learned from evolution.

EvoDiff bridges sequence-based and structure-based protein design by exploiting the enormous information content of evolutionary sequence data. Generating proteins with controllable disorder and region-specific properties expands the design space well beyond globular folds, opening new avenues in cell signaling, regulation, and degradation-targeting therapeutics. Within Microsoft AI for Science, EvoDiff is a flagship example of how diffusion models trained on natural data can yield genuinely novel therapeutic modalities while remaining grounded in the rules of biology.