EquiDesign: a versatile computational framework for de novo protein sequence design

Abstract

Designing amino acid sequences that uniquely fold into target three-dimensional conformations defines the protein inverse folding problem, a formidable bottleneck in de novo protein design. Current computational paradigms, including deep learning-based approaches, often struggle to reconcile high sequence accuracy with structural stability, computational efficiency, and generalizability across diverse protein families. Here, we present EquiDesign, a deep learning framework that surmounts these constraints via a novel multistep Equiformer-enhanced architecture. By disentangling geometric features into invariant scalars and equivariant vectors, EquiDesign preserves the intrinsic directional topology of the protein backbone during message passing. This is achieved through spatial K-nearest neighbor graph construction and geometric attention mechanisms that hierarchically fuse multiscale contextual information, yielding a physically grounded representation of protein conformations. Extensive benchmarking demonstrates that EquiDesign achieves superior sequence recovery on standard and challenging test sets compared with state-of-the-art methods. Critically, the generated sequences exhibit exceptional structural fidelity, folding stability, and sequence diversity, as validated by comprehensive controls against training-set memorization, compositional bias, and physical energy landscapes, alongside targeted diversity sampling. These results establish EquiDesign as a robust, scalable, and generalizable inverse folding solution, driven by SE(3)-equivariant encoding and multiscale geometric attention, thus propelling de novo protein design toward unprecedented structural precision.