Building Confidence in Deep Generative Protein Design

Abstract

Deep generative models show promise for de novo protein design, but their effectiveness within specific protein families remains underexplored. In this study, we evaluate two 3D rigid-body generative methods, score matching and flow matching, to generate monomeric protein backbones in SE(3) space. Our goal is to provide new insights and build confidence in the broader applicability of deep generative models for protein design. The optimal amino acid sequences were predicted from the generated backbones, followed by side-chain homology modeling. Results demonstrated high structural integrity, with conserved key residues aligning with known proteins. Structural phylogenetic analysis shows evolutionary links between the generated samples and their protein family members. Further molecular dynamics simulations and protein-ligand docking confirm the dynamic stability and functional potential of these samples, with ligand binding inducing conformational changes consistent with those in wild-type proteins.