ProteinSGM: Score-based generative modeling forde novoprotein design

Abstract

Score-based generative models are a novel class of generative models that have shown state-of-the-art sample quality in image synthesis, surpassing the performance of GANs in multiple tasks. Here we present ProteinSGM, a score-based generative model that produces realistic de novo proteins and can inpaint plausible backbones and functional sites into structures of predefined length. With unconditional generation, we show that score-based generative models can generate native-like protein structures, surpassing the performance of previously reported generative models. We apply conditional generation to de novo protein design by formulating it as an image inpainting problem, allowing precise and modular design of protein structure.

Broadly applicable and accurate protein design by integrating structure prediction networks and diffusion generative models

“…For all in silico benchmarks in this paper, we use the AF2 structure prediction network 21 for validation and define an in silico “success” as an RF diffusion output for which the AF2 structure predicted from a single sequence is (1) of high confidence (mean predicted aligned error, pAE, < 5), (2) globally within 2Å backbone-RMSD of the designed structure, and (3) within 1Å backbone-RMSD on any scaffolded functional-site. This definition of success is significantly more stringent than those described elsewhere (refs [ 5,8,16,25 ], Fig. S3A-B) but is a good predictor of experimental success 4,7,26 .…”

“…In RF diffusion , we input motifs as 3D coordinates (including sequence and sidechains) both during conditional training and inference, and RF diffusion builds scaffolds that hold the motif atomic coordinates in place. A number of deep learning methods have been developed recently to address this problem, including RF joint Inpainting 4 , constrained Hallucination 4 , and other DDPMs 5,8,25 . To rigorously evaluate the performance of these methods in comparison to RF diffusion across a broad set of design challenges, we established an in silico benchmark test comprising 25 motif-scaffolding design problems addressed in six recent publications encompassing several design methodologies 4,5,25,36–38 .…”

“…A number of deep learning methods have been developed recently to address this problem, including RF joint Inpainting 4 , constrained Hallucination 4 , and other DDPMs 5,8,25 . To rigorously evaluate the performance of these methods in comparison to RF diffusion across a broad set of design challenges, we established an in silico benchmark test comprising 25 motif-scaffolding design problems addressed in six recent publications encompassing several design methodologies 4,5,25,36–38 . The benchmark includes 25 challenges that span a broad range of motifs, including simple “inpainting” problems, viral epitopes, receptor traps, small molecule binding sites, binding interfaces and enzyme active sites.…”

From sequence to function through structure: Deep learning for protein design

“…GENESIS [86] implemented a VAE that takes secondary structure sketches and outputs contact maps with finer definitions of secondary structural elements. ProteinSGM [87] uses stochastic differential equations (SDE) to generate matrices that capture distance and torsional angles, which are then passed to Rosetta to produce 3D folded structures. FoldingDiff [88] merges this two-step approach by using a set of six internal angles, directly producing good quality structures without needing other methods like Rosetta for refinement.…”

An integrative approach to protein sequence design through multiobjective optimization

“…The field of computational protein design has achieved major breakthroughs in recent years [1] in terms of its ability to design proteins and protein assemblies with diverse folds and functions (see, e.g., [2][3][4][5][6][7][8][9]), which has already found application in the design of therapeutically relevant biomolecules such as vaccines [10] and antibodies [11,12]. Such breakthroughs are built upon improvements in atomistic modeling techniques, such as the Rosetta software suite [13], and recent advances in machine learning-based structure prediction models [14][15][16][17][18], sequence design (or inverse folding) models [19][20][21][22][23][24][25], protein language models [26][27][28][29][30][31][32][33][34], and denoising diffusion probabilistic models [35][36][37][38][39][40][41][42][43][44].…”