Diffusion probabilistic modeling of protein backbones in 3D for the motif-scaffolding problem

Abstract

Construction of a scaffold structure that supports a desired motif, conferring protein function, shows promise for the design of vaccines and enzymes. But a general solution to this motif-scaffolding problem remains open. Current machine-learning techniques for scaffold design are either limited to unrealistically small scaffolds (up to length 20) or struggle to produce multiple diverse scaffolds. We propose to learn a distribution over diverse and longer protein backbone structures via an E(3)equivariant graph neural network. We develop SMCDiff to efficiently sample scaffolds from this distribution conditioned on a given motif; our algorithm is the first to theoretically guarantee conditional samples from a diffusion model in the largecompute limit. We evaluate our designed backbones by how well they align with AlphaFold2-predicted structures. We show that our method can (1) sample scaffolds up to 80 residues and (2) achieve structurally diverse scaffolds for a fixed motif.

ProteinSGM: Score-based generative modeling forde novoprotein design

“…We observe that 50.5% of the generations are designable given by scTM > 0.5, where a sequence (and predicted structure) with the same fold as the starting backbone can be generated. This is significantly higher than 11.8% reported in [16], which also faces helix chirality issues that further affects backbone designability. Chirality is a non-issue in our case since we use L-alanines in backbone minimization, and therefore all generations maintain proper handedness.…”

“…Though here we design the full-length protein from polyalanines, we note that for inpainting tasks where partial sequence and structure information are known, the Rosetta pipeline can easily be adapted to specifically design the inpainted region and retain the native domain(s). We also provide results on an alternative approach to sequence and full-atom structure generation from backbones by employing sequence design followed by AlphaFold2 prediction as in [16]. This allows one to circumvent the expensive FastDesign step and only require the relatively inexpensive…”

“…However, alternative approaches such as direct 3D coordinate generation with (E)3 equivariant networks [20] as in [16] extended to sequence and all-atom generation may be explored. One drawback of current generative models is the inability to model proteins larger than 256 residues, which limits practical applicability since the average eukaryotic protein length is > 400 residues.…”

“…As an alternative method to survey designability of generated backbones, we adapt the self-consistency TM (scTM) metric from [16]. After backbone selection from MinMover, instead of passing the backbone to FastDesign, we use the ESM-IF1 [3] sequence design model with temperature = 1 to sample 10 sequences for each backbone, and predict structures for each sequence with AlphaFold2 using the model_1_ptm parameters, 10 recycling iterations, and without any MSA input information.…”

“…Sampling is performed by using the estimated scores with Langevin dynamics, which maps the Gaussian prior to the data distribution, thereby generating data from noise. Despite its success in many domains, its application to protein design is still limited [15,16].…”