A Deep SE(3)-Equivariant Model for Learning Inverse Protein Folding

Abstract

In this work, we establish a framework to tackle the inverse protein design problem; the task of predicting a protein's primary sequence given its backbone conformation. To this end, we develop a generative SE(3)-equivariant model which significantly improves upon existing autoregressive methods. Conditioned on backbone structure, and trained with our novel partial masking scheme and side-chain conformation loss, we achieve state-of-the-art native sequence recovery on structurally independent CASP13, CASP14, CATH4.2, and TS50 test sets. On top of accurately recovering native sequences, we demonstrate that our model captures functional aspects of the underlying protein by accurately predicting the effects of point mutations through testing on Deep Mutational Scanning datasets. We further verify the efficacy of our approach by comparing with recently proposed inverse protein folding methods and by rigorous ablation studies.

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

“…Following the success of deep language models, Ferruz et al [11] developed protein sequence models to generate new proteins, but these models do not allow specification of structural motifs. Another class of methods, referred to as fixed backbone sequence design [12,20,40,27,18], attempts to solve the problem of identifying a sequence that folds into any given designable backbone structure. In the present work, we utilize a particular sequence design method, ProteinMPNN [8], but in principle any other fixed-backbone sequence design method could be used in its place.…”

End-to-End deep structure generative model for protein design

“…Fixed backbone sequence design We use the pre-trained fixed backbone design model of (McPartlon et al, 2022) to generate sequences from backbone structures. For each test target, we generate 500 conformational decoys and sequences are deranged for each of the decoys independently.…”

“…This approach comes with several drawbacks, such as the exponential search space and limited capability to model higher order interactions. Recently, many deep learning based models have explored the use of rotation equivariant frameworks (McPartlon et al, 2022; Jing et al, 2020; Hsu et al, 2022) and invaraint 3-dimensional voxel-based approaches (Qi & Zhang, 2020; Anand et al, 2022). Although our model does not directly design the sequence alongside the conformation decoys, we show that sequences designed from the resulting templates can be more reliable than those designed from a single backbone structure.…”

“…These advancements have also played a part in aspects of the computational protein design problem. For fixed backbone sequence design, a series of deep learning methods have emerged to improve conventional energy based approaches (Alford et al, 2017; Khatib et al, 2011) by directly incorporating structural information using SE(3)-equivaraint frameworks (McPartlon et al, 2022; Jing et al, 2020; Hsu et al, 2022). An array of recent works have studied the use of generative models for structure generation (Lin et al, 2021; Anand et al, 2019; Anand & Huang, 2018), however, these methods often generate topological constraints and rely on downstream tools for 3-dimensional structure determination.…”

“…In recent years, progress in machine learning and deep learning research has contriburted to significant advances for protein modeling such as mutation effect estimation (Frazer et al, 2021;Meier et al, 2021;Riesselman et al, 2018), protein function prediction (Lai & Xu, 2021;Gligorijević et al, 2021), and structure prediction (Jumper et al, 2021;Baek et al, 2021;Wang et al, 2017).These advancements have also played a part in aspects of the computational protein design problem. For fixed backbone sequence design, a series of deep learning methods have emerged to improve conventional energy based approaches (Alford et al, 2017;Khatib et al, 2011) by directly incorporating structural information using SE(3)-equivaraint frameworks (McPartlon et al, 2022;Jing et al, 2020;Hsu et al, 2022). An array of recent works have studied the use of generative models for structure generation Anand et al, 2019;Anand & Huang, 2018), however, these methods often generate topological constraints and rely on downstream tools for 3-dimensional structure determination.…”