Deep learning extends de novo protein modelling coverage of genomes using iteratively predicted structural constraints

Joe G Greener,Shaun M Kandathil,David T Jones

doi:10.1038/s41467-019-11994-0

Joe G Greener, Shaun M Kandathil + Show 1 more

Open Access

https://doi.org/10.1038/s41467-019-11994-0

Copy DOI

Abstract

The inapplicability of amino acid covariation methods to small protein families has limited their use for structural annotation of whole genomes. Recently, deep learning has shown promise in allowing accurate residue-residue contact prediction even for shallow sequence alignments. Here we introduce DMPfold, which uses deep learning to predict inter-atomic distance bounds, the main chain hydrogen bond network, and torsion angles, which it uses to build models in an iterative fashion. DMPfold produces more accurate models than two popular methods for a test set of CASP12 domains, and works just as well for transmembrane proteins. Applied to all Pfam domains without known structures, confident models for 25% of these so-called dark families were produced in under a week on a small 200 core cluster. DMPfold provides models for 16% of human proteome UniProt entries without structures, generates accurate models with fewer than 100 sequences in some cases, and is freely available.

Highlights

The inapplicability of amino acid covariation methods to small protein families has limited their use for structural annotation of whole genomes
At the recent CASP13 experiment, methods using deep learning approaches to predict distances to use in model building appeared for the first time
We show that DMPfold produces more accurate models than CONFOLD2 and Rosetta for the CASP12 free modelling (FM) domains, with good performance when asked to generate just a single best model

Summary

Introduction

The inapplicability of amino acid covariation methods to small protein families has limited their use for structural annotation of whole genomes. CONFOLD2 integrates secondary structure and predicted contacts in a two-stage modelling approach, where unsatisfied contacts are filtered out after initial model generation These methods are computationally cheaper than fragment assembly but produce poor models without a large number of sufficiently accurate predicted contacts. At the recent CASP13 experiment, methods using deep learning approaches to predict distances to use in model building appeared for the first time. Two recently proposed deep learning methods attempt to generate model coordinates directly from sequence data by end-to-end training[15,16] Whilst promising, these end-to-end trained methods have not yet shown anything close to state-of-the-art performance in protein modelling, probably because they do not make use of the recent advances in sequence covariation analysis

Methods

Results

Conclusion