Structure-based protein function prediction using graph convolutional networks.

Gligorijević V, Renfrew PD, Kosciolek T, Leman JK, Berenberg D, Vatanen T, Chandler C, Taylor BC, Fisk IM, Vlamakis H, Xavier RJ, Knight R, Cho K, Bonneau R. · 2021

Read Structure-based protein function prediction using graph convolutional networks. by Gligorijević V, Renfrew PD, Kosciolek T, Leman JK, Berenberg D, Vatanen T, Chandler C, Taylor BC, Fisk IM, Vlamakis H, Xavier RJ, Knight R, Cho K, Bonneau R. (2021) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc8155034.txt.

Introduction

Gligorijević and co-authors present DeepFRI, a graph method that predicts protein function from structure. Protein databases grow fast, and known functions lag far behind. Their method pairs sequence features from a protein language model with structure graphs.

The authors report that DeepFRI beats leading methods and sequence-only networks. Adding homology models to the training set widened the functions it can predict. Performance held up well when models replaced measured structures. Site-level maps let users see which residues drive a prediction. The tool scales to the size of current sequence collections.

Catalog record

CREATORGligorijević V, Renfrew PD, Kosciolek T, Leman JK, Berenberg D, Vatanen T, Chandler C, Taylor BC, Fisk IM, Vlamakis H, Xavier RJ, Knight R, Cho K, Bonneau R.
YEAR2021
COLLECTIONPAPERS
FORMATTXT
FILE SIZE90 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC8155034

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Structure-based protein function prediction using graph convolutional networks By Vladimir Gligorijević, P. Douglas Renfrew, Tomasz Kosciolek, Julia Koehler Leman, Daniel Berenberg, Tommi Vatanen, Chris Chandler, Bryn C. Taylor, Ian M. Fisk, Hera Vlamakis, Ramnik J. Xavier, Rob Knight Published in: Nature Communications DOI: 10.1038/s41467-021-23303-9 ABSTRACT The rapid increase in the number of proteins in sequence databases and the diversity of their functions challenge computational approaches for automated function prediction. Here, we introduce DeepFRI, a Graph Convolutional Network for predicting protein functions by leveraging sequence features extracted from a protein language model and protein structures. It outperforms current leading methods and sequence-based Convolutional Neural Networks and scales to the size of current sequence repositories. Augmenting the training set of experimental structures with homology models allows us to significantly expand the number of predictable functions. DeepFRI has significant de-noising capability, with only a minor drop in performance when experimental structures are replaced by protein models. Class activation mapping allows function predictions at an unprecedented resolution, allowing site-specific annotations at the residue-level in an automated manner. We show the utility and high performance of our method by annotating structures from the PDB and SWISS-MODEL, making several new confident function predictions. DeepFRI is available as a webserver at https://beta.deepfri.flatironinstitute.org/. INTRODUCTION Proteins fold into 3-dimensional structures to carry out a wide variety of functions within the cell1. Even though many functional regions of proteins are disordered, the majority of domains fold into specific and ordered three-dimensional conformations2–6. In turn, the structural features of proteins determine a wide range of functions: from binding specificity and conferring mechanical stability, to catalysis of biochemical reactions, transport, and signal transduction. There are several widely used classification schemes that organize these myriad protein functions including the Gene Ontology (GO) Consortium7, Enzyme Commission (EC) numbers8, Kyoto Encyclopedia of Genes and Genomes (KEGG)9, and others. For example, GO classifies proteins into hierarchically related functional classes organized into three different ontologies: Molecular Function (MF), Biological Process (BP), and Cellular Component (CC), to describe different aspects of protein functions. The advent of efficient low-cost sequencing technologies and advances in computational methods (e.g., gene prediction) have resulted in a massive growth in the number of sequences available in key protein sequence databases like the UniProt Knowledgebase (UniProtKB)10. UniProt currently contains over 100 million sequences, only about 0.5% of which are manually annotated (UniProtKB/Swiss-Prot). Due to considerations of scale, design, and costs of experiments to verify a function, it is safe to posit that most proteins with unknown function (i.e., hypothetical proteins) are unlikely to be experimentally characterized. Understanding the functional roles and studying the mechanisms of newly discovered proteins is one of the most important biological problems in the post-genomic era. In parallel to the growth of sequence data, advances in experimental and computational techniques in structural biology has made the three-dimensional structures of many proteins available11–18. The Protein Data Bank (PDB)19, a repository of three-dimensional structures of proteins, nucleic acids, and complex assemblies, has experienced significant recent growth, reaching almost 170,000 entries. Large databases of comparative models such as SWISS-MODEL also provide valuable resources for studying structure–function relationships13,20. To address the sequence-function gap many computational methods have been developed with the goal to automatically predict protein function. Further, related work is directed at predicting function in a site- or domain-specific manner21–24. Traditional machine learning classifiers, such as support vector machines, random forests, and logistic regression have been used extensively for protein function prediction. They have established that integrative prediction schemes outperform homology-based function transfer25,26 and that integration of multiple gene- and protein-network features typically outperform sequence-based features even though network features are often incomplete or unavailable. Systematic blind prediction challenges, such as the Critical Assessment of Functional Annotation (CAFA127, CAFA228, and CAFA329) and MouseFunc30, are critical in the development of these methods and have shown that integrative machine learning and statistical methods outperform traditional sequence alignment-based methods (e.g., BLAST)26. However, the top-performing CAFA methods typically rely strongly on manually-engineered features constructed from either text, sequence, biological networks, or protein structure31. In most cases, for newly sequenced proteins, or proteins of poorly studied organisms these features are difficult to obtain because of limited information (e.g., no text features or biological network available). Here, we focus on methods that take sequence and sequence-based features (such as predicted structure) as inputs and do not focus on, or compare to, the many methods that rely on protein networks like GeneMANIA32, Mashup33, DeepNF34, and other integrative network prediction methods. As a result, we present a method applicable to hundreds of thousands of sequences of proteins from unknown organisms, lacking the required network data. In the last decade, deep learning has led to unprecedented improvements in performance of methods tackling a broad spectrum of problems, ranging from learning protein sequence embeddings for contact map prediction35 to predicting protein structure36,37 and function38. In particular, convolutional neural networks (CNN)39, the state-of-the-art in computer vision, have shown tremendous success in addressing problems in computational biology. They have enabled task-specific feature extraction directly from protein sequence (or the corresponding 3D structure), overcoming the limitations of standard feature-based machine learning (ML) methods. The majority of sequence-based protein function prediction methods use 1D CNNs, or variations thereof, that search for recurring spatial patterns within a given sequence and converts them hierarchically into complex features using multiple convolutional layers. Recent work has employed 3D CNNs to extract features from protein structural data40,41. Although these works demonstrate the utility of structural features, storing and processing explicit 3D representations of protein structure at high resolution is not memory efficient, since most of the 3D space is unoccupied by protein structure. In contrast, geometric deep learning methods42,43, and more specifically graph convolutional networks (GCNs)44, overcome these limitations by generalizing convolutional operations on more efficient graph-like molecular representations. GCNs have shown tremendous success in various problems ranging from learning features for quantitative structure-activity relationship (QSAR) models45, to predicting biochemical activity of drugs46, to predicting interfaces between pairs of proteins47. Here, we describe a method based on GCNs for functionally annotating proteins and detecting functional regions in proteins, termed Deep Functional Residue Identification (DeepFRI), that outperforms current methods and scales to the size of current repositories of sequence information. Our model has a two-stage architecture that takes as input a protein structure and a sequence representation from a pre-trained, task-agnostic language model, represented as graphs derived from amino acid interactions in the 3D structure. The model outputs probabilities for each function (see Fig. 1) and identifies residues important for function prediction by using the gradient-weighted Class Activation Map (grad-CAM)48 approach, that we adapted for post-training analysis of GCNs. We provide several examples where we automatically and correctly identify functional sites for various functions where binding and catalytic sites are known.Fig. 1Schematic method overview.a LSTM language model, pre-trained on ~10 million Pfam protein sequences, used for extracting residue-level features of PDB sequence. b Our GCN with three graph convolutional layers for learning complex structure–function relationships. RESULTS DeepFRI combines protein structure and pre-trained sequence embeddings in a GCN In the past few years, it has been shown that features extracted from pre-trained, task-agnostic, language models can significantly increase classification performance in many natural language processing49 and biological problems35. Here, we use a similar approach

Read the complete text