Explainable protein function annotation using local structure embeddings

Abstract

The rapid expansion of protein sequence and structure databases has resulted in a significant number of proteins with ambiguous or unknown function. While advances in machine learning techniques hold great potential to fill this annotation gap, current methods for function prediction are unable to associate global function reliably to the specific residues responsible for that function. We address this issue by introducing PARSE (Protein Annotation by Residue-Specific Enrichment), a knowledge-based method which combines pre-trained embeddings of local structural environments with traditional statistical techniques to identify enriched functions with residue-level explainability. For the task of predicting the catalytic function of enzymes, PARSE achieves comparable or superior global performance to state-of-the-art machine learning methods (F1 score > 85%) while simultaneously annotating the specific residues involved in each function with much greater precision. Since it does not require supervised training, our method can make one-shot predictions for very rare functions and is not limited to a particular type of functional label (e.g. Enzyme Commission numbers or Gene Ontology codes). Finally, we leverage the AlphaFold Structure Database to perform functional annotation at a proteome scale. By applying PARSE to the dark proteome—predicted structures which cannot be classified into known structural families—we predict several novel bacterial metalloproteases. Each of these proteins shares a strongly conserved catalytic site despite highly divergent sequences and global folds, illustrating the value of local structure representations for new function discovery.

Limitations of Current Machine-Learning Models in Predicting Enzymatic Functions for Uncharacterized Proteins

“…Recently, there has been an explosion of publications reporting the use of pre-trained Protein Language Models (PLM) to predict protein functions 29–33,78–80 ( Fig. 2 ).…”

“…2 legend), models have been developed to link protein sequences to vocabularies such as GO terms 28,81,82 or to EC numbers 83 . Some models predicting EC numbers have integrated structural data to identify active site signature residues, improving the separation of non-isofunctional paralogous subgroups when these harbor experimentally validated members 17,29,84 . While EC prediction models outperform BLAST 78,80,83 , they are less accurate than the curated annotation databases such as SwissProt or KEGG 85 and not accurate enough to reliably annotate enzymes for many practical purposes, particularly when trying to reach the level of substrate specificity (or the fourth EC number 86 ).…”

Sequence-Structure Analysis Unlocking the Potential Functional Application of the Local 3D Motifs of Plant-Derived Diterpene Synthases

“…The advent of AlphaFold2 and its continuous optimization have significantly improved this situation, allowing for effective analysis of enzyme product specificity through structural analysis. This approach has been extended to the large-scale annotation of enzyme functions [92,93], showcasing the advantages and potential of structural analysis in understanding PdiTPSs. Of course, our analysis has some disadvantages.…”

Opportunities and Challenges for Machine Learning-Assisted Enzyme Engineering

“…If these proteins could be accurately annotated, protein engineers would have access to a wealth of diverse candidates for engineering. While enzyme engineers have long been using multiple sequence alignments (MSAs) and homology to predict the functions of unannotated protein sequences, ML classification models extend these approaches and draw from more complete features describing protein sequences and structures to predict more specific functions, such as type of reactivity and k cat . ,− Focusing on known sequences without annotations, many of these methods aim to classify enzyme sequences based on their enzyme commission (EC) numbers, which is a hierarchical classification scheme that divides enzymes into general classes and then further subclasses, based on their catalytic activities (Figure A).…”