Discovering molecular features of intrinsically disordered regions by using evolution for contrastive learning

Abstract

A major challenge to the characterization of intrinsically disordered regions (IDRs), which are widespread in the proteome, but relatively poorly understood, is the identification of molecular features, such as short motifs, amino acid repeats and physicochemical properties that mediate the functions of these regions. Here, we introduce a proteome-scale feature discovery method for IDRs. Our method, which we call "reverse homology", exploits the principle that important functional features are conserved over evolution as a contrastive learning signal for deep learning: given a set of homologous IDRs, the neural network has to correctly choose a randomly held-out homologue from another set of IDRs sampled randomly from the proteome. We pair reverse homology with a simple architecture and interpretation techniques, and show that the network learns conserved features of IDRs that can be interpreted as motifs, repeats, and other features. We also show that our model can be used to produce specific predictions of what residues and regions are most important to the function, providing a computational strategy for designing mutagenesis experiments in uncharacterized IDRs. Our results suggest that feature discovery using neural networks is a promising avenue to gain systematic insight into poorly understood protein sequences.

Sequence- and chemical specificity define the functional landscape of intrinsically disordered regions

“…With this in mind, a general model for IDR evolution has emerged in which substantial sequence variation can be tolerated assuming SLiMs are conserved or the amino acid composition and patterning (i.e. bulk sequence properties) are maintained (5,(23)(24)(25).…”

The return of the rings: Evolutionary convergence of aromatic residues in the intrinsically disordered regions of RNA‐binding proteins for liquid–liquid phase separation

“…47 Moreover, the "spacers" between the stickers are also essential factors in driving LLPS, 47,48 and current sequence alignment algorithms for IDRs cannot easily pick the importance of spacers as well. Although a recent work using machine learning approaches to collect the "features" of IDRs might be an alternative approach to find their traits, 49 it is still limited in capturing the distal interaction, such as the prevailing aromatic residues.…”

Discovering molecular features of intrinsically disordered regions by using evolution for contrastive learning

“…In previous work, we explained the theoretical principles behind using evolutionary homology as a basis for contrastive learning [ 57 ]: our method is expected to learn conserved features of protein sequences, which we argue are likely important for the conserved function of rapidly-diverging IDRs ( S1 Methods ).…”

“…From a technical perspective, reverse homology employs a self-supervised approach, as many emerging representation learning approaches for protein sequences do [ 50 – 54 ]. Unlike these methods, which are mostly based on methods adapted from natural language processing, we proposed a novel proxy task that purposes principles of evolutionary proteomics as a learning signal instead [ 57 ]. Another distinction in our study is that previous approaches primarily focus on representation learning, with the aim of optimizing the performance of the representation on downstream regression tasks reflecting protein design or classification problems.…”

“…This task, which we call reverse homology, requires our model to learn conserved features of IDRs, in order to distinguish them from non-homologous background sequences. Our method is a contrastive learning method, a strategy that is now frequently employed in self-supervised learning [ 54 , 57 – 60 ]. We show that reverse homology can be applied on a proteome-wide scale to learn a large set of diverse features.…”