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 that mediate functions of these regions, such as short motifs, amino acid repeats and physicochemical properties. Here, we introduce a proteome-scale feature discovery approach for IDRs. Our approach, which we call “reverse homology”, exploits the principle that important functional features are conserved over evolution. We use this as a contrastive learning signal for deep learning: given a set of homologous IDRs, the neural network has to correctly choose a held-out homologue from another set of IDRs sampled randomly from the proteome. We pair reverse homology with a simple architecture and standard interpretation techniques, and show that the network learns conserved features of IDRs that can be interpreted as motifs, repeats, or bulk features like charge or amino acid propensities. We also show that our model can be used to produce visualizations of what residues and regions are most important to IDR function, generating hypotheses for uncharacterized IDRs. Our results suggest that feature discovery using unsupervised neural networks is a promising avenue to gain systematic insight into poorly understood protein sequences.
Introduction
Intrinsically Disordered Regions (IDRs) comprise approximately 40% of the residues in eukaryotic proteins. Unlike structured domains, these flexible, "spaghetti-like" segments do not adopt a stable 3D shape, yet they are critical for cell signaling, regulation, and phase separation. Because IDRs evolve rapidly, traditional sequence alignment tools like BLAST often fail to identify functional similarities. While IDRs are known to contain functional "features"—such as short linear motifs or distributed physical properties like net charge—our current understanding is limited by a bias toward previously studied examples. There is a pressing need for unbiased, proteome-scale methods to discover the molecular features that drive IDR function.
Methods
To address this, researchers developed Reverse Homology, a self-supervised contrastive learning framework that utilizes evolutionary conservation as a training signal. The core idea is that if a molecular feature is functionally important, it will be conserved across species even if the surrounding sequence drifts.
The model is trained on a proxy task: given a "query set" of homologous IDRs, it must identify a held-out homologue from a "target set" containing hundreds of random distractor IDRs. To succeed, the network must learn to ignore evolutionary noise and focus on conserved functional signatures. The architecture employs a lightweight 5-layer Convolutional Neural Network (CNN) with a dual-pooling strategy: max pooling to detect local motifs and average pooling to capture bulk properties like hydrophobicity or charge propensities.
Datasets
The model was trained and validated on large-scale datasets, including 5,306 homology sets from 15 yeast species and over 16,000 homology sets from vertebrates (including humans). Performance was benchmarked against 92 "GO Slim" biological process terms and a library of 66 literature-curated IDR features.
Results and Key Insights
The reverse homology approach successfully discovered a wide array of molecular features without requiring manual labels. The model learned 512 distinct features, 70 of which could be directly mapped to known motifs or physical properties, while others represent potentially novel functional patterns.
Key findings include:
- Dual-Nature of IDRs: The model confirmed that IDR function is mediated by both "islands" of conservation (motifs) and "oceans" of distributed bulk properties.
- Superior Performance: Representations learned via reverse homology outperformed expert-curated features in predicting biological processes, such as identifying proteins belonging to the Golgi membrane or translation machinery.
- High Precision: In-silico mutational scanning of the human protein p27 precisely highlighted known tyrosine modification sites and nuclear localization signals, demonstrating the model's ability to pinpoint functional residues in uncharacterized sequences.
- New Biological Hypotheses: The model identified S/T-rich repeats in the yeast protein Uth1 as critical for cell wall biogenesis, suggesting a primary role different from its previously assumed mitochondrial function.
Discussion and Future Work
While powerful, the current CNN-based approach is limited to local sequence windows and cannot capture long-range interactions within an IDR. Future iterations may utilize Transformers or Recurrent Neural Networks to handle variable sequence lengths and distal dependencies. Additionally, while the model generates high-confidence hypotheses, "wet lab" experimental validation remains essential to confirm the biological roles of newly discovered features. This work demonstrates that unsupervised deep learning, guided by evolutionary principles, is a transformative tool for decoding the "dark matter" of the proteome.
Summary updated