DisoPatho: A Cross-View Feature-Adaptive Interaction Encoding Framework for Predicting Disease-Associated Variants in...

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Predicting variants in intrinsically disordered regions (IDRs) is vital for disease diagnosis but challenging due to IDRs' structural flexibility and sequence variability.
• Existing prediction tools struggle with the unique characteristics of IDRs, limiting their performance in these critical genomic regions.

Purpose of the Study:

• To develop DisoPatho, a novel deep learning framework for accurate prediction of disease-associated variants specifically within intrinsically disordered regions (IDRs).
• To enhance the biomedical interpretation of variants in IDRs by improving prediction accuracy and coverage.

Main Methods:

• Introduced DisoPatho, a deep learning framework with a mutation-centric architecture for variant prediction in IDRs.
• Employed a cross-view adaptive-feature interaction mechanism integrating IDR-specific energy and protein language model embeddings (xTrimoPGLM, Evolutionary Scale Modeling).
• Focused on capturing evolutionary constraints and physicochemical patterns without relying on explicit structures or multiple sequence alignments.

Main Results:

• DisoPatho demonstrated superior performance over existing methods across multiple IDR datasets, achieving high AUCs (0.899, 0.840) and ACCs (0.862, 0.860).
• Achieved a 50.2% relative improvement in MCC over AlphaMissense on a challenging independent test set, showcasing enhanced discriminative power.
• Showcased broader prediction coverage and effectiveness in IDR-specific scenarios, even with limited phylogenetic signals.

Conclusions:

• DisoPatho offers a robust and accurate solution for predicting disease-associated variants in intrinsically disordered regions.
• The framework's innovative approach overcomes limitations of traditional methods, advancing disease diagnosis and biomedical interpretation.
• DisoPatho provides a valuable, accessible tool for researchers studying variants in IDRs.