Enhancing missense variant classification in predicted intrinsically disordered regions

Classifying disease-causing missense variants in intrinsically disordered regions (IDRs) remains a significant challenge, with over 25% of known deleterious variants occurring in these regions. Existing in silico missense variant predictors that predict variant classification generally perform better in ordered regions of the protein, limiting their effectiveness. To address this, we developed a machine learning methodology that integrates global IDR conformation (gIDRc) features from ALBATROSS, phase separation (PS) features from BioPython, and 1024-dimensional protein embeddings from ProtTransBertBFD generated for both wild-type (WT) and mutant IDR sequences. IDR boundaries were defined using the AlphaFold-RSA predictions, which identifies disordered regions based on AlphaFold2 pLDDT scores and relative solvent accessibility. Using ClinVar variant classifications as ground truth, AlphaMissense, EVE, and ESM1b were the highest scoring unsupervised in silico missense predictors for IDR variants. Our baseline model, using only IDR-specific features achieved competitive performance on the hold-out test set with a PR-AUC of 0.817. Critically, when these IDR features were combined with these methods we saw significant overall improvement. The AlphaMissense-Enhanced model increased its PR-AUC from 0.807 to 0.919. Similarly, ESM1b-Enhanced improved PR-AUC from 0.679 to 0.845 and EVE increased from 0.591 to 0.910. These results demonstrate the effectiveness of our enhancements for classifying missense variants in IDRs and highlight its ability to complement existing in silico missense predictors.