Reinforcement learning-based dynamic ensemble for missense variant effect prediction and tiered prioritization of VUS

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Accurate classification of missense variants is crucial but challenging.
• Existing computational models and ensemble methods have limitations in capturing complex predictive signal interactions.

Purpose of the Study:

• To develop a dynamic ensemble framework for improved missense variant pathogenicity prediction.
• To overcome limitations of static weighting in ensemble models by employing adaptive strategies.

Main Methods:

• GenixRL reformulates model fusion as a reinforcement learning optimization problem.
• A Q-learning agent dynamically weights outputs from predictors like BayesDel, ClinPred, and MetaRNN.
• Policy learning replaces static weighting for adaptive optimization.

Main Results:

• GenixRL achieved an AUROC of 0.9644 on an independent ClinVar dataset, outperforming 25 state-of-the-art predictors.
• Demonstrated superior performance in saturation genome editing assays for BRCA1 and BRCA2, ranking highest on 14/17 clinically significant genes.
• Enabled high-confidence prioritization of hundreds of thousands of uncertain/conflicting ClinVar variants using population data.

Conclusions:

• GenixRL represents an advancement in missense variant pathogenicity prediction.
• Provides an adaptive ensemble for prioritizing variants of uncertain significance for further validation.