badmut

badmut predicts the deleteriousness of nonsynonymous single nucleotide variations (nsSNVs) by integrating multiple deleteriousness prediction scores using deep learning-based meta-estimation.

Key Features:

• Integration of Multiple Predictors: Combines outputs from predictors such as PolyPhen and SIFT via meta-estimators that model complex nonlinear relationships.
• Deep Learning Application: Employs supervised and unsupervised deep learning to enhance classification performance and extract hierarchical features from heterogeneous genetic datasets.
• Optimization Techniques: Uses a genetic algorithm for hyper-parameter optimization across generations with hardware acceleration to support computationally intensive searches.
• Overfitting Mitigation: Applies noise injection and dropout to reduce coadaptation of hidden units and improve generalization.
• Performance Evaluation: Benchmarks predictions against multiple popular predictors, including those recommended by the American College of Medical Genetics and Genomics (ACMG), reporting improved accuracy and coverage.

Scientific Applications:

• nsSNV deleteriousness inference: Supports inference of phenotypical effects for nonsynonymous SNVs used in genetic analyses.
• Theoretical and medical research: Aids studies in personalized medicine and disease risk assessment by improving classification accuracy.
• Low-data scenarios: Enhances handling of records with limited data to increase reliability of variant effect predictions.

Methodology:

Integrates PolyPhen and SIFT outputs via meta-estimators; applies supervised and unsupervised deep learning; optimizes hyper-parameters with a genetic algorithm using hardware acceleration; uses noise injection and dropout for regularization; and compares performance to ACMG-recommended and other predictors.