AmazonForest

AmazonForest predicts the pathogenic potential of genetic variants using machine learning to support interpretation of variants with uncertain or conflicting clinical significance.

Key Features:

• Data Source: Uses a curated ClinVar dataset comprising approximately 774,000 variant entries as the primary training and reference resource.
• Machine Learning Algorithms: Employs classical algorithms including Naive Bayes, Random Forest, and Support Vector Machine (SVM) to construct a meta-prediction model.
• Functional Annotation: Annotates variants with eight functional impact predictors via SnpEff/SnpSift v4.3 to enrich feature information for modeling.
• Encoding and Classifiers: Applies one-hot encoding combined with tree-based classifiers such as Random Forest to improve predictive performance.
• Model Evaluation: Evaluates models using 10-fold cross-validation and metrics including accuracy, F1-Score, Receiver Operating Characteristic (ROC), and Area Under the Curve (AUC), with top models achieving AUC ≥ 0.93.
• High-Probability Variant Query: Identifies and queries a subset of 5,000 variants predicted with high pathogenic probability (RFprob ≥ 0.9).

Scientific Applications:

• Clinical Variant Interpretation: Provides pathogenicity predictions to aid interpretation of variants of uncertain or conflicting significance.
• Genetic Counseling and Diagnosis: Supports decision-making in genetic counseling, diagnosis, and personalized medicine by prioritizing likely pathogenic variants.
• Research on Complex Phenotypes: Facilitates research into complex phenotypes and underlying genetic mechanisms by enabling large-scale variant prioritization and analysis.

Methodology:

Processes a ClinVar dataset (~774,000 entries); annotates variants with eight functional impact predictors using SnpEff/SnpSift v4.3; trains Naive Bayes, Random Forest, and SVM models combined into a meta-prediction framework; applies one-hot encoding and tree-based classifiers (e.g., Random Forest) achieving AUC ≥ 0.93; evaluates performance with 10-fold cross-validation using accuracy, F1-Score, ROC, and AUC; and queries a subset of 5,000 variants with RFprob ≥ 0.9.