Catapult

Catapult predicts gene-trait associations using Positive-Unlabeled Learning within a biased support vector machine and derives features from walks through a heterogeneous gene-trait network to enable cross-species prediction.

Key Features:

• Positive-Unlabeled Learning: Handles datasets with only labeled positive examples and unlabeled instances to address class imbalance.
• Biased Support Vector Machine (SVM): Implements a biased SVM framework to accommodate PU learning and emphasize separation between positives and unlabeled examples.
• Feature Derivation from Network Walks: Extracts features by analyzing walks through a heterogeneous gene-trait network to capture complex biological relationships.
• Functional Gene Associations: Incorporates functional gene associations derived from known interactions and functions within model organisms.
• Gene-Phenotype Associations in Model Organisms: Integrates gene-phenotype association data from model organisms to inform cross-species predictions.
• Performance Evaluation: Benchmarks predictive performance using OMIM phenotypes and drug-target interaction datasets and compares results to the Katz measure.

Scientific Applications:

• Gene-disease association prediction: Predicts associations between genes and diseases using integrated network-derived features and PU learning.
• Prioritization of gene-trait links: Ranks candidate gene-phenotype associations from large-scale gene-phenotype datasets.
• Cross-species comparative genomics: Transfers information across model organisms by leveraging functional and gene-phenotype associations for cross-species inference.

Methodology:

Uses supervised Positive-Unlabeled Learning within a biased support vector machine, derives features from walks on a heterogeneous gene-trait network, integrates functional gene associations and gene-phenotype associations from model organisms, and evaluates performance on OMIM phenotype and drug-target interaction datasets with comparison to the Katz measure.