mantis-ml

mantis-ml prioritizes protein-coding genes associated with human disease using a stochastic semi-supervised machine-learning framework to rank disease-associated genes across the exome.

Key Features:

• Multi-step stochastic semi-supervised learning: An automated framework that iteratively learns from random balanced datasets to assess gene relevance.
• Multi-dimensional feature integration: A multi-dimensional, multi-step machine-learning approach that integrates diverse features for gene prioritization.
• Generic Mantis-ML Score (GMS): A score trained using over 1,200 features to estimate generic-disease likelihood and reported to outperform previously published gene-level scores.
• Exome-wide ranking: Ranks and triages genes across the protein-coding exome for genome-wide analyses.
• High prediction performance: Reported average area under the curve (AUC) of 0.81–0.89 when applied to chronic kidney disease (CKD), epilepsy, and amyotrophic lateral sclerosis (ALS).
• Validation with cohort-level studies: Predictions were overlapped with published cohort-level association studies and showed statistically significant enrichment surpassing state-of-the-art methods.
• Gene Prioritization Atlas: Provides predictions across ten different disease areas as a consolidated resource of mantis-ml outputs.

Scientific Applications:

• Disease research: Facilitates identification and prioritization of genes associated with specific diseases such as CKD, epilepsy, and ALS.
• Genomic data interpretation: Assists interpretation of exome-wide association studies by providing ranked lists of candidate disease genes.
• Biomarker discovery: Supports discovery of potential biomarkers through prioritized gene candidates derived from predictive models.
• Hypothesis-free genome-wide analyses: Enables objective, data-driven triaging of large-scale genomic discovery studies.

Methodology:

Uses stochastic semi-supervised machine learning with iterative training on random balanced datasets, integrates multi-dimensional features including a GMS trained on >1,200 features, validates predictions by overlap with published cohort-level association studies, and reports performance using AUC metrics.