HaForest

HaForest identifies haploinsufficient genes from epigenomic data to enable discovery of genetic contributors to diseases such as cancers and neurodevelopmental disorders.

Key Features:

• Deep forest model: Uses a deep forest (cascade forest) architecture for tree-based ensemble learning on epigenomic features.
• Multiscale scanning with LDA: Extracts local contextual representations from input features using a multiscale scanning approach combined with Linear Discriminant Analysis (LDA).
• Cascade forest and feature concatenation: Integrates decision-tree-based forests in a cascade structure that concatenates features across layers to capture complex patterns and dependencies.
• LightGBM feature extraction: Incorporates LightGBM to reveal highly expressive features by exploiting intricate dependency structures among haploinsufficient genes.
• Robustness to bias and noise: Aims to mitigate study bias, experimental noise, and instability through its multiscale and ensemble modeling strategies.
• Benchmarking: Validated against several computational methods and four deep learning algorithms across five epigenomic datasets.

Scientific Applications:

• Haploinsufficiency prediction: Predicts haploinsufficient genes from epigenomic datasets.
• Disease genetics: Supports investigation of genetic contributors to cancers and neurodevelopmental disorders linked to haploinsufficiency.
• Epigenomic pattern discovery: Identifies epigenomic patterns and dependencies associated with haploinsufficient genes.

Methodology:

Multiscale scanning to extract local contextual representations, Linear Discriminant Analysis (LDA) for feature representation, a cascade deep forest integrating decision-tree-based forests with feature concatenation, and LightGBM for extracting expressive dependency-aware features; validated by comparisons with several computational methods and four deep learning algorithms across five epigenomic datasets.