Enhancer-IF

Enhancer-IF identifies cell-specific enhancers in the human genome using an integrative machine learning framework.

Key Features:

• Enhancer prediction: Identifies enhancer loci in human DNA sequences as regulatory DNA fragments that, when bound by transcription factors, enhance gene transcription.
• Cell-type specificity: Incorporates cell-specific modelling to distinguish enhancers across different cellular contexts.
• Addresses fixed-length sequence limitation: Explicitly targets limitations of using fixed-length sequences for model development.
• Multi-cell datasets: Leverages datasets constructed from eight distinct cell types for training and evaluation.
• Machine learning methods: Implements random forest, extremely randomized trees, multilayer perceptron, support vector machines, and extreme gradient boosting classifiers.
• Sequence encodings: Trains models using seven different sequence encodings.
• Baseline models: Generates 35 baseline models from combinations of five classifiers and seven encodings.
• Meta-model construction: Integrates outputs of baseline models and feeds them back into the same classifiers to build meta-models.
• Ensemble learning: Combines five meta-models into an ensemble to improve robustness.
• Evaluation: Assesses performance on both training and independent datasets across various cell types.

Scientific Applications:

• Genome-wide enhancer annotation: Predicts enhancer locations genome-wide in the human genome.
• Cell-type specific enhancer discovery: Identifies enhancers specific to different cell types using models trained on eight cell-type datasets.
• Functional annotation of enhancers: Supports annotation of potential enhancer functions and their physiological roles in cellular contexts.
• Comparative model benchmarking: Enables comparison of predictive accuracy between baseline models, meta-models, and the final ensemble using training and independent datasets.

Methodology:

Five classifiers (random forest, extremely randomized trees, multilayer perceptron, support vector machines, extreme gradient boosting) are trained with seven sequence encodings to produce 35 baseline models; baseline outputs are integrated and re-input into the same classifiers to build five meta-models, which are combined into an ensemble and evaluated on training and independent datasets derived from eight distinct cell types.