Enhanced prediction of recombination hotspots using input features extracted by class specific autoencoders

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Recombination hotspots and coldspots are genomic regions with high and low recombination frequencies, respectively.
• Understanding these regions is crucial for insights into meiosis, sequence variation effects, and genome evolution.
• Accurate identification of recombination patterns aids in comprehending genetic diversity and evolutionary trajectories.

Purpose of the Study:

• To develop and evaluate a machine learning approach for predicting recombination hotspots and coldspots.
• To explore the efficacy of autoencoders for feature extraction in genomic prediction tasks.
• To compare the performance of different classifiers trained on autoencoder-derived features.

Main Methods:

• Utilized class-specific autoencoders for feature extraction and dimensionality reduction.
• Trained gradient boosting machines, random forest, and deep learning neural networks on extracted deep features.
• Conducted a comparative performance analysis using features from various autoencoder configurations.

Main Results:

• Deep features extracted from a combined class-specific autoencoder yielded superior performance compared to other feature sets.
• The autoencoder-based feature extraction method demonstrated enhanced prediction accuracy for recombination hotspots and coldspots.
• The study validates the effectiveness of the proposed feature extraction technique across different learning algorithms.

Conclusions:

• Combined class-specific autoencoder-based feature extraction significantly improves the prediction of genomic recombination hotspots and coldspots.
• This methodology offers a powerful tool for biological problems requiring accurate feature representation and prediction.
• The approach has broad applicability in genomics and can advance our understanding of genome evolution and meiotic recombination.