AIKYATAN: mapping distal regulatory elements using convolutional learning on GPU.

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• The non-coding genome presents a data deluge, necessitating sophisticated machine learning (ML) techniques for functional annotation.
• Accurately annotating regulatory elements like enhancers is challenging due to their distal nature and diverse signatures.
• Selecting appropriate ML classifiers under hardware and complexity constraints is crucial for genome annotation.

Purpose of the Study:

• To develop and evaluate ML models for classifying genome sequences as distal regulatory regions based on histone modification signatures.
• To identify the most effective ML approach for distal regulatory element (DRE) detection, considering both accuracy and computational efficiency.

Main Methods:

• Development of a suite of ML models, including Support Vector Machines (SVMs), random forest variants, and deep learning architectures, under the AIKYATAN system.
• Utilizing combinatorial signatures of histone modifications to train binary classifiers for DRE detection.
• Employing convolutional neural networks (CNNs) for their ability to model image-like properties and spatial features in epigenomic datasets.

Main Results:

• Deep learning approaches, particularly CNNs, demonstrate significant computational advantages over traditional ML models for DRE detection.
• CNNs achieved a 97.9% accuracy in classifying DREs using the H1 human embryonic cell line, with an order of magnitude lower runtime than kernel SVM.
• AIKYATAN-CNN exhibited superior prediction performance, achieving a 40% higher validation rate than CSIANN and matching the accuracy of RFECS.

Conclusions:

• Exhaustive experiments validate the need for expressive and scalable ML models to handle large and diverse epigenomic datasets.
• CNNs, with optimized activation and pooling functions, are well-suited for modeling integrated epigenomic datasets to decipher distal regulatory elements.
• The AIKYATAN suite and its source code are made publicly available to facilitate research in functional genome annotation.