A Comparative Study of Supervised Machine Learning Algorithms for the Prediction of Long-Range Chromatin Interactions

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Three-dimensional genome organization critically regulates gene expression.
• Chromatin Conformation Capture (3C) techniques map long-range interactions but are laborious and costly.
• In silico prediction offers an alternative for generating chromatin interaction maps.

Purpose of the Study:

• To model cohesin-mediated chromatin interactions using 1D sequencing signals.
• To evaluate and compare the performance of six machine learning algorithms for predicting chromatin interactions.
• To identify key features driving accurate prediction of long-range chromatin interactions.

Main Methods:

• Utilized publicly available 1D sequencing signals from two human cell lines.
• Modeled cohesin-mediated chromatin interactions.
• Evaluated six machine learning algorithms: decision trees, random forests, gradient boosting, support vector machines, multi-layer perceptron, and deep learning.

Main Results:

• The developed approach accurately predicted long-range chromatin interactions.
• Gradient boosting significantly outperformed the other five algorithms, achieving ~95% accuracy.
• Chromatin features near interaction anchors contained most predictive information.
• Gradient boosting models showed robustness with varying feature subsets, unlike other methods.
• Transcription factor binding sites were identified as highly informative features.

Conclusions:

• Gradient boosting is the optimal algorithm for predicting long-range chromatin interactions.
• The study provides a framework for systematic prediction of chromatin interactions.
• Cell-type specific transcription factor binding at anchors significantly influences cohesin-mediated chromatin organization.