ML-Peaks: Chip-seq peak detection pipeline using machine learning techniques

Abstract

CHIP-Seq data is critical for identifying the locations where proteins bind to DNA, offering valuable insights into disease molecular mechanisms and potential therapeutic targets. However, identifying regions of protein binding, or peaks, in CHIP-seq data can be challenging due to limitations in peak detection methods. Current computational tools often require manual human inspection using data visualization, making it challenging and resource demanding to detect all peaks, particularly in large datasets. CHIP-seq data poses difficulties in detecting peaks due to its high background noise, low signal-to-noise ratio, and variation in the size and shape of the peaks. To overcome these challenges, we propose a data preprocessing approach using sliding window and feature reduction techniques, and the resulting features can be further used in machine learning methods. Our machine learning methodology can accurately identify peaks using a small training set, which represents a distinct advantage over commonly used statistical approaches, as it has a greater capacity for learning from data. We tested our methodology on the H3K9me3_TDH_BP CHIP-Seq dataset exploring a range of different machine learning methods, sliding window settings, and feature reduction techniques to detect peak values without human intervention. Our pipeline efficiently detected the peaks, and achieved an F1-score of 0.9644 and a false positive rate of 0.1030.