Prediction accuracy of regulatory elements from sequence varies by functional sequencing technique.

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• * Identifying cis-regulatory elements (CREs) is crucial for understanding genome regulation.
• * Various sequencing techniques (ChIP-seq, ATAC-seq, DNase-seq, STARR-seq) are used, relying on direct sequence or indirect markers.
• * CRE activity depends on DNA sequence and secondary epigenetic processes.

Purpose of the Study:

• * To evaluate the predictive accuracy of CREs based solely on DNA sequence using machine learning.
• * To differentiate sequence-driven regulatory activity from that influenced by secondary processes.
• * To compare the suitability of different sequencing datasets for training sequence-based prediction models.

Main Methods:

• * Machine learning models were trained and evaluated on cis-regulatory element sequences from *D. melanogaster*.
• * Datasets included DNase-seq, STARR-seq, and various ChIP-seq (H3K4me1, H3K4me3, H3K27ac), FAIRE-seq, and ATAC-seq data.
• * Experimental validation using luciferase assays compared enhancer activity of selected sequences.

Main Results:

• * Models trained on DNase-seq and STARR-seq data showed significantly higher accuracy than those using ChIP-seq, FAIRE-seq, or ATAC-seq.
• * DNase-seq and STARR-seq activities are largely explained by DNA sequence, independent of secondary epigenetic marks.
• * Luciferase assays confirmed STARR-seq sequences are enriched for enhancer activity, while DNase-seq and H3K4me1 ChIP-seq sequences are not.

Conclusions:

• * DNase-seq identifies a broad range of regulatory elements, suitable for general sequence-based activity prediction.
• * STARR-seq data are optimal for training models to predict enhancer-specific sequences.
• * H3K4me1 ChIP-seq data are not well-suited for training sequence-based models for CRE prediction.