Improving the Quantification of DNA Sequences Using Evolutionary Information Based on Deep Learning.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Over 98% of the human genome is non-coding DNA.
• 93% of disease-associated variants are located in non-coding regions.
• The function of non-coding DNA remains largely unknown and challenging to elucidate.

Purpose of the Study:

• To introduce a novel computational model, DQDNN, for quantifying non-coding DNA function.
• To improve the accuracy of predicting the functional impact of non-coding genomic regions.
• To enhance the prioritization of disease-associated variants within non-coding DNA.

Main Methods:

• Developed a deep neural network model (DQDNN) integrating convolution and recurrent layers.
• Utilized raw genomic sequences and evolutionary information as input.
• Employed Area Under the Receiver Operating Characteristic Curve (AUC) and Precision-Recall Curve (PRC) for performance evaluation.

Main Results:

• DQDNN significantly outperforms existing state-of-the-art models.
• The model achieved improvements of 96.9% in AUC and 98% in PRC.
• DQDNN demonstrated superior prioritization of functional variants, including expression quantitative trait loci (eQTLs).

Conclusions:

• DQDNN is an effective deep learning model for quantifying non-coding DNA function.
• Integrating evolutionary information enhances predictive performance.
• The model offers a significant advancement in understanding non-coding genome function and prioritizing disease variants.