LOGOWheat: deep learning-based prediction of regulatory effects for noncoding variants in wheats

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Identifying regulatory effects of noncoding genetic variants is crucial but challenging.
• Advancements in wheat epigenomic data offer new avenues for functional variant modeling.

Purpose of the Study:

• To introduce Language of Genome for Wheat (LOGOWheat), a deep learning tool for predicting regulatory effects of noncoding variants in wheat.
• To enable accurate identification and prioritization of functional genetic variations in wheat.

Main Methods:

• Utilized a self-attention-based, contextualized pretrained language model for wheat genome representation.
• Fine-tuned the model with epigenomic profiling data to discern regulatory codes in genomic sequences.
• Employed deep learning to predict chromatin features and variant impacts.

Main Results:

• LOGOWheat achieved high accuracy in predicting chromatin features, with an average AUROC of 0.8531 and AUPRC of 0.7633.
• Demonstrated LOGOWheat's utility in scoring and prioritizing causal variants and in silico mutagenesis mapping.
• Showcased the tool's ability to identify high-impact sites and functional motifs.

Conclusions:

• LOGOWheat is an effective deep learning tool for predicting regulatory effects of noncoding variants in wheat.
• The tool facilitates the discovery of functional variations and aids in understanding wheat genome regulation.
• Integration with evolutionary conservation information can enhance the extraction of potential functional variations from wheat populations.