rbpCNN: a biophysics-informed deep learning model for predicting piRNA and mRNA interactions.

Area of Science:

• Molecular Biology
• Bioinformatics
• Genetics

Background:

• Accurate prediction of piRNA-mRNA interactions is crucial for understanding post-transcriptional gene regulation in the germline.
• PIWI-guided silencing is a key mechanism, and its modulation is vital for medical applications.

Purpose of the Study:

• To develop a novel, lightweight convolutional neural network (CNN) for predicting piRNA-mRNA interactions.
• To enhance prediction accuracy by incorporating biophysically motivated interaction channels.

Main Methods:

• The proposed rbpCNN model augments nucleotide-pair encoding with five specific interaction channels: one compatibility, two helix-run, one positional, and one structural channel.
• The network architecture is designed to be lightweight while supporting CNN layer predictions.

Main Results:

• rbpCNN achieved an AUC of 96.55% and 90.66% accuracy on fivefold cross-validation.
• On an independent external dataset, rbpCNN obtained an AUC of 94.19% and 86.74% accuracy.
• Performance metrics were competitive with and, in some cases, superior to existing methods.

Conclusions:

• The rbpCNN model demonstrates high accuracy and efficiency in predicting piRNA-mRNA interactions.
• The integration of biophysically motivated channels significantly improves prediction performance.
• This lightweight CNN offers a promising tool for advancing RNA sequence interaction analysis and related medical fields.