Explainable Deep Learning for Augmentation of Small RNA Expression Profiles.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Publicly available RNA expression data is rapidly expanding.
• Lack of structured metadata hinders data reusability and interpretation.
• Machine learning-based metadata prediction (data augmentation) can enhance annotation quality.

Purpose of the Study:

• To benchmark deep learning (DL) and random forest (RF) for metadata augmentation of small RNA (sRNA) expression profiles.
• To assess prediction accuracy for tissue, age, and sex using sRNA expression data.
• To identify biologically relevant sRNAs influencing DL predictions using feature importance scores.

Main Methods:

• Systematic benchmarking of DL and RF algorithms.
• Training and testing on 4243 annotated sRNA-Seq samples from the sRNA expression atlas database.
• Utilizing DeepLIFT method for backpropagation-based feature importance analysis.

Main Results:

• DL models generally outperformed RF models across tested metadata.
• Average cross-validated prediction accuracies: 96.5% (tissue), 77% (sex), 77.2% (age) for DL.
• Average tissue prediction accuracy on new datasets: 83.1% (DL) and 80.8% (RF).

Conclusions:

• DL-based metadata augmentation significantly improves annotation quality for sRNA expression data.
• The study provides insights into the biological relevance of specific sRNAs through feature importance analysis.
• Enhanced metadata facilitates better interpretation and reusability of public RNA expression datasets.