Improving plant miRNA-target prediction with self-supervised k-mer embedding and spectral graph convolutional neural...

Area of Science:

• Plant molecular biology
• Bioinformatics
• Genomics

Background:

• MicroRNA (miRNA) regulation is crucial for plant development and phenotype variation.
• Current computational methods for miRNA target prediction often require large datasets, limiting their use in plant breeding.
• Cell-specific miRNA activity necessitates refined prediction strategies.

Purpose of the Study:

• To develop a novel and efficient computational framework, kmerPMTF, for predicting plant microRNA-target interactions.
• To overcome limitations of existing methods by reducing data requirements.
• To enhance the understanding of miRNA regulatory networks in plants.

Main Methods:

• Utilized k-mer splitting and deep self-supervised neural networks to extract sequence embeddings.
• Constructed similarity networks based on k-mer embeddings.
• Employed graph convolutional networks to derive miRNA and target representations for association probability calculation.

Main Results:

• Achieved high Area Under the Precision-Recall Curve (AUPRC) values: 84.9% (Arabidopsis thaliana), 91.0% (Oryza sativa), 80.1% (Solanum lycopersicum), and 82.1% (Prunus persica).
• Demonstrated superior performance compared to state-of-the-art methods on threshold-independent evaluation metrics.
• Validated the framework's effectiveness across diverse plant species.

Conclusions:

• The kmerPMTF framework offers an efficient and simplified approach for plant miRNA-target prediction.
• This methodology can aid in improving plant breeding through a better understanding of miRNA functions.
• Contributes to deciphering complex miRNA regulatory mechanisms in plants.