An AI-guided framework reveals conserved features governing microRNA strand selection.

MicroRNAs (miRNAs) are central regulators of gene expression, yet how cells choose between the two strands (5p or 3p) of a miRNA duplex during biogenesis remains unresolved. Here, we present a comprehensive, experimentally grounded framework that decodes the logic of miRNA strand selection. Using Caenorhabditis elegans as a model system, we developed a high-throughput platform enabling precise quantification of strand usage across developmental stages and in specific somatic tissues. To uncover the molecular grammar guiding this process, we built a predictive machine learning model trained on experimentally validated strand usage data. This AI-driven model, integrating 77 biologically informed features, accurately predicts strand preference not only in nematodes but also across vertebrates, including humans, revealing compositional and structural biases that are conserved yet functionally repurposed. Our analysis shows that strand selection is not stochastic but follows conserved, context-dependent rules shaped by cellular and developmental cues. To support the research community, we provide open-access resources: a database of strand usage profiles, predictive scores across species, and code and protocols via GitHub. This work offers the first unified, generalizable model for miRNA strand selection, establishing a paradigm that combines large-scale experimentation with AI to reveal a conserved, programmable layer of gene regulation.