Abstract
The limited repertoire of experimentally validated RNA-targeting nucleases has constrained both mechanistic studies and the efficient discovery of novel enzymes for RNA biotechnology. This challenge is particularly pronounced for prokaryotic Argonaute (Ago) proteins, where the scarcity of confirmed RNA-targeting members and a lack of clarity regarding RNA specificity determinants hinder systematic exploration. Although machine learning offers a potential solution, its application is often impeded by the scarcity of labeled training data in this field. To address these limitations, we developed the self-iterative hierarchical ensemble model (SIM), which integrates hierarchical ensemble learning with a self-training strategy. This approach bypasses the dependency on large-scale experimental datasets, allowing SIM to iteratively expand its predictive capability from minimal initial labeled data. When applied to prokaryotic Agos, SIM identified six high-confidence RNA-targeting candidates, five of which were experimentally validated (83% success rate). Notably, SIM identified three uncharacterized Agos harboring a novel N-terminal domain, defining a previously unrecognized subclass. Biochemical and in vivo validations of Haloferax profundi Ago (HpAgo) confirmed its RNA cleavage activity and a distinctive RNA modification-sensing capability. We leveraged this latter finding to develop a rapid, cost-effective method for quantifying modified RNAs. Our study not only expands the repertoire of RNA-targeting tools but also establishes SIM as a generalizable framework for protein function prediction under data-scarce conditions. This work has broad implications for both RNA biotechnology and the application of machine learning in data-limited fields.