From rules to foundation models: a comprehensive review of machine learning approaches for siRNA design.

Small interfering RNAs (siRNAs) are a clinically validated therapeutic modality with eight FDA-approved drugs, yet designing effective siRNAs remains computationally challenging due to complex dependencies on sequence composition, thermodynamic properties, target-site accessibility, and off-target interactions. Over two decades, computational approaches have evolved from empirical heuristics to deep learning systems integrating physical priors with learned representations. We review the complete landscape of machine learning methods for siRNA design, spanning classical scoring rules, pretrained RNA foundation models, transformer-based efficacy predictors, graph neural networks encoding siRNA/messenger RNA interaction topology, off-target prediction frameworks, and chemical modification-aware architectures. Across over 40 studies, we identify convergent findings: hybrid models integrating thermodynamic features with learned representations are among the strongest performers, although this evidence rests largely on single-model ablations and does not establish that foundation-model embeddings specifically are required; graph neural networks with leakage-aware data splitting address pervasive benchmark inflation; and off-target prediction has matured through empirical RNA-seq frameworks and structure-based features. We distinguish throughout between chemically unmodified siRNAs, which dominate public benchmarks, and the fully modified siRNAs used therapeutically, whose efficacy data remain scarce and whose prediction is correspondingly harder. We provide a taxonomy of methods, head-to-head performance comparisons, benchmark dataset descriptions, code availability, biology-informed interpretability analysis with formal saliency validation protocols, and concrete recommendations for advancing siRNA design. Critical gaps in uncertainty quantification, active learning, and prospective experimental validation are identified as priorities for clinical translation.