用于小干扰 RNA 药物设计的核酸基础模型

BACKGROUND: Small interfering RNA (siRNA) therapeutics have become an important class of nucleic acid medicines. Meanwhile, nucleic acid foundation models and RNA language models have demonstrated promising performance in selected RNA-related tasks and may support siRNA drug design through transferable sequence representation learning.

APPROACH: This mini-review selectively surveys and critically evaluates representative nucleic acid foundation models and their potential roles in siRNA design, including representation learning, lowdata adaptation, multitask prediction, generative design, and multi-objective optimization. Particular attention is given to RNA language models such as RNA-FM and RiNALMo, while DNA, regulatory, unified DNA-RNA, and multimodal models are positioned according to their direct or contextual relevance to siRNA development.

RESULTS: Current models can be broadly grouped into DNA genomic language models, RNA language models, epigenomic/regulatory models, unified DNA-RNA models, and multimodal biomolecular foundation models. Across these categories, existing models provide complementary capabilities for sequence, structure, regulatory context, and cross-modal representation, but none yet constitutes a complete framework for therapeutic siRNA design.

DISCUSSION: Although these models show strong transferability across genomic and RNA tasks, their value for siRNA-specific therapeutic design remains insufficiently established. A central challenge is the mismatch between the natural RNA grammar and the therapeutic oligonucleotide grammar. Real siRNA development requires simultaneous consideration of efficacy, specificity, chemical modification, stability, immunogenicity, off-target effects, delivery compatibility, and manufacturability.

CONCLUSION: Future progress will depend on modification-aware representations, therapeutic oligonucleotide datasets, multimodal integration, interpretable and experimentally testable outputs, robust cross-dataset and cross-protocol evaluation, and Pareto-based multi-objective selection. These requirements define a practical roadmap for moving from transferable RNA representations to therapeutically credible siRNA design systems.