Background/Aims: Small interfering RNAs (siRNAs) represent a powerful strategy for targeted gene silencing therapies across a range of genetic disorders. However, the design of highly effective siRNAs that achieve strong target inhibition with minimal toxicity remains a major challenge, owing to the intricate experimental conditions and the dynamic nature of RNA molecules. An AI-driven pipeline can capture these complexities, enabling the precise identification of potent siRNA candidates and significantly accelerating the drug discovery process. Clinically approved siRNA drugs often incorporate chemical modifications to enhance their stability and activity. Yet, the limited availability of in silico research and the vast chemical modification space make it difficult to design clinically viable siRNAs. In this work, we investigate the use of various molecular fingerprints and large language models to represent chemical modification fragments, achieving an average precision of 0.80. Integrating these chemical modifications into the base model adds a novel dimension to siRNA design, offering deeper insights into how specific modifications influence binding affinity and overall therapeutic efficacy.

Finally, we implement SHAP analysis on the trained models to interpret the chemical features and substructures that drive gene knockdown efficiency. The results highlight specific features correlated with higher knockdown performance, providing valuable guidance for the design of future experimental studies.