Fast activity prediction of chemically modified siRNAs via structure-based energy calculations and inference-augmented tabular deep learning.

Chemical modification is essential for the clinical application of small interfering RNAs (siRNAs), as it improves their stability and specificity. However, predicting the activity of chemically modified siRNAs remains challenging owing to the scarcity of high-quality datasets and the computational expense of molecular dynamics (MD) simulations. In this study, we propose fast and robust activity prediction of chemically modified siRNAs via structure-based energy (FRAMEs), a novel framework that combines rapid structural prediction via deep learning with physics-based energy calculations for feature engineering of siRNA modifications. To address data scarcity, FRAMEs employs inference-augmented tabular deep learning to achieve robust activity prediction. The total energy score correlates strongly with experimental IC 50 and melting temperature, achieving performance comparable to MD-based metrics. Under both leave-one-out and stratified 5-fold cross-validation, inference-augmented TabPFN consistently outperformed all classical machine learning baselines, with the two evaluation schemes yielding mutually reinforcing conclusions. Furthermore, by exploiting physically meaningful stochasticity, FRAMEs stabilizes predictions on small datasets and exhibits strong generalization to an independent real-world dataset, outperforming existing methods. Guided by FRAMEs, several fully modified siRNA candidates targeting oncogenes relevant to cancer therapy were designed and experimentally verified. Cell-based gene-silencing assays confirmed their potent knockdown activity, validating the practical utility of FRAMEs for the rational design of therapeutic modified siRNAs.