Deep Genomics adds a million-basepair context to siRNA activity prediction: removing the foundation models nearly halves performance

DeepRNAi extends siRNA activity prediction from the local hybridization site alone to long-range context from the genomic region around the transcript, producing on- and off-target predictions across the transcriptome. Models that take only local target context work in some cases but largely do not reach the accuracy useful for drug discovery, because they miss global factors such as transcript stability, expression level, localization, site position and cell state. On seed-mediated knockdown prediction, DeepRNAi without the biological foundation models performs in aggregate about as well as the random forest baseline with manual transcript context features, and incorporating the foundation models nearly doubles performance.

The model carries global context through high-dimensional per-base-pair embeddings from biological foundation models, then trains on data generated specifically for siRNAs. The foundation models on their own are of minimal use for predicting siRNA activity and need fit-for-purpose siRNA data to contribute. Deep Genomics' BigRNA, Enigma and REPRESS read up to one million base pairs of sequence context and predict how a DNA or RNA sequence behaves in a real cell across thousands of outputs.

In real deployment, the long-range context changed the model's judgment about specific candidate molecules. The version without foundation models made very different predictions for siRNAs in the same distal 3' UTR, while a standard bioinformatics pipeline avoids tiling over regions with common variants but cannot account for variant effects that alter siRNA efficacy. For Alnylam's PLG program, DeepRNAi predicted near-zero on-target effect for molecules targeting the distal 3' UTR of the annotated transcript, downstream of the primary poly(A) site and poorly expressed in liver cell lines, and the experimental data showed those siRNAs were indeed inactive; at IRF5, a common variant with 58% global allele frequency creates a proximal poly(A) signal yielding an isoform lacking about 600 bp of the distal 3' UTR, and the foundation-model version flagged that siRNAs tiling that region are unlikely to work for much of the population.

The report covers Deep Genomics' public description of DeepRNAi: transcriptome-wide on- and off-target prediction, a performance comparison presented as an ablation on seed-mediated knockdown, and two cases involving Alnylam's PLG program and the IRF5 gene. The author states that current targets cannot be disclosed and that a DeepRNAi preprint with fuller methods and results will be released soon.

The performance gain is given only as an aggregate "nearly double," with no specific metric values; in the PLG case the author can only speculate why Alnylam screened those molecules; in the IRF5 case the variant frequency is 58% globally and its effect on particular populations is not addressed; full methods and results await the preprint.