DeepRNAi predicts siRNA off-target knockdown near experimental reproducibility, and a genome-wide scan finds off-target risk in most potent molecules
Synopsis
DeepRNAi is a deep learning model that predicts on-target and off-target knockdown directly from siRNA and mRNA sequence, trained on roughly 16 million siRNA-transcript pairs from 2,031 siRNAs and augmented with transcript context from RNA biological foundation models. On a held-out set of 240 siRNAs, its accuracy for fully complementary targets, near-perfect matches and seed matches approaches the agreement between independent RNA-seq measurements in the same cell line, and it transfers zero-shot to unseen cell lines, rat cells and rat liver in vivo. A genome-wide scan of 24.4 million candidate siRNAs against 7,946 genes finds that only 9.9% of the top 100 most potent siRNAs per gene combine no active near-perfect-match off-targets with low off-target burden.
Interpretation
Off-target knockdown by siRNA drugs can now be predicted transcript by transcript before any experiment, rather than only judging whether sequences match. Earlier siRNA predictors mostly estimated on-target potency, while bioinformatic off-target tools mainly found sequence matches, of which only 4% to 8% validate as functional off-target sites; assessing off-target burden still relied on qPCR and RNA-seq screening. On a held-out test set containing neither the training siRNAs nor other guides to the same target transcripts or paralogs, DeepRNAi classified near-perfect matches and seed matches significantly better than a random forest, TargetScan and a seed melting-temperature baseline (p < 0.001).
The model encodes the siRNA guide sequence, the on- or off-target mRNA sequence and categorical variables for cell line, siRNA format and chemistry, adds single-nucleotide-resolution embeddings from a biological foundation model, and outputs a log2 fold change for each siRNA-mRNA pair. Before training, an unsupervised, sequence-agnostic preprocessing step reduces indirect effects and noise so that learning focuses on direct, RISC-mediated, hybridization-dependent knockdown. A version retrained without the foundation-model embeddings performed significantly worse on the on-target task (p = 0.004) and on the near-perfect-match and seed-match tasks (p < 0.001), indicating the model uses that context.
Predictive accuracy carries over to cellular contexts and species absent from training, including rat liver in vivo. The model was trained only on human HepG2 and Huh7 data, with no data from other species used for training or fine-tuning. For two siRNAs targeting Hao1 and Ttr, average precision was 0.55 for near-perfect-match off-targets and 0.39 for seed matches in primary rat hepatocytes, against a baseline hit rate of 0.20; in rat liver in vivo, average precision was 0.40 across 173 near-perfect matches against a baseline of 0.16, comparable to the 0.41 obtained by using rat hepatocyte measurements to predict the in vivo result.
Perspective
The demonstrated scope is a held-out test on human HepG2 and Huh7 training data, plus zero-shot prediction on Hep3B, primary human hepatocytes from two donors, primary rat hepatocytes and rat liver in vivo; the intended reader is a researcher working on siRNA sequence design and preclinical assessment. The authors propose next steps of adding siRNAs with diverse chemical modifications, incorporating in vivo screens and clinical trial outcomes, and prospectively evaluating candidates advanced through a drug-discovery pipeline.
The authors state that the comparison separating approved from hepatotoxic siRNAs involves only a small number of molecules and warrants caution; model architecture details are proprietary and undisclosed; and the clearest test, a prospective evaluation on candidates advanced through development, has not yet been done.
