Virtual-cell models treat the whole perturbation response as the answer; Deep Genomics inverts the task, learning the roughly 1% direct effects and holding up on unseen transcripts
Synopsis
Deep Genomics subtracts the whole-transcriptome perturbation response that virtual-cell models usually try to predict, treating it as noise, and trains a sequence model, DeepRNAi, only on the direct effects between an siRNA and its target transcript. Direct effects account for about 1% of all siRNA-RNA interactions, and uncalibrated virtual-cell objectives had treated that signal as a rounding error. On a held-out-transcript split, DeepRNAi performs about the same as on seen transcripts, while a conventional machine-learning approach without a direct-effect focus loses roughly 50% of its above-baseline performance.
Interpretation
Deep Genomics changes the modeling target for siRNA perturbation experiments from the full transcriptome response to the direct effect between an siRNA and its target transcript, and trains the sequence model DeepRNAi on that target. The usual approach trains a model to predict the overall perturbation response; in that response, up- and down-regulated transcripts appear in roughly equal numbers and similar patterns recur across diverse siRNAs, so most of the signal cannot be attributed to sequence-specific direct action. Across 15 million siRNA-transcript RNA-seq pairs, built from about 2,200 siRNAs and about 7,000 transcripts, the signal left after removing the overall response is strongly enriched for seed matches, where positions 2 to 7 of the siRNA pair with the transcript, and for near-perfect matches, where the 19-mer siRNA and the transcript have an edit distance below 2, indicating that the residual carries the direct effects.
DeepRNAi takes only the siRNA sequence and the target transcript sequence as input and predicts the direct effects isolated in this way. The overall perturbation response is treated as a nuisance variable to be removed first, by filling in a missing response value from the patterns of neighboring siRNAs and transcripts, a naive fill-in-the-blanks model. In ablation comparisons on held-out transcripts, DeepRNAi shows no substantial difference in performance between seen and unseen transcripts; a conventional machine-learning approach trained without a direct-effect focus drops roughly 50% in above-baseline performance on the same split, from an already-poor starting point.
Perspective
The report covers the direct-effect separation and DeepRNAi training built on about 15 million RNA-seq pairs from roughly 2,200 siRNAs and about 7,000 transcripts, together with ablation comparisons under held-out-siRNA and held-out-transcript splits; the authors say later posts will discuss how DeepRNAi relates to RNA-seq experiments and how much biological foundation model features contributed to its performance.
The authors state that direct effects, the interactions producing at least 50% knockdown, are about 1% of all siRNA-RNA interactions, so the learnable signal is itself sparse; indirect effects reproduce poorly across cell lines, whereas the direct effects DeepRNAi predicts reproduce better, and the details of that comparison are not laid out in the text.