Saliency maps are increasingly used as \emph{design guidance} in siRNA efficacy prediction, yet attribution methods are rarely validated before motivating sequence edits. We introduce a \textbf{pre-synthesis gate}: a protocol for \emph{counterfactual sensitivity faithfulness} that tests whether mutating high-saliency positions changes model output more than composition-matched controls. Cross-dataset transfer reveals two failure modes that would otherwise go undetected: \emph{faithful-but-wrong} (saliency valid, predictions fail) and \emph{inverted saliency} (top-saliency edits less impactful than random). Strikingly, models trained on mRNA-level assays collapse on a luciferase reporter dataset, demonstrating that protocol shifts can silently invalidate deployment. Across four benchmarks, 19/20 fold instances pass; the single failure shows inverted saliency. A biology-informed regularizer (BioPrior) strengthens saliency faithfulness with modest, dataset-dependent predictive trade-offs. Our results establish saliency validation as essential pre-deployment practice for explanation-guided therapeutic design. Code is available at https://github.com/shadi97kh/BioPrior.
Small interfering RNAs (siRNAs) enable programmable, sequence-specific gene silencing and have become a practical modality in both therapeutic development and functional genomics (Elbashir et al., 2001;Fire et al., 1998). The clinical success of FDA-approved siRNA drugs, including patisiran, givosiran, and inclisiran (Adams et al., 2018;Balwani et al., 2020;Ray et al., 2020;Setten et al., 2019), has intensified interest in computational methods for predicting siRNA efficacy. In discovery settings, researchers routinely screen many candidate oligonucleotides and prioritize those expected to achieve strong knockdown. This has driven sustained interest in machine learning models that predict siRNA efficacy directly from nucleotide sequence and related descriptors (Han et al., 2018;Bai et al., 2024). In practice, this means: pick sequences, edit motifs or seed composition, adjust GC balance, and re-screen, so explanation quality directly affects experimental cost and iteration speed. Reliable saliency maps would enable practitioners to rationally edit candidate siRNA sequences at positions most likely to improve knockdown, accelerating the design of effective therapeutic oligonucleotides while reducing costly experimental iterations.
Modern deep predictors can be accurate on standard benchmarks, but an increasingly important question is whether they are trustworthy as decision-support tools. In practice, investigators do not only use predicted efficacy scores; they often inspect saliency maps or other attribution visualizations to infer which nucleotides “matter,” and then use those attributions to motivate sequence edits (e.g., adjusting seed composition, GC balance, or motif avoidance). If these explanations are not faithful (meaning interventions at highlighted positions do not produce larger prediction changes than controls), then explanation-guided design can be misleading (Adebayo et al., 2018;Kindermans et al., 2019), especially under the protocol and distribution shifts that are common across assays, labs, and readouts.
Figure 1: Positioning saliency validation in the lab-in-the-loop decision pipeline. In both therapeutic lead selection and functional genomics knockdown screens, researchers rely on predicted efficacy and position-level saliency (“important positions”) to decide which siRNA sequences to synthesize or prioritize for experimental validation. However, explanation methods can appear plausible while failing basic perturbation tests, a risk that compounds under assay or protocol shift across laboratories, cell lines, and readout technologies. This paper introduces a standardized faithfulness check (expected-effect perturbations with a nucleotide-matched baseline) that practitioners can apply as a pre-synthesis gate before acting on saliency maps in a new dataset or experimental setting. When validation passes, saliency-guided decisions (sequence edits, candidate ranking) can be trusted; when it fails, predictions may still be useful but position-importance reasoning should be avoided. Downstream wet-lab validation and clinical development (dashed region) are outside the scope of this work.
We therefore focus on a concrete, testable desideratum: a saliency method is useful for design only if mutating high-saliency positions changes the model’s prediction more than mutating appropriate controls (Samek et al., 2017). We term this counterfactual faithfulness: the saliency map correctly identifies positions where the model is sensitive to interventions. We propose this test as a pre-synthesis gate, a validation step practitioners should run before acting on saliency maps in explanation-guided siRNA design. Crucially, this is a model-centric guarantee: saliency tracks where the model is sensitive, not necessarily what biology truly cares about. This is distinct from biological causality (whether position changes affect true efficacy) and from distributional faithfulness (whether saliency reflects learned correlations). Our test validates model sensitivity under single-base perturbations, which is the operationally relevant property for explanation-guided sequence editing.
We introduce a perturbation-based validation protocol that operationalizes this idea for nucleotide sequence predictors. Given a trained model and a held-out siRNA, we (i) compute position-wise saliency on nucleotide identity channels (Simonyan et al., 2014;Sundararajan et al., 2017), (ii) select the top-k salient positions, (iii) quantify an expected-effect mutation score by averaging the prediction change under all single-base substitutions at those positions, and (iv) compare this score to a nucleotide-matched random baseline to control for compositional bias. This yields a simple pass/fail faithfulness test that can be run before saliency maps are used for design guidance (Amorim et al., 2023). Our protocol’s key innovations over standard in-silico mutagenesis and faithfulness metrics are detailed in Section 2 and Table 1.
This content is AI-processed based on open access ArXiv data.