acceptodds
Under review as a conference paper at ICLR 2027

siDiff: Efficacy-Guided Discrete Diffusion for siRNA Target-Site Discovery

Abstract

High-quality experimental data for siRNA efficacy remain scarce, limiting the generalization of discriminative predictors to unseen genes. Although conventional pipelines enumerate all complementarity-valid target windows, prediction errors may concentrate top-ranked candidates within a few local regions while leaving other experimentally active regions under-ranked. Recovering such regions is important for therapeutic discovery, where a diverse portfolio of non-redundant candidates is needed to withstand downstream attrition caused by accessibility, off-target effects, chemical modification, and safety constraints. We propose siDiff, an efficacy-guided discrete diffusion framework for siRNA target-site discovery. siDiff first learns a target-conditioned distribution over functional siRNA duplexes and then uses efficacy guidance to reshape this learned distribution during sampling, increasing the probability of efficacy-favored sequence patterns and target regions without retraining the generator. Its coarse-to-fine sampler progressively reveals the siRNA sequence: position guidance prioritizes coordinates with high efficacy potential, while delayed base guidance steers nucleotide assignments only after sufficient sequence context has emerged. By conditioning each decision on an increasingly informative partial sequence, this progressive process captures combinatorial motifs and higher-order intra-siRNA dependencies more explicitly than single-pass candidate scoring. The guided active patterns consequently recover promising regions that may be under-ranked by discriminative predictors, while reverse-complement construction and spatial clustering preserve duplex validity and regional diversity. Experiments on the public Takayuki benchmark and three patent-derived targets demonstrate improved positional recovery and coverage of distinct active regions on unseen genes.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
siRNA Design, Discrete Diffusion Models, Efficacy-Guided Generation, Biological Sequence Design

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_1yY9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the problem of identifying effective siRNA target sites on unseen genes, where limited experimental data can cause conventional efficacy predictors to overlook promising target regions. To address this challenge, the paper proposes siDiff, an efficacy-guided discrete diffusion framework that learns a target-conditioned distribution of functional siRNA sequences and guides candidate generation toward potentially effective regions. The proposed method combines coarse-to-fine sampling with position-level and nucleotide-level efficacy guidance, followed by reverse-complementary candidate construction and cluster-aware selection to preserve regional diversity. The main contributions include a generative framework for siRNA target-site discovery, a biology-aware guided sampling strategy, and three patent-derived evaluation benchmarks with positional accuracy and regional coverage metrics. Experiments on the Takayuki benchmark and three unseen patent-derived targets demonstrate improved target-site recovery and coverage of distinct high-efficacy regions compared with conventional enumeration-based methods and diffusion baselines.

Critical strengths

Applying a discrete masked diffusion process to the siRNA candidate ranking problem is a fresh and interesting approach. Cluster-Aware Redundancy Mitigation was impressive, as it helps to reduce redundant candidates whilst maintaining diversity across different target domains.

Critical weaknesses

Although this may be a limitation of the task itself, the contribution of this work to therapeutic discovery seems limited to computational target-site recovery, rather than actual therapeutic discovery. The baselines used in this study are somewhat outdated, and comparisons with more recent SOTA models are missing. It is also disappointing that the paper does not provide a detailed case study of individual target sites identified by the proposed model, although regional coverage analyses are presented.

To make the manuscript more concrete, I believe it would be helpful to evaluate the proposed model on an siRNA efficacy prediction task and compare it with SOTA methods. Demonstrating strong prediction performance or showing that the model learns biologically meaningful representations would make this work much more convincing.

The authors claim that generative proposals can recover effective mRNA regions that existing predictors fail to identify. However, the main experimental results primarily measure whether the model correctly recovers specific annotated active positions. There is no direct analysis showing how siDiff recovers regions that were incorrectly ranked by existing predictors. Therefore, there is a gap between the authors' main claim and the experimental evidence presented in the paper.

[1] Gao, Bowen, et al. "Drugclip: Contrastive protein-molecule representation learning for virtual screening." Advances in Neural Information Processing Systems 36 (2023): 44595-44614.

Questions for authors
  1. The latent diffusion decoder of RNAGenesis[1] can generate RNA sequences based on context. I believe comparisons with models such as RNAGenesis, RiNALMo[2], RNA-FM, or even protein foundation models such as the ESM family would be valuable. Several categories of recent SOTA models have been introduced that could potentially be applied to siRNA-related tasks. As an additional suggestion, have the authors considered incorporating therapeutic feature? Could the authors provide a head-to-head performance comparison with these models, adapted to the task addressed in this paper? I believe such results would make the manuscript more concrete and convincing.

  2. If the proposed model performs well in position recovery and clustering, it is anticipated that it will also be able to learn representations useful for predicting siRNA efficacy.

  3. According to Appendix Table 10, OligoFormer shows a relatively stronger correlation with experimentally measured efficacy but achieves lower target-site recovery performance. In contrast, siDiff achieves higher position recovery performance despite using a predictor with weaker efficacy correlation. Could the authors explain why this happens and which components contribute most to this difference?

  4. Regarding the dataset construction, please provide a reference or justification for choosing the inhibition threshold of 0.7. It would also be helpful to show whether the results remain consistent under different thresholds.

[1] Zhang, Zaixi, et al. "RNAGenesis: a generalist foundation model for functional RNA therapeutics." bioRxiv (2024): 2024-12.

[2] Penić, Rafael Josip, et al. "RiNALMo: general-purpose RNA language models can generalize well on structure prediction tasks." Nature Communications 16.1 (2025): 5671.

3 comments

Response to Reviewer 3

We sincerely thank the reviewer for the constructive assessment. These comments help us make both the scope and the empirical evidence of the work more concrete.

1. [W1, Q1] siDiff proposes diverse binding regions for downstream screening for therapeutic development instead of efficacy prediction

The intended contribution of siDiff identifying a compact and spatially diverse set of promising target regions and valid siRNA candidates for downstream experimental screening.

The output is the siRNA binding regions rather than an efficacy score. Predicting efficacy requires a ranking model that outputs the exact siRNA efficacy score.

Our goal is to recover several promising and non-redundant target regions rather than to concentrate the full candidate portfolio in one local region. siDiff is designed to accelerate candidate proposal and experimental prioritization at this stage.

2. [W1, Q1] siRNADiscovery is compared directly; adapted generators are in the common comments, and the remaining suggested models do not fit this task

We appreciate the recommendation to include more recent methods. We will extend the empirical comparison to models that can be reproducibly adapted to conditional siRNA generation or target-site proposal shown in the common comments section 1 :

  • siRNADiscovery, a recent siRNA-specific method;
  • an autoregressive Transformer;
  • a conditional autoencoder;
  • a conditional variational autoencoder;

We also add siRNADiscovery into our benchmark

MethodTaka Hit0@1Taka Hit3@3ANGPTL7 Hit0@1ANGPTL7 Hit3@3CTNNB1 Hit0@1CTNNB1 Hit3@3GSK3A Hit0@1GSK3A Hit3@3
siRNADiscovery0.0210.4190.0320.4840.0330.5210.0260.603
siDiff0.2420.6640.1850.6120.2000.6540.1010.679

For the comparison with the autoregressive Transformer, the conditional autoencoder, and the conditional VAE under the shared downstream protocol, please refer to Section 1 of the Common Official Comments.

We also evaluated RNAGenesis. Its original configuration is not designed for 21-nt siRNA generation conditioned on a target mRNA. To adapt it, we modified its denoiser to support the required sequence length and fine-tuned it on the Huesken data with mRNA conditioning. Under the same reverse-complement matching protocol with a mismatch tolerance of 5, RNAGenesis produced no valid siRNA candidate on any of the four external test datasets. The generation success rate is 0% for Takayuki, ANGPTL7, CTNNB1, and GSK3A. We report this result for completeness. It indicates that adapting a general RNA generator to target-complementary siRNA proposal requires more than changing its output length and fine-tuning it on this task.

We also investigated the suggested siDPT comparison. Its executable implementation and source code are not publicly available.

The other suggested foundation models do not provide a compatible siRNA-generation interface. RiNALMo and RNA-FM are RNA representation models rather than mRNA-conditioned 21-nt siRNA generators. ESM operates on proteins, and DrugCLIP is designed for protein–small-molecule screening. Their native outputs cannot generate and rank reverse-complementary siRNA candidates for a target mRNA, so they are not direct proposal-model baselines.

3. [W3] Hit0@1, Hit3@3, and EWC measure recovery of active sites and regions that discriminative rankings do not place first

The improved Hit0@1, Hit3@3, and EWC results show that siDiff recovers experimentally active sites and regions more effectively than the evaluated discriminative ranking pipelines on unseen targets. In particular:

  • Hit0@1 measures exact recovery by the highest-ranked candidate;
  • Hit3@3 measures whether a compact top-three portfolio recovers an experimentally active position within three nucleotides; and
  • EWC measures recovery of distinct experimentally supported high-efficacy regions across the mRNA.

Together, these results show that the generative proposal-and-selection process can recover active positions and regions that the discriminative baselines do not place at the top. This recovery is practically useful because it identifies localized regions for subsequent experimental validation rather than relying only on a single predictor ranking.

4. [W1, W3] Potent sites ranked near the bottom by iScore, OligoFormer, and siRNADiscovery are recovered within 1–3 nt, as detailed in the common comments

Please refer to Section 1 of the Common Official Comments for the coordinate-verified strict Hit3@3 cases. All three sites have normalized inhibition above 0.91, yet the corresponding baseline ranks each in the bottom three of the 41 local candidates, at 40/41, 39/41, and 39/41. siDiff places a final top-three proposal within 3, 1, and 2 nt, respectively. The OligoFormer and siRNADiscovery sites are already within 1–2 nt of siDiff's top-1 prediction. For the iScore site, the top-1 distance is 14 nt, and the top-three portfolio supplies the candidate at 3 nt. The corresponding baseline scores are 29.3, 0.388, and 0.429. These are strict Hit3@3 successes under the unchanged evaluation rule, and they show the benefit of a small, diverse proposal set over a single pointwise maximum.

5. [W2, Q2] A regression head on frozen Diffusion representations predicts held-out inhibition with PCC 0.536 and SCC 0.548

We thank the reviewer for asking whether Diffusion learns an efficacy-aware representation, rather than only sequence regularities sufficient for masked-nucleotide reconstruction. We tested this with a frozen-representation transfer experiment. Clean siRNA inputs were passed through the trained CondDiT, its internal representations were pooled, all Diffusion parameters were frozen, and only a lightweight regression head (two linear layers and one activation function) was trained to predict continuous experimental inhibition. Evaluation was performed on the four held-out external targets.

Evaluation scopePCCSCCRMSE
Four held-out external targets0.5360.5480.0920

The frozen Diffusion representation achieves a PCC of 0.536 and an SCC of 0.548. It therefore retains a clear efficacy-related signal that transfers to unseen target contexts.

6. [Q3] Table 10 evaluates only the lightweight post-hoc scorer; siDiff's recovery comes from the full proposal pipeline

Appendix Table 10 evaluates the lightweight predictor used only for rapid post-generation scoring and candidate filtering. OligoFormer uses a longer 57-nt mRNA context and thermodynamic features, and it is optimized as a dedicated efficacy-prediction model. It is therefore reasonable that OligoFormer obtains stronger standalone PCC/SCC values in Table 10.

The complete siDiff framework still achieves stronger target-site recovery than the OligoFormer enumeration baseline, because its output is not obtained by globally sorting all 41 candidates with this lightweight full-sequence predictor. It comes from the diffusion model during multi-step generation.

The predictor serves only as a rapid downstream scoring and filtering component within a broader proposal-and-selection process that combines:

+ \text{partial-state efficacy guidance} + \text{reverse-complement candidate construction} + \text{region-level consensus selection}.$$ ## 7. [Q4] The 0.7 inhibition cutoff follows prior siRNA design practice and public implementations, rather than test-set tuning The normalized inhibition threshold of $0.7$ was adopted from established siRNA efficacy practice rather than selected post hoc from our test results. The convention is explicit in both classical design work and recent predictors. OligoWalk defines an efficient siRNA as one with **greater than 70% silencing efficacy** [Lu and Mathews, *Nucleic Acids Research*, 2008](https://doi.org/10.1093/nar/gkn250). More recently, the official OligoFormer implementation converts measured efficacy into the positive class using `label > 0.7` ([official](https://github.com/lulab/OligoFormer/blob/main/scripts/test.py) `test.py`), and the official siRNADiscovery implementation uses the same `efficacy > 0.7` boundary for its binary efficacy evaluation ([siRNA-split code](https://github.com/BertramLoong/siRNADiscovery/blob/main/siRNA_split/siRNADiscovery.py); [mRNA-split code](https://github.com/BertramLoong/siRNADiscovery/blob/main/mRNA_split/siRNADiscovery.py)). These recent models are built on the influential 2,431-siRNA Huesken benchmark reported in *Nature Biotechnology* [Huesken et al., 2005](https://doi.org/10.1038/nbt1118), which was produced in an industrial drug-discovery setting by researchers from Novartis and Compugen. Thus, the 70% activity boundary has precedent across classical siRNA design, modern efficacy prediction, and industry-generated screening data.

Thank you for providing the additional experiments and detailed explanations. The new comparative analyses with recent models, including RNAGenesis, and the additional case studies demonstrating that siDiff can recover active sites that were ranked lower by existing prediction models, were most appreciated. However, I remain concerned that the scope of the problem formulation is relatively narrow. The three case studies are interesting, but I still believe they are insufficient to demonstrate that the model can consistently recover promising regions overlooked by existing prediction models across a diverse range of unknown genes. I also appreciate the explanation regarding Q3, but the relative contributions of individual components to the performance improvement remain somewhat unclear. Therefore, the additional results have addressed some of my concerns, but my main concerns regarding the scope of this study and the empirical evidence remain. I therefore maintain my original score.

Reviewer_MjRTOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes siDiff, a framework for proposing siRNA target sites using target-conditioned masked discrete diffusion. The generator is trained on high-activity Huesken-derived examples. At inference, a partial-state efficacy predictor guides the order of nucleotide unmasking and, after a delay, nucleotide selection. The resulting partial patterns are mapped to valid reverse-complementary target windows; a separate full-sequence predictor and spatial clustering then select complete candidates. The objective is to recover diverse active regions that may be under-ranked by conventional enumeration and scoring.

The evaluation uses Takayuki and three patent-derived datasets targeting ANGPTL7, CTNNB1, and GSK3A, which the authors state are excluded from training and hyperparameter selection. The paper reports improved local target-site recovery and efficacy-window coverage, together with ablations of guidance and clustering. These are retrospective computational evaluations; the paper does not report new experimental validation of proposed siRNAs.

Critical strengths

1. A relevant discovery objective with a coherent proposal-and-selection design.

Recovering multiple active target regions is a meaningful objective when candidates may subsequently fail downstream screening. siDiff combines a learned proposal prior, guidance over partial sequences, and spatial selection while ensuring that final candidates are constructed from valid reverse-complementary windows. This makes the method more directly relevant to target-site selection than unconstrained sequence generation.

2. Encouraging external-dataset results and informative component analyses.

Table 1 reports substantial improvements under the stated evaluation protocol. For example, on Takayuki, Hit0@1 increases from the strongest enumeration result of 0.048 to 0.242, and Hit3@3 from 0.401 to 0.664. Reported coverage reaches 28/29 windows on CTNNB1 and 9/9 on GSK3A. Evaluating on datasets held out from the proposed models' training, and examining guidance schedules and clustering, provides useful initial evidence that merits further investigation.

Critical weaknesses

1. The comparisons do not isolate the value of guided diffusion from the effects of scoring and diversity-aware selection.

The central comparison changes several components simultaneously: the proposal mechanism, the efficacy predictor, and the selection policy. The enumeration baselines use iScore or OligoFormer, whereas siDiff uses separately trained predictors and cluster-based filtering and quotas. The paper does not report exhaustive enumeration scored by the same full-sequence predictor and processed using the same clustering and selection policy. This is especially important because each evaluated 61-nt context contains only 41 valid starts, and clustering has a large reported effect: in Table 2, removing it reduces Takayuki Hit0@1 from 0.242 to 0.016. Consequently, the current results establish a difference between complete pipelines, but do not establish that diffusion is responsible for the improvement over a simpler enumeration pipeline with comparable selection.

The EWC comparison also depends on each method's own predicted scores and thresholding (Appendix D). Although these scores are not directly pooled, applying the same numerical high-efficacy criteria can make coverage sensitive to calibration differences between predictors. EWC is a recall measure, as the authors acknowledge, and the total candidate budget used to obtain it is not sufficiently explicit. A comparison at a common final candidate budget, with coverage determined by independent experimental annotations, would make the claimed accuracy–diversity advantage substantially more persuasive.

2. The evaluation supports a narrower claim than general discovery of experimentally potent target sites, and the patent-label interpretation is not established.

The positional benchmark is constructed from retained active sites, with up to five 61-nt contexts around each site (Section 4.1 and Table 5). This is a useful test of local recovery when an active site is known to lie within the input neighborhood. It does not by itself establish successful screening across an entire previously unseen transcript. It remains unclear whether the EWC analysis independently scans complete transcripts or aggregates predictions from these activity-conditioned contexts. The distinction affects the strength of the target-region discovery claim.

Appendix A further assumes that unmodified efficacy upper-bounds efficacy under every chemical modification, then uses the maximum measured value over modifications as a proxy for unmodified efficacy. The paper does not validate this assumption, and an upper-bound assumption alone does not establish that the observed maximum estimates that bound accurately. Taking a maximum can also favor sequences with more tested modification variants. A fixed modification pattern for ANGPTL7 and GSK3A supports comparisons within that pattern, but does not turn the measurements into unmodified efficacy. In addition, the activity thresholds follow per-mRNA min–max normalization and therefore describe relative activity rather than a uniform absolute knockdown threshold. These choices need sensitivity analysis or a narrower interpretation as recovery of sites with high observed activity under the recorded assay conditions.

Questions for authors

1. Controlled comparison — additional analysis. Could you compare siDiff with exhaustive enumeration using the same full-sequence predictor, cluster definition, filtering thresholds, quotas, and final candidate budget? Please also include a simple diversity-aware selection baseline and report inference cost. Does the advantage remain when proposal generation is the principal changed component?

2. Inference inputs and evaluation scope — clarification. Appendix E.1 states that the guidance predictor receives a 21-nt mRNA target segment, whereas the discovery input is a 61-nt context with the target position withheld. How is this 21-nt segment selected at each sampling step without access to the annotated target position? Is EWC obtained by independently scanning full transcripts, or by aggregating outputs from contexts constructed around known active sites? Please specify how these inputs are generated.

3. Patent labels — clarification and sensitivity analysis. What evidence supports the unmodified-efficacy upper bound and the maximum-over-modifications proxy in Equations 18–19? Please provide patent identifiers and the mapping from assay records to sequence-level labels, including modification patterns and assay conditions. Do the main conclusions persist when CTNNB1 is analyzed within a matched modification/assay setting, or when the maximum aggregation is replaced by another justified summary?

4. Coverage and candidate budgets — clarification and additional analysis. How many unique candidates does each method retain for EWC, and how are the numerical thresholds made comparable across predictors? Please report experimentally annotated active-site or active-window recovery at common candidate budgets, with the fraction of selected candidates having measured outcomes. Unmeasured candidates should remain unknown rather than being treated as experimentally inactive or active.

5. Validation units and uncertainty — clarification and additional analysis. Appendix F describes an 80/20 split after constructing augmented context instances. Can views of the same siRNA or target site appear in both partitions, and how are the two efficacy predictors split? Please clarify the independent unit used for evaluation and uncertainty estimation, report the seed/repeat protocol for the headline comparisons, and account for repeated contexts from the same site. The stated external holdouts are valuable; this question concerns internal model selection and the precision of the reported estimates, rather than asserting external-test leakage.

5 comments

Response to Reviewer 2

We sincerely thank the reviewer for the careful and constructive evaluation. The clarifications and additional controlled analyses are given below.

1. [W1, Q1, Q2] Sampling guidance and post-matching scoring use two different predictors, so iScore and OligoFormer cannot replace guidance

The siDiff pipeline uses two separately trained efficacy predictors with different inputs and roles:

a 21-nt partial-state siRNA inhibition predictor, used only during masked-diffusion sampling which input is 21-nt partial-state siRNA; and

  1. a full-sequence efficacy predictor, used only after partial patterns have been mapped to valid reverse-complementary mRNA windows and complete siRNA candidates have been constructed.

1.1 [W2, Q2] Guidance receives the partially masked 21-nt siRNA, not an annotated mRNA segment

The 61-nt mRNA window is only the conditioning input to the diffusion model. The guidance predictor does not receive an annotated 21-nt mRNA target segment. Its input is the current 21-nt partially masked siRNA state produced by the diffusion model.

1.2 [W1, Q1] iScore and OligoFormer score complete siRNAs and cannot evaluate masked intermediate states

The guidance predictor evaluates incomplete siRNAs containing varying numbers and arrangements of [MASK] tokens. In contrast, complete-candidate predictors such as iScore and OligoFormer score clean, complete siRNAs for which the target relationship, the required sequence context, and the thermodynamic features are already defined.

1.3 [W1, Q1] siRNADiscovery is compared directly, and the alternative generators are in the common comments

For the controlled comparison of proposal mechanisms (an autoregressive Transformer, a conditional autoencoder, and a conditional VAE), please refer to Section 1 of the Common Official Comments.

We also compare siDiff with the recent siRNA-specific baseline siRNADiscovery. Results are reported separately for the four targets, without macro-averaging. Hit0@1 uses K=1,δ=0K=1,\delta=0, and Hit3@3 uses K=3,δ=3K=3,\delta=3.

MethodTaka Hit0@1Taka Hit3@3ANGPTL7 Hit0@1ANGPTL7 Hit3@3CTNNB1 Hit0@1CTNNB1 Hit3@3GSK3A Hit0@1GSK3A Hit3@3
siRNADiscovery4/191 = 0.02180/191 = 0.4191/31 = 0.03215/31 = 0.4848/242 = 0.033126/242 = 0.5212/78 = 0.02647/78 = 0.603
siDiff0.2420.6640.1850.6120.2000.6540.1010.679

Together with Section 1 of the Common Official Comments, this comparison tests whether masked diffusion and its efficacy guidance improve target-site proposal when the rest of the evaluation pipeline is held fixed.

2. [W1, Q1] Clustering consolidates unordered diffusion matches, and the shared selection results are in the common comments

The cluster-aware stage follows the structure of masked-diffusion output. siDiff does not generate mRNA sites from left to right. It progressively unmasks a 21-nt siRNA pattern, and the resulting partial pattern may reverse-complement match one or more target positions. Different sampling trajectories can also map to identical or neighboring mRNA starts. The matched proposals are therefore unordered in mRNA coordinates and locally redundant.

Exhaustive enumeration and left-to-right generation have a different output structure. Each complete candidate is already tied to a unique, ordered mRNA start, and there are no independently sampled partial patterns that must first be mapped and consolidated. The original siDiff clustering step is therefore a natural part of the masked-diffusion pipeline, not a required processing step for conventional enumeration.

For the cluster-level selection rule, the enumeration-compatible regional-consensus analysis on iScore and OligoFormer, and the cases in which potent sites are recovered despite low baseline ranks, please refer to Section 1 of the Common Official Comments.

3. [W1, Q4] After the shared EWC normalization, the same window counts also give regional precision and F1

For the EWC score normalization and the cross-method F1 comparison, please refer to Section 2 of the Common Official Comments.

To make the spatial agreement in Figure 3 explicit, we summarize the existing EWC outputs below. Because EWC measures regional recall, we also derive regional precision and F1 from these same window counts:

R_{\mathrm{reg}}=\mathrm{EWC}=\frac{|W_{\mathrm{pred}}\cap W_{\mathrm{GT}}|}{|W_{\mathrm{GT}}|},\qquad F1_{\mathrm{reg}}=\frac{2P_{\mathrm{reg}}R_{\mathrm{reg}}}{P_{\mathrm{reg}}+R_{\mathrm{reg}}}.$$ | Dataset | All windows | GT-positive | Predicted-positive | Overlap | Prediction-only | GT-only | Predicted-positive rate | Regional precision | Regional recall/EWC | Regional F1 | | ------- | ----------: | ----------: | -----------------: | ------: | --------------: | ------: | ----------------------: | -----------------: | ------------------: | ----------: | | | | | | | | | | | | | | ANGPTL7 | 22 | 1 | 5 | 1 | 4 | 0 | 0.227 | 0.200 | 1.000 | **0.333** | | CTNNB1 | 36 | 29 | 33 | 28 | 5 | 1 | 0.917 | 0.848 | 0.966 | **0.903** | | GSK3A | 21 | 9 | 14 | 9 | 5 | 0 | 0.667 | 0.643 | 1.000 | **0.783** | These window counts and F1 scores show that siDiff retains a clear advantage in siRNA binding-site selection under matched candidate budgets relative to other SOTA methods. ## 4. [Q5] The 80/20 split precedes augmentation, and the main comparisons share seeds All model training and internal validation, including the proposal generator, the partial-state guidance predictor, and the full-sequence candidate evaluator, use the Huesken dataset. Takayuki, ANGPTL7, CTNNB1, and GSK3A are external test targets only. They do not affect Huesken validation or model selection. Within Huesken, preprocessing proceeds as follows: $$\text{unique Huesken target sites} \rightarrow \text{site-level 80/20 train/validation split} \rightarrow \text{random-window augmentation within each partition} \rightarrow \text{masked-state construction within each partition}.$$ The split is therefore performed on independent Huesken target sites **before** context augmentation or mask-state generation. All random-window context views derived from one target site remain in the same partition. The main experiment, the ablation study, the guidance-gate sensitivity analysis, and the guidance-mode comparison all use the same four random seeds. ## 5. [W2, Q3] The maximum over modifications removes unfavorable modification artifacts and keeps each site's functional ceiling The patent dataset is used **only** as a held-out evaluation benchmark and is never seen during model training. In this evaluation, sequence-intrinsic properties (mRNA target accessibility and RISC cleavage thermodynamics) set the functional ceiling of a target site, whereas suboptimal chemical modifications often cause false-negative readouts in the patent tables. If raw or mean values were used as ground truth, a model would be penalized for unfavorable modification variants. Taking the maximum across modifications (Equations 18–19) isolates each sequence's realized functional ceiling after those unfavorable modification artifacts are removed.

I had a heavy reviewing workload this round, so I’m sorry for following up later than planned. Thank you to the authors for the additional results. I revisited the main concerns from my initial review. The new baselines strengthen the evidence for local target-site recovery, but the exhaustive enumeration baselines still use different scoring and selection rules, making it difficult to isolate the contribution of guided diffusion. It remains unclear whether the predictions used to compute EWC came from independent full-transcript scans. The rebuttal also states that the guidance predictor uses only the partially masked siRNA, while Appendix E.1 of the revised manuscript still lists an mRNA input. This is causing some confusion here. I therefore maintain my rating of 2 (Weak rejection).

Reviewer_YRvnOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes siDiff, a discrete diffusion approach for finding potent siRNA target sites on mRNA. Instead of scoring all valid 21-nt windows directly with a predictor, the authors train a masked diffusion transformer on high-efficacy siRNA-mRNA pairs and use a separate efficacy predictor to guide the unmasking process (first guiding which position to unmask, and later guiding the nucleotide choice). The partially unmasked sequence is then matched back to the target mRNA to extract valid 21-nt candidates, which are grouped by position and ranked by a full-sequence predictor. The method is tested on the Takayuki dataset and three newly curated patent datasets.

Critical strengths

1.Practical motivation and useful patent benchmarks. In practice, standard siRNA predictors often rank overlapping windows in a single region at the top, which is risky if that region turns out to be inaccessible or toxic. Curating three external test sets (ANGPTL7, CTNNB1, GSK3A) from patents and trying to control for chemical modification effects gives a nice benchmark for testing generalization to unseen genes. 2.Reasonable coarse-to-fine guidance design. Waiting until more bases are unmasked before applying base-level guidance makes intuitive sense, since efficacy predictions on heavily masked sequences are unreliable. The curriculum training experiments in the appendix back this up well.

Critical weaknesses

1.The ablation study contradicts the paper's main claims, showing that the gains come from post-hoc clustering rather than guided diffusion. Looking at Table 2, removing the clustering step (siDiff w/o clustering) causes Hit0@1 to collapse down to 0.02, which is no better than uniform random guessing and worse than the simple iScore baseline. Conversely, keeping the clustering step but turning off all diffusion guidance (No guidance in Table 2) actually gets better Hit0@1 than the full siDiff model on 3 out of the 4 datasets. It also makes no sense why a clustering step—which is meant to improve diversity across top-K predictions—would boost top-1 accuracy by 10x, unless sorting clusters by start position is exploiting some positional bias in how the 61-nt test windows were cropped around the true site. Moreover, Table 10 shows that OligoFormer and iScore are much stronger predictors than the authors' own full-sequence predictor on most targets, yet the enumeration baselines in Table 1 were not given the same clustering and filtering post-processing.

2.Using discrete diffusion and fuzzy string matching to search a space of 41 valid windows is poorly motivated, and key inference parameters are hidden. Because a functional siRNA must be the reverse complement of a 21-nt window on the target mRNA, a 61-nt test input only has 41 valid candidate sequences in total. Running a 7.6M-parameter diffusion model for 16 steps and calling a transformer guidance predictor hundreds of times per trajectory—only to discard the invalid generated bases and align the partial pattern back to those 41 windows via fuzzy string matching—is hard to justify over simply scoring the 41 windows directly. Worse, the paper never states how many trajectories are sampled per context or what the mismatch tolerance is in ReverseComplementMatch. Since Table 7 shows that the candidate pool before ranking already covers 92–97% of target sites within 2 nt, it appears the matching step is loose enough to let most of the 41 windows through anyway, making the diffusion step largely redundant.

3.The EWC metric is flawed, and important baselines and related work are missing. In Appendix D, EWC defines predicted active windows by applying fixed cutoffs to each method's own uncalibrated scores, without penalizing false-positive windows. Because OligoFormer outputs lower raw score values on average, it gets 0 EWC on most targets simply due to a score scale mismatch, whereas siDiff (trained only on >=0.7 sites) flags almost the entire transcript as positive and gets near 100% recall for free. Finally, the paper misses core literature on classifier/value-guided discrete diffusion (e.g., Nisonoff et al., ICLR 2025; Schiff et al., ICLR 2025; SVDD, NeurIPS 2024) and DNA/RNA discrete diffusion (e.g., D3, ICLR 2025; DRAKES, ICLR 2025). It also omits recent open-source siRNA baselines already cited in the text, such as siRNADiscovery (2024) and siDPT (2025)—notably, the margin-based pairwise ranking loss in Eqs. 7–8 is identical to the one in siDPT but is presented here without attribution.

Questions for authors

1.Why does removing clustering (siDiff w/o clustering in Table 2) drop Hit0@1 from 0.24 to 0.016, and how exactly are candidates ordered across different clusters when picking the single top-1 prediction? 2.What do the results in Table 1 look like if you apply the exact same spatial clustering and staircase selection step to the enumeration baselines (iScore, OligoFormer, and your own full-sequence predictor)? 3.In Table 2, why does No guidance beat the full siDiff model on Hit0@1 on 3 out of 4 datasets? 4.How many diffusion samples are drawn per 61-nt context, what is the mismatch tolerance when matching back to the mRNA, and what are the exact thresholds/quotas used in the staircase filter? 5.Since EWC relies on uncalibrated score cutoffs and only measures recall, how do the methods compare if you simply take the top-N candidates from each method and evaluate both regional recall and precision?

1 comment

Response to Reviewer 1

Thank you for the careful reading and for raising these important questions. We appreciate the opportunity to clarify these points.

1. [W1, Q3] Guidance gives lower Hit0@1 on three datasets, but raises Hit3@3 and multi-region EWC by avoiding single-mode sampling

The unguided generator is trained on high-inhibition siRNA–mRNA pairs. Without guidance, it tends to overfit to seen patterns, which substantially reduces the diversity of recovered target regions.

In practical drug discovery, the objective is a small, diverse set of promising candidates covering multiple potentially active target regions. Hit3@3 and EWC are therefore needed to interpret this ablation. Guidance reshapes the learned diffusion prior so that sampling explores a broader set of efficacy-favored patterns and regions on out-of-distribution targets, rather than repeatedly concentrating on a single mode. Table 2 shows Hit3@3 and EWC are largely imporved by introducing the guidance.

2. [W2, W3] siDiff outperforms siRNADiscovery; the alternative-generator comparison is in the common comments

We additionally compare siDiff with the recent siRNA-specific baseline siRNADiscovery. Results are reported separately for the four targets, without macro-averaging. Hit0@1 uses K=1,δ=0K=1,\delta=0, and Hit3@3 uses K=3,δ=3K=3,\delta=3.

MethodTaka Hit0@1Taka Hit3@3ANGPTL7 Hit0@1ANGPTL7 Hit3@3CTNNB1 Hit0@1CTNNB1 Hit3@3GSK3A Hit0@1GSK3A Hit3@3
siRNADiscovery0.0210.4190.0320.4840.0330.5210.0260.603
siDiff0.2420.6640.1850.6120.2000.6540.1010.679

For the controlled comparison with alternative conditional proposal models (an autoregressive Transformer, a conditional autoencoder, and a conditional VAE), please refer to Section 1 of the Common Official Comments.

3. [W1, Q1, Q2, Q4] Clustering, the enumeration controls, and the recovered high-activity sites are reported in the common comments

Please refer to Section 1 of the Common Official Comments for (1) the clustering and selection rule, which ranks clusters by efficacy-score aggregates rather than genomic start position (2) the adapted regional-consensus controls (Cluster A and Cluster B) applied to the enumeration baselines iScore and OligoFormer; futher more, we replace the diffsuion model with other conditional sequence generators, and (3) the strict Hit3@3 cases in which siDiff recovers experimentally potent sites ranked near the bottom by discriminative baselines.

4. [W1, Q1] Unifrom sampling over the 41 valid offsets keeps the annotated site off center, so selection has no positional cue

The 61-nt contexts are constructed by random-window augmentation. For each known 21-nt target site, the relative start is sampled uniformly from the 41 valid offsets in the 61-nt context (indices 0–40). The upstream and downstream flanks are adjusted accordingly, while the complete target site is retained. The annotated site is therefore not systematically centered, and the window does not provide a privileged center position that candidate selection could exploit. This construction introduces no positional bias.

5. [W3, Q5] EWC normalizes scores within each target, and regional F1 shows that the predictions stay localized

Please refer to Section 2 of the Common Official Comments for the EWC score normalization, the localized spatial patterns in Figure 3, and the regional F1 comparison across methods and test datasets.

6. [W3] The suggested guided-diffusion and biological-sequence diffusion references will be added to Related Work

We thank the reviewer for pointing us to the literature on classifier- and value-guided discrete diffusion, including Nisonoff et al., Schiff et al., and SVDD, and to recent discrete diffusion methods for biological sequences such as D3 and DRAKES. We will add these references to the revised Related Work section.

Comments on the paper

3 comments

1. What the clustering and selection stage actually does

For each cluster GkG_k, the full-sequence efficacy predictor produces a score qaq_a for every complete candidate aa. We compute both its mean score and the fraction of candidates reaching the high-efficacy score level:

\qquad \rho_k=\frac{1}{|G_k|}\sum_{a\in G_k}\mathbf{1}[q_a\ge 0.7].$$ Clusters are assigned to three efficacy grades: **high** if $\mu_k\ge0.7$ and $\rho_k\ge0.5$, **medium** if $\mu_k\ge0.55$ but the high-grade condition is not met, and **low** otherwise. Grades are prioritized from high to medium to low; within a grade, clusters are ranked by $\mu_k$. For the metrics in Table 2: - for $K=1$, we select the highest-scoring candidate from the best cluster in the best available grade; - for $K=3$, we select the two highest-scoring candidates from the best-scoring cluster and one candidate from the second-best cluster. Our clustering aiming at the positional group based on the generated siRNA. And evaluating the groups based on the full-seqeunce predictor. To make fair comparison, we induct the other conditional sequence generators as the baseline to compare, such as transformer, autoencoder, and varitional autoencoder. siDiff outperforms other generative models and ranking methods, benefiting from dynamic representation learning during the diffusion process. Incorporating varying lengths and unmasked base combinations provides richer information than relying solely on a single, fully unmasked 21-nt sequence. | Proposal model | gate0 | gate1 | Mismatch threshold | Taka Hit0@1 | Taka Hit3@3 | Taka EWC | ANGPTL7 Hit0@1 | ANGPTL7 Hit3@3 | ANGPTL7 EWC | CTNNB1 Hit0@1 | CTNNB1 Hit3@3 | CTNNB1 EWC | GSK3A Hit0@1 | GSK3A Hit3@3 | GSK3A EWC | |---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:| | Autoregressive Transformer | 5 | 12 | 5 | 0.0000 | 0.0209 | 0/1 | 0.0000 | 0.0645 | 1/1 | 0.0207 | 0.1116 | 7/29 | 0.0000 | 0.0513 | 0/9 | | Conditional autoencoder | 5 | 12 | 5 | 0.0105 | 0.0628 | 1/1 | 0.0000 | 0.0000 | 0/1 | 0.0124 | 0.0661 | 3/29 | 0.0128 | 0.0385 | 0/9 | | Conditional VAE | 5 | 12 | 5 | 0.0000 | 0.0105 | 0/1 | 0.0000 | 0.0645 | 0/1 | 0.0289 | 0.3058 | 16/29 | 0.0000 | 0.0128 | 0/9 | | siDiff | 5 | 12 | 5 | 0.242 | 0.664 | 1/1 | 0.185 | 0.612 | 1/1 | 0.200 | 0.654 | 28/29 | 0.101 | 0.679 | 9/9 | At the same time, we test two simple clustering methods regardering the case of prediction method such as iScore and Oligofromer. Based on the all postions' prediction score, we select couples anchors as the cluster centers. Based on the method of selection in section above. We obtain the final siRNA condidate. **Cluster A (threshold anchors)** retains every position whose normalized inhibition score is at least 0.7; if none reaches the threshold, the highest-scoring position is retained. **Cluster B (top-B anchors)** always retains the eight highest-scoring positions, making the number of anchors independent of score calibration. In both variants, anchors within three start positions are treated as the same local peak. Each peak is expanded by two positions on both sides, overlapping regions are merged, and the mean score of all enumerated candidates in each resulting region is used to rank the regions. | Predictor | Selection method | Anchor method | Takayuki Hit0@1 | Takayuki Hit3@3 | ANGPTL7 Hit0@1 | ANGPTL7 Hit3@3 | CTNNB1 Hit0@1 | CTNNB1 Hit3@3 | GSK3A Hit0@1 | GSK3A Hit3@3 | |---|---|---|---:|---:|---:|---:|---:|---:|---:|---:| | iScore | Original enumeration (Table 1) | N/A | 0.038 | 0.401 | 0.032 | 0.309 | 0.022 | 0.385 | 0.046 | 0.476 | | iScore | Cluster A | Threshold = 0.7 | 0.057264 | 0.285260 | 0.032258 | 0.380645 | 0.028099 | 0.378512 | 0.064103 | 0.397436 | | iScore | Cluster B | Top-B = 8 | 0.063627 | 0.535525 | 0.019355 | 0.419355 | 0.029752 | 0.480165 | 0.048718 | 0.528205 | | OligoFormer | Original enumeration (Table 1) | N/A | 0.048 | 0.318 | 0.032 | 0.283 | 0.023 | 0.275 | 0.018 | 0.358 | | OligoFormer | Cluster A | Threshold = 0.7 | 0.057264 | 0.314952 | 0.032258 | 0.380645 | 0.033058 | 0.354545 | 0.017949 | 0.379487 | | OligoFormer | Cluster B | Top-B = 8 | 0.058324 | 0.487805 | 0.032258 | 0.535484 | 0.028926 | 0.448760 | 0.017949 | 0.584615 | | siDiff | Original cluster-level selection | N/A | **0.242** | **0.664** | **0.185** | **0.612** | **0.200** | **0.654** | **0.101** | **0.679** | Cluster B improves Hit3@3 over the original Table 1 enumeration baseline for both predictors on all four targets, while the fixed-threshold variant is less consistent. Regional consensus therefore contributes independently to tolerant top-K recovery, but it does not account for siDiff's overall advantage: siDiff remains substantially stronger in Hit0@1 and retains the highest Hit3@3 on every target. This controlled analysis separates the benefit of regional selection from that of target-conditioned guided diffusion.

We additionally inspected coordinate-verified individual sites to determine whether siDiff's aggregate gains translate into recovery of experimentally potent sequences that conventional predictors place near the bottom of a local candidate ranking.

Strict Hit3@3 cases. These examples are placed in the bottom three by the indicated baseline but are recovered within the formal 3-nt Hit3@3 tolerance by siDiff:

Under-ranking baselineTargetGT startExperimental 21-nt mRNA segmentNormalized inhibitionBaseline ranksiDiff top-3 minimum distance
iScoreGSK3A641CACUGCAAUAUUGUGAGGCUG0.94340/413 nt
OligoFormerCTNNB1436ACUCAAGAACAAGUAGCUGAU0.91639/411 nt
siRNADiscoveryGSK3A687AGAAGAAAGACGAGCUUUACC0.96039/412 nt

Thus, three experimentally strong sites, all assigned to the bottom three of their local rankings, are recovered by siDiff's compact top-three portfolio within 1–3 nt. This directly complements the aggregate Hit3@3 results: target-conditioned generation can rescue potent sites that would be excluded by taking the top candidates of a single discriminative predictor.

2. EWC normalization and the regional patterns in Figure 3

Before selecting predicted high-efficacy windows, each method's predicted scores are normalized to get rid of the effact of value scale. EWC is the fraction of ground-truth-positive windows overlapped by predicted-positive windows.

EWC measures recovery of experimentally supported high-efficacy regions, while Figure 3 provides the corresponding spatial profiles and is important for interpreting the aggregate score. The figure shows that siDiff does not label nearly the entire transcript as active.

F1 across methods and test data:

methodANGPTL7CTNNB1GSK3A
OligoFormer0.1820.7060.182
iScore0.0910.9030.696
siDiff0.3330.9030.783

Thus, the high EWC values do not arise from blindly enumerating or marking all regions as positive. They reflect localized predictions that are largely consistent with the experimentally measured regional activity profiles.

We have carefully addressed the reviewers’ comments by conducting additional experiments and revising the manuscript accordingly. All major changes in the revised paper are highlighted in blue for ease of reference. We welcome any further questions or suggestions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.