acceptodds
Under review as a conference paper at ICLR 2027

GIDER: GUIDE-INFORMED DUAL-STAGE MODELING OF CRISPR–CAS9 EDITING RESPONSES IN SINGLE CELLS

Abstract

Predicting single-cell transcriptional responses to specific CRISPR–Cas9 editing designs can support the assessment of intervention outcomes before experimentation. However, target-gene identity alone cannot distinguish different single-guide RNAs (sgRNAs) targeting the same gene under identical cellular backgrounds,limiting design-specific response prediction. Sequence information provides finer-grained conditions, but how to organize sgRNA–site relationships and their corresponding DNA contexts into predictive conditions for unseen editing designs remains underexplored. We therefore introduce GIDER, a two-stage framework inspired by gene-editing mechanisms that models targeting relationships and the corresponding site contexts as interdependent conditioning information. First, GIDER incorporates target-gene priors to encode sgRNA relationships with intended target sites and computationally predicted candidate off-target sites. It then uses each site’s query to read its own long-range DNA context and reuses the first-stage site weights for aggregation. Auxiliary supervision is provided by observed differences in mean expression between perturbed and background-matched control populations in the training data. Conditioned on these representations and control expression, GIDER combines discrete change-state prediction with continuous magnitude generation to learn perturbed expression distributions from unpaired data. Across unseen-sgRNA and unseen-target-gene tasks on Frangieh and Papalexi, GIDER achieves the lowest all-gene EMD, defined as the mean gene-wise Wasserstein-1 distance, in all four settings, with reductions of 6.00%–19.62% relative to the best baseline in each setting. On Papalexi, it leads all six error metrics and the perturbation discrimination score (PDS cosine) in both tasks. Controlled ablations assess the contributions of editing information, conditioning structure, and response supervision. These results suggest that structured editing representations improve prediction of the transcriptional consequences of specific editing designs, providing a computational basis for virtual gene editing to inform experimental design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
CRISPR–Cas9, single-cell perturbation prediction, conditional generative modeling, mechanism-inspired learning

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_gYUCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents GIDER, a novel method for modelling the effect of CRISPR-Cas9 perturbations on single-cell gene expression distributions. The method utilises information from the pairing of the sgRNA to target and off target sites in addition to using information from the DNA sites themselves. Across a few benchmarks, they demonstrate their method outperforms others in the field in addition to a few simple baselines and ablations,

Critical strengths

The method leverages novel information that other approaches lack, explicitly incorporating sgRNA-specific targeting relationships and potential off-target sites to guide the integration of genomic context.

Critical weaknesses

The empirical validation of GIDER is insufficient to conclude it outperforms other methods or simple baselines. Of the two benchmarks reported, GIDER only consistently outperforms other methods on one. Additionally, the benchmarks lack Pearson delta metrics and simple mean-predictor baselines, which Systema (Viñas Torné et al., 2025) shows are important for assessing whether models capture perturbation-specific effects. Also, no error bars or cross validation is performed, so the significance of the improvements cannot be evaluated.

The method is very complex, with some of the design decisions lacking ablations or clear justifications. The model introduces numerous heuristic design choices, including two-stage attention, shared site weights, graph-based gene grouping, and a complex state-conditioned generator. The ablations performed do not isolate these components and so the reason behind their use remains unclear.

Questions for authors

What motivates the construction and reuse of the gene graph for both message passing and fixed gene grouping? Have the authors tested the effect of sgRNA-site information in a simple baseline?

Can the authors report results across multiple random seeds with error bars or confidence intervals, and extend the architectural ablations to Frangieh to demonstrate that the improvements are robust?

Reviewer_TXgaOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GIDER is a two-stage framework for predicting single-cell transcriptional response distributions from a specific CRISPR-Cas9 editing design and control expression. Stage 1 encodes short sgRNA-site features, a target-gene prior from a frozen scGPT GeneEncoder, and ON/OFF roles, producing site queries and relevance weights. Stage 2 uses each query to read only that site's long-range DNA context from a frozen Flashzoi encoder and reuses the Stage 1 weights for cross-site aggregation. Two response heads map targeting and context representations to signed gene-wise population responses, supervised by observed differences between training perturbed and control population means. Training uses entropy-regularized optimal transport within each sgRNA-background condition to construct statistical endpoint pairs from unpaired control and perturbed cells. At inference, states and magnitudes are sampled and combined with control expression to reconstruct nonnegative perturbed expression. Experiments on Frangieh and Papalexi datasets show GIDER achieves the lowest values across all six error metrics plus the highest PDS on Papalexi SG. Ablation experiments support editing information, long DNA context, the two-stage design, and auxiliary RNA supervision are important.

Critical strengths
  1. It looks like GIDER identifies a concrete limitation of gene-level perturbation conditions. The framing is biologically grounded and makes the prediction task more relevant to evaluating alternative editing designs.

  2. GIDER uses observed population mean-expression differences as auxiliary supervision is interesting. Both targeting and context response heads predict complete signed gene-wise responses, and both are supervised by the difference between training perturbed and training control population means under the same background. This avoids requiring paired single-cell before/after measurements, which are usually unavailable.

  3. GIDER achieves great performance with the lowest All EMD and DE20 EMD in all four setting.

  4. The ablation study is focused and informative. It compares the full model with ON-only, removal of long-range DNA context, Joint Fusion, and removal of RNA auxiliary supervision. These variants test the contributions of candidate off-target information, DNA context, the two-stage query-guided design, and population-response supervision.

  5. The appendix is detailed and documents all details needed for reproducing the results. The paper also states that source code will be made public upon publication. This level of detail makes the results easier to assess and reproduce.

Critical weaknesses
  1. The AI novelty in this work is limited and mostly we have an engineering work with results that are not extensive. The evaluation uses only two datasets and only RNA expression is predicted. There is no evaluation on other cell types, CRISPRi/CRISPRa modalities, perturbations outside Cas9, multimodal readouts, or cross-dataset transfer. Candidate off-target sites are computational predictions, not verified editing events. Downstream biomedical applications are acknowledged to require independent experimental validation and ethical review. This limits the generality of the claims.

  2. GIDER relies on frozen scGPT GeneEncoder features for the target-gene prior and frozen Flashzoi features for long-range DNA context. These representations may contain biases or miss target-relevant factors, and the generator never adapts them.

  3. OT coupling endpoints are statistical, not biological trajectories. GIDER obtains state and magnitude supervision by coupling control and perturbed cells within each condition using entropy-regularized optimal transport. The paper explicitly states that sampled endpoints provide state and magnitude targets, not measured single-cell editing trajectories. This is appropriate, but it also means the supervision depends on the coupling’s plausibility.

  4. For magnitude generation, GIDER takes gene-wise absolute values of the context response, aggregates them over 16 fixed gene groups, and uses a learnable group embedding as a global additive condition. This compression discards response signs and within-group gene-specific magnitude patterns.

  5. Evaluation and baseline adaptation raise fairness and reproducibility concerns. DE20 genes are ranked using the observed test perturbed cells and evaluation controls, so the evaluation gene set is test-dependent, even though all methods share it. The paper states this is not a multiple-testing-corrected set of significant DEGs, which is honest, but it still means the metric is not a fixed gene panel.

Questions for authors
  1. How much do computationally predicted off-target sites actually contribute, and how sensitive is GIDER to their quality? relationships or from additional sequence/context features.

  2. Are the OT-derived state and magnitude targets biologically plausible? The OT coupling provides statistical endpoint pairs, not measured trajectories.

  3. Why does GIDER improve EMD but not consistently improve PDS or predictive-mean metrics?

  4. Would adapting or jointly learning the frozen representations improve performance? GIDER freezes scGPT and Flashzoi features. If these representations miss target-relevant factors, no downstream module can recover them. Could a lightweight adaptation improve results without destabilizing the site correspondence?

Reviewer_Nm4POpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces GIDER, a two-stage framework for predicting single-cell transcriptional responses to CRISPR–Cas9 editing designs. Unlike existing methods that mainly rely on target-gene information, GIDER also considers sgRNA sequences, potential off-target sites, and their surrounding DNA contexts. The first stage models the relationships between sgRNAs and their target sites, while the second stage uses these relationships to guide the integration of DNA-context information. The resulting representations are then used in a generative model that predicts discrete expression-change states and their magnitudes, with additional supervision from observed population-level RNA responses. Experiments on the Frangieh and Papalexi datasets show that GIDER performs well on both unseen-sgRNA and unseen-target-gene prediction tasks, achieving the lowest all-gene EMD in all four settings.

Critical strengths
  1. The paper makes a convincing case for going beyond target-gene information, as different sgRNAs targeting the same gene can lead to different transcriptional responses. I find the two-stage design reasonable, particularly the idea of using sgRNA–site relationships to guide how the model incorporates surrounding DNA-context information, rather than simply combining these features.

  2. GIDER achieves the lowest all-gene EMD across all four settings on the Frangieh and Papalexi datasets, covering both unseen-sgRNA and unseen-target-gene prediction. The experiments also include comparisons with several established methods and the simple but competitive SameTarget-shift baseline, making the reported improvements more convincing.

Critical weaknesses

1.The paper presents the two-stage design and long-range DNA context as key contributions, but removing or replacing these components leads to only small changes in All EMD, which is also the paper's main evaluation metric. In contrast, removing RNA auxiliary supervision has a much larger impact on several other metrics. Some ablations also change multiple things at once. For example, removing long-range DNA context also removes the corresponding auxiliary loss, making it difficult to tell which component is actually responsible for the improvement.

  1. One of the main motivations is to distinguish different sgRNAs targeting the same gene, but the unseen-sgRNA evaluation only holds out one sgRNA per target gene. As a result, the current evaluation does not directly test this ability. The parameter-free SameTarget-shift baseline also remains competitive and even outperforms GIDER on some metrics, suggesting that the benefits of guide-specific modeling are not consistent across all evaluation metrics. In addition, Papalexi UG includes only five held-out target genes, which makes the evidence for generalization to unseen genes somewhat limited.
Questions for authors
  1. This is one of the main motivations of the paper, but Appendix F states that each target gene contributes only one test sgRNA in the SG splits. As a result, PDS cosine measures discrimination across different perturbations rather than between sgRNAs targeting the same gene. The SameTarget-shift baseline also performs surprisingly well. On Frangieh SG, GIDER improves All EMD from 0.0721 to 0.0645, but the parameter-free baseline achieves better PDS cosine (0.6041 vs. 0.5839), DE20 RMSE, and DE20 E-distance. This suggests that the advantage of guide-specific modeling is not consistent across metrics. I think an evaluation with multiple held-out sgRNAs targeting the same gene is important to support the paper's central motivation.

  2. In Figure 2, removing or replacing the proposed components only leads to small changes in All EMD, roughly 0.7%–2.0% compared with the full model. This is worth noting because All EMD is also the metric on which the paper reports its most consistent improvements. In particular, the two-stage design and long-range DNA context seem to contribute relatively little to this metric. On the other hand, removing RNA auxiliary supervision has a much larger effect on several other metrics, increasing All RMSE by 13.8% and decreasing PDS cosine by 17.4%. While this supervision is clearly useful, it is not the main architectural novelty of the paper. I am also concerned that some ablations change multiple components at once. For example, w/o Long-DNA also removes the context-response auxiliary loss LCL_C, so the effect of DNA context cannot be separated from that of supervision. Similarly, Joint Fusion changes both query-guided retrieval and gated aggregation. The current experiments therefore do not clearly show which parts of the proposed architecture are actually driving the improvements. More controlled ablations would make the methodological contribution more convincing.

  3. According to Table 4, Papalexi UG contains only 20 training target genes and 5 held-out target genes. This is a relatively small number for evaluating generalization to unseen genes, especially since this setting also shows the largest improvement in All EMD (19.62%). While the result is encouraging, it is unclear whether the improvement would remain consistent with a different selection of held-out genes. Additional gene-level splits would provide stronger evidence, or at least the authors should acknowledge this limitation when discussing the results.

  4. The paper does not report results across multiple random seeds or provide confidence intervals. This matters particularly for the ablation study, where some of the differences are quite small, and for Papalexi UG, where only five target genes are held out. Without some measure of variability, it is difficult to judge how stable these improvements are. Reporting results across multiple seeds or providing appropriate confidence intervals would help establish whether the observed differences are reliable.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.