REVIEW 3 major objections 5 minor
CellPrism shows that a visual analytics system can turn thousands of predicted gene perturbation strategies into a single wet-lab-ready candidate.
Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →
T0 review · deepseek-v4-flash
2026-08-04 23:12 UTC pith:XX6JHLKQ
load-bearing objection CellPrism is a genuine design contribution for exploring virtual cell perturbation spaces, but its effectiveness claim leans on a single co-designed case study and unvalidated upstream model predictions. the 3 major comments →
CellPrism: A Visual Analytics System for Exploring AI-Driven Virtual Cells in Drug Discovery
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
CellPrism's central claim is that a visual analytics system designed around clustering-based overviews, a radar-based glyph encoding, and coordinated comparison views can make AI-driven virtual cell models actionable in drug discovery. The system organizes predicted perturbation outcomes into clusters, uses radar glyphs to compactly show expression differences across cell types and the SHAP-based contribution of each perturbed gene, and supports progressive screening from global patterns to individual comparisons. The paper demonstrates this through a case study in which an expert narrows from a large perturbation space to a five-candidate collection and finally selects the STAT3+PCNA combin
What carries the argument
The core object is the radar-based glyph, which encodes a single perturbation as a polygon whose axes are cell types and whose vertices are the differential expression levels of selected observed genes; superimposed arcs show SHAP values, the contribution of each perturbed gene to expression changes. This glyph is embedded in a global-to-local workflow: a cluster projection plot and cluster summary table give a coarse overview of the perturbation space, the glyph supports local inspection after semantic zoom, and the Expression Comparison and Collection Views support detailed side-by-side comparison and final trade-off analysis. The SHAP-based contribution layer is used not just as an explan
Load-bearing premise
The entire exploration rests on the accuracy of the virtual cell model's predicted gene expression changes and of the SHAP attributions derived from them; if those predictions are wrong for the cell types studied, the selected STAT3+PCNA strategy would not produce the intended effects in wet-lab validation.
What would settle it
Run the STAT3+PCNA perturbation in the cell lines studied (A375, HepG2, K562, and iPSC) and compare the measured differential expression to PerturbNet's predictions; if the predicted suppression of IER3 and activation of JUN do not appear, or the iPSC side-effect profile differs, the pruning logic that selected this strategy is not reliable. Alternatively, on held-out perturbation experiments from the same datasets, benchmark PerturbNet's top-K predicted strategies against random or SHAP-dissimilar strategies for matching observed outcomes.
If this is right
- Experts can reduce a combinatorial perturbation space of thousands of strategies to a small set of wet-lab candidates in a single interactive session, cutting the cost of experimental screening.
- The same workflow can be retargeted to other therapeutic goals by replacing the selected observed genes, e.g., anti-proliferative or differentiation-inducing screens.
- Feature-contribution (SHAP) explanations become an operational part of hypothesis pruning, not just a post-hoc explanation, which can inform the design of other model-driven visual analytics systems.
- Virtual cell models become more useful in practice because their high-dimensional outputs are organized and interpreted through a decision-oriented interface.
Where Pith is reading between the lines
- The system's trustworthiness is only as strong as the underlying perturbation model's accuracy; if PerturbNet's predictions are biased for these cell lines, the STAT3+PCNA recommendation would fail in wet-lab validation, so a natural next step is to couple the visualization with explicit uncertainty displays.
- The glyph's encoding of many-to-many relationships (multiple perturbed genes, observed genes, cell types) could transfer to other combinatorial intervention problems, such as drug combinations, CRISPR screens, or multi-tissue toxicity predictions.
- Because SHAP contributions are used to decide which gene to remove, the system implicitly assumes that model attributions reflect causal effects; this is a testable assumption and could be checked by validating the pruned perturbations in vitro.
- The expert selection process (choosing gene subsets, cluster thresholds, and filter criteria) embeds subjective priors into the final recommendation; documenting those choices explicitly would make the system's output more reproducible and easier to audit.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. CellPrism is a visual analytics system for exploring gene perturbation strategies generated by AI-based virtual cell models, aimed at drug-discovery tasks such as identifying anti-tumor perturbations while protecting normal cells. The system combines a candidate-gene filtering view, a cluster navigation view with radar-based glyphs, an expression comparison view, and a collection view, all driven by predicted differential expression and SHAP-based contribution attributions from the PerturbNet model. The authors derive five design requirements from a six-month co-design process with three domain experts, then evaluate the system through a single case study performed by one of the co-designers (E1) and semi-structured interviews with three independent experts. The central claim is that CellPrism effectively supports expert-driven narrowing of a large perturbation space to a promising candidate (STAT3+PCNA) for wet-lab validation.
Significance. If the effectiveness claim is substantiated, CellPrism addresses a real and timely gap: virtual cell models produce high-dimensional, combinatorial perturbation outputs that are difficult for biologists to interpret and compare. The hierarchical exploration design, the radar glyph that compactly encodes multi-gene, multi-cell relationships, and the use of SHAP contributions as an interactive retrieval and pruning mechanism are all well-motivated and potentially valuable for the visual analytics community. The design requirements are grounded in an extended co-design process with domain experts, and the system is demonstrated on real single-cell perturbation data across multiple cell lines. The paper also explicitly acknowledges the important limitation that the system cannot suggest perturbations outside the user-specified space. However, the current evaluation evidence is qualitative and partly confounded: the central case study is conducted by a co-designer, and the interviews are reported mostly as positive quotes. The paper would be stronger as a system/design contribution than as a demonstrated effectiveness study.
major comments (3)
- [§6 and §7 (evaluation of effectiveness)] The abstract and Section 1 claim that CellPrism's effectiveness is demonstrated through a real-world case study and expert interviews. In §6, however, the case study is performed by E1, who co-designed the system over six months (§4). This creates a confirmation-bias risk and does not constitute an independent demonstration. Moreover, §7 reports only qualitative impressions and quotes, with no task-based metrics, no baseline comparison against an existing workflow or tool, and no systematic coding/thematic results. For a system whose contribution is about human decision-making, this does not support the 'effectiveness' claim beyond usability anecdotes. I recommend adding an independent analyst case study and/or a structured user study with quantitative measures such as task completion time, accuracy of identifying desired perturbations, coverage of the perturbation space, and comparison
- [§6 and §8 (upstream model accuracy)] The entire analytical narrative—from candidate-gene filtering to the final STAT3+PCNA recommendation—depends on PerturbNet's predicted differential expression values and the SHAP attributions derived from them. Section 6 calls PerturbNet a 'high-precision virtual cell simulation' and §8 states that 'our models achieve high accuracy in the current study,' yet no accuracy metrics, held-out validation, or ground-truth comparison are reported for the A375, HepG2, K562, or iPSC contexts used here. If the underlying predictions are inaccurate for these cell types or for the specific multi-gene combinations considered, the pruning of VIM from STAT3_VIM_PCNA is an artifact of model error rather than biological insight. The authors should either report relevant validation metrics or explicitly reframe the case study as a demonstration of interactive exploration of model outputs, not as biological
- [§5.2 and Fig. 7 (SHAP contribution computation)] The SHAP-based contribution layer is load-bearing for the main result: it is what leads E1 to remove VIM from STAT3_VIM_PCNA. Yet the manuscript does not specify how SHAP values are computed for PerturbNet's outputs, what reference/background distribution is used, how the values are aggregated across cells, or whether the attributions were validated against any biological or model-based ground truth. SHAP values are not model-independent; they depend on the chosen explainer, baseline, and approximation algorithm. Without this information, the central analytical step cannot be reproduced or assessed. Please add a methods subsection describing the SHAP computation and any sensitivity checks.
minor comments (5)
- [§6] The dataset footnotes for scPerturb and Zenodo are empty. Please provide URLs or accession identifiers so that the data sources are reproducible.
- [Throughout] Notation is inconsistent: 'Cell-A375' appears as 'Cell-A375' in some places, and gene names are written variously as 'STAT3', 'Gene-STAT3', and 'STAT3_PCNA'. Please standardize the notation.
- [Fig. 2] The figure uses callouts A1, A2, B1–B4, but the caption does not explain what these stages are. Adding a one-sentence description of the data pipeline and workflow would improve readability.
- [§7] The text says interview data were analyzed through thematic analysis [6], but no themes, codes, or inter-rater procedure are reported. Either provide the thematic results or describe the analysis as qualitative feedback rather than formal thematic analysis.
- [§8] The phrase 'our models achieve high accuracy in the current study' is inaccurate because the models were developed in prior work, not in this paper. Please rephrase to refer to the upstream model and to caveat the absence of context-specific validation.
Circularity Check
No circularity: CellPrism is a visual analytics system whose displayed quantities are external model outputs (PerturbNet predictions and SHAP attributions), not parameters fitted to the system's own conclusions.
full rationale
CellPrism does not derive any claimed result from a quantity defined in terms of that result. The data pipeline (Section 2, Fig. 2-A1/A2) takes a baseline expression matrix and perturbation labels, passes them to the virtual-cell model PerturbNet [50] to obtain predicted differential expression ΔX, and separately computes SHAP attributions. These values drive the cluster projection, radar glyphs, heatmaps, and comparison views; they are not estimated from the visualization or from the expert's final choice. The case study (Section 6) is a narrative of E1 using the system to prune STAT3_VIM_PCNA to STAT3_PCNA based on SHAP contribution arcs; the pruning is an expert decision made from displayed attributions, not a fitted parameter renamed as a prediction. The evaluation (Section 7) uses three experts not involved in the design, and the case-study expert is disclosed to be a co-designer, which is an evaluation limitation rather than a circular derivation. The accuracy-related assertions—'This high-precision virtual cell simulation provides a reliable data foundation' (Section 6) and 'Although our models achieve high accuracy in the current study' (Section 8, Limitation)—are unsupported by reported metrics, but unsupported premises are not circularity: the paper's outputs are not defined in terms of those premises, and PerturbNet is an external model whose published benchmarks are not re-established for A375, HepG2, K562, or iPSC. That is a correctness/evidence risk, not a circularity. Self-citations ([37] TrialCompass, [38] CellScout, [43] TrajLens) appear only in related-work comparisons and do not carry the load of the effectiveness claim. No uniqueness theorem, ansatz, or imported formulation is invoked to forbid alternatives. The system's central claim—that the visualization workflow supports expert exploration and pruning—is demonstrated through the interface design and qualitative evaluation; it does not reduce to a fitted input or to a self-citation chain.
Axiom & Free-Parameter Ledger
axioms (4)
- domain assumption PerturbNet predictions of differential gene expression are accurate enough to support drug discovery decisions.
- domain assumption SHAP values faithfully attribute expression differences to individual perturbed genes.
- domain assumption The three experts' design requirements and positive feedback are representative of domain needs.
- domain assumption The public datasets used (scPerturb, Zenodo) were preprocessed and integrated correctly.
Cite this review
Pith. "Pith review of CellPrism: A Visual Analytics System for Exploring AI-Driven Virtual Cells in Drug Discovery." pith.science (2026). https://pith.science/paper/XX6JHLKQ
@misc{pith2026260801669,
author = {Pith},
title = {Pith review of: CellPrism: A Visual Analytics System for Exploring AI-Driven Virtual Cells in Drug Discovery},
year = {2026},
howpublished = {\url{https://pith.science/paper/XX6JHLKQ}},
note = {Machine review of arXiv:2608.01669}
}
read the original abstract
Gene perturbation analysis plays a critical role in drug discovery by enabling researchers to investigate how interventions on specific genes influence global gene expression patterns within cells. Recent advances in artificial intelligence-driven virtual cell models have made it possible to predict gene expression outcomes for a wide range of perturbation strategies in silico, substantially reducing reliance on costly and time-consuming biological experiments. However, effectively exploring and interpreting the high-dimensional perturbation spaces produced by these models remains challenging because of the combinatorial nature of perturbations and the complex cell-specific gene expression responses they generate. In this work, we present CellPrism, a visual analytics system designed to support the systematic exploration of gene perturbation strategies for drug discovery. Specifically, CellPrism integrates clustering-based overviews to summarize perturbation outcomes, a glyph-based representation to compactly encode gene expression patterns across cell types, and coordinated views that enable fine-grained comparison and interpretation of perturbation effects. We demonstrate the effectiveness of CellPrism through a real-world case study and expert interviews. This work highlights the value of visual analytics in bridging virtual cell modeling with expert-driven decision making in drug discovery.
Figures
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.