acceptodds
Under review as a conference paper at ICLR 2027

CellReprogrammer: Phenotype-Guided Therapeutic Target Discovery for Cell Reprogramming

Abstract

Therapeutic target discovery requires interventions that improve disease-associated cellular phenotypes while remaining feasible and limiting disruption to other cellular programs. Existing methods often stop at gene ranking or response prediction, while combination approaches may rely on full-transcriptome matching. To address these limitations, we introduce , a differentiable framework that directly optimizes combinatorial target sets while jointly considering effectiveness, feasibility, and toxicity. Specifically, CellReprogrammer couples a disease-state classifier that defines phenotype improvement with a perturbation surrogate and a differentiable combination optimizer. The optimizer selects a fixed-size target set, incorporates druggability, penalizes high-risk targets, and caps the magnitude of downstream transcriptional changes, without requiring paired disease–treatment profiles from the target cohort. Across five disorders and three cell types, phenotype-guided optimization substantially reduced downstream transcriptional disruption relative to full-transcriptome matching at matched predicted effectiveness. Across the broader benchmark, CellReprogrammer outperformed 20 baselines in predicted reprogramming effectiveness while selecting compact target sets with 93.7% druggability and lower predicted target and downstream toxicity. Finally, the selected interventions were disease- and cell-type-specific and recovered coherent biological programs and established cross-disease relationships. Together, these results establish phenotype-guided combinatorial optimization as a practical approach to therapeutic target discovery.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Single-cell transcriptomics, Disease-state reprogramming, Perturbation modeling, Combinatorial intervention design, Differentiable optimization

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_b27eOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents CellReprogrammer, a framework for selecting small gene intervention sets using a disease classifier, a perturbation surrogate, and differentiable top-K optimization. The objective combines predicted disease-state reversal with drug-realizability and target-risk penalties, while capping downstream expression changes. Experiments across five disorders and three cell types report improvements over 20 baselines. However, the evidence remains computational, and the current evaluation does not sufficiently establish therapeutic relevance or reliable prediction of combinatorial effects.

Critical strengths
  1. The problem is relevant: selecting compact intervention sets while considering effectiveness, feasibility, and risk is a useful research direction.
  2. The framework integrates these objectives coherently and includes several sensible controls, such as donor-level aggregation and perturbation-response differencing.
  3. The experiments cover multiple disease–cell-type settings and include ablations of the objective, surrogate, and optimization constraints.
Critical weaknesses
  1. The evaluation lacks independence. Interventions are optimized and assessed through the same predictive pipeline, leaving open whether the optimizer exploits model errors rather than identifies meaningful disease reversal. No experimental intervention validation is provided.
  2. Surrogate validation does not support the full search space or combination claims. Measured perturbations cover only a small subset of the 3,000 candidates per cell type, and held-out single-target prediction does not establish accuracy for multi-target interventions. The description of weighted response vectors also leaves the modeling of non-additive interactions unclear.
  3. The baseline comparisons are not sufficiently controlled. Training protocols and donor holdout procedures differ across methods, and some baselines use zero-shot predictions or surrogate-generated training data. Matched computational budgets and stronger statistical comparisons are needed.
  4. The presentation significantly limits accessibility. Figure 1 can not be smoothly read.
Questions for authors
  1. Can the authors evaluate selected combinations using an independent predictor?
  2. How accurately does the surrogate predict unseen multi-target interventions, including response magnitudes and non-additive effects?
  3. Do the reported gains persist under consistent donor-held-out protocols, matched training and computational budgets, and an analysis that includes unsuccessful effect-matching cases?
1 comment
AuthorsOpenReview ↗

Independent Evaluation, Multigene Prediction, and Effect Matching

Q1: Can the authors evaluate selected combinations using an independent predictor?

A1: Thank you for this suggestion. We added two complementary evaluations: (i) rescoring selected target sets in the original cohort with four alternative evaluators—RF, MLP, scPRINT2, and CellWhisperer—and (ii) evaluating selected target sets in independent cohorts with five evaluators—LR and the same four alternatives. These comparisons separately examine sensitivity to the disease-scoring model and transfer to new cohorts.

Alternative evaluators. At K = 3, we retained the selected target sets and rescored the predicted perturbed profiles using RF, MLP, scPRINT2, and CellWhisperer without reselecting targets. Table R1 reports mean control-adjusted logit changes; more negative values indicate stronger predicted effects. CellReprogrammer outperformed the strongest baseline under each evaluator (55–78% improvement, Table R1).

Independent cohorts. We evaluated selected targets on external disease-donor profiles (ASD: GSE298243; BPD: GSE306819; MDD: GSE213982; PTSD: GSE313257; SCZ: GSE254569), using the same control-adjusted scoring procedure with LR and the four alternative evaluators. CellReprogrammer outperformed the strongest baseline under each evaluator (57–72% improvement, Table R1).

Table R1. Predicted reprogramming effects across evaluators and cohorts

CohortEvaluatorsOursBest associationBest perturbationBest combinatorial
OriginalRF, MLP, scPRINT2, CW-0.42 to -0.53-0.04 to -0.05-0.19 to -0.27-0.25 to -0.31
ExternalLR, RF, MLP, scPRINT2, CW-0.43 to -0.55-0.03 to -0.05-0.17 to -0.28-0.25 to -0.32

Q2: How accurately does the surrogate predict unseen multi-target interventions, including response magnitudes and non-additive effects?

A2: Thank you for raising this question. We evaluated the surrogate on 26 held-out two-gene perturbations in K562 cells. Joint prediction reduced MSE by 52.4% relative to the additive baseline across all 3,000 genes (0.003102 vs. 0.006510) and increased mean Pearson r from 0.7643 to 0.8458 (Table R2). The joint model also achieved lower MSE and higher Pearson r on the top 20, 50, and 100 response genes. Genes were ranked by absolute measured expression change, with metrics averaged equally across the 26 combinations.

For our three benchmark cell types (excitatory neurons, microglia, and astrocytes), suitable Perturb-seq datasets with matched single-gene and multigene measurements remain unavailable, limiting direct validation in these settings.

Table R2. Prediction accuracy by response gene set

Gene setAdditive MSE ↓Joint MSE ↓Additive Pearson r ↑Joint Pearson r ↑
All 3,000 genes0.0065100.0031020.76430.8458
Top 200.2233210.0912840.89700.9399
Top 500.1495490.0594450.88160.9317
Top 1000.1039600.0421530.86790.9192

Q3: Do the reported gains persist under consistent donor-held-out protocols, matched training and computational budgets, and an analysis that includes unsuccessful effect-matching cases?

A3: Thank you for raising these points. Across all 510 folds, phenotype guidance reached the predicted-effect threshold more often than expression matching (492/510 vs. 322/510; 96.5% vs. 63.1%, a 33.3-percentage-point improvement).

Consistent donor holdout. CellReprogrammer and direct-search baselines use the same training donors for target selection and the same held-out donor for evaluation. PDGrapher uses surrogate-predicted post-treatment profiles for training because paired treatment measurements were unavailable.

Effect-matching coverage. At Δℓadj=−0.1\Delta\ell_{\mathrm{adj}}=-0.1, all 510 folds are included in Table R3. Figure 4A compares non-target expression changes within the 314 folds where both methods reached the same predicted effect.

Table R3. Effect-matching outcomes across all 510 folds

OutcomeFolds
Both methods reached threshold314
Phenotype guidance only178
Expression matching only8
Neither method10
Total510

Across four matched-effect levels (Δℓadj=−0.05,−0.1,−0.2,−0.3\Delta\ell_{\mathrm{adj}}=-0.05,-0.1,-0.2,-0.3), phenotype guidance showed fewer large non-target expression changes within jointly successful folds. The all-fold analysis separately reports how often each effect was attainable.

Computational budgets. On one MDD–EXN donor-held-out fold (3,000 candidates, RTX 3090), CellReprogrammer took 1.04–2.09 s for K = 1–20, versus 0.41–8.38 s for greedy. Random and genetic search took less than 0.4 s, and simulated annealing ~1.1 s. PDGrapher required ~26 min for proxy-data preparation and retraining, reused across K.

Reviewer_6N51OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper picks a small set of genes (three) to turn up or down so that diseased brain cells look healthier. A disease classifier scores how "diseased" a cell profile looks, a fine-tuned scGPT model predicts how gene expression changes after each intervention, and an optimizer selects the genes that lower the disease score, with extra penalties favoring druggable and low-risk targets. Across five psychiatric disorders and three cell types, it reports a stronger predicted effect than 20 baselines and a much higher fraction of druggable targets.

Critical strengths
  • It directly optimizes disease improvement rather than matching the whole healthy profile, and builds druggability and target risk into the selection.
Critical weaknesses
  • The effect is scored by the same classifier and response model that the method optimizes, so the main result is not independently validated.
  • The druggability and risk advantages come from penalties only this method uses; without them, its numbers fall within the baseline range (Fig. 4E).
Questions for authors
  1. Does the gain hold with an independent evaluator, e.g., a different classifier or cohort?
  2. How do the search baselines (e.g., greedy) perform with the same penalties?
1 comment
AuthorsOpenReview ↗

Independent Evaluation and Search Baselines with Matched Penalties

Q1: Does the gain hold with an independent evaluator, e.g., a different classifier or cohort?

A1: Thank you for this suggestion. We added two complementary evaluations: (i) rescoring selected target sets in the original cohort with four alternative evaluators—RF, MLP, scPRINT2, and CellWhisperer—and (ii) evaluating selected target sets in independent cohorts with five evaluators—LR and the same four alternatives. These comparisons separately examine sensitivity to the disease-scoring model and transfer to new cohorts.

Alternative evaluators. At K = 3, we retained the selected target sets and rescored the predicted perturbed profiles using two conventional classifiers (RF and MLP) and two pretrained-model-based evaluators (scPRINT2 and CellWhisperer). This comparison changes the evaluator without reselecting targets. Table R1 reports mean control-adjusted logit changes for CellReprogrammer with penalties and the best competing baseline; more negative values indicate stronger predicted effects. CellReprogrammer outperformed the strongest baseline under each evaluator (55–78% improvement in predicted reprogramming effect, Table R1).

Independent cohorts. We obtained external donor profiles (GSE298243 for ASD, GSE306819 for BPD, GSE213982 for MDD, and GSE313257 for PTSD, and GSE254569 for SCZ). We evaluated the selected target sets on disease-donor profiles from external cohorts using the same control-adjusted scoring procedure. LR was included alongside the four alternatives. CellReprogrammer outperformed the strongest baseline under each evaluator (57–72% improvement in predicted reprogramming effect, Table R1).

Table R1. Predicted reprogramming effects across evaluators and cohorts

CohortEvaluatorOursBest associationBest perturbationBest combinatorial
OriginalRF-0.42clf-w/DisGeNET -0.04linear-rank -0.19random/greedy -0.25
OriginalMLP-0.53clf-w -0.05linear-rank -0.27random/greedy -0.31
OriginalscPRINT2-0.45clf-w -0.04linear-rank -0.23greedy -0.29
OriginalCellWhisperer-0.48DisGeNET -0.04linear-rank -0.22greedy -0.27
ExternalLR-0.47DisGeNET -0.05linear-rank -0.23greedy -0.30
ExternalRF-0.43clf-w/DisGeNET -0.03linear-rank -0.17random -0.25
ExternalMLP-0.55clf-w/DisGeNET -0.03linear-rank -0.28random -0.32
ExternalscPRINT2-0.47clf-w/DisGeNET -0.04linear-rank -0.24random/greedy -0.28
ExternalCellWhisperer-0.48DisGeNET -0.05linear-rank -0.24greedy -0.29

Q2: How do the search baselines (e.g., greedy) perform with the same penalties?

A2: Thank you for this suggestion. Under matched penalties, CellReprogrammer achieved the strongest mean predicted effect, highest mean druggability, and lowest mean target risk among the five methods.

We applied the same druggability and target-risk penalties used by CellReprogrammer to four combinatorial search baselines (greedy, random search, simulated annealing, and genetic search) at K = 3. The comparison covered 510 donor-held-out folds across all 15 disease–cell-type settings, with results averaged across settings.

All four baselines achieved higher druggability and lower target risk after adding the penalties. Greedy, the strongest baseline in predicted effect under matched penalties, increased druggability from 19.7% to 92.2% and reduced the target-risk score from 35.95 to 5.28. Its predicted effect weakened, with the adjusted logit change shifting from -0.338 to -0.307 (more negative indicates a stronger predicted effect).

Under the same penalties, CellReprogrammer achieved 93.7% druggability, a target-risk score of 2.36, and an adjusted logit change of -0.473.

Reviewer_LFDMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper presents CellReprogrammer, a phenotype-guided framework for combinatorial therapeutic target discovery. The framework combines a disease-state classifier, a perturbation surrogate based on fine-tuned single-cell foundation models, and a differentiable top-K optimizer to identify gene combinations that reduce predicted disease-state scores while considering druggability and toxicity. The authors evaluate the method across five psychiatric disorders and three cell types, comparing it with 20 baselines. The reported results show improved predicted reprogramming effectiveness, higher druggability, and lower toxicity-related scores.

Critical strengths

The paper addresses an important problem in therapeutic target discovery. Optimizing combinations of targets rather than ranking individual genes is a meaningful direction, particularly for complex diseases involving multiple interacting biological pathways.

The phenotype-guided optimization is well motivated. Rather than requiring the entire disease transcriptome to match the healthy state, the framework focuses on reducing disease-associated signals while limiting unnecessary transcriptional changes. Incorporating druggability and toxicity constraints also makes the formulation more relevant to therapeutic applications.

The experiments cover multiple diseases and cell types, with extensive baseline comparisons and ablation studies. The manuscript is generally well organized and easy to follow.

Critical weaknesses
  1. The effectiveness evaluation relies on the same disease classifier and perturbation surrogate used for target optimization. Although leave-one-donor-out evaluation is performed, it does not provide independent evidence that the selected interventions improve disease-associated cellular states. Moreover, the disease classifiers achieve relatively modest AUROCs of 0.62-0.69 (Table 1), raising concerns about the reliability of the optimization objective.

  2. The perturbation surrogate is primarily trained and validated on single-gene perturbations, but the proposed method optimizes multigene combinations. The accuracy of predicting combinatorial effects, particularly non-additive genetic interactions, has not been demonstrated. In addition, the reported correlations are calculated using only the 20 genes with the largest observed responses.

  3. The candidate pool contains 3,000 genes per cell type, but only 284, 76, and 62 genes have measured perturbation responses for EXN, MG, and AST, respectively. Most candidate interventions therefore rely on predictions for unmeasured perturbations. It is unclear how reliably the surrogate generalizes to these targets and how prediction errors affect target selection.

  4. The fairness of the baseline comparisons needs clarification. The combinatorial search methods use limited candidate-evaluation budgets, while CellReprogrammer uses gradient-based optimization across multiple seeds. The computational costs are not compared. Furthermore, PDGrapher is adapted using surrogate-generated responses rather than measured disease-treatment profiles, which may affect its performance.

  5. The comparison between phenotype guidance and expression matching in Figure 4A uses the same classifier-based effectiveness metric optimized by the proposed method. Moreover, only 314 of 510 folds achieve the matched-effect threshold. It is unclear whether the conclusions hold across all folds or under independent measures of phenotype improvement.

Questions for authors
  1. Can the authors provide an independent evaluation of the predicted reprogramming effectiveness?

  2. How accurately does the surrogate predict multigene perturbations, especially non-additive effects?

  3. Can the authors compare the computational costs and optimization budgets of the combinatorial methods?

1 comment
AuthorsOpenReview ↗

Independent Evaluation, Multigene Prediction, and Computational Costs

Q1: Can the authors provide an independent evaluation of the predicted reprogramming effectiveness?

A1: Thank you for this suggestion. We added two complementary evaluations: (i) rescoring selected target sets in the original cohort with four alternative evaluators—RF, MLP, scPRINT2, and CellWhisperer—and (ii) evaluating selected target sets in independent cohorts with five evaluators—LR and the same four alternatives. These comparisons separately examine sensitivity to the disease-scoring model and transfer to new cohorts.

Alternative evaluators. At K = 3, we retained the selected target sets and rescored the predicted perturbed profiles using two conventional classifiers (RF and MLP) and two pretrained-model-based evaluators (scPRINT2 and CellWhisperer). This comparison changes the evaluator without reselecting targets. Table R1 reports mean control-adjusted logit changes for CellReprogrammer with penalties and the best competing baseline; more negative values indicate stronger predicted effects. CellReprogrammer outperformed the strongest baseline under each evaluator (55–78% improvement in predicted reprogramming effect, Table R1).

Independent cohorts. We obtained external donor profiles (GSE298243 for ASD, GSE306819 for BPD, GSE213982 for MDD, GSE313257 for PTSD, and GSE254569 for SCZ). We evaluated the selected target sets on disease-donor profiles from external cohorts using the same control-adjusted scoring procedure. LR was included alongside the four alternatives. CellReprogrammer outperformed the strongest baseline under each evaluator (57–72% improvement in predicted reprogramming effect, Table R1).

Table R1. Predicted reprogramming effects across evaluators and cohorts

CohortEvaluatorsOursBest associationBest perturbationBest combinatorial
OriginalRF, MLP, scPRINT2, CellWhisperer-0.42 to -0.53-0.04 to -0.05-0.19 to -0.27-0.25 to -0.31
ExternalLR, RF, MLP, scPRINT2, CellWhisperer-0.43 to -0.55-0.03 to -0.05-0.17 to -0.28-0.25 to -0.32

Q2: How accurately does the surrogate predict multigene perturbations, especially non-additive effects?

A2: Thank you for raising this question. We evaluated the surrogate on 26 held-out two-gene perturbations in K562 cells. Joint prediction reduced MSE by 52.4% versus the additive baseline across 3,000 genes (0.003102 vs. 0.006510) and improved mean Pearson r from 0.7643 to 0.8458 (Table R2). Improvements also held for the top 20, 50, and 100 response genes, ranked by absolute measured expression change. Metrics were averaged equally across combinations.

For the three cell types in our main benchmark (excitatory neurons, microglia, and astrocytes), we have not identified suitable Perturb-seq datasets with matched single-gene and multigene perturbation measurements. As such datasets become available, they will enable direct evaluation of predicted response and non-additive effects in these cell types.

Table R2. Prediction accuracy by response gene set

Gene setAdditive MSE ↓Joint MSE ↓Additive Mean Pearson r ↑Joint Mean Pearson r ↑
All 3,000 genes0.0065100.0031020.76430.8458
Top 200.2233210.0912840.89700.9399
Top 500.1495490.0594450.88160.9317
Top 1000.1039600.0421530.86790.9192

Q3: Can the authors compare the computational costs and optimization budgets of the combinatorial methods?

A3: Thank you for this suggestion.

Runtime. We added a timing experiment on one MDD–EXN donor-held-out fold with 3,000 candidate genes on an RTX 3090. CellReprogrammer and the direct-search baselines all completed target selection within a few seconds. As K increased from 1 to 20, CellReprogrammer’s runtime rose from 1.04 to 2.09 s, while greedy’s rose from 0.41 to 8.38 s. Random and genetic search each took less than 0.4 s, and simulated annealing took approximately 1.1 s. PDGrapher required about 26 min for proxy-data preparation and retraining, and the resulting model was reused across all values of K.

Computational complexity. For C candidates and K targets, greedy performs O(KC) set evaluations. CellReprogrammer uses 150 gradient updates, each with O(KC) selection overhead. With the update count fixed, both workload measures scale as O(KC) in K and C. Exhaustive search evaluates (CK)\binom{C}{K} sets. At the measured pairwise throughput of approximately 7,509 sets/s, enumerating all triples in the timing experiment is projected to take about 6 days. Selecting 20 targets from 3,000 candidates yields approximately 1.35×10511.35 \times 10^{51} sets, corresponding to roughly 5.7×10395.7 \times 10^{39} years at the same throughput. Even with substantially larger computational budgets, exhaustive enumeration is impractical at large K.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.