acceptodds
Under review as a conference paper at ICLR 2027

GenInv: Generative Inverse Learning for Cell Perturbation Prediction

Abstract

Predicting how drugs alter gene expression across cellular contexts can guide the selection of treatments for experimental screening. Large single-cell perturbation screens provide a basis for learning these responses, but measure separate control and treated populations without observing individual cells before and after treatment. Recent models address this constraint by learning transitions between populations. Their forward objective directly supervises the treated distribution, yet does not require the predictor to recover the untreated context represented in those observations. This suggests a complementary learning signal: predicting controls from observed treated cells. We introduce GenInv, a generative inverse learning approach that jointly supervises forward response prediction and reverse control prediction through a shared network. Both tasks use the same experimental measurements, and inference retains the original forward procedure. On five evaluation cell lines in State-Tahoe-Filtered, GenInv improves all four primary response metrics over a matched forward-only model. Mean response correlation (Pearson-) increases from 0.922 to 0.931, and differential-expression log-fold-change Spearman correlation increases from 0.612 to 0.665. Label interventions show that reverse prediction retains an expression-level benefit under shuffled inverse labels, while correct conditioning further improves differential-expression recovery. These results establish generative inverse supervision as an effective training principle in this setting, extending the predictive value of existing perturbation measurements. Code will be available at https://anonymous.4open.science/r/geninv/.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
single-cell perturbation prediction, chemical perturbation response, dual learning, gene expression modeling

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_qWr8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper adds an aux task to the State transition model: besides predicting treated cells from controls plus a drug label, the same set transformer also predicts controls from treated cells with a separate linear label encoder and the same energy-MMD loss at weight 1. Evaluation uses the State few-shot split of State-Tahoe-Filtered in 2000-HVG space: training covers 45 cell lines plus part of the conditions in five evaluation lines, and the rest are held out (Sec 4.1). One GenInv run is compared with one forward-only run ("State-ST") at 200k steps. With shuffled inverse labels (100k steps) the Pearson-delta gain stays, but roughly half of the direction and LFC gains are lost (Tab. 2).

Critical strengths
  • The idea could hardly be simpler! And the comparison is like for like: one extra pass through the same network, no change at inference (Sec 3.3), same architecture, split, optimizer and step budget (L245). The listed settings also match Arc's released HVG few-shot config, so State-ST isn't a strawman. If the gain is real, anybody training a State-style model gets it in a few lines which would be cool.
  • The authors test their own story and report the result even where it hurts, e.g. the shuffled-label control keeps the whole Pearson-delta gain(L370). They also state single seed (L247), DE overcalling (L430).
Critical weaknesses
  • My main concern is noise. The contribution is one loss term, and it's backed by one run per arm, all seed 42 (L247). The only scale for variation I can find is State-ST's own drift, where Pearson-delta drops from 0.9275 at 100k to 0.9220 at 200k (L313). That isn't seed variance but it's the size of the gains (0.0047, 0.0092) and the gap widens only bc State-ST falls faster than GenInv. The recall gain comes with more positive calls (Sec 4.5). To be fair, LFC Spearman looks more solid: +0.036 at 100k, +0.052 at 200k in all five lines both times, and GenInv itself improves too.
  • The controls test the label but not the direction. Shuffled labels give the same Pearson-delta as correct ones (0.9323 vs 0.9322, L370), so as the authors say that benefit isn't treatment-specific. They mention a "regularizing effect" (L317), which a second pass in any direction might provide on a recipe without weight decay (L249, which I believe comes from the State release?). There's no State-ST run with standard regularization or with a treated-to-treated pass, which keeps the second pass and treated inputs and removes only the reverse direction, and Sec 4.6 isn't a matched ablation (L462). The label-specific part (LFC +0.020, L372) is again one run against one.

Overall a simple, honestly reported idea that may well work, but "establish" (L028), even "in this setting" (one backbone, one dataset), is more than one seed per arm and no control for direction support. I'm close to the border on this one tbh.

Questions for authors
  1. (experiment) What's the across-seed spread of Pearson-delta and DE LFC Spearman for both models? 100k steps would do. If the LFC gain clearly exceeds it over three seeds, I'd raise my score.
  2. (clarification) What does Tab. 1 look like with best-validation checkpoints for both models, as in the State release?
  3. (experiment) Does State-ST with weight decay, or with a treated-to-treated pass (disjoint samples), close the LFC gap? If not, that would move me up.
  4. (experiment) Tab. 3's baselines never see the five evaluation lines (Sec 4.5). How do your linear models score when refit on the few-shot data, or State's context-mean baseline? Pearson-delta against controls rewards shifts shared across perturbations (Viñas Torné et al., 2025). The label swap (L362) shows the label matters, but not how much of 0.92 is shared shift!
1 comment

Thank you for proposing concrete tests of the gains in our tables.


Response to Q1:

We agree that the gain needs an across-seed comparison. At 100,000 steps, GenInv improved log-fold-change (LFC) correlation in each seed pair. Even the smallest gain exceeded either model's across-seed SD.

We resumed State-ST and GenInv from their 50,000-step checkpoints for seeds 43, 44 and 45, retaining optimizer state and batch size 64. We scored all six final checkpoints under the submission protocol, weighting five test lines equally. Values are seed means ± SD.

ModelPearson-deltaLFC Spearman
State-ST.92704 ± .00121.61157 ± .00624
GenInv.93063 ± .00144.63621 ± .00056

We will report the paired LFC gains of +.02848, +.02776 and +.01767, averaging +.02464. With validation selection, mean LFC and Pearson gains were +.01945 and +.00348. For both metrics at both endpoints, the exact two-sided sign-flip test gave p=.25, its minimum with three pairs.


Response to Q2:

Your question about State's checkpoint rule prompted us to recompute Table 1. When both models select the lowest forward-validation loss, GenInv scores higher on all four metrics.

Candidates included the best checkpoints retained from the 10,000- and 50,000-step stages and checkpoints every 2,000 steps from 52,000 to 100,000. In seed order, State-ST selected 96k/98k/92k and GenInv 72k/92k/92k. The table gives three-seed test means ± SD.

ModelPearson-deltaDirection agreementLFC SpearmanSignificant-gene recall
State-ST.92881 ± .00099.90080 ± .00214.61689 ± .00575.71874 ± .00277
GenInv.93228 ± .00061.90906 ± .00196.63634 ± .00550.72855 ± .00435

This will be our main table, with final checkpoints reported as a sensitivity check. Scoring the selected checkpoints with cell_eval2 also favored GenInv in Delta Pearson and average precision, although State-ST had higher LFC (.16901 versus .16561). In a separate 50,000-step analysis, switching from predicted to observed DMSO controls reversed the LFC ranking with samples, averaging and significant genes held fixed.


Response to Q3:

Thank you for suggesting these two alternatives. Neither closed the LFC gap. GenInv had higher LFC than both controls with validation selection and at the final checkpoint.

We trained both variants to 100,000 steps. One used AdamW with weight decay 5e-4 instead of zero. The other added a pass predicting one treated population from disjoint cells under the same treatment. Multi-seed values are means ± SD.

ModelValidation-selected LFCFinal LFC
State-ST.61689 ± .00575.61157 ± .00624
State-ST with AdamW.61816.61816
State-ST with treated-to-treated training.58009 ± .00663.58239 ± .00338
GenInv.63634 ± .00550.63621 ± .00056

The final gaps to GenInv's three-seed mean were .01805 for AdamW and .05382 for treated-to-treated training. We will add both controls to the ablation table and document their different training histories and the checkpoints available for selection.


Response to Q4:

You are right that access to target-line data matters. All three Table 3 baselines improved after refitting, and both neural models still led in Pearson-delta and LFC. A separate context-mean comparison also showed that a shared shift alone fell short of State-ST's correlation.

For Additive, Ridge and Diagonal we scored the original fit and two adjustments (nine settings). Held-out treatments had no target-line training data, so their original coefficients stayed fixed. Using 348 training conditions per target line, we fitted gene-wise offsets, or scales and offsets, shared across treatments. We excluded validation and test treatments, applied scaling and offsets to population means before clipping, and retained the original DMSO predictions and test references. Entries give five-line mean Pearson-delta / LFC Spearman on the same test conditions.

ModelWithout target-line adjustmentWith target-line offsetsWith target-line scaling and offsets
Additive.42858 / .25828.55468 / .34698.53842 / .48537
Ridge.53697 / .36995.62807 / .41095.70732 / .54104
Diagonal.61496 / .43292.65673 / .45721.77218 / .53975

Diagonal had the best linear Pearson-delta (.77218) and Ridge the best LFC (.54104). With common validation selection, State-ST reached .92881/.61689 and GenInv .93228/.63634 across three seeds.

The context mean predicts the target line's mean treated training expression for every perturbation. It scored .41345 in Pearson-delta versus State-ST's .87324 in a matched seed-43, 10,000-step comparison. We will add the refits and context mean to the appendix and clarify that neither this comparison nor the label swap assigns an additive fraction of .92 to the shared shift.

Reviewer_zBFROpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Authors propose GeneInv wwhich adds reverse pass to state perturbation model. the perturbed cells with conditioning labels are mapped to control cells with energy MMD loss using a shared network and individual inverse encoder. the claimed novel contribution is bidirectional supervision from control-perturbed pairs, with evaluation on 5 held out cell lines and label shuffling analysis.

Critical strengths

Adding one small encoder is simple and cheap in terms of parameters. The label shuffling mechanism is sound. Held out results are strong and biologically sound method of evaluation.

Critical weaknesses

I believe the gains arent separated from noise well enough on 1 seed. Authors do mention it in limitations so I believe they will run multi seed experiments during rebuttal period along with statistical tests to show significance of gains.

mechanism is also not isolated enough with ablations, as reverse pass doesnt use labels, so the benefit might be from regularization rather than pure inverse supervision. There is no theoretical grounding of the mechanism. There should be a label free, control to control, ablations as baseline controls.

Finally i believe the novelty is weak for a conference like ICLR. bidirectional control-perturbed learning for unpaired sc data is already explored (cycleCDR Predicting single-cell cellular responses to perturbations using cycle consistency learning ) which is not discussed in related works. for an application paper i think the novelty aspect is fine though as long as the code is visible (currently broken link).

Also the state paper is old and many new sota have shown improvement over it like cellflow, perturbdiff, scDFM. Please test against those backbones too. This will give the readers confidence that the method proposed works universally in any backbone.

Questions for authors

Can you report multiple seeds with confidence intervals with significance of gains using statistical tests? please also report metrics from new cell-eval 2 package.

Does the gain hold with a label-free reverse pass, control-to-control and treated-to-treated auxiliaries, and a cell-line-identity auxiliary? this would be an important baseline and mechanism understanding experiment.

How does a cycle consistency objective compare on the same backbone? just to show that GeneInv outperforms cycleCDR method.

How does the label-swap probe show a forward pearson delta well below the main tables? please explain or justify.

important note for the authors: the github anonymous link is broken so theres no way to see if the code matches the paper. if this is fixed during rebuttal phase and the code matches the paper's method i will raise the score by 1 point.

overall i think this is a good paper for application track. If all my concerns (apart from anonymous code link) are addressed or justified i will raise the score by additional 2 points.

Authors can get a clear accept from me if all the above points agree with the paper's claim.

2 comments

Thank you for the thoughtful review.


Response to Q1:

We agree that gains should recur across training runs. Under the submission protocol, GenInv improved Pearson-delta and log-fold-change (LFC) Spearman in every seed at both validation-selected and final checkpoints.

We continued State-ST and GenInv to 100,000 steps with three seeds and weighted five lines equally. Selecting by minimum forward-validation loss gave paired gains of +.01945 in LFC (95% paired-t interval [-.00710,.04600]) and +.00348 in Pearson-delta ([-.00049,.00744]). Final LFC gained +.02464. With three pairs, even consistent gains yield p=.25 in the exact two-sided sign-flip test, above the 5% threshold.

With cell_eval2 v0.21.0, GenInv led in Delta Pearson, AUROC and average precision, and State-ST led in LFC. We will add the intervals and this table (validation-selected mean ± SD).

ModelDelta PearsonLFC SpearmanAUROCAverage precision
State-ST.91134 ± .00053.16901 ± .00499.64911 ± .00070.37374 ± .00137
GenInv.91777 ± .00164.16561 ± .01237.65293 ± .00487.38041 ± .00659

Expression metrics use all test cells. DE metrics use common conditions with at least two observed cells.


Response to Q2:

Your controls helped distinguish the value of the reverse pass from that of the label. Label-free inverse still surpassed State-ST. Control-to-control exceeded GenInv in LFC but had lower average precision.

With the inverse treatment vector zeroed, final seed-43 Pearson-delta/LFC/average precision at 100,000 steps were .9321/.6478/.6635 for label-free inverse, .9261/.6081/.6578 for State-ST and .9313/.6366/.6673 for GenInv.

Control-to-control and treated-to-treated training used disjoint input and target cells. At 50,000 steps (two seeds, common validation selection), control-to-control scored .6409 LFC/.6321 average precision versus GenInv's .6330/.6520. At 100,000 steps, treated-to-treated final LFC was .5824 ± .0034 versus .63609 for GenInv on the same two seeds. Training histories differed.

In the preliminary 10,000-step test with three seeds, State-ST predicted cell-line identity from mean-pooled Transformer features of treated cells. Under the submission protocol, final five-line means for Pearson-delta/LFC/average precision were .8260/.4488/.5407 versus GenInv's .8756/.5239/.5882. GenInv led on all three metrics in every seed. We will tabulate each auxiliary with its budget and seeds.


Response to Q3:

Your same-backbone test isolates the training objective. GenInv led cycle consistency on all three metrics and cycleCDR on Pearson and LFC.

On State, both objectives used the same data, two seeds, batch size 64, 50,000 updates and minimum forward-validation loss selection. GenInv matches observed controls directly. The cycle objective composes predictions to reconstruct both populations. Five-line submission means:

Objective on StatePearson-deltaLFC SpearmanAverage precision
GenInv.9255.6330.6520
Cycle consistency.8125.4307.5032

GenInv used two backbone passes per update versus five for our cycle implementation.

We also trained four models with three seeds on a 24-drug Tahoe subset. Five target lines supplied only controls during training and validation. Validation-selected means:

ModelPearson-deltaLFC SpearmanAverage precision
State-ST.4276.3851.1575
GenInv.3726.3728.1653
scPreGAN-AC.2885.2192.2058
cycleCDR.2046.1784.2503

Test cells, conditions, observed DMSO and metrics were shared. State-ST/GenInv used 50,000 updates and validation pseudobulk-MSE selection. External methods kept their own schedules and selection, and led in average precision. State-ST led GenInv in Pearson and LFC. We will report the two comparisons separately.


Response to Q4:

We agree the raw scores invite comparison. The probe uses a different reference and averaging rule. On its non-DMSO conditions, main scores still averaged .910 on C32 and .928 on PANC-1.

The probe averages correlations over 40 batches centered on input controls. The main evaluation pools cells by treatment, centering predictions and observations on their respective line-wide DMSO means. Excluding DMSO raised the correct-label probe score only to about .840.

Swapping labels within the probe lowered correlation from .834 to .382 on C32 and .831 to .335 on PANC-1. We will revise the caption to distinguish this label-sensitivity test from pooled performance.


Response to Q5:

Thank you for flagging this. We fixed the permissions and verified anonymous access to the code. The source uses a separate inverse treatment encoder, matches observed controls directly and disables cycle loss by default. The README lists scripts and setup steps.

Reviewer_zBFROpenReview ↗

Acknowledgement

I thank the authors for the thorough rebuttal, but the new results weaken the paper’s central claims.

The headline lfc gain shrinks under 3 seed validation. Observed vs predicted DMSE lfc advantage reverses and STATE leads lfc.

label free inverse and control to control match or exceed GeneInv so neither treatment label nor treated inputs are required.

all these contradict the conclusion that predicting controls from treated observation drives the gain.

same backbone cycle falls below STATE. Eval on newer backbones I mentioned are still missing.

I appreciate the effort and raise the score by 1 but I request the reviewers to reconsider and revise the method and evaluations before resubmission next time.

Reviewer_kr6rOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

GenInv augments a State transition predictor with an auxiliary task that predicts control-cell populations from observed treated populations. Forward and inverse passes share the cell projector, transformer and output head, but use separate treatment encoders. Both directions are trained with energy-MMD against observed target populations, with unit inverse-loss weight and no cycle-consistency term. Inference retains only the forward pass.

The evaluation uses one seed per configuration on State-Tahoe-Filtered, holding out cell-line–treatment combinations within five cell lines that are also partially represented in training. Against a matched forward-only State-ST model at 200k steps, mean Pearson-Δ increases from 0.9220 to 0.9312 and DE LFC Spearman from 0.6121 to 0.6646. All four primary metrics improve in each evaluation line. At 100k steps, shuffled inverse labels retain the Pearson-Δ benefit, while correct labels further improve DE direction and LFC agreement. The reported gains are restricted to this setting; the evaluation does not establish a substantive methodological advance or a robust practical advantage.

Critical strengths
  1. Straightforward implementation and unchanged forward inference. GenInv reuses existing control and treated measurements without requiring cell-level correspondence. The architecture and objective are clearly described, and the inverse branch is unnecessary at inference.

  2. Improved reported DE effect prediction. All four primary metrics improve in each evaluation line at 200k steps, with DE LFC Spearman increasing by 0.0525. This is the strongest empirical result. However, the single-seed protocol limits the strength of the evidence.

Critical weaknesses
  1. The methodological contribution is incremental, and its significance is not demonstrated. GenInv retains the State backbone and its energy-MMD loss, adds a second loss with the observed source and target populations exchanged, and couples the two tasks through shared parameters and separate treatment encoders. I do not find a sufficiently substantive methodological advance in this construction as currently justified. There is a directly relevant domain precedent that the paper omits: scPreGAN generates both control-to-treated and treated-to-control expression using a shared encoder and separate generators. Its published-version implementation includes an adversarial loss on treated-to-control predictions, evaluated by a discriminator trained on real controls, and both directions update the shared encoder. Thus, reverse generation constrained by an observed control distribution and coupled through shared representations is already present in single-cell perturbation modeling. Dual Supervised Learning, which the manuscript does cite, also establishes the broader bidirectional-learning motivation. A discussion centered on how GenInv differs from cycle consistency does not adequately address this closer precedent. GenInv is not identical to scPreGAN: it shares more of the prediction network, uses direction-specific treatment embeddings and direct energy-MMD, and excludes cycle consistency. However, these are implementation differences whose independent value is not established. The main ablation only compares the complete auxiliary task against its absence. It cannot show whether extensive parameter sharing, separate encoders or the particular inverse objective is necessary, nor whether simpler reconstruction/denoising auxiliaries would achieve the same benefit. The exploratory cycle experiments are not matched final-checkpoint comparisons. Consequently, the paper demonstrates an incremental training modification without establishing why that modification constitutes a significant methodological contribution. The mechanism is also unresolved. At 100k steps, shuffled inverse labels yield Pearson-Δ of 0.9323 versus 0.9322 with correct labels. Correct labels improve DE effects, so this does not imply that conditioning is entirely irrelevant. It does show that the expression-correlation benefit does not require correct inverse conditioning. Recovering a familiar context-specific control distribution remains a plausible explanation. The paper does not adequately distinguish a treatment-dependent inverse-learning benefit from generic auxiliary regularization. The maturity of the task is not itself grounds for rejection; the issue is the limited demonstrated advance within that task.

  2. The experimental evidence is insufficient to compensate for the limited methodological novelty. Each configuration uses only one training seed. The gains in Pearson-Δ (+0.0092), DE AUROC (+0.0072) and average precision (+0.0071) are modest and reported without training-variability estimates. The larger DE LFC gain merits recognition, but results across five cell lines from the same trained models are not independent training replications. The manuscript therefore does not establish that the improvements are reliably reproducible. This is an unresolved uncertainty, not proof that randomness caused the gains. The comparison also matches steps rather than compute, although GenInv adds an inverse training pass. It therefore does not establish superior performance for the same training resources. Checkpoint dependence further weakens the comparison: between 100k and 200k steps, State-ST Pearson-Δ falls from 0.9275 to 0.9220, whereas GenInv falls from 0.9322 to 0.9312. GenInv already leads at 100k, but the larger late gap partly reflects greater baseline deterioration. A common validation-based selection rule and a compute-matched baseline are needed before interpreting the final-checkpoint margin as a practical advantage. Finally, the evidence covers only one backbone, one benchmark and a few-shot split in which the evaluation cell lines are partly represented in training. It does not establish transfer to unseen cellular contexts or broader applicability. Linear and VAE baselines exclude those five lines during fitting, so their results cannot support a controlled cross-model ranking. Gene discovery remains mixed: predicted significant-gene counts rise from 567.3 to 597.0 against 425.1 observed, while overlap/precision do not consistently improve. Taken together, the narrow evaluation, missing variance estimates, unequal compute and unresolved mechanism leave the central contribution insufficiently supported for acceptance.

Questions for authors
  1. Novelty clarification. What advances beyond scPreGAN's bidirectional generation, control-distribution constraint and shared encoder? Which of full backbone sharing, direction-specific encoders and direct energy-MMD is essential?

  2. Robustness evidence. Can the State-ST/GenInv comparison be repeated across independent training seeds, reporting paired differences and variability? This could materially affect my assessment, especially for the smaller gains. Condition-level resampling alone cannot resolve training-seed uncertainty.

  3. Fairness of comparison. Do gains persist with a common validation-based checkpoint-selection rule and equal training compute? Please quantify the additional cost rather than relying on parameter or step counts.

  4. Mechanism and scope. Do fixed-label inverse prediction or reconstruction/denoising auxiliaries recover the benefit under matched budgets? Does evidence on a second backbone or held-out cellular context support broader usefulness? These are targeted tests of the central contribution, not a request for an exhaustive benchmark.

1 comment

Thank you for helping us test what inverse training contributes.


Response to Q1:

We agree that bidirectional prediction alone does not distinguish GenInv from scPreGAN. scPreGAN shares an expression encoder and fits the treated and control distributions with separate generators. In GenInv, an inverse pass from observed treated cells to observed controls trains the same expression projector, Transformer and output head used for forward prediction. It uses State's energy-MMD loss, and inference requires only the forward pass.

Our ablation directly supports separate treatment encoders. We trained GenInv with either a shared treatment encoder or separate forward and inverse encoders, keeping the backbone and loss fixed. After 10,000 steps, separate encoders raised final-checkpoint mean Pearson-delta from .8541 to .8757 and log-fold-change (LFC) Spearman from .5197 to .5338 across two seeds under the submission protocol. In the Methods we will distinguish this result from the rationale for full sharing and energy-MMD, whose individual effects this experiment leaves open.


Response to Q2:

You are right that resampling test conditions leaves training variability unresolved. Across three independent seed pairs, GenInv improved both metrics under validation selection and at the final checkpoint.

We continued State-ST and GenInv from 50,000 to 100,000 steps for seeds 43, 44 and 45, retaining optimizer state and batch size 64. For each run we averaged the submission metrics equally over five cell lines. The table reports validation-selected means ± SD across seeds.

ModelPearson-deltaLFC Spearman
State-ST.92881 ± .00099.61689 ± .00575
GenInv.93228 ± .00061.63634 ± .00550

In seed order, paired LFC gains were +.00726, +.02721 and +.02388 (mean +.01945). Pearson gains were +.00164, +.00434 and +.00445 (mean +.00348). With three pairs, even this agreement gives p=.25 in the exact two-sided sign-flip test. We will include the paired gains for each seed beside the main results.

Under cell_eval2, GenInv led in Delta Pearson and differential-expression average precision, while State-ST led in validation-selected LFC (.16901 versus .16561). In a separate 50,000-step diagnostic, switching from predicted to observed DMSO controls reversed the LFC ranking with samples, genes and averaging held fixed.


Response to Q3:

Thank you for distinguishing checkpoint choice from cost. A common validation rule preserved both gains. At approximately matched wall time, the short comparison favored GenInv in average precision and State-ST in Pearson-delta and LFC.

Both models kept earlier best checkpoints and were validated every 2,000 continuation steps. Selecting the lowest forward-validation loss gave GenInv gains of +.01945 in LFC and +.00348 in Pearson-delta, versus +.02464 and +.00360 at the final checkpoints.

Across three seeds, 100,000 steps took 20.70 ± .72 hours for State-ST and 22.83 ± .22 for GenInv on the shared A100 host, including data waits and validation. GenInv uses an extra pass per update and took 10.3% longer. Mean peak allocated memory was 12.05 versus 22.17 GiB. Within the time for 10,000 GenInv steps, State-ST reached about 12,000. Validation-selected average precision averaged .5717 for State-ST and .5882 for GenInv across three seeds. We will report wall time and memory alongside the step budgets.


Response to Q4:

Your controls ask whether a simpler training target can recover the same benefit. Fixed inverse labels and denoising recovered less LFC than GenInv at 10,000 steps. Predicting controls from other controls did recover the LFC gain at 50,000 steps, though GenInv retained higher average precision.

Each auxiliary used the same backbone, batch size 64, updates and passes per update as GenInv at its budget. At 10,000-step final checkpoints across two seeds, LFC was .5338 for GenInv, .5253 with a fixed DMSO inverse label and .5078 when predicting an unmasked treated population from disjoint treated cells with masked genes. State-ST scored .5426. At 50,000 steps (two seeds, common validation selection), control-to-control and GenInv scored .6409 versus .6330 in LFC and .6321 versus .6520 in average precision.

For the transfer test, we trained four models with three seeds each on a 24-drug Tahoe subset. Five target lines supplied controls but no treated cells during training or validation. GenInv led scPreGAN-AC and cycleCDR in Pearson and LFC, while both led in average precision. State-ST led GenInv in Pearson and LFC. Test conditions were shared. State-ST and GenInv used 50,000 steps and validation pseudobulk-MSE selection, while external methods kept their own budgets and selection rules. The appendix will report these controls and describe the scope as transfer to lines with observed controls. A second backbone and wholly unseen lines remain open.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to All Reviewers: Additional Experiments and Key Findings

Thank you for suggesting these experiments. GenInv improved Pearson-delta and LFC Spearman in all three seed pairs under the submission protocol.

Replication. We continued State-ST and GenInv from 50k to 100k steps with seeds 43, 44 and 45 and batch size 64. We averaged five test lines equally per run. Values are means ± SD across seeds. Selected checkpoints minimize forward-validation loss.

Submission metricState-STGenInv
Pearson-delta, selected.92881±.00099.93228±.00061
Direction agreement, selected.90080±.00214.90906±.00196
LFC Spearman, selected.61689±.00575.63634±.00550
Significant-gene recall, selected.71874±.00277.72855±.00435
Pearson-delta, final.92704±.00121.93063±.00144
LFC Spearman, final.61157±.00624.63621±.00056

With validation selection, paired gains averaged +.01945 for LFC (95% paired-t CI [-.00710,.04600]) and +.00348 for Pearson-delta ([-.00049,.00744]). Final LFC gains were +.02848, +.02776 and +.01767 (mean +.02464). The exact two-sided sign-flip test gave p=.25, the minimum with three pairs.

On the shared A100 host, 100k steps took 20.70±.72 h for State-ST and 22.83±.22 h for GenInv (+10.3%, including data waits and validation). Mean peak allocated memory was 12.05 and 22.17 GiB. At similar wall time (State-ST ~12k steps, GenInv 10k), selected average precision (AP) was .5717 and .5882. State-ST led in the two correlations.

cell_eval2 v0.21.0, selectedDelta PearsonLFC SpearmanAUROCAP
State-ST.91134±.00053.16901±.00499.64911±.00070.37374±.00137
GenInv.91777±.00164.16561±.01237.65293±.00487.38041±.00659

Expression metrics use all test cells. DE metrics use common conditions with at least two observed cells. In a separate 50k diagnostic, switching from predicted to observed DMSO changed GenInv's LFC advantage from +.02536 to -.00958 on fixed samples and genes.

Auxiliary objectives on the State backbone. P denotes Pearson-delta. Rows share the last stated budget, seeds and checkpoint. The identity test is preliminary.

Objective (steps, seeds, checkpoint)PLFCAP
State-ST (10k, 43,44, final).8739.5426.5595
GenInv, separate treatment encoders.8757.5338.5869
Tied treatment encoders.8541.5197.5655
Fixed DMSO inverse label.8519.5253.5732
Shuffled inverse label.8659.5220.5777
Masked treated-cell denoising.8477.5078.5523
Control-to-control.8768.5949.5488
State-ST (10k, 43–45, final).8718.5379.5579
GenInv.8756.5239.5882
Cell-line identity.8260.4488.5407
GenInv (50k, 43,44, selected).9255.6330.6520
Control-to-control.9277.6409.6321
Cycle consistency.8125.4307.5032
State-ST (100k, 43, final).9261.6081.6578
GenInv.9313.6366.6673
Label-free inverse.9321.6478.6635

At 50k, GenInv used two backbone passes per update versus five for cycle consistency. The 100k controls had different training histories. GenInv's final LFC was .63609 on the two treated-to-treated seeds. In a separate 10k screen with disjoint forward and inverse perturbations (seed 43), batch-level inverse quotas of 0/25/50/100% gave final LFC .2772/.3065/.2996/.2894.

State-ST controlSeedsSelected LFCFinal LFC
AdamW, weight decay 5e-41.61816.61816
Treated-to-treated, disjoint cells2.58009±.00663.58239±.00338

Transfer. We trained four models on a 24-drug Tahoe subset with three seeds each. Five target lines supplied only controls during training/validation. Test cells and observed DMSO were shared. State-ST and GenInv used 50k steps and validation pseudobulk MSE. External methods used native budgets and selection. GenInv led both in P and LFC, while they led in AP. At the subset's final 50k checkpoints, forward and inverse gradients on GenInv's shared weights aligned positively on all 12 sampled training batches.

Selected modelPLFCAP
State-ST.4276.3851.1575
GenInv.3726.3728.1653
scPreGAN-AC.2885.2192.2058
cycleCDR.2046.1784.2503

Few-shot linear baselines. We fitted gene-wise offsets or scales and offsets on 348 distinct non-DMSO conditions per target line, excluding the union of validation and test. Held-out treatment coefficients stayed fixed. Entries give P/LFC.

ModelOriginalOffsetScale + offset
Additive.42858/.25828.55468/.34698.53842/.48537
Ridge.53697/.36995.62807/.41095.70732/.54104
Diagonal.61496/.43292.65673/.45721.77218/.53975

At 10k (seed 43), the context mean (one treated training mean per line) scored .41345 versus State-ST's .87324 in P. In the batch-centered probe, swapping labels reduced P from .834 to .382 on C32 and .831 to .335 on PANC-1. Neither result decomposes the pooled score near .92.

We will use common validation selection in Table 1, add the controls to the appendix and correct Table 11 for duplicate test entries and two labels shared with validation, listing DMSO separately.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.