{"id":"a38464e5-f2ed-4b42-a9b4-d61eabbe99aa","arxiv_id":"2412.19228","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A cross-transfer autoencoder learns drug perturbation vectors that are swapped between cell profiles, improving reconstruction of held-out drug responses at single-cell level.","lead":"XTransferCDR is a machine learning model that separates cells' baseline gene activity from the changes drugs cause, then swaps those drug effects between profiles to learn reusable drug signatures. It reports higher accuracy than existing methods on held-out drugs, but the evaluation feeds the drug's own measured response into the model, so it is not a true zero-shot prediction.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-time protocol feeds the held-out drug's actual perturbed profile into the encoder, turning claimed zero-shot prediction into target reconstruction; reported gains do not demonstrate transferable prediction.","rationale":"The paper proposes a disentanglement-based cross-transfer framework and claims state-of-the-art performance for predicting single-cell transcriptional responses to novel perturbations. The reader's verdict was CONDITIONAL, with the linear-additivity assumption as the weakest point. I identify a more fundamental and more load-bearing flaw: the evaluation protocol violates the zero-shot requirement of the central claim. The test-time input includes the actual perturbed profile of the held-out drug, so the model is reconstructing the target, not predicting it. This affects the main drug and single-gene genetic experiments; the combinatorial experiment is a legitimate extrapolation but only tests additivity. A simple re-run with the target profile withheld would determine whether the reported gains persist. Because the primary evidence for the headline claim is invalidated, the manuscript should be rejected as-is, pending a proper zero-shot evaluation.","tokens_in":11300,"tokens_out":5692,"duration_ms":54611,"concrete_test":"Re-run the sci-Plex3 held-out-drug evaluation without feeding the test drug's perturbed expression profile into Ep. At inference, provide only the unperturbed expression profile and a drug representation derived from molecular structure (or from a training-only drug embedding), as in ChemCPA. If the held-out-drug R2 (all genes) drops from 0.81 toward the 0.50 baseline, the reported improvement is an artifact of target leakage.","verdict_should_be":"REJECT","load_bearing_attack":"Section 'Experimental Setup' states: 'In the testing stage, the perturbed expression profile by a perturbation “unseen” in training stage was fed into the encoder to extract its perturbation representation, which is then combined with the basal state derived from corresponding unperturbed expression profiles.' Thus, for every test cell, the model receives the actual perturbed expression profile (the prediction target) as input to Ep, and the decoder reconstructs D(S + P), a function of the target itself. This is a transductive reconstruction task, not a zero-shot prediction of an unseen drug's response. The baselines in Tables 1–3 use only unperturbed profiles, so they are denied access to the target while XTransferCDR is given it. Consequently, the large R2/PCC/EV improvements over baseline and over chemCPA/cycleCDR (which do not receive the test drug's profile) primarily reflect the model's ability to compress and reproduce the input, not the transferability of the learned perturbation representation. The central claim in the abstract—'better performance than current state-of-the-art methods' for transferable drug perturbations—is therefore unsupported by the drug and single-gene genetic experiments. Only the combinatorial experiment (predicting dual-gene profiles from single-gene profiles) is a genuine extrapolation, and it depends entirely on the unvalidated linear-additivity assumption.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes XTransferCDR, a generative framework that disentangles perturbed single-cell expression profiles into a basal-state representation and a perturbation representation via two encoders, then enforces a linear additivity constraint in the latent space and a cross-transfer reconstruction loss to learn transferable perturbation representations. The authors evaluate on sci-Plex drug-response data, single-gene Perturb-seq data, and combinatorial perturbation data, reporting improved R2, EV, and PCC over several baselines, and conclude that their method advances phenotypic drug discovery by predicting responses to novel drugs and perturbations.","tokens_in":11512,"tokens_out":6148,"duration_ms":61316,"significance":"If the claimed ability to predict transcriptional responses to unseen drugs were real, the method would be a useful tool for drug screening and for interpreting perturbation effects. The combinatorial perturbation experiment (Table 4) is a genuine extrapolation, and the cross-domain disentanglement idea is a reasonable contribution to the representation-learning toolbox. The paper also ships a public code repository, which aids reproducibility. However, the central evaluation protocol for the drug and single-gene experiments feeds the target perturbed expression profile into the model at test time, making those experiments circular and invalidating the headline claims of transferable prediction. The net significance is therefore substantially below what the abstract and introduction assert.","major_comments":[{"comment":"The test-time protocol is circular for the drug and single-gene experiments. The paper states: “In the testing stage, the perturbed expression profile by a perturbation ‘unseen’ in training stage was fed into the encoder to extract its perturbation representation, which is then combined with the basal state derived from corresponding unperturbed expression profiles.” This means the model receives the actual perturbed profile (the prediction target) as input to Ep, and the decoder reconstructs a function of that target. The comparisons in Tables 1–3 against chemCPA, GEARS, and cycleCDR are therefore not apples-to-apples: those baselines do not receive the test drug’s perturbed profile and are being asked to predict it, while XTransferCDR is doing transductive reconstruction. The high R2/PCC/EV values are largely a measure of the autoencoder’s ability to compress and reproduce its input, not of transferable prediction. This undermines the abstract’s central claim that the model “achieved better performance than current state-of-the-art methods” for transferable drug perturbations.","section":"Experimental Setup"},{"comment":"The architecture cannot, as described, produce a perturbation representation for a truly unseen drug without already having that drug’s perturbed expression profile. The only path to P(a) is Ep(X(a)), i.e., the encoder requires the perturbed expression profile itself. For a novel drug, by definition, no such profile is available. Thus the method is not a predictor of cellular responses to novel drugs; it is a profile autoencoder/reconstructor. The claim in “Experimental Setup” that the strategy evaluates “the model’s ability to generalize to novel drugs” is contradicted by the protocol. To claim zero-shot drug prediction, the paper would need a separate drug representation source (e.g., a molecular-structure encoder like ChemCPA’s) or a method to derive P for an unseen drug from training drugs; neither is present.","section":"XTransferCDR / Representation Disentanglement"},{"comment":"Table 4 is presented as a robust comparison with GEARS (“these dual-gene perturbations were also chosen for testing in the GEARS study”), but the table only reports Baseline and XTransferCDR metrics. No GEARS results are shown or discussed beyond that sentence. Moreover, this experiment is the only non-circular evaluation, yet it relies entirely on the unvalidated linear-additivity assumption: the model predicts S + P(a) + P(b) for a dual-gene perturbation from single-gene perturbation vectors. The cross-transfer loss in Eq. (5) is built on the same additivity assumption, so the strong Table 4 R2 values are partially an in-sample fit to that assumption. The paper does not provide direct evidence that dual-gene transcriptional effects are additive in latent space (e.g., by comparing predicted dual effects against a non-additive baseline or by quantifying the error of additivity in held-out combinations).","section":"Evaluation on Combinatorial Genetic Perturbations"},{"comment":"The ablation study in Table 5 uses the same circular test protocol as Tables 1–3. Since the model receives the target profile at test time, the observed improvements from adding Lcross and the other losses could reflect better reconstruction of the given target rather than better transfer of perturbation representations. The ablations therefore do not provide independent support for the claim that the cross-transfer constraint promotes generalizable representations.","section":"Model Ablations"}],"minor_comments":[{"comment":"The notation in Eq. (1) is unclear: P_i^(a) and S_i^(a) are presumably vectors, so their dot product is a scalar, and the squared Frobenius norm reduces to an absolute square. The authors should rewrite this as ‖(P_i^(a))^T S_i^(a)‖^2 or clearly define the operation.","section":"Representation Disentanglement, Eq. (1)"},{"comment":"The text says the basal states “should be aligned to each other as much as possible” but then writes “we try to minimize the similarity between them.” Since the objective is KL divergence, minimizing it makes the distributions more similar. The sentence should say “minimize the divergence” or “maximize the similarity.”","section":"Representation Disentanglement, Eq. (2)"},{"comment":"The description of the Basline model is ambiguous: “calculates performance metrics directly using the unperturbed and actual perturbed expression profiles.” If the baseline is simply predicting the unperturbed profile for every perturbation, this should be stated explicitly. If it is using the actual perturbed profile as its own prediction, the R2 values in Tables 1–3 would be nonsensical.","section":"Experimental Setup"},{"comment":"The paper does not report error bars or significance tests for any of the tables. Since the datasets are split once at random, all reported numbers are point estimates from a single split; without multiple seeds or confidence intervals, it is hard to assess whether the gaps between methods are meaningful.","section":"Experimental Setup"},{"comment":"The phrase “to our knowledge, this is the first time that cross-domain disentanglement representation learning has been proposed to model perturbation-induced cellular responses” is overclaimed. CycleCDR, which is cited, already uses cycle-consistency for cross-domain perturbation transfer; the authors should temper this claim and position their contribution more precisely relative to cycleCDR.","section":"Introduction"}],"recommendation":"reject","confidential_remarks":"The test-time protocol is not a minor oversight; it is the core of the empirical evaluation. Because the model receives the actual perturbed profile for every held-out drug, the claimed superiority over methods that predict without that profile does not demonstrate transferable prediction. The architecture lacks any mechanism to obtain a perturbation representation for a truly novel drug. The combinatorial experiment is interesting, but it is presented with an incomplete comparison and depends on an assumption that is not independently validated. I would suggest the authors consider reframing the work as a cross-domain representation-learning method for reconstruction and combinatorial extrapolation, and then resubmitting with a fully rewritten evaluation protocol. As it stands, the central claims are not supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is a clean adaptation of cross-domain disentanglement from image style transfer to single-cell perturbation response prediction. The model is well-specified, the code is promised, and the combination of basal-state and perturbation encoders with a cross-transfer reconstruction loss is new for this problem. The ablation study gives real evidence that the cross-transfer loss helps, and the combinatorial perturbation experiment is a genuine extrapolation: the model predicts dual-gene profiles from single-gene profiles, which is the right kind of test for transferable representations. That part deserves credit.\n\nThe soft spot is load-bearing. The drug and single-gene experiments do not test prediction of unseen perturbations. The test protocol explicitly feeds the perturbed expression profile of the held-out drug into the perturbation encoder, then reconstructs from S + P. That is target reconstruction, not zero-shot prediction. The baselines (baseline, chemCPA, cycleCDR) do not receive the target profile, so the large metric gains in Tables 1–3 largely reflect the autoencoder's compression ability, not transferable representation quality. The paper's own text confirms this, so it is not a hidden artifact; the authors just mislabel it as generalization to novel drugs. This flaw undermines the abstract's claim of outperforming state-of-the-art for transferable drug perturbations.\n\nSecondary issues: no error bars or repeated runs, no comparison to a fair transductive baseline, and the linear additivity assumption is asserted rather than validated. The novelty relative to cycleCDR is incremental, though the cross-transfer formulation is a distinct contribution.\n\nWho should read this: researchers working on representation learning for perturbation prediction might find the cross-transfer mechanism useful for building better extrapolating models. But it should not be cited as evidence that zero-shot drug response prediction works.\n\nMy recommendation: send it to peer review, because the method is reproducible and the combinatorial experiment is meaningful, but the authors will need to redo the drug and single-gene evaluation under a protocol that does not leak the target, and temper the claims accordingly.","headline":"A sensible extension of cycleCDR with a cross-transfer loss, but the drug-response evaluation is transductive, not zero-shot, and the headline overclaims.","tokens_in":611,"tokens_out":742,"would_cite":false,"duration_ms":23019,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Cross-domain swapping lets a model predict single-cell responses to drugs and gene knockouts it has never seen.","keywords":["single-cell transcriptomics","drug response prediction","cross-domain transfer learning","disentangled representation learning","linear additive latent space","perturbation representation","combinatorial genetic perturbations","phenotypic drug discovery"],"falsifier":"Take a set of gene pairs with documented epistatic or non-additive interactions, train XTransferCDR on single-gene knockouts only, and compare its predicted double-knockout profiles $S + P^{(a)} + P^{(b)}$ to actual measured profiles; if DEG-level R2 collapses specifically on the epistatic pairs while staying high on additive ones, the linear-additivity assumption is falsified.","tokens_in":1821,"feed_emoji":"💊","tokens_out":2589,"duration_ms":77450,"temperature":0.7,"pith_summary":"The paper claims that single-cell transcriptional responses to drugs and genetic perturbations can be predicted for unseen perturbations by learning a perturbation representation that transfers across cellular contexts. XTransferCDR encodes each perturbed expression profile into a basal cell state and a perturbation vector, then forces the model to reconstruct the paired expression profile after the two perturbation vectors are swapped. This cross-transfer constraint is what makes the perturbation representations transferable. On three benchmark families, including sci-Plex drug screens, Perturb-seq single-gene knockouts, and combinatorial genetic perturbations, the reported R2, explained variance, and Pearson correlation beat the comparison methods, with the largest gains on differentially expressed genes. If the claim holds, drug screening could query many more drugs and drug combinations in silico before committing to wet-lab experiments.","feed_headline":"Swapped latent vectors predict unseen drug and knockout responses","feed_subtitle":"By treating perturbations as additive vectors, the model forecasts held-out drug and gene-combination effects from single-cell data.","key_machinery":"The load-bearing mechanism is the cross-transfer constraint encoded in the loss $L_{cross}$: the decoder must turn the sum of perturbation $P^{(a)}$ and basal state $S^{(b)}$ into expression profile $X^{(a)}$, and vice versa. The loss is the mean squared error between the cross-transferred reconstructions and the original perturbed profiles. This is supported by an orthogonality loss $L_{orth}$ that separates perturbation from basal information, a similarity loss $L_{sim}$ that aligns basal states extracted from two perturbations, and two reconstruction losses; the total objective is $L = L_{sim} + L_{orth} + L^{(1)}_{reco} + L^{(2)}_{reco} + L_{cross}$. The linear-additivity rule, that the effect of a perturbation on the latent cellular state is a vector addition to the basal state, is what justifies both the cross-transfer reconstruction and the combinatorial prediction $S + P^{(a)} + P^{(b)}$.","core_discovery":"XTransferCDR's central discovery is that a cross-transfer reconstruction loss is sufficient to learn perturbation representations that remain valid when moved to a different basal state. Given expression profiles $X^{(a)}$ and $X^{(b)}$ from two perturbations applied to the same cell context, two encoders extract basal states $S^{(a)}$, $S^{(b)}$ and perturbation vectors $P^{(a)}$, $P^{(b)}$. The model reconstructs the original profiles from $S+P$ and, crucially, also reconstructs $X^{(a)}$ from $S^{(b)}+P^{(a)}$ and $X^{(b)}$ from $S^{(a)}+P^{(b)}$, with a shared decoder. Under the stated assumption that perturbations add linearly to the latent cellular state, this forces $P$ to encode only the perturbation, not the context; after training, a novel drug's response is predicted by adding its extracted $P$ to the basal state of a target cell population and decoding the sum. The paper evaluates this on held-out drugs and gene knockouts, and also predicts dual-gene knockout profiles by summing two single-gene perturbation vectors, with R2 on differentially expressed genes rising from 0.166 for the baseline to 0.837.","pith_inferences":["A testable extension the authors do not run: systematically scan gene pairs with documented epistatic or non-additive effects and compare $S + P^{(a)} + P^{(b)}$ predictions to measured double-perturbation profiles, which would show where the linear-additivity representation breaks.","Because the model predicts only from expression profiles, it should also propose perturbation vectors for entirely unseen cell types by swapping in that cell type's basal state, a direct extension to patient-derived or tissue-specific contexts.","The same framework could be pointed at drug combination synergy: treating two drugs' vectors as additive would allow ranking candidate combinations by the distance between predicted and desired phenotype before wet-lab screening.","Interpreting $P$ vectors as drug directions lets researchers cluster drugs by mechanism of action without pathway knowledge, and could flag off-target effects when a drug's vector points along an unexpected biological axis."],"forward_implications":["A model trained on existing single-cell drug screens can predict expression profiles for held-out drugs, since the test protocol separated drugs by identity between training and test.","The same learned perturbation vectors predict single-gene knockout responses in K562 and RPE-1 cells, with DEG-level R2 of 0.68 and 0.78 respectively, roughly double the best comparison method.","Summing two single-gene perturbation vectors predicts dual-gene knockout profiles: on the ten held-out combinatorial perturbations, DEG-level R2 reaches 0.837 versus 0.166 for the baseline.","Ablations show the cross-transfer loss is the main driver: removing it drops all-gene R2 on sci-Plex3 from 0.817 to 0.660 and DEG-level R2 from 0.626 to 0.183.","Because perturbation vectors live in a shared latent space, the model offers an interpretable arithmetic: similar drugs should have similar $P$ vectors, which can be compared directly."],"supporting_citations":[{"why":"Supplies the sci-Plex3 and sci-Plex4 single-cell drug response datasets used for the main drug experiments.","marker":"Srivatsan et al. 2020"},{"why":"Provides the genome-scale Perturb-seq single-gene knockout datasets in K562 and RPE-1 cells used for evaluation.","marker":"Replogle, Saunders, and et al. 2021"},{"why":"Supplies the A549 combinatorial dual-gene knockout dataset used to test additive prediction of combined perturbations.","marker":"Norman et al. 2019"},{"why":"Proposes the linear additive latent-space modeling of perturbations that this paper adopts.","marker":"Lotfollahi et al. 2021"},{"why":"Provides the ChemCPA baseline and the related assumption of linear additivity in latent space.","marker":"Hetzel et al. 2022"},{"why":"Provides the cycleCDR baseline, the most directly related method also built on latent linear additivity and cycle consistency.","marker":"Huang and Liu 2024"},{"why":"Provides the GEARS baseline for genetic perturbations and the held-out dual-gene test set used for comparison.","marker":"Roohani, Huang, and Leskovec 2022"},{"why":"Supplies the cross-domain disentanglement architecture, DRANet, that the framework adapts for swapping perturbation representations.","marker":"Lee, Cho, and Im 2021"}],"fun_headline_variants":["Cross-transfer loss forces transferable drug perturbation vectors","Swapping latent states predicts unseen drug and knockout responses","Decoupling context from perturbation enables cross-domain drug prediction","XTransferCDR: context-free perturbation vectors via cross-transfer","Summing single-gene vectors predicts dual-knockout expression profiles"],"cache_read_input_tokens":14208,"weakest_assumption_plain":"The model's transferability rests on the assumption that a perturbation's effect is a vector that adds linearly to the cell's latent basal state, so that swapping or summing vectors in the learned space corresponds to actually applying the perturbations.","fun_headline_variants_meta":{"raw":{"variants":["Cross-transfer loss forces transferable drug perturbation vectors","Swapping latent states predicts unseen drug and knockout responses","Decoupling context from perturbation enables cross-domain drug prediction","XTransferCDR: context-free perturbation vectors via cross-transfer","Summing single-gene vectors predicts dual-knockout expression profiles"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000436,"raw_usage":{"total_tokens":2230,"prompt_tokens":969,"completion_tokens":1261,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":1180}},"tokens_in":585,"tokens_out":1261,"duration_ms":11543,"temperature":1.0,"reasoning_tokens":1180,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:48:26.915742+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of gene pairs with documented epistatic or non-additive interactions, train XTransferCDR on single-gene knockouts only, and compare its predicted double-knockout profiles $S + P^{(a)} + P^{(b)}$ to actual measured profiles; if DEG-level R2 collapses specifically on the epistatic pairs while staying high on additive ones, the linear-additivity assumption is falsified.","supporting_citations":[{"cited_title":"R.; McFaline-Figueroa, J","cited_arxiv_id":null,"evidence_quote":"Supplies the sci-Plex3 and sci-Plex4 single-cell drug response datasets used for the main drug experiments."},{"cited_title":"M.; Saunders, R","cited_arxiv_id":null,"evidence_quote":"Provides the genome-scale Perturb-seq single-gene knockout datasets in K562 and RPE-1 cells used for evaluation."},{"cited_title":"M.; Horlbeck, M","cited_arxiv_id":null,"evidence_quote":"Supplies the A549 combinatorial dual-gene knockout dataset used to test additive prediction of combined perturbations."},{"cited_title":"K.; De Donno, C.; Ji, Y.; Ibarra, I","cited_arxiv_id":null,"evidence_quote":"Proposes the linear additive latent-space modeling of perturbations that this paper adopts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the ChemCPA baseline and the related assumption of linear additivity in latent space."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the cycleCDR baseline, the most directly related method also built on latent linear additivity and cycle consistency."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the GEARS baseline for genetic perturbations and the held-out dual-gene test set used for comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the cross-domain disentanglement architecture, DRANet, that the framework adapts for swapping perturbation representations."}],"review_version":1}