12 major claims checked against the paper's own evidence: all adequately supported.
partialAbstract, Section 4.3, Figure 4Reviewer 1
Current predictive models of cellular responses fail primarily due to insufficient coverage of biological contexts, rather than limitations in data volume or model capacity.
supported
Evidence: The paper presents empirical evidence from its own analysis on a 22-million-cell T cell dataset (Figure 4) showing that DEG-F1 improves with the number of training contexts, not with cell count. It also reviews published benchmarks (Ahlmann-Eltze 2025) showing simple baselines match deep learning.
“We take the position that current predictive models of cellular responses fail primarily due to insufficient coverage of biological contexts, rather than limitations in data volume or model capacity.”
partialSection 4.1Reviewers 1, 2
Simple baselines perform on par with sophisticated architectures within a given biological context.
supported
Evidence: The paper cites Ahlmann-Eltze 2025 and other systematic benchmarks showing that no deep learning method consistently outperforms simple additive or mean baselines.
“A simple additive baseline achieved lower error than all deep models.”
partialSection 4.3Reviewers 1, 3
Context diversity, not just cell count, drives cross-context generalization.
supported
Evidence: The core empirical result from the scLDM analysis on the T cell dataset (Section 4.3, Figure 4E-F) shows that at controlled cell counts, higher context diversity yields significantly higher DEG-F1 (p=0.002) and lower L2-Δ (p=0.01).
“Even at controlled data volume, context diversity improves generalization.”
partialSection 3.4Reviewer 1
The perturbation prediction community faces a causal transport problem.
supported
Evidence: The paper formalizes this using Proposition 3.2 and connects the empirical failures to the statistical genetics literature and causal transportability theory.
“Proposition 3.2 (Informal). When changes in context modulate the mechanism linking perturbation P to response Y, predictions learned in source domain π need not transfer to target domain π* with different contexts...”
partialSection 4.3Reviewer 1
Standard evaluation metrics (e.g., correlation, L2 distance) can mask biologically meaningful failures.
supported
Evidence: The paper shows a dissociation between aggregate metrics and DEG-F1 (Section 4.3, Figure 3E-F): correlation between Correlation-Δ and DEG-F1 is r=0.26, and many configurations achieve high Correlation-Δ but low DEG-F1.
“DEG-F1 is weakly correlated with aggregate metrics (r = 0.26 and r = −0.05 with Correlation-Δ and L2-Δ).”
partialSection 7Reviewer 1
The Virtual Cell remains a worthy aspiration; achieving it will require respecting context.
supported
Evidence: The paper provides concrete recommendations for data generation, benchmarking, and modeling (Section 6) that follow from the theoretical and empirical analysis, without claiming impossibility.
“The Virtual Cell remains a worthy aspiration. Achieving it will require respecting the biological reality that context matters.”
partialAbstract, Section 3.4Reviewers 2, 3
The community faces a causal transport problem that cannot be solved by accumulating more data from the same distributions.
The theoretical framework (Section 3) clearly defines the transportability problem, and the empirical evidence shows diminishing returns, but the claim is partially supported because the empirical evidence is limited to one dataset and one model.
Evidence: Section 3.4 and empirical results showing that context coverage, not cell count, improves generalization.
“This perspective explains why accumulating more data from the same contexts yields diminishing returns: it refines performance for the training distribution of contexts, but does not inform the context-specific mechanism f_c* in a new target context.”
supportedAbstractReviewer 2
Current predictive models of cellular responses fail primarily due to insufficient coverage of biological contexts, not limitations in data volume or model capacity.
The claim is supported by the new computational experiment showing DEG-F1 improves with context coverage but not with cell count, and by cited benchmarks showing deep learning fails to outperform baselines.
Evidence: Section 4.3: Figure 4 and associated text showing DEG-F1 improves with number of training contexts (r=0.11 for cell count, but significant improvement with context diversity).
“This position paper argues that scaling model capacity is insufficient to solve the Virtual Cell problem because the primary failure mode is a lack of adequate coverage over diverse biological contexts , not insufficient model expressivity.”
supportedSection 4.3Reviewer 2
Context diversity, not cell count, drives cross-context generalization in perturbation prediction.
The new experiment on the T-cell dataset shows DEG-F1 improves with context coverage (mean DEG-F1 increase from <0.1 to ~0.19 when moving from ≤3 to 8 contexts) and only weakly correlates with cell count (r=0.11).
Evidence: Section 4.3: 'DEG-F1 improves substantially with context coverage ... In contrast, DEG-F1 shows only a weak association with cell count ( r = 0.11 ; ).'
“DEG-F1 improves substantially with context coverage (mean DEG-F1 increases from < 0.1 to ~ 0.19 when moving from ≤ 3 to 8 contexts; ), indicating that perturbations observed across more training contexts yield better biological recovery in held-out contexts”
supportedSection 1Reviewer 2
AlphaFold's success does not generalize to cellular response prediction because protein folding is context-invariant while cellular responses are context-dependent.
The paper provides a clear contrast (Figure 1) and discusses why the AlphaFold analogy fails, citing the Anfinsen principle and citing benchmarks showing context-dependence.
Evidence: Section 1 (Figure 1) and Section 5: 'Protein folding obeys relatively context-invariant physical laws ... Cellular response, by contrast, is exquisitely context-dependent.'
“Protein folding obeys relatively context-invariant physical laws: the same sequence folds to the same structure regardless of whether the protein resides in a T cell or a neuron (). Cellular response, by contrast, is exquisitely context-dependent.”
supportedSection 4.3Reviewer 2
Standard evaluation metrics (Δ-based) overestimate model performance and dissociate from biologically meaningful DEG recovery.
The paper shows a weak correlation (r=0.26, r=−0.05) between Δ-based metrics and DEG-F1 in its experiment, and cites Vinas Torne et al. on systematic variation driving metric inflation.
Evidence: Section 4.3: 'DEG-F1 correlates weakly with aggregate metrics ( r = 0.26 and r = − 0.05 with Correlation-Δ and L2-Δ; –).'
“DEG-F1 correlates weakly with aggregate metrics ( r = 0.26 and r = − 0.05 with Correlation-Δ and L2-Δ; –).”
supportedAbstract, Section 4.3Reviewer 3
Scaling model capacity is insufficient to solve the Virtual Cell problem; the primary failure mode is lack of context coverage.
The paper provides theoretical reasoning (causal transportability) and empirical evidence (context coverage improves DEG-F1) that directly support this claim.
Evidence: Figure 4 shows DEG-F1 improves with number of training contexts, and controlled comparison (p=0.002) supports the benefit of context diversity.
“This position paper argues that scaling model capacity is insufficient to solve the Virtual Cell problem because the primary failure mode is a lack of adequate coverage over diverse biological contexts, not insufficient model expressivity.”