acceptodds
Under review as a conference paper at ICLR 2027

The Platonic gene: convergent representations across biological models learn cell plausibility

Abstract

Protein coding genes can be represented in myriad ways, including their nucleic acid or protein sequences, molecular networks, textual annotations, co-expression in cells, and perturbation phenotypes. The Platonic representation hypothesis proposes that increasingly capable models converge toward shared representations of the underlying world, even when trained on different modalities. Here, we test this hypothesis for gene biology. We find that representations from different modalities, architectures, and training objectives can agree on the relationships between genes. This agreement is stronger among models that predict biological processes better, in accordance with the Platonic representation hypothesis. Nonetheless, convergence is not perfect, raising the question whether differences between representations capture complementary biological information that can be used to obtain a more holistic representation. We show that combining six models into a shared PlantaGenet consensus representation improves gene-property prediction. We use the geometry of genes in the PlantaGenet consensus space to assess the biological plausibility of single-cell expression profiles. In a combined human blood and brain evaluation, cell coherence and dominant-dimensionality scores distinguish real observed cells from gene-shuffled, independently sampled gene counts, and cross-tissue chimeric expression profiles, achieving AUROCs of 1.000, 0.992, and 0.845, respectively, outperforming individual models. These results suggest that biological models capture both shared and complementary gene relationships, and that integrating these relationships provides a geometric approach to assess gene-expression coherence, with potential implications for biological discovery and cell engineering.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Computational genomics, Platonic representation, gene embeddings, single-cell, Gene Biology, Cell Biology, Foundation Models

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

The paper tests the Platonic representation hypothesis using 33 gene embeddings from text, sequence, network, expression, perturbation, and imaging models. Mutual-kNN agreement is above chance and correlates with gene-property prediction (r = 0.75). The authors combine six models into a PLANTAGENET consensus that improves gene-property prediction. They also derive two cell scores, ρc and Dc, from the consensus. These distinguish real cells from shuffled, marginal, and chimeric profiles, with reported AUROCs of 1.000, 0.992, and 0.845, respectively.

Critical strengths

The comparison covers 33 gene representations across very different modalities. I also like the simplicity of the cell scores: they require no training and provide a novel way to use gene-level representations to assess cells.

Critical weaknesses
  1. The headline cell results come from the easiest setting. In the combined experiment, chimeras and marginals mix blood with brain, so the experiment mostly tests whether the scores detect cross-tissue mixing. Within tissue, Table S4 shows that the consensus is near chance for chimeras (0.55 to 0.63) and brain marginals (0.49 to 0.54). Shuffled profiles are trivial for every model.

  2. The paper lacks baselines for the cell scores. There is no comparison with detected-gene count, co-expression likelihood, or the authors' own atlas PCA views. Doublet detectors such as DoubletFinder are also missing, even though the chimeras are synthetic doublets. The stated use case is evaluating generative-model outputs, but none are tested.

  3. The novelty is limited by earlier work on multimodal gene-embedding integration. The paper does not cite arXiv 2507.07367, and PRESAGE already concatenates seven views. Kan-Tor et al. (2024), whose benchmark is used here, also reported complementarity across model families.

  4. Several confounds affect the convergence analysis. Fold-over-chance (≈ k/|G|) favors pairs that share more genes. PRESAGE overlaps with its own components, inflating the correlation, while text and network embeddings share sources with the benchmark labels.

  5. The consensus gains are small. VI and VIII differ by 0.004 AUROC, with no variance reported. The headline numbers also mix results from the concatenation and Procrustes variants.

Questions for authors
  1. How do ρc and Dc perform within tissue compared with doublet-detection and co-expression baselines?
  2. Does the correlation between agreement and performance hold when you use a common gene set and exclude PRESAGE?
  3. How does PLANTAGENET compare with existing integration methods or with concatenating the top-k models?
5 comments
AuthorsOpenReview ↗

Response to reviewer yjQN (1/3)

(part 1/3)

We thank the reviewer for recognizing the breadth of the representation analysis and the potential value of the cell-level geometric scores. In response, we expanded the evaluation with conventional and reference-dependent cell-scoring baselines, analyses of tissue-specific chimera difficulty, and controls examining alternative consensus constructions. We also clarified the relationship to previous multimodal gene-embedding integration work.

We are happy to engage in further discussion (:

Q1: How do ρc\rho_c and DcD_c perform within tissue compared with doublet-detection and co-expression baselines?

We expanded the evaluation to compare our geometric scores with conventional QC metrics, co-expression likelihood, expression-space methods, and Scrublet across both combined and tissue-specific settings (Table S14).

Importantly, ρc\rho_c and DcD_c require only the expression profile of the evaluated cell and a pretrained gene representation, whereas co-expression and doublet-detection methods rely on a reference population. Despite this distinction, our scores perform competitively across most settings. Our scores also detect gene-shuffled profiles with near-perfect accuracy across tissues, while conventional QC statistics perform at near chance. Specialized reference-dependent methods achieve higher performance on certain chimeric profiles, particularly in the brain setting. Our graded-chimera analysis (Figure S33) helps explain this behavior; rejection becomes more challenging as the transcriptional similarity between source cell types increases. We believe these comparisons highlight the value of assessing cellular coherence directly through gene-space geometry, without requiring a representative reference population or task-specific model fitting. And wish to clarify we designed the metrics as complimentary and not as the unique evaluation system.

Manuscript: We expanded Sec. 4.3 with the tissue-specific and cell-scoring comparisons, supported by Appendix F.7 (Tables S14) and the graded-chimera analysis in Appendix F.8 (Figure S33).

W1: The headline cell results come from the easiest setting [...] the experiment mostly tests whether the scores detect cross-tissue mixing. [...]

We agree that cross-tissue chimeras are easier to reject than within-tissue mixtures. We therefore evaluated chimeras generated from 406 source-cell-type pairs spanning different transcriptional similarities (Figure S33).Rejection decreases as source-type similarity increases (Spearman rs=−0.65r_s = −0.65 for coherence and −0.58−0.58 for dominant dimensionality). Varying the mixing ratio has a smaller effect (Figure S33). These findings explain part of the performance gap between combined and tissue-specific settings (Table S12) and are now mentioned in section Sec 4.3.

Manuscript: We expanded the tissue-specific interpretation in Sec. 4.3 and added the graded-chimera analysis in Appendix F.8 (Figure S33).

AuthorsOpenReview ↗

Response to reviewer yjQN (2/3)

(part 2/3)

W2: The paper lacks baselines for the cell scores. [...]

We expanded the cell-level evaluation with conventional QC statistics and reference-dependent methods, including PCA reconstruction, expression-space distance, co-expression likelihood, and doublet detector (Scrublet; Table S14).

Our geometric scores outperform the four conventional QC statistics on several corruption tasks, including marginal and chimeric profiles in blood and combined blood-and-brain settings. Specialized reference-dependent methods, however, can perform better on chimeras: Scrublet reaches AUROC 0.871 versus 0.637 for coherence in blood, and 0.923 versus 0.556 in brain; expression-space distance reaches 0.950 versus 0.845 on combined blood-and-brain chimeras (Table S14). The distinction is that our scores use pretrained gene-space geometry without requiring a reference-cell population at evaluation time. Additional comparisons with same-modality and atlas-PCA representations are reported in Table S12.

The atlas-PCA views score higher in tissue-specific settings, but only because they are fit on the very atlases the test cells come from (which our scores don’t have access to); the multimodal consensus remains best on the combined blood-and-brain setting (Table S12)

Manuscript: We expanded Sec. 4.3 and Appendix F.7 with the cell-scoring comparisons (Table S14), and report the representation controls in Appendix F.4 (Table S12).

W2: The stated use case is evaluating generative-model outputs, but none are tested.

We agree that detecting selected synthetic corruptions does not establish biological validity for arbitrary generated or experimentally perturbed cellular states. The current results establish a reference-free geometric scoring approach and characterize its behavior on several constructed corruption types. Evaluating previously unseen cellular states against independent biological validity criteria remains a future application, beyond the scope of this work which intends to set the foundations.

Manuscript: We clarify the demonstrated scope of Sec. 4.3 and the prospective generative-model motivation in the Discussion, Sec. 5.

W3: The novelty is limited by earlier work on multimodal gene-embedding integration. The paper does not cite arXiv 2507.07367, and PRESAGE already concatenates seven views. Kan-Tor et al. (2024), whose benchmark is used here, also reported complementarity across model families.

We agree that multimodal gene-embedding integration and complementarity across modalities are established concepts and thank the reviewer for highlighting the missed reference to PRISME (Zheng et al., 2025). We would like to highlight that in each of the previous works the multimodality was explored from a different lens and towards a different objective:

  • Kan-Tor et al. (2024) demonstrated modality-specific strengths in gene-property prediction;
  • PRESAGE (Littman et al., 2025) integrates multiple gene-embedding sources for perturbation-response prediction; and
  • PRISME examines cross-modal representational similarity and integrates embeddings through an autoencoder.

We believe our contribution is different, it is not a new concatenation or alignment algorithm. Rather, we connect the systematic analysis of cross-modal gene agreement, its use in selecting representations for a training-free consensus space, and the application of the resulting geometry to cell-level analysis.

Manuscript: We expanded the related-work discussion in Sec. 2 and clarified the selection rationale in Sec. 3.5.

AuthorsOpenReview ↗

Response to reviewer yjQN (3/3)

(part 3/3)

Q3: How does PlantaGenet compare with existing integration methods or with concatenating the top-k models?

We evaluated top-kk combinations across gene-property prediction and cell rejection (Figures S27 and S32). Gene-property prediction largely saturates at approximately four representations, while cross-tissue chimera rejection continues to benefit from additional models. In contrast, adding views can reduce performance for brain-only marginal corruptions. These findings demonstrate that contributions are task-dependent, rather than establishing a universally optimal number of representations. We also added an ensemble averaging predictions from the six individual models (Table S8). The ensemble achieves an all-task AUROC of 0.892, comparable to PlantaGenet VI Procrustes (0.886). This results motivate the integration of complementary biological information across representations. PlantaGenet builds on this principle by integrating the different views into a single, reusable gene representation space, whose geometry can be leveraged directly for downstream cell-level analysis (which the ensemble method can’t provide).

Manuscript: We added the top-k sensitivity analysis to Sec. 4.2 and 4.3 and Appendix E.5 (Figure S27) and F.5 (Figure S32) and the ensemble method to Sec 4.2 and Appendix E.3 (Table S8)

Q2: Does the correlation between agreement and performance hold when you use a common gene set and exclude PRESAGE?

We now evaluate agreement on two common-gene panels (Table S4). Excluding the two integrated PRESAGE representations preserves the positive association between biological-process prediction and calibrated agreement after controlling for model characteristics (β=0.53\beta= 0.53, p=1.1×10−5p = 1.1 \times 10^-5).

The relationship also persists in the broader 16,385-gene evaluation (β=0.39\beta = 0.39, p=4×10−4p = 4 \times 10^-4). We hope these results reduce concerns about differing gene coverage and dependence on integrated views.

Manuscript: We clarified the common-gene evaluation in Sec. 3.2 and expanded the controlled analyses in Sec. 4.1 and Appendices A-C (Table S4; Figures S18–S21).

W5: The consensus gains are small [...] The headline numbers also mix results from the concatenation and Procrustes variants.

We corrected the variant-specific comparisons in Sec. 4.2 and added uncertainty estimates across 10 CV seeds and tasks composition (Table S6). We also reported confidence intervals for the cell rejection task (Table S11). We do not interpret the VI–VIII differences as improvements – the motivation for VI is to reduce dependence on tissue-specific atlas representations (by removing the PCA-views fitted on these altases) rather than maximize benchmark AUROC.

Manuscript: We revised Sec. 4.2 and provide the uncertainty analysis in Appendix E.1 (Table S6) and F.3 (Table S11)

Reviewer_91ynOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper tests the Platonic representation hypothesis for gene embeddings. It compares 33 gene representations from 11 model families and the agreement metric is mutual 10-nearest-neighbor (mNN) overlap, divided by a permutation null. Agreement is above chance for all families. It also correlates with performance on previous gene-property benchmark.

The authors then combine six models into a consensus space , by concatenation or by generalized Procrustes alignment. The consensus gives a higher benchmark AUROC than each single model. Finally, the authors define two training-free cell scores on the consensus geometry: coherence ρc\rho_c and number of dominant directions DcD_c. These scores separate real blood and brain cells from shuffled, marginal, and chimeric profiles.

Critical strengths
  • The paper tackle an important question, multimodality is a cornerstone of biology, but the combination of modalities is still a challenge. Also, the scale of the data is impressive, overall this is a good effort to tackle an important question.
  • The Algorithms are fully specify the consensus and the cell scores.
  • The paper presents the tissue-specific results, including those where the consensus loses.
  • The consensus improves the results at three levels: gene properties (Table 1), lineage clustering (Table S3, NMI 0.156 vs. 0.130 for scGPT), and cells (Table 2). The margins are small.
Critical weaknesses
  • The title, the abstract, and S4.1 claim convergence. however, the Platonic representation hypothesis predicts a trend. That is- alignment increases with scale and competence (Huh et al., 2024; l.313–314). While the paper shows two different results (agreement is above a random null, agreement correlates with benchmark AUROC), neither result presents a trend. For example : l.314 excludes scale from the analysis, and Fig. S4 shows the number of parameters only as a color. Also - l.474–478 proposes that "Platonic representation" should mean a consensus representation, not convergence. This changes the hypothesis; it does not test it. Also, the paper shows that the differences between views add information (l.362–365). This result is evidence against strict convergence, not for it.

  • Regarding the null hypothesis, in Kan-Tor et al. the biological-process and protein-property labels come from UniProt keywords and Reactome top-level pathways. Hence, some views use the same sources, and this shared source can violate the null hypothesis been a confounder for similarity. Also, the uniform permutation null removes all structure. Any shared gene-level property can give above-chance agreement, for example how well a gene is studied, its expression level, or its protein length. Does the agreement remain after the removal of such covariates, or against a covariate-matched null?

  • The paper cites Gröger et al. (2026) (l.083–085), who show that model width and depth inflate mNN. The paper does not apply their calibration. Here, the embedding dimension goes from 30 to 9,208 (Table S1). After calibration, Gröger et al. find that local neighborhood agreement remains, but global convergence does not. mNN with k=10k=10 is a local metric. Thus, the results support the weaker "Aristotelian" hypothesis, not the Platonic one.

  • The claim that PlantaGenet VI beats the best single model on all five task families (l.358–360) uses different variants per column. Concatenation wins biological processes and protein properties. Procrustes wins localization. Neither variant alone beats the best single model on all five families (Table 1). The VI vs. VIII claims (l.360–361) also take the concatenation result for one metric and the Procrustes result for the other. The margins are 0.004–0.005, from a single 5-fold cross-validation with no variance. Please use one fixed variant per claim, and report the standard deviation across folds or seeds.

Questions for authors

In Weaknesses

3 comments
AuthorsOpenReview ↗

Response to reviewer 91yn (1/3)

(part 1/3)

We thank the reviewer for the careful reading and for highlighting important distinctions between shared representational structure, the stronger predictions of the Platonic representation hypothesis, and consensus-space construction.

Following the reviewer’s comments we expanded the analysis with calibrated similarity metrics, model-scale associations, controls for potential confounders, and statistical uncertainty as detailed below.

We are happy to engage in further discussion (:

W1: While the paper shows two different results (agreement is above a random null, agreement correlates with benchmark AUROC), neither result presents a trend.

We tested whether bigger and more capable models show higher agreement (figures S10, S11). We see that there is a strong positive correlation between the number of model parameters and mNN agreement (Pearson’s r=0.44,p=0.014r = 0.44, p = 0.014 for 415 well-studied genes’ representations across 31 models with known parameter count, Pearson’s r=0.75,p=3.8×10−5r = 0.75, p = 3.8 \times 10^{-5} for 16,385 genes across 23 models). Similarly, bigger models have higher performance in gene property prediction benchmark (Pearson’s r=0.60,p=3.5×10−4r = 0.60, p = 3.5 \times 10^{-4} for 415 genes across 31 models, Pearson’s r=0.80,p=4×10−6r = 0.80, p = 4 \times 10^{-6} for 16,385 genes across 23 models). From these results, it follows that more powerful models are more convergent and predict biological properties better, in agreement with the Platonic representation hypothesis.

Manuscript: We expanded section 4.1 to mention model scale analysis, and added figures S10, and S11 to the appendix.

W1: l.474–478 proposes that "Platonic representation" should mean a consensus representation, not convergence. This changes the hypothesis; it does not test it.

Thank you for this point. We are sorry our intention was not clear – we rephrased this line in the text. We did not aim to redefine the Platonic Representation Hypothesis, but to propose a complementary perspective; suggesting that shared representational structure need not capture all concept relevant information – modality-specific differences may encode additional information, motivating their integration into a more comprehensive representation in some cases.

Manuscript: we revised Sec. 5 to clarify this distinction.

W1: the paper shows that the differences between views add information (l.362–365). This result is evidence against strict convergence, not for it.

We partially agree and this is also a point raised in the original paper Huh et al., 2024: “One immediate objection to our hypothesis is: what about the information that is unique to a given modality?”. We hypothesize that the true picture is more intricate than the absolute convergence. There is shared information between modalities – more powerful models capture it better and their representations converge to a certain limit. But there is also modality-specific information. The second part of our paper dives deeper into how we can build better representations, bringing additional value from modality-specific information on top of shared information.

W2: some views use the same sources, and this shared source can violate the null hypothesis been a confounder for similarity

Thank you for the great points. We extended our paper with extensive robustness analysis (section C). To account for shared data between representations, we built a linear model for predicting representation similarity (measured by calibrated mNN overlap) from gene properties benchmark performance with correction for shared modality between models, shared data in the training set, using the same architecture and using the same training objective. The association between gene mNN overlap and gene properties prediction AUROC persists after adjusting for all confounders (β=0.58,p=5.5×10−6\beta = 0.58, p = 5.5 \times 10^{-6} for 415 genes, Figure S18; β=0.39,p=4×10−4\beta = 0.39, p = 4 \times 10^{-4} for 16,385 genes, Figure S19). Completely excluding the views that contain information from the Kan-Tor et al. benchmark (text-based embeddings, gene network representations, integrated PRESAGE embeddings) does not break the association between mNN overlap and biological processes AUROC (r=0.54,p=0.011r = 0.54, p = 0.011, appendix C).

Manuscript: we added subsections “Model and training data confounders” and “annotation-derived representations” and figures S18, S19 to appendix C.

AuthorsOpenReview ↗

Response to reviewer 91yn (2/3)

(part 2/3)

W2: Any shared gene-level property can give above-chance agreement, for example how well a gene is studied, its expression level, or its protein length. Does the agreement remain after the removal of such covariates, or against a covariate-matched null?

We thoroughly tested the robustness of our results against how well genes are studied, their expression level and protein length (appendix C). First, we restrict results reported in the main paper to 415 well-studied genes to make sure each representation contains the same set of genes and these genes are well studied. In the appendix, we additionally report results for an extended set of 16,385 genes that are shared by 23 models (appendix B.2). Our results hold on both sets of genes.

We then tested whether gene properties can affect our results. We characterized each gene by logarithm of its publication counts, expression level and logarithm of protein length. Median calibrated mNN agreement of cross-family gene representations decreases from 0.039 to 0.031 after covariate-matched null and to 0.028 after regressing out gene covariates from gene embeddings (Figures S22, S23). However, overall results remain robust with 95% and 93% of cross-family gene representation pairs remaining higher than expected at random for 415 genes panel. To evaluate the effects of gene annotation quality deeper, we stratified 16,385 genes into well-studied, partially characterized and mostly unknown. Median calibrated mutual-10NN was 0.0095, 0.0078, and 0.0059, respectively, showing that well-studied genes have higher mNN agreement, but even mostly unknown genes have higher agreement than expected at random.

Overall, our conclusions remain robust to correction for gene-related confounders.

Manuscript: we added subsection “Gene-level covariates” and figures S22, S23 to appendix C. We expanded main section 4.1 to mention results robustness.

W3: The paper cites Gröger et al. (2026) (l.083–085), who show that model width and depth inflate mNN. The paper does not apply their calibration.

Thank you for raising this point. We now apply the calibration of Gröger et al. and report calibrated mNN everywhere in the paper.

Manuscript: we recomputed figure 2 with calibrated mNN overlap, updated numbers in section 4.1, and described the calibration in appendix B.1. Figures S3-S4 contain calibrated mNN values between all pairs of representations, and Figure S5 shows analogue of Figure 2 with extended set of genes.

W3: After calibration, Gröger et al. find that local neighborhood agreement remains, but global convergence does not. mNN with is a local metric. Thus, the results support the weaker "Aristotelian" hypothesis, not the Platonic one.

We extended our analysis with a centered kernel alignment (CKA) metric calibrated with correction from Gröger et al (appendix B.3). We see very similar results to our mNN analysis:

  • Higher than expected CKA across distinct models (statistically significant for 98.1% of cross-family pairs)
  • High correlation between gene property benchmark and mean cross-model calibrated CKA (Pearson’s r=0.60,p=2×10−4r = 0.60, p = 2 \times 10^{-4} for 415 genes, Figure S12)

For 16,385 genes, we see a clear linear trend between gene property prediction and mean cross-model CKA (Figure S13, Pearson’s $r = 0.30, p = 0.16), however, it is broken by outliers. Protein language model and text embeddings of gene descriptions that have low CKA and high gene property benchmark performance.

Overall, our results show that gene embeddings show both local and global similarity, supporting the Platonic representation hypothesis for well-characterized genes.

Manuscript: we added section B.3 with the description of CKA and figures S12-S15 to the appendix. We added reference to these results in section 4.1 of the main text.

W4: The claim that PlantaGenet VI beats [..] uses different variants per column. [...] also take the concatenation result for one metric and the Procrustes result for the other” .

We thank the reviewer for identifying this inconsistency. We now report each construction separately: VI concatenation outperforms the best individual representation in four of five task families, whereas VI Procrustes does so in three. The VI–VIII comparisons are also reported using matched construction methods, rather than selecting the best result from different variants.

Manuscript: We corrected the variant-specific comparisons in Sec. 4.2.

AuthorsOpenReview ↗

Response to reviewer 91yn (3/3)

(part 3/3)

W4: The margins are 0.004–0.005, from a single 5-fold cross-validation with no variance [...] report the standard deviation across folds or seeds.

We do not interpret the VI–VIII differences as improvements – the motivation for VI is to reduce dependence on tissue-specific atlas representations rather than maximize benchmark AUROC.

As for the standard deviation across folds or seeds, we repeated gene-property prediction across ten CV seeds and computed nested 95% confidence intervals incorporating variability across both seeds and benchmark tasks (Table S6). Across-seed variability is small (SD ≤\leq 0.003), while task variability contributes wider confidence intervals.

PlantaGenet VI Procrustes achieves higher point estimates than all six individual models. However, its 95% CI overlaps AlphaGenome's on the all-task aggregate, and those of ESM-2 and GenePT-protein on the family-mean aggregate (Table S6). The individual models nevertheless retain task-specific strengths: GenePT-protein performs better on biological processes (0.948 vs. 0.935), and ESM-2 on protein properties (0.951 vs. 0.915), whereas VI Procrustes leads on localization, regulatory functions, and genomic properties (Table 1).

Separately, cell-level performance also varies across tissues: AlphaGenome achieves the highest brain-only mean AUROC (0.791 vs. 0.684 for VI Procrustes), whereas VI Procrustes leads on blood-only and combined blood-and-brain evaluations (0.848 and 0.938, respectively; Table S10). These results highlight the complementary, context-dependent strengths of individual representations and the consensus.

Manuscript: We added uncertainty reporting to Sec. 4.2 and Appendix E.1 (Table S6).

Reviewer_FdY9OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether gene representations learned from different biological modalities, model architectures, and training objectives exhibit convergence consistent with the Platonic Representation Hypothesis. The authors systematically analyze 33 gene representations derived from DNA and protein sequences, textual annotations, biological networks, single-cell transcriptomics, and perturbation phenotypes. Using mutual nearest-neighbor (mNN) agreement, they demonstrate that independently learned gene embeddings exhibit non-random representational overlap, with stronger agreement associated with better biological-property prediction performance. Motivated by the observation that different representations also capture complementary biological information, the authors introduce PlantaGenet, a consensus gene-embedding framework that integrates six pretrained models through normalized concatenation or generalized Procrustes alignment. The resulting consensus representations improve performance on a gene-property prediction benchmark covering biological processes, protein properties, localization, regulatory functions, and genomic properties. The authors further propose two geometric measures, cell coherence and the effective number of dominant directions, to evaluate the biological plausibility of single-cell expression profiles.

Critical strengths
  • The first critical strength is the systematic investigation of representational convergence across heterogeneous biological modalities. The paper addresses an important and relatively underexplored question in biological foundation modeling: whether models trained on different measurements of gene biology recover common underlying relationships between genes.

  • The second critical strength is the demonstration that integrating complementary gene representations improves downstream biological prediction and enables a potentially useful cell-level application. Rather than assuming that representational convergence implies redundancy, the authors recognize that modality-specific differences may encode complementary biological information.

Critical weaknesses
  • The first major weakness is that the evidence for the Platonic Representation Hypothesis is primarily correlational and does not sufficiently distinguish meaningful biological convergence from shared annotation structure, model-family effects, or differences in representation quality. Although the authors report substantial mutual nearest-neighbor agreement above random expectations, the absolute agreement between many representations from different modalities remains relatively low. For example, Figure S6 shows that several cross-modal model pairs share only approximately 1–4% of their ten nearest neighbors, despite exhibiting multiple-fold enrichment over random expectation. Such results demonstrate statistically non-random overlap but do not necessarily establish convergence toward a common biological representation. Moreover, the reported correlation between representation agreement and biological-property prediction performance may be influenced by shared supervision, correlated training data, gene annotation density, or model-family membership. Natural-language representations derived from UniProt annotations, for example, may have substantial overlap with the biological knowledge used to define downstream prediction labels. The analysis also does not sufficiently separate model capability from modality-specific characteristics or evaluate whether the reported correlation persists after controlling for these factors. This distinction is consequential because the central scientific claim concerns representational convergence as a general property of biological models, whereas the current evidence more directly supports shared local gene relationships among heterogeneous pretrained representations.

  • The second major weakness is that the biological plausibility evaluation relies on relatively simple synthetic corruption procedures and does not establish that the proposed geometric scores capture biologically meaningful cellular validity beyond statistical regularities of the reference datasets. The main experiments distinguish real blood and brain cells from gene-shuffled profiles, independently sampled gene counts, and synthetic chimeras. However, gene shuffling and independent sampling destroy fundamental expression dependencies and may introduce distributional artifacts that can be detected without sophisticated biological representations. The near-perfect AUROCs for these corruption types therefore provide limited evidence that PlantaGenet can identify more realistic but biologically invalid cell states. The chimera experiment is more challenging, but the reported improvement is primarily observed in the combined blood-and-brain setting, where artificial profiles are constructed by mixing cells from highly distinct tissues. Importantly, Appendix Table S4 shows substantially weaker performance in tissue-specific experiments. For example, PlantaGenet VI with Procrustes alignment achieves chimera coherence AUROCs of only 0.632 for blood and 0.554 for brain, compared with 0.845 in the combined blood-and-brain experiment. Furthermore, AlphaGenome outperforms the consensus representations on the aggregate brain-only evaluation, suggesting that the proposed integration does not consistently improve biological plausibility assessment across settings. Since the manuscript motivates this approach as a potential discriminator for generative models of cells, the absence of evaluation on realistic generated expression profiles, subtle perturbation-induced states, or experimentally validated abnormal cellular states substantially limits the strength of this application claim.

Questions for authors
  • Could the authors clarify what evidence establishes genuine cross-modal representational convergence rather than agreement induced by shared biological annotations, training-data overlap, or model-family characteristics? The reported correlation between mutual nearest-neighbor agreement and biological-property prediction performance is interesting, but it remains unclear whether this relationship persists after controlling for modality, model architecture, gene annotation density, and other potential confounders. In particular, representations derived from textual gene descriptions may encode biological knowledge that overlaps with the downstream benchmark labels, potentially contributing to both high prediction performance and strong agreement with other biologically informed models. A more controlled analysis separating within-family and cross-family agreement, together with partial-correlation or regression analyses accounting for these factors, would materially strengthen the central claim. It would also be useful to report absolute nearest-neighbor overlap alongside enrichment over random expectation and examine whether the findings are robust to different neighborhood sizes and alternative representation-similarity metrics. These analyses would help distinguish the weaker claim that biological representations share some local structure from the stronger Platonic hypothesis that increasingly capable models converge toward a common representation of gene biology.

  • Could the authors provide stronger evidence that the improvements of PlantaGenet arise from biologically complementary information rather than increased representation dimensionality or the aggregation of multiple strong predictors? The proposed consensus representations are constructed using relatively straightforward concatenation or Procrustes alignment, making it important to isolate the contribution of multimodal complementarity from the benefits of combining multiple feature spaces. In particular, the concatenation-based PlantaGenet VI representation has 768 dimensions, whereas the Procrustes representation has 128 dimensions, and the individual models are first reduced to 128 principal components. Although Table 1 shows improvements over individual representations, the manuscript does not provide sufficiently controlled comparisons against dimension-matched alternatives, simple ensemble predictions, or consensus representations constructed from multiple models within the same modality. A leave-one-modality-out analysis would also clarify which biological information sources contribute most strongly to each downstream task. This is particularly relevant because the DepMap CRISPR representation performs relatively poorly on the gene-property benchmark, while the benefits of including it in the consensus are not directly established. Such experiments would help determine whether the proposed integration captures genuinely complementary biological information or primarily benefits from combining several informative feature sets.

  • Could the authors clarify how the gene-property prediction benchmark controls for annotation overlap and potential information leakage between pretrained representations and evaluation labels? The benchmark includes biological processes, protein properties, localization, regulatory functions, and genomic properties, while several evaluated representations are derived from textual annotations, pathway information, or models pretrained on biological measurements related to these labels. For example, GenePT-protein uses UniProt protein summaries, which may explicitly describe biological processes and protein functions represented in the evaluation tasks. Consequently, strong linear-probe performance may partly reflect the recovery of previously encoded annotations rather than generalization to independent biological properties. The five-fold cross-validation procedure evaluates prediction across held-out genes, but it does not necessarily address overlap between pretrained knowledge sources and downstream labels or correlations among closely related genes. Could the authors provide an evaluation using temporally held-out annotations, independent experimental measurements, or gene-family-disjoint splits? It would also be valuable to report variability across cross-validation folds and statistical uncertainty for the differences between consensus and individual models. These analyses would materially affect the interpretation of the reported improvements as evidence of more comprehensive biological understanding.

  • Could the authors explain why the consensus representations show substantially weaker performance in tissue-specific corrupted-cell evaluations and whether the combined blood-and-brain results are driven primarily by cross-tissue separation? Table 2 reports strong performance for PlantaGenet VI with Procrustes alignment, including a chimera coherence AUROC of 0.845. However, Appendix Table S4 shows that this value decreases to 0.632 for blood-only chimeras and 0.554 for brain-only chimeras, approaching random discrimination in the latter setting. The same table also shows that AlphaGenome achieves the strongest overall performance in the brain-only evaluation, outperforming both consensus variants. These observations raise the possibility that the consensus geometry is particularly effective at detecting mixtures of distinct tissue-level transcriptional programs but less sensitive to biologically incompatible combinations within a tissue. Could the authors analyze performance as a function of the transcriptional similarity between the source cell types used to construct chimeras? Comparisons involving closely related cell types, matched cell-type proportions, and chimeras with different mixing ratios would help establish whether the geometric scores capture fine-grained biological incompatibility rather than broad tissue differences. This distinction is central to the proposed application of assessing cellular plausibility.

  • Could the authors provide a more controlled evaluation of the proposed coherence and dominant-dimensionality scores against established single-cell quality-control and biological-consistency baselines? The two proposed metrics are intuitive geometric statistics derived from weighted gene embeddings, but it remains unclear whether their predictive performance depends specifically on the biological structure of PlantaGenet or on simpler properties of gene-expression profiles, such as the number of detected genes, library size, expression sparsity, or the concentration of highly expressed genes. Although the authors use log-transformed normalized counts, the synthetic corruption procedures may alter these statistical properties in ways that affect the resulting scores. Comparisons against conventional single-cell quality-control features, PCA-based reconstruction errors, expression-space distances, and existing doublet-detection methods would help establish the incremental value of the proposed geometric formulation. In addition, the authors could evaluate whether performance remains stable after matching real and corrupted cells for library size, detected-gene count, and other relevant expression statistics. Such controls are important for determining whether the reported discrimination reflects biologically meaningful gene relationships rather than differences in basic transcriptomic distributions.

4 comments
AuthorsOpenReview ↗

Response to reviewer FdY9 (1/4)

(part 1/4)

We thank the reviewer for the thorough evaluation and constructive suggestions, particularly for recognizing the value of both our cross-modal representation analysis and the PlantaGenet consensus space.

Following the reviewer's comments, we substantially expanded the evaluation and clarified the distinction between the two contributions of our work.

First, we investigate whether independently learned gene representations recover shared biological structure. Through additional similarity metrics and controls for model- and gene-level confounders, we strengthen the evidence for shared local gene relationships, without equating this with complete representational convergence.

Second, building on both shared and complementary information across representations, PlantaGenet constructs a training-free, reusable gene space. Unlike task-specific prediction ensembles, this shared geometry enables direct cell-level analysis through reference-free coherence and dominant-dimensionality scores. We expanded the evaluation with integration controls, conventional baselines, and tissue-specific analyses to characterize their capabilities and limitations.

We address the reviewer's specific concerns and detail the corresponding manuscript revisions below.

We are happy to engage in further discussion (:

Q1: What evidence establishes genuine cross-modal representational convergence rather than agreement induced by shared biological annotations, training-data overlap, or model-family characteristics?

To address this point, we now evaluate agreement on two common-gene panels (Table S4): small (415 genes) represented by all 33 models (main text; Appendix B, Figure S3) and large (16,385 genes) shared by 23/33 models (Appendix B, Figures S4–S5, S9).

To assess the influence of biological annotation density, we stratified the extended panel into three annotation-coverage classes (Oprea et al., 2018); (1) well-studied, (2) partially characterized, and (3) mostly unknown genes. Cross-family agreement is significantly higher than expected at random across all three groups (despite being lower for less characterized genes). The median calibrated mutual-10NN is: 0.0095, 0.0078, and 0.0059, for well-studied, partially characterized and mostly unknown genes respectively (Appendix C, Figure S23a).

To examine model-level confounders, we tested whether pairs of models with higher biological-process prediction performance also exhibit greater representational agreement. We regressed their calibrated mutual-10NN agreement against the average AUROC of each model pair, controlling for shared modality, training data, architecture, and training objective. The association remained significant after these controls in both gene panels (small: β=0.58,p=5.5×10−6\beta = 0.58, p=5.5 \times 10^{-6}. large: β=0.39,p=4×10−4\beta = 0.39, p = 4 \times 10^{-4}; Appendix C, Figures S18, S19)). Shared modality and training data contribute to agreement, whereas architecture and training objectives do not show significant independent effects.

To make sure that annotation-derived representations do not bias our results, we excluded text and gene network-based representations as well as integrated PRESAGE gene representations, and repeated the analysis. On the remaining 24 representations, the correlation between calibrated mNN overlap and biological process AUROC remained significant (r=0.54r = 0.54, p=0.011p = 0.011).

We thank the reviewer for raising this concern and believe that these analyses support cross-modal representational agreement beyond the measured differences in annotation coverage, training data, and model characteristics.

Manuscript: We expanded Sec. 3.2 and Sec. 4.1 to clarify the common-gene evaluation and controlled agreement analysis, with additional results in Appendix B (Figures S3-S5, S7; Table S4) and Appendix C (Figures S18, S19, S23).

Q1: Examine whether the findings are robust to different neighborhood sizes and alternative representation-similarity metrics

We added evaluation at k=5,10,20,50k = 5, 10, 20, 50, and 100100. The association between calibrated mNN agreement and biological-process prediction remains significant across these settings (r=0.86r = 0.86 at k=5k = 5, 0.810.81 at k=10k = 10, and 0.500.50 at k = 100; Figures S16–S17).

We additionally evaluated calibrated centered kernel alignment (CKA). On the subset 415-gene panel, CKA correlates with biological-process prediction (r=0.61r = 0.61, p=1.6×10−4p = 1.6 \times 10^{-4}) and calibrated mNN (r=0.91r = 0.91; Figure S12), supporting a distinction between local agreement and global representational similarity.

Manuscript: We updated the similarity metrics in Sec. 3.2 and summarized their robustness in Sec. 4.1, with additional analyses in Appendices B.1, B.3, and C (Figures S12–S17).

AuthorsOpenReview ↗

Response to reviewer FdY9 (2/4)

(part 2/4)

W1: The absolute agreement between many representations from different modalities remains relatively low [...] Figure S6 shows that several cross-modal model pairs share only approximately 1–4% of their ten nearest neighbors

We agree that absolute agreement between some cross-modal representations is limited. However, this is not unique to our setting; in the original Platonic Representation Hypothesis publication, cross-modal mNN alignment reaches only 0.16 (Huh et al., 2024, Figure 3).

Our findings rely not only on the absolute overlap, but also on its significance relative to a calibrated null and its association with biological-property prediction, which persists after controlling for model and training characteristics. We therefore interpret the results as evidence of shared local gene relationships alongside complementary modality-specific information. And the gaps motivate the construction of PlantaGenet.

Manuscript: We clarified this interpretation in Sec 4.1 and Figure 2, report raw and calibrated agreement in Appendix B (Figures S1–S4), and provide the controlled analyses in Appendix C (Figures S18–S19).

Q1: Report absolute nearest-neighbor overlap alongside enrichment over random expectation

Thank you for raising this point. Following the comment from reviewer 91yn we updated the main text to report null-calibrated mNN agreement following Gröger et al. (2026). This calibration accounts for the expected overlap of randomly matched genes and provides a more comparable measure across representations. Raw mNN values before the calibration can be found in Figures S1, S2.

Manuscript: We report raw and null-calibrated mNN agreement in Appendix B.1–B.2 (Figures S1–S5) and updated the metric definition in Sec. 3.2.

W2: biological plausibility evaluation relies on relatively simple synthetic corruption procedures and does not establish that the proposed geometric scores capture biologically meaningful cellular validity [...].

Q5: a more controlled evaluation of the proposed coherence and dominant-dimensionality scores against established single-cell quality-control and biological-consistency baselines? [...] In addition, evaluate whether performance remains stable after matching real and corrupted cells for library size, detected-gene count, and other relevant expression statistics.

We expanded the evaluation to include conventional QC statistics and reference-dependent cell-scoring methods (Table S14). Importantly, our suggested scores, ρc\rho_c and DcD_c, assess coherence through pretrained gene-space geometry, without fitting a model to a reference population of cells.

Gene-shuffled profiles preserve each cell's library size and detected-gene count, yet our geometric scores detect these corruptions with near-perfect accuracy, while conventional QC metrics remain near chance. On blood-only and combined-tissue marginal and chimeric profiles, our scores also better discriminate compared to the four conventional QC statistics. Specialized reference-dependent methods perform better on some chimeric profiles, particularly in the brain setting.

Overall, our reference-free scores outperforms comparable reference-free QC statistics and stays competitive with the reference-dependent detectors (which require a real-cell background our scores never use).

These comparisons demonstrate the value of gene-space geometry beyond several basic expression statistics. Importantly, we see the scores as complementary measures of cellular coherence, and do not claim they are unique indicators of biological validity.

Manuscript: We expanded Sec, 4.3 and Appendix F.7 (Table S14). The corruption procedures are described in Sec. 3.8.

W2: the manuscript motivates this approach as a potential discriminator for generative models of cells, the absence of evaluation [...] limits the strength of this application claim.

We agree that distinguishing observed cells from synthetic corruptions does not establish the ability to assess the biological validity of generative-model outputs. Our motivation is to provide a reference-free geometric measure that could support exploration of novel cellular states, including those proposed by generative models in the future. Exploring such cases is beyond the scope of this work, which is meant to set the foundations for future work which shall include both in-silico generation as well as validation with dedicated experiments.

Manuscript: We clarified that the current cell-level evaluation is based on synthetic corruptions in Sec. 4.3 and expanded the Discussion in Sec. 5 to distinguish the demonstrated capabilities from the envisioned application to generative models, explicitly acknowledging the need for independent biological validation.

AuthorsOpenReview ↗

Response to reviewer FdY9 (3/4)

(part 3/4)

Q4: Explain why the consensus representations show substantially weaker performance in tissue-specific corrupted-cell evaluations and whether the combined blood-and-brain results are driven primarily by cross-tissue separation? [...] analyze performance as a function of the transcriptional similarity between the source cell types.

We agree that cross-tissue chimeras are easier to reject than within-tissue mixtures. We therefore evaluated chimeras generated from 406 source-cell-type pairs spanning different transcriptional similarities (Figure S33).Rejection decreases as source-type similarity increases (Spearman rs=−0.65r_s = −0.65 for coherence and −0.58−0.58 for dominant dimensionality). Varying the mixing ratio has a smaller effect (Figure S33). These findings explain part of the performance gap between combined and tissue-specific settings (Table S10)

Manuscript: We expanded the tissue-specific interpretation in Sec. 4.3, Appendix F.2 (Table S10), and added the graded-chimera analysis in Appendix F.8 (Figure S33).

Q2: The concatenation-based PlantaGenet VI representation has 768 dimensions, whereas the Procrustes representation has 128 dimensions [...] making it important to isolate [...] increased representation dimensionality.

We added a comparison of the original 768-dimensional VI concatenation with a version reduced to 128 principal components. The reduced representation achieves a higher all-task gene-property AUROC (0.889 versus 0.876; Table S8). We also evaluated this dimensionality-matched control on cell rejection and found that the reduced representation performs slightly better on blood and blood-and-brain (mean of 0.826 vs 0.822, and 0.923 vs 0.896) and slightly worse on brain (mean of 0.673 vs 0.699) (Table S12). These experiments indicate that the larger concatenated dimension is not the source for better impact performance.

Manuscript: We added the dimensionality controls to Sec. 4.2 and Appendix E.3 (Table S8) and F.4 (Table S12).

Q2: Controlled comparisons against [...] simple ensemble predictions, or consensus representations constructed from multiple models within the same modality.

Thank you for making this suggestion, we added an ensemble averaging predictions from the six individual models, alongside RNA-only and DNA-only consensus representations (Table S8).

The prediction ensemble achieves an all-task AUROC of 0.892, comparable to PlantaGenet VI Procrustes (0.886). We view this as further evidence of the value of integrating complementary biological information. We wish to clarify that while the ensemble combines task-specific predictions, PlantaGenet integrates pretrained representations into a reusable, training free, gene space. Its geometry can then be used directly to represent cells and compute the reference-free coherence and dominant-dimensionality scores (ρc\rho_c, DcD_c), without requiring a reference-cell population.

The same-modality controls also illustrate the contribution of integrating diverse biological views. RNA-4 and DNA-4 achieve lower overall gene-property performance (0.875 and 0.845, respectively), although RNA-4 performs well on localization (AUROC 0.933; Table S8). This task dependence also appears at the cell level: RNA-4 nearly matches VI Procrustes in the blood and combined blood-and-brain settings, while DNA-4 performs well in the brain setting (0.798 vs. 0.684; Table S12). These results suggest that different representations contribute complementary information, with the benefits of integration varying across tasks and tissues.

We believe that these controls support the motivation for PlantaGenet; construct a shared gene representation that uses integrated information but is training-free, not optimized for a particular task, and whose geometry supports analysis beyond gene-property prediction.

Manuscript: We introduce the same-modality controls in Sec. 3.5, summarize the integration results in Sec. 4.2, and provide detailed gene-property and cell-level comparisons in Appendices E.3 and F.4 (Tables S8 and S12).

AuthorsOpenReview ↗

Response to reviewer FdY9 (4/4)

(part 4/4)

Q2: A leave-one-modality-out analysis would also clarify which biological information sources contribute most strongly to each downstream task.

We performed leave-one-model-out (LOMO) analyses for both gene-property prediction and cell rejection (Tables S9 and S13). PlantaGenet selects one model per family, hence the LOMO represents a modality-based ablation, but we realize it can't strictly isolate modality contributions as some models overlap in sources.

Following our intuition, we notice that model contributions differ across tasks. Removing GenePT-protein produces the largest reduction in gene-property family-mean AUROC (−0.037), whereas removing DepMap CRISPR slightly improves it (+0.006; Table S9). At the cell level, removing scGPT substantially reduces combined blood-and-brain performance (mean AUROC −0.125), but improves the brain-only result (+0.063; Table S13).

Together with the top-kk analyses (Figures S27 and S32), these results strengthen the motivation for PlantaGenet integration, showing that individual representations contribute differently across gene-property and cell-level tasks.

Manuscript: We added leave-one-model-out analyses in Appendices E.4 and F.6 (Tables S9 and S13), and top-kk analyses in Appendices E.5 and F.5 (Figures S27 and S32), with the key findings summarized in Sec. 4.2.

Q3: Could the authors clarify how the gene-property prediction benchmark controls for annotation overlap and potential information leakage between pretrained representations and evaluation labels?

The gene-property benchmark introduced by Kan-Tor et al. (2024) evaluates how well pretrained gene representations capture established biological knowledge, using linear probes to predict known gene properties. Since some representations are trained using curated biological annotations, overlap between their training sources and the benchmark properties is possible. We therefore interpret improvements in benchmark performance as demonstrating the value of integrating biological information across representations, rather than annotation-independent discovery of new gene properties.

Manuscript: We clarified the benchmark's objective and potential annotation-source overlap in Sec. 3.3, and the interpretation of integration gains in Sec. 4.2.

Q3: Provide an evaluation using temporally held-out annotations, independent experimental measurements, or gene-family-disjoint splits.

We added five-fold HGNC gene-family-disjoint cross-validation to examine whether correlations between related genes contribute to benchmark performance. PlantaGenet VI Procrustes retains an all-task AUROC of 0.881, compared with 0.886 under the original split; VI concatenation changes from 0.876 to 0.856 (Table S7). The consensus therefore retains strong performance when related genes are held together.

Manuscript: We describe the family-disjoint evaluation in Sec. 3.3 and Appendix E.2 (Table S7), with the results summarized in Sec. 4.2.

Q3: report variability across cross-validation folds and statistical uncertainty for the differences between consensus and individual models.

We repeated gene-property prediction across ten CV seeds and computed nested 95% confidence intervals incorporating variability across both seeds and benchmark tasks (Table S6). Across-seed variability is small (SD ≤\leq 0.003), while task variability contributes wider confidence intervals.

PlantaGenet VI Procrustes achieves higher point estimates than all six individual models. However, its 95% CI overlaps AlphaGenome's on the all-task aggregate, and those of ESM-2 and GenePT-protein on the family-mean aggregate (Table S6). The individual models nevertheless retain task-specific strengths: GenePT-protein performs better on biological processes (0.948 vs. 0.935), and ESM-2 on protein properties (0.951 vs. 0.915), whereas VI Procrustes leads on localization, regulatory functions, and genomic properties (Table 1).

Separately, cell-level performance also varies across tissues: AlphaGenome achieves the highest brain-only mean AUROC (0.791 vs. 0.684 for VI Procrustes), whereas VI Procrustes leads on blood-only and combined blood-and-brain evaluations (0.848 and 0.938, respectively; Table S10). These results highlight the complementary, context-dependent strengths of individual representations and the consensus.

Manuscript: We added uncertainty estimates in Sec. 4.2 and Appendix E.1 (Table S6), with task-family comparisons in Table 1 and tissue-specific cell-level comparisons in Sec. 4.3 and Appendix F.2 (Table S10).

Manuscript: We added the uncertainty evaluation in Sec. 4.2 and Appendix E.1 (Table S6).

Comments on the paper

2 comments
AuthorsOpenReview ↗

Scientific contributions and new analyses (thank you for the suggestions).

We thank all three reviewers for their careful evaluations and constructive suggestions. Their feedback motivated substantial additional analyses and helped us better demonstrate the two connected contributions of this work: investigating shared gene representations across biological modalities, which motivated using complementary biological information to construct a reusable consensus gene space, PlantaGenet.

We wish to highlight that unlike prior integration approaches optimized for specific predictive tasks, PlantaGenet defines a training- and reference-free space, which is designed to serve as a foundation for exploring and engineering new biology. By integrating complementary biological knowledge into a shared geometry, we envision using it to explore and assess novel cellular states, including those proposed by generative models or arising through cell engineering, without requiring a reference population representing those states. Furthermore, its modular construction makes it easily extensible with new modalities, and also adaptable to a specific settings when required.

As initial steps toward this vision, we demonstrate the recovery of established gene properties and introduce two intuitive reference-free geometric scores ($\rho_c$, $D_c$) to assess cellular expression coherence. We evaluate these scores on deliberately corrupted single-cell profiles, demonstrating their sensitivity to disruptions in expression structure. Extending this approach to genuinely novel or engineered cellular states, and validating their biological plausibility experimentally, IS the exciting direction for our future work (:

Below we detail the new analyses and manuscript changes (marked in BLUE in the updated PDF).

1. New analyses

Representational agreement and robustness

  • Similarity metrics (FdY9, 91yn). We adopted null-calibrated mNN (Gröger et al., 2026), added CKA as a complementary metric, and examined neighborhood-size sensitivity.
  • Gene coverage and confounders sensitivity analysis (FdY9, yjQN). We extended the analysis to encompass two gene sets (smaller (415); shared between all models and a larger (16,385) shared by 23/33 models); examined annotation coverage, shared training data, model families, architectures, objectives, and modalities; and repeated analyses excluding specific representations (PRESAGE and annotation-derived).
  • Gene-level properties and model scale (91yn) We introduced covariate-matched nulls and controls for publication coverage, expression, and protein length, alongside analyses of the relationship between model size, biological-property prediction, and representational agreement.

Consensus integration and gene-property evaluation

  • Integration controls (FdY9, yjQN). We conducted extensive analysis for both gene-property and cell-level evaluations, including dimension-matched representations, a prediction ensemble, same-modality (RNA-only and DNA-only) consensus spaces, leave-one-model-out, and top-k performance.
  • Statistical robustness (FdY9, 91yn). We added repeated cross-validation, task-bootstrap confidence intervals, HGNC gene-family-disjoint evaluation, and uncertainty estimates for cell-rejection AUROCs.

Cell-level geometry

  • Cell-scoring comparisons (FdY9, yjQN). We expanded the evaluation with conventional QC metrics, atlas-PCA representations, PCA reconstruction, expression-space distance, co-expression likelihood, and Scrublet.
  • Tissue-specific evaluation (FdY9, yjQN). We assessed cell rejection within individual tissues and across 406 source-cell-type pairs with varying transcriptional similarity, including mixing-ratio sensitivity.

The corresponding methods and results are organized in Appendices A–F, with updated references throughout the manuscript.

2. Manuscript revisions

  • Abstract and Sec. 1. Introduction: Highlighting the contributions of PlantaGenet.
  • Sec. 2. Prior work: Expanded the discussion of Kan-Tor et al., PRESAGE, and PRISME, clarifying how their objectives differ from ours.
  • Sec. 3. Methods: Clarified the common-gene panels, calibrated similarity metrics, gene-property benchmark, consensus construction, and cellular corruption procedures.
  • Sec. 4.1. Representational agreement: Added the controlled analyses and refined the interpretation of shared local gene relationships.
  • Sec. 4.2. Consensus evaluation: Corrected the mixed-variant comparisons and added integration controls, ablations, family-disjoint evaluation, and uncertainty estimates.
  • Sec. 4.3. Cell-level evaluation: Added comparisons with established scoring metrics and analyses of tissue-specific performance and chimera difficulty.
  • Sec. 5. Discussion: Clarified the scope of the demonstrated cell-level results and distinguished them from prospective applications to novel or generated cellular states.

Hi reviewers,

Thank you for your reviewing service. The authors have provided fairly extensive rebuttals. Please confirm that you have read them, state which of your concerns are or are not addressed, and whether that changes your recommendations.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.