acceptodds
Under review as a conference paper at ICLR 2027

Biologically informed generation of protein levels with flow matching

Abstract

Cellular protein levels are governed by RNA and other complex factors making them hard to predict computationally. Limited protein experimental data, particularly at single cell (sc) and spatial resolution, often necessitates the use of scRNA-seq as a proxy despite poor direct correlation. Current methods estimating single cell protein levels largely use discriminative predictors while omics data is inherently stochastic. Since proteins measured are often few and widely differ across experiments, existing methods require repeated training and tuning with limited generalizability. They fail to incorporate batch and sample effects; phenotypic priors such as cell type, disease, or condition; and mechanistic priors such as structure or function encoded in sequence. Most methods fail to incorporate spatial locations of cells in tissue slices. We address these issues with RnaProFlo, a conditional generative foundation model using rectified flow matching. It generates surface or intracellular protein levels at single cell or spatial spot resolution. Protein generation is conditioned on RNA expression, cell technical and biological priors, sequence and spatial location, if present. We develop a tokenizer that modulates expression levels based on protein language model derived sequence embeddings which inherently contain structure and functional information. It enables learning from non-overlapping protein feature sets across experiments and supports generalization to proteins unseen during training. Batch, disease, cell type and data source abstract metadata are embedded using a large language model (LLM) and globally condition the transformer ensuring dataset specific technical effects and biological conditions are incorporated. We further introduce gated graph structured token level cross attention with spatial distance bias during fine-tuning to enable generation of spatially informed proteomic data despite limited spatial training data. RnaProFlo outperforms other methods on paired single cell and spatial RNA-protein data effectively predicting held-out protein levels. Detailed ablation tests, zero-shot and fine-tuning experiments across seen and unseen proteins, tissues, biological conditions, sequencing technologies and species, collectively show RnaProFlo's superiority in estimating cellular and spatial protein from scRNA-seq or spatial transcriptomics (ST) expression and metadata.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
flow matching, single cell, spatial, proteomics, transformers

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_M76nOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper presents RnaProFlo, a conditional generative model that produces single-cell and spatial spot protein levels from paired RNA expression using rectified flow matching in a learned latent token space. Gene and protein tokens are built by modulating expression embeddings with FiLM parameters derived from ESM2 sequence embeddings, which the authors argue allows transfer to proteins and genes absent from training. Dataset and cell metadata are written as text prompts, embedded with Qwen3-Embedding-8B, and injected into DiT-style transformer blocks through AdaLN-Zero. Training proceeds in three stages (encoder/decoder pretraining, latent flow matching, and LoRA joint tuning with a rollout expression loss), and a spatial variant adds gated cross-attention to neighbouring spots’ gene tokens with a Fourier distance bias. The model is trained on about 3.3M cells from 38 CITE-seq and ECCITE-seq datasets and evaluated in zero-shot and fine-tuned settings, including new species, new technologies, a perturbation screen, and spatial CITE-seq, against CAPTAIN, scTranslator, totalVI, scProTrans, and an RNA-as-proxy baseline.

Critical strengths
  • Using sequence embeddings to modulate expression gives a principled way to handle protein panels that differ across experiments, and it allows predictions for proteins with no training data (the NEAT-seq experiment), which lookup-table approaches such as CAPTAIN and totalVI cannot make.
  • Encoding heterogeneous metadata as text embeddings is a practical answer to the inconsistent metadata fields across public datasets, and the metadata ablations make some attempt to measure its contribution.
  • The model is small (1.1M parameters, about 6 hours on one GPU), and the inclusion of a discriminative variant trained under the same pipeline is a useful control that more papers should report.
Critical weaknesses
  • The claim to be the first flow-matching single-cell foundation model rests only on a recent non-peer-reviewed review [1]. The paper does not discuss earlier flow-based single-cell generative models such as CFGen.
  • The Dataset103_1 zero-shot test count (161,764) equals the training count (145,587) plus the held-out count (16,177), suggesting training cells were evaluated. Zero-shot rows in Tables 2 and 10 also match Table 1 exactly despite supposedly using smaller subsets.
  • The paper claims top performance on all datasets, but scTranslator and even the RNA baseline beat RnaProFlo on GSE194122. CAPTAIN-FT and scTranslator-FT also win on the macaque and Perturb-CITE-seq generalization tests.
  • Cell prompts include test-set cell type labels and Seurat clusters, which in CITE-seq are often derived from protein data. Baselines get no comparable information, and test-set ablations without these fields are missing.
  • Removing the niche attention changes validation cosine only from 0.945 to 0.944 (Table 14). The spatial gains likely come from fine-tuning alone.

[1] Jiang, Shiyu, et al. "The Landscape of Single-Cell Foundation Models: Design Principles, Applications, and Open Challenges." (2026).

Questions for authors
  • Since RnaProDisc matches RnaProFlo on every task, what does the flow-matching formulation add? Can you show that generated protein distributions match observed ones, for example with a distributional distance or per-cell variance calibration?
  • Why is scProTrans compared only in the zero-shot setting, and how were the abstract’s headline improvements (13.9% cosine, 48.2% MSE) computed?
3 comments
AuthorsOpenReview ↗

Addressing Critical Weaknesses

We thank the reviewer for their appreciation of our sequence-modulated expression tokenizer, which enables predictions for proteins unseen during training. We agree that our LLM-based metadata embedder overcomes inconsistencies in metadata across datasets. Finally, we thank the reviewer for their appreciation of the small size and runtime requirements of RNAProFlo and the detailed comparisons against its discriminative variant.

1. The claim to be the first flow-matching single-cell foundation model rests only on a recent non-peer-reviewed review [1]. The paper does not discuss earlier flow-based single-cell generative models such as CFGen.

We thank the reviewer for pointing us to CFGen. We have cited it in Section A.1.3 and removed our claim to be the first flow-matching single-cell foundation model.

2. The Dataset103_1 zero-shot test count (161,764) equals the training count (145,587) plus the held-out count (16,177), suggesting training cells were evaluated. Zero-shot rows in Tables 2 and 10 also match Table 1 exactly despite supposedly using smaller subsets.

We thank the reviewer for pointing out this reporting error. No training cells were used for evaluation in any of our experiments. A summary of the cell counts for the datasets in Tables 1, 2 and 10 is provided below.

FT denotes fine-tuning; test counts correspond to the evaluated cells. Dashes indicate settings not evaluated.

DatasetTotal cellsPretraining cellsZero-shot test cellsSetting 0.9: FT cellsSetting 0.9: test cellsSetting 0.99: FT cellsSetting 0.99: test cells
Dataset103_1161,764145,58716,1761,63014,52516615,844
Dataset86218,3310218,02421,806196,2142,177213,870
GSE19412270,988070,9887,09963,888——
Dataset10837,819031,7303,14728,44432030,971
Dataset78_1.31,25401,2341351,10091,217

3. The paper claims top performance on all datasets, but scTranslator and even the RNA baseline beat RnaProFlo on GSE194122. CAPTAIN-FT and scTranslator-FT also win on the macaque and Perturb-CITE-seq generalization tests.

Thank you for your comment. We agree that our claim was too broad and have revised it to state that we achieve competitive performance across diverse evaluation settings. In addition, we have substantially revised our evaluation framework across metrics and datasets, separating shared evaluations from those not shared by all methods and highlighting evaluations on proteins not seen during training.

Across the eleven sc-datasets (9 single-cell and 2 single-cell spatial) evaluated on common cell and protein panels, RNAProFlo shows broad improvements over external baselines.

  • Zero-shot: RNAProFlo exceeds scTranslator on all eleven datasets for protein cosine, Pearson, Spearman, cell cosine and relative MSE. It also exceeds scProTrans in protein Pearson and Spearman on every dataset, and CAPTAIN in protein cosine, cell cosine and relative MSE on every dataset.

  • Fine-tuned: RNAProFlo outperforms the RNA baseline on all six metrics across all eleven datasets. It exceeds scProTrans in protein Pearson and Spearman on nine datasets and CAPTAIN in protein cosine, cell cosine and relative MSE on all eleven.

  • Overall ranking after fine-tuning: Among all evaluated methods, including RNAProDisc, RNAProFlo ranks among the top two for relative MSE on ten datasets, protein cosine on nine, and each of the remaining four metrics on seven.

These results demonstrate consistent competitiveness across complementary metrics, while acknowledging that other methods lead individual comparisons.

4. Cell prompts include test-set cell type labels and Seurat clusters, which in CITE-seq are often derived from protein data. Baselines get no comparable information, and test-set ablations without these fields are missing.

We thank the reviewer for pointing this out. We reviewed the original publications to determine how cell-type annotations were obtained. Approximately 77% of training datasets and 45% of test datasets used annotations informed by protein measurements, either directly or through multimodal or reference-based analysis.

To address this concern, we removed all cell-type-associated metadata during testing and evaluation and reran the experiments, updating the tables, results and conclusions in the manuscript. In this scenario, training data contains cell types but test/evaluation data does not.

AuthorsOpenReview ↗

Addressing critical weaknesses (Continued)

5. Removing the niche attention changes validation cosine only from 0.945 to 0.944 (Table 14). The spatial gains likely come from fine-tuning alone.

The original spatial niche ablation was performed on spot-based spatial CITE-seq data. We now compare zero-shot RNAProFlo, fine-tuned RNAProFlo without niche information, and the spatial variant under matched evaluation conditions in new single-cell paired spatial multiomic (RNA–protein) datasets, using held-out protein levels in test subsets.

On renal carcinoma, fine-tuning alone reduces relative MSE from 0.774 to 0.169; adding niche attention further reduces it to 0.160, a 5.3% reduction, while improving all reported metrics. A retrained model using permuted niches obtains relative MSE 0.170, close to the no-niche model.

Thus, fine-tuning accounts for the larger gain, but the matched fine-tuning and permutation controls support a clear additional contribution from spatial neighborhood information.

AuthorsOpenReview ↗

Response to Reviewer's Questions

  • Since RnaProDisc matches RnaProFlo on every task, what does the flow-matching formulation add? Can you show that generated protein distributions match observed ones, for example with a distributional distance or per-cell variance calibration?

Utility of RNAProFlo over RNAProDisc

We expanded our study to evaluate the quality and diversity of the learned conditional protein distributions showing the benefit of stochastic protein generation over deterministic prediction.

We evaluate population distribution fidelity and conditional probabilistic predictive accuracy to assess the contribution of flow matching beyond point prediction. Comparative tables (Table 8-9) and a summarizing figure (Figure 15) is included in the manuscript along with details in A.11 and the added metrics in A.12.

RNAProFlo's population distribution fidelity

The results are shown in Table 8. Protein-wise 1-Wasserstein distance measures agreement between generated and observed marginal distributions across cells, whereas joint MMD2\mathrm{MMD}^2 compares populations of complete protein profiles. Across the evaluated flow variants, flow-based predictions achieved lower 1-Wasserstein distance than RnaProDisc on 18 of 19 datasets in zero-shot evaluation and 17 of 19 after fine-tuning. MMD2\mathrm{MMD}^2 was lower on 18 and 15 datasets, respectively, with two additional fine-tuned comparisons tied. These results indicate improved population distribution fidelity.

RNAProFlo's probabilistic predictive accuracy

The results are shown in Table 9. We additionally evaluated continuous ranked probability score (CRPS) and the multivariate energy score against each cell’s measured protein profile, retaining the correspondence between conditioning RNA and predicted proteins. Raw flow samples achieved lower CRPS and energy scores than deterministic RnaProDisc predictions on all 19 datasets in both settings. They also outperformed Gaussian-noise-augmented RnaProDisc on 18 of 19 datasets zero-shot and 16 of 19 after fine-tuning for both metrics. The reported raw-sample scores were consistently lower than those of the flow mean predictions. Together, these findings support the value of retaining sampled variability for probabilistic prediction, alongside improved population distribution fidelity.

Usage guidelines

Across twelve datasets (10 single cell, 2 single cell spatial), RNAProFlo achieved higher dataset-averaged protein cosine, Pearson and Spearman correlations than RNAProDisc by 2.10%, 18.75% and 17.69%, respectively, in zero-shot evaluation. Cell cosine was 8.98% higher and relative MSE was 11.08% lower. After fine-tuning, the corresponding improvements were 0.17%, 8.16% and 6.16% for protein cosine, Pearson and Spearman, alongside 1.31% higher cell cosine, 3.32% higher cell Spearman and 11.32% lower relative MSE. Thus, fine-tuning narrows the cosine differences while RNAProFlo retains advantages in average protein correlations and relative MSE.

RNAProFlo is the default mode for point predictions, sampling protein profiles or probabilistic predictions, given its stronger proper scoring-rule performance and population distribution fidelity. RNAProDisc provides a faster alternative for point prediction, particularly after fine-tuning, where performance differences are often small.

  • Why is scProTrans compared only in the zero-shot setting, and how were the abstract’s headline improvements (13.9% cosine, 48.2% MSE) computed?

scProTrans is now included in all relevant benchmarking experiments assessing zero-shot, fine-tuning, limited-data fine-tuning, generalizability and spatial performance.

The abstract's original headline improvements were calculated from the four zero-shot datasets in the original Table 1. For each dataset, we selected the strongest external comparator separately for cosine (highest value) and MSE (lowest value), considering CAPTAIN, scTranslator, scProTrans and totalVI, and excluding RNAProDisc. We then averaged the dataset-specific relative improvements:

ΔCos=1004∑d=14cd−cd⋆cd⋆\Delta_{\mathrm{Cos}} = \frac{100}{4}\sum_{d=1}^{4}\frac{c_d-c_d^{\star}}{c_d^{\star}}

ΔMSE=1004∑d=14ed⋆−eded⋆\Delta_{\mathrm{MSE}} = \frac{100}{4}\sum_{d=1}^{4}\frac{e_d^{\star}-e_d}{e_d^{\star}}

Here, cdc_d and ede_d denote RNAProFlo's cosine similarity and MSE, respectively. The quantities cd⋆c_d^{\star} and ed⋆e_d^{\star} denote the highest cosine similarity and lowest MSE among external comparators for dataset dd. Both improvements are expressed as percentages.

This yields 13.9% higher cosine and 48.2% lower MSE. However, we have now revised our evaluations and therefore exclude these numbers.

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

The paper introduces RnaProFlo, a conditional generative model based on rectified flow matching for predicting protein abundance from single-cell RNA-seq and spatial transcriptomics. The method addresses the scarcity of paired RNA-protein measurements and the variability of protein panels across experiments. Its sequence-aware tokenizer uses ESM2-derived embeddings to modulate RNA and protein expression, while a Transformer combines protein self-attention and RNA-protein cross-attention. LLM-derived cell and dataset metadata condition the model through adaptive layer normalization. Training consists of three stages: encoder-decoder training, latent-space flow matching, and joint adaptation with an expression rollout loss. The spatial extension, RnaProFloS, adds gated spatial niche cross-attention with Fourier distance bias. The model is pretrained on approximately 3.3 million cells from 38 datasets and evaluated in zero-shot and fine-tuned settings across tissues, sequencing technologies, species, perturbations, and unseen proteins. The authors compare its predictive performance with existing methods and RnaProDisc, a discriminative variant, and present ablations of the main architectural and training components.

Critical strengths
  1. RnaProFlo achieves substantial improvements over established RNA-to-protein prediction methods across several datasets, with good results in both zero-shot and fine-tuned settings. The evaluation covers different tissues, sequencing technologies, species, perturbations, and spatial proteomics. The ability to predict proteins absent from training is particularly interesting, although this capability is demonstrated on a relatively small panel of seven proteins in a single dataset.

  2. The proposed tokenization approach uses protein sequence embeddings and shared expression-modulation mechanisms to accommodate different protein feature sets across experiments, without relying on a fixed protein vocabulary. This is a meaningful architectural contribution for cross-dataset prediction. The ablation results show that protein sequence representations and their FiLM modulation play an important role in the model's predictive performance.

Critical weaknesses
  1. Although the paper motivates flow matching through the stochastic nature of RNA-to-protein relationships, the experiments primarily assess point-prediction accuracy using cosine similarity and MSE computed from the average of ten generated profiles per cell. The reported visualizations and sampling diagnostics do not establish whether individual generated samples reproduce realistic conditional protein distributions and variability. Moreover, RnaProDisc, the discriminative variant sharing essentially the same architecture, achieves comparable predictive performance, particularly after fine-tuning. Consequently, the experiments do not convincingly demonstrate the added value of flow matching over deterministic prediction, despite conditional generation being a central contribution of the paper.

  2. While the ablations demonstrate the importance of protein ESM2 embeddings and sequence-based expression modulation, other components, including RNA sequence-based expression modulation (FiLM(R)), dataset metadata embeddings, and the spatial niche attention mechanism, show relatively small improvements. The ESM(P) ablation removes the protein embeddings entirely, including protein identity information. As a result, it does not show whether the biological information captured by ESM2 improves prediction over simpler protein-identity embeddings. Similarly, the LLM-derived metadata embeddings are not compared with simpler categorical representations. For the spatial extension, the main experiments do not isolate the contribution of spatial niche cross-attention from target-dataset fine-tuning, while the spatial ablations show only modest improvements. Finally, the external baselines use different and generally more limited conditioning information, so their performance comparisons do not isolate the contribution of the proposed model architecture.

Questions for authors
  1. For the ablations in Table 5, were the models retrained after removing the corresponding components, or were the components disabled only during evaluation?
4 comments
AuthorsOpenReview ↗

Predicting proteins unseen during training and RNAProFlo vs RNAProDisc,

Thank you for recognizing the quantitative improvements in RNAProFlo over current methods, its performance in zero-shot and fine-tuned settings as well as its generalizability across tissues, sequencing technologies, species, perturbations, and spatial proteomics. We appreciate your recognition of the architectural contribution of our FiLM-based tokenization approach.

Prediction of proteins absent from training

Thank you for highlighting this capability. We have expanded the evaluation beyond the original seven-protein panel to include proteins absent from pretraining across multiple datasets. The revised results include 22 proteins in Dataset108, two in Dataset78_1.3, six intracellular proteins in GSE178707, and 27 and ten proteins in the newly added breast and renal sc-spatial datasets, respectively. We report these panels separately from proteins encountered during pretraining and distinguish zero-shot prediction from subsequent fine-tuning.

We additionally performed controlled protein-holdout experiments, excluding 1%, 10%, 20%, 50%, and 75% of protein identities from training, corresponding to 4–304 held-out proteins, and compared with scProTrans, another method that claims predictive power for proteins unseen during training. On targets supported by both methods, RNAProFlo achieves higher protein Pearson and Spearman correlations than scProTrans at every holdout fraction. We also report predictions for additional targets unsupported by scProTrans. These experiments broaden the evidence for generalization to untrained protein identities, while showing that recovering their cell-to-cell expression variation remains more challenging than predicting training-seen proteins.

Utility of RNAProFlo over RNAProDisc

We expanded our study to evaluate the quality and diversity of the learned conditional protein distributions showing the benefit of stochastic protein generation over deterministic prediction.

We evaluate population distribution fidelity and conditional probabilistic predictive accuracy to assess the contribution of flow matching beyond point prediction. Comparative tables and a summarizing figure is included in the manuscript.

RNAProFlo's population distribution fidelity

The results are shown in Table 8 of the updated manuscript. Protein-wise 1-Wasserstein distance measures agreement between generated and observed marginal distributions across cells, whereas joint MMD2\mathrm{MMD}^2 compares populations of complete protein profiles. Across the evaluated flow variants, flow-based predictions achieved lower 1-Wasserstein distance than RnaProDisc on 18 of 19 datasets in zero-shot evaluation and 17 of 19 after fine-tuning. MMD2\mathrm{MMD}^2 was lower on 18 and 15 datasets, respectively, with two additional fine-tuned comparisons tied. These results indicate improved population distribution fidelity.

RNAProFlo's probabilistic predictive accuracy

The results are shown in Table 9 of the updated manuscript. We additionally evaluated continuous ranked probability score (CRPS) and the multivariate energy score against each cell’s measured protein profile, retaining the correspondence between conditioning RNA and predicted proteins. Raw flow samples achieved lower CRPS and energy scores than deterministic RnaProDisc predictions on all 19 datasets in both settings. They also outperformed Gaussian-noise-augmented RnaProDisc on 18 of 19 datasets zero-shot and 16 of 19 after fine-tuning for both metrics. The reported raw-sample scores were consistently lower than those of the flow mean predictions. Together, these findings support the value of retaining sampled variability for probabilistic prediction, alongside improved population distribution fidelity.

Usage guidelines

Across twelve datasets (10 single cell, 2 single cell spatial), RNAProFlo achieved higher dataset-averaged protein cosine, Pearson and Spearman correlations than RNAProDisc by 2.10%, 18.75% and 17.69%, respectively, in zero-shot evaluation. Cell cosine was 8.98% higher and relative MSE was 11.08% lower. After fine-tuning, the corresponding improvements were 0.17%, 8.16% and 6.16% for protein cosine, Pearson and Spearman, alongside 1.31% higher cell cosine, 3.32% higher cell Spearman and 11.32% lower relative MSE. Thus, fine-tuning narrows the cosine differences while RNAProFlo retains advantages in average protein correlations and relative MSE.

RNAProFlo is the default mode for point predictions, sampling protein profiles or probabilistic predictions, given its stronger proper scoring-rule performance and population distribution fidelity. RNAProDisc provides a faster alternative for point prediction, particularly after fine-tuning, where performance differences are often small.

AuthorsOpenReview ↗

Ablations and spatial conditioning vs fine-tuning

Ablations

We have added controls that more directly isolate the contributions of the representations and spatial conditioning. We had previously removed various other components of the model and retrained it, isolating the contribution of each component during training and evaluation (ablation table).

Pretrained representations versus identity embeddings

We agree that zeroing ESM2 embeddings removes protein identity as well as sequence information. We therefore added identity-preserving lookup-embedding ablations and a one-hot metadata baseline. Replacing protein ESM2 embeddings with learned lookup embeddings reduces protein Pearson correlation from 0.301 to 0.215 and cell Spearman from 0.555 to 0.386, while relative MSE increases from 0.465 to 0.571. Replacing LLM-derived metadata embeddings with one-hot encodings reduces these correlations to 0.213 and 0.311, respectively, and increases relative MSE to 0.587. These comparisons support the value of the pretrained representations beyond identity information alone.

Spatial conditioning versus target-dataset adaptation

We now compare zero-shot RNAProFlo, fine-tuned RNAProFlo without niche information, and the spatial variant under matched evaluation conditions in new single-cell spatial omics datasets. On renal carcinoma, fine-tuning alone reduces relative MSE from 0.774 to 0.169; adding niche attention further reduces it to 0.160, a 5.3% reduction, while improving all reported metrics. A retrained model using permuted niches obtains relative MSE 0.170, close to the no-niche model. Thus, adaptation accounts for the larger gain, while the matched and permutation controls support an additional contribution from spatial neighborhood information.

AuthorsOpenReview ↗

Ablation clarification from manuscript

For the ablations in Table 5, were the models retrained after removing the corresponding components, or were the components disabled only during evaluation?

The components were removed during both training and evaluation. Each ablated model was retrained without the corresponding component and then evaluated under the same configuration.

Thank you for the detailed response and for the additional experiments and ablations. They address many of my initial concerns regarding the differences in performance between RnaProFlo and RnaProDisc and the interpretation of some of the ablations. The added evaluations of distributional fidelity and probabilistic predictive accuracy indeed show the benefits of RnaProFlo over the deterministic variant. The extended evaluation of proteins absent from training also provides stronger evidence for the model's generalization ability. Additionally, the new ablations on the identity-preserving lookup embeddings and one-hot metadata encodings, along with the ablations on the roles of spatial niche attention and fine-tuning in RnaProFloS, help to clarify the contributions of each component. There are still, however, specific components, namely FiLM(R), dataset metadata and, to some extent, spatial niche attention, whose benefit still remains a bit unclear, as the performance improvement in the corresponding ablations is relatively small. Overall, the additional results strengthen the evidence supporting the proposed model and clarify different aspects of its contributions. I therefore maintain my recommendation to accept this paper.

Reviewer_Tz5cOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper presents a novel conditional generative method for RNA-to-protein prediction, RnaProFlo, along with its deterministic variant RnaProDisc and spatial variant RnaProS. The framework integrates sequence embedding from protein language model (PLMs) with gene and protein expression values, incorporates condition embeddings from LLMs which contains the dataset and cell metadata to strengthen the model generalizability, and utilizes a rectified flow matching to generates the protein tokens. Although the benchmark results reported by authors show higher values for the selected metrics, I have concerns about the comparison fairness and would recommend further check of the results.

Critical strengths
  1. The method’s motivation is clear presented: existing RNA-to-protein prediction methods usually exhibit limited generalizability, and to address this limitation, the authors propose a generative framework conditioned not only on the RNA expressions, but also the cell technical and biological priors and protein sequences, adapting from PLMs and LLMs, respectively.
  2. The method details and the training procedures are clearly described. Moreover, the evaluation settings are well designed and presented, which progressing from the held-out samples to more challenging generalization contexts including unseen techniques, species, and perturbations assays.
Critical weaknesses
  1. A central contribution of the paper is modeling protein expression as a conditional distribution using flow matching. However, RnaProDisc, which replaces flow matching with a deterministic reconstruction objective, achieves very similar and sometimes better performance, particularly after fine-tuning (Tables 1–3). This raises questions about whether the generative framework is necessary for the reported prediction improvements. Moreover, the experiments primarily evaluate the average of multiple generated predictions, rather than the quality or diversity of the learned conditional protein distributions. The results do not sufficiently establish the benefit of stochastic protein generation over deterministic prediction.
  2. Another main concern is that the selected evaluation metrics, cosine similarity and MSE, are insufficient on their own to conclude that RnaProFlo outperforms existing methods. Specifically, cosine similarity compares the predicted and observed expression vector based on angle between them, without subtracting their means. Consequently, when both vectors are large positive values relative to their variation, predictions may achieve high cosine similarity despite poorly capturing variations around the mean. In contrast, Pearson correlation coefficient (PCC) measures the correlation after subtracting the respective means. MSE is also sensitive to the scales of evaluated values. Therefore, these comparisons maybe unfair if different methods use different normalizations, resulting in vectors at different scale. I therefore recommend that the authors to additionally report Pearson and Spearman correlation, particular for each protein across cells, to assess whether the predictions capture variations in protein expression. Additional metrics that support a fair comparison in this setting would also be welcome. These analyses are essential to support the manuscript’s central claim of superior performance.
Questions for authors
  1. Figure 1 does not illustrate the full pipeline, particularly how the transformer operates within the flow matching process and each component trained by the three stages.
  2. Are the contrastive loss uses pairs of the same cell’s RNA and protein token or different cells? The description in Line 247 does not match the supplementary.
  3. What RNA and protein expression normalization used?
  4. Could the authors additionally report the PCC and Spearman correlation across cells and proteins, respectively, and any additional fair metrics?
  5. For results in table 1 & 2, the proposed methods perform well on the first three datasets but not optimal on the fourth dataset. Moreover, the difference between zero-shot and finetuned versions is relatively small. Could the better performance on the first three datasets be partly explained by the presence of closely related data in the training set, such as samples from the same donors or cohorts? If so, the author should explicitly describe these relationships, as other methods may have not been pre-trained on the same datasets.
  6. In ablation study, the way to test the influence of ESM2 embeddings and LLMs derived condition embeddings is to replace the embeddings with all zeros. I don’t think that replacement is reasonable, especially for the protein token head, which will lose the protein identity. Could the authors try to replace with a simple embedding instead, such as the one-hot encoding, to still preserve the identity but lose the prior knowledge from PLMs and LLMs?
  7. In addition to the quantitative metrics, it would be informative to compare the predicted and observed protein expression distributions and spatial patterns for spatial datasets. Per-protein scatter plots across cells or spots would also help assess whether the predictions capture expression variability and reveal any systematic biases.
  8. Could the author check whether the metadata involved lead to data leakage. Specifically, any of the annotations contains the information derived from protein expression, such as the cell type? If so, those metadata should be excluded before training the model.
  9. Are the metrics calculated over the same sets of proteins for all benchmarking methods? Please clarify how proteins that cannot be predicted by a particular method are handled.
5 comments
AuthorsOpenReview ↗

Utility of RNAProFlo over RNAProDisc

We thank the reviewer for their response. We are glad the reviewer appreciated the generalizability of RNAProFlo and its ability to generate protein levels conditioned on RNA expression while additionally leveraging PLMs and LLMs to condition on protein sequences, technical and biological priors. We appreciate the positive response to our description of the method and training procedure as well as our design and presentation of experiments.

We expanded our study to evaluate the quality and diversity of the learned conditional protein distributions showing the benefit of stochastic protein generation over deterministic prediction.

We evaluate population distribution fidelity and conditional probabilistic predictive accuracy to assess the contribution of flow matching beyond point prediction. Comparative tables and a summarizing figure is included in the manuscript.

RNAProFlo's population distribution fidelity

The results are shown in Table 8. Protein-wise 1-Wasserstein distance measures agreement between generated and observed marginal distributions across cells, whereas joint MMD2\mathrm{MMD}^2 compares populations of complete protein profiles. Across the evaluated flow variants, flow-based predictions achieved lower 1-Wasserstein distance than RnaProDisc on 18 of 19 datasets in zero-shot evaluation and 17 of 19 after fine-tuning. MMD2\mathrm{MMD}^2 was lower on 18 and 15 datasets, respectively, with two additional fine-tuned comparisons tied. These results indicate improved population distribution fidelity.

RNAProFlo's probabilistic predictive accuracy

The results are shown in Table 9. We additionally evaluated continuous ranked probability score (CRPS) and the multivariate energy score against each cell’s measured protein profile, retaining the correspondence between conditioning RNA and predicted proteins. Raw flow samples achieved lower CRPS and energy scores than deterministic RnaProDisc predictions on all 19 datasets in both settings. They also outperformed Gaussian-noise-augmented RnaProDisc on 18 of 19 datasets zero-shot and 16 of 19 after fine-tuning for both metrics. The reported raw-sample scores were consistently lower than those of the flow mean predictions. Together, these findings support the value of retaining sampled variability for probabilistic prediction, alongside improved population distribution fidelity.

Usage guidelines

Across twelve studies (10 single cell, 2 single cell spatial), RNAProFlo achieved higher dataset-averaged protein cosine, Pearson and Spearman correlations than RNAProDisc by 2.10%, 18.75% and 17.69%, respectively, in zero-shot evaluation. Cell cosine was 8.98% higher and relative MSE was 11.08% lower. After fine-tuning, the corresponding improvements were 0.17%, 8.16% and 6.16% for protein cosine, Pearson and Spearman, alongside 1.31% higher cell cosine, 3.32% higher cell Spearman and 11.32% lower relative MSE. Thus, fine-tuning narrows the cosine differences while RNAProFlo retains advantages in average protein correlations and relative MSE.

RNAProFlo is the default mode for point predictions, sampling protein profiles or probabilistic predictions, given its stronger proper scoring-rule performance and population distribution fidelity. RNAProDisc provides a faster alternative for point prediction, particularly after fine-tuning, where performance differences are often small.

We added Section A.11 explaining these differences. Tables 8 and 9 contain the evaluations across MMD2\mathrm{MMD}^2, 1-WD, CRPS and energy score with a visualization of the overall superior performance of RNAProFlo on these distributional metrics in Figure 15.

AuthorsOpenReview ↗

Updated Evaluation Framework, predicting proteins unseen during training, spatial niche utility and ablations

We revised the evaluation to compare methods on common cells and protein panels, reporting protein-level cosine Cos(P)\mathrm{Cos(P)}, Pearson PCC(P)\mathrm{PCC(P)} and Spearman SCC(P)\mathrm{SCC(P)} correlations, alongside cell-level cosine Cos(C)\mathrm{Cos(C)}, Spearman SCC(C)\mathrm{SCC(C)} and relative MSE rMSE\mathrm{rMSE}.

Proteins outside the shared panel are evaluated separately as training-seen or training-unseen targets. We excluded cell-type information from evaluation metadata. We included scProTrans in all relevant evaluations. All relevant tables are updated or added to the manuscript.

Predicting Proteins Absent During Training

We expanded our evaluations by holding out 1%, 10%, 20%, 50% and 75% of proteins from all 3838 training datasets and evaluated the zero-shot performance on the held-out proteins on cells containing those proteins from the validation set.

We compare with scProTrans, the only method explicitly designed to predict protein levels completely unseen during training. RNAProFlo outperforms scProTrans in the held-out setting on protein-level Pearson and Spearman correlation, though there is a significant performance drop from the model trained on all proteins.

RNAProFlo predicts more cell–protein pairs than scProTrans, likely due to mapping issues between ADT protein names and UniProt protein identifiers. The results are in Table 19 of the updated manuscript.

Spatial Niche

We updated the spatial evaluations to include two single-cell paired RNA–protein spatial multiomic datasets from human breast cancer and renal carcinoma, in addition to the previous 9 spot-based spatial CITE-seq datasets.

The new datasets are larger and are at single-cell resolution and better highlight the advantages of spatial niche attention in the spatial variant RNAProFloS over a fine-tuned RNAProFlo (RNAProFlo-FT). Comparisons are in Table 25.

Ablations

We update the ablations to include learned lookups for RNA, protein or both (instead of ESM2 or zeroed-out embeddings) and one-hot encodings for metadata instead of LLM-generated or zeroed-out embeddings in Table 15.

Metadata ablations during training are already included in the manuscript. In the updated evaluations, cell-type associated metadata has been removed from all test/evaluation datasets since it was largely generated from the hidden protein expression in the original source publications or its origins were unclear.

AuthorsOpenReview ↗

Addressing the reviewers questions

1. Figure 1 does not illustrate the full pipeline, particularly how the transformer operates within the flow matching process and each component trained by the three stages.

We thank the reviewer for their question and have updated Figure 1 in the manuscript.

2. Are the contrastive loss uses pairs of the same cell’s RNA and protein token or different cells? The description in Line 247 does not match the supplementary.

Thank you for identifying this ambiguity. As specified in Appendix A.6, Eqs. (36)–(39), RNA and protein token representations are separately mean-pooled and ℓ2\ell_2-normalized to obtain cell-level embeddings. We compute similarities between all RNA–protein embedding pairs within a minibatch and apply a symmetric InfoNCE loss. Same-cell pairs (i=ji=j) are positives, while different-cell pairs (i≠ji\neq j) are negatives. The indices ii and jj in Lines 247–248 refer to this pairwise comparison; they do not imply that different cells are positive pairs. We have revised these lines to explicitly state the positive and negative pairing, consistent with the supplementary equations.

3. What RNA and protein expression normalization used?

We apply total-count normalization followed by log⁡(1+x)\log(1+x) transformation to RNA and protein expression separately within each dataset. Each cell is scaled to the median total count of its corresponding modality within that dataset. After normalization, we retain RNA genes with available ESM2 embeddings and select the top 2,000 highly variable genes. Protein ADTs are retained if they map to the protein embedding table. The same normalization procedure is used for the spot and single-cell spatial datasets.

4. Could the authors additionally report the PCC and Spearman correlation across cells and proteins, respectively, and any additional fair metrics?

We have revised our evaluation framework to report protein-level cosine Cos(P)\mathrm{Cos(P)}, Pearson PCC(P)\mathrm{PCC(P)} and Spearman SCC(P)\mathrm{SCC(P)} correlations alongside cell-level cosine Cos(C)\mathrm{Cos(C)}, Spearman SCC(C)\mathrm{SCC(C)} and relative MSE rMSE\mathrm{rMSE} across all datasets and benchmarked methods. We additionally compare RNAProFlo and RNAProDisc, the generative and discriminative variants of our model, with 1-Wasserstein distance (1-WD), maximum mean discrepancy MMD2\mathrm{MMD}^2, continuous ranked probability score (CRPS) and multivariate energy score.

5. For results in Tables 1 & 2, the proposed methods perform well on the first three datasets but not optimally on the fourth dataset. Moreover, the difference between zero-shot and fine-tuned versions is relatively small. Could the better performance on the first three datasets be partly explained by the presence of closely related data in the training set, such as samples from the same donors or cohorts? If so, the authors should explicitly describe these relationships, as other methods may not have been pre-trained on the same datasets.

Thank you for raising this point. We agree that training-data relationships are important for interpreting these comparisons. In RNAProFlo, Dataset103_1 uses a cell-level training/test split with 145,587 training and 16,177 test cells and therefore evaluates held-out cells from a previously encountered dataset. Related datasets (same publication, different experiment) of Dataset78_1.3 are also included in RNAProFlo's training data. These settings are different from Dataset108 and GSE194122, which are completely excluded from RNAProFlo pretraining.

Importantly, related-data exposure is not unique to RNAProFlo. scProTrans was trained by us from scratch on the same set of cells as RNAProFlo. However, CAPTAIN and scTranslator used Dataset103_1 in pretraining, so in fact their results are not zero-shot, which explains the superior performance of CAPTAIN on this dataset. CAPTAIN also used GSE194122 during pretraining, so those results are not entirely zero-shot either. We show a training exposure table here: Training-data exposure across evaluated methods

“Yes” indicates use of the dataset during training; related-data exposure is specified explicitly. “No” indicates no identified training exposure.

PoolDatasetRNAProFlo / -discCAPTAINscTranslatorscProTrans
Test setDataset103_1Same dataset, other cellsYesYesSame dataset, other cells
Dataset108NoNoNoNo
Dataset78_1.3Same study, other samplesNoNoSame study, other samples
Dataset86NoYesNoNo
MoreGSE194122NoYesNoNo
TechnologyGSE100501_1NoYesNoNo
GSE100501_2NoYesNoNo
GSE178707NoNoNoNo
SpeciesDataset76_5Same study, other speciesNoNoSame study, other species
Dataset85NoNoNoNo
Spatial (Spot)GSE213264_1–9NoNoNoNo
Spatial (Single Cells)Breast / renalNoNoNoNo

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of manuscript changes during the revision

We thank the reviewers for their valuable feedback. We have extensively revised the manuscript, expanding the evaluation framework, adding experiments and clarifying the presentation. Our underlying methodology remains unchanged.

Updated and new figures

We revised Figure 1 to illustrate the flow-matching pipeline and staged training, and added 12 new figures (Figures 4–15). Figure 4 illustrates spatial niche attention. Figures 5–14 qualitatively compare measured protein expression with fine-tuned RNAProFlo generated proteins using UMAP embeddings or spatial maps. Figure 15 summarizes the distributional-metric comparison between RNAProFlo and its discriminative variant, RNAProDisc.

Revised benchmarking framework

We added three main tables (Tables 3, 5 and 6) and nine appendix tables (Tables 8, 9, 12, 13, 15, 16, 18, 19 and 25), and updated three main tables (Tables 1, 2 and 4) and one appendix table (Table 20).

  • Tables 1, 2 and 4 report revised comparisons on shared sets of cells and proteins. We exclude cell-type-associated metadata at evaluation time because many original annotations were informed by the protein measurements being predicted. We also add three benchmarking metrics across all datasets and methods.

  • Tables 3 and 5 report predictions for proteins absent from training. Tables 12, 13 and 15 provide additional evaluations outside the shared protein panels, distinguishing training-seen from training-unseen proteins.

  • Table 6 introduces two paired single-cell spatial RNA–protein datasets: human breast cancer and renal carcinoma. We predict and evaluate held-out protein expression, extending the spatial evaluation from spot-based measurements to single-cell resolution.

  • Tables 8 and 9 compare RNAProFlo with RNAProDisc using population distribution and probabilistic prediction metrics: MMD2\mathrm{MMD}^2, 1-Wasserstein distance, CRPS and energy score.

Additional ablations and generalization experiments

  • Table 16 compares ESM2 representations with learned RNA and protein lookup embeddings, and LLM-derived metadata representations with one-hot encodings. These controls preserve identity information while testing the contribution of pretrained representations.

  • Table 18 clarifies the numbers of cells used for pretraining, fine-tuning and testing.

  • Table 19 presents controlled protein-holdout experiments comparing RNAProFlo with scProTrans across different numbers of proteins excluded from training.

  • Table 20 updates limited-data fine-tuning results after excluding cell-type-associated evaluation metadata.

  • Table 25 isolates the contribution of spatial niche attention from target-dataset fine-tuning on the new renal carcinoma dataset.

Together, these updates address the reviewers’ concerns about comparison fairness, metadata leakage, stochastic generation, unseen-protein prediction and component contributions. The revised evaluations separate shared-panel comparisons from predictions on additional targets, assess generated distributions beyond point predictions, and distinguish the benefits of spatial conditioning from fine-tuning alone.

All manuscript changes are marked in red. We believe these revisions address the reviewers’ comments and substantially strengthen the manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.