Integrating Single-Cell Foundation Models with Graph Neural Networks for Drug Response Prediction

Jonas Balke; Ming Tang; Nikoo Salehfard; Till Rossner; Tom Seifert; Ziteng Li

REVIEW 3 major objections 6 minor 12 references

Integrating Single-Cell Foundation Models with Graph Neural Networks for Drug Response Prediction

T0 review · 3 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read scGPT embeddings improve cancer drug response prediction when swapped into a graph-network framework.

desk verdict Potentially invalid cell embeddings from a gene-ordering mismatch undermine an otherwise honest incremental comparison. read the letter →

arxiv 2504.14361 v2 pith:CI2XM2CD submitted 2025-04-19 cs.LG cs.CLq-bio.QM

Till Rossner , Ziteng Li , Jonas Balke , Nikoo Salehfard , Tom Seifert , Ming Tang This is my paper · ORCID

classification cs.LGcs.CLq-bio.QM

keywords drugresponsepredictionsingle-cellfoundationmodelsscGPTgraphneuralnetworksIC50transferlearningcancercelllines

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper asks whether cell representations from a pretrained single-cell foundation model, scGPT, can replace multi-omics profiles in an existing drug-response prediction framework. It claims they can: a pipeline that encodes drug structures with a graph neural network and cancer cell lines with scGPT embeddings predicts IC50 values (the drug concentration needed to inhibit a response by half) with higher correlation between predicted and measured values than the same pipeline using scFoundation embeddings or the original multi-omics encoder. The improvement appears across cell-line, cancer-type, and drug-level filters, extends to unseen drugs in leave-one-drug-out tests, and is accompanied by smoother training curves. The paper's bet is that pretrained single-cell knowledge compensates for the small, expensive drug-response datasets that would otherwise limit accuracy.

What carries the argument

The load-bearing element is the pretrained scGPT cell encoder: a generative transformer trained on 33 million single cells, repurposed here as a fixed embedding function that maps bulk RNA-seq gene-expression values for a cancer cell line into a 512-dimensional vector. That vector replaces the multi-omics feature set used in the original DeepCDR and is concatenated with a drug representation produced by a graph neural network over the drug's molecular graph; the combined vector is passed through a neural network with dropout and batch normalization to regress IC50. The paper's comparisons attribute the performance gap to the quality of these scGPT embeddings rather than to any change in the drug-encoding side.

What would settle it

Re-run the same three-model comparison with a randomly sampled 95/5 split, or a stratified split that preserves drug and cancer-type coverage, drawn from the full dataset and using the same 90,000-instance cap selected uniformly at random. If the scGPT-based model's Pearson correlation no longer exceeds the scFoundation-based model across the cell-line, cancer-type, and drug filters, the claimed ordering rests on the slicing artifact rather than on scGPT embeddings.

Watch

Extended reading notes

Core claim

The paper's central claim is that replacing the cell-side representation in DeepCDR—normally built from methylation, mutation, and expression data—with a 512-dimensional scGPT embedding of bulk gene expression improves cancer drug response prediction. In its evaluation, the scGPT-based model outperforms both an scFoundation-based adaptation and the original DeepCDR for IC50 prediction, as measured by Pearson correlation between predicted and observed responses at three filter levels: cell lines, cancer types, and drugs. It also shows higher and more consistent correlation gains than scFoundation in leave-one-drug-out runs, indicating better generalization to unseen drugs, and its validation score climbs steadily from about 0.80 to above 0.90 over 20 epochs while the scFoundation baseline fluctuates. The paper interprets this as evidence that scGPT's pretrained gene representations capture cellular state more effectively for this task, despite using a lower-dimensional embedding than scFoundation.

Load-bearing premise

The evaluation assumes that the 95/5 split obtained by slicing ordered data, with training capped at 90,000 instances, still produces a test set representative enough to rank the three models fairly; the paper itself notes that this slicing may have introduced bias.

Editorial extensions

If this is right

Future drug-response models can treat scGPT as a drop-in cell encoder, replacing multi-omics preprocessing with a single gene-expression input without sacrificing accuracy.
Because the advantage persists in leave-one-drug-out tests, the approach is directly applicable to screening candidate drugs that have no prior response data in the training set.
The smoother training trajectory means less sensitivity to epoch selection and could reduce computational tuning for new datasets.
Since scGPT's 512-dim embeddings beat scFoundation's 768-dim embeddings, the choice of pretraining objective may matter more than embedding size when transferring to drug response.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: if the ranking survives a randomized evaluation split, the strongest claim would shift from 'scGPT works here' to 'cell foundation model choice systematically changes transfer quality,' making model selection a standard axis of comparison in drug-response pipelines.
Editorial inference: a natural paired test is to swap the drug encoder for a molecular foundation model, which the paper lists as future work, and ask whether cell-side and drug-side gains compound.
Editorial inference: because the paper omits methylation and mutation data in all compared variants, an ablation that re-adds those modalities would reveal whether scGPT embeddings capture enough biology to make them redundant or whether the two are complementary.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Main thing to know: the headline result—scGPT beats scFoundation and DeepCDR—rests on scGPT embeddings produced from inputs prepared with scFoundation's gene list (Sections 3.2–3.3). scGPT's checkpoint uses its own vocabulary ordering. If the bulk expression vector was ordered and zero-padded to scFoundation's list, the values are permuted relative to scGPT's gene embeddings, and the resulting 512-dim vectors are not meaningful scGPT representations. The paper never says it used scGPT's vocab.json, and no code is released. This would invalidate the central comparison, not just weaken it.

What's new: it's the first public attempt to use scGPT specifically inside DeepCDR, and the comparison to scFoundation is a reasonable test. The authors also flag the non-random 90,000-instance slicing themselves, which is honest. The training stability curve is a nice extra observation.

Soft spots in proportion: the ordering issue is load-bearing and needs an explicit correction. Second, the evaluation is single-run PCC with no error bars or repeated seeds; the acknowledged slicing bias means the test set may not be representative. The leave-one-drug-out covers only 20 of 223 drugs, randomly chosen but without seeds, so it's suggestive rather than conclusive. These are fixable but currently make the margin between scGPT and scFoundation look thinner than the paper implies.

Who it's for: people working on drug response prediction who want to know whether scGPT embeddings transfer to bulk expression. If the ordering issue is resolved, it's a useful incremental data point.

Recommendation: I would not publish as is. I'd send it to peer review only if the authors first clarify the gene ordering and ideally release code; without that, the central claim is unverifiable. A serious referee could check this quickly, so if you're considering it, ask for the implementation details before investing more.

Referee Report

3 major / 6 minor

Summary. The manuscript adapts the DeepCDR drug-response framework by replacing its cell representation with embeddings from the pretrained scGPT single-cell foundation model, using bulk RNA-seq from CCLE and IC50 labels from GDSC. It compares against two baselines—DeepCDR and scFoundation, both stripped to gene expression only—using Pearson correlation across cell lines, cancer types, and drugs, plus a leave-one-drug-out experiment on 20 drugs. The paper reports that scGPT outperforms both baselines and exhibits smoother training curves, and concludes that scGPT embeddings improve cancer drug response prediction.

Significance. If the central comparison is valid, the paper provides a useful, if incremental, benchmark of single-cell foundation model embeddings for bulk drug-response prediction, and it is one of few direct comparisons of scGPT and scFoundation in this setting. The evaluation is externally grounded: CCLE/GDSC data, public pretrained checkpoints, and two implemented baselines are used, and no target-label information enters the embedding generation, so the reported gains are not circular. The authors also state that their implementation builds on publicly available codebases. However, the evaluative basis is thin: a single run, a non-random data split, no released code, and a potentially fatal ambiguity about gene ordering.

major comments (3)

[3.2–3.3] The manuscript states that gene expression data were made compatible with scGPT and scFoundation 'using a gene list provided by scFoundation' and by zero-padding to that list. scGPT's human checkpoint has its own pretrained gene vocabulary and ordering. If the CCLE expression values were ordered and zero-padded according to scFoundation's gene list and then passed to scGPT, the values would be permuted relative to the positions learned by scGPT's gene embeddings, making the resulting 512-dimensional vectors meaningless. The paper never states that scGPT's own vocab.json or gene ordering was used, and no code is released to verify this. Because the headline claim is that scGPT embeddings outperform scFoundation on exactly this representation, the paper must clarify and verify the gene-order handling before the results can be interpreted.
[3.4] The evaluation rests on a single run of each configuration on a 95/5 split in which the 90,000-instance training subset was 'selected by slicing rather than random sampling, which may have introduced bias and affected the evaluation results.' No error bars, confidence intervals, or repeated-seed results are reported for Figures 2–5. Since the scGPT-versus-scFoundation differences are described as 'small but consistent,' the absence of uncertainty estimates and the acknowledged bias mean the ranking could plausibly change under a randomized split or a different seed. I request repeated runs with different seeds, a randomized or stratified split, and reporting of per-category variability.
[3.4 and 4] The baselines are not 'original DeepCDR': Section 3.4 states that both baselines were modified to exclude methylation and mutation data, relying solely on gene expression inputs. The abstract and conclusion nevertheless claim superiority over 'the original DeepCDR framework' and describe 'the original DeepCDR model' as failing to generalize to unseen drugs. This conflates a gene-expression-only DeepCDR variant with the published multi-omics DeepCDR model. The comparisons are reasonable for isolating the cell-embedding contribution, but the wording should be corrected, and ideally the published multi-omics DeepCDR should be included or its omission explicitly justified.

minor comments (6)

[1] There are multiple typos in the introduction, including 'heterogenity', 'vaiability', 'reserach', and 'predcition'; these should be corrected.
[Figure 2] The caption says each dot represents a cell line, cancer type, or drug, and the red dashed identity line y=x is used, but the axes are not clearly identified; the reader cannot tell which model is on the x-axis and which is on the y-axis.
[Figure 4] The x-axis is described as 'the rank of drugs based on their improvement,' but it is unclear whether the rank is based on scGPT, scFoundation, or both, and no error bars or per-drug sample sizes are shown.
[3.4] The leave-one-drug-out analysis covers only 20 of 223 drugs, with no description of how the random selection was performed; a seed or selection criterion should be reported.
[4] The claim that the 90,000-instance cap 'did not seem to have too big of an impact' is unsupported because no comparison against a full-data run is provided.
[3.3] Details of the normalization and preprocessing are incomplete: the text mentions CPM and log1p, but not whether the scGPT checkpoint's own recommended preprocessing (including any gene filtering) was applied.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the central claim is an external benchmark comparison whose inputs (pretrained scGPT and scFoundation checkpoints, CCLE/GDSC data, DeepCDR baselines) are not derived from the target labels.

full rationale

The paper's central claim is empirical: integrating pretrained scGPT embeddings into DeepCDR yields higher Pearson correlation and better leave-one-drug-out performance than scFoundation-based and original DeepCDR baselines. Nothing in this claim is constructed from the IC50 test labels. The pretrained embeddings come from public checkpoints, the drug graphs and expression data come from external CCLE/GDSC sources, and both baselines are independently implemented from prior work. No parameter is fitted to a subset and then renamed a prediction, and no uniqueness theorem or self-citation is invoked to force the choice of scGPT. The acknowledged non-random 90000-instance slicing and the separate question of whether scGPT's gene vocabulary was correctly aligned are experimental validity concerns, not circular-derivation concerns: they affect whether the comparison is fair or technically correct, but they do not make the reported gain an artifact of the paper's own definitions or fitted inputs. Accordingly, no self-definitional, fitted-input, self-citation, or ansatz-smuggling step is present, and a non-circularity score of 0 is appropriate.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim depends on standard domain assumptions about the transferability of single-cell pretrained embeddings to bulk RNA-seq data and about the reliability of the CCLE/GDSC labels. The most fragile assumption is that the non-random data slice gives a fair comparison; the paper itself acknowledges this risk. No new entities are introduced. The only hand-chosen numbers are experimental settings (train fraction, instance cap, number of evaluated drugs).

free parameters (3)

train_fraction = 0.95
95% of data used for training, 5% for testing, chosen by the authors following the scFoundation evaluation setup (Section 3.4).
training_instance_cap = 90000
Training limited to 90000 instances by slicing due to memory constraints (Section 3.4); the subset selection is non-random and may bias results.
leave_one_drug_count = 20
Only 20 of 223 drugs evaluated in leave-one-drug-out due to computational limits (Section 3.4).

assumptions (4)

domain assumption Pretrained scGPT checkpoint, trained on 33 million single cells, transfers to bulk RNA-seq expression profiles from CCLE.
The method relies on scGPT embeddings of bulk CCLE data being informative for IC50 prediction; zero-padding and CPM/log1p preprocessing are applied (Section 3.3).
ad hoc to paper The non-randomly sliced 90000-instance training subset and the 95/5 split do not systematically favor scGPT over baselines.
Authors acknowledge slicing 'may have introduced bias and affected the evaluation results' (Section 3.4); the comparison's validity depends on this assumption.
domain assumption GDSC IC50 measurements are reliable ground-truth labels for drug response.
Ground truth taken from GDSC without quality filtering or uncertainty modeling (Section 3.2).
domain assumption One training run per configuration suffices to compare model performance and stability.
No multiple seeds or statistical tests; the stability comparison in Figure 5 is a single trajectory per model.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Integrating Single-Cell Foundation Models with Graph Neural Networks for Drug Response Prediction." pith.science (2026). https://pith.science/paper/CI2XM2CD

@misc{pith2026250414361,
  author       = {Pith},
  title        = {Pith review of: Integrating Single-Cell Foundation Models with Graph Neural Networks for Drug Response Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CI2XM2CD}},
  note         = {Machine review of arXiv:2504.14361}
}

abstract

AI-driven drug response prediction holds great promise for advancing personalized cancer treatment. However, the inherent heterogenity of cancer and high cost of data generation make accurate prediction challenging. In this study, we investigate whether incorporating the pretrained foundation model scGPT can enhance the performance of existing drug response prediction frameworks. Our approach builds on the DeepCDR framework, which encodes drug representations from graph structures and cell representations from multi-omics profiles. We adapt this framework by leveraging scGPT to generate enriched cell representations using its pretrained knowledge to compensate for limited amount of data. We evaluate our modified framework using IC$_{50}$ values on Pearson correlation coefficient (PCC) and a leave-one-drug out validation strategy, comparing it against the original DeepCDR framework and a prior scFoundation-based approach. scGPT not only outperforms previous approaches but also exhibits greater training stability, highlighting the value of leveraging scGPT-derived knowledge in this domain.

Figures

Figures reproduced from arXiv: 2504.14361 by the authors.

**Figure 2.** Comparison of scGPT, scFoundation and Baseline results: PCC values between predicted and actual responses [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Relationship between predicted and observed IC [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Comparison of models’ performance in the leave-one-drug-out test: PCC value difference between two [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: Comparison of PCC values per Epoch for scGPT-based model, scFoundation-based model, and the baseline [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 10 canonical work pages

[1]

Deepcdr: a hybrid graph convolutional network for predicting cancer drug response

Qiao Liu, Zhiqiang Hu, Rui Jiang, and Mu Zhou. Deepcdr: a hybrid graph convolutional network for predicting cancer drug response. Bioinformatics, 36(Supplement_2):i911–i918, 2020

work page 2020
[2]

Large-scale foundation model on single-cell transcriptomics

Minsheng Hao, Jing Gong, Xin Zeng, Chiming Liu, Yucheng Guo, Xingyi Cheng, Taifeng Wang, Jianzhu Ma, Xuegong Zhang, and Le Song. Large-scale foundation model on single-cell transcriptomics. Nature Methods, pages 1–11, 2024

work page 2024
[3]

scgpt: towards building a foundation model for single-cell multi-omics using generative ai

Haotian Cui, Chloe Wang, Hassaan Maan, Kuan Pang, Fengning Luo, Nan Duan, and Bo Wang. scgpt: towards building a foundation model for single-cell multi-omics using generative ai. Nature Methods, pages 1–11, 2024

work page 2024
[4]

scfoundation github repository - https://github.com/biomap-research/scfoundation, 2023

biomap research. scfoundation github repository - https://github.com/biomap-research/scfoundation, 2023. URL https://github.com/biomap-research/scFoundation

work page 2023
[5]

The cancer cell line encyclopedia enables predictive modelling of anticancer drug sensitivity

Jordi Barretina, Giordano Caponigro, and Nicolas Stransky. The cancer cell line encyclopedia enables predictive modelling of anticancer drug sensitivity. Nature, 483:603–607, 2012

work page 2012
[6]

A landscape of pharmacogenomic interactions in cancer

Francesco Iorio, Theo A Knijnenburg, Daniel J Vis, Graham R Bignell, Michael P Menden, Michael Schubert, Nanne Aben, Emanuel Gonçalves, Syd Barthorpe, Howard Lightfoot, Thomas Cokelaer, Patricia Greninger, Ewald van Dyk, Han Chang, Heshani de Silva, Holger Heyn, Xianming Deng, Regina K Egan, Qingsong Liu, Tatiana Mironenko, Xeni Mitropoulos, Laura Richard...

work page 2016
[7]

scgpt github repository - https://github.com/bowang-lab/scgpt, 2024

Haotian Cui. scgpt github repository - https://github.com/bowang-lab/scgpt, 2024. URL https://github.com/ bowang-lab/scGPT

work page 2024
[8]

Chemberta-2: Towards chemical foundation models, 2022

Walid Ahmad, Elana Simon, Seyone Chithrananda, Gabriel Grand, and Bharath Ramsundar. Chemberta-2: Towards chemical foundation models, 2022. URL https://arxiv.org/abs/2209.01712

arXiv 2022

Show all 12 references

[9]

Bidirectional generation of structure and properties through a single molecular foundation model

Jinho Chang and Jong Chul Ye. Bidirectional generation of structure and properties through a single molecular foundation model. Nature Communications, 15:2323, 2024

2024
[10]

Large-scale chemical language representations capture molecular structure and properties

Jerret Ross, Brian Belgodere, Vijil Chenthamarakshan, Inkit Padhi, Youssef Mroueh, and Payel Das. Large-scale chemical language representations capture molecular structure and properties. Nature Machine Intelligence, 4: 1256–1264, 2022

2022
[11]

Mole: a foundation model for molecular graphs using disentangled attention

Oscar Méndez-Lucio, Christos A Nicolaou, and Berton Earnshaw. Mole: a foundation model for molecular graphs using disentangled attention. Nature Communications, 15:9431, 2024

2024
[12]

Self-supervised graph transformer on large-scale molecular data, 2020

Yu Rong, Yatao Bian, Tingyang Xu, Weiyang Xie, Ying Wei, Wenbing Huang, and Junzhou Huang. Self-supervised graph transformer on large-scale molecular data, 2020. URL https://arxiv.org/abs/2007.02835. 7

2020 arXiv

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.

[1] [1]

Deepcdr: a hybrid graph convolutional network for predicting cancer drug response

Qiao Liu, Zhiqiang Hu, Rui Jiang, and Mu Zhou. Deepcdr: a hybrid graph convolutional network for predicting cancer drug response. Bioinformatics, 36(Supplement_2):i911–i918, 2020

work page 2020

[2] [2]

Large-scale foundation model on single-cell transcriptomics

Minsheng Hao, Jing Gong, Xin Zeng, Chiming Liu, Yucheng Guo, Xingyi Cheng, Taifeng Wang, Jianzhu Ma, Xuegong Zhang, and Le Song. Large-scale foundation model on single-cell transcriptomics. Nature Methods, pages 1–11, 2024

work page 2024

[3] [3]

scgpt: towards building a foundation model for single-cell multi-omics using generative ai

Haotian Cui, Chloe Wang, Hassaan Maan, Kuan Pang, Fengning Luo, Nan Duan, and Bo Wang. scgpt: towards building a foundation model for single-cell multi-omics using generative ai. Nature Methods, pages 1–11, 2024

work page 2024

[4] [4]

scfoundation github repository - https://github.com/biomap-research/scfoundation, 2023

biomap research. scfoundation github repository - https://github.com/biomap-research/scfoundation, 2023. URL https://github.com/biomap-research/scFoundation

work page 2023

[5] [5]

The cancer cell line encyclopedia enables predictive modelling of anticancer drug sensitivity

Jordi Barretina, Giordano Caponigro, and Nicolas Stransky. The cancer cell line encyclopedia enables predictive modelling of anticancer drug sensitivity. Nature, 483:603–607, 2012

work page 2012

[6] [6]

A landscape of pharmacogenomic interactions in cancer

Francesco Iorio, Theo A Knijnenburg, Daniel J Vis, Graham R Bignell, Michael P Menden, Michael Schubert, Nanne Aben, Emanuel Gonçalves, Syd Barthorpe, Howard Lightfoot, Thomas Cokelaer, Patricia Greninger, Ewald van Dyk, Han Chang, Heshani de Silva, Holger Heyn, Xianming Deng, Regina K Egan, Qingsong Liu, Tatiana Mironenko, Xeni Mitropoulos, Laura Richard...

work page 2016

[7] [7]

scgpt github repository - https://github.com/bowang-lab/scgpt, 2024

Haotian Cui. scgpt github repository - https://github.com/bowang-lab/scgpt, 2024. URL https://github.com/ bowang-lab/scGPT

work page 2024

[8] [8]

Chemberta-2: Towards chemical foundation models, 2022

Walid Ahmad, Elana Simon, Seyone Chithrananda, Gabriel Grand, and Bharath Ramsundar. Chemberta-2: Towards chemical foundation models, 2022. URL https://arxiv.org/abs/2209.01712

arXiv 2022