Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Cell-o1: Training LLMs to Solve Single-Cell Reasoning Puzzles with Reinforcement Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A 7B language model, trained with batch-level reinforcement learning, outperforms OpenAI's o1 on a new cell-type annotation puzzle benchmark.

desk verdict A genuinely new batch-level annotation benchmark with a strong 7B result, but the paper must audit that every batch really has N distinct cell types before the 73% gain over o1 can be trusted. read the letter →

arxiv 2506.02911 v1 pith:Q3L3P6AH submitted 2025-06-03 cs.CL cs.AIcs.CEcs.HCcs.LG

Yin Fang , Qiao Jin , Guangzhi Xiong , Bowen Jin , Xianrui Zhong , Siru Ouyang , Aidong Zhang , Jiawei Han
show 1 more author
Zhiyong Lu
This is my paper
classification cs.CLcs.AIcs.CEcs.HCcs.LG
keywords celltypeannotationsingle-cellRNA-seqbatch-levelreasoningreinforcementlearningdistillationlargelanguagemodelsPuzzlesCell-o1
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

The paper tries to establish that cell type annotation is best framed as a batch-level reasoning puzzle, not a per-cell classification, and that a 7-billion-parameter language model trained for this puzzle can beat much larger frontier models. To make the case, the authors build CellPuzzles, a benchmark in which each instance is a batch of cells from the same donor, each cell comes from a distinct cell type, and the model must assign every candidate cell type to exactly one cell while writing out its reasoning. They train Cell-o1 in two stages: supervised fine-tuning on reasoning traces distilled from a strong model, then reinforcement learning with a reward that pays only when the whole batch is correct. On the held-out test set, Cell-o1 reaches 68.5% cell-level and 32.9% batch-level accuracy, versus 64.8% and 19.0% for OpenAI's o1, a 73% relative improvement at batch level. The paper argues this shows that joint reasoning and batch-level context, not raw model scale, are what drive performance.

What carries the argument

The load-bearing object is CellPuzzles itself, defined by the one-to-one mapping requirement between $N$ cells and $N$ candidate labels, together with the batch-level reward function used in training: a format-valid answer earns $1$ only if every predicted label matches the ground truth, and any error earns $0$ (with $-1$ for invalid output). GRPO, the group-relative policy optimization algorithm, turns this sparse reward into a learning signal without a critic network by normalizing rewards within a group of rollouts. The standardized prompt with `<think>` and `<answer>` tags gives distillation, SFT, and RL a shared output space, so the reasoning trace is both the explanation and the scaffold the policy learns on.

What would settle it

Inspect the CellPuzzles construction for a random set of test batches and count any batch where two cells share the same annotated cell type or where a cell's annotated type is absent from its candidate set; a non-trivial violation rate would mean batch-level accuracy is not measuring one-to-one assignment quality. Alternatively, run Cell-o1 on batches deliberately built with duplicate cell types and observe whether the forced unique assignment produces cascading errors.

Watch

Extended reading notes

Core claim

The central claim is that the all-or-nothing batch reward is what produces the gain. The paper shows that when the same models annotate cells independently, even the strongest ones stay near 50% cell-level accuracy and almost never reconstruct a full batch correctly, because they cannot exploit mutual exclusivity or compare cells against each other. Cell-o1's pipeline, which distills o1's reasoning traces, fine-tunes on them for a cold start, then runs GRPO with the reward $R_{\text{batch}} = \prod_i \mathbb{1}(\hat{y}_i = y_i)$ within a strict format, yields a 7B model that outperforms every baseline, including o1 itself, and whose traces exhibit self-correction and an easy-cases-first ordering. The paper interprets this as evidence that a small model can emulate expert annotation workflows when the training signal is aligned with the batch-level objective.

Load-bearing premise

The benchmark assumes that in every batch, the $N$ cells have $N$ distinct ground-truth cell types, so the correct answer is always a perfect one-to-one matching; noisy, duplicate, or missing cell type labels in the source data would corrupt that target.

Editorial extensions

If this is right

  • Cell-o1's 73% relative improvement over o1 on batch-level accuracy positions a 7B open model ahead of frontier closed models on this specific benchmark.
  • Decomposing the task into per-cell classification drops even the best models to near-zero batch-level accuracy, so batch context is necessary for the reported performance.
  • On four unseen disease conditions, Cell-o1 maintains or extends its advantage, suggesting the learned reasoning transfers beyond the training distribution.
  • Open-ended generation without a candidate set gives Cell-o1 0% batch-level accuracy, so the constrained candidate-label format is essential to the puzzle.
  • Without the SFT cold start, GRPO training fails to make progress, indicating that distillation provides the scaffolding the RL stage needs.

Reading between the lines

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

  • If applied to real, uncurated batches that contain duplicate or ambiguous cell types, the hard one-to-one reward could penalize correct biological assignments; a soft matching reward would be a natural robustness test.
  • The same distillation-plus-sparse-reward recipe may transfer to other structured prediction problems with permutation constraints, such as coreference resolution, entity linking, or multi-object caption alignment.
  • A direct comparison against a model trained with cell-level rewards only, while keeping batch context in the prompt, would isolate how much of the gain comes from the batch reward rather than the batch context.
  • Adding retrieval of cell-type marker information during reasoning could extend the method beyond fixed candidate sets, where open-ended generation currently fails.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces CellPuzzles, a benchmark that casts single-cell RNA-seq cell type annotation as a batch-level reasoning task: given N cells from the same donor, their top expressed genes, and a shuffled candidate set of N cell types, a model must assign each cell a unique label and provide a reasoning trace. The authors then propose Cell-o1, a 7B LLM (Qwen2.5-7B base) trained in two stages: supervised fine-tuning on o1-generated reasoning traces filtered by exact-match correctness, followed by GRPO reinforcement learning with a rule-based batch-level reward. On a held-out test set of 1,095 batches, Cell-o1 reaches 0.6849 cell-level and 0.3288 batch-level accuracy, compared with 0.6479 and 0.1900 for OpenAI's o1, a 73.05% relative improvement in batch-level accuracy. The paper also reports zero-shot generalization to four unseen disease conditions and qualitative analyses of reasoning behaviors such as self-reflection and curriculum reasoning.

Significance. If the benchmark construction is valid, CellPuzzles is a useful and timely task formulation that moves LLM-based cell type annotation from independent per-cell classification toward joint, context-aware assignment with interpretable reasoning. The two-stage training recipe (distilled SFT cold start followed by RL with an exact-match reward) is clear and reproducible, and the public release of code and data is a concrete strength. The zero-shot disease generalization results, if contamination-free, are a meaningful demonstration of robustness. The main risks are empirical rather than conceptual: the benchmark's one-to-one label assumption is not audited, the train/test split is not shown to be donor-disjoint, and the headline numbers lack uncertainty quantification.

major comments (4)
  1. [Sec. 3 and Sec. A.4] The benchmark construction asserts that each batch of N cells is sampled from N distinct cell types within the same donor (p.3, top) and that the candidate label set Y contains the ground-truth types. The appendix (Sec. A.4) describes assembly of donor metadata, top-expressed genes, and 'curated cell type annotations' from CellxGene but provides no audit for duplicate, synonymous, or inconsistent cell type labels. If two cells in a batch share the same CellxGene label, or if the same biological type appears under different label granularities (e.g., 'T cell' vs. 'CD4-positive, alpha-beta T cell'), then no one-to-one assignment can match the ground truth, and the exact-match reward in Eq. (2) is structurally zero for those instances. The reported 0.3288 batch-level accuracy would then conflate annotation skill with benchmark solvability. Please add a label-canonicalization and uniqueness audit, report the fraction of raw batches discarded or repaired, and apply the same audit explicitly to the 1,095 test batches.
  2. [Sec. 5.1 and Sec. A.2] The paper does not specify a donor-level split. Section 5.1 states that 10,155 instances are used for distillation, 6,912 for RL, and 1,095 for held-out testing, but it does not say whether the same donors or CellxGene collections can appear in both training and test splits. If a donor appears in both, the model can exploit donor-level metadata (age, sex, disease, tissue) as a shortcut, and the zero-shot disease numbers in Table 3 would be inflated. Please state the split criterion explicitly (e.g., donor-disjoint, collection-disjoint) and report the number of unique donors per split and any overlap statistics.
  3. [Tables 2 and 3] All accuracy numbers are reported as single point estimates without variance, confidence intervals, or significance tests. For the main test set of 1,095 batches, the binomial standard error of a batch-level accuracy of 0.33 is about 1.4 percentage points, so the 13.9-point gap over o1 is likely robust; however, the zero-shot disease results in Table 3 are based on much smaller samples (e.g., 539 total instances across four diseases), and the melanoma batch-level accuracy of 0.5397 could be a small-sample artifact. Please provide bootstrap confidence intervals or per-disease sample sizes, and, where feasible, evaluate across multiple seeds or test-set resamplings.
  4. [Sec. 4.3.2 and Eq. (2)] The RL reward is identical to the batch-level evaluation metric, which is aligned by design, but the paper does not analyze the reward distribution across rollouts or the variance of the advantage estimates in Eq. (1). With G=5 rollouts and an all-or-nothing binary reward, a single lucky rollout can dominate the group-normalized advantage and drive the policy update. Reporting reward variance, the fraction of rollouts with zero reward, and any observed reward-hacking behavior (e.g., format-valid but trivial assignments) would strengthen the claim that the RL stage improves biological reasoning rather than merely optimizing a sparse signal.
minor comments (5)
  1. [Figure 1] The figure embeds numeric labels such as '0.33 (+73.05%)' and '0.68 (+5.65%)' without a clear textual description of whether these are batch-level or cell-level accuracies; please add a caption explanation or refer to them in the text.
  2. [Table 5] The entry 'Age Related Macular Degeneration 7 13' appears to be a formatting error; the intended number of samples should be clarified.
  3. [Appendix D] The open-ended QA evaluation uses a MedCPT threshold of 0.8 to determine correctness, but no sensitivity analysis or rationale for this threshold is given; a brief justification or a small threshold sweep would be helpful.
  4. [Sec. 6.2 and Appendix E] The human evaluation reports 100 sampled instances and two expert annotators, but no inter-annotator agreement statistic (e.g., Cohen's kappa) is provided; adding one would make the qualitative claims more reproducible.
  5. [Sec. 4.2.1] The acceptance rate of 38.52% is reported, but it is unclear whether this is the fraction of the 8 sampled responses per instance that pass rejection filtering or the fraction of instances with at least one accepted response; please clarify.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: benchmark labels are external, distillation is filtered by exact ground-truth match, and reward-metric alignment is standard rather than a constructed equivalence.

full rationale

The paper's derivation chain is self-contained against external evidence. CellPuzzles ground-truth labels come from the CellxGene portal's curated cell type annotations, not from the model or the authors' prior work, so the benchmark is an external target. The SFT distillation step accepts an o1 trace only if its assignment exactly matches the ground-truth labels (Section 4.2.1: "A response is accepted only if it (1) adheres to the expected format and (2) yields a cell-type assignment that exactly matches the ground-truth labels"), which prevents the teacher from being the source of truth. The batch-level RL reward in Equation 2 is identical in form to the batch-level evaluation metric; this is standard reward-metric alignment for a rule-based reward and does not constitute a fitted parameter being renamed as a prediction. The one-to-one uniqueness assumption in Section 3 is a benchmark-design premise that defines the task, not a conclusion derived from itself, and the paper explicitly acknowledges in Appendix D that the constrained QA setup is "somewhat artificial." Self-citations (e.g., Search-R1, RAG-Gym, MedCPT) appear only as related-work context or evaluation tools, not as load-bearing justification for the central claim that Cell-o1 outperforms o1. No step reduces, by the paper's own equations or by self-citation, to its inputs, so no circularity is present.

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

The central claim rests on the correctness and practical realism of the CellPuzzles benchmark construction, which is described at a high level in Section 3 and detailed only in the supplementary. No numeric constant is fitted to the data for the main claim; the free parameters are design choices M and N, plus an auxiliary MedCPT threshold.

free parameters (3)
  • M (number of top expressed genes per cell)
    Hand-chosen input truncation; the numeric value is never stated, but all prompts use 'top expressed genes'. It determines the signal available to the model.
  • N (batch size range) = 8 to 15
    Benchmark instances sample 8 to 15 cells per batch; larger N empirically lowers accuracy (Figure 4A), so the difficulty is partly set by this range.
  • MedCPT threshold = 0.8
    Used only in Appendix D to convert open-ended predictions into accuracy; a hand-set semantic similarity cutoff that does not affect the main constrained-QA evaluation.
assumptions (4)
  • domain assumption CellxGene ground truth cell type labels are correct and authoritative for all training, reward, and evaluation.
    The SFT labels, the RL reward (Equation 2), and the test metrics are all computed against these labels; any label noise directly corrupts the reported accuracies.
  • domain assumption Each cell in a batch is a proxy for a distinct cluster centroid, and every batch contains exactly N distinct ground-truth cell types.
    The one-to-one uniqueness requirement and the candidate label set derive from this assumption (Section 3, Figure 2). Duplicate or misassigned types would make exact-match rewards unattainable or arbitrary.
  • domain assumption Top expressed genes approximate the differentially expressed marker genes experts use for annotation.
    The entire input representation for each cell is a ranked list of top M genes; if these do not capture discriminative markers, the task is underdetermined.
  • domain assumption Exact-match against ground truth is a valid measure of annotation quality for the batch-level reward.
    Equation (2) uses a product of indicator functions, which ignores partial credit and semantically equivalent labels; the paper acknowledges this limitation in Appendix D.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Cell-o1: Training LLMs to Solve Single-Cell Reasoning Puzzles with Reinforcement Learning." pith.science (2026). https://pith.science/paper/Q3L3P6AH

@misc{pith2026250602911,
  author       = {Pith},
  title        = {Pith review of: Cell-o1: Training LLMs to Solve Single-Cell Reasoning Puzzles with Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q3L3P6AH}},
  note         = {Machine review of arXiv:2506.02911}
}
read the original abstract

Cell type annotation is a key task in analyzing the heterogeneity of single-cell RNA sequencing data. Although recent foundation models automate this process, they typically annotate cells independently, without considering batch-level cellular context or providing explanatory reasoning. In contrast, human experts often annotate distinct cell types for different cell clusters based on their domain knowledge. To mimic this workflow, we introduce the CellPuzzles task, where the objective is to assign unique cell types to a batch of cells. This benchmark spans diverse tissues, diseases, and donor conditions, and requires reasoning across the batch-level cellular context to ensure label uniqueness. We find that off-the-shelf large language models (LLMs) struggle on CellPuzzles, with the best baseline (OpenAI's o1) achieving only 19.0% batch-level accuracy. To fill this gap, we propose Cell-o1, a 7B LLM trained via supervised fine-tuning on distilled reasoning traces, followed by reinforcement learning with batch-level rewards. Cell-o1 achieves state-of-the-art performance, outperforming o1 by over 73% and generalizing well across contexts. Further analysis of training dynamics and reasoning behaviors provides insights into batch-level annotation performance and emergent expert-like reasoning. Code and data are available at https://github.com/ncbi-nlp/cell-o1.

Figures

Figures reproduced from arXiv: 2506.02911 by the authors.

Figure 1
Figure 1. Overview of this work. Cell-o1 achieves state-of-the-art on the CellPuzzles task. ∗ Equal Contribution. † Corresponding Author. arXiv:2506.02911v1 [cs.CL] 3 Jun 2025 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. CellPuzzles formulates cell type annotation as a batch-level reasoning task that integrates gene expression and contextual metadata, inspired by how experts annotate cells in practice. To emulate this expert-driven process, we introduce CellPuzzles, a novel benchmark shown in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Data distribution of [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Fine-grained analysis of model performance. (A) Cell-level and batch-level accuracy across [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Comparing training and inference dynamics across different configurations. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Analysis of model reasoning behavior. (A) Step-by-step reasoning traces generated by each [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Overview of data construction. Each batch consists of [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Gene-R1: Reasoning with Data-Augmented Lightweight LLMs for Gene Set Analysis

    q-bio.GN 2025-09 conditional novelty 5.0 of 10

    Gene-R1 combines domain pre-training, GPT-o1 distilled reasoning examples, and GRPO reinforcement learning so lightweight Llama models match commercial LLMs on gene set analysis.

Reference graph

Works this paper leans on

76 extracted references · 50 canonical work pages · cited by 1 Pith paper

  1. [1]

    Challenges in unsupervised clustering of single-cell rna-seq data

    Vladimir Yu Kiselev, Tallulah S Andrews, and Martin Hemberg. Challenges in unsupervised clustering of single-cell rna-seq data. Nature Reviews Genetics, 20(5):273–282, 2019

    work page 2019

  2. [2]

    Current best practices in single-cell rna-seq analysis: a tutorial

    Malte D Luecken and Fabian J Theis. Current best practices in single-cell rna-seq analysis: a tutorial. Molecular systems biology, 15(6):e8746, 2019

    work page 2019

  3. [3]

    Integrating single-cell transcriptomic data across different conditions, technologies, and species

    Andrew Butler, Paul Hoffman, Peter Smibert, Efthymia Papalexi, and Rahul Satija. Integrating single-cell transcriptomic data across different conditions, technologies, and species. Nature biotechnology, 36(5):411–420, 2018

    work page 2018

  4. [4]

    From louvain to leiden: guaranteeing well-connected communities

    Vincent A Traag, Ludo Waltman, and Nees Jan Van Eck. From louvain to leiden: guaranteeing well-connected communities. Scientific reports, 9(1):1–12, 2019

    2019

  5. [5]

    Com- prehensive integration of single-cell data

    Tim Stuart, Andrew Butler, Paul Hoffman, Christoph Hafemeister, Efthymia Papalexi, William M Mauck, Yuhan Hao, Marlon Stoeckius, Peter Smibert, and Rahul Satija. Com- prehensive integration of single-cell data. cell, 177(7):1888–1902, 2019

    work page 1902

  6. [6]

    Reference-based analysis of lung single-cell sequencing reveals a transitional profibrotic macrophage

    Dvir Aran, Agnieszka P Looney, Leqian Liu, Esther Wu, Valerie Fong, Austin Hsu, Suzanna Chak, Ram P Naikawadi, Paul J Wolters, Adam R Abate, et al. Reference-based analysis of lung single-cell sequencing reveals a transitional profibrotic macrophage. Nature immunology, 20(2):163–172, 2019

    work page 2019

  7. [7]

    Eleven grand challenges in single-cell data science

    David Lähnemann, Johannes Köster, Ewa Szczurek, Davis J McCarthy, Stephanie C Hicks, Mark D Robinson, Catalina A Vallejos, Kieran R Campbell, Niko Beerenwinkel, Ahmed Mahfouz, et al. Eleven grand challenges in single-cell data science. Genome biology, 21:1–35, 2020

    work page 2020

  8. [8]

    A comparison of automatic cell identification methods for single-cell rna sequencing data

    Tamim Abdelaal, Lieke Michielsen, Davy Cats, Dylan Hoogduin, Hailiang Mei, Marcel JT Reinders, and Ahmed Mahfouz. A comparison of automatic cell identification methods for single-cell rna sequencing data. Genome biology, 20:1–19, 2019

    work page 2019

Show all 76 references
  1. [9]

    scbert as a large-scale pretrained deep language model for cell type annotation of single-cell rna-seq data

    Fan Yang, Wenchuan Wang, Fang Wang, Yuan Fang, Duyu Tang, Junzhou Huang, Hui Lu, and Jianhua Yao. scbert as a large-scale pretrained deep language model for cell type annotation of single-cell rna-seq data. Nature Machine Intelligence, 4(10):852–866, 2022

    2022

  2. [10]

    Transfer learning enables predictions in network biology

    Christina V Theodoris, Ling Xiao, Anant Chopra, Mark D Chaffin, Zeina R Al Sayed, Matthew C Hill, Helene Mantineo, Elizabeth M Brydon, Zexian Zeng, X Shirley Liu, et al. Transfer learning enables predictions in network biology. Nature, 618(7965):616–624, 2023

    2023

  3. [11]

    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, 21(8):1481–1491, 2024

    2024

  4. [12]

    Assessing gpt-4 for cell type annotation in single-cell rna-seq analysis

    Wenpin Hou and Zhicheng Ji. Assessing gpt-4 for cell type annotation in single-cell rna-seq analysis. Nature Methods, pages 1–4, 2024

    2024

  5. [13]

    Cell2sentence: Teaching large language models the language of biology

    Daniel Levine, Syed A Rizvi, Sacha Lévy, Nazreen Pallikkavaliyaveetil, David Zhang, Xingyu Chen, Sina Ghadermarzi, Ruiming Wu, Zihe Zheng, Ivan Vrkic, et al. Cell2sentence: Teaching large language models the language of biology. In International Conference on Machine Learning,...

    2024

  6. [14]

    scelmo: Embeddings from language models are good learners for single-cell data analysis

    Tianyu Liu, Tianqi Chen, Wangjie Zheng, Xiao Luo, and Hongyu Zhao. scelmo: Embeddings from language models are good learners for single-cell data analysis. bioRxiv, pages 2023–12, 2023. 10

    2023

  7. [15]

    Simple and effective embedding model for single-cell biology built from chatgpt

    Yiqun Chen and James Zou. Simple and effective embedding model for single-cell biology built from chatgpt. Nature Biomedical Engineering, pages 1–11, 2024

    2024

  8. [16]

    Langcell: Language- cell pre-training for cell identity understanding

    Suyuan Zhao, Jiahuan Zhang, Yushuai Wu, Yizhen Luo, and Zaiqing Nie. Langcell: Language- cell pre-training for cell identity understanding. In International Conference on Machine Learning, pages 61159–61185. PMLR, 2024

    2024

  9. [17]

    Multimodal learning of transcriptomes and text enables interactive single-cell rna-seq data exploration with natural-language chats

    Moritz Schaefer, Peter Peneder, Daniel Malzl, Mihaela Peycheva, Jake Burton, Anna Hakobyan, Varun Sharma, Thomas Krausgruber, Joerg Menche, Eleni M Tomazou, et al. Multimodal learning of transcriptomes and text enables interactive single-cell rna-seq data exploration with natu...

    2024

  10. [18]

    A multi-modal ai copilot for single-cell analysis with instruction following

    Yin Fang, Xinle Deng, Kangwei Liu, Ningyu Zhang, Jingyang Qian, Penghui Yang, Xiaohui Fan, and Huajun Chen. A multi-modal ai copilot for single-cell analysis with instruction following. arXiv preprint arXiv:2501.08187, 2025

    2025 arXiv

  11. [19]

    Multimodal language modeling for high-accuracy single cell transcriptomics analysis and generation

    Yaorui Shi, Jiaqi Yang, Sihang Li, Junfeng Fang, Xiang Wang, Zhiyuan Liu, and Yang Zhang. Multimodal language modeling for high-accuracy single cell transcriptomics analysis and generation. arXiv preprint arXiv:2503.09427, 2025

    2025 arXiv

  12. [20]

    Au- tomated methods for cell type annotation on scrna-seq data

    Giovanni Pasquini, Jesus Eduardo Rojo Arias, Patrick Schäfer, and V olker Busskamp. Au- tomated methods for cell type annotation on scrna-seq data. Computational and Structural Biotechnology Journal, 19:961–969, 2021

    2021

  13. [21]

    Opening the black box: interpretable machine learning for geneticists

    Christina B Azodi, Jiliang Tang, and Shin-Han Shiu. Opening the black box: interpretable machine learning for geneticists. Trends in genetics, 36(6):442–455, 2020

    2020

  14. [22]

    Openai o1 system card

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720, 2024

    2024 arXiv

  15. [23]

    Self-reflection in llm agents: Effects on problem-solving performance

    Matthew Renze and Erhan Guven. Self-reflection in llm agents: Effects on problem-solving performance. arXiv preprint arXiv:2405.06682, 2024

    2024 arXiv

  16. [24]

    Evaluating large language models through role-guide and self-reflection: A comparative study

    Lili Zhao, Yang Wang, Qi Liu, Mengyun Wang, Wei Chen, Zhichao Sheng, and Shijin Wang. Evaluating large language models through role-guide and self-reflection: A comparative study. In The Thirteenth International Conference on Learning Representations

  17. [25]

    A mathematical model for curriculum learning for parities

    Elisabetta Cornacchia and Elchanan Mossel. A mathematical model for curriculum learning for parities. In International Conference on Machine Learning, pages 6402–6423. PMLR, 2023

    2023

  18. [26]

    On curriculum learning for commonsense reasoning

    Adyasha Maharana and Mohit Bansal. On curriculum learning for commonsense reasoning. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 983–992, 2022

    2022

  19. [27]

    Defining cell types and states with single-cell genomics

    Cole Trapnell. Defining cell types and states with single-cell genomics. Genome research, 25 (10):1491–1498, 2015

    2015

  20. [28]

    A scalable scenic workflow for single-cell gene regulatory network analysis

    Bram Van de Sande, Christopher Flerin, Kristofer Davie, Maxime De Waegeneer, Gert Hulsel- mans, Sara Aibar, Ruth Seurinck, Wouter Saelens, Robrecht Cannoodt, Quentin Rouchon, et al. A scalable scenic workflow for single-cell gene regulatory network analysis. Nature protocols, ...

    2020

  21. [29]

    sctenifoldnet: a machine learning workflow for constructing and comparing transcriptome-wide gene regulatory networks from single-cell data

    Daniel Osorio, Yan Zhong, Guanxun Li, Jianhua Z Huang, and James J Cai. sctenifoldnet: a machine learning workflow for constructing and comparing transcriptome-wide gene regulatory networks from single-cell data. Patterns, 1(9), 2020

    2020

  22. [30]

    scgen predicts single-cell perturbation responses

    Mohammad Lotfollahi, F Alexander Wolf, and Fabian J Theis. scgen predicts single-cell perturbation responses. Nature methods, 16(8):715–721, 2019

    2019

  23. [31]

    Machine learning for perturbational single-cell omics

    Yuge Ji, Mohammad Lotfollahi, F Alexander Wolf, and Fabian J Theis. Machine learning for perturbational single-cell omics. Cell Systems, 12(6):522–537, 2021. 11

    2021

  24. [32]

    scgpt: toward 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: toward building a foundation model for single-cell multi-omics using generative ai. Nature Methods, 21(8):1470–1480, 2024

    2024

  25. [33]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023

    2023 arXiv

  26. [34]

    Llama: Open and efficient foundation language models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timo- thée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023

    2023 arXiv

  27. [35]

    Large language model instruction following: A survey of progresses and challenges

    Renze Lou, Kai Zhang, and Wenpeng Yin. Large language model instruction following: A survey of progresses and challenges. Computational Linguistics, 50(3):1053–1095, 2024

    2024

  28. [36]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901, 2020

    1901

  29. [37]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837, 2022

    2022

  30. [38]

    Large language models are zero-shot reasoners

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199–22213, 2022

    2022

  31. [39]

    Cz cellxgene discover: a single-cell data platform for scalable exploration, analysis and modeling of aggregated data

    CZI Cell Science Program, Shibla Abdulla, Brian Aevermann, Pedro Assis, Seve Badajoz, Sidney M Bell, Emanuele Bezzi, Batuhan Cakir, Jim Chaffer, Signe Chambers, et al. Cz cellxgene discover: a single-cell data platform for scalable exploration, analysis and modeling of aggrega...

    2025

  32. [40]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025

    2025 arXiv

  33. [41]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024

    2024 arXiv

  34. [42]

    Qwen technical report

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023

    2023 arXiv

  35. [43]

    Proximal policy optimization algorithms

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017

    2017 arXiv

  36. [44]

    Search-r1: Training llms to reason and leverage search engines with reinforcement learning

    Bowen Jin, Hansi Zeng, Zhenrui Yue, Jinsung Yoon, Sercan Arik, Dong Wang, Hamed Za- mani, and Jiawei Han. Search-r1: Training llms to reason and leverage search engines with reinforcement learning. arXiv preprint arXiv:2503.09516, 2025

    2025 arXiv

  37. [45]

    Deepretrieval: Hacking real search engines and retrievers with large language models via reinforcement learning

    Pengcheng Jiang, Jiacheng Lin, Lang Cao, Runchu Tian, SeongKu Kang, Zifeng Wang, Jimeng Sun, and Jiawei Han. Deepretrieval: Hacking real search engines and retrievers with large language models via reinforcement learning. arXiv preprint arXiv:2503.00223, 2025

    2025 arXiv

  38. [46]

    Rag-gym: Optimizing reasoning and search agents with process supervision

    Guangzhi Xiong, Qiao Jin, Xiao Wang, Yin Fang, Haolin Liu, Yifan Yang, Fangyuan Chen, Zhixing Song, Dengyu Wang, Minjia Zhang, et al. Rag-gym: Optimizing reasoning and search agents with process supervision. arXiv preprint arXiv:2502.13957, 2025

    2025 arXiv

  39. [47]

    Research: Learning to reason with search for llms via reinforcement learning

    Mingyang Chen, Tianpeng Li, Haoze Sun, Yijie Zhou, Chenzheng Zhu, Haofen Wang, Jeff Z Pan, Wen Zhang, Huajun Chen, Fan Yang, et al. Research: Learning to reason with search for llms via reinforcement learning. arXiv preprint arXiv:2503.19470, 2025

    2025 arXiv

  40. [48]

    The role of ontologies in biological and biomedical research: a functional perspective

    Robert Hoehndorf, Paul N Schofield, and Georgios V Gkoutos. The role of ontologies in biological and biomedical research: a functional perspective. Briefings in bioinformatics, 16(6): 1069–1080, 2015. 12

    2015

  41. [49]

    Gene ontology: tool for the unification of biology

    Michael Ashburner, Catherine A Ball, Judith A Blake, David Botstein, Heather Butler, J Michael Cherry, Allan P Davis, Kara Dolinski, Selina S Dwight, Janan T Eppig, et al. Gene ontology: tool for the unification of biology. Nature genetics, 25(1):25–29, 2000

    2000

  42. [50]

    The impact of reasoning step length on large language models.arXiv preprint arXiv:2401.04925, 2024

    Mingyu Jin, Qinkai Yu, Dong Shu, Haiyan Zhao, Wenyue Hua, Yanda Meng, Yongfeng Zhang, and Mengnan Du. The impact of reasoning step length on large language models.arXiv preprint arXiv:2401.04925, 2024

    2024 arXiv

  43. [51]

    Opportunities and challenges for chatgpt and large language models in biomedicine and health

    Shubo Tian, Qiao Jin, Lana Yeganova, Po-Ting Lai, Qingqing Zhu, Xiuying Chen, Yifan Yang, Qingyu Chen, Won Kim, Donald C Comeau, et al. Opportunities and challenges for chatgpt and large language models in biomedicine and health. Briefings in Bioinformatics, 25(1):bbad493, 2024

    2024

  44. [52]

    Single cell rna sequencing of human microglia uncovers a subset associated with alzheimer’s disease

    Marta Olah, Vilas Menon, Naomi Habib, Mariko F Taga, Yiyi Ma, Christina J Yung, Maria Cimpean, Anthony Khairallah, Guillermo Coronas-Samano, Roman Sankowski, et al. Single cell rna sequencing of human microglia uncovers a subset associated with alzheimer’s disease. Nature comm...

    2020

  45. [53]

    Single-cell rna-seq analysis reveals cell subsets and gene signatures associated with rheumatoid arthritis disease activity

    Marie Binvignat, Brenda Y Miao, Camilla Wibrand, Monica M Yang, Dmitry Rychkov, Emily Flynn, Joanne Nititham, Whitney Tamaki, Umair Khan, Alexander Carvidi, et al. Single-cell rna-seq analysis reveals cell subsets and gene signatures associated with rheumatoid arthritis diseas...

    2024

  46. [54]

    High- resolution single-cell atlas reveals diversity and plasticity of tissue-resident neutrophils in non-small cell lung cancer

    Stefan Salcher, Gregor Sturm, Lena Horvath, Gerold Untergasser, Christiane Kuempers, Geor- gios Fotakis, Elisa Panizzolo, Agnieszka Martowicz, Manuel Trebo, Georg Pall, et al. High- resolution single-cell atlas reveals diversity and plasticity of tissue-resident neutrophils in...

    2022

  47. [55]

    Cells of the human intestinal tract mapped across space and time

    Rasa Elmentaite, Natsuhiko Kumasaka, Kenny Roberts, Aaron Fleming, Emma Dann, Hamish W King, Vitalii Kleshchevnikov, Monika Dabrowska, Sophie Pritchard, Liam Bolt, et al. Cells of the human intestinal tract mapped across space and time. Nature, 597(7875): 250–255, 2021

    2021

  48. [56]

    Persistent t cell unresponsiveness associated with chronic visceral leishmaniasis in hiv-coinfected patients

    Nicky de Vrij, Julia Pollmann, Antonio M Rezende, Ana V Ibarra-Meneses, Thao-Thy Pham, Wasihun Hailemichael, Mekibib Kassa, Tadfe Bogale, Roma Melkamu, Arega Yeshanew, et al. Persistent t cell unresponsiveness associated with chronic visceral leishmaniasis in hiv-coinfected pa...

    2024

  49. [57]

    Single-cell multi-omics analysis of human pancreatic islets reveals novel cellular states in type 1 diabetes

    Maria Fasolino, Gregory W Schwartz, Abhijeet R Patil, Aanchal Mongia, Maria L Golson, Yue J Wang, Ashleigh Morgan, Chengyang Liu, Jonathan Schug, Jinping Liu, et al. Single-cell multi-omics analysis of human pancreatic islets reveals novel cellular states in type 1 diabetes. N...

    2022

  50. [58]

    An integrated cell atlas of the lung in health and disease

    Lisa Sikkema, Ciro Ramírez-Suástegui, Daniel C Strobl, Tessa E Gillett, Luke Zappia, Elo Madissoon, Nikolay S Markov, Laure-Emmanuelle Zaragosi, Yuge Ji, Meshal Ansari, et al. An integrated cell atlas of the lung in health and disease. Nature medicine, 29(6):1563–1577, 2023

    2023

  51. [59]

    Single-cell transcriptomics of the human retinal pigment epithelium and choroid in health and macular degeneration

    Andrew P V oigt, Kelly Mulfaul, Nathaniel K Mullin, Miles J Flamme-Wiese, Joseph C Gi- acalone, Edwin M Stone, Budd A Tucker, Todd E Scheetz, and Robert F Mullins. Single-cell transcriptomics of the human retinal pigment epithelium and choroid in health and macular degeneratio...

    2019

  52. [60]

    An atlas of healthy and injured cell states and niches in the human kidney

    Blue B Lake, Rajasree Menon, Seth Winfree, Qiwen Hu, Ricardo Melo Ferreira, Kian Kalhor, Daria Barwinska, Edgar A Otto, Michael Ferkowicz, Dinh Diep, et al. An atlas of healthy and injured cell states and niches in the human kidney. Nature, 619(7970):585–594, 2023

    2023

  53. [61]

    Ovarian cancer mutational processes drive site-specific immune evasion

    Ignacio Vázquez-García, Florian Uhlitz, Nicholas Ceglia, Jamie LP Lim, Michelle Wu, Neeman Mohibullah, Juliana Niyazov, Arvin Eric B Ruiz, Kevin M Boehm, Viktoria Bojilova, et al. Ovarian cancer mutational processes drive site-specific immune evasion. Nature, 612(7941): 778–78...

    2022

  54. [62]

    Single-cell multiomics reveals increased plasticity, resistant populations, and stem-cell–like blasts in kmt2a-rearranged leukemia

    Changya Chen, Wenbao Yu, Fatemeh Alikarami, Qi Qiu, Chia-hui Chen, Jennifer Flournoy, Peng Gao, Yasin Uzun, Li Fang, James W Davenport, et al. Single-cell multiomics reveals increased plasticity, resistant populations, and stem-cell–like blasts in kmt2a-rearranged leukemia. Bl...

    2022

  55. [63]

    Single-cell atlas of common variable immunodefi- ciency shows germinal center-associated epigenetic dysregulation in b-cell responses

    Javier Rodríguez-Ubreva, Anna Arutyunyan, Marc Jan Bonder, Lucía Del Pino-Molina, Stephen J Clark, Carlos de la Calle-Fabregat, Luz Garcia-Alonso, Louis-François Handfield, Laura Ciudad, Eduardo Andrés-León, et al. Single-cell atlas of common variable immunodefi- ciency shows ...

    2022

  56. [64]

    Single-cell resolution characterization of myeloid-derived cell states with implication in cancer outcome

    Gabriela Rapozo Guimarães, Giovanna Resk Maklouf, Cristiane Esteves Teixeira, Leandro de Oliveira Santos, Nayara Gusmão Tessarollo, Nayara Evelin de Toledo, Alessandra Freitas Serain, Cristóvão Antunes de Lanna, Marco Antônio Pretti, Jéssica Gonçalves Vieira da Cruz, et al. Si...

    2024

  57. [65]

    Distribution-independent cell type identification for single-cell rna-seq data

    Yuyao Zhai, C Liang, and Minghua Deng. Distribution-independent cell type identification for single-cell rna-seq data. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence (IJCAI-24), pages 6143–51, 2024

    2024

  58. [66]

    Celler: A ge- nomic language model for long-tailed single-cell annotation

    Huan Zhao, Yiming Liu, Jina Yao, Ling Xiong, Zexin Zhou, and Zixing Zhang. Celler: A ge- nomic language model for long-tailed single-cell annotation. arXiv preprint arXiv:2504.00020, 2025

    2025 arXiv

  59. [67]

    Trl: Transformer reinforce- ment learning

    Leandro von Werra, Younes Belkada, Lewis Tunstall, Edward Beeching, Tristan Thrush, Nathan Lambert, Shengyi Huang, Kashif Rasul, and Quentin Gallouédec. Trl: Transformer reinforce- ment learning. https://github.com/huggingface/trl, 2020

    2020

  60. [68]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1 (2):3, 2022

    2022

  61. [69]

    Qwen2.5: A party of foundation models, September 2024

    Qwen Team. Qwen2.5: A party of foundation models, September 2024. URL https://qwenlm. github.io/blog/qwen2.5/

    2024

  62. [70]

    Qwen2 technical report

    An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng...

    2024 arXiv

  63. [71]

    Hybridflow: A flexible and efficient rlhf framework

    Guangming Sheng, Chi Zhang, Zilingfeng Ye, Xibin Wu, Wang Zhang, Ru Zhang, Yanghua Peng, Haibin Lin, and Chuan Wu. Hybridflow: A flexible and efficient rlhf framework. arXiv preprint arXiv: 2409.19256, 2024

    2024 arXiv

  64. [72]

    Gonzalez, Hao Zhang, and Ion Stoica

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large lan- guage model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Syste...

    2023

  65. [73]

    chain-of-thought

    Qiao Jin, Won Kim, Qingyu Chen, Donald C Comeau, Lana Yeganova, W John Wilbur, and Zhiyong Lu. Medcpt: Contrastive pre-trained transformers with large-scale pubmed search logs for zero-shot biomedical information retrieval. Bioinformatics, 39(11):btad651, 2023. 14 Appendix A D...

    2023

  66. [74]

    These genes alone are not sufficient for cell type identification

    Gene Expression Patterns: • Many cells express ribosomal protein genes (e.g., RPLP1, RPL41, RPL10, RPS27, RPS18, etc.), which are common housekeeping genes. These genes alone are not sufficient for cell type identification. • Some cells express immune-related genes (e.g., IGKC...

  67. [75]

    • Some cell types are more specific (e.g., IgA plasma cell, activated CD4-positive T cell), while others are broader (e.g., plasma cell, vein endothelial cell)

    Candidate Cell Types: • The list includes immune cells (e.g., T cells, B cells, plasma cells), endothelial cells, and intestinal epithelial cells (e.g., goblet cells). • Some cell types are more specific (e.g., IgA plasma cell, activated CD4-positive T cell), while others are ...

  68. [76]

    T-helper 1 cell

    Contextual Relevance: • Crohn disease is an inflammatory bowel disease, so immune cells (e.g., T cells, B cells, plasma cells) and intestinal epithelial cells (e.g., goblet cells) are likely to be present. • The ileum is rich in immune cells (e.g., germinal center B cells, pla...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.