Pith. sign in

REVIEW 3 major objections 7 minor 69 references

Splitting single-cell target discovery into decision-aligned agents with structured biological evidence yields more stable, interpretable gene shortlists than general-purpose automation.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-30 11:20 UTC pith:7U3NM3EN

load-bearing objection Solid methods paper on decision-structured agents for scRNA-seq target shortlists; the CP ablation is real evidence, but stability claims need a chance/pool baseline before you lean hard on the numbers. the 3 major comments →

arxiv 2607.23821 v1 pith:7U3NM3EN submitted 2026-07-26 cs.LG q-bio.GN

SCTA: An Agentic Framework for Stable and Interpretable Target Gene Discovery from Single-Cell RNA Sequencing

Shuyu Chen , Chen Zhu , Ye Zhang , Yang Li , Qiqi Xie , Haohan Wang This is my paper
classification cs.LG q-bio.GN
keywords single-cell RNA sequencingtarget gene discoveryagentic frameworkcomputational biologydifferential expressionmulti-agent systemstherapeutic targetsanalytical stability
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Finding therapeutic target genes from single-cell RNA data is fragile: small choices in preprocessing, cell selection, and differential expression can reshuffle the final shortlist even when a pipeline still looks coherent. General-purpose analysis agents automate tools and code but do not lock the decision points that decide which genes reach validation, so their recommendations often jump across runs. This paper introduces SCTA, which decomposes target discovery into specialized agents for preprocessing, annotation, disease-relevant population selection, and evidence-constrained filtering, with unidirectional handoffs and external biological evidence. In a hereditary chronic pancreatitis study with upstream data held fixed, the full evidence configuration produced the highest run-to-run agreement among tested setups and recovered a coherent inflammatory core matching known disease mechanisms. A sympathetic reader cares because labs can validate only a few candidates; a shortlist that stays put and points to real biology is more usable than a faster list that keeps changing.

Core claim

SCTA claims that decision-centric multi-agent orchestration tailored to single-cell analysis—specialized stages, constrained tools, and multi-stream biological evidence—improves repeated-run stability and interpretability of final target gene sets relative to evidence-ablated variants and broad automation baselines, while recovering disease-relevant mechanisms such as the hereditary chronic pancreatitis inflammatory chemokine program.

What carries the argument

SCTA (Single-Cell Target Agent): a fixed-order multi-agent pipeline (Preprocessing, Annotation, Insight, Filter, plus a failure-only Reviewer) with disjoint toolboxes, unidirectional handoffs, and final prioritization grounded in differential expression, enrichment, interaction neighbors, and gene summaries.

Load-bearing premise

That higher run-to-run overlap of a fixed-size gene shortlist under frozen preprocessing, plus pathway coherence scores, is a good enough stand-in for real translational reliability of the targets.

What would settle it

Under the same fixed-artifact six-run protocol on hereditary chronic pancreatitis, the full-evidence setup fails to beat the bio-knowledge and enrichment ablations on mean pairwise Jaccard, or its recurrent genes fail to match the known CCR6–CCL20 and neutrophil-chemokine inflammatory programs for that disease.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Target-discovery agents can be scored by repeated-run consensus of the final shortlist, not only single-run plausibility.
  • Dropping network/functional knowledge or pathway enrichment both lowers stability and can redirect the agent onto a different biological program.
  • Etiology-specific immune axes (hereditary versus idiopathic chronic pancreatitis) can surface without explicit supervision of those axes.
  • When only a few candidates will be validated, slower multi-evidence runs can be preferable to faster automation that reshuffles the shortlist.
  • Other single-cell agents can adopt decision-centric stage design when reproducibility of the final targets is the goal.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Repeated-run consensus depth plus external pathway checks could become a dual gate any stochastic biomedical agent must pass before wet-lab handoff.
  • Because the ablation freezes preprocessing, the next decisive test is whether decision-centric design still wins when normalization, clustering, and differential expression are deliberately perturbed.
  • The enrichment ablation’s stable-but-reoriented gene set suggests product agents should treat numerical recurrence and biological coherence as separate failure modes, not interchangeable metrics.
  • The same orchestration pattern may transfer to other decision-sensitive omics pipelines where upstream choices silently rewrite the candidate space.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. The manuscript presents SCTA, a multi-agent LLM framework that decomposes scRNA-seq target gene discovery into five specialized agents (Preprocessing, Annotation, Insight, Filter, Reviewer) with disjoint tool sets, unidirectional handoffs, and a failure-handling loop. The central empirical claim is that SCTA's full evidence integration (STRING neighbors, gene summaries, TF/pathway enrichment) yields more stable target shortlists across repeated runs than ablated configurations (mean pairwise Jaccard 0.679±0.157 vs. 0.333 and 0.436 over six runs on hereditary chronic pancreatitis, with preprocessing artifacts frozen), and higher repeated-run overlap than Biomni across three datasets. The authors additionally show that the stable recurrent core (CCR7, CSF1, CXCL2, CXCL8 at 6/6; CCL20/CCR6 at 4/6) is consistent with the published CCR6–CCL20 axis in hereditary CP (Lee et al., Gut 2022), and that the EnrichmentFunction ablation remains recurrent but converges to a biologically distinct program (ANXA1, C5AR1) — a finding they appropriately use to argue that stability and biological coherence are complementary criteria. The paper is honest about its scope: downstream-only stability under fixed preprocessing, a single mechanistic case study for the ablation, and no wet-lab validation.

Significance. If the claims hold, the paper makes a useful contribution to the evaluation methodology of biomedical LLM agents: it shifts the success criterion from single-run plausibility to repeated-run convergence of the final decision, which is the right metric when downstream validation is expensive. Specific strengths worth naming: (1) a fair downstream-only ablation that freezes preprocessing artifacts to isolate evidence-module effects — a cleaner design than the mixed protocol the authors describe and discard; (2) six independent runs per setting with mean±std Jaccard and a full gene-selection matrix (Figure 2), which is more transparency than most agent papers provide; (3) public code; (4) a genuinely informative negative result — the EnrichmentFunction ablation shows that high recurrence can coincide with biological reorientation (ANXA1/C5AR1 program), which is arguably the most instructive finding in the paper; (5) commendably scoped claims, with explicit disclaimers that end-to-end stability and wet-lab validation are out of scope. The contribution is methodological and incremental rather than a validated discovery tool, but the framing is valuable to this community.

major comments (3)
  1. [§4.4, Table 3; §3.3] The headline stability result (Normal 0.679±0.157 vs. 0.333±0.119 and 0.436±0.153) is reported as raw mean pairwise Jaccard of fixed-size shortlists with no chance-level calibration. The expected Jaccard for random draws of size k from a candidate pool of size N is approximately k/(2N−k), which is negligible for a large pool but substantial (≥0.2) if the Filter Agent's effective pool after the Insight stage is only a few dozen genes. More importantly, adding STRING/NCBI evidence could raise Jaccard mechanically by constraining the LLM toward annotation-rich genes (a smaller effective pool) rather than by improving selection reliability — precisely the mechanism the ablation is meant to exclude. The paper never reports the candidate-pool size available to the Filter Agent, a random-selection null, or a chance-corrected agreement statistic. This is load-bearing for the central claim in §4.
  2. [§4.3.1–4.3.3, Table 2] The single-pass baseline is executed only once per dataset ("The baseline is executed once per dataset"), and Table 2 accordingly shows '–' for its overlap rate. Yet the abstract, §4.3.3, and the Conclusions repeatedly claim SCTA improves stability "relative to single-pass baselines and broad automation frameworks." No stability comparison against the single-pass baseline exists in the data; the claim is currently supported only for Biomni. Additionally, the Biomni comparison (R=10 runs) reports point overlap rates (0.33/0.26, 0.43/0.12, 0.60/0.44) with no uncertainty intervals or significance test, and the same chance-baseline issue from the ablation applies here. Note also that SCTA's own overlap on GSE193337 is 0.33 — meaning two random SCTA runs disagree on roughly two-thirds of the shortlist — which sits awkwardly next to the paper's stability framing and deserves explicit discussio
  3. [§5.1.1; §6] §5.1.1 states SCTA recovered the CCR6–CCL20 axis "without explicit supervision," and this recovery is used to externally ground the stability signal (§6: "links SCTA's stability signal to a known hereditary CP inflammatory mechanism rather than to a purely numerical recurrence artifact"). However, the Filter Agent queries NCBI Gene summaries and the LLM backend is GPT-4o, whose training data almost certainly includes the Lee et al. 2022 Gut paper and the surrounding CP literature. Recovery of a published axis is therefore consistent with literature recall rather than data-driven prioritization, and cannot serve as independent validation of the pipeline's discoveries. This does not invalidate the descriptive stability result, but it weakens the coherence argument that currently carries much of the paper's biological weight. A concrete control would be one or more of: (i) mask the disease
minor comments (7)
  1. [Figure 4 caption] Caption states consensus genes are "selected in 2/3 runs within each setting," but the ablation protocol in §3.3 uses six runs and Table 3 uses ≥3/6 thresholds. Please reconcile — if Figure 4 uses a different threshold, say so explicitly and justify.
  2. [§3.3 / Table 3] The fixed shortlist size k for the ablation experiments is never stated (k=10 is mentioned only for the baseline in §4.3.1). Since Jaccard magnitudes depend directly on k, please state k for Table 3/Figures 2–3 and confirm it is identical across settings.
  3. [§3.2, §4] Reproducibility details are thin for a stability-focused paper: LLM sampling temperature, retry budget for the Reviewer Agent, prompt templates, and per-run runtime variance are not reported. Please add these (supplement is fine) and pin a commit hash or release tag for the GitHub repository cited in §3.2.
  4. [Table 2, §4.3.2–4.3.3] Table 2 needs clarification: (a) on GSE165045, Biomni's mean enrichment score (2984.04) is ~4× SCTA's (730.05) while SCTA has more significant pathways (94 vs 55) — the direction of "better" for the score and its units/definition should be stated; (b) the overlap metric |A∩B|/|A| is asymmetric — confirm sets are equal-sized so it coincides with a symmetric measure, or use Jaccard consistently with §4.4; (c) Q3 values of 0.009 vs 0.003 (GSE165045) are near the floor and the practical meaning of such differences should be tempered.
  5. [§4.2, Table 1] Table 1 reports results from "one representative run per dataset," but the selection criterion for the representative run is unspecified (median-stability run? first run?). Given the paper's own argument that single runs are unreliable, please state how the representative run was chosen or report cross-run summaries.
  6. [Abstract] The abstract says SCTA recovers mechanisms "validated in prior studies" — since this work contributes no new validation and (per the leakage concern) recovery may reflect prior literature, "consistent with prior studies" would be more accurate.
  7. [Figure 2] Figure 2 is information-dense; the N/B/E frequency notation in row labels is explained only in the caption. Consider a legend within the panel and ordering rows by Normal-run frequency for readability.

Circularity Check

1 steps flagged

No load-bearing circular derivation: stability is measured by external repeated-run overlap; only a mild, non-central enrichment-coherence loop exists.

specific steps
  1. self definitional [§3.1 Insight/Filter Agents; §4.4 Ablation Analysis (post-hoc pathway claim)]
    "Insight Agent. Selects disease-informative cell populations and forms a candidate gene pool by combining differential expression with enrichment-level evidence. Filter Agent. Aggregates multiple evidence streams... enrichment support... Consistent with this interpretation, the Normal-stable genes are enriched for Inflammatory Response and TNF-alpha/NF-kB signaling in post-hoc pathway analysis, matching the inflammatory immune programs previously reported in hereditary CP."

    Pathway/TF enrichment is an explicit upstream selection signal for the candidate pool and final shortlist. Claiming biological coherence because the recurrent Normal genes are enriched for inflammatory/TNF–NF-κB pathways therefore partly restates the enrichment criterion used to prefer those genes, rather than an independent derivation of coherence. External Lee et al. CCR6–CCL20 recovery and Open Targets scores remain non-circular; this loop only mildly inflates the enrichment-coherence narrative, not the stability metric.

full rationale

SCTA’s central quantitative claim is empirical repeated-run stability (mean pairwise Jaccard / overlap of fixed-size shortlists under fixed preprocessing artifacts), compared across ablated evidence modules and against Biomni/baseline. That metric is not optimized, fitted, or defined in terms of the reported targets, so the stability result is not true by construction. Biological grounding likewise rests on external anchors (Lee et al. CCR6–CCL20 in hereditary CP; Open Targets; Reactome), not on a self-citation uniqueness theorem or a renamed known formula. The only mild circularity is secondary: Insight/Filter agents already use TF/pathway enrichment when forming and ranking candidates, so post-hoc statements that Normal-stable genes are enriched for Inflammatory Response / TNF–NF-κB partly restate a selection input. That does not force the Jaccard ranking, the ablation gaps, or the independent ligand–receptor recovery, and does not rise above a minor non-load-bearing loop. No fitted parameter is relabeled as a prediction; no author-unique theorem forbids alternatives. Score 1 reflects that single minor coherence loop with an otherwise self-contained empirical evaluation.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 3 invented entities

The central stability claim rests on standard single-cell and LLM-agent assumptions, author-chosen evaluation knobs (run counts, k, consensus thresholds, frozen artifacts), and the premise that structured external evidence plus stage constraints cause reproducible shortlists. No new physical entities; invented pieces are system modules. Free parameters are design hyperparameters, not curve-fits to a scientific constant.

free parameters (5)
  • target shortlist size k = 10
    Baseline and comparisons use k=10 prioritized genes; stability and consensus depend on this choice.
  • repeated-run counts R = 10 (main); 6 (ablation)
    SCTA/Biomni compared with R=10; ablation uses 6 independent downstream runs—stability estimates depend on these sample sizes.
  • consensus frequency thresholds = ≥3/6, 4/6, 6/6
    Stable genes defined at ≥3/6, 4/6, 6/6; shapes which genes count as recurrent core.
  • LLM backend and sampling = GPT-4o
    GPT-4o with stochastic decoding; temperature and seeds not fully pinned in text, yet drive run-to-run variation being measured.
  • Reviewer retry budget
    Failure loop terminates after unspecified retry budget; affects whether runs complete or abort.
axioms (6)
  • domain assumption scRNA-seq analytical choices (normalization, clustering, annotation, DE) materially reshape downstream candidate target sets.
    Stated in Introduction and Related Work citing benchmarks (Duò, Luecken, Squair, etc.); motivates the whole framework.
  • ad hoc to paper Fixed agent order, disjoint tool boxes, and unidirectional handoffs improve auditability and stability versus unconstrained general agents.
    Architecture principles in §3.2; not independently proven outside this system design.
  • domain assumption STRING neighbors, NCBI gene summaries, and TF/pathway enrichment are valid stabilizing evidence for therapeutic target prioritization.
    Filter/Insight agents in §3.1; ablation treats their removal as the experimental factor.
  • ad hoc to paper Average pairwise Jaccard/overlap of top-k lists across runs is an appropriate primary metric of translational target reliability when preprocessing is fixed.
    Evaluation protocol §3.3 and §4.3; central to the strongest claim.
  • domain assumption Merging samples without explicit batch integration preserves disease-associated differences useful for target discovery.
    Datasets §4.1 design choice; could alternatively confound biology with batch.
  • domain assumption Open Targets Q3 and Reactome enrichment counts indicate disease relevance and functional coherence of computational shortlists.
    Comparison metrics in §4.3.2.
invented entities (3)
  • SCTA multi-agent pipeline (Preprocessing, Annotation, Insight, Filter, Reviewer agents) no independent evidence
    purpose: Decompose scRNA-seq target discovery into constrained decision stages with structured evidence aggregation.
    Core system introduced in §3; software architecture rather than a physical entity.
  • Binary failure indicator I_fail and Reviewer-controlled repair loop no independent evidence
    purpose: Localize execution errors and retry without changing biological objective.
    Defined in §3.2; engineering control structure.
  • Fair downstream-only ablation protocol with frozen Pancreatitis preprocessing artifacts independent evidence
    purpose: Isolate evidence-module effects on target-selection stability.
    §3.3 experimental construct used to support the main stability result.

pith-pipeline@v1.2.0-grok45-kimik3 · 21451 in / 3954 out tokens · 71928 ms · 2026-07-30T11:20:40.417884+00:00 · methodology

0 comments
read the original abstract

Identifying therapeutic target genes from single-cell RNA sequencing (scRNA-seq) data remains a fundamental challenge in translational biology. Unlike bulk assays, scRNA-seq captures heterogeneous cellular states and rare subpopulations, but this same heterogeneity makes target discovery highly sensitive to analytical choices throughout the pipeline, including preprocessing, cell population selection, differential expression analysis, and downstream biological interpretation. As a result, existing workflows and general-purpose analysis agents often produce unstable or difficult-to-interpret target hypotheses, limiting their reliability for disease-focused discovery. We present SCTA (Single-Cell Target Agent), a decision-centric agentic framework for stable and interpretable target gene discovery from scRNA-seq data. Rather than treating analysis as a single general-purpose reasoning task, SCTA decomposes target discovery into specialized agents aligned with key decision points in the single-cell pipeline and constrains downstream reasoning with structured biological evidence. In a representative ablation study on hereditary chronic pancreatitis, we demonstrate that SCTA's full evidence integration yields the most stable target selection across independent runs among the tested configurations, while recovering biologically coherent, disease-relevant mechanisms validated in prior studies. These results suggest that decision-aware agent orchestration tailored to the structure of single-cell analysis can improve the robustness, interpretability, and practical utility of target discovery in precision medicine.

Figures

Figures reproduced from arXiv: 2607.23821 by Chen Zhu, Haohan Wang, Qiqi Xie, Shuyu Chen, Yang Li, Ye Zhang.

Figure 1
Figure 1. Figure 1: SCTA framework architecture. SCTA structures single-cell target discovery around explicit, biologically constrained [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Gene-selection matrix for hereditary CP under the fair six-run downstream-only ablation. Rows are genes recurring [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Stability summary for the fair six-run downstream-only ablation. (A) Mean pairwise Jaccard similarity across settings. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Evidence profiles for representative recurrent [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

69 extracted references · 3 canonical work pages · 1 internal anchor

  1. [1]

    Wilk, and James Zou

    Samuel Alber, Bowen Chen, Eric Sun, Alina Isakova, Aaron J. Wilk, and James Zou

  2. [2]

    Suryanarayanan Balaji, Rishikesh Magar, Yayati Jadhav, and Amir Barati Farimani

  3. [3]

    Tanya Barrett, Sarah E Wilhite, Philippe Ledoux, Chris Evangelista, Irina F Kim, Marina Tomashevsky, Katherine A Marshall, Kara H Phillippy, Paul M Sherman, Mark Holko, Alex Yefanov, Hyeseung Lee, Ni Zhang, Christine L Robertson, Natalia Serova, Sean Davis, and Alexandra Soboleva. 2013. NCBI GEO: archive for functional genomics data sets—update.Nucleic Ac...

    doi:10.1093/nar/gks1193 2013

  4. [4]

    Daniil A Boiko, Robert MacKnight, Ben Kline, and Gabe Gomes. 2023. Au- tonomous chemical research with large language models.Nature624, 7992 (2023), 570–578. doi:10.1038/s41586-023-06792-0

    doi:10.1038/s41586-023-06792-0 2023

  5. [5]

    T Boutros, A Nantel, A Emadali, G Tzimas, S Conzen, E Chevet, and PP Metrakos

  6. [6]

    Anabel Liyen Cartelle, Rachel Bocchino, Ishani Shah, Awais Ahmed, Steven D Freedman, and Sunil G Sheth. 2023. Natural History, Clinical Characteristics, and Outcomes in Idiopathic Chronic Pancreatitis.Gastro Hep Advances2, 4 (2023), 449–453

    2023

  7. [7]

    Jing Chen, Jun Zhang, Hao Yao, and Yu Li. 2025. CellTypeAgent: Trustworthy Cell Type Annotation with Large Language Models.arXiv preprint arXiv:2505.08844 (2025). https://doi.org/10.48550/arXiv.2505.08844

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2505.08844 2025

  8. [8]

    Erin Dann, Nicholas C Henderson, Sarah A Teichmann, Michael D Morgan, and John C Marioni. 2022. Differential abundance testing on single-cell data using k-nearest neighbor graphs.Nature Biotechnology40, 2 (2022), 245–253

    2022

  9. [9]

    P Di Sebastiano, Fabio Francesco di Mola, C Di Febbo, Giovanna Baccante, Ettore Porreca, Paolo Innocenti, H Friess, and MW Büchler. 2000. Expression of inter- leukin 8 (IL-8) and substance P in human chronic pancreatitis.Gut47, 3 (2000), 423–428

    2000

  10. [10]

    Clara Domínguez Conde, Chenling Xu, Lloyd B Jarvis, et al. 2022. Cross-tissue immune cell analysis reveals tissue-specific features in humans.Science376 (2022), eabl5197. doi:10.1126/science.abl5197

    doi:10.1126/science.abl5197 2022

  11. [11]

    Angelo Duò, Mark D Robinson, and Charlotte Soneson. 2018. A systematic performance evaluation of clustering methods for single-cell RNA-seq data. F1000Research7 (2018), 1141

    2018

  12. [12]

    Leonard P Freedman, Iain M Cockburn, and Timothy S Simcoe. 2015. The economics of reproducibility in preclinical research.PLoS Biology13, 6 (2015), e1002165

    2015

  13. [13]

    Qing Ge, Yifan Sheng, Jun Lu, Yuting Yang, and Ming Pan. 2025. Single-cell RNA-seq data normalization: A benchmarking study.PLOS ONE20, 12 (2025), e0335102

    2025

  14. [14]

    Marc Gillespie, Bijay Jassal, Ralf Stephan, Marija Milacic, Karen Rothfels, Andrea Senff-Ribeiro, Johannes Griss, Cristoffer Sevilla, Lisa Matthews, Chuqiao Gong, Chuan Deng, Thawfeek Varusai, Eliot Ragueneau, Yusra Haider, Bruce May, Veronica Shamovsky, Joel Weiser, Timothy Brunson, Nasim Sanati, Liam Beck- man, Xiang Shao, Antonio Fabregat, Konstantinos...

    2022

  15. [15]

    Juliane Glaubitz, Anika Wilden, Janine Golchert, Georg Homuth, Uwe Völker, Barbara M Bröker, Thomas Thiele, Markus M Lerch, Julia Mayerle, Ali A Aghdassi, et al. 2022. In mouse chronic pancreatitis CD25+ FOXP3+ regulatory T cells control pancreatic fibrosis by suppression of the type 2 immune response.Nature communications13, 1 (2022), 4502

    2022

  16. [16]

    Esteban N Gurzov, Fernanda Ortis, Latifa Bakiri, Erwin F Wagner, and Decio L Eizirik. 2008. JunB inhibits ER stress and apoptosis in pancreatic beta cells.PLoS One3, 8 (2008), e3030

    2008

  17. [17]

    Mark A Hauser and Daniel F Legler. 2016. Common and biased signaling pathways of the chemokine receptor CCR7 elicited by its ligands CCL19 and CCL21 in leukocytes.Journal of Leucocyte Biology99, 6 (2016), 869–882

    2016

  18. [18]

    I Heidegger, A Pircher, A Noureen, A Krogsdam, G Fotakis, J Goveia, and C Plattner. 2022. Single-cell and bulk RNA-seq of human tumor-benign prostate adenocarcinoma samples (GSE193337). NCBI Gene Expression Omnibus dataset. https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE193337 Expression profiling by high-throughput sequencing of prostate cancer tu...

    2022

  19. [19]

    Kathleen M Hertzer, Graham W Donald, and O Joe Hines. 2013. CXCR2: a target for pancreatic cancer treatment?Expert opinion on therapeutic targets17, 6 (2013), 667–680

    2013

  20. [20]

    Lukas Heumos, Anna C Schaar, Christopher Lance, Anastasia Litinetskaya, Felix Drost, Luke Zappia, Malte D Lücken, Daniel C Strobl, Juan Henao, Fabiola Curion, et al. 2023. Best practices for single-cell analysis across modalities.Nature Reviews Genetics24, 8 (2023), 550–572. doi:10.1038/s41576-023-00586-w

    doi:10.1038/s41576-023-00586-w 2023

  21. [21]

    PM Hobbs, GW Johnson, and YD Graham. 2016. Management of pain in chronic pancreatitis with emphasis on exogenous pancreatic enzymes.World Journal of Gastrointestinal Pharmacology and Therapeutics7, 3 (2016), 370–386. doi:10.4292/ wjgpt.v7.i3.370

    2016

  22. [22]

    Sirui Hong, Yizhang Lin, Bang Liu, Bangbang Liu, Binhao Wu, Ceyao Zhang, Chenxing Wei, Danyang Li, Jiaqi Chen, Jiayi Zhang, Jinlin Wang, Li Zhang, Lingyao Zhang, Min Yang, Mingchen Zhuge, Taicheng Guo, Tuo Zhou, Wei Tao, Xiangru Tang, Xiangtao Lu, Xiawu Zheng, Xinbing Liang, Yaying Fei, Yuheng Cheng, Zhibin Gou, Zongze Xu, and Chenglin Wu. 2024. Data Inte...

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.18679 2024

  23. [23]

    Sirui Hong, Mingchen Zhuge, Jiaqi Chen, Xiawu Zheng, Yuheng Cheng, Ceyao Zhang, Jinlin Wang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, Chenyu Ran, Lingfeng Xiao, Chenglin Wu, and Jürgen Schmidhuber. 2024. MetaGPT: Meta Programming for a Multi-Agent Collaborative Framework.arXiv preprint arXiv:2308.00352(2024). https://doi.org/10.48550/arXiv...

    Pith/arXiv arXiv doi:10.48550/arxiv.2308.00352 2024

  24. [24]

    Kangru Huang, Sheng Zhang, Hao Wang, Yihang Qu, Yu Lu, Yashar Roohani, Rui Li, Long Qiu, Guoqing Li, Jun Zhang, Dong Yin, Sandeep Marwaha, James N Carter, Xin Zhou, Matthew Wheeler, Jonathan A Bernstein, Ming Wang, Peng He, Jie Zhou, Michael Snyder, Le Cong, Aviv Regev, and Jure Leskovec. 2025. Biomni: A general-purpose biomedical AI agent.bioRxiv(2025). ...

    2025

  25. [25]

    Bomi Lee, Hyewon Namkoong, Yong Yang, Hong Huang, et al . 2021. Single- cell sequencing unveils distinct immune microenvironment in human chronic pancreatitis (GSE165045). NCBI Gene Expression Omnibus dataset. https://www. ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE165045 CITE-seq and scRNA-seq of pancreatic immune cells from chronic pancreatitis and contr...

    2021

  26. [26]

    BioKDD ’26, August 2026, TBD Chen et al

    Bomi Lee, Hong Namkoong, Yan Yang, Huang Huang, David Heller, Gregory L Szot, Mark M Davis, Sohail Z Husain, Stephen J Pandol, Melena D Bellin, et al. BioKDD ’26, August 2026, TBD Chen et al

    2026

  27. [27]

    Sirpa Leppä and Dirk Bohmann. 1999. Diverse functions of JNK signaling and c-Jun in stress response and apoptosis.Oncogene18, 45 (1999), 6158–6162

    1999

  28. [28]

    Arthur Liberzon, Aravind Subramanian, Reid Pinchback, Helga Thorvaldsdóttir, Pablo Tamayo, and Jill P Mesirov. 2011. Molecular signatures database (MSigDB) 3.0.Bioinformatics27, 12 (2011), 1739–1740. doi:10.1093/bioinformatics/btr260

    doi:10.1093/bioinformatics/btr260 2011

  29. [29]

    Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Nikita Smetanin, Robert Verkuil, Ori Kabeli, Yaniv Shmueli, Allan Dos Santos Costa, Maryam Fazel-Zarandi, Tom Sercu, Salvatore Candido, and Alexander Rives. 2023. Evolutionary-scale prediction of atomic-level protein structure with a language model.Science379, 6637 (2023), 1123–1130....

    doi:10.1126/science.ade2574 2023

  30. [30]

    Romain Lopez, Jeffrey Regier, Michael B Cole, Michael I Jordan, and Nir Yosef

  31. [31]

    Malte D Luecken, Maren Büttner, Kridsadakorn Chaichoompu, Anna Danese, Marta Interlandi, Michaela F Müller, Daniel C Strobl, Luke Zappia, Martin Dugas, Maria Colomé-Tatché, et al. 2022. Benchmarking atlas-level data integration in single-cell genomics.Nature Methods19, 1 (2022), 41–50. doi:10.1038/s41592- 021-01336-8

    doi:10.1038/s41592- 2022

  32. [32]

    Malte D Luecken and Fabian J Theis. 2019. Current best practices in single- cell RNA-seq analysis: a tutorial.Molecular Systems Biology15, 6 (2019), e8746. doi:10.15252/msb.20188746

    doi:10.15252/msb.20188746 2019

  33. [33]

    Renqian Luo, Liai Sun, Yingce Xia, Tao Qin, Sheng Zhang, Hoifung Poon, and Tie-Yan Liu. 2022. BioGPT: generative pre-trained transformer for biomedical text generation and mining.Briefings in Bioinformatics23, 6 (2022), bbac409. doi:10.1093/bib/bbac409

    doi:10.1093/bib/bbac409 2022

  34. [34]

    Bran, Sam Cox, Oliver Schilter, Carlo Baldassari, Andrew D White, and Philippe Schwaller

    Andres M. Bran, Sam Cox, Oliver Schilter, Carlo Baldassari, Andrew D White, and Philippe Schwaller. 2024. Augmenting large language models with chemistry tools.Nature Machine Intelligence6 (2024), 525–535. doi:10.1038/s42256-024- 00832-8

    doi:10.1038/s42256-024- 2024

  35. [35]

    Ali Madani, Ben Krause, Kevin Greene, Chandrasekaran Subramanian, Nicholas Mohr, Christopher Holton, Meng Liu, Shilpa Phatnani, Alex M McCandlish, Nick F Rubinsteyn, Robert J Alley, Lawrence C McGee, Navdeep Jaitly, Michael R Jones, Petras J Kundrotas, Jui-Hung Yang, and Kevin Beaumont. 2023. Large language models generate functional protein sequences acr...

    work page doi:10.1038/s41587-023-01772-5 2023

  36. [36]

    Yuren Mao, Yu Mi, Peigen Liu, Mengfei Zhang, Hanqing Liu, and Yunjun Gao

  37. [37]

    Noboru Nakatsuka, Daniel Adler, Li Jiang, et al. 2025. Improving reproducibility of differentially expressed genes in single-cell transcriptomic studies of neurode- generative diseases through meta-analysis.Nature Communications16 (2025), 7436

    2025

  38. [38]

    H C T Nguyen, Bora Baik, Seungsoo Yoon, et al. 2023. Benchmarking integration of single-cell differential expression.Nature Communications14 (2023), 1570

    2023

  39. [39]

    Daniel Ochoa, Arthur Hercules, Manuel Carmona, Dusan Suveges, Aitor Gonzalez-Uriarte, Chiara Malangone, Ariadna Miranda, Ludovic Fumis, Denise Carvalho-Silva, Miriam Spitzer, et al. 2021. Open Targets Platform: supporting systematic drug–target identification and prioritisation.Nucleic Acids Research 49, D1 (2021), D1302–D1310. doi:10.1093/nar/gkaa1027

    work page doi:10.1093/nar/gkaa1027 2021

  40. [40]

    OpenAI. 2024. GPT-4o: System Card and Technical Report. https://openai.com/ research/gpt-4o-system-card

    2024

  41. [41]

    S Park, J Lee, I Jung, and E Shin. 2020. Immunophenotyping of COVID-19 and influenza underscores type I IFN response in severe COVID-19 (GSE149689). NCBI Gene Expression Omnibus dataset. https://www.ncbi.nlm.nih.gov/geo/ query/acc.cgi?acc=GSE149689 scRNA-seq of PBMCs from COVID-19, influenza patients, and healthy controls

    2020

  42. [42]

    https://doi.org/10.48550/arXiv.2504.04698

    scAgent: Universal Single-Cell Annotation via a LLM Agent.arXiv preprint arXiv:2504.04698(2025). https://doi.org/10.48550/arXiv.2504.04698

    Pith/arXiv arXiv doi:10.48550/arxiv.2504.04698 2025

  43. [43]

    Remo C Russo, Cristiana C Garcia, Mauro M Teixeira, and Flavio A Amaral. 2014. The CXCL8/IL-8 chemokine family and its receptors in inflammatory diseases. Expert review of clinical immunology10, 5 (2014), 593–619

    2014

  44. [44]

    Wouter Saelens, Robrecht Cannoodt, Hristo Todorov, and Yvan Saeys. 2019. A comparison of single-cell trajectory inference methods.Nature Biotechnology37, 5 (2019), 547–554

    2019

  45. [45]

    Raghuwansh P Sah, Sushil K Garg, Ajay K Dixit, Vikas Dudeja, Rajinder K Dawra, and Ashok K Saluja. 2014. Endoplasmic reticulum stress is chronically activated in chronic pancreatitis.Journal of Biological Chemistry289, 40 (2014), 27551–27561

    2014

  46. [46]

    Rahul Satija, Jeffrey A Farrell, David Gennert, Alexander F Schier, and Aviv Regev. 2015. Spatial reconstruction of single-cell gene expression data. R package documentation. https://satijalab.org/seurat/

    2015

  47. [47]

    Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. Toolformer: Language models can teach themselves to use tools. InAdvances in Neural Information Processing Systems

    2023

  48. [48]

    Yifan Qi, Jun Wang, Wei Liu, Lei Yang, Xin Zhang, Zhi Wang, and Yu Gao. 2024. Scientific Hypothesis Generation with Large Language Models: Benchmarks, Methods, and Evaluations.arXiv preprint arXiv:2407.08940(2024). https://arxiv. org/abs/2407.08940

    Pith/arXiv arXiv 2024

  49. [49]

    Karan Singhal, Shekoofeh Azizi, Tao Tu, S Sara Mahdavi, Jason Wei, Hyung Won Chung, Nathan Scales, Ajay Tanwani, Heather Cole-Lewis, Stephen Pfohl, et al

  50. [50]

    Jordan W Squair, Matthieu Gautier, Claudia Kathe, Mark A Anderson, Nicholas D James, Thomas H Hutson, Rémy Hudelle, Taha Qaiser, Kaya J E Matson, Quentin Barraud, et al . 2021. Confronting false discoveries in single-cell differential expression.Nature Communications12 (2021), 5692. doi:10.1038/s41467-021- 25960-2

    doi:10.1038/s41467-021- 2021

  51. [51]

    Colin W Steele, Saadia A Karim, Mona Foth, Loveena Rishi, Joshua DG Leach, Ross J Porter, Colin Nixon, TR Jeffry Evans, C Ross Carter, Robert JB Nibbs, et al

  52. [52]

    Luyi Tian, Xueyi Dong, Saskia Freytag, Kim-Anh Le Cao, Shian Su, Abolfazl JalalAbadi, Daniela Amann-Zalcenstein, Tom S Weber, Azadeh Seidi, Jafar S Jabbari, et al. 2019. Benchmarking single cell RNA-sequencing analysis pipelines using mixture control experiments.Nature Methods16, 6 (2019), 479–487. doi:10. 1038/s41592-019-0425-8

    2019

  53. [53]

    Cole Trapnell. 2014. Monocle: Clustering, differential expression, and trajectory analysis for single-cell RNA-seq experiments. R package documentation. https: //cole-trapnell-lab.github.io/monocle-release/

    2014

  54. [54]

    Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: Language agents with verbal reinforcement learning. InAdvances in Neural Information Processing Systems

    2023

  55. [55]

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023. Self-consistency improves chain of thought reasoning in language models. InInternational Conference on Learning Representations

    2023

  56. [56]

    doi:10.1038/s41586-023-06291-2

    Large language models encode clinical knowledge.Nature620, 7972 (2023), 172–180. doi:10.1038/s41586-023-06291-2

    doi:10.1038/s41586-023-06291-2 2023

  57. [57]

    F Alexander Wolf, Philipp Angerer, and Fabian J Theis. 2018. SCANPY: large- scale single-cell gene expression data analysis. Python package documentation. https://scanpy.readthedocs.io

    2018

  58. [58]

    Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Zhang, Erkang Zhu, Beibin Li, Li Jiang, Xiaoyun Zhang, and Chi Wang. 2023. AutoGen: Enabling next- gen LLM applications via multi-agent conversation framework.arXiv preprint arXiv:2308.08155(2023)

    Pith/arXiv arXiv 2023

  59. [59]

    Yifan Xiao, Jun Liu, Yu Zheng, Xiaojun Xie, Jing Hao, Ming Li, Rui Wang, Feng Ni, Yu Li, et al. 2024. CellAgent: An LLM-Driven Multi-Agent Framework for Automated Single-cell Data Analysis.arXiv preprint arXiv:2407.09811(2024). https://doi.org/10.48550/arXiv.2407.09811

    Pith/arXiv arXiv doi:10.48550/arxiv.2407.09811 2024

  60. [60]

    Chenling Xu, Romain Lopez, Elior Mehlman, Jeffrey Regier, Michael I Jordan, and Nir Yosef. 2021. Probabilistic harmonization and annotation of single-cell transcriptomics data with deep generative models.Molecular Systems Biology17, 1 (2021), e9620. doi:10.15252/msb.20209620

    doi:10.15252/msb.20209620 2021

  61. [61]

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. ReAct: Synergizing reasoning and acting in language models. InInternational Conference on Learning Representations

    2023

  62. [62]

    Natalia Vydra, Patryk Janus, Agnieszka Toma-Jonik, Tomasz Stokowy, Katarzyna Mrowiec, Joanna Korfanty, Anna Długajczyk, Bartosz Wojtaś, Bartłomiej Giel- niewski, and Wiesława Widłak. 2019. 17 𝛽-Estradiol activates HSF1 via MAPK signaling in er𝛼-positive breast cancer cells.Cancers11, 10 (2019), 1533

    2019

  63. [64]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H Chi, Quoc V Le, and Denny Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models. InAdvances in Neural Information Processing Systems

    2022

  64. [2008]

    The MAP kinase phosphatase-1 MKP-1/DUSP1 is a regulator of human liver response to transplantation.American Journal of Transplantation8, 12 (2008), 2558–2568

    2008

  65. [2015]

    The Journal of pathology237, 1 (2015), 85–97

    CXCR2 inhibition suppresses acute and chronic pancreatic inflammation. The Journal of pathology237, 1 (2015), 85–97

    2015

  66. [2018]

    doi:10.1038/s41592-018-0229-2

    Deep generative modeling for single-cell transcriptomics.Nature Methods 15, 12 (2018), 1053–1058. doi:10.1038/s41592-018-0229-2

    doi:10.1038/s41592-018-0229-2 2018

  67. [2022]

    Single-cell sequencing unveils distinct immune microenvironments with CCR6-CCL20 crosstalk in human chronic pancreatitis.Gut71, 9 (2022), 1831– 1842

    2022

  68. [2023]

    https://doi.org/10

    GPT-MolBERTa: GPT molecular features language model for molecular property prediction.arXiv preprint arXiv:2310.03030(2023). https://doi.org/10. 48550/arXiv.2310.03030

    Pith/arXiv arXiv doi:10.48550/arxiv.2310.03030 2023

  69. [2025]

    doi:10.1101/2025.06.03.657517

    CellVoyager: AI CompBio Agent Generates New Insights by Autonomously Analyzing Biological Data.bioRxiv(2025). doi:10.1101/2025.06.03.657517

    doi:10.1101/2025.06.03.657517 2025