Pith. sign in

REVIEW 1 major objections 68 references

LLM methods for predicting gene responses to cellular perturbations produce plausible but inaccurate results by relying on general gene tendencies rather than specific effects.

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-4.3

2026-06-28 17:59 UTC pith:AYOEGCUM

load-bearing objection The paper shows LLM perturbation predictors mostly latch onto gene-intrinsic tendencies rather than specific effects, and CORE's contrastive KG evidence lifts results, but lacks the ablation needed to pin the gain on contrast itself. the 1 major comments →

arxiv 2606.01042 v1 pith:AYOEGCUM submitted 2026-05-31 cs.LG cs.AI

Plausibility Is Not Prediction: Contrastive Evidence for LLM-Based Cellular Perturbation Reasoning

Xinyu Yuan , Xixian Liu , Jianan Zhao , Yashi Zhang , Hongyu Guo , Jian Tang This is my paper
classification cs.LG cs.AI
keywords LLM perturbation reasoningcellular perturbation predictioncontrastive evidence organizationgene expression responsedifferential expressionbiomedical knowledge graphvirtual cell simulation
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.

The paper establishes that current LLM-based approaches to simulating cellular perturbations generate biologically plausible step-by-step explanations yet fail to predict actual differential expression outcomes. These methods systematically overestimate changes, often lose to a basic gene-frequency baseline in overall scores, and drop to chance performance when assessed gene by gene. The authors trace the problem to isolated evaluation of each perturbation-gene pair, which allows the model to default to intrinsic response patterns instead of learning perturbation-specific logic. They introduce CORE, a reframing that retrieves related perturbations from a knowledge graph and organizes them as positive and negative evidence for direct comparison. This contrastive organization raises accuracy in both LLM reasoning and simpler voting setups across drug and generic perturbation datasets.

Core claim

Despite producing biologically plausible explanations, these methods fail to capture perturbation-specific effects: systematically overestimating differential expression, often underperforming a simple gene-frequency baseline in aggregate evaluations, and collapsing to chance-level performance at the per-gene level. This reveals a reliance on intrinsic gene response tendencies rather than true perturbation reasoning. Reframing prediction as a comparison task by organizing evidence into positive and negative outcomes from related perturbations using a biomedical knowledge graph improves calibration and substantially boosts perturbation-specific prediction.

What carries the argument

CORE (Contrastive Organization of Relational Evidence), which retrieves related perturbations from a biomedical knowledge graph and structures their outcomes as positive and negative examples to turn isolated prediction into explicit comparison.

Load-bearing premise

The failure of existing methods stems specifically from evaluating perturbation-gene pairs in isolation without contrast to related perturbations, and that reframing as a comparison task with positive and negative outcomes from a knowledge graph will reliably address this without introducing new selection biases.

What would settle it

Run the same LLM reasoning pipeline on the same datasets but replace CORE's related-perturbation contrasts with randomly selected unrelated perturbations; if the performance gains disappear, the contrastive mechanism is necessary.

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

If this is right

  • CORE-Reasoning raises aggregate metrics by up to 28.6 percent for a 9B-parameter model on drug-perturbation data.
  • CORE-Voting lifts macro per-gene AUROC from chance level to 0.703 on average across four cell lines in generic perturbation data.
  • The same contrastive organization improves results in both LLM-based reasoning pipelines and non-LLM voting baselines.
  • Reliable perturbation reasoning requires explicit comparison across related conditions rather than isolated evidence.

Where Pith is reading between the lines

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

  • Contrastive evidence organization may transfer to other LLM reasoning tasks where isolated prompts produce plausible but non-specific answers, such as causal attribution in other scientific domains.
  • The method could be tested on larger or more diverse cell-line panels to check whether the reported AUROC lift holds when the knowledge-graph coverage varies.
  • If the knowledge-graph retrieval step introduces its own selection effects, an ablation that substitutes random or lexical matches for graph neighbors would isolate that contribution.

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

1 major / 0 minor

Summary. The manuscript claims that LLM-based methods for predicting gene expression responses to cellular perturbations produce biologically plausible explanations but fail to capture perturbation-specific effects, as shown by systematic overestimation of differential expression, frequent underperformance relative to a gene-frequency baseline in aggregate metrics, and chance-level performance at the per-gene level. It attributes this to isolated evaluation of perturbation-gene pairs and introduces CORE (Contrastive Organization of Relational Evidence), which retrieves positive and negative outcomes from related perturbations via a biomedical knowledge graph and reframes prediction as an explicit comparison task. CORE is reported to improve calibration and boost performance in both LLM-based and non-LLM settings, with examples including up to 28.6% gains in aggregate metrics for drug-perturbation data and lifting average macro-per-gene AUROC from chance to 0.703 across cell lines.

Significance. If the results hold, the work demonstrates that contrastive evidence organization can substantially advance knowledge-grounded LLM reasoning for perturbation prediction beyond plausible but non-specific outputs, with the empirical baseline comparisons and use of an external KG for evidence retrieval providing falsifiable grounding. This could support more reliable virtual cell models, though the gains must be shown to stem specifically from contrastive presentation rather than evidence curation.

major comments (1)
  1. [Abstract] Abstract: the central claim that contrastive evidence organization is 'essential' to reliable perturbation reasoning is not supported by an ablation that holds the retrieved evidence set fixed while varying only presentation format (contrastive positive/negative pairs versus a flat list). Without this control, performance improvements could arise from KG evidence selection or curation biases rather than the contrastive structure itself.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their careful reading and constructive feedback. We address the major comment below and will revise the manuscript accordingly to strengthen the claims.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that contrastive evidence organization is 'essential' to reliable perturbation reasoning is not supported by an ablation that holds the retrieved evidence set fixed while varying only presentation format (contrastive positive/negative pairs versus a flat list). Without this control, performance improvements could arise from KG evidence selection or curation biases rather than the contrastive structure itself.

    Authors: We agree that the current experiments do not fully isolate the contribution of contrastive presentation format from the effects of evidence curation via the knowledge graph. To address this, we will add a controlled ablation in the revised manuscript that holds the retrieved evidence set fixed (same positive and negative outcomes from related perturbations) while varying only the presentation: contrastive positive/negative pairs versus a flat concatenated list. This will allow us to quantify the specific benefit of the contrastive structure. We will also revise the abstract to replace the word 'essential' with a more precise statement supported by the new results (e.g., 'substantially improves' rather than 'essential'). revision: yes

Circularity Check

0 steps flagged

No circularity: empirical evaluations and external KG remain independent of inputs

full rationale

The paper's core argument—that isolated evaluation leads to reliance on gene-intrinsic tendencies while contrastive KG evidence improves perturbation-specific prediction—is supported by direct empirical comparisons (overestimation vs. gene-frequency baseline, per-gene AUROC collapse to chance, and reported lifts such as +28.6% aggregate metrics and macro AUROC 0.703). No equations, fitted parameters, or self-citations are invoked to derive the claimed improvements; the KG is external, results are measured on held-out perturbation data, and the contrastive reframing is presented as an organizational change whose effect is tested rather than assumed by definition. The derivation chain therefore does not reduce to its own inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that the biomedical knowledge graph supplies unbiased relational evidence for constructing positive and negative perturbation outcomes; no free parameters or invented entities are described in the abstract.

axioms (1)
  • domain assumption Biomedical knowledge graph supplies unbiased relational evidence for constructing positive and negative perturbation outcomes
    Invoked for evidence retrieval in CORE to enable contrastive organization.

pith-pipeline@v0.9.1-grok · 5838 in / 1240 out tokens · 26114 ms · 2026-06-28T17:59:47.604973+00:00 · methodology

0 comments
read the original abstract

Perturbation experiments are central to understanding cellular mechanisms, but remain costly and sparse, motivating prediction of gene expression responses for unobserved conditions. A promising recent direction leverages large language models (LLMs) as "virtual cell" simulators-using stepwise, knowledge-grounded mechanistic reasoning to infer differential expression-pointing toward an interpretable, knowledge-driven paradigm that transcends purely data-driven approaches. However, we find that plausibility is not prediction: despite producing biologically plausible explanations, these methods fail to capture perturbation-specific effects: systematically overestimating differential expression, often underperforming a simple gene-frequency baseline in aggregate evaluations, and collapsing to chance-level performance at the per-gene level. This reveals a reliance on intrinsic gene response tendencies rather than true perturbation reasoning. We trace this failure to how evidence is presented: existing methods evaluate perturbation-gene pairs in isolation, without exposing how related perturbations differ in their effects on the same gene. To address this limitation, we introduce CORE (Contrastive Organization of Relational Evidence), which reframes prediction as a comparison task by organizing evidence into positive and negative outcomes from related perturbations. Using a biomedical knowledge graph for evidence retrieval, CORE improves calibration and substantially boosts perturbation-specific prediction in both LLM-based and non-LLM settings: for example, on drug-perturbation data, CORE-Reasoning improves Qwen3.5-9B aggregate metrics by up to 28.6%, while on generic perturbation data, CORE-Voting raises macro-per-gene AUROC from chance to 0.703 in average across four cell lines. This highlights contrastive evidence organization as essential to reliable LLM-based perturbation reasoning

Figures

Figures reproduced from arXiv: 2606.01042 by Hongyu Guo, Jianan Zhao, Jian Tang, Xinyu Yuan, Xixian Liu, Yashi Zhang.

Figure 1
Figure 1. Figure 1: ReasonKG harmonizes five biomedical resources into a canonical graph for CORE retrieval. To address this limitation, we propose CORE (Contrastive Organization of Relational Evidence), which reframes pre￾diction as comparative task. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Predicted positive-rate on Tahoe100M C32, showing VC￾World’s yes bias and CORE’s closer calibration to the true yes ratio. We first examine whether isolated virtual-cell prompts yield calibrated binary decisions. Across VCWorld-style prompts using different base LLMs, we observe a consistent positive￾rate bias ( [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: CORE framework. KG-based perturbation signatures retrieve same-gene positive and [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of Gene-prior Voting and CORE-Voting across support sweeps. Gene-prior [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: LLM comparison at fixed-budget (k = 50, m = 200) on Tahoe100M C32. CORE￾Reasoning substantially improves aggregate metrics across LLM backbones and improves macro per-gene discrimination most clearly for Qwen3.5 models and GPT-5.4 mini, while CORE-Voting remains a strong non-LLM reference. On PerturbQA K562, CORE-Voting improves every reported metric over Gene-prior Voting, including overall AUPRC from 0.8… view at source ↗
Figure 6
Figure 6. Figure 6: ReasonKG node composition and source integration. The graph is centered on genes, [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: ReasonKG relation vocabulary and edge coverage. The harmonized graph contains 83 [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Connectivity and degree distribution of ReasonKG. The largest weakly connected com [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Source and type distribution of isolated nodes in ReasonKG. Most isolated nodes originate [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Query-set connectivity comparison between ReasonKG and the VCWorld KG. In the ana [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Full metrics for comparing Gene-prior Voting and CORE-Voting on PerturbQA K562 [PITH_FULL_IMAGE:figures/full_fig_p023_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Calibration and output coverage for C32 fixed-budget LLM runs. The dashed line in [PITH_FULL_IMAGE:figures/full_fig_p024_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Calibration and output coverage for PerturbQA K562 cell line. CORE evidence changes [PITH_FULL_IMAGE:figures/full_fig_p024_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: PerturbQA K562 comparison between SUMMER-style prompting, CORE-Reasoning, and [PITH_FULL_IMAGE:figures/full_fig_p025_14.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

68 extracted references · 2 canonical work pages

  1. [1]

    A multiplexed single-cell crispr screening platform enables systematic dissection of the unfolded protein response.Cell, 167(7):1867–1882, 2016

    Britt Adamson, Thomas M Norman, Marco Jost, Min Y Cho, James K Nuñez, Yuwen Chen, Jacqueline E Villalta, Luke A Gilbert, Max A Horlbeck, Marco Y Hein, et al. A multiplexed single-cell crispr screening platform enables systematic dissection of the unfolded protein response.Cell, 167(7):1867–1882, 2016

    2016

  2. [3]

    Predicting cellular responses to perturbation across diverse contexts with state.BioRxiv, pages 2025–06, 2025

    Abhinav K Adduri, Dhruv Gautam, Beatrice Bevilacqua, Alishba Imran, Rohan Shah, Mohsen Naghipourfar, Noam Teyssier, Rajesh Ilango, Sanjay Nagaraj, Mingze Dong, et al. Predicting cellular responses to perturbation across diverse contexts with state.BioRxiv, pages 2025–06, 2025

    2025

  3. [4]

    Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines.Nature Methods, 22:1657–1661, 2025

    Constantin Ahlmann-Eltze, Wolfgang Huber, and Simon Anders. Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines.Nature Methods, 22:1657–1661, 2025

    2025

  4. [5]

    Controlling the false discovery rate: A practical and powerful approach to multiple testing.Journal of the Royal Statistical Society: Series B (Methodological), 57(1):289–300, 1995

    Yoav Benjamini and Yosef Hochberg. Controlling the false discovery rate: A practical and powerful approach to multiple testing.Journal of the Royal Statistical Society: Series B (Methodological), 57(1):289–300, 1995

    1995

  5. [6]

    Language models are few-shot learners.Advances in neural information processing systems, 33:1877–1901, 2020

    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

  6. [7]

    Building a knowledge graph to enable precision medicine.Scientific data, 10(1):67, 2023

    Payal Chandak, Kexin Huang, and Marinka Zitnik. Building a knowledge graph to enable precision medicine.Scientific data, 10(1):67, 2023

    2023

  7. [8]

    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

  8. [9]

    Perturb-seq: dissecting molecular circuits with scalable single-cell rna profiling of pooled genetic screens

    Atray Dixit, Oren Parnas, Biyu Li, Jenny Chen, Charles P Fulco, Livnat Jerby-Arnon, Ne- manja D Marjanovic, Danielle Dionne, Tyler Burks, Raktima Raychowdhury, et al. Perturb-seq: dissecting molecular circuits with scalable single-cell rna profiling of pooled genetic screens. cell, 167(7):1853–1866, 2016

    2016

  9. [10]

    Systematic integration of biomedical knowledge prioritizes drugs for repurposing.elife, 6:e26726, 2017

    Daniel Scott Himmelstein, Antoine Lizee, Christine Hessler, Leo Brueggeman, Sabrina L Chen, Dexter Hadley, Ari Green, Pouya Khankhanian, and Sergio E Baranzini. Systematic integration of biomedical knowledge prioritizes drugs for repurposing.elife, 6:e26726, 2017

    2017

  10. [11]

    The connectivity map: using gene-expression signatures to connect small molecules, genes, and disease.science, 313(5795):1929–1935, 2006

    Justin Lamb, Emily D Crawford, David Peck, Joshua W Modell, Irene C Blat, Matthew J Wrobel, Jim Lerner, Jean-Philippe Brunet, Aravind Subramanian, Kenneth N Ross, et al. The connectivity map: using gene-expression signatures to connect small molecules, genes, and disease.science, 313(5795):1929–1935, 2006

    1929

  11. [12]

    Jiachang Liu, Dinghan Shen, Yizhe Zhang, William B Dolan, Lawrence Carin, and Weizhu Chen. What makes good in-context examples for gpt-3? InProceedings of Deep Learning Inside Out (DeeLIO 2022): The 3rd workshop on knowledge extraction and integration for deep learning architectures, pages 100–114, 2022

    2022

  12. [13]

    Predicting cellular responses to complex perturbations in high-throughput screens.Molecular systems biology, 19(6):MSB202211517, 2023

    Mohammad Lotfollahi, Anna Klimovskaia Susmelj, Carlo De Donno, Leon Hetzel, Yuge Ji, Ignacio L Ibarra, Sanjay R Srivatsan, Mohsen Naghipourfar, Riza M Daza, Beth Martin, et al. Predicting cellular responses to complex perturbations in high-throughput screens.Molecular systems biology, 19(6):MSB202211517, 2023

    2023

  13. [14]

    Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity

    Yao Lu, Max Bartolo, Alastair Moore, Sebastian Riedel, and Pontus Stenetorp. Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, pages 8086–8098, Dublin, Ireland, 2022. Association for Computational Linguistics. 11

    2022

  14. [15]

    Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. Rethinking the role of demonstrations: What makes in-context learning work? InProceedings of the 2022 conference on empirical methods in natural language processing, pages 11048–11064, 2022

    2022

  15. [16]

    Identifying compound-protein interactions with knowledge graph embedding of perturbation transcriptomics.Cell genomics, 4(10), 2024

    Shengkun Ni, Xiangtai Kong, Yingying Zhang, Zhengyang Chen, Zhaokun Wang, Zunyun Fu, Ruifeng Huo, Xiaochu Tong, Ning Qu, Xiaolong Wu, et al. Identifying compound-protein interactions with knowledge graph embedding of perturbation transcriptomics.Cell genomics, 4(10), 2024

    2024

  16. [17]

    Qwen3.5: Towards native multimodal agents

    Qwen Team. Qwen3.5: Towards native multimodal agents. https://qwen.ai/blog?id= qwen3.5, February 2026. Qwen3.5-9B model card: https://huggingface.co/Qwen/ Qwen3.5-9B. Accessed: 2026-05-07

    2026

  17. [18]

    Predicting transcriptional outcomes of novel multigene perturbations with gears.Nature Biotechnology, 42(6):927–935, 2024

    Yusuf Roohani, Kexin Huang, and Jure Leskovec. Predicting transcriptional outcomes of novel multigene perturbations with gears.Nature Biotechnology, 42(6):927–935, 2024

    2024

  18. [19]

    Learning to retrieve prompts for in-context learning

    Ohad Rubin, Jonathan Herzig, and Jonathan Berant. Learning to retrieve prompts for in-context learning. InProceedings of the 2022 conference of the North American chapter of the association for computational linguistics: human language technologies, pages 2655–2671, 2022

    2022

  19. [20]

    Biokg: A knowledge graph for relational learning on biological data

    Brian Walsh, Sameh K Mohamed, and Vít Nováˇcek. Biokg: A knowledge graph for relational learning on biological data. InProceedings of the 29th ACM International Conference on Information & Knowledge Management, pages 3173–3180, 2020

    2020

  20. [21]

    VCWorld: A biological world model for virtual cell simulation

    Zhijian Wei, Runze Ma, Zichen Wang, Zhongmin Li, Shuotong Song, and Shuangjia Zheng. VCWorld: A biological world model for virtual cell simulation. InInternational Conference on Learning Representations, 2026

    2026

  21. [22]

    Individual comparisons by ranking methods

    Frank Wilcoxon. Individual comparisons by ranking methods. InBreakthroughs in statistics: Methodology and distribution, pages 196–202. Springer, 1992

    1992

  22. [23]

    Contextualizing biological perturbation experiments through language

    Menghua Wu, Russell Littman, Jacob Levine, Lin Qiu, Tommaso Biancalani, David Rich- mond, and Jan-Christian Huetter. Contextualizing biological perturbation experiments through language. InThe Thirteenth International Conference on Learning Representations, 2025

    2025

  23. [24]

    arXiv preprint arXiv:2602.19685

    Xinyu Yuan, Xixian Liu, Ya Shi Zhang, Zuobai Zhang, Hongyu Guo, and Jian Tang. Perturbdiff: Functional diffusion for single-cell perturbation modeling.arXiv preprint arXiv:2602.19685, 2026

    work page arXiv 2026

  24. [25]

    CellVerse: Do large language models really understand cell biology? InAdvances in Neural Information Processing Systems, 2025

    Fan Zhang, Tianyu Liu, Zhihong Zhu, Hao Wu, Haixin Wang, Donghao Zhou, Yefeng Zheng, Kun Wang, Xian Wu, and Pheng-Ann Heng. CellVerse: Do large language models really understand cell biology? InAdvances in Neural Information Processing Systems, 2025. Datasets and Benchmarks Track

    2025

  25. [26]

    Tahoe-100m: A giga- scale single-cell perturbation atlas for context-dependent gene function and cellular modeling

    Jesse Zhang, Airol A Ubas, Richard De Borja, Valentine Svensson, Nicole Thomas, Neha Thakar, Ian Lai, Aidan Winters, Umair Khan, Matthew G Jones, et al. Tahoe-100m: A giga- scale single-cell perturbation atlas for context-dependent gene function and cellular modeling. BioRxiv, pages 2025–02, 2025

    2025

  26. [27]

    Support” denotes the training split available to retrieval, voting, and prompt construction; “test

    Lianmin Zheng, Liangsheng Yin, Zhiqiang Xie, Chuyue Sun, Jeff Huang, Cody Hao Yu, Shiyi Cao, Christos Kozyrakis, Ion Stoica, Joseph E. Gonzalez, Clark Barrett, and Ying Sheng. SGLang: Efficient execution of structured language model programs. InAdvances in Neural Information Processing Systems, volume 37, pages 62557–62583, 2024. 12 A Dataset Details Tabl...

    2024

  27. [28]

    Task and query context.The prompt states the fixed cell context, query perturbation, and query gene, together with textual descriptions of the perturbation and gene when available

  28. [29]

    Gene prior.The support-set positive rate for the target gene is displayed as a calibration reference, followed by an explicit warning that the prior alone is insufficient because the prediction must distinguish perturbations for the same gene

  29. [30]

    It then lists up to three top-weighted positive and up to three top-weighted negative support perturbations for the same target gene, including their KG similarity and vote weight

    Same-gene support evidence.The packet reports the same-gene contrast score, positive vote mass, negative vote mass, and observed vote mass. It then lists up to three top-weighted positive and up to three top-weighted negative support perturbations for the same target gene, including their KG similarity and vote weight

  30. [31]

    Mapped endpoints are explicitly marked so that the LLM can down-weight indirect evidence

    KG path evidence.The packet includes up to five shortest KG paths between the grounded or mapped query endpoints. Mapped endpoints are explicitly marked so that the LLM can down-weight indirect evidence

  31. [32]

    In the main experiments, the prompt displays three anchor genes, while the numerical anchor score is computed from the top ten anchors

    Gene-anchor evidence.The packet displays the anchor score, the final blended score, and the top displayed anchor genes. In the main experiments, the prompt displays three anchor genes, while the numerical anchor score is computed from the top ten anchors

  32. [33]

    The prompt asks the model to compare positive and negative same-gene cases before constructing a biological rationale

    Output contract.Before the final deterministic yes/no sentence, the model must output a one-line JSON object with fields probability, label, default_final_score, support_score, anchor_score, probability_adjustment, anddecisive_evidence. The prompt asks the model to compare positive and negative same-gene cases before constructing a biological rationale. T...

    work page arXiv

  33. [34]

    **Mechanism & Analogue Identification:** Identify drugs in the evidence set that share the same *Mechanism of Action (MoA)* or target specific pathway nodes as Gemfibrozil.,→

  34. [36]

    **Downstream Signaling Cascade Simulation:** Trace the signaling cascade initiated by Gemfibrozil. When Gemfibrozil inhibits/activates its target, which specific downstream kinases, transcription factors, or stress responses are modulated?,→ *Constraint:* Ensure this simulation aligns with the enhanced context of C32 cells

  35. [37]

    Analogue Cases

    **Causal Bridge & Evidence Synthesis:** Connect the drug's downstream effect to the gene's regulatory requirements. - Construct a logical bridge: Drug -> Target -> Pathway -> TF -> Gene. - Cite specific "Analogue Cases" that support this link. - *Soft Reference:* Briefly refer to "Contrast Cases" (if available) as supplementary context to see if they offe...

  36. [38]

    End your response with exactly one of the following options: - No

    **Final Deterministic Prediction:** Based on the analysis above, determine if the drug effectively perturbs the gene in this specific cell line. End your response with exactly one of the following options: - No. Perturbation of Gemfibrozil does not impact GADD45B. - Yes. Perturbation of Gemfibrozil results in differential expression of GADD45B. - There is...

  37. [39]

    Its primary biological effect is the modulation of lipid metabolism, specifically increasing fatty acid oxidation and lipoprotein clearance

    **Mechanism & Analogue Identification:** Gemfibrozil functions as a potent agonist of the nuclear receptor Peroxisome Proliferator-Activated Receptor-alpha (PPARalpha). Its primary biological effect is the modulation of lipid metabolism, specifically increasing fatty acid oxidation and lipoprotein clearance. To identify relevant analogues within the provi...

  38. [40]

    While PPARalpha is primarily metabolic, its activation in cancer cells can trigger metabolic stress responses

    **Specificity & Relevance Analysis (Drug-Gene-Cell Triad):** * **Specificity:** Gemfibrozil's effect is specific to the PPARalpha pathway, whereas GADD45B is a downstream effector of cellular stress and DNA damage (regulated by p38/JNK MAPKs). While PPARalpha is primarily metabolic, its activation in cancer cells can trigger metabolic stress responses. ,→...

  39. [41]

    **Downstream Signaling Cascade Simulation:** Upon perturbation by Gemfibrozil:

  40. [42]

    **Activation:** Gemfibrozil binds and activates **PPARalpha**

  41. [43]

    **Metabolic Shift:** PPARalpha transcriptional activity increases the expression of genes involved in fatty acid oxidation (e.g., *CPT1A*, *ACO*).,→

  42. [44]

    **Stress Induction:** In the context of C32 melanoma cells (which have altered lipid handling due to oncogenic signaling), this shift can lead to an accumulation of Reactive Oxygen Species (ROS) or mitochondrial uncoupling.,→

  43. [45]

    This is distinct from the constitutive ERK activation driven by BRAF V600E.,→

    **MAPK Activation:** ROS and metabolic stress activate the **p38 and JNK MAPK pathways**. This is distinct from the constitutive ERK activation driven by BRAF V600E.,→

  44. [46]

    **Gene Regulation:** Activated p38/JNK phosphorylate transcription factors (e.g., ATF2, MEF2) that bind to the **GADD45B** promoter.,→

  45. [47]

    **Outcome:** GADD45B expression increases, leading to cell cycle arrest (via CDK1 interaction) and apoptosis, serving as a counter-regulatory mechanism to the metabolic stress induced by the drug.,→

  46. [48]

    Result: B

    **Causal Bridge & Evidence Synthesis:** * **Logical Bridge:** Gemfibrozil (PPARalpha Agonist) -> Lipid Metabolism Perturbation -> Metabolic Stress/ROS -> p38/JNK MAPK Activation -> GADD45B Upregulation.,→ * **Analogue Cases:** **Example 9 (Oleic acid)** provides direct support. As a fatty acid ligand for PPARalpha, Oleic acid resulted in differential expr...

  47. [49]

    ,→ ,→ ,→ Yes

    **Final Deterministic Prediction:** Based on the strong mechanistic alignment between Gemfibrozil and Oleic acid (Example 9), the specific metabolic stress pathway triggered in C32 melanoma cells (BRAF/PTEN background), and the downstream activation of p38/JNK MAPK pathways known to regulate GADD45B, the perturbation is predicted to successfully induce di...

  48. [50]

    Explain why Gemfibrozil looks closer to one side.,→

    **Same-Gene Contrast:** Compare Gemfibrozil against the positive and negative support perturbations for the same fixed gene GADD45B. Explain why Gemfibrozil looks closer to one side.,→

  49. [51]

    **Specificity & Relevance Analysis (Drug-Gene-Cell Triad):** Analyze the potential associations between the Drug, Gene, and Cell Line: - **Specificity:** Is the drug's effect broad (e.g., general stress) or specific (e.g., targeted kinase inhibition)? Is the gene's expression tissue-specific?,→ - **Relevance:** Given the enhanced cell line context (e.g., ...

  50. [52]

    Distinguish mechanism-bearing paths from generic/noisy graph connectivity.,→ *Constraint:* Ensure this simulation aligns with the enhanced context of C32 cells

    **KG Path Quality Judgment:** Trace the signaling cascade initiated by Gemfibrozil using the provided KG path evidence. Distinguish mechanism-bearing paths from generic/noisy graph connectivity.,→ *Constraint:* Ensure this simulation aligns with the enhanced context of C32 cells

  51. [53]

    **Gene-Anchor Adjustment:** If direct same-gene evidence is weak or conflicting, use the gene-anchor evidence to explain whether biologically similar genes suggest a positive or negative response under Gemfibrozil.,→ Treat GeneAnchor as mathematical smoothing, not only as narrative support: explicitly compare support_score, anchor_score, and final_score =...

  52. [54]

    - Construct a logical bridge: Drug -> Target -> Pathway -> TF -> Gene

    **Causal Bridge & Evidence Synthesis:** Connect the drug's downstream effect to the gene's regulatory requirements. - Construct a logical bridge: Drug -> Target -> Pathway -> TF -> Gene. - Cite specific same-gene support cases and KG paths that are decisive. - Explicitly mention whether the final judgment is driven more by same-gene support transfer, KG p...

  53. [55]

    probability

    **Calibrated Final Prediction:** Based on the analysis above, estimate a calibrated probability that GADD45B is differentially expressed under Gemfibrozil in this cell line, then convert it into a deterministic answer.,→ This probability will be evaluated as a continuous score for within-gene ranking across perturbations. Use the provided final blended sc...

  54. [56]

    Oleic acid | KG similarity=0.146 | vote weight=0.382 | label=positive Top negative same-gene supports:

  55. [57]

    Ritonavir | KG similarity=0.401 | vote weight=0.633 | label=negative

  56. [58]

    Clopidogrel | KG similarity=0.343 | vote weight=0.585 | label=negative

  57. [59]

    Mapped-endpoint paths are indirect evidence for the original query

    Methylprednisolone succinate | KG similarity=0.327 | vote weight=0.572 | label=negative KG path evidence: - Shortest paths are searched between grounded query endpoints; if an endpoint is not in KG, the displayed path explicitly uses the KG-support mapped drug or anchor gene endpoint. Mapped-endpoint paths are indirect evidence for the original query. ,→ ...

  58. [60]

    INV::upregulates | INV::drug_drug_interaction (count=2)

  59. [61]

    - Evidence scope: direct query endpoints

    INV::upregulates | drug_drug_interaction (count=1) Mapped endpoint shortest paths: Path 1: - KG path: mRNA:GADD45B -[INV::upregulates]-> PERT:BUDESONIDE -[drug_drug_interaction]-> PERT:GEMFIBROZIL - Endpoint mapping: Gemfibrozil -> Gemfibrozil (query_drug); GADD45B -> GADD45B (query_gene). - Evidence scope: direct query endpoints. - Path length: 2 hops. -...

  60. [62]

    GADD45B | similarity=1.000 | predicted score under current drug=0.103

  61. [63]

    SPATA19 | similarity=0.452 | predicted score under current drug=0.000

  62. [64]

    It functions by interacting with cell cycle components like CDK1 and positively regulating stress-activated JNK and p38MAPK signaling cascades

    TRBV20OR9-2 | similarity=0.452 | predicted score under current drug=0.000 Contrasting fixed-gene examples: Example 1: - Drug: Oleic acid - Gene: GADD45B - Drug Description: Drug'Oleic acid'description not found - Gene Description: is a nuclear and cytoplasmic protein that regulates the cell cycle and promotes apoptosis. It functions by interacting with ce...

  63. [65]

    Both Gemfibrozil and Oleic acid modulate lipid metabolism (Gemfibrozil via PPARalpha activation, Oleic acid as a fatty acid ligand for PPARs)

    **Same-Gene Contrast:** Gemfibrozil is mechanistically closer to the positive support perturbation **Oleic acid** than the negative supports (Ritonavir, Clopidogrel, Methylprednisolone). Both Gemfibrozil and Oleic acid modulate lipid metabolism (Gemfibrozil via PPARalpha activation, Oleic acid as a fatty acid ligand for PPARs). In contrast, the negative s...

  64. [66]

    GADD45B is a stress/DNA damage response gene regulated by MAPK/JNK pathways

    **Specificity & Relevance Analysis (Drug-Gene-Cell Triad):** - **Specificity:** Gemfibrozil is a specific PPARalpha agonist affecting lipid clearance. GADD45B is a stress/DNA damage response gene regulated by MAPK/JNK pathways. There is no direct mechanistic link between PPARalpha activation and GADD45B induction in the provided context. ,→ ,→ - **Relevan...

  65. [67]

    For example,`GADD45B -[INV::upregulates]-> Budesonide -[drug_drug_interaction]-> Gemfibrozil`

    **KG Path Quality Judgment:** The retrieved KG paths are of length 2 (Gene -> Drug -> Drug). For example,`GADD45B -[INV::upregulates]-> Budesonide -[drug_drug_interaction]-> Gemfibrozil`. These paths rely on`drug_drug_interaction`relations, which represent generic connectivity rather than specific mechanistic causation. The directionality`GADD45B -> Budes...

  66. [68]

    - **Anchor Score:** 0.088 (Low)

    **Gene-Anchor Adjustment:** - **Support Score:** 0.103 (Low). - **Anchor Score:** 0.088 (Low). - **Final Blended Score:** 0.095 (Calculated as 0.5 * 0.103 + 0.5 * 0.088). - **Analysis:** The gene anchor score is very low, and the similar genes (SPATA19, TRBV20OR9-2) show 0.000 predicted scores under Gemfibrozil. This mathematical smoothing reinforces the ...

  67. [69]

    - **Evidence:** The only strong positive same-gene support is Oleic acid (Lipid/PPAR), but the negative vote mass is 9x higher

    **Causal Bridge & Evidence Synthesis:** - **Bridge:** Gemfibrozil -> PPARalpha -> Lipid Metabolism -> [Weak Link] -> Cellular Stress -> GADD45B. - **Evidence:** The only strong positive same-gene support is Oleic acid (Lipid/PPAR), but the negative vote mass is 9x higher. The KG paths are noisy. The C32 context (BRAF/PTEN) does not amplify a PPARalpha-GAD...

  68. [70]

    probability

    **Calibrated Final Prediction:** The default final score is 0.095. Given the low support and negative bias, I will preserve this score as it accurately reflects the weak evidence and within-gene ranking.,→ Probability: 0.095. Label: no. {"probability": 0.095, "label": "no", "default_final_score": 0.095, "support_score": 0.103, "anchor_score": 0.088, "prob...