Agentic Discovery Transfer Enables Cross-Cancer Discovery at Scale

Our plain-English summary of Agentic Discovery Transfer Enables Cross-Cancer Discovery at Scale, by L. Cheng, J. Shireman, A. Gogate, et al (bioRxiv, 2026).

The study

• Authors: L. Cheng, J. Shireman, A. Gogate, M. Stangis, Y. Cai, Y. Xie, P. Khatri, M. Dey, C. Kendziorski
• Journal: bioRxiv
• Published: 2026
• Read it: PubMed · Publisher · Free full text

Abstract

Cancer research generates tens of thousands of studies each year, yet whether a published finding holds in another cancer type or patient population is rarely tested as doing so is labor-intensive and difficult to scale. We define this task as discovery transfer and present SMART, a multi-agent large language model framework for discovery transfer in single-cell RNA-sequencing studies. Given a source manuscript and a target dataset, SMART extracts major clinical and mechanistic findings, generates and executes analyses for the testable findings, reports which findings transfer, and suggests follow-up. Applied to 121 studies across six cancer types, SMART identified findings that transfer across cancers, including a lung cancer prognostic signature that we confirmed to be associated with survival in liver cancer, pancreatic cancer and pleural mesothelioma. SMART also generated new mechanistic hypotheses, demonstrating how agentic discovery transfer can extend published findings beyond their original setting.

In plain English

This preprint, not peer reviewed, did not study asbestos exposure. The authors built SMART, a computer system of 20 large language model agents. It tests whether a published cancer study's findings hold in a different single-cell RNA-sequencing dataset. When the authors checked 10 cases by hand, 92–97% of audited outputs were correct. They then applied SMART to 121 studies of six cancers. Of the 605 reports pairing a study with a dataset from a different cancer, 46% were judged to transfer overall. One extra target dataset was pleural mesothelioma. In that dataset, expression of the gene CYP27A1 correlated with the proportion of CD8 T cells in epithelioid tumors. In a separate cohort of 69 mesothelioma patients (TCGA-MESO), high scores on ORACLE, a 23-gene lung cancer signature, were associated with worse survival (hazard ratio 3.04, 95% CI 1.78–5.18). High ORACLE scores were also associated with poorer survival in liver and pancreatic cancer. The authors cautioned that they used only one target dataset per cancer type, and that support in this kind of data is not the same as functional validation.

We wrote this plain-English summary from the full article with the help of AI, and a second AI model checked each sentence against the article. It is not the authors' wording; the abstract and the article are their own account.

What the study did and found

• Design: Computational methods study (bioRxiv preprint, not peer reviewed): development, validation and large-scale application of a multi-agent large language model framework, with follow-up survival, drug-response and knockdown re-analyses of public data
• Who was studied: 121 published single-cell RNA-sequencing studies of glioblastoma or colorectal, liver, lung, ovarian or pancreatic cancer, published in Nature Portfolio journals in 2024–2025. Each was run against six public target datasets (one per cancer type, 726 reports) and against a public pleural mesothelioma dataset (121 reports). Follow-up used public TCGA patient cohorts, including TCGA-MESO with n = 69.
• Exposure: No asbestos or other fiber exposure was studied. The study tested whether published gene-expression findings transferred between cancer types, including the 23-gene ORACLE lung cancer prognostic signature.
• Main results:
  • In manual validation of 10 cases, 92–97% of audited outputs were correct.
  • All 15 decoy reports built from reversed study summaries were scored Middle-low or Low, including the 12 whose original reports had been scored Middle-high or High.
  • Of the 605 reports (source study and target dataset pairs) in which the source and target cancers differed, 46% transferred, meaning they received an overall score of Middle-high or High. The rate ranged from 17% to 67% across cancer pairs. When source and target were the same cancer, 75–88% of reports transferred.
  • In TCGA cohorts, higher ORACLE scores were significantly associated with poorer survival in liver cancer (HR = 1.92, 95% CI 1.31–2.82, P = 0.00065; n = 336) and pancreatic cancer (HR = 1.73, 95% CI 1.14–2.64, P = 0.0097; n = 171). They were not significantly associated with survival in ovarian cancer, glioblastoma or colorectal cancer.
  • In TCGA-MESO (pleural mesothelioma, n = 69), high ORACLE scores were associated with worse survival (HR = 3.04, 95% CI 1.78–5.18, P = 2.2 × 10⁻⁵). ORACLE was more strongly associated with survival than all 1,000 proliferation-matched random signatures.
  • In the pleural mesothelioma single-cell dataset, CYP27A1 expression correlated positively with the LXR target gene ABCA1 and with the proportion of CD8 T cells in epithelioid tumors.
• The authors' conclusion: The authors concluded that SMART can reveal cross-cancer similarities, distinguish conserved from context-dependent mechanisms, and prioritize mechanistic and prognostic hypotheses for follow-up.
• Limitations the authors state:
  • Analyses used only one target dataset per cancer type, so transfer rates should be interpreted with caution.
  • Support in single-cell RNA-sequencing data is not equivalent to functional validation, particularly for mechanisms acting through protein activity or spatial organization.
  • Transfer rates refer only to studies whose findings were testable in single-cell RNA-sequencing data.
  • Literature searches used to call findings 'not previously reported' were not systematic reviews.
  • Because of memory constraints, the ovarian cancer dataset was restricted to a random 11 of its 34 patients.

Topics

Medical terms in this study, explained in our Lexicon: Discovery, Lung Cancer, Pleural Mesothelioma.

Related studies

• Discovery and application of the Yoshida effect: nano-sized acicular materials enable penetration of bacterial cells by sliding friction force — N. Yoshida (2007)
• Identifying the pulmonary hazard of high aspect ratio nanoparticles to enable their safety-by-design — K. Donaldson, F. Murphy, et al. (2011)
• Chimeric antigen receptor T cells engineered to recognize the P329G-mutated Fc part of effector-silenced tumor antigen-targeting human IgG1 antibodies enable modular targeting of solid tumors — S. Stock, M.R. Benmebarek, et al. (2022)
• Cisplatin-induced expression of Gb3 enables verotoxin-1 treatment of cisplatin resistance in malignant pleural mesothelioma cells — D. Johansson, C. Andersson, et al. (2010)
• Pleural Fluid Has Pro-Growth Biological Properties Which Enable Cancer Cell Proliferation — R. Asciak, N.I. Kanellakis, et al. (2021)
• Tissue transglutaminase (TG2) enables survival of human malignant pleural mesothelioma cells in hypoxia — S. Zonca, G. Pinton, et al. (2017)

Browse return to medical research.