Agentic discovery transfer

Every published finding is a hypothesis for another cancer.

AI agents test published findings in new data.

121 papers · 6 cancers · 726 automated tests

What AI agents found

New prognostic signature

Proven in lung. Overlooked in mine?

New leads for a rare cancer

Few studies in my cancer? Borrow from richly studied ones.

New drug lead

Which cancer is mine most like? Not the one you would guess.

New mechanism

Pieces from separate papers. A new mechanism for mine.

New hypotheses

One gene, every cancer. What holds in mine?

New explanation

A whole mechanism from one paper. Does it hold in mine?

New network

Every paper adds a piece. What do they build together?

ORACLE, a lung cancer signature, scored in TCGAoverall survival, high vs low score
Pancreatic TCGA-PAAD, n = 171
Kaplan-Meier curves, pancreatic cancer: ORACLE-high patients survive shorter
HR 1.73 (1.14–2.64), log-rank P = 0.0097
Liver TCGA-LIHC, n = 336
Kaplan-Meier curves, liver cancer: ORACLE-high patients survive shorter
HR 1.92 (1.31–2.82), log-rank P = 0.00065
Mesothelioma (TCGA-MESO): HR 3.04 (1.78–5.18), P = 2.2 × 10⁻⁵
colorectalglioblastomaliverlungovarianpancreatic
pleural mesothelioma121 papers tested
explains

A good prognosis, not yet explained

CYP27A1, from colorectal cancer

Potential explanation: T cell infiltration

rediscovers

A forgotten survival signature

ORACLE, from lung cancer

TCGA: worse survival in TCGA-MESO (HR 3.04)

What if we test at scale? 121 papers × 6 cancers
OVAPANCCRCGBMLUNGLIVER
Which two of our 6 cancers are most alike?
Ovarian⇄Pancreatic
63%of findings transfer between them
vs 36–56% for each of the other 14 pairs
Drug leadGPX4-inhibitor resistance by PCSK9 expression in 38 ovarian cancer cell lines, P = 0.044

GPX4 inhibitors for PCSK9-low ovarian cancer

P = 0.044 · 38 cell lines
pancreatic paperEZH2✓STAT3
another pancreatic paperSTAT3✓ZEB1✓EMT
✓ holds in glioblastoma data
joined at shared genes
EZH2reported in GBMSTAT3not yet reported in GBMZEB1reported in GBMEMT
One new link to test in GBM: STAT3 → ZEB1.
Colorectal
C1→MYC
C2→MYC
MYC→C3
…
Pancreatic
P1→MYC
P2→MYC
MYC→P3
…
Lung
Lu1→MYC
Lu2→MYC
MYC→Lu3
…
Liver
Lv1→MYC
Lv2→MYC
MYC→Lv3
…
17 claims gathered into one MYC hub, each tested in glioblastoma single-cell data
MYC
13 / 17MYC claims hold in glioblastoma
7 already known
6 new
4 don’t hold
UPP1 in glioblastoma, link by linkmechanism from a lung cancer study
UPP1TGF-β1Tregs, M2 macrophagesPI3K/AKTNF-κBPD-L1✓✗NF-κB route: proposed by SMART,matches the GBM literature
✓holds in GBM✗doesn’t holdproposed by SMART
Does UPP1 logic hold? Five target cancers
GBM ✓pancreatic ✓liver ✓colorectal ✗ovarian ✗
All 5 calls match the published literature, including the NF-κB route in GBM
colorectalliverlungovarianpancreaticseveral cancers
Explore the 3D network →
How agentic discovery transfer works

From tens of thousands of papers a year to one connected map.

Every lab does discovery transfer by hand: take a finding from a paper and check whether it holds in new data. Most findings are never tested outside their original setting. SMART gives the task to AI agents, so every finding can be tested everywhere.

01

Pull

Each month, gather new single-cell cancer studies from Nature Portfolio journals across six cancer types.

02

Summarize

Extract each paper’s clinical findings and mechanistic claims, then check every claim against the paper’s text and figures.

03

Test

Check whether the target dataset can test each claim, write and run the analysis, and report whether it holds, differs, or fails.

04

Release

Results go into cancer-wiki: the cancer-connection map and the glioblastoma mechanism network.

Explore

See what transfers: between cancers, and inside one.

Which cancers learn from each other

Each of 121 source studies was tested in six target datasets (726 reports). Averaging by source cancer gives the share of reports that transfer from each source (row) to each target (column).

OVAPANCCRCGBMLUNGLIVEROVAPANCCRCGBMLUNGLIVER
HighMiddle-HighMiddle-LowLowInfeasible
source ↓ · target →
ovarian
pancreatic
colorectal
glioblastoma
lung
liver
ovarian
67%
33%
50%
42%
17%
pancreatic
60%
60%
53%
33%
53%
colorectal
59%
53%
59%
47%
29%
glioblastoma
46%
42%
29%
33%
33%
lung
33%
63%
50%
58%
50%
liver
48%
38%
52%
45%
41%
below the column averageabove the column averageown cancer
67%60%

Ovarian and pancreatic cancer are the most mutually transferable pair.

Ovarian findings transfer into pancreatic cancer more than any other source’s (67%), and pancreatic findings transfer into ovarian cancer more than any other source’s (60%). No other pair ranks first in both directions.

A surprising pair. So why?

1

Cell of origin does not predict it

Ranked against pancreatic cancer, ovarian cancer is fourth by developmental origin but first by transfer.

Similarity in cell of origin
  1. 1liverendoderm, foregut
  2. 2lungendoderm, foregut
  3. 3colorectalendoderm, hindgut
  4. 4ovarianmesoderm
  5. 5glioblastomaectoderm
Transfer rate into pancreatic cancer
  1. 1ovarian 67%
  2. 2lung 63%
  3. 3colorectal 53%
  4. 4glioblastoma 42%
  5. 5liver 38%
2

Lipids and stroma do

All eight lipid-metabolism and fibroblast studies transfer between the pair, against at most 50% into the other cancers. Tumor cell-intrinsic findings mostly do not.

OVA ↔ PANCCRCGBMLIVERLUNG
Lipid metabolism (4)100%50%25%50%25%
Fibroblasts, CAF (4)100%50%0%0%50%
Immune environment (7)71%43%71%29%71%
Tumor cell-intrinsic (12)33%50%67%50%17%

Share of the 27 ovarian and pancreatic studies that transfer, by topic (number of studies in brackets).

3

A testable drug lead

From pancreatic cancer: PCSK9-high cells are more resistant to GPX4 inhibitors. SMART found the mechanism transfers to ovarian cancer, where PCSK9-high cells show higher cholesterol biosynthesis.

GPX4-inhibitor resistance by PCSK9 expression in 38 ovarian cancer cell lines, P = 0.044

Independent check: among 38 ovarian cancer cell lines in CTRP, PCSK9-high lines are more resistant to GPX4 inhibitors (P = 0.044).

A rare cancer borrows from the well-studied ones

Pleural mesothelioma is aggressive, asbestos-linked and has few studies of its own. Instead of waiting for its own literature, SMART tested all 121 source papers from six well-studied cancers against one mesothelioma single-cell dataset.

colorectalglioblastomaliverlungovarianpancreaticpleural mesothelioma121 papers testedTwo of the findings, zoomed in below ↓
from colorectal cancer

CYP27A1

A known marker of favorable prognosis in mesothelioma, not yet explained. SMART proposes a potential explanation: T cell infiltration.

CD8 T cell percentage in CYP27A1-low versus CYP27A1-high epithelioid pleural mesothelioma
from lung cancer

ORACLE

In TCGA-MESO, patients with high ORACLE scores have worse survival.

Kaplan–Meier curves in TCGA-MESO: ORACLE-high patients survive shorter
HR 3.04 (1.78–5.18), log-rank P = 2.2 × 10⁻⁵, n = 69

Glioblastoma mechanism network

Claims from papers on five other cancers that hold in glioblastoma data, joined wherever two claims share a gene or cell state. Drag to rotate, click a node for its claims, or start the guided tour.

Loading the network…

The network and its claims are released under CC BY-NC 4.0: free for research use with credit. Please cite our paper when you use them.

Search the network’s links

216 claims that hold in glioblastoma, forming 206 unique links. Each comes from a published paper on another cancer.

Documentation

SMART, the engine for agentic discovery transfer, and how to run it on your own data.

01What you can use it for

The same machinery answers a range of everyday questions, depending on what you point it at.

Is this paper worth my time?

Before committing bench work to a finding from a paper, see which of its claims survive in your own data. Point one paper at your dataset.

Where does my system differ from theirs?

A claim that fails is as informative as one that holds. It marks where two settings genuinely diverge. Read the negatives, not just the positives.

Does this signature work in my cancer?

Prognostic signatures built in one tumour type can retain value in others that nobody thought to check. Test it outside its home cancer.

My cancer barely has a literature.

Borrow systematically from the well-studied ones instead of by hand. Point many papers at one dataset.

What’s already known about this gene here?

Claims about one gene are scattered across papers in different cancers. Collect the ones that still hold in your data. Group tested claims by gene.

I need a hypothesis narrow enough to test.

Join claims from separate papers into a chain, then test only the link nobody has reported. Assemble across reports.

02You have a paper. Does it hold in your data?

That question is easy to ask and slow to answer. You read the paper, work out what it actually claims, decide which of those claims your dataset could speak to, write the analyses, run them, and read the output. None of it is hard. Doing it for every paper worth checking is what makes it impossible.

SMART does that for you. It takes the paper’s mechanism, tests each claim in it against your data, and reports which ones held.

the paper’s mechanism
A→B→C→D
What the study reports, across its figures and text.
tested in your dataset
A→B
✓ holds
B→C
✓ holds
C→D
✗ doesn’t
C→E
+ new
Each claim gets its own verdict, and signals the paper never reported can surface here.
SMART report illustration · this paper × your datasetTransferability level: Middle-High
A → B✓Holds in your data.
B → C✓Holds, in one group of samples.
C → D✗Does not hold.
Beyond the paper. C triggered E instead of D, with suggested follow-up.
Read a complete reportA full SMART report for one paper tested in one dataset.
Open example report

03How SMART works: six modules, 20 agents

ModuleWhat it does
Results ExtractionWorks out what the study actually claimed, checks it against the paper’s own figures and text, and pins down what each biological term means in single-cell terms.
Data InspectionDescribes your dataset as it is: the cell types present, every metadata field, and what can be measured.
TransferabilityDecides whether this mechanism can be tested here at all. If it cannot, the run stops rather than spending compute on a pairing the data cannot speak to.
Task PlanDecides which claims are testable in your data and exactly which analysis tests each one, with real parameters read from your dataset.
Code & AnalyzeWrites and runs the analysis, debugging the run until it completes.
ReportDecides whether each claim held, what contradicts the paper, what is new, and what to do next.

01Check our repo

The repository has the full instructions for the 20 agents and how to run them directly as a workflow.

02Hot papers and gold-standard datasets

We host:

  • Hot papers, with summaries and designed tests you can reuse.
  • Gold-standard datasets, with notes on how to use them well.
  • New fields. Discovery transfer is not limited to cancer, so vote for the next field to apply it to.

Vote for what you want next, or suggest your own.

Community forum

Choose what agents test next.

We host hot papers, with summaries and designed tests you can reuse, and gold-standard datasets, with notes on how to use them well. Vote for the papers and datasets you want next, or suggest your own. Discovery transfer is not limited to cancer, so you can also vote for the next field to apply it to.

Source papers

Recent papers with a clear mechanism that agents can test in single-cell data from another cancer. Star the ones worth testing next.

Gold-standard datasets

Vote for reference datasets to include in upcoming runs.
Beyond cancer

Where should agentic discovery transfer go next?

SMART isn’t cancer-specific. Any field works if it has:
  • Extractable logic: papers that state mechanisms or claims
  • Testable data: a public dataset to check them against
  • A computational test: an analysis that can confirm or refute each claim
Vote for the next field, or suggest your own.

Votes are anonymous: one per item per browser. New suggestions appear right away, marked new, and are reviewed weekly.

A message to the community

Published results are only the beginning.

Tens of thousands of cancer papers appear every year. Most findings are tested once, in the setting they came from, then cited and left behind. The field produces results far faster than it connects them.

01

Findings should enrich each other.

A mechanism found in lung may explain an outcome in liver. A signature built for one cancer may predict survival in another. A finding gains value every time it is tested somewhere new, and every test, whether it holds or fails, tells us something about both cancers.

02

Only agents can do this at scale.

Mapping every finding against every other cancer means reading, extracting, testing and reporting far more times than any lab or consortium could by hand. Agents can, systematically and with every step on record.

03

So agents should.

Not to replace investigators, but to make sure no finding goes unused. Let agents map the literature onto itself, and let people choose which leads to pursue.

Our call: agentic discovery transfer

Let agents enrich what we already know.

Treat published results as a shared map that agents keep testing and connecting. Share your papers and datasets, state your mechanisms in a form agents can read, and help decide what gets tested next.

Want to apply the same logic beyond cancer? We will actively expand SMART to new fields, so tell us where it should go next.

About

Every finding, tested beyond the cancer it came from.

Cancer research publishes tens of thousands of papers a year, and almost every finding is tested only where it was found. cancer-wiki tests them elsewhere, automatically, and keeps every claim that transfers.

Highlighted connections

  • ORACLEBuilt for lung cancer. Predicts survival in liver, pancreatic cancer and mesothelioma. None of 161 citing papers had checked.
  • Ovarian ↔ pancreaticThe most mutually transferable pair, explained by lipids and stroma. Drug lead: GPX4 inhibitors for PCSK9-low ovarian cancer.
  • Pleural mesotheliomaA rare cancer borrows from 121 papers: an explanation for a good prognosis and a survival signature.
  • UPP1Same role in glioblastoma, different route: NF-κB instead of PI3K/AKT, as the literature later confirmed.

Built with SMART, our AI agent system for agentic discovery transfer. Glioblastoma is the first full target; each release adds papers and cancers.

How to cite

Cheng L.*, Shireman J.*, Gogate A., Stangis M., Cai Y., Xie Y., Khatri P., Dey M., Kendziorski C. Agentic Discovery Transfer Enables Cross-Cancer Discovery at Scale. bioRxiv (2026). doi:10.64898/2026.10.28.762506

*These authors contributed equally.

cancer-wiki data are released under CC BY-NC 4.0: free to use and share with credit, not for commercial use.