Cybergenic

Cancer gene co-occurrence and exclusivity in tumour cohorts, with confound controls and exact tests.

View on cybergenic.imTry this MCPSponsored

Hosted MCP Server

npx add-mcp 'https://cybergenic.im/mcp?via=mcpservers'

Installs into Claude Code, Codex, Cursor and more

Documentation

Ask Cybergenic from your AI assistant

Cybergenic's evidence is available to AI assistants and agents over the Model Context Protocol. Add one address and your assistant can check whether two genes' mutations co-occur or avoid each other across tumour cohorts, read the engine's findings with their verdicts, and run exact statistical tests. Free and read-only, with no account and no key.

Server address

https://cybergenic.im/mcp

Streamable HTTP, no authentication. Speaks the 2026-07-28 protocol and the 2025 revisions clients still use.

Listed in the Official MCP Registry as im.cybergenic/mcp and on Glama.

Connect in a minute

Claude (claude.ai and the desktop app) Settings, then Connectors, then Add custom connector. Give it a name and paste the address.
```
https://cybergenic.im/mcp
```

Claude Code One command in a terminal.

claude mcp add --transport http cybergenic https://cybergenic.im/mcp

Cursor Add it to ~/.cursor/mcp.json (or.cursor/mcp.json in a project).

{
  "mcpServers": {
    "cybergenic": {
      "url": "https://cybergenic.im/mcp"
    }
  }
}

VS Code (GitHub Copilot agent mode) Add it to.vscode/mcp.json in your workspace.

{
  "servers": {
    "cybergenic": {
      "type": "http",
      "url": "https://cybergenic.im/mcp"
    }
  }
}

ChatGPT, and any other MCP client Add a custom connector or MCP server with this address and the Streamable HTTP transport, and no authentication.
```
https://cybergenic.im/mcp
```

Test it from a terminal List the tools with one request, no set-up.

curl -s https://cybergenic.im/mcp \
  -H 'content-type: application/json' \
  -H 'accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list","params":{}}'

Then ask it things like

“Do KEAP1 and STK11 mutations co-occur in lung cancer, and does that hold within tumour types?”
“Which genes are mutually exclusive with TP53 in breast cancer tumours?”
“What has Cybergenic found about KRAS in pancreatic cancer, and which findings held up?”
“How reliable is Cybergenic's track record so far?”
“Run Fisher's exact test on 12, 5, 29 and 954, and give me the odds ratio with its interval.”

The tools

5 tools read the engine's evidence and 4 compute exact statistics. Each returns a plain-language summary your assistant can quote and the structured numbers behind it; evidence results carry a permanent address to cite.

Evidence

lookup_gene_pair Look up a gene pair across tumour cohorts Co-occurrence or mutual exclusivity of two genes in every cohort that tested them, with controls. Example: {"gene_a":"KEAP1","gene_b":"STK11"}
gene_partners Find a gene's strongest co-alteration partners The genes that co-occur with, or avoid, one gene in patient tumours: pairs the scan keeps, strongest first. Example: {"gene":"KEAP1","cancer":"NSCLC"}
search_findings Search Cybergenic's published findings The engine's findings with measured verdicts, searched like the site: genes, cancers, abbreviations. Example: {"query":"KRAS pancreatic","sort":"signal","limit":5}
get_finding Read one finding's full record One finding in full: data, statistics, verdict and why, checks, prediction lock, citation.
get_track_record Read the engine's live track record Locked predictions, out-of-sample replication against matched controls, and the frontier, with limits.

Statistics

fisher_exact_test Fisher's exact test (2x2) Exact p-values, sample and conditional odds ratios with an exact interval. Example: {"a":12,"b":5,"c":29,"d":954}
chi_square_test Chi-square test of independence Pearson chi-square, Yates and G-test for 2x2 up to 10x10 tables, with residuals. Example: {"table":[[20,15],[30,35]]}
odds_ratio_relative_risk Odds ratio, relative risk and NNT Odds ratio, relative risk, risk difference and NNT with confidence intervals. Example: {"a":15,"b":85,"c":30,"d":70}
fdr_correction FDR and multiple-testing correction Benjamini-Hochberg q-values plus BY, Holm, Hochberg and Bonferroni for up to 10,000 p-values. Example: {"p_values":[0.0001,0.004,0.019,0.03,0.2,0.5],"alpha":0.05}

The statistics tools run the same code as the calculators on this site, validated against SciPy, statsmodels and R.

What the numbers are, and are not

Associations, not mechanisms. Two genes that co-occur or avoid each other in tumours are a statistical pattern. It can suggest biology; it does not show one causes or protects against the other, and it is not clinical evidence.
Corrected for the number of tests. Each q-value is a Benjamini-Hochberg correction over every pair tested in that cohort, and the test accounts for how many mutations each tumour carries, so a gene-rich tumour does not manufacture co-occurrence.
Controlled, with the limits stated. Each result names the control it used (molecular subtype, lineage or histology strata) and what that control leaves open, and whether the effect held, faded or reversed within strata.
Verdicts, not just p-values. Findings carry the verdict the engine measured and the reason for it. A hypothesis attached to a finding is written by a language model: it is an annotation, never a result.
Counted by independent source. MSK-IMPACT 50K and MSK-CHORD share most of their patients, so agreement between them is not replication. Every result counts independent sources separately from cohorts, and states a direction only where the data establish one.
The scan's own gates. A pair can be significant yet too weak to count: the scan keeps it only with at least a two-fold effect, a consistent direction and enough tumours. Results say which pairs pass and why the others do not.

The full account is in the methodology, including the live track record the get_track_record tool reads.

Questions researchers ask

What is an MCP server?

The Model Context Protocol is the open standard AI assistants use to call outside tools. An MCP server publishes a set of tools; once you add its address, your assistant can call them while it answers you. Cybergenic's server is a remote one, so there is nothing to install.

Is it free, and do I need an account or an API key?

It is free and needs no account and no key. Every tool is read-only: it can look things up and compute, never change anything.

What data does it answer from?

Driver mutations in public tumour sequencing cohorts read through cBioPortal: MSK-IMPACT 50K (2026), MSK-CHORD (2024), the TCGA PanCancer Atlas (2018), OrigiMed China (2020) and METABRIC. The gene-pair tools read every pair the engine's scan tested in each cohort, significant or not; the findings tools read the engine's published findings with their verdicts.

How current is it?

The same as the website: the discovery scan re-reads the cohorts on a rolling schedule, and findings and verdicts change as the engine re-measures them, several times an hour. Each result names the dataset and scan version it came from.

Can I cite a result?

Yes. Every finding has a permanent address on this site, and get_finding returns a BibTeX entry with the date the finding was first seen and, when the engine locked a prediction for it, the registry hash. For a gene-pair statistic, cite the dataset and the scan version the result names.

Is this medical advice?

No. The statistics are associations in tumour sequencing data, not causal or clinical evidence, and a finding's hypothesis text is a model-written annotation, not a result. Research use only.

What do you record when my assistant calls a tool?

The name of the tool, whether it succeeded, how long it took and the name of the client (for example Claude), counted like a page view: your IP address is hashed with a salt that changes daily and is never stored. We never record what you asked: the genes, tables and p-values in a call are not logged.

Prefer the website or a file?

Research use only, not medical advice. Statistics derive from public cohorts read through cBioPortal; each study has its own terms of use.