cBioChatA Conversational Interface for cBioPortal and HTAN Data

HTAN-TBEL Data Science Mini-Symposium
October 30, 2026

cBioChat answering a question about HTAN polyp data, with a link to cBioPortal

chat.cbioportal.org

Ino de Bruijn, PhDData Visualization Team Lead, Memorial Sloan Kettering Cancer CenterHTAN Data Coordinating Center (HTAN DCC)

cBioPortal for Cancer Genomics

Make complex cancer genomic data accessible and interpretable

cBioPortal views: cohort exploration, group comparison, gene-centric queries, patient visualization

Open source and open data, maintained by MSK, Dana-Farber, UHN, CHOP, The Hyve and SE4BIO

15+ Years of cBioPortal

One of the most used tools in cancer genomics

Weekly cBioPortal users from 2010 to 2026, about 12,000 per week recently, with milestones: Start of TCGA, cBioPortal launch, AACR Project GENIE, HTAN, MSK-CHORD, cBioChat + MCP

~12kusers per week

36k+citations of the 3 core papers

550+curated studies

~36MAPI requests per month

100+registered installations

Researchers Already Ask AI Assistants

Visits to cbioportal.org referred by AI assistants

Sessions per day referred by AI assistants, Sep 2023 to Oct 2026, mostly ChatGPT and Gemini, peaking above 200 per day

From near zero to 100–200 visits a day.

How do we bring cBioPortal into this era?

Google Analytics 4, sessions by referrer (monthly mean)

Two Directions, Launched August 2026

AI in cBioPortal, and cBioPortal in any AI

AI in the website

cBioChat

Ask in plain language at chat.cbioportal.org; every answer links back to the cBioPortal view that shows it.

cBioPortal in any AI

MCP servers

Connect Claude, ChatGPT, Copilot or your own agent, and use cBioPortal data in broader workflows.

cBioPortal What's New, August 14, 2026: AI integrations, cBioPortalChat and MCP server

cBioChat: Everyday Questions

Query TCGA and look up treatments in MSK-CHORD, each confirmed in cBioPortal

cBioPortal in Your Own AI

Add the MCP connector once, then ask from the assistant you already use

mcp.cbioportal.org/db/mcp · mcp.cbioportal.org/navigator/mcp · illustration with real query results

cBioChat Usage

About 100 people a week use chat.cbioportal.org (1.2K so far this year)

Weekly active users of chat.cbioportal.org in 2026, rising from about 20 in January to about 100 per week since the August launch

MCP tool calls per week since June 2026, about 2,800 per week recently

MCP tool calls per week, from Claude, Codex, Copilot, Cursor and custom agents

Left: weekly active users, Google Analytics (chat.cbioportal.org), Jan – Oct 2026. Right: Datadog, our own benchmark runs excluded

How Do We Know the Answers Are Right?

146 curated questions, graded on every release

Example questions

Data

“How many patients are in MSK-CHORD?”→ 24,950

Analysis

“Immunotherapy response in melanoma in MSK-CHORD?”→ must say there are no melanoma patients

Out of scope

“Any flaws in the MSK-CHORD paper’s methods?”→ must decline

Pass rate, latest run per model

Claude Sonnet 5.5

74%

Claude Haiku 5.5

66%

Claude Haiku 4.5

64%

Qwen3.8 27B (open)

57%

Most common failure

Asking a clarifying question instead of acting

github.com/cBioPortal/cbioportal-mcp-qa

Beyond Bulk Genomics

More data types in cBioPortal

Number of studies by data type: clinical 548, genomic 545, mRNA 229, methylation 94, protein 84; single-cell cell-type composition 6; imaging 35, 17 and 5 studies; 229 studies (42%) with 3 or more modalities

3 studies in cBioPortalmore coming

Breast cancerOHSU · Cell Rep Med 20225 samples

Colorectal polypsVanderbilt · Cell 202161 samples

Colorectal cancerMSK · Nature 202583 samples

Linked to Dedicated Viewers

Multimodal data in cBioPortal, single-cell data in Cell Explorer, multiplex imaging in Minerva

cBioPortal for multimodal data, Cell Explorer for single-cell data, Minerva for multiplex imaging data

HTAN cBioChat Example: Two Routes to Colorectal Cancer

Chen et al., Cell 2021: Differential pre-malignant programs and microenvironment chart distinct paths to malignancy in human colorectal polyps

Conventional adenomasfrom WNT-driven stem cells

vs

Serrated polypsfrom differentiated cells via metaplasia

Two observations to reproduce (HTAN Vanderbilt, 61 polyp samples in cBioPortal)

1

Different driver mutations

APC in 85% of tubular adenomas, BRAF V600E in 89% of sessile serrated lesions

2

Different neoplastic cells (scRNA-seq)

Adenoma-specific cells (ASCs) in adenomas, serrated-specific cells (SSCs) in serrated polyps

Can cBioChat reproduce both from the data in cBioPortal?

Asking About HTAN Data

Both observations, reproduced by asking and confirmed in cBioPortal

What Reproduced

Chen et al., Cell 2021	cBioChat, and the linked cBioPortal view
“APC mutations were detected in 85% of the TAs”; BRAF V600E in 89% of SSLs	APC 11/13 tubular adenomas; BRAF V600E 8/9 sessile serrated lesions (mutation-profiled samples only)	Exact match
ASCs enriched in adenomas, SSCs in serrated polyps (scRNA-seq)	Median ASC fraction 0.30 vs 0.04, SSC 0.005 vs 0.21; cBioPortal’s test: q = 1.6e-5 and 2.5e-3	Same direction

AI in the Browser

Claude in Chrome reading a Minerva H&E slide of a tubular adenoma (HTAN Vanderbilt, HTA11_2951)

Features Shown

Bringing cBioPortal into the LLM era

1

AI in the cBioPortal Website

cBioChat at chat.cbioportal.org: ask in plain language, with answers that link back to cBioPortal

2

cBioPortal in Your AI of Choice

Connect Claude, ChatGPT, Copilot or your own agent through the MCP at mcp.cbioportal.org

3

AI in the Browser

Visualize and interpret with Claude in Chrome; with the MCP it can also query the data

Acknowledgements

cBioPortal contributors and funders

DCC team

Aaron Lisman · Xiang Li · Subhi Nandakumar