CHILDHOOD CANCER MODEL ATLAS

Hudson's Atlas already held the research. Ippon made it reachable in plain language, and built it so an answer never says more than the data does.

Client Stories
Hudson Institute of Medical Research
Share:

WHAT THE ATLAS NEEDED

A multidisciplinary team reviews one patient on a Wednesday morning.
A child with a rare cancer, scans, a pathology report, and the next case waiting. Somewhere in the Childhood Cancer Model Atlas there may be a cell line that matches that tumour, or a compound already tested against it. Finding it means working through the Atlas, and nobody in that room has the time, so a match that already exists may go unseen.

That room is where Hudson is heading rather than where the Atlas is today, and Hudson iscareful about the distinction. The clinic aims to be a leading indicator, weighed alongside everything else on the table. Today the people asking are researchers, and they hit the same wall for different reasons.

The Atlas is open, free and multi-omics, built with partner institutes around the world. It already has a dashboard and an analytical portal, and both reward someone with an afternoon to spare and the skills to use them.
The research was already there. What it needed was a way in that fits the time people actually have.
182
cell lines in the published release
126
of them paediatric, kept apart from adult lines and controls
6
assay types, from drug screens and CRISPR through to methylation

Why Hudson, and Why Us?

Hudson built the Atlas for one purpose, to make paediatric cancer data usable for clinical translation.

Ippon's Managing Director has served on Hudson's Patient and Family Advisors Committee since 2015, and the work grew out of that relationship rather than a sales process. It shaped how we ran it: when people care this much about the outcome, the temptation is to move fast and impress, and we took the slower route instead.

Governance First

We asked the uncomfortable questions before any architecture

Before we designed anything, we went after what the data actually was, what the existing ethics approvals permitted, whether secondary use for AI processing was covered, and where the data was allowed to live. Some of those answers took weeks, and the build waited on them rather than running ahead. Hudson's answers set its boundaries.

The data

De-identified cell-line and model data, with no patient-level clinical annotation.

The ethics

Use sits inside the Atlas's existing research-only ethics scope.

The scope

Published open data only. Unpublished work stays with Hudson's researchers.

The framing

A leading indicator, read the way a clinician reads Therapeutic Guidelines.

What we built

A plain-language way into the Atlas, with the sources attached

We built a Model Context Protocol server over the Atlas. A researcher asks a question in plain language from their own Claude account and the Atlas answers, so there is no new portal to learn, no new credentials to issue, and no research data moved into somebody else's platform.

A real answer.

Asked how many cell lines there are for each tumour type, the system states the whole field first, then separates the paediatric cancer lines from the adult lines and controls.

What we deliberately did not build

Most of the design is what the system is not allowed to do

Hallucination was named as the primary risk in the very first conversation. In work that feeds clinical thinking that is the correct fear, and it shaped what follows.

Citations are the product

Every figure traces to the row it came from. The system retrieves and shows its working rather than concluding, and it does not summarise away the evidence.

The cost survives the grant cycle

The query runs on the user's own account, so Hudson carries no per-query model cost however many partner institutes pick it up.

Built to be handed over

Designed from the first week for Hudson to own and host. A not-for-profit research institute is better served owning the system outright than depending on a consultancy to run it.

How we know the citations hold

We wrote a golden dataset of 15 test questions covering query routing and citation accuracy, and validated the system against it before anyone relied on it. We use the same set as the regression test for every future release.

An agent you cannot evaluate is a demo. This one was built to be evaluated, because the people using it make decisions that matter more than any of ours.

What Hudson Can Do Now

Put a question to the Atlas in the language researchers already use, inside tools they already have. Check any figure against the source row behind it. Run it with no per-query model cost to Hudson, and own and operate it on Hudson's own environment with no ongoing Ippon dependency.
Technology
Company logoCompany logoCompany logoCompany logoCompany logoCompany logoCompany logo
"The Childhood Cancer Atlas holds years of vital paediatric cancer data, but accessing its insights once required specialist expertise. Now, our AI-powered interface puts those insights within reach of researchers, clinicians, patients and families. They can ask questions in plain language and get answers in seconds. This opens the door to new discoveries, better-informed care and greater understanding for families. The potential is enormous."
Professor Ron Firestein, MD/PhD
Head, Cancer Centre
Deputy Director, Hudson Institute of Medical Research
Product & Experience

Experience design and product strategy services that turn customer needs and business opportunities into valuable, intuitive and successful digital products.

Read more
Technology & Engineering

Software engineering, cloud and quality engineering services to build secure, scalable and reliable digital products and platforms.

Read more
Data & AI

Data and AI consulting services that help organisations build trusted data foundations, unlock intelligence and turn emerging technologies into business value.

Read more