Data-Driven Life Sciences · KTH & SciLifeLab · Autumn 2026
We send a forward-deployed AI scientist to work on your data.
Two 1-hour meetings is the whole ask. In between, someone from the course builds a working analysis on your data with an AI agent and hands back a report and an app you can run yourself.
2 hours of your time · ≈ 3 weeks end to end
Who is asking
Data-Driven Life Sciences (SK2538 / FSK3538) is a course at KTH, run with SciLifeLab by the AICell Lab for the DDLS community — master's students, PhD students, postdocs and researchers alike. We are its participants and its teaching team.
The course trains one specific role, and it can only be practised on someone else's real problem. That is the entire reason we are writing to you.
The name is borrowed from software. A forward-deployed engineer — a role made well known by Palantir, and since 2024 used at the major AI labs — is the engineer who is sent to sit with the customer, rather than the one who stays with the codebase. They learn the customer's problem first-hand and build against it directly, in days rather than quarters.
The life-science version is what this course trains:
A scientist who can sit down with someone else's biological problem, interview them until the question is precise, and come back within days with a working, checked answer — with an AI agent doing the labour and themselves supplying the judgement.
What makes it new is the agent. An AI agent can now execute almost any analysis you can specify, which moves the scarce skill from implementation to specification, direction and verification. The failure mode stops being "the code didn't run" and becomes "the code ran beautifully on the wrong question" — so the person's job is to choose the right question, then check the answer hard enough to sign it.
More on the role, and how the course is built around it, further down this page.
This is deliberately short
It is not an open-ended collaboration and we are not asking for a grant, a co-authorship or a commitment. Two 1-hour meetings, about three weeks, then it is done — and you keep whatever was built.
The problem we are both in
One person owns the question. Another owns the methods. They are rarely the same person, and the queue between them is longer than the project's patience.
What changed
Answering a real question means climbing down a ladder, from a human sentence to something a machine can run and a human can check. AI agents can now execute every rung of it.
What they cannot do is tell you that rung two was the wrong quantity to measure — that the thing you chose to count was not the thing that matters. That judgement is yours, and it is why this still needs a scientist in the loop.
Which is exactly why our students need your problem: the judgement only gets trained on questions that are real.
“Does this drug change how the cells look?”
Mean nuclear area per well, from DAPI. Only a person who knows the biology can pick this one.
Segment nuclei → measure → compare to control.
Effect size + 95% CI, FDR across 84 genes.
Hold out plate 4 · shuffle labels · buffer wells.
An AI agent can run all five. It cannot tell you that rung 02 was the wrong quantity to measure.
What we actually run
The whole course is about one thing: getting real scientific work out of an AI agent reliably. Not prompting tricks — specification, direction and verification.
Not a chatbot in a browser tab. An agent that reads your files, writes and runs code, plots, and iterates on its own output until the task is done.
Every agent run is specified, steered and checked by a course participant. They sign the result — “the agent said so” is not a methods section.
Held-out splits, shuffled labels, buffer wells. The agent is fast and fallible, so checking it is part of the work rather than an afterthought.
This is what the course teaches, and what your project is a live exercise in. You do not need to know anything about agents to take part.
The procedure
Everything in blue is time you spend. Everything else happens without you.
Your total time: two 1-hour meetings — about 2 hours, over three weeks
Tell us the question. Attach a sample file now if you like.
They interview you — and this is where you hand the data over.
An AI agent does the work; a person checks every number.
You say what is wrong. More rounds only if you both want them.
A written report and an analysis app you can run yourself.
We match you with a participant. If nobody is free, our AI data scientist runs it — every project gets an answer.
Your part, precisely
No preparation, no slides, no code. Five things, in order — and two of them take an hour.
Not a form — an assistant asks you a few short questions and writes the details down for you. You can attach a sample file there if one is handy, but it is optional and most people do not.
We pair your problem with a participant on the course and introduce you by email. If nobody is free for your project, our AI data scientist takes it on instead — either way, you get an answer.
They ask you what you want to find out, how the data was produced, what you have already tried, and what would make you distrust a result. This is also where you hand the data over — so come with it ready, or with a way to share it. Recorded for teaching only if you agree.
An AI agent does the implementation; the participant directs it and checks every number. You hear nothing for a couple of days, which is normal.
They show you what they built, and your job is to say what is wrong, what is missing and what to drop. Being told it is wrong is the expected outcome. Further rounds are optional — only if you and they both want them.
A written report, plus a small analysis app or the code itself — something that runs on your data after they are gone. Yours to keep, use, ignore or build on.
You need no programming and no AI experience. You bring the biology and the judgement about whether an answer is believable — the half of this that cannot be automated.
Why bother
Code, figures and a pipeline that still runs after they are gone — plus a short written report.
An outsider has to ask what you never say out loud. That is often worth as much as the analysis itself.
If no student takes your project, our AI data scientist runs it. Nobody who volunteers goes away empty-handed.
Free, no claim on your data, results or authorship. Use what they build, or don't.
Deadline to sign up a project.
We match you and introduce you by email.
Meeting 1 — the interview · 1 hour.
Meeting 2 — the prototype · 1 hour.
Hand-over — report, figures, app.
Our side of the bargain
We would rather have more projects than students. If nobody takes yours, our AI data scientist runs it — it interviews you in the same meeting, does the same work, and a member of the teaching team checks everything before it reaches you.
Register
Signing up takes about two minutes and gives you a private project page — upload your sample file there, watch the work move through its stages, and raise anything that comes up while it runs.
Not sure yet? The assistant answers questions too — ask it there before you commit to anything.
Deadline: Friday 25 September 2026 · contact the course team