DDLS 2026 · SK2538 / FSK3538 · Module 1

Data-Driven
Life Sciences

Using AI agents to answer biological questions

Wei Ouyang · KTH Royal Institute of Technology / SciLifeLab · 25 August 2026
Science for Life Laboratory
a national research infrastructure for the life sciences, shared by KTH, KI, Stockholm University and Uppsala

This course is the teaching end of the DDLS programme

Four research areas
2020 → 2034
Cell and molecular biology · evolution and biodiversity · precision medicine · epidemiology
Eleven universities
SEK 3.55 billion
Fellows, PhD students and postdocs recruited across the country
AICell Lab, KTH
one DDLS group
AI agents for imaging and cell biology — aicell.io
This course
SK2538 · 7.5 hp
The teaching end of that programme, run since 2022

The SciLifeLab & Wallenberg National Program for Data-Driven Life Science exists to build exactly the skills this course teaches — and to find the people who will use them.

DDLS programme: scilifelab.se/data-driven · a 14-year, SEK 3.55 bn initiative funded by the Knut and Alice Wallenberg Foundation · AICell Lab: aicell.io
Module 1

By the end of today you should be able to…

data agents controls the role
  • Explain why modern biology produces more data than it can interpret
  • Describe what an AI agent is, and how it differs from a chatbot
  • Name three ways an agent fails on biological data — and the control for each
  • Describe the role this course trains you for, and how you will be assessed
These map onto the course intended learning outcomes · ddls.aicell.io/course/ddls-2026
Intended learning outcomes, DDLS 2026 course syllabus · ddls.aicell.io/course/ddls-2026
Roadmap

Six parts, eighty minutes, then discussion

IBiology asmeasurement18 min IIWhat agentschanged20 min IIIHow itfails8 min IVThe newrole15 min VHow thecourse runs12 min VIWrap-up &discussionthe rest

Interrupt me. This is the one session where asking the obvious question helps everyone.

DDLS 2026 · SK2538 / FSK3538 · Module 1 · ddls.aicell.io/course/ddls-2026
I
Part one

Biology as measurement,
and the
interpretation gap

Where the data came from

Measurement got cheap faster than computing did

$100M$10M$1M$100k$10k$1k$100 200120072010201520202026 Moore's law next-generation sequencing $525 announced: $200 → $100
Fig 1. Cost per human genome, log scale. Amber = vendor announcement. NHGRI Sequencing Cost, genome.gov

In 2008 the cost fell off a cliff and never came back. Sequencing stopped being a project and became an assay.

So what
A genome went from a fifteen-year international project to an overnight assay, and in 2026 several platforms list it at $100. Nothing else in the pipeline moved that fast.
NHGRI Genome Sequencing Program, genome.gov/sequencingcostsdata · $200 and $100 are vendor list prices at full capacity, reagents only — Illumina 2023; MGI, Ultima, Element by 2026
The single fact this whole course is a response to

We generate data exponentially. We interpret it linearly.

10⁰10²10⁴10⁶10⁸ 2001 2013 2025 people who can analyse it genomes sequenced the gap
Fig 2. Blue: real, from the sources below. Grey: schematic.log scale

Sequencing output has risen by roughly eight orders of magnitude since the first human genome. The number of people trained to analyse it has not.

The consequence
Most biological data that gets generated is never fully analysed. Not because the methods do not exist — because nobody has the time.
Genome counts: NHGRI · Stephens et al., PLOS Biology 2015, e1002195, projecting 100 million–2 billion human genomes and 2–40 exabytes of storage by 2025. Analyst curve is illustrative.
And it is not only sequencing

One afternoon on a microscope

384-well plate 384 wells × 9 fields per well × 4 channels
Fig 3. A routine high-content screen, one plate.schematic · arithmetic

384 × 9 × 4 = 13 824 images.

From one plate, in one afternoon. Nobody is going to look at these by eye, and no one is going to count cells by hand.

Plate · well
A plate is a tray of 384 tiny wells, each holding one condition — one gene silenced, one drug, one control. A whole plate is prepared and imaged together, so "which plate" is a fact about the handling, not the biology.
Plate arithmetic, not a citation: 384 wells × 9 fields × 4 channels. Typical of a high-content screen at the SciLifeLab imaging facilities
The data you will meet

The six modules, and the data each one uses

WEEK 1
Agents and data
WEEK 2
Images and microscopy
WEEK 3
Patients and cohorts
WEEK 4
Protein structure
WEEK 5
Single cells and transcriptomes
WEEK 6
Automated discovery

The domains differ. The workflow does not.

The six module topics, DDLS 2026 · guest lecturers from SciLifeLab and DDLS Fellows
And the object itself is not getting simpler

One human cell is the hardest thing we routinely try to model

~37 × 1012
cells in one human body
~20 000
protein-coding genes, expressed in different combinations
> 200
recognised cell types, and the number keeps rising

No individual can hold this. The question is not whether we need computation — it is who directs it.

Cell count: Bianconi et al., Annals of Human Biology 2013, 40(6):463–71 · gene and cell-type counts: Human Protein Atlas and Human Cell Atlas · background illustration is AI-generated, not a micrograph
Why "data-driven"

Two ways to run the scientific loop

Hypothesis-driven ideaexperimentone measurementtestData-driven measure everythingmodelmany hypothesesrank and test
Fig 4. Two ways of running the loop.schematic

Data-driven work does not replace hypothesis testing. It generates the hypotheses that hypothesis-driven experiments are good at killing.

Framing after Kell & Oliver, “Here is the evidence, now what is the hypothesis?”, BioEssays 2004, 26:99–105
Fifty years of method, in one figure

Each step moved the human further from the mechanics

Statistics you choose the model t-test on expression Classical ML you choose the features random forest on cell shape Deep learning you choose the architecture CNN on raw pixels Foundation models you choose the question describe it, it builds it

…and closer to the question.

Cellpose: Stringer et al., Nature Methods 2021 · transformers: Vaswani et al., NeurIPS 2017 · foundation models: Bommasani et al., arXiv:2108.07258
What "deep learning" actually bought biology

Learned features: you stopped having to say what to measure

area = πr² perimeter circularity …and 40 more a human decides what matters
Before. Hand-engineered features.schematic
raw pixels learned representation the network decides
After. Learned from the data. e.g. Cellpose, Stringer et al., 2021
Cellpose: Stringer, Wang, Michaelos & Pachitariu, Nature Methods 2021, 18:100–106
The proof that prediction can be trusted

AlphaFold: a fifty-year problem, closed in one competition

M K T A Y I A K Q R Q I S F V K S H F S R Q L E E R L G L I E V Q A P I amino acid sequence predicted structure
Fig 5. Sequence to structure — schematic. Jumper et al., Nature 2021

At CASP14, AlphaFold 2 reached a median GDT_TS of 92.4 — roughly the agreement you would expect between two experimental structures of the same protein.

The 2024 Nobel Prize in Chemistry went to Baker, Hassabis and Jumper. The field has made up its mind about whether this counts.

Jumper et al., “Highly accurate protein structure prediction with AlphaFold”, Nature 2021, 596:583–589 · Nobel Prize in Chemistry 2024: Baker, Hassabis, Jumper
Not hypothetical, and not far away

Alpha Cell: SciLifeLab's predictive model of the human cell

It aims to go from describing cells to predicting how they change — and then to intervening. It is coordinated a few floors from this lecture.

1Human Protein Atlasa generative model ofcell metabolic function2Space3D molecular-resolutionreference maps3Timelive imaging of statetransitions4Futurepredict the nextcell state5Controlintervene beforepathology
Fig 6. The five declared phases.scilifelab.se/alpha-cell

Programmes at this scale do not fail for lack of models. They fail for lack of people who can turn a biological question into one.

Alpha Cell, SciLifeLab · SEK 590 million, extended through 2033, funded by the Knut and Alice Wallenberg Foundation · coordinated by Jan Ellenberg and Mathias Uhlén
And yet

None of it shortened the queue

experiment waiting for someone who can analyse it analysis "can you also try…" week 0 week 20
Fig 7. The shape of a typical collaboration — schematic, not measured.schematic

The biologist owns the question; the computational scientist owns the method; and between them sits a misunderstood file format. Most projects do not fail at the model. They fail in translation.

Schematic. The pattern is described in Attwood et al., “A global perspective on evolving bioinformatics and data science training needs”, Briefings in Bioinformatics 2019, 20:398–404
The other half of why this course exists

The entry level is where AI is landing first

0%-10%-20%≈ 0%Experiencedworkers−9%Juniors atadopting firms−16%Ages 22–25,AI-exposed−20%Developersaged 22–25
Fig 8. Relative employment change since 2022, by seniority.measured

Two large studies, different data, same shape: junior hiring falls, senior employment does not. The mechanism is slower hiring rather than layoffs.

Read it carefully
This is relative change, in AI-exposed occupations, and the interpretation is contested. It is a signal, not a prophecy.
Brynjolfsson, Chandar & Chen, “Canaries in the Coal Mine?”, Stanford Digital Economy Lab, rev. Nov 2025 · Hosseini Maasoum & Lichtinger, “Seniority-Biased Technological Change”, SSRN 5425555, 2025
What that means for you, concretely

The scarce thing is not the model, and it is not the biology

A biologist alonecan ask the question,cannot build the answer An agent alonecan build anything,does not know what to build A biologist directing an agentasks the question andbuilds the answer, in a week
Fig 9. Why the third box is the one worth being.schematic

You will not be out-competed by a model. You may be out-competed by someone who directs one better — and unlike the model, that is a skill, which means it can be taught. That is what the next six weeks are.

The studies above find the decline concentrated where AI automates rather than augments a task — the distinction this course is built on
II
Part two

What an AI agent is,
and how it
differs from a chatbot

Start from what you already know

Almost everyone here has used a chatbot

your question text model an answer text
Fig 10. Text in, text out. Nothing happens in the world.schematic

It answers, it drafts, it explains. Then it stops, and you go and do the work.

Most people's mental model of AI stopped here, in about 2023. The rest of this lecture is about what it missed.

ChatGPT released 30 November 2022 · GPT-4, Claude and Gemini families followed 2023–2026
One slide of mechanism

Underneath, it is next-token prediction

MKTAYIAKQRQIEVQ…and the next symbol, and the nextI0.58L0.24V0.12probability over the next symbol
Fig 11. Schematic — illustrative probabilities, not measured.schematic
Why you should care
The same machinery reads protein sequence. ESM, AlphaFold's language-model cousins — same idea, different alphabet.

A model this simple has no notion of true. It has a notion of plausible. Hold on to that — it explains every failure later in this lecture.

Protein language models: Lin et al., “Evolutionary-scale prediction of atomic-level protein structure”, Science 2023, 379:1123–1130 (ESM-2)
The term this whole course rests on

An agent is three things added to a model

A language modeltext in, text outlanguage modeltexttext+ toolsit can act on the worldlanguage model + a loopit can look and try againlanguage model+ a goal = an agentit keeps going until donelanguage model goal
Fig 12. Each panel adds one capability. Only the last one is an agent.schematic
And the analogy, for the rest of the lecture
A very fast rotation student who has read everything, can run any software, works for hours — and remembers nothing between sessions unless you write it down. Your job is to supervise it.
Definition follows common usage in Anthropic’s and OpenAI’s agent documentation, 2024–2026 · loop formalised in Yao et al., “ReAct”, ICLR 2023
“Tools” is doing a lot of work in that definition

Tool use: giving it hands

agent run codepython, R, bash read filesyour CSVs and TIFFs searchdocs and literature write outputscripts, plots, reports
Fig 13. Every call goes out and a result comes back — that is the whole trick.schematic
The lab analogy
A student who can only talk to you is a tutor. A student with bench access is a colleague — and also a liability. Both halves of that sentence matter.
Tool use and the act–observe loop: Yao et al., “ReAct: Synergizing Reasoning and Acting in Language Models”, ICLR 2023
The mechanism

The agent loop is the whole difference

goal + context plan act run code read a file search observe correct an artefact code · figures · a report repeat

A chatbot answers. An agent acts, looks at what happened, and tries again.

Tool use and the act–observe loop: Yao et al., “ReAct: Synergizing Reasoning and Acting in Language Models”, ICLR 2023
The same loop, with a real request in it

What one turn actually looks like

You“Count nuclei per well,compare to control.” Planreads AGENTS.md, liststhe plate folder Actwrites segment.pyand runs it ObserveTypeError: plate_3 is16-bit, not 8-bit Correctfixes the dtype,reruns Delivercounts per well,one figureretry until it runs clean
Fig 14. One turn on a real high-content screen.schematic

Nobody typed the fix. Nobody checked whether plate 3 should have been rescaled or excluded, either.

Schematic, based on a real high-content screening session. Error text is illustrative
Before we can talk about memory, look at what it is given

The context window: one document, rebuilt every turn

assembled from scratch, every turn system promptthe harness: its tools, its rules AGENTS.mdthe rules you wrote for this projectfiles & tool resultswhatever it opened or ran the conversation so farevery earlier turn, replayed in full— and it is longer every turn your new messagethe only part you just typed languagemodelsent whole,every turnthe answer is appended,and the whole thing is sent againIt keeps nothing of its own.
Fig 15. The context window is not a memory. It is a document, rebuilt from scratch each turn.schematic
Context engineering
Deciding what goes into that document, and when. Anthropic calls the thing you are spending an attention budget — every token you add takes some of it. This is now most of the job.
Anthropic Engineering, “Effective context engineering for AI agents”, 29 Sept 2025 · transformer attention is n² in context length: Vaswani et al., NeurIPS 2017
Why “just paste everything in” is not a strategy

Context rot: the window is enormous, the usable part is not

MeasuredThe advertised window1,000,000 tokens ≈ 750,000 words ≈ eight novels ≈ 16,000 tokensthe part GPT-4.1 actually uses reliably,measured on a task it aces at 1,000 tokens1.6% of what the box says.SchematicWhat happens on the way there100%0accuracy1K32K128Ktokens in the contextAt 32K, 10 of 12 tested models score below halfof what they score on the same task at 1K.Same question. Same needle. Just more around it.
Fig 16. Left: two measured numbers, same model. Right: the shape of the decline between them.left measured · right schematic
Context rot
Accuracy falls as the context grows, long before the window is full. A full window is not a used one.
Window: OpenAI GPT-4.1 model card 2025 · effective length and the 32K result: Modarressi et al., “NoLiMa”, ICML 2025 · Liu et al., TACL 2024, 12:157–173 · “Context Rot”, Chroma Research, 2025
What happens when you keep going anyway

Compaction: it does not stop, it quietly throws things away

context limitthe conversation growsyouplate 3 was re-imaged on Tuesdayplate 3 = re-imagedagentreads the folder — 384 wells, 16-bityoucolumn 24 is buffer, drop itcol 24 = bufferagentwrites segment.py, runs ityou“obviously” means log₂ here“obviously” = log₂agent1,200 lines of output, 40 minutes summary of the earlier conversationyounow compare the two conditionsagentwhich column was the control? auto-compaction: the harness summarises and continues no longer in the contextNobody told you these were dropped.A model chose what was worth keeping,and it does not know your experiment.
Fig 17. Auto-compaction: the earlier turns are replaced by a summary so the session can continue.schematic
The consequence
Anything you said only in the chat has a shelf life — it lasts until the window fills, and no longer.
Anthropic Engineering, “Effective context engineering for AI agents”, 29 Sept 2025
The part everyone skips, and then suffers

Persistence: what matters goes on disk, not in the chat

It remembers fine inside one conversation. Nothing survives the compaction and the closed tab except what was written to a file — and it does not matter what kind of file.

Next session · nothing on disknothing written down “Which column is the control?”“Should I drop plate 3?” you answer it againNext session · a few files on disk AGENTS.md docs/traps.md notes/plate-qc.md reads them, then startswith the traps already known it picks up where you left off
Fig 18. The same agent, the same task, one week later.schematic

Notes, a data dictionary, a script, a README — all of it counts. In tomorrow's lab you start from an empty folder, and what you leave on disk is what survives the week.

Writing state out to files as the durable memory of an agent: Anthropic Engineering, “Effective context engineering for AI agents”, 2025 · the same practice underlies AGENTS.md, agent skills and project docs
One of those files is not like the others

AGENTS.md: the briefing you would give a new lab member

On disk · lay it out however you like AGENTS.mdthe project, the rules, and where to lookloaded every turn docs/data.mdwhere it lives, how to load it skills/plate-qc/SKILL.mda reusable procedure, and its scripts docs/traps.mdwhat has already bitten us notes/2026-08-25.mdwhat happened in today's sessionany layout works — opened only when the task needs themIn the context windowsystem promptalwaysAGENTS.mdalwaysthe conversationalwaysdocs/traps.mdon demandAGENTS.md is read before you have askedanything at all. Every turn. Keep itunder about 200 lines.
Fig 19. The brief is always loaded; what it points at is opened on demand.schematic
Three jobs, one file
What we are doing — the project in a paragraph. How we work here — the conventions, and what it must never do. Where the rest lives — pointers to the files below, opened only when needed.
AGENTS.md open format (Agentic AI Foundation / Linux Foundation), 60,000+ public repositories, 2026 · just-in-time retrieval: Anthropic Engineering, 2025 · the 200-line ceiling is this course’s convention, not the spec
Some of those files are more than notes

Agent skills: a folder that teaches it a procedure

A skill is a folder on disk: a SKILL.md saying when to use it and how, plus the scripts it needs. The agent always sees the name and one line; it opens the folder only when the task matches.

skills/ · on diskthe context windowplate-qc/flag wells that failed QCnuclei-seg/segment nuclei, then spot-checkfigure-style/our figure conventionsthe indexalwaysplate-qcflag failed wellsnuclei-segsegment nucleifigure-stylefigure conventionsnuclei-seg/on demandSKILL.md · run_seg.py · examples.mdwhen to use it, how to do it, and the scriptnames onlythe whole folder
Fig 20. Ask for segmentation and only that folder opens. Fifty skills cost fifty lines, not fifty procedures.schematic
Why you should care
The tacit things — your plate layout, your QC threshold, how your microscope names channels — become a skill once, and every session after that already has them.
Agent Skills and the open SKILL.md format: Anthropic, “Equipping agents for the real world with Agent Skills”, 2025 · progressive disclosure is the same idea as just-in-time retrieval on the previous slide
Model, tools, files, skills — something has to hold them together

The harness: everything wrapped around the model

the harnessthe loopact, look, try againthe toolswhat it may callthe contextwhat goes in each turnthe filesAGENTS.md, docs, notesthe skillsprocedures it can loadthe permissionswhat it may never dolanguage modelClaudeGPTGeminiyou shapemost of thisswap the model,keep the harness
Fig 21. Claude Code, Cursor, a notebook agent: the same models, different harnesses.schematic
Why the word matters — this is the debugging question
When the answer is wrong, ask which layer failed. The model reasoning badly? Rare. A missing file, a tool it was never given, a rule nobody wrote into AGENTS.md? Almost always. You will train a model or two in this course — but you will spend far more of it shaping the harness around one.
“Harness” follows common usage in agent engineering, 2024–2026 · in this course the harness is Claude Code + your AGENTS.md + your skills + the tools you allow
Now put all of it — agent, context, files, harness — on one question

"Does knocking down this gene change nuclear shape?"

Six steps, one afternoon, and you write none of the code. Watch which parts the harness supplies — and which part only you can.

01
Look at the data
tool · reads your TIFFs
02
Segment nuclei
skill · segment, then spot-check
03
Measure shape
tool · area and eccentricity
04
Correct for plate
context · you wrote this rule down
05
Test
tool · effect size, CI, correction
06
One figure
output · a file on disk, not chat
Step 04 is the whole lecture
Five of these six the agent will reach for on its own. Step 04 it will not — plate correction is in your AGENTS.md or it does not happen. The model was never the weak link. The harness was missing a line.
Schematic of a standard high-content imaging workflow · plate correction: Caicedo et al., Nature Methods 2017, 14:849–863
III
Part three

Three failure modes,
and the control
for each

Before anyone gets carried away

Three ways it goes wrong on your data

Hallucination
It invents things that fit
A gene symbol that does not exist. A column that was never in your file. A citation with a fake DOI.
Sycophancy
It agrees with you
Tell it your hypothesis and it will help you support it. Your bad idea gets excellent engineering.
It has not met your data
No lab knowledge
It does not know plate 3 was imaged a week later, or that your "control" means something specific here.
The one that will actually get you
Code that runs is not code that is right. It will report a beautiful p = 0.001 driven entirely by which plate a cell came from, and nothing will turn red.
plate 1plate 3uncorrected“p = 0.001”corrected for plateno effect
Fig 22. The same cells, before and after correcting for plate.schematic
Sycophancy: Sharma et al., “Towards Understanding Sycophancy in Language Models”, ICLR 2024 · hallucination remains an open problem across all current models
The good news

Controls: you already know how to handle an unreliable instrument

effect size real labels strong shuffled gone — good shuffled still there — leak
Fig 23. The label-shuffle control — schematic.schematic
  • Shuffle the labels. If the effect survives, you have leakage.
  • Beat a dumb baseline. Mean predictor, majority class.
  • Check the split. Same patient, same plate, both sides — the classic disaster.
  • Ask it to argue against itself. "Three reasons this is an artefact."

None of this is new. It is the scientific method, pointed at a new instrument.

Batch effects in high-throughput biology: Leek et al., Nature Reviews Genetics 2010, 11:733–739
What this does to your skill set

The hard part used to be the code

WAS THE HARD PART IS THE HARD PART NOW writing the syntax knowing which library implementing the method making the plot look right asking a question worth answering specifying it without ambiguity knowing what the data can support designing the test that proves you wrong the scarce part moved

Every item on the right is domain judgement or communication — neither is taught by a Python tutorial, and both are exactly what the six labs make you practise.

DDLS 2026 · SK2538 / FSK3538 · Module 1 · ddls.aicell.io/course/ddls-2026
IV
Part four

The
forward-deployed
scientist

Back to the queue

The ownership gap: nobody owns both ends of the problem

The scientistowns the questionowns the sampleknows what wrong looks like The analystowns the methodsowns the pipelinesbooked for four monthsthe gapwhere most projects die

The role below exists to collapse this gap into one person.

DDLS 2026 · SK2538 / FSK3538 · Module 1 · ddls.aicell.io/course/ddls-2026
A job title borrowed from software, and adapted

The forward-deployed scientist

The scientistowns the question The methodsand the compute one person, both endsinterviewsdirects
Fig 24. The gap from the last slide, closed by one person.schematic
Definition
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.
“Forward-deployed engineer” as practised at Palantir and, since 2024, at the major AI labs
The shape of every lab, and of the final project

Six stations

01 Interview record, probe, steer 02 Specify goal → machine task 03 Build prototype, fast 04 Validate controls and splits 05 Deliver show it, be corrected 06 Close report and talk the next interview starts from a higher floor

Every computer lab rehearses 01 to 05. The final project runs all six, over several weeks, with a real person on the other end.

DDLS 2026 · SK2538 / FSK3538 · Module 1 · ddls.aicell.io/course/ddls-2026
Station 01 · the skill the whole course is built on

You are not filling in a form. You are steering a conversation.

The callZoom or in person —with the recorder on You steerquestions aimed at whatthe agent will need later Transcribethe audio, written outverbatim — not your notesWork it with the agentextract, summarise, andname what is missingAGENTS.mdthe briefspec.mdevery detailTwo documentswritten from the transcript,never from memory
Fig 25. Station 01, end to end. Every arrow is a step you actually perform.schematic
What you leave with
You do not write either file from memory. You paste the transcript to the agent and work through it together: AGENTS.md is the short brief it reads at the start of every session; spec.md holds every number, path and exception behind it — including the throwaway sentence about plate 3 that turns out to be the whole problem.
DDLS 2026 · SK2538 / FSK3538 · Module 1 · ddls.aicell.io/course/ddls-2026
Station 01 · what you must not leave without

The four probes you come back with

THE GOAL
  • What decision changes when you know?
  • What would convince you it is real?
  • Who reads the answer, and in what form?
THE DATA
  • Table, images, sequences, something else?
  • How much, and where does it live?
  • What is in the metadata — and who typed it?
  • How do I load it? Is there a parser already?
WHAT YOU TRIED
  • What have you already run?
  • What worked, and what failed?
  • Why do you think it failed?
WHAT DONE LOOKS LIKE
  • What outcome do you expect?
  • Do you have a hypothesis, or is this open?
  • What would make you distrust the result?

If you cannot answer all four afterwards, you have not finished the interview — book the second call.

DDLS 2026 · SK2538 / FSK3538 · Module 1 · ddls.aicell.io/course/ddls-2026
Station 01 · they will not adapt to you

Every vague answer has one follow-up that fixes it

Data owners speak their own language. Assume nothing is defined until you have made it definite.

“It's just some imaging data.”
“How many files, what format, and can you send me one right now?”
“We tried machine learning, it didn't work.”
“Which method, on which subset, and what did the output look like?”
“We want to know if the treatment does anything.”
“Anything compared to what — and measured how?”
“The metadata is in the spreadsheet.”
“Who fills it in, and has the format changed over the years?”

Ask for one real file before the call ends. Everything you were told will turn out to be slightly different from what is in it.

DDLS 2026 · SK2538 / FSK3538 · Module 1 · ddls.aicell.io/course/ddls-2026
Station 02 · the translation nobody teaches

A scientific goal is not yet something a machine can optimise

The scientific question“Does this drug change how the cells look?”A measurable quantitymean nuclear area per well, from DAPIA machine tasksegment nuclei → measure → compare to controlAn objective and a metriceffect size + 95% CI, FDR across 84 genesA split and a controlhold out plate 4 · shuffle labels · buffer wells
Fig 26. Five rungs from a question to something an agent can run and you can check.schematic

The agent can do every rung. It cannot tell you that rung two was the wrong quantity to measure.

Worked example follows the high-content imaging workflow used in the Module 2 computer lab
Station 02 · the single most valuable skill in this course

AGENTS.md: human-shaped question → machine-shaped brief

“We've got these plates from thesiRNA screen and the nuclei just look…different in some of them? Bea thoughtmaybe it's real. Could you have a look?”from the transcript — human-shaped, hedged, out of orderwith the agentAGENTS.mdGOAL Does knockdown of any of 84 genes change mean nuclear area vs control?DATA plate_1..4/ ch1 = DAPI meta.csvMUST NOT Pool plates without correcting.machine-shaped: complete, checkable — the rest lives in spec.md
Fig 27. Nothing was added. It was made definite.schematic

Everything on the right was already implied on the left. Your job is to make it explicit before the agent has to guess.

DDLS 2026 · SK2538 / FSK3538 · Module 1 · worked example from the Module 2 imaging lab
Station 02 · what makes that translation possible

You cannot look this up. You get it by doing it.

Nothing on the last two slides was typing. Every rung of that ladder was a judgement — and three things have to already be in your head before you can make one.

PRINCIPLES
What a method assumes
Not the derivation — the assumptions. What it needs from the data, what it gives back, and the condition under which its answer means nothing.
INTUITION
Whether this data fits that method
You notice the mismatch before you can prove it — the sample size, the confound, the label that is really a batch. Then you go and check.
JUDGEMENT
Which rung is the risky one
Seeing that the result turns on rung two — the quantity you chose to measure — and not on which model you put on top of it.
Where it comes from
Reps, and nothing else. Six labs, six domains, six families of method — each one an interview you have to translate yourself, on data you have never seen. The labs are not exercises about the lecture. They are the training set for this.
DDLS 2026 · SK2538 / FSK3538 · Module 1 · ddls.aicell.io/course/ddls-2026
Station 02 · the objection everybody has at this point

“Do I have to be an expert in all of them?” No.

You will translate onto methods you have never run, in fields you do not work in. That is the normal case, not the failure case — and it is exactly what six different domains in six weeks is for.

SIX LABS · SIX DOMAINS · SIX FAMILIES OF METHODWEEK 1Agents and dataWEEK 2Images and microscopyWEEK 3Patients and cohortsWEEK 4Protein structureWEEK 5Single cellsWEEK 6Automated discoveryTransferable judgementthe questions, not the answersA field you have neverworked in. A method youhave never run. You can still do the job.
Fig 28. Enough cases, and the skill stops being about the case.schematic
The goal of the training
To walk into an unfamiliar domain, with a method you have not used, and still scope it, check it and contribute something real — and be able to say “this will work”, and give the reason. Nobody gets there in six weeks. Six labs and one project is where it starts.
DDLS 2026 · SK2538 / FSK3538 · Module 1 · ddls.aicell.io/course/ddls-2026
Stations 03 to 06

Build, validate, deliver, close

03 · Build
Smallest thing that works
Make it explain the data back to you first. One plate, one gene, a stupid baseline. End-to-end before good.
04 · Validate
Try to break it
Controls, shuffles, the split. Read the code where the data is filtered and joined — that is where silent errors live.
05 · Deliver
Runs without you
A notebook that executes on their machine, the figure they asked for, and the caveats in writing.
06 · Close
What opens next
Separate the result from the story about the result. The best delivery ends with a better experiment.
The line you do not cross
You sign the result. "The agent said so" is not a methods section, and it is not a defence in a viva.
DDLS 2026 · SK2538 / FSK3538 · Module 1 · ddls.aicell.io/course/ddls-2026
A course rule, and the reason for it

You may not use your own problem

Many of you have your own dataset. Using it is the worst way to learn this: the context is already in your head.

Your own problemplate 3 was re-imagedcolumn 24 is bufferBea's replicate is bad“obviously” means log₂what gets said out loudgoal: compareconditions the agent guessesSomeone else's problemplate 3 was re-imagedcolumn 24 is bufferBea's replicate is bad“obviously” means log₂what gets said out loudgoal, data, traps,controls, done-when the agent can work
Fig 29. The curse of knowledge: what you already know, you do not say.schematic

Being an outsider forces the context out into the open, where the agent can read it.

The curse of knowledge: Camerer, Loewenstein & Weber, Journal of Political Economy 1989, 97:1232–1254
V
Part five

How the DDLS course
runs, and how
you are assessed

Six weeks · 25 August – 2 October

Your week: lecture, computer lab, seminar

TUESDAY Lecture mostly recorded · the week's domain WEDNESDAY 13:00–17:00 Computer lab live · mandatory · where the skill is built FRIDAY 10:00–12:00 Seminar live · mandatory · you present Final project — a real problem owner, running underneath from week 2
Fig 30. The weekly rhythm.schematic
7.5 ECTS · labs 2.0 hp · project 3.0 hp · oral exam 2.5 hp · all sessions online via Zoom
DDLS 2026 schedule · teaching 25 August – 2 October 2026, course period to 23 October · ddls.aicell.io/course/ddls-2026/schedule
Wednesday · mandatory · the heart of the course

In the lab you will talk to two agents

Agent Aproblem owner Youthe only link Agent Bthe analystyou interviewyou directnothing passes between them except what you write down
Fig 31. One plays a researcher who does not volunteer the important parts. The other is an empty folder.schematic

Agent A behaves like a real scientist: it answers a good question honestly and a vague one uselessly. Agent B becomes whatever you configure it to be.

DDLS 2026 · SK2538 / FSK3538 · Module 1 · computer lab, Wednesdays 13:00–17:00
Both of them, inside one four-hour block

The computer lab: five steps, four hours

Agent A problem owner 01 INTERVIEW the brief goal · data · traps 02 TRANSLATE in the project, from a recording empty folder AGENTS.md 03 CONFIGURE Agent B the analyst 04 DIRECT a report with your caveats on it 05 DELIVER
Fig 32. One computer-lab session, end to end.schematic
DDLS 2026 · SK2538 / FSK3538 · Module 1 · ddls.aicell.io/course/ddls-2026
Assessment · pass/fail

You submit the transcript, not just the result

interview log analysis log the report AI-written, checked
Fig 33. One submission, three parts.schematic

The transcript is where the skill is visible. We read it for:

  • Did you find the question behind the question?
  • Where did you refuse what the agent gave you?
  • What did you check, and what did you take on faith?
DDLS 2026 · SK2538 / FSK3538 · Module 1 · ddls.aicell.io/course/ddls-2026
Friday 10:00–12:00 · mandatory

The seminar: everyone prepares, we draw the presenters

everyone prepares · 7–10 present · 7 min + 3 discussion
Fig 34. Random selection, every week.schematic · n = 26 signed up

You present the work you did in that week's lab: the problem, what you built, a critical read of the method, and what the literature says about it.

The blunt bit
Drawn with nothing prepared is a fail for that seminar. It exists so that everyone spends Thursday thinking critically about their own work.
Cohort figure: 26 participants signed up for DDLS 2026 · 7–10 drawn per seminar
The final project · starts in week two

Where your client comes from

A volunteer from the poolwe put out a call to SciLifeLab;you pick one from the list Someone you finda group you know, a facility,anyone with a real dataset A simulated ownerlast resort only — an agentplaying a data owner
Fig 35. Three routes, in order of preference.schematic
The rule, again
It must be someone else's problem, and it must be different from what you already work on. If you are stuck finding one, tell us early — the pool exists for exactly that.
Call for problems goes to the SciLifeLab community · scilifelab.se
The final project · how it actually runs

Two calls, a prototype, and a few rounds of being wrong

Call 1~2 h, recordedunderstand the problemand the data Build< 1 weekbrief → agent →working prototype Call 2show the prototypewhat is wrong, what ismissing, what to drop Iterate2–3 roundsrefine against theirreal judgement Hand overreport + talksomething they can runwithout youeach round the question gets sharper
Fig 36. The project is a client relationship, not a submission.schematic

Along the way you learn the things nobody examines directly: judging whether the agent is doing the right thing, letting it run long without losing control, and turning a result into a report and a talk.

DDLS 2026 · SK2538 / FSK3538 · Module 1 · ddls.aicell.io/course/ddls-2026
What the project produces, and how it is graded

A product, a talk, a report

The product
Something that works
Code, figures, a pipeline that runs on the client's data and can be handed over. The primary deliverable.
The talk
You, defending it
What the problem was, what you built, what you checked, and what you would not claim.
The report
Written, AI-assisted
Drafted with your agent, disclosed as such, and accurate because you checked every number in it.
TrackOral presentationProject report and labsCourse grade
Master's studentsMandatoryPass / failA–F, from the oral
PhD students and othersOptional, welcomePass / failPass / fail
Credit breakdown: labs 2.0 hp · project 3.0 hp · oral exam 2.5 hp · KTH course SK2538 (Master’s) and FSK3538 (PhD)
Attendance, and what to do when you cannot make it

Two of the three sessions are mandatory

Tuesday 10:00–12:00
Lecture — optional
Mostly recorded, and the slides go up either way. Come if you can: it is the one place to ask live.
Wednesday 13:00–17:00
Computer lab — mandatory
Live. This is where the skill is built, and it does not transfer from a recording.
Friday 10:00–12:00
Seminar — mandatory
Live. Everyone prepares and we draw the presenters, so an empty chair is an empty slot.
If you have to miss one
You may miss one mandatory session in the whole course — a lab or a seminar, not one of each. Tell us before the session, not after — illness, fieldwork, a conference are all normal, and we judge case by case. What we cannot work around is finding out afterwards.
Master’s students: the final oral presentation is mandatory — it carries your A–F grade.
Absence: email ddls-course@scilifelab.se in advance · DDLS 2026 · SK2538 / FSK3538
All of that runs under one rule

The AI policy: use everything, disclose everything, own everything

use it disclose it own it protect the data verify refs
  • Any model, any tool. There is no cheating here — the tools are the subject of the course.
  • Attach your chat history to graded work. We want to see how you worked, not only what came out.
  • Every number in your report is yours. If it is wrong, it is wrong under your name.
  • Never upload data you do not have the right to upload — with a real client, ask them explicitly what may leave their machine.
  • Verify every reference. A plausible DOI is not a real one.
KTH policy on generative AI in examination · disclosure is required, use is not penalised
VI
Part six

Wrap-up
and discussion

Back to where we started

You should now be able to…

data agents controls the role
  • Say why biology produces more data than it can interpret — measurement got cheap; interpretation did not
  • Define an agent — a model in a loop, with tools and a goal
  • Name three failure modes and their controls — hallucination, sycophancy, batch effects; shuffle, baseline, split
  • Describe the role and the assessment — forward-deployed scientist; labs, seminars, one real project

In six weeks you should be able to sit down with a scientist at SciLifeLab for an hour and leave with a working, checked analysis by the end of the week.

DDLS 2026 · SK2538 / FSK3538 · Module 1 · ddls.aicell.io/course/ddls-2026
Before 13:00 tomorrow

Four things before tomorrow

01
Get an agent running
Claude Code, Codex, Gemini CLI, Cursor — your choice. Install it and make it do one small real task today.
02
Check the prerequisites
You will not write much code, but you must be able to read it.
03
Clone the material repo
Read the week 1 folder before the lab.
04
Bring a machine
One you are allowed to install software on.
ddls.aicell.io/course/ddls-2026 · github.com/aicell-lab/ddls-course-2026-material · ddls-course@scilifelab.se
DDLS 2026 · SK2538 / FSK3538 · Module 1 · ddls.aicell.io/course/ddls-2026
Practising what the AI policy slide preaches

AI disclosure: how these slides were made

Disclosure
This deck was built with custom agents running on Claude Code: one to draft and lay out the slides, one to generate the illustrations, and one to validate every slide for overflowing text, colliding labels, unreadable font sizes and uncited figures.
The agent did
Layout, figures, checking
Every diagram is SVG it authored. The validator it wrote caught layout bugs I could not see by eye.
I did
The argument and the numbers
Structure, claims, and which numbers earn a slide. Data plots come from published series, never from a model.
Not delegated
The wording and the claims
Every slide proof-read, tuned and signed off by hand. Sources are printed so you can check them — a plausible DOI is not a real one.

The whole toolchain is in the course repository. You are expected to work the same way, and to disclose it the same way.

Deck source, figure scripts and validator: github.com/aicell-lab/ddls-course-2026-material — week-1/introduction-lecture
Now the useful part

Questions —
and one I
have for you

What is a problem in your own lab that you would bring to a forward-deployed scientist?

Wei Ouyang · wei.ouyang@scilifelab.se · DDLS 2026 · Module 1