DDLS 2026 · SK2538 / FSK3538 · Module 4 · Live lab walkthrough

Trust a Protein
Structure

Interview a structural-biology data owner. Direct an agent. Decide whether the model actually answers the question.

Computer Lab 4 · Wednesday 13:00–17:00 · advance with →  ·  press N for speaker notes
Welcome to Week 4

Same job — new data: a protein structure

You run the exact forward-deployed-scientist loop from weeks 1–3. What changes: the data is now a protein sequence and a 3D structure with a confidence score. The bar goes up in one specific way — a model can look convincing and still be confident about the wrong thing for the owner's question.

The skill on trial
Trust a structure — read the right confidence and check it's their protein — not just fetch one.
You produce
A small viewer app (structure + confidence), a structured report, and a public GitHub repo.
Graded on
Pass/fail — and we read the transcript: did you find the question behind the question?
Computer Lab 4 · full instructions: ddls.aicell.io/course/ddls-2026/module-4/lab/
How today runs · 13:00–17:00

One loop, once — checked, and defensible

  • 13:15 · Set up Pi and open the portal (do this first — it installs while you interview)
  • 13:30 · Interview your data owner (Agent A)
  • 14:00 · Brief & direct Pi — write AGENTS.md / spec.md, review by hand, then direct the analysis
  • 15:45 · Hard switch — stop analysing, build the viewer
  • 16:45 · Wrap — write the report, publish to GitHub, hand in

The analysis loop is cheap — a couple of minutes of agent work for a few cents. The clock goes to your thinking, reviewing and checking. A checked partial result you can show beats an unfinished perfect one.

Treat 15:45 as a hard switch to building · seminar prep is a separate job on Thursday
The workflow · you are the only link between the two agents

Interview → direct → check → build → hand in

01
Interview A
The data owner. Pin down the molecule, the claim, and what they trust.
02
Brief & direct Pi
Turn the interview into AGENTS.md + spec.md; direct the analysis.
03
Check results
Read the right confidence; look at the model; is it their protein?
04
Build
A 3D structure + confidence viewer, and a report.
05
Publish & hand in
Push to GitHub; submit transcripts, app, report, repo URL.

Nothing passes between the two agents except what you write down. That channel — you — is exactly the examined skill.

You get one data owner, assigned at random · you work theirs, never your own protein
Step 1 · Interview the data owner (Agent A)

Find what they intend to trust — and for what decision

Pin down
The molecule (one chain or a complex? how long?), the files, the exact claim and which residues it touches, and how sure they are — right there.
Their confidence is the clue
A structure viewer makes everything look solid. Get them to describe why they trust the model — without tipping them off that you doubt it.

Use the Interview playbook and the two coaching helpers (Coach me · Suggest a question) in the portal. A vague question earns a vague answer; a precise one earns the truth. Then play it back and let them correct you.

The chat is logged · download the transcript at the end — your analyst reads the transcript, never your notes
Step 2 · Brief & direct Pi (Agent B) — your empty-folder analyst

Translate the interview into a brief the agent follows

  • AGENTS.md — how Pi operates here (env, where the data is, the fold-service instruction, a git commit rule)
  • spec.md — the protein, the exact claim, the confidence that matches it, and the check that could break it
  • Review both by hand against the transcript and the files — this is the step that separates a pass from a fail
  • Direct with the recipe: add a CONFIDENCE CHECK and a STRUCTURE CHECK, and gate the agent round by round

You don't do the analysis — you direct it and judge what comes back. Same shape as every week.

Pi runs on your machine, talks to the course model through the portal gateway · every call is metered
Get the folding service from the portal → give it to Pi

Fold the owner's own construct on our GPU — one URL

ddls-portal…/week/4
▸ Fold a structure on our GPU · free
Click to reveal your fold API access.
To fold a sequence that isn't in the AlphaFold DB, read the instructions at …/skill.md and follow them. The fold key is in $DDLS_FOLD_KEY (from .env).
the key lives in .env — never in AGENTS.md
⧉ Copy
  • Copy the one instruction and paste it to Pi — it stores the key in .env (gitignored) and notes how to fold in AGENTS.md. You don't edit files.
  • Pi reads the key from $DDLS_FOLD_KEY — the key never goes in AGENTS.md or any committed file
  • Then just tell Pi "fold this and read the pLDDT / PAE" — it does the rest
  • Free · ≤ 400 aa · a fold takes ~1–5 s · open source: github.com/aicell-lab/ddls-structure-api
AlphaFold DB gives the canonical model (a cross-check); our service folds their actual construct/complex
Step 3 · Check the results — don't trust the cartoon

Read the confidence that matches the claim

Is a region a real fold?
Read per-residue pLDDT for exactly those residues — not the global score.
Do two chains bind?
Read the interface PAE. High per-chain confidence is not evidence of binding.
Is it even their protein?
Align the model's sequence to their construct; confirm the assembly (monomer vs complex).

Look at the model (vision is on): a low-confidence region or a shaky interface is obvious once you know to look. A finding that deflates the owner's plan is a win, not a failure.

Gate Pi round by round — never let "the answer" arrive before the confidence that must back it
Step 4 · Build the deliverable

A 3D structure + confidence viewer, and a report

The viewer app (Core)
FastAPI + Tailwind: the headline answer, a 3D model coloured by pLDDT, the confidence panel that matches the claim, a sequence/assembly check line, and the one caveat.
The report
Lead with the answer; state the right confidence next to the claim; name the trap and the honest truth; disclose the AI use. Agent-drafted, then you fix every line.

The viewer is an interrogation instrument — build it so a sceptic (you first, the owner on Friday) can see whether the model supports the claim.

Core first · reach for the Strong tier only once Core runs
This is what an embedded, rotating viewer looks like
demo: crambin (1CRN) — an unrelated example protein

Your seminar deck must embed one too

  • A live, spinning 3D model reads far better than a static screenshot
  • Colour it by pLDDT; highlight the residues the owner cares about
  • One self-contained slides.html — 3Dmol.js from a CDN, the PDB embedded inline
  • Direct Pi: "embed a 3Dmol viewer, colour by B-factor, and call viewer.spin()"
Same idea in your lab viewer app · in the seminar it's the centrepiece of your talk
Version control — new, and required this week

Commit as you go · publish to GitHub · submit the repo URL

  • Tell Pi to init git and add a commit-before-big-changes rule to AGENTS.md — your undo button
  • No GitHub account? Make a free one now — Pi can set up the SSH key and push for you
  • Put the repo URL in the "note to teachers" box on the Hand-in form
.gitignore
.env .venv/ __pycache__/ *.pyc
⚠ Never commit your key. A public repo is public — check .env is not in any commit.
git is what makes bold direction safe — you can only let the agent try things freely if you can undo them
What to hand in · deadline Friday 10:00, before the seminar

Your transcripts, your analysis, your deliverable

  • Interview transcript (Agent A) + analysis transcript(s) (Pi) + the code Pi wrote
  • Your app, zipped to app.zip, and the report (report.md)
  • AGENTS.md + spec.md
  • Your GitHub repo URL — in the note-to-teachers box

Everything through the portal Hand in button. Re-submit any time — we grade your latest. Submit even if you can't attend. A modest, honest result with a transcript that shows real checking is exactly what we're after.

Graded pass/fail · we read the transcript · fill the short feedback panel too
Starting now · your final-project data owner

Find a real owner, dataset and question — this week

  • Your dashboard has a personal invite link — share it in your network (labs, friends, anyone with data + a question)
  • They open it, describe their problem, and submit a project; a link through your invite gives you first claim
  • We approve each project before you start — don't leave it to the last minute
  • Scope: not too easy, not too ambitious — one clear question, answerable in ~a week with the data they have
ddls-portal…/dashboard
Your invite link
…/signup-projects?ref=your-code
⧉ Copy invite   Share
Full guidance: ddls.aicell.io/course/ddls-2026/final-project/
Let's go

A confident structure is
not the same as an answer

Read the confidence that matches the claim, check the model is really their protein, and hand back the honest truth.

Start with the Part 2 setup, then interview · lab page: ddls.aicell.io/course/ddls-2026/module-4/lab/ · good luck
press N for speaker notes