View on GitHub
Auckland site · Aotearoa New Zealand
● Mon 19 – Tue 20 October 2026 · Free

How far can a chain of surrogates take us in biomolecular interactions?

Systems, proteins, ligand binding, drug discovery, pharmacology. Biology has a ladder of scales, and machine learning now has one running alongside it. Two days at the University of Auckland finding out how far up you have to climb before the answer can be trusted.

Register — free The themes The two days

Two steps, both free: an RSVP for Auckland, and the global event's registration.

When
Mon 19 Oct, 10am – Tue 20 Oct, 4pm
Where
Digital Research Innovation & AI Lab · Level 10, 70 Symonds Street
Cost
Free
Bring
A laptop

The question

Physiological outcomes come from molecules interacting. Those interactions produce functional change. Function is driven by spatial and structural change. Pharmacology is the effect of drugs on that system.

Association models over sequence, 3D-aware models over structure, interatomic potentials over atoms and motion, multi-scale models over cells and tissue. Most are open-weight and most arrived recently. Most of us know our own rung.

The global event

OSI Hackathon

Open Scientific Intelligence Hackathon
for the Physical Sciences & Mathematics

This is the thing that got us to run a site at all.

Now in its fourth year. It began as the LLM Hackathon for Materials Science and Chemistry and has widened to the physical sciences and mathematics — agents, language models, datasets, benchmarks, models and scientific software. Free, hybrid, and run from sites around the world on the same two days.

26Oct Osi OSI Hackathon October 21–22, 2026
120projects submitted in 2025
16on-site locations, plus virtual
4community papers, every team credited
Freeto take part, at every site

Auckland runs Mon 19 – Tue 20 October; the global days are 21–22 October. Their registration is the one that makes a project visible to the judges, and Wednesday the 21st is open if a group wants to carry the work across. Every edition is written up with every contributor named, which for two days' work is a real line on a CV.

Register with the global event llmhackathon.github.io Join their Slack

Built at past editions, close to our themes

All open source, all from the 2025 event. Worth half an hour before October: it's the fastest way to see what two days actually produces, and several of these are live starting points rather than sketches.

Near theme 1 · Screening at scale

  • MIDAS — an agentic interface for structure-based drug design, steering DiffSBDD inside a pocket and analysing the poses it gets back.Glasgow
  • SmeLLMap — ESM-2 embeddings over receptors with voxelised binding cavities: a different answer to how you represent a pocket.Duke

Near theme 2 · Molecules in motion

  • UMADock — docking with an MLIP as the scoring function, with desolvation and ligand-strain terms.CafChem
  • DynaMate — autonomous protein–ligand MD over GROMACS and AMBER with MM/PB(GB)SA, and quality checks that retry when a step fails. Started at the hackathon as DynaAgent and still being developed.EPFL LIAC · preprint
  • LARA-HPC — an agent that submits electronic-structure workflows to a cluster, with a rehearsal mode that validates before anything real runs.CEA, CNR, INRIA, RIKEN and others
  • DFTPilot — retrieval plus crystal GNNs to set up and preview a DFT calculation before you pay for it.DTU, UCL, Cambridge, NTU

Near themes 3 and 4

  • Thin, which is the interesting part. Nothing in 2025 carried a molecular signal up to cells or physiology, and nothing took on drug repurposing directly.
  • AssemblAI — designing peptide self-assembly protocols, the nearest thing to working up a scale.Harms Informatics, MIT
  • ARIA — causal knowledge graphs for inverse design: the machinery theme 4 would use, pointed somewhere else.Johns Hopkins

Useful whatever you work on

  • ACME — literature to structured data, end to end. Fifth overall in 2025, and the quickest route to a dataset when you don't have one.Argonne
  • AtomBridge — papers to validated structures, with physical sanity checks on the way out.UIUC, Northwestern, Chicago, Georgia Tech
  • ATOMS Lab — property prediction on 50–300 examples, with the support set ablated properly. Small-data method, done carefully.UMBC
  • MCP4SDL — MCP interfaces to lab instruments.UBC, Tennessee, ETH Zürich, Empa

Themes

Four to start from. They describe the kind of work we're interested in, and they come from where we happen to sit. Bring your problem and take a space. If none of them fit, add one — we're making room for people to gather, and the shape of the room is still open.

Theme 1, screening at scale: models that know something about shape, used to filter before you pay for docking. Boltz-2, OpenFold3, Chai-1r, IntFold Theme 2, molecules in motion: machine-learned interatomic potentials in MD. MACE, UMA, eSEN, Orb-v3, NequIP Theme 3, binding to whole system: carrying a molecular signal up to functional and physiological change. STATE, scGPT, Geneformer, scFoundation Theme 4, repurposing what we have: genotype, structure and approved drugs with deliberately small models. TxGNN, PrimeKG, lightweight KG embeddings An open theme. Bring a problem that doesn't sit in the four, and open an issue. An open theme. A dataset, an evaluation or a comparison you think ought to exist.

And the wider scope

The global event takes submissions across the physical sciences and mathematics, in any of these:

Autonomous agentsLanguage modelsDatasets BenchmarksModelsScientific software

Anything in that scope works here. A benchmark nobody has built, a dataset that should exist, a piece of software your field keeps rewriting, an agent that does a job you currently do by hand — all of it counts, and none of it has to be about ligands.

The two dotted tiles are real. Propose a theme and we'll make space for it. We also have default targets and chemistry ready if you'd rather pick something up.

What a group does

01

Find out what open-weight models exist in your area. The landscape moved a long way in the last eighteen months.

02

Get one or two running on our hardware. Being able to run it is the main filter.

03

Work out what to measure, and report it back at the end of day one.

That third one is the move we care most about. A recent benchmark found single-cell foundation models level with a simple linear baseline at predicting perturbation effects, and another found the choice of metric changes which model comes out on top. How to evaluate these things is genuinely unsettled, and there's a lot of room to do useful work there.

Each group sources its own data and picks its own benchmark. Groups stay comparable by answering the same four questions about whatever they picked:

  1. What's your ground truth?
  2. What are you comparing against, and is there a simple baseline in there?
  3. What would change your mind?
  4. Where would it break?

Who it's for

Come and explore the broader opportunity while getting hands-on with AI in your own area.

You don't need to be able to code

Come prepared to work with AI agents and with code. The agent writes the code. You bring the judgement about whether the answer is any good. Everyone is welcome here, whatever you last wrote and whenever you last wrote it.

Chemists and structural biologists
Shape, pockets, poses
Physicists and computational chemists
Potentials, forces, motion
Systems biologists and pharmacologists
From a binding event to a phenotype
Statisticians and ML researchers
Evaluation, baselines, what the numbers support
Bioinformaticians and clinicians
Variants, approved drugs, repurposing
Research software engineers
Making any of it run, twice

We're inviting project teams. A group leader brings their team, their expertise, and a problem or candidate. If you can only come for the opening, come for the opening and leave your team to it. Coming on your own works too.

The two days

Day 1 — get running, do something, work out what to measure, report back. You can pick a benchmark once you've seen a model behave, so that comes after the running.

Day 2 — measure it, and report what you got, including the negatives.

Bookends each day, lightning talks each day, work in ones, twos and threes, and food and a social occasion both days. Setup is agentic support plus our team, and at least one model per theme will be running on our hardware before you arrive.

No stupid questions. It's safe to experiment. A negative is a result.

Registering

Two steps, and both are free. The first one gets you into the room in Auckland. The second is the global event's, and it is what makes a project visible to their judges, so a group that means to submit needs both.

01 · AUCKLAND

RSVP for the Auckland site. Name, email, what you work on, what you want out of the two days, and which days you're coming.

Auckland RSVP
02 · GLOBAL

Register with the global event. One free sign-up, and the one the judges work from.

Global registration

You can come for one day or both, and you can join remotely instead if getting to Symonds Street doesn't work. The form asks which, so say so there.

Submitting

Judging is the global event's, and so is the submission. Teams submit through the main site, which is where the judges see it. There is nothing to hand in to us.

What that has meant, and what to plan for: a video of two minutes or less, posted somewhere public, and one form for the team. The judges weigh potential for impact, how innovative the approach is, whether it would scale, and how it lands in its field. Past years also show that the write-up is what gets a project into the community paper — 32 of the 120 submissions in 2025 were left out of it for incomplete documentation.

The 2026 deadline has not been announced. It falls at the end of the global Day 2, once our two days are over, and the organisers say it goes to Slack before it goes anywhere else. Their submission page still carries last year's date, so take the deadline from Slack or from their agenda.

Running on the 19th and 20th leaves that window open: our two days finish before the global days begin, so a group has time to cut the video and write the thing up.

Questions

Open an issue in the repo, or ask in the hackathon Slack — the invitation comes with registration.

github.com/drai-inn/osi-hackathon-auckland

There's an A3 poster too, for putting up somewhere your group will walk past it.