dbluns  /  showcase  /  research · kuala lumpur · 2024

AU-Bio

One model that predicts how living tissue evolves across space and time. Given the state of a tissue now, it produces the whole trajectory ahead in one shot — and it is trained on several biologies at once, and differentiable, so it can be steered as well as read.

Research · v0 IP held · licence by enquiry Under a university technology-transfer proxy Two manuscripts in draft

Patterns forming

textbook, 1984
↔ drag to disturbSeed new spots and watch them divide and settle.

Waves in excitable tissue

textbook, 1991
↔ drag to disturbFire new wavefronts and watch them collide, break and curl.

Two classical systems from the textbooks, running live in this page. Drag your mouse across a dish (or your finger, on a phone) to disturb it, then watch how it evolves from there. They stand for the kind of thing the operator predicts: patterns forming, waves crossing tissue. The operator itself is not on this page.

01 · the bet

Physics first, then living systems.

The foundation models that matter next are not the ones that predict the next word. They take the state of a physical system on a grid and predict how it evolves across space and time — weather, fluids, materials — with neural operators rather than transformers, and one model serving many kinds of physics.

AU-Bio is that bet, made about biology. Cell densities, chemical concentrations, voltages and tissue maps on a grid; morphogenesis, cardiac and neural waves, tumour invasion and chemotaxis as the first four biologies; one operator across all of them.

Read it, then steer it.

Because the whole trajectory comes out at once, errors do not compound frame by frame the way they do in a simulator that steps forward. And because the model is differentiable end to end, the same gradient that trained it can run backwards through it: recover the hidden parameters of an experiment from what was observed, or ask which tissue map, initial condition or intervention would produce the outcome wanted.

Simulate, improve, repeat — on living systems, at the speed of a forward pass.

02 · what has been shown

At laboratory scale, the three claims hold.

Every result below was produced on one workstation, from synthetic test benches first and then from public recordings and real organs, with held-out data, several seeds, and the same budget for every model that was compared. Stated here at the level a licence conversation starts at; the tables are in the manuscripts.

I

One operator learns four biologies at once.

Pattern formation, excitable waves, invasion through a heterogeneous tissue and chemotaxis — four different kinds of equation — share one model, and it predicts each of them on held-out experiments.

II

Training across biologies helps, and harms nothing.

Matched for exposure, the shared model cut the error on invasion dynamics by about a quarter, on every seed, and tied its specialist on the other three. The gain appears from the first checkpoint and no biology shows sustained negative transfer.

III

The gradient steers a real experiment.

The hidden parameters of a held-out trajectory were recovered by gradient descent through the operator, then confirmed by re-running the true simulator with the recovered values, which reproduced what was observed.

IV

It has met real recordings, and real organs.

Optical maps of beating rodent hearts, time-lapse microscopy of wounds closing, and a third-party solver's data in two dimensions; then a brain-tumour cohort, hearts of unseen anatomy and a developing zebrafish in three. A start pretrained across biologies was never worse than starting from nothing at full data, and steadier every time. The head start is small; the steadiness is the reliable part.

△

Where it fails, and why.

Long-horizon prediction of spiral-wave breakup in excitable tissue is the open case. It was diagnosed rather than papered over: the error is set by how far ahead the model is asked to look, not by how densely the data is sampled, and the fixes are known. This is v0 research, and it is barely mature. The page says so because the licence conversation should start from what is true.

1 operator 4 biologies, trained together 3 real organs, in 3-D 3 seeds behind every number 2 manuscripts in draft

03 · what is not on this page

Sealed, by decision.

AU-Bio is held for licensing, and it is early enough that the method is most of the value. So this page carries the claims and never the recipe. What follows is disclosed to institutions under agreement, in a data room, with the authorship and the terms stated first.

In the room, not on the site

  • The operator: its architecture, conditioning, and size
  • The one tensor layout every biology shares
  • The corpus: simulators, parameter ranges, sampling
  • The training and evaluation protocol
  • The benchmark tables, all arms, per seed
  • The public-data and organ pipelines, with provenance
  • The code, the checkpoints, the walkthroughs
  • The review and benchmark manuscripts

How a conversation starts. An institution writes with who it is and what it wants to evaluate. The author replies with a mutual non-disclosure agreement and the licensing frame; the technical dossier follows signature, and the university technology-transfer office that acts as proxy is party to the terms.

admin@dbluns.com · subject AU-Bio licence enquiry

04 · authorship and priority

Who made it, and when.

AU-Bio is the work of S. Mustappa, working as dbluns in Kuala Lumpur, who holds it through a university technology-transfer office as proxy. Two manuscripts are in draft: a systematic review of machine-learned surrogates for spatiotemporal biology, and the cross-biology benchmark that produced the claims above. Their digests are published here so that priority can be shown without the manuscripts being published before their time.

Manuscript fingerprints

SHA-256 of the draft files as they stood on 14 September 2026. The files that produce these digests are held by the author and can be presented on request.

Review paper · draft 0.161b560576041cc989275dc5c4b7a9f0325eb90488f4ab4aaea3be9c90b7ec601
Benchmark paper · draft 0.2248e7c2495a0a0ad5b1b6d10eb4069f9eb7b2a14b9d9eff1d546a28dc73dce9f

Terms of this page

© 2024 S. Mustappa (dbluns). AU-Bio is a working name. The claims on this page are made at the level of the manuscripts’ abstracts and are the author’s. Nothing here is a licence to use, reproduce or build upon the method, which is not disclosed here; the two dishes above are classical textbook systems and are not the model.

Institutions that are already in conversation should refer to the dated dossier they were given rather than to this page.