- Treat simulators as the object of study, not the tool. Mech interp methodology applied to mechanistic simulators is almost completely unexplored, and it's a natural fit. Each simulator (CorticalSim, Cytosim, Tubulaton) makes different assumptions and produces different array behaviors. The mech interp move: don't just compare their outputs to data — intervene inside them. Ablate the collision rule and see what global statistic changes. Patch the dynamic instability module from Tubulaton into Cytosim and see if it rescues a phenotype. Treat each model as a circuit and ask which subcircuits are doing the work. This is exactly activation patching, just on a hand-built mechanistic model instead of a learned one. I don't think the plant MT community has framed it this way and the framing alone would be useful.
- Learned surrogates of the simulators. Cytosim takes hours to run; you can't easily do MCMC over its parameters or run gradient-based optimization. Train a neural surrogate that maps (parameters, initial conditions) → (array statistics). Suddenly you can do Bayesian inference, sensitivity analysis, and design of experiments at the speed of inference rather than simulation. This is a standard move in scientific ML but the plant MT field hasn't really adopted it. If you build the surrogate well, you can then do mech interp on the surrogate — find which input dimensions matter, what internal representations the network learns about the array, whether it discovers the same structure the simulators were built to express. That's a genuinely novel research direction.
- SAEs / probing on simulation trajectories. A long microtubule simulation produces a high-dimensional trajectory. Train a sparse autoencoder on the state at each timestep. What features does it find? Probably things like "local nematic order," "defect density," "regions of active zipping." These are exactly the order parameters the field cares about, but currently they're hand-designed. Letting unsupervised methods discover them and then comparing to the hand-designed ones is the kind of project that would interest both communities.
- Inference from microscopy data. The biggest open methodological problem in the field: given a movie of MTs in a real cell, infer what mechanistic rules generated it. This is a simulation-based inference problem (SBI, ABC, neural posterior estimation). The ML community has tooling for this — sbi library, normalizing flows, etc. — that the plant MT community is largely not using. Bringing modern SBI to bear on the "which model explains this microscopy data" question is a clean, high-value contribution.
- The AI safety / interpretability connection that's actually deep. Self-organization in MT arrays is mathematically the same family of problem as: how do emergent capabilities arise in trained networks? How do mesa-objectives form? You have a system of many local interactions producing global structure, and you want to predict and control the global structure from the local rules. The plant MT case has the enormous advantage that the local rules are known (it's a designed simulator), so it's a sandbox for developing interp methods that you then carry back to neural nets where local rules are emergent. I would actually pitch this to your AI safety colleagues. Cortical MTs are a beautiful "interpretability easy mode" testbed: nontrivial emergent organization, known generating rules, rich data, fast iteration. If your methods can't recover the mechanism here, they probably won't on a transformer either. How to actually engage with this community Concretely, this week (since you're at the workshop): Find the two or three people working on inference or model comparison and have coffee. The names to look for are anyone presenting on parameter fitting, anyone comparing simulators, anyone working with microscopy data quantitatively. Tell them you do mech interp, explain the analogy in one sentence, and ask what their hardest inference problem is. They will tell you, and it will probably be something you have tooling for. Don't lead with "I do AI safety." That phrase means different things to different people and you'll spend the conversation explaining instead of collaborating. Lead with "I work on understanding the internals of trained neural networks, and there's a methodological overlap with what you're doing." Then if it's relevant, mention safety. Offer something specific and small. Not "let's collaborate" — that goes nowhere. "I could spend a week training a surrogate on your Cytosim outputs and seeing if it recovers your order parameters" is concrete and useful. Or "send me 100 of your simulation trajectories and I'll run SAEs on them and tell you what features it finds." Tractable proof-of-concept first, then a real project. The right collaborator is probably a postdoc, not a PI. PIs have full plates and slow timelines. A postdoc who's stuck on an inference problem and would love a methods collaborator is your best bet for actually shipping something. The longer-term pitch If you wanted to make this a real research thread rather than a one-off contribution: there's a paper to be written titled something like "Mechanistic interpretability methods applied to biological self-organization" that takes one cortical MT simulator, treats it as the object of study, applies the full mech interp toolkit (activation patching, probing, SAEs, causal interventions), and shows which methods recover known mechanism and which don't. That paper would land well in both communities and would be a natural bridge piece for someone with your background. It would also be a real contribution to AI safety, not a detour from it — because validating interp methods on systems where ground truth is known is exactly what the field needs and is exactly what biological simulators offer. So: you have more to offer them than you might think, and they have a cleaner testbed than most things you're working with. Worth pushing on.
Last active
May 26, 2026 18:03
-
-
Save bigsnarfdude/9b32446c585f458c4a7f196d5c36cba0 to your computer and use it in GitHub Desktop.
plant MT mechinterp
Author
Author
BIRS workshop 26w5658 · Banff · May 2026 · post-workshop sprint
The MT Benchmark Nobody Built
Three simulation engines, thirty years of work, zero cross-validation — until now
- Three groups built three simulators for plant microtubule dynamics. Nobody had ever run them against the same parameters and checked if they agree.
- A one-day sprint compared all three on isolated MT dynamics. They agree. The DI physics layer is consistent across all three codebases.
- The no-interaction condition revealed that the collision rule isn't doing alignment — it's doing population control. Remove it and the array explodes.
- All three engines are built on the 2-state DL model. STADIA 2022 showed a third state — the stutter — precedes 80% of catastrophes. Nobody has it.
- Cancer drug discovery targets MT dynamics. The models drugs are designed against are incomplete. The heavy lifting won't be done by individual labs.
The Setup
The BIRS workshop on plant microtubule modelling brought together the three groups who built the main simulation engines. The frustration in the room: thirty years of work, increasingly detailed biology, and nobody had ever just run the engines against the same parameters and compared outputs. No shared benchmark. No canonical test. No MNIST for microtubule simulators.
Cytosim
François Nédélec · C++
Continuous space, explicit mechanics, forces between filaments. Knows where every MT is physically at every moment. ~24 hrs per run.
CorticalSim
Eva Deinum · C++
Event-driven cortical array. Jumps to next state change. Collision outcomes as decision rules. Seconds per run.
elastic-mt
Tim Tian · Python
Event-driven, BY-2 parameters, PNAS 2025. Same decision-rule architecture. Seconds per run.
Three levels of abstraction. François built the physics. Eva and Tim built rules that approximate the physics. Each one is a bet that the level below it matters less than it looks like.
What We Ran
Tier-1: 200 isolated MTs, Dogterom-Leibler 2-state model, BY-2 plant cell parameters (v+=0.08 µm/s, v-=0.16 µm/s, kcat=0.003/s, kres=0.005/s), 1800s. No interactions. Do the engines agree on basic DI statistics?
Tier-2: Full cortical array on a 125×20 µm cylinder, same DI parameters, three collision conditions: 40° zipper rule, no interaction, always catastrophe. Where do the engines diverge?
What We Found
CONFIRMED Tier-1 passes — all three engines agree on DI physics
Every pairwise KS test passes at p>0.05. Growth and shrink velocities match spec to 3 sig figs across all three codebases. Any future disagreement at tier-2 is collision handling, not the growth/shrink model.
FINDING The collision rule controls population, not just alignment
No-interaction under BY-2 parameters: population grew from 2,800 MTs at t=180s to 14,000 at t=1800s with no sign of stabilizing. BY-2 has rescue rate > catastrophe rate — without collision-induced catastrophe the array has no equilibrium. The zipper rule's primary job isn't alignment, it's keeping the array alive. Remove it and you don't get a disordered array. You get a population explosion.
GAP All three engines are missing a state that precedes 80% of catastrophes
The Dogterom-Leibler model has two states: grow and shrink. STADIA 2022 (Goodson lab) showed that a third state — the stutter, a period of slowed or halted dynamics — precedes 78% of catastrophes in simulation and 86% in experiment. None of the three engines implement it. The benchmark is testing an incomplete model.
Why It Matters Beyond Plant Biology
Microtubule dynamic instability is a primary target for cancer chemotherapy. Taxol, vincristine, and the broader class of MT-targeting agents work by disrupting the grow/shrink cycle. The models these drugs are designed and tested against are built on the same 2-state DL framework.
If the stutter state is mechanistically involved in catastrophe — and STADIA says it is — then the model cancer drug discovery is running on is missing the step it's trying to target.The simulators aren't wrong. They're incomplete in a specific, now-measurable way.
Individual labs won't fix this. Each one is optimized for their own questions. The benchmark work — agreeing on parameters, running cross-validation, establishing what "passing" looks like, updating the canonical model when new states are discovered — is coordination work. It needs to live outside any single lab.
The Abstraction Stack
François built physics from observations through a microscope. Eva and Tim built decision rules from the same observations. The mech-interp work on Tim's simulator built a model of the rules to understand what they're actually doing (answer: density control, not alignment).
The next step isn't building a better rule. It's removing the handcrafted translation entirely — generate enough trajectories from the event-driven sims (seconds each, 100k overnight) and train something that learns the structure directly. The simulator becomes training data, not the artifact. The stutter state becomes a feature the model discovers, not a state someone has to add by hand.
That's the gap between 1995 and now. The microscopes got better. The simulators didn't change architecture.
What's Still Running
Cytosim tier-1 finishes tonight. Cytosim tier-2 (bundling, crossing, nonsteric) is running on nigel — days away. That's the comparison worth waiting for: does the force-based physics agree with the decision-rule engines on cortical array organization? If yes, the rules are a good approximation. If no, the forces are capturing something the rules miss.
Tier-1 data:
nigel:/home/vincent/bench/ · STADIA: Vemu et al. 2022, PMC9250389 · BIRS 26w5658 · 2026-05-23
Author
MT Simulator Benchmark — Tier 1 & Tier 2 Comparison Report
Date: 2026-05-26 | Machine: nigel (RTX 4070 Ti) | Seed: 26058
Engines
┌─────────────┬──────────────────┬──────────┬───────────────────────────────────┐
│ Engine │ Author │ Language │ Type │
├─────────────┼──────────────────┼──────────┼───────────────────────────────────┤
│ Cytosim │ François Nédélec │ C++ │ Physics, continuous forces │
├─────────────┼──────────────────┼──────────┼───────────────────────────────────┤
│ elastic-mt │ Tim │ Python │ Event-driven, zippering │
├─────────────┼──────────────────┼──────────┼───────────────────────────────────┤
│ corticalsim │ Eva │ C++ │ Cortical array, cylinder geometry │
└─────────────┴──────────────────┴──────────┴───────────────────────────────────┘
Tier 1 — Isolated Dynamic Instability
Single MTs, no collisions. Tests whether engines agree on basic DI physics (grow/shrink/catastrophe/rescue) using BY-2 parameters.
Engines compared: Cytosim vs Tim DI wrapper (Tim's full elastic-mt can't run isolated MTs without refactoring — collision detection is baked into the main loop, so a standalone wrapper reimplementing his exact DI parameters was used.)
Results (200 MTs, seed 26058, t=1518s)
┌───────────────────────────┬─────────┬────────┬────────┬──────────┐
│ Statistic │ Cytosim │ Tim DI │ Spec │ Verdict │
├───────────────────────────┼─────────┼────────┼────────┼──────────┤
│ Growth velocity (µm/s) │ 0.080 │ 0.080 │ 0.080 │ ✅ │
├───────────────────────────┼─────────┼────────┼────────┼──────────┤
│ Shrink velocity (µm/s) │ 0.159 │ 0.159 │ 0.160 │ ✅ │
├───────────────────────────┼─────────┼────────┼────────┼──────────┤
│ Mean cats/MT │ 1.7 │ 1.7 │ — │ ✅ │
├───────────────────────────┼─────────┼────────┼────────┼──────────┤
│ Mean final length (µm) │ 64.9 │ 59.6 │ — │ ✅ │
├───────────────────────────┼─────────┼────────┼────────┼──────────┤
│ Max final length (µm) │ 122 │ 121 │ — │ ✅ │
├───────────────────────────┼─────────┼────────┼────────┼──────────┤
│ Alive MTs at t=T │ 52/200 │ 63/200 │ — │
├───────────────────────────┼─────────┼────────┼────────┼──────────┤
│ KS p-value (length dist.) │ — │ — │ p=0.37 │ ✅ PASS │
└───────────────────────────┴─────────┴────────┴────────┴──────────┘
Verdict: PASS. Both engines agree on all four canonical DI statistics at p > 0.05.
One known artifact: Catastrophe frequency reads ~5× the input rate (0.017/s vs spec 0.003/s). This is a 1 Hz sampling artifact — fast sub-second catastrophe+rescue cycles are invisible to the logger. Both engines are equally affected so they
still agree with each other. Fix: sub-second sampling or event-log counting.
Alive MT count (52 vs 63): Within expected stochastic variance for n=1 seed. Not a bug.
Tier 2 — Cortical Array with Collisions
MTs nucleated on a surface, MT-MT interactions enabled. Tests whether engines agree on how collisions shape the steady-state array.
Engines compared: Cytosim (3 variants) × corticalsim/Eva (3 variants).
Variants
┌──────────────────┬─────────┬───────────────────────────────────────────┐
│ Variant │ Engine │ Collision physics │
├──────────────────┼─────────┼───────────────────────────────────────────┤
│ cyto:nonsteric │ Cytosim │ No MT-MT interaction (reference) │
├──────────────────┼─────────┼───────────────────────────────────────────┤
│ cyto:crossing │ Cytosim │ Steric repulsion only │
├──────────────────┼─────────┼───────────────────────────────────────────┤
│ cyto:bundling │ Cytosim │ Steric attraction + repulsion │
├──────────────────┼─────────┼───────────────────────────────────────────┤
│ csim:nozipper │ Eva │ No zipper, MTs cross freely │
├──────────────────┼─────────┼───────────────────────────────────────────┤
│ csim:baseline │ Eva │ 40° zipper rule + 50% induced catastrophe │
├──────────────────┼─────────┼───────────────────────────────────────────┤
│ csim:alwayscatas │ Eva │ Always catastrophe on collision │
└──────────────────┴─────────┴───────────────────────────────────────────┘
▎ Geometry note: Not directly comparable in absolute numbers — Cytosim uses a 50×50 µm flat periodic patch (2,500 µm²), Eva's sim uses a cylinder (18,221 µm², ~7.3× larger) with ~36× higher nucleation flux. Qualitative trends are what matter
▎ here.
Steady-state results (last 50% of simulation time, t=900–1800s)
┌──────────────────┬──────────────────┬──────────────────┬───────────┐
│ Variant │ Density (µm/µm²) │ Mean length (µm) │ MTs alive │
├──────────────────┼──────────────────┼──────────────────┼───────────┤
│ cyto:nonsteric │ 3.83 │ 29.8 │ 317 │
├──────────────────┼──────────────────┼──────────────────┼───────────┤
│ cyto:crossing │ 2.51 │ 20.8 │ 301 │
├──────────────────┼──────────────────┼──────────────────┼───────────┤
│ cyto:bundling │ 3.44 │ 26.1 │ 327 │
├──────────────────┼──────────────────┼──────────────────┼───────────┤
│ csim:nozipper │ 20.80 │ 30.3 │ 12,383 │
├──────────────────┼──────────────────┼──────────────────┼───────────┤
│ csim:baseline │ 0.86 │ 7.1 │ 2,203 │
├──────────────────┼──────────────────┼──────────────────┼───────────┤
│ csim:alwayscatas │ 0.54 │ 5.7 │ 1,718 │
└──────────────────┴──────────────────┴──────────────────┴───────────┘
Order parameter S2 (Eva's sim only — cylinder geometry required)
┌──────────────────┬───────┬──────────────────────────────────────────────────────────────────┐
│ Variant │ S2 │ Interpretation │
├──────────────────┼───────┼──────────────────────────────────────────────────────────────────┤
│ csim:nozipper │ 0.004 │ Isotropic, no preferred direction │
├──────────────────┼───────┼──────────────────────────────────────────────────────────────────┤
│ csim:baseline │ 0.101 │ Modest alignment from zipper rule │
├──────────────────┼───────┼──────────────────────────────────────────────────────────────────┤
│ csim:alwayscatas │ 0.159 │ Strongest alignment — fast turnover lets stable bundles dominate │
└──────────────────┴───────┴──────────────────────────────────────────────────────────────────┘
Key Finding
Both engines agree qualitatively, disagree quantitatively by ~24×.
Within each engine, adding collision effects reduces density and mean length. The ordering is consistent. But the magnitude of the effect is completely different:
┌───────────┬─────────────────────────────────┬───────────────────┐
│ Engine │ No collisions → With collisions │ Density reduction │
├───────────┼─────────────────────────────────┼───────────────────┤
│ Cytosim │ 3.83 → 2.51 µm/µm² │ ~35% │
├───────────┼─────────────────────────────────┼───────────────────┤
│ Eva's sim │ 20.8 → 0.86 µm/µm² │ ~24× │
└───────────┴─────────────────────────────────┴───────────────────┘
Root cause: The engines implement different biology.
- Cytosim collision = mechanical force only (MTs push/pull, no state change)
- Eva's sim collision = 50% chance of catastrophe (MT dies on contact)
Collision-induced catastrophe is the dominant mechanism. Mechanical repulsion alone barely affects steady-state density. Any simulator without it cannot reproduce in-vivo BY-2 MT densities.
Cytosim oddity — bundling > crossing: Adding attraction (bundling) gives higher density than repulsion alone (crossing). Mechanism: co-alignment from attraction reduces crossing angles, so fewer effective collisions occur. This is actually
consistent with the biology — zippering reduces interference, it doesn't amplify it.
Summary
┌────────┬────────────────────────────────────────┬─────────────────────────────────────────────────────────────────────────────────────────┐
│ Tier │ Question │ Answer │
├────────┼────────────────────────────────────────┼─────────────────────────────────────────────────────────────────────────────────────────┤
│ Tier 1 │ Do engines agree on isolated DI? │ Yes — engines are interchangeable at the single-MT level │
├────────┼────────────────────────────────────────┼─────────────────────────────────────────────────────────────────────────────────────────┤
│ Tier 2 │ Do engines agree on collision effects? │ Qualitatively yes, quantitatively no — 24× gap driven by catastrophe-on-contact physics │
└────────┴────────────────────────────────────────┴─────────────────────────────────────────────────────────────────────────────────────────┘
The benchmark is working as designed. Tier 1 establishes the floor (DI physics is shared). Tier 2 reveals that the dominant biological question is what happens when two MTs meet — and the two engines encode different answers to that question.
Next Steps (v0.3)
- Add collision-induced catastrophe to Cytosim — implement a custom Hand that triggers catastrophe on fiber contact; would close the engine gap and confirm the mechanism
- Normalize geometry/nucleation — rerun with matched area and nucleation rate so absolute numbers are directly comparable
- Run Tim's elastic-mt in tier-2 — adds a third engine with its own zippering model (already on nigel at ~/elastic-mt/)
- Multi-seed (n=5) — Poisson CIs on density and MT count for all variants
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Plant MT Simulation Benchmark — What We Actually Agreed On
BIRS Workshop 26w5658 follow-up
Status: Draft for sign-off
Owner of this doc: [name]
Deadline for sign-off: 4 weeks from workshop close
Why This Document Exists
The workshop ended without a shared artifact. Two years of planning produced great conversations and no concrete deliverable beyond a planned special issue with no agreed content.
This document fixes that. It proposes the minimum thing the field needs to stop talking past itself: a shared benchmark that any simulator must pass before its results can be compared to another simulator's results.
This is not a research paper. This is a checkerboard. The point is not to be novel. The point is to be agreed upon.
If you sign this, you commit to:
What We Agree On (Sign at the Bottom)
Tier 1 — Isolated Microtubule Dynamic Instability
Setup: 200 non-interacting MTs in free space. BY-2 parameters (growth velocity 0.08 μm/s, shrink velocity 0.16 μm/s, catastrophe rate 0.003/s, rescue rate to be specified). 1800 second simulation. 5 seeds minimum.
Reported statistics:
Pass criterion: All pairwise KS tests p > 0.05 on the four classical statistics.
Tier 1.5 — Stutters (added based on Mahserejian et al. 2022): Report whether your simulator produces stutter behavior. If yes, report frequency and duration distributions. If no, state explicitly that your model is two-state and does not capture pre-catastrophe slowdown. This is not a pass/fail criterion yet — it is a disclosure requirement. We need to know which simulators can even ask the stutter question before we decide whether stutter agreement matters.
Tier 2 — Collective Behavior with Collision Rules
Setup: Cylindrical cell, 125 × 20 μm (BY-2-like). Same DI parameters as Tier 1. Three collision conditions:
zipper_40always_catasno_interaction10 seeds per condition. 3 hours simulation time minimum.
Reported statistics (all three, jointly):
Pass criterion: Each simulator reports all five metrics for all three conditions. There is no pass/fail across simulators yet — Tier 2 is the diagnostic. The point is to see where engines diverge and why.
Tier 3 — Orientation-Dependent Catastrophe (ε Condition)
Setup: Same cylinder as Tier 2. Default zippering. Add orientation-dependent catastrophe rate per Tian et al. 2025:
r_c(φ) = r_0 (1 + ε sin φ).Sweep ε ∈ {0, 0.2, 0.4, 0.8}. 10 seeds each.
Reported statistics: Same as Tier 2, plus:
Pass criterion: Same as Tier 2 — disclosure, not validation. This tier exists because the ε mechanism is where engines are expected to disagree most.
Tier 4 — Branched Nucleation Control
Setup: Tier 2 baseline + 2×2 factorial on (collision rule × nucleation mode):
no_interaction10 seeds per cell. The gap (A − D) is the total ordering contribution. The gap (B − D) isolates collision-rule contribution. The gap (A − B) isolates nucleation contribution.
Why this matters: Workshop finding — S₂ as currently reported is a mixture signal of branched nucleation + collision-induced alignment. Without controlling for nucleation mode, "my simulator agrees with experiment on S₂" is not a meaningful claim.
What We Explicitly Do Not Agree On Yet
These are listed here so they don't get smuggled in later as if they were settled:
Each of these gets its own benchmark when there's enough infrastructure to test it. Not before.
Deliverables
Per simulator, due 8 weeks from sign-off:
git clone && run.sh.These four artifacts together become the QPB special issue contribution. The special issue is not a place for papers about your favorite simulator. It is the proceedings of this benchmark.
Process Rules
Tier Owners (Fill In)
Signatures
By signing, you commit to running this benchmark on your simulator within 8 weeks and publishing the results in the QPB special issue.
Appendix A — What Already Exists
The following work was done at the workshop and around it. Some of it is partial; all of it is a starting point:
/home/vincent/bench/.mt_phase1/. Reproducible.collision/decision.py+collision/geometry.py+collision/api.py. 81 tests pass. Backward-compatible. Branchrefactor/split-zip-catonbigsnarfdude/elastic-mt-refactor.accept_MT=0.24contributes substantially to S₂ even when collision-based alignment is removed. See ablation results.This is not "Vincent's work that you can ignore." This is the existing tier-1 calibration data. If your simulator's tier-1 numbers don't agree with this, that is the conversation to have.
Appendix B — Why We Need This (One Paragraph)
The workshop description called for a "Stanford Bunny" for plant MT simulation. We didn't produce one. This document is the minimum viable Stanford Bunny: an agreed isolated-MT test, an agreed collective-behavior test, and an honest disclosure of what each simulator can and cannot do. Without this, every cross-paper comparison in the field is unmoored. With this, we can argue about science instead of about whose code is right.