Hodgkin's Razor

A mechanistic digital twin for neural organ-on-chip recordings: from a baseline and a treated recording, the probability that each of ten molecular mechanisms moved, read against untreated cultures, or a refusal.

AI4S Open Innovation: AI for Life Science, 5th Pazhou Algorithm Competition. Category: End-to-End System. Apache-2.0.

Live demo (no login, no install): https://marc-dvci.github.io/Hodgkin-s-Razor/ · Technical report: docs/TECHNICAL_REPORT.pdf · 中文说明: README.zh.md

A microelectrode array under a neural organ-on-chip produces a rich recording and a thin answer: firing fell, bursts shortened, synchrony dropped. It does not say whether the compound blocked AMPA receptors, opened chloride channels or slowed vesicle recycling, and those have different consequences for a drug programme.

Hodgkin's Razor fits a biophysical network model to the recording itself. It reads a baseline and a treated recording (one well twice, or two sister cultures) and returns, for ten mechanisms, the probability that each moved, the size of the shift if it did, and an interval. When no simulation resembles the recording, or the fitted twin cannot reproduce it, it says so and names nothing. The razor is the sparse prior: among the mechanisms that could explain a change, prefer the fewest.

Results

Every result is scored against untreated recordings read the same way: pre-drug stretches of the same well, sister cultures neither of which was treated, and vehicle wells. A method's chance level is its own hit rate when nothing was applied.

Test (pre-registered, blind)	What was asked	Result
Chronic NMDA blockade, mouse hippocampal sister cultures (Charlesworth et al. 2015), version 3	does p(NMDA) separate 29 treated from 23 untreated preparations of the same genotypes? bar AUROC 0.70	0.86 [0.74, 0.96]: met. Doorn et al. estimator 0.49; same twin unpaired 0.66
same	is NMDA the single top-ranked mechanism?	2/29 against 1/23: not met
same, secondary	does the reading fade as the cultures compensate, as the authors report?	yes, early > late, p = 3e-5
Dynasore, human iPSC networks (Doorn et al. 2024), version 2	named mechanism (u_rel or tau_d) in at least 5 of 10 wells	2/10: not met
same, secondaries	detection; mechanism class	drug called in 9/10 wells and 0/10 untreated pairs; class 8/10 (p = 0.0016)
Recorded microchannel chips (Mateus et al. 2024), version 2	channel statistic separates diodes (AUROC >= 0.75)	0.62 [0.28, 0.88]: not met, and underpowered (power 0.48); the twin says 20 chips per design are needed
Recorded four-compartment chip (Lassers et al. 2023, hippocampal EC, DG, CA3, CA1), version 4	do an axon's spikes follow the compartment it grows from? (twin: at least 0.73 of axon pools)	0.82, 49/60: met; 0.79 and 0.86 after two stimulation patterns
same	does compartment timing barely track the direction of axonal traffic? (twin: Spearman in [-0.16, 0.41])	-0.30: not met, opposite sign; inside the interval after stimulation

The failures are reported in full in results/v3/RESULTS.md and results/v2/RESULTS.md, and discussed in section 9 of the report.

---

Quick start

pip install -r requirements.txt          # CPU is enough
python demo.py                           # the pre-registered evidence, well by well
python demo.py --serve                   # web application on 127.0.0.1:8000
python -m pytest tests/ -q
python scripts/verify.py                 # hashes, checksums, frozen models, written numbers

A static version of the application, with no server, is in site/ and is published by .github/workflows/pages.yml. It has three views: Analyse a recording pair (the mechanism report on bundled recordings), Blind test (every preparation of the pre-registered test, with method, untreated group and mechanism switchable; each AUROC is recomputed in the browser from the scored values, written by scripts/export_evidence.py) and Chip planner (which readout resolves which chip property, and power against chips per design). A CPU notebook is in notebooks/hodgkins_razor_demo.ipynb.

With a CUDA device, the whole pipeline rebuilds from public data:

pip install -r requirements-gpu.txt
python scripts/fetch_tampere.py          # Tampere recordings, CC BY 4.0, checksummed
python scripts/fetch_external.py         # Doorn, Mateus and Charlesworth et al., checksummed
python scripts/run_all.py

docs/REPRODUCTION.md lists every stage with its measured time.

---

How it works

 baseline + treated recording of one well (Axion, MCS or a two-column CSV)
              |
   recording system: electrode layout and detection dead time,
   applied to the recording and to every simulation alike
              |
   40 summary statistics        one function for recorded and simulated events
              |
   5 presence heads             probability that each of 10 mechanisms moved
   conditional flow             culture parameters; size of each shift if it moved
              |
   guard: typicality + predictive check   (either firing -> "outside the model")
              |
   mechanism report             called mechanism, direction, effect, interval,
                                class, rivals, culture parameters, hash;
                                next experiment when two mechanisms tie

The simulator. 256 conductance-based neurons on a 16 x 16 grid: excitatory and inhibitory populations, AMPA, NMDA and GABA-A synapses, a tonic GABA-A conductance for bath agonists, short-term depression, asynchronous release (after Doorn et al.), slow after-hyperpolarisation and distance-dependent delays. One CUDA block per network, the whole time loop in shared memory, and bit-reproducible: spikes are collected with a warp ballot and read back in neuron order, so floating-point sums never reorder.

Recording systems. The kernel records at 0.2 ms. Each system is a view with its own electrode layout and dead time: the Axion 48-well plate (16 electrodes, 2 ms), the MCS 24-well plate (12 electrodes, 0.3 ms) and the MCS 60-electrode array read as four quadrants (1.08 ms).

Two designs. One well recorded twice holds wiring and pickup fixed. Two sister cultures share a preparation but not a network, so the bank draws the second sister's wiring afresh and adds a drift between sisters, measured on sister pairs that were never treated.

The domain. A twin is trained on the cultures its own recording system produces: the range the system's untreated baselines span, a living network-driven culture, and activity that collapses when AMPA is blocked. scripts/fit_domain.py finds where the simulator produces that domain, from baselines only.

The prior. A compound moves one to three mechanisms. Directions are drawn where the baseline leaves room, and a shift the bounds would clip is labelled inactive.

Inference. Five presence heads, averaged and calibrated on held-out cultures, and a masked autoregressive flow conditioned on the recording and on the active set, so each mechanism's effect is reported given that it moved.

The guard. Typicality: distance to the nearest training simulations, which needs no GPU. Predictive check: re-simulate 48 posterior draws and count the statistics outside their band. Thresholds at the 97.5th percentile of held-out simulations.

The chip twin. The same kernel simulates two chambers joined by directional microchannels and reads them four ways: chamber electrodes, 2 Hz calcium, channel electrodes and a chamber perfusion. For each chip property it says which readout resolves it, before the experiment is run.

---

Validation

Four pre-registrations, each hashed before its data were read.

• PREREGISTRATION.md (version 1) fixed the Tampere answer key. The pre-registered model scored at chance, and after three label-free corrections the plates had been scored four times. They are the development set now. The record is in results/v1/ and the git tag v1.

• PREREGISTRATION_v2.md was written after version 2 was frozen on simulation evidence and before any version 2 output existed on two blind tests: Dynasore on human iPSC networks from another laboratory (Doorn et al. 2024), and the chip twin's readout prediction on recorded microchannel chips (Mateus et al. 2024). hodgkins_razor/doorn.py refuses to return a treated Dynasore window until that file exists and matches its hash, and scripts/evaluate_v2.py checks it and every frozen model's digest before it runs.

• PREREGISTRATION_v3.md was hashed after the version 3 twin passed a stop rule committed before it was trained (docs/STOP_RULE_v3.md; the first twin failed it and both attempts are listed). Its null group is matched on genotype. hodgkins_razor/charlesworth.py refuses to return a treated sister until that file matches its hash.

• PREREGISTRATION_v4.md holds the chip twin's predictions for a recorded four-compartment chip (Lassers et al. 2023), simulated by scripts/brewer_prediction.py, then hashed and pushed before a spike was counted. hodgkins_razor/brewer.py refuses to return a spike time until that file matches its hash.

Every outcome, including what failed, is in results/v4/RESULTS.md, results/v3/RESULTS.md and results/v2/RESULTS.md.

---

Data

Dataset	Use	Licence
Tampere comparative MEA dataset	development set; domain of the Axion twin	CC BY 4.0
Charlesworth et al. 2015, sister-array recordings	blind test of version 3; domain and drift (untreated sisters only)	CC0 1.0
Doorn et al. 2024, Dynasore peak trains	blind test; domain of the MCS twin (baselines only)	Apache-2.0
Mateus et al. 2024, microchannel chips	blind test of the chip readout prediction	CC BY-NC-ND, read in place, not redistributed
Lassers et al. 2023, four-compartment hippocampal chips	test of the chip twin, version 4	CC0 1.0
Doorn et al. 2025, SBI estimator	prior art, scored beside the twin	Apache-2.0

No restricted, clinical or personal data. docs/DATA.md has the details.

---

Repository layout

hodgkins_razor/
  params.py      the parameter table, read by everything
  kernel.cu      the CUDA network kernel
  simulator.py   batched simulator, paired mode, recording-system views
  features.py    the 40 statistics, one path for simulation and experiment
  shift.py       the sparse prior over what a compound does
  nde.py         presence ensemble, conditional flow, unpaired baseline
  ppc.py         predictive check and typicality
  report.py      the mechanism report
  design.py      the next experiment, when two mechanisms tie
  chip.py        two-compartment chip twin and its four readouts
  tampere.py doorn.py mateus.py charlesworth.py brewer.py   readers for the five datasets
  examples_key.py  the known target of each bundled example
scripts/
  fit_domain.py make_bank.py train.py sim_eval.py guard_sim_check.py
  calibrate_drift.py pharmacology_check.py gaba_check.py chip_study.py chip_power.py
  lassus_study.py design_study.py null_controls_v2.py
  freeze_v2.py evaluate_v2.py mateus_check.py prior_art_doorn.py
  freeze_v3.py evaluate_v3.py evaluate_v3_lowmem.py
  brewer_prediction.py freeze_v4.py evaluate_v4.py render_v4.py base_rate.py
  render_v2.py render_v3.py figures_v2.py render_report_v3.py make_examples.py
  export_evidence.py export_static.py make_notebook.py verify.py run_all.py
  control_referenced.py   post hoc: each mechanism read against untreated controls
app/        FastAPI service and the single-page interface
site/       the same page as a static site
tests/      pytest suite
docs/       technical report, data, reproduction, committed predictions

The version 1 scripts (evaluate.py, figures.py, render_results.py, render_report.py, development.py, fit_regime.py, sensitivity.py) are kept for the record; they describe the version 1 model at the tag v1.

---

Documents

docs/TECHNICAL_REPORT.md	the full report (PDF in docs/)
results/v4/RESULTS.md	the four-compartment chip test (version 4), its predictions and both outcomes
results/v3/RESULTS.md	every version 3 outcome, the stop rule, the chip and design studies
results/v2/RESULTS.md	every version 2 outcome
PREREGISTRATION_v4.md, PREREGISTRATION_v3.md, PREREGISTRATION_v2.md	the tests, fixed before scoring
docs/DATA.md	sources, licences, compliance
docs/REPRODUCTION.md	how to rebuild everything

Author: Marc Donovici. Built with public data and open-source software only.