🧬 anatomical-compiler

A higher-order network instrument for synthetic morphology & synthetic multicellularity

Treat a tissue's regulome as a hypergraph — then ask, quantitatively: did the engineered program execute, and how distinct is its modular organization from a primary-tissue blueprint?

---

🔭 What this is

Two things, tightly coupled:

1. A whitepaper — “Engineering Fully-Biologic Tissue Systems: A Higher-Order Network Instrument for Synthetic Morphology and Synthetic Multicellularity” (the fully-biologic tissue systems framing nods to Shiwarski & Feinberg's FRESH/CHIPS work — the wet-lab endpoint this instrument is built to serve). It connects a cluster of frontier areas rarely treated as one subject — self-organizing organoids, 3D/4D/freeform collagen bioprinting (FRESH, SWIFT), synthetic-morphogen / synNotch circuits, computer-designed biobots, bioelectric pattern control, single-cell regulome inference, the theory of biological modularity & identifiability, the call for predictive (not just descriptive) developmental biology, and the systems/tissue-organization view of cancer — and argues they share one missing capability: a way to read out whether an engineered tissue executed its intended program and how stable & distinct its modules are vs a primary blueprint. Then it builds the instrument, benchmarks it across real datasets, and lays out an experimental programme — toward what Levin calls the anatomical compiler (desired anatomy in → intervention out).
2. The computational engine behind it — hgx, a JAX/Equinox framework for hypergraph neural networks (+ devograph developmental extensions), which treats a regulome as a hypergraph of co-regulated gene batteries and yields three reusable readouts:

Readout	What it measures	Built on
🎯 Fidelity score	regulon-overlap (Jaccard, Fisher) + perturbation direction concordance vs primary tissue	hypergraph signal propagation, multi-hop in-silico KO
🧩 Module Identifiability Index	how sharp & stable a system's regulatory modules are	Hodge-Laplacian spectral gap on the regulon hypergraph
🌊 Hypergraph Neural ODE	splits a tissue's dynamics into stable structural drivers vs transient stress responders	latent neural ODE/SDE (diffrax), per-TF rollout error

flowchart LR
  subgraph Inputs["📥 Engineered & primary systems"]
    O["Organoids<br/>Fleck 2023 regulome"]
    B["3D/4D bioprinted<br/>kidney · liver · brain · GBM"]
    P["Programmed<br/>synNotch · synthetic morphogens"]
    E["Embodied<br/>Anthrobots · xenobots"]
    X["Perturbed / dynamic<br/>CRISPRi · kidney IRI · light-stim"]
  end
  subgraph HGX["⚙️ hgx · higher-order network instrument"]
    H1["regulome → hypergraph<br/>(incidence H, PPCA features)"]
    H2["🎯 fidelity score"]
    H3["🧩 Module Identifiability Index<br/>(Hodge Laplacian)"]
    H4["🌊 Hypergraph Neural ODE<br/>(drivers vs stress)"]
    H5["↩︎ SBI inverse<br/>(CellFlow flow-matching, Jacobian)"]
    H6["🎮 jaxctrl: controllability<br/>+ LQR/MPC steer-to-target<br/>(the anatomical compiler)"]
  end
  subgraph Blueprint["📐 Primary-tissue blueprints / targets"]
    NA["Neocortex Atlas (Sonthalia 2026)<br/>7 conserved mammalian patterns"]
    FK["Fetal kidney / cortex refs"]
    PC["Primary cortex CRISPRi (Pollen 2026)"]
  end
  Inputs --> H1 --> H2 & H3 & H4
  Blueprint --> H2 & H3
  H4 --> H6
  Blueprint -. target state .-> H6
  H2 & H3 & H4 & H6 --> R["📊 figures/*_results.json"]
  H5 & H6 -.design loop / intervention.-> Inputs
  R --> W["📄 paper.Rnw → paper.pdf<br/>(live knitr tables)"]

Loading

---

📄 The whitepaper

Canonical source: publication/paper.Rnw — a literate knitr / Sweave document. It knits to publication/paper.tex → publication/paper.pdf. Result tables and inline numbers are pulled live from figures/*_results.json (Python/JAX analyses) and data/cropseq/*.csv (R/Seurat differential expression), so the manuscript can never drift from the artifacts.

# build the whitepaper (R 4.x + tectonic)
Rscript -e 'knitr::knit("publication/paper.Rnw", output="publication/paper.tex")' && tectonic publication/paper.tex
# (or from inside publication/:  Rscript -e 'knitr::knit("paper.Rnw")' && tectonic paper.tex)

All knitr chunks are resilient: a missing result file produces a “[run scripts/benchmark_*.py]” note instead of failing — the document always knits.

Structure  |  §1 Introduction = the background review (synthetic morphology from Davies 2008 → tissue engineering → computationally designed tissues/organs/organisms → regulomes → modularity & identifiability → complexity science in oncology → synthetic multicellularity & Solé et al. 2024's open problems)  ·  §2 Computational Methods (hgx, regulome→hypergraph, Module Identifiability Index, Hypergraph Neural ODE/SDE, projectR-in-JAX, fidelity metrics, CellFlow SBI, TDA, dataset table)  ·  §3 Results (14 subsections, multi-dataset)  ·  §4 Conclusions & Outlook (what the modeling demonstrates + a 6-experiment forward programme)  ·  full bibliography.

📚 The background review, the synthetic-morphogenesis primer, and the manuscript prose are all in paper.Rnw (they used to be separate FOUNDATIONS.md / SYNTHETIC_MORPHOGENESIS.md / MANUSCRIPT.md files — now consolidated). The human-readable master bibliography is kept alongside it at REFERENCES.md.

🧾 Provenance & CURE alignment

• MODEL_CARD.md — index of the project's 8 major computational artifacts (Hypergraph Neural ODE, MII, fidelity-triple predictor, Lab-6 controllability, anatomical compiler, FM-prior caches, BETSE-JAX, cpjax) with intended use / source data / metrics / limitations / references per Mitchell et al. 2019.
• docs/cure-audit.md — compliance audit against Sauro et al. 2025 (FAIR → CURE), the COMBINE-community guidelines for Credible / Understandable / Reproducible / Extensible computational biology models (ref. 99a). Includes the priority list of remaining gap-closure actions.
• Dockerfile — two-stage reproducible container:

  docker build -t anatomical-compiler:baseline .              # CPU, stub-mode + all ablations
  docker build --target fm -t anatomical-compiler:fm .        # adds Geneformer/scGPT/UCE/Evo/Borzoi + JASPAR

  The baseline image runs the full educational track + ablations in stub mode without a GPU; the fm target is the DGX Spark / real-mode build per docs/dgx-spark-setup.md. Build-time smoke tests (ablate_edge_priors.py + ablate_perturb_eig.py) bake into the image so a broken environment fails the build.

---

🧫 From measurement to experiments

The point of the survey-plus-tool is the programme it enables (whitepaper §4.3) — a model-in-the-loop, engineering-biology agenda aimed at the core question of synthetic morphology (build predictably, with the cells' own competencies; know when you have). Abstractly, every item is one optimal-control problem on the Hypergraph Neural ODE — given a target tissue state, find the actuation (print geometry, synNotch input, light schedule, dose, bioelectric set-point) that gets there — i.e. Levin's anatomical compiler, with jaxctrl (differentiable LQR/MPC, controllability Gramians, structural & hypergraph controllability) as the solver:

#	Experiment	Instrument readout in the loop	Tech
(i)	4D-bioprinting maturation assay — sweep geometry/curvature/matrix relaxation; let an optimiser pick the next print	pattern-projection + Module Identifiability Index as live objective	FRESH collagen I/II/III · CHIPS perfusable scaffolds · SWIFT vascular channels · model-guided vasculature · open-source PRINTESS (Skylar-Scott lab, built in-lab)
(ii)	Hybrid programmed-plus-printed tissues — synNotch circuit × bioprinted geometry	“hypergraph state” alignment + ODE driver-stability (is it an attractor?)	synNotch / synthetic morphogens + bioprinting
(iii)	Optogenetic morphogenesis — designed WNT/SHH/BMP schedule	Hodge-Laplacian “stop-signal” / commitment threshold	optogenetic induction + scRNA-seq
(iv)	Microglia / vasculature titration — dose series into organoids	Hypergraph Neural ODE: dose at which mature drivers cross transient → stable	iPSC-microglia · engineered vasculature
(v)	Bioelectric control layer — perturb Vmem / gap-junctional coupling	shift in Identifiability Index & ODE driver set	bioelectric reprogramming
(vi)	Cancer-as-loss-of-module-identifiability assay — primary → organoid → tumour organoid → cancer line	Index falls, driver set collapses, unicellular↔multicellular gene balance shifts	tissue-organization / atavism / attractor views, operationalised

Two passes at the control layer ship now (jaxctrl is a uv sync dependency):

• Linear warm-up — scripts/benchmark_network_control.py: jaxctrl on the Pando TF co-regulation graph — Kalman/structural controllability from the master-regulator set, per-TF control-leverage (single-input Gramians), a steer-to-target (early→late pseudotime) energy + LQR law. Finding: the regulome is steerable with broad actuation but not from the master regulators alone (the static graph collapses most directions into a slow, weakly-actuated subspace; the master TFs are the privileged handles of the nonlinear flow, not the linearisation) → so the real control problem is on the Hypergraph Neural ODE. Output: figures/network_control{,_results.json}.
• The anatomical compiler, end-to-end — scripts/benchmark_anatomical_compiler.py: fit a Hypergraph Neural ODE to the kidney injury-repair timecourse, then — by direct shooting through that learned ODE (diffrax adjoints + Adam, jaxctrl LQR on a linear surrogate as warm-start) — compute the TF-actuation schedule that drives the early-injury state to the recovered state. The optimised schedule cuts the distance to the target on the actuated TFs by ~73%. Learned tissue-dynamics model + target state → explicit intervention schedule, all differentiable — the minimal proof of the loop. Output: figures/anatomical_compiler{,_results.json}.

Both are read live by paper.Rnw §3. For the smallest standalone worked examples see jaxctrl's examples/repressilator_control_demo.py (quench a 3-gene oscillator), examples/irma_sindy_lqr.ipynb (SINDy → LQR on an IRMA-topology GRN), and examples/grn_hypergraph_drivers.ipynb (minimum driver TFs / control-energy on a GRN-as-hypergraph).

The bioprinting lineage behind (i): Feinberg lab FRESH freeform collagen printing → Shiwarski open-source bioprinting hardware (CHIPS/VAPOR) → Skylar-Scott lab SWIFT & the $250 open-source PRINTESS → model-guided synthetic vasculature fed straight to the printer. (Citations in REFERENCES.md / paper.Rnw.)

---

📊 Results at a glance

Across ~20 datasets mapped onto Solé's open problems (full table in paper.Rnw §2.10; per-dataset numbers in figures/*_results.json):

• Organoid regulatory logic is conserved. Organoid GRN topology predicts CRISPRi targets in primary human cortex — 8 TFs survive Bonferroni, ~91% direction concordance — and concordance rises with tissue context (2D screen → 3D slice → in vivo); key regulons conserved to marmoset and mouse.
• The organoid “fidelity gap” is specific & locatable. It sits in the mature / layer-specific / human-specific programs (Neocortex-Atlas Pattern 2; the oRG and synaptic / ASD-enriched patterns), with a regulatory correlate — the Early-Stage Buffer (early master regulators depleted for mature disease genes).
• Bioprinting closes much of it. 3D bioprinted brain tissue: ~10× higher synaptic + outer-radial-glia pattern activity. 4D conformation control: ~3.2× kidney proximal-tubule maturity. Liver hepatorganoids: master regulators HNF4A / FOXA2 strongly induced in 3D vs 2D.
• The instruments measure organisation itself. Module Identifiability Index ranks brain organoid ≈ fetal kidney > bioprinted kidney (the biofabricated construct converging toward primary). The Hypergraph Neural ODE on a kidney injury-repair time course cleanly separates stable regenerative drivers (Lhx1, Cdh1, Pax8/2, Six1/2, Wt1, Foxc2; rollout MSE ≤ 0.11) from transient stress markers (Fos 4.4, Jun 1.5, Cd44 0.93, Atf3 0.80).

---

⚡ Speed & accuracy

hgx vs DHG — 5–120× faster inference (Fleck organoid GRN, NVIDIA GB10)

Model	Framework	Inference	Train (200 ep)
UniGCNConv	hgx / JAX	1.48 ms	5.39 s
UniGATConv	hgx / JAX	2.09 ms	4.84 s
UniGINConv	hgx / JAX	3.22 ms	6.55 s
HGNN+	DHG / PyTorch	10.77 ms	3.65 s
HyperGCN	DHG / PyTorch	256.50 ms	53.97 s

2,792 nodes, 720 hyperedges. Inference averaged over 100 forward passes with CUDA sync.

hgx matches published HGNN baselines (Cora / Citeseer / Pubmed)

Cora (2,708 nodes, 7 classes, 1-hop construction, Planetoid splits, 5 seeds, early stopping):

Model	Accuracy	Source
HGNN	79.39%	Feng et al. 2019
hgx UniGCNConv	78.72 ± 1.06%	this work
UniGCN	78.95%	Huang & Yang 2021
AllSet	78.58%	Chien et al. 2022
HyperGCN	78.45%	Yadati et al. 2019
hgx UniGATConv	77.96 ± 0.76%	this work
hgx UniGINConv	72.70 ± 1.86%	this work

Citeseer (3,327 / 6): hgx UniGCNConv 64.80 ± 0.82% (7-pt gap consistent across normalizations — likely a construction difference in published HGNN). Pubmed (19,717 / 3, 60 train labels): hgx UniGCNConv 76.10%, processed in 15.5 s / 6.7 GB.

Accuracy: task matters more than architecture

The 720-regulon task has 258 singleton classes (unlearnable). With balanced tasks, hgx performs as expected:

Task	Classes	Best hgx	Accuracy	vs random
Spectral clusters	20	UniGINConv	94.6%	19×
Lineage prediction	3	UniGINConv	77.2%	2.3×
TF vs target	2	UniGINConv	77.1%	1.5×
Regulon assignment	720	UniGINConv	9.1%	36×

The apparent HyperGCN accuracy lead disappears under proper class balance.

---

✅ Biological validation

All 5/5 Fleck et al. (2023) checks pass:

Check	Result	Detail
TF centrality	✅ PASS	5/8 master regulators in top 100/720 TFs (composite rank)
Regulon coherence	✅ PASS	within-regulon genes 6.5× more correlated than between
GLI3 KO direction	✅ PASS	4/5 genes correct via multi-hop hypergraph propagation; cross-checked vs real CROP-seq DE (GLI3↔TBR1 log2FC correlated, R/Seurat)
Pseudotime patterns	✅ PASS	TBR1, NEUROD6 show late-stage increase
Fate probabilities	✅ PASS	DF ↑ (r = 0.80), MH ↓ (r = −0.74) along pseudotime

---

🔁 Simulation-based inference & CellFlow

Forward: hgx Hypergraph Neural ODEs simulate TF knockouts. Inverse: CellFlow (flow matching) learns velocity fields from real CRISPRi distributions; Jacobian analysis (∂V/∂X) attributes regulatory drivers and compares them to the biological GRN — a posterior estimate of the regulome benchmarked against Pando/GRN-VAE-style methods. Details: docs/sbi_integration.md.

---

🚀 Quick start

GPU server (DGX Spark / A100) — the compute pipeline

git clone https://github.com/m9h/anatomical-compiler.git
git clone https://github.com/m9h/hgx.git
git clone https://github.com/m9h/devograph.git
pip install -e hgx -e devograph
pip install "jax[cuda12]" equinox diffrax optax scanpy anndata ripser
cd anatomical-compiler
# download data from Zenodo — see docs/data_preprocessing.md
python scripts/00_preprocess.py        # PPCA features, pseudotime bins, fate probs (~13 s)
python scripts/generate_figures.py     # 8 publication figures on GPU (~175 s)
python scripts/validate_against_pando.py   # 5 biological checks
python scripts/benchmark_comparison.py     # hgx vs DHG (speed + accuracy)
python scripts/accuracy_ablation.py        # hyperparameter sweep + task comparison
# ... plus the synthetic-multicellularity tracks:
python scripts/benchmark_toda_morphogenesis.py     # programmed patterning (synNotch)
python scripts/benchmark_anthrobot_fidelity.py     # embodied self-assembly
python scripts/benchmark_vorganoid_crosstalk.py    # vascularization / metabolic wall
python scripts/benchmark_regenerative_flow.py      # kidney IRI Hypergraph Neural ODE
python scripts/test_nitmb_modularity.py            # Module Identifiability Index
python scripts/benchmark_network_control.py        # linear controllability / steer-to-target (jaxctrl — a `uv sync` dependency)
python scripts/benchmark_anatomical_compiler.py    # nonlinear optimal control on the learned Hypergraph Neural ODE (the anatomical compiler)

Build the whitepaper

# needs R 4.x (knitr, jsonlite) + tectonic; reads figures/*_results.json + data/cropseq/*.csv
Rscript -e 'knitr::knit("publication/paper.Rnw", output="publication/paper.tex")'
tectonic publication/paper.tex          # → publication/paper.pdf

Google Colab

Open scripts/organoid_hgx_colab.ipynb with an A100 runtime.

---

🗂️ Where things live

publication/
  paper.Rnw                 ← 📄 canonical whitepaper (knitr/Sweave; edit this, then re-knit)
  paper.tex                 ←   generated — do not hand-edit
  paper.pdf                 ←   compiled (tectonic)
  *.png, presentation/      ←   figures + Beamer deck
scripts/
  00_preprocess.py          ← Zenodo data → modeling arrays (PPCA k=97)
  generate_figures.py       ← 8 core figures (GRN, modules, trajectory, eigenspectrum, spectral, ODE/SDE, perturbation, persistence)
  validate_against_pando.py ← reproduce Fleck et al. (2023)
  benchmark_comparison.py   ← hgx vs DHG (standard + organoid datasets)
  accuracy_ablation.py      ← LR/depth/hidden/dropout sweep × 4 tasks
  benchmark_*.py            ← synthetic-multicellularity tracks (Toda, Anthrobot, vOrganoid, regenerative flow, learning regulome, disease enrichment, bioprinting, ...)
  test_nitmb_modularity.py  ← Hodge-Laplacian Module Identifiability Index
  benchmark_network_control.py     ← linear network controllability + steer-to-target (jaxctrl)
  benchmark_anatomical_compiler.py ← nonlinear optimal control on the learned Hypergraph Neural ODE (the anatomical compiler)
figures/
  *_results.json            ← per-dataset artifacts (consumed live by paper.Rnw)
  *.png                     ← per-track figures
data/
  cropseq/*.csv             ← real CROP-seq DE (R/Seurat) — consumed live by paper.Rnw
  bioprinting/, choose/, krienen/, mouse/, ...   ← processed h5ad per dataset
notebooks/                  ← Colab notebook + the seed of an educational "course" track (see notebooks/README.md)
hgx_prep/                   ← hgx-prep CLI: standardized GRN dataset preparation
docs/                       ← data_preprocessing.md · benchmark_datasets.md · sbi_integration.md
README.md · ROADMAP.md · REFERENCES.md    ← (root) overview · plan · master bibliography

---

📦 Data

Processed data from Zenodo 10.5281/zenodo.5242913: 74,448 Pando GRN edges · 720 TFs · 2,792 genes · 34,088 cells with pseudotime, lineage, CellRank fate probabilities (DF/VF/MH) · PPCA/MELODIC consensus k = 97 (AIC 168, BIC 26). External resources: SJD joint matrix decomposition (GitHub), projectR (Bioconductor), Neocortex Development compendium (GitHub).

---

📚 References

Full theme-grouped bibliography with DOIs: REFERENCES.md (mirrored in the thebibliography block of paper.Rnw). Load-bearing:

• Davies JA (2008) — Synthetic morphology: prospects for engineered, self-constructing anatomies. J Anat 212(6):707–719. doi
• Davies JA (2022) — Synthetic Morphogenesis: introducing IEEE journal readers to programming living mammalian cells to make structures. Proc IEEE 110(5):688–707. doi
• Davies J & Levin M (2023) — Synthetic morphology with agential materials. Nat Rev Bioeng 1:46–59. doi
• Solé R, et al. (2024) — Open problems in synthetic multicellularity. npj Syst Biol Appl 10:151. doi
• Hartwell LH, et al. (1999) — From molecular to modular cell biology. Nature 402:C47–C52. doi
• Fleck JS, et al. (2023) — Inferring and perturbing cell fate regulomes in human brain organoids. Nature 621:365–372. doi
• Sonthalia S, et al. (2026) — Neocortex Atlas: a compendium of transcriptomic data. Nat Neurosci. doi
• Lee A, et al. (2019) — 3D bioprinting of collagen to rebuild components of the human heart (FRESH v2.0). Science 365(6452):482–487. doi
• Sexton ZA, et al. (2025) — Rapid model-guided design of organ-scale synthetic vasculature for biomanufacturing. Science 388(6752):1198–1204. doi
• Levin M (2022) — Technological approach to mind everywhere. Front Syst Neurosci 16:768201 (the "anatomical compiler"). doi
• Liu YY, Slotine JJ & Barabási AL (2011) — Controllability of complex networks. Nature 473:167–173. doi
• hgx: github.com/m9h/hgx · devograph: github.com/m9h/devograph · cellflow: github.com/m9h/cellflow · jaxctrl: github.com/m9h/jaxctrl

_{Biopunk Lab · built with hgx (JAX/Equinox) · manuscript via knitr + tectonic}