Plexus

A unified, differentiable framework for inverse problems in living tissue.

Plexus re-converges four sibling frameworks that all forked from ParticleGraph and then drifted:

Framework	Domain	Becomes (in Plexus)
CellGraph (cell-gnn)	collective cell interactions	Lateral operators @ cell level
NeuralGraph	neural / signaling networks	Lateral operators @ cell level
MPMGraph (MPM_pytorch)	tissue mechanics, deformable matter	Exchange (P2G/G2P) @ particle level
MetabolismGraph	bipartite reaction networks	Exchange (stoichiometry) @ molecule level

The unification is built on one missing primitive: a hierarchical graph container in which entities are sets at nested scales and dynamics are operators (GNNs implementing ODE vector fields + Laplacians) dispatched by relation. See paper/plexus.pdf for the concept, schematics, and glossary.

The abstraction in one paragraph

A tissue is a stack of sets linked by containment (a particle belongs to a cell, a cell to a population). Continuous quantities (metabolites, morphogens, mechanical momentum) are fields, each on its own grid, each bound to one level. Operators move state: Lateral (within a set), Aggregate ↑ / Broadcast ↓ (across containment), Exchange (set ↔ field). A model is a Schedule — an ordered, multi-rate list of operators — declared in config, not hand-coded per system.

Layout

src/plexus/
  models/
    base.py       # Level, Field, Hierarchy, Schedule, Operator  (the container)
    registry.py   # @register_entity / @register_operator / @register_field
    catalog.py    # the filled registry: operators that run at every level
  generators/     # ground-truth simulators per domain
config/           # schedule + hierarchy declarations (yaml)
paper/            # concept paper: plexus.tex / .md / .pdf + figures/
prototype/        # working 2-level MPM tissue prototype + rendered mp4s
*.qmd, _quarto.yml # Quarto website source
docs/             # rendered website (GitHub Pages output) — do not edit by hand
graphs_data/      # generated datasets

Website

The blueprint is published as a Quarto site (sources are the *.qmd files at the repo root; rendered output lives in docs/ for GitHub Pages).

quarto render          # build docs/
quarto preview         # live preview

Status

Architecture / scaffolding stage. base.py defines the abstraction; catalog.py registers operators at all levels as stubs to be filled by porting the existing repos.