This repository studies spatial symmetry breaking in an explicit Delta–Notch–activated-Notch–proneural (D–N–A–ASC) feedback architecture, using linear stability analysis, spatial eigenmode decomposition, control theory, and nonlinear deterministic simulation. It contains the frozen, validated deterministic analysis behind a preprint, not the exploratory history that produced it.
State variables: D = Delta, N = Notch, A = activated Notch /
NICD-like downstream activity, S = ASC / proneural activity. The
biological loop: S_i -> D_i (proneural drives Delta); Delta on a
neighbouring cell activates that neighbour's Notch/A (neighbour D -> Notch/A, trans interaction); A_i represses S_i (A_i -| S_i),
closing the loop. A separate cis interaction couples D and N within
the same cell.
The exact implemented equations (model/equations.py,
model/jacobian.py), unchanged from the source repository:
D_dot = lam_D * h_D(S) - d_D*D - f_D*D*s_N - k_cis*N*D
N_dot = lam_N - d_N*N - f_N*N*s_D - k_cis*N*D
A_dot = -d_A*A + f_N*N*s_D
S_dot = -d_S*S + lam_S*h_S(A)
h_D(S) = 1 / (1 + (Kd/S)^n_d) (increasing in S)
h_S(A) = 1 / (1 + (A/Kd)^n_ASC) (decreasing in A)
s_D = sum_{j~i} D_j, s_N = sum_{j~i} N_j (six-neighbour contact sums)
Cells are represented on a periodic six-neighbour triangular contact
graph corresponding to a hexagonal cellular tiling (geometry/lattice.py,
geometry/graph_spectrum.py). Its adjacency operator W is diagonalized
by its own spatial eigenmodes, W v_mu = mu v_mu. The uniform mode has
mu = 6; the dominant, three-sublattice pattern-forming mode has
mu = -3 (the graph's minimum eigenvalue, attained only at the two
three-fold-periodic "K-like" points).
A perturbation along one spatial eigenmode decouples the full
multicellular linearization into one 4x4 ODE per graph eigenvalue
(model/jacobian.py, stability/mode_scan.py):
d xi_mu / dt = J_mu xi_mu
det(sI - J_mu) = 0
The leading eigenvalue/pole of J_mu determines whether a spatial
perturbation at mode mu decays or grows.
Partitioning x = (D, N, A, S) into z = (D, N, A) and the scalar S
gives the exact state-space form actually used in the analysis
(control/state_space.py):
z_dot = A_mu z + b S
S_dot = c^T z + a_S S
P_mu(s) = c^T (sI - A_mu)^{-1} b
L_mu(s) = P_mu(s) / (s - a_S)
det(sI - J_mu) = det(sI - A_mu) * (s - a_S) * [ 1 - L_mu(s) ]
verified against the direct 4x4 Jacobian eigenvalues to machine precision
(max coefficient difference 1.4e-13, control/schur_verification.py).
Key result: the stationary pattern-forming bifurcation occurs when the exact DC loop gain of the dominant three-sublattice mode reaches unity:
G_exact(-3) = L_{-3}(0) ~= 1 at lam_N ~= 1.305
This was derived from the exact Schur-complement representation above, not
assumed, and independently confirmed by an unconstrained root-locus pole
crossing at the same lam_N (control/root_locus.py). Full derivation:
docs/CONTROL_THEORY.md.
- Downstream Notch/proneural regulatory kinetics (
d_A, lam_N, Kd, n_ASC, d_ASC, lam_ASC) — not the classical cis/trans Delta-Notch coupling (k_cis, f_D, f_N) — dominate the spatial stability margin in this realization. - The dominant pattern-forming mode is the
mu = -3three-sublattice mode on the six-neighbour contact graph, not a square-checkerboard mode. - Geometry enters as a mode-dependent spatial gain: the same biochemical
S->D->A->Sloop is net-stabilizing atmu=6and net-destabilizing atmu=-3, purely through the sign of the graph eigenvalue. - The bifurcation is an exact unity-DC-loop-gain condition,
G_exact(-3)=1, not merely a numerically-observed sign change. - Nonlinear deterministic growth from a tiny (1e-4 relative) perturbation
matches the predicted unstable pole and saturates into a persistent,
finite-amplitude three-sublattice pattern (validated at
lam_N=2.0and5.623, zero numerical-integration clipping).
pip install -r requirements.txt
pytest # verifies the refactored code reproduces the frozen numbers
python figures/plot_fig1.py # and plot_fig2.py .. plot_fig5.pyEach figures/plot_fig*.py reads only the small CSV/NPZ files already
included in figures/data/ — no simulation or parameter search is
rerun. Figures are written to figures/main/.
python control/explain_geff.py # optional: recomputes figures/data/geff_comparison.csv
# from the included raw global sample (no new search)To rerun a piece of the deterministic analysis itself (fast; no search):
from model.params import P0, LAM_N_P2
from model.fixed_point import find_fixed_points
from stability.mode_scan import lambda_pattern
import dataclasses
p = dataclasses.replace(P0, lam_N=LAM_N_P2)
fp = find_fixed_points(p)[0]
print(lambda_pattern(fp, p)) # {'nonuniform_max_re_lambda': 0.103..., ...}| directory | contents |
|---|---|
model/ |
kinetic parameters, drift equations, homogeneous fixed-point solver, mode-resolved Jacobian, order-parameter metrics |
geometry/ |
six-neighbour contact lattice, graph spectrum / dispersion relation, reciprocal-space growth map |
stability/ |
spatial-mode scan (lambda_pattern), local sensitivity, 1-D parameter continuation |
control/ |
Schur-complement state-space form, algebraic verification, root locus, exact/heuristic loop gain |
simulation/ |
validated nonlinear deterministic ODE integration and controlled eigenmode perturbations |
figures/ |
scripts for the five main figures, plus the small source data they read |
docs/ |
condensed control-theory and geometry derivations, manuscript numbers |
tests/ |
fidelity checks against the frozen source analysis |
This repository contains the frozen deterministic analysis used for
the preprint: validated equations, fixed points, linear stability, control
theory, geometry, and nonlinear confirmation at two validated parameter
points (lam_N = 2.0, 5.623). It intentionally excludes the exploratory
history that produced it — in particular, an earlier dt=0.02
Euler-Maruyama result at a different (k_cis-only) baseline was found to
be a numerical integration artifact and is not reproduced here. One
preliminary stochastic check has since been added (stochastic_check/,
scripts only — results live in the main development repository per
stochastic_check/README.md): at validated point P2 (lam_N=2.0),
additive white noise on Delta shortens formation time (t50 -45% to -66%
across sigma=1e-4..1e-2) without changing final pattern amplitude, but
collapses final three-sublattice mode purity from 1.0 to ~0.13-0.22
(verdict S2); a deterministic random-IC control found ~60% of that purity
loss already occurs with zero noise, with ongoing noise contributing a
further, significant ~40% of the acceleration (verdict R3). This is one
check at one parameter point, not a systematic noise-timing study across
the bifurcation, which remains future work. The frozen
4000-point global parameter search that identified the dominant drivers
(figures/data/global_parameter_samples.csv) is included in full; its
root-finding step itself is not rerun here (no new search), but
control/explain_geff.py recomputes the G_exact-vs-G_eff comparison
from it and reproduces figures/data/geff_comparison.csv byte-for-byte.