A quantitative systems pharmacology model of MOGAD built around one awkward observation: rituximab lowers the MOG-IgG titre more than IVIG does, and prevents far fewer relapses. Any model in which relapse risk simply tracks antibody concentration gets that backwards. This one gets it right, and the two structural choices that make it work — a short-lived CD20-negative antibody source, and a threshold hazard in functionally active antibody — then also predict why the oligodendrocyte survives, why serum GFAP stays low, and why steroid tapers buy time rather than remission.

Every simulated number on this page is reproduced by python3 mogad_reference_impl.py — a dependency-free RK4 twin of the same 34 ODEs with the same parameters. Nothing here is asserted from intuition.

MOGAD is an antibody-mediated demyelinating disease of the central nervous system. Class-switched IgG1 against myelin oligodendrocyte glycoprotein reaches the CNS through a permeabilised blood–brain barrier and binds MOG, which sits on the outermost lamella of the myelin sheath — so the epitope is directly accessible without the antibody ever entering a cell. Bound IgG1 fixes complement (C1q → C4 → C3 → C5b-9) and opsonises myelin for phagocytosis by microglia and monocyte-derived macrophages. Attacks present as optic neuritis (often bilateral, with disc oedema and severe acuity loss), longitudinally extensive myelitis with conus predilection, ADEM in children, or unilateral cortical encephalitis with seizures (FLAMES). Recovery is usually good, relapses occur in roughly half of patients, and — unlike multiple sclerosis — there is no secondary progressive phase: disability accrues only in steps, from attacks.

Everything in the model follows from three choices. Each is stated here so it can be attacked.

The pathogenic IgG1 comes from a plasmablast pool with a 7-day half-life (KOUT_PB = 0.14/day) sustained by an IL-6-dependent survival niche. Only part of its generation is fed by CD20+ memory B cells; the parameter FRAC_ESC = 0.60 is the share that rituximab structurally cannot reach.

Consequences that fall out rather than being imposed: titres drop within days of steroids or plasma exchange; they rebound over about a month once suppression stops; and rituximab has a hard ceiling on how far it can lower them no matter how completely it depletes CD20+ cells.

Myelin (MYEL_*) and axons (AXON_*) are separate states, and the remyelination term is multiplied by the surviving oligodendrocyte fraction OL. In MOGAD the oligodendrocyte death coefficient KOLD is small, matching the neuropathology (Höftberger 2020, Takai 2020: demyelination with preserved oligodendrocytes and intact AQP4).

Setting AQP4_MODE = 1 multiplies oligodendrocyte death and astrocytic GFAP release, converting the same model into an AQP4-IgG NMOSD comparator. This is the model's built-in control experiment: identical antibody kinetics, identical barrier, identical effectors, identical steroids — one switch.

The recovery gap and the GFAP discrimination — both well documented clinically (Stiebel-Kalish 2017; Marignier 2025) — come entirely from whether the target cell survives the attack. Note the NfL columns are nearly identical: the model says these two diseases differ in glial injury, not axonal injury, which is exactly the reported dissociation.

Commitment 2 also forces a subtler feature. To fit a severe nadir and good final acuity and marked RNFL thinning simultaneously, visual deficit has to be split into reversible conduction block (proportional to demyelination) plus irreversible axonal loss beyond a functional reserve (AXRES = 0.15). Without that reserve term no parameterisation fits all three.

Hazard is LAM0 × Hill(AB_EFF, AB50 = 1.8, h = 3) × amplification, where AB_EFF is the functionally active antibody, not the measured titre. IVIG lowers AB_EFF far more than it lowers the titre, because it also blocks Fcγ receptors, competes for FcRn and scavenges complement.

Because the untreated titre sits on the Hill plateau, partial suppression of the source buys very little. This is where the model earns its keep:

Rituximab lowers the titre more than IVIG and prevents far fewer relapses. The C5 inhibitor lowers the titre not at all and still helps. Titre reduction and relapse protection are not the same axis, and which node a drug acts on decides the outcome.

Falsifiable consequences. Serial MOG-IgG titres should be a poor surrogate for treatment benefit. IVIG's benefit should persist in patients whose titres do not fall. Both are testable in cohorts that already exist.

Annualised relapse rate, 365-day horizon, index optic neuritis, IV methylprednisolone from day 4 and a guideline-like 90-day oral wean:

LAM0 was set from the 0.64 figure; the rest of the column is a consequence of the structure rather than of per-drug fitting.

Attack course under standard treatment: visual acuity nadir logMAR 1.60, recovering to 0.37 (~20/47) at one year; RNFL 98 → 75 µm; serum NfL 8 → peak 160 pg/mL; CSF peak 72 cells/µL; surviving oligodendrocytes 0.88. Total IgG at steady state runs 15.0 g/L on maintenance IVIG, 5.7 g/L on rituximab (hypogammaglobulinaemia) and 4.7 g/L on the FcRn inhibitor.

The model was built expecting taper duration to matter. It does not. At a matched cumulative dose of 840 mg prednisone-equivalent:

Stretching the wean from two weeks to six months moves the one-year relapse probability from 0.500 to 0.479. Raising the total dose helps somewhat more (2700 mg over 90 days → 0.46) at a measurable bone cost. The mechanistic reading: oral steroids buy time roughly in proportion to exposure and cannot deliver remission — which is an argument for starting a steroid-sparing agent early, not for tapering slowly.

Steroid dependency itself is emergent. Nothing in the model says "relapse when steroids stop". The hazard depends on antibody; antibody depends on a plasmablast pool that glucocorticoids suppress reversibly; when the wean ends, plasmablasts recover in a fortnight and the titre follows over a month, so the hazard climbs back on its own.

Mechanistic map — 181 nodes, 263 edges, 17 subgraph clusters: host susceptibility and triggers · peripheral B lineage and the plasmablast source · antibody biology and FcRn handling · T-cell and cytokine amplification · blood–brain barrier transit · target antigen and oligodendrocyte biology · complement and phagocytic effectors · tissue outcome and repair · the AQP4 contrast arm · optic neuritis · myelitis · brain phenotypes (ADEM, FLAMES, brainstem) · acute therapy · maintenance therapy · biomarkers · clinical endpoints · treatment toxicity.

mrgsolve model — 34 ODE compartments, 119 parameters, 33 captured outputs, 17 prebuilt scenarios. Drug arms: IV methylprednisolone, oral prednisone with a continuously tunable linear taper, plasma exchange, acute and maintenance IVIG, rituximab, IL-6R blockade, antimetabolites, FcRn inhibition, C5 inhibition, plus the AQP4 comparator switch and an intercurrent-infection trigger.

Shiny dashboard — 12 tabs, including a taper explorer that re-runs the matched-dose sweep live and a side-by-side MOGAD/NMOSD comparator.

library(mrgsolve)
mod <- mread_cache("mogad", project = ".", file = "mogad_mrgsolve_model.R")
# MOGAD_simulate_scenarios() and MOGAD_taper_sweep() are supplied as commented
# R code at the foot of the model file (so that mread() sees a clean spec) —
# copy that block into your session, then:
df <- MOGAD_simulate_scenarios(365)
shiny::runApp("mogad_shiny_app.R")

python3 mogad_reference_impl.py        # regenerates every number quoted above
dot -Tsvg mogad_qsp_model.dot -o mogad_qsp_model.svg
dot -Tpng -Gdpi=150 mogad_qsp_model.dot -o mogad_qsp_model.png

Stated here rather than buried, because they are where the model is most likely to be wrong.

1. IL-6R blockade is under-predicted — model 0.23 against a reported 0.09. Two readings, and the model cannot distinguish them: either the observed incidence-rate ratio of 0.08 is a pre-versus-on-treatment comparison inflated by regression to the mean (the trustworthy contemporaneous comparison, against IVIG ≥1 g/kg, was not significant — in which case the model is right), or the IL-6 axis carries more weight than EMAX_IL6 and KAMP allow. Randomised data would settle it.
2. FRAC_ESC = 0.60 is fitted, not measured. It is the single parameter that makes rituximab weaker than IVIG, so the whole source-versus-clearance conclusion rests on it. The competing explanation — that relapses on rituximab cluster in B-cell repopulation windows rather than arising from CD20-negative escape — corresponds to a low FRAC_ESC plus under-dosed redosing, and predicts that strict CD19-guided redosing rescues rituximab. That is the experiment which separates the two.
3. Azathioprine and mycophenolate share one arm, so the reported ordering between them (azathioprine apparently better) cannot be reproduced. That ordering is mechanistically unexplained and may be channelling in retrospective series; the model takes no position.
4. Attack nadir is under-predicted and recovery is pessimistic — nadir logMAR 1.60 against a reported average of count-fingers, final acuity 20/47 against a reported average 20/30. Both would improve with an explicit conduction-block state separate from myelin content, which is not implemented.
5. Relapses are a hazard, not events. No attack actually fires from the hazard, so step-wise disability accrual across multiple attacks is not simulated. Drawing attack times from the hazard and injecting TRIG is the natural extension.
6. No paediatric ADEM physiology. Encephalopathy, seizures and cognitive outcome appear in the mechanistic map but have no ODE counterpart, and the much higher rate of transient seropositivity in children is not modelled.
7. Titre-to-assay mapping is linear. Real live cell-based assay dilutions are ordinal and step in twofold increments, so comparisons of simulated titres to reported dilutions are order-of-magnitude only.

Educational and research model only. Parameters are approximations fitted to published aggregate cohort data, not to individual patients; the model has not been independently validated, and it must not be used for clinical decisions, prescribing or regulatory submissions.