analysis/t1d_mechanism.png
Author: Seth Cope¹ (¹ Independent researcher. ORCID: 0009-0000-5520-915X)
Preprint draft — illustrative modeling study / hypothesis generation. Not validated experimental
or clinical findings. Convert to PDF (e.g. pandoc MANUSCRIPT.md -o manuscript.pdf) for posting.
Background. Two disease-modifying immunotherapies are advancing in early (pre-symptomatic) type 1 diabetes (T1D): anti-CD3 monoclonal antibody (teplizumab-class), which delays clinical onset, and antigen-specific tolerance therapy (peptide / mRNA / tolerogenic constructs), which re-educates the autoimmune response. Both are being pushed toward combination use, so how to combine them is a live question. Foster et al. (2025, NOD mice) reported an unexplained result — anti-CD3 reduces the efficacy of antigen-specific immunotherapy — that no mechanistic model addresses. Methods. We build a three-state within-host ODE in which autoreactive effector T cells (E) and antigen-specific regulatory T cells (R) form a mutual-repression bistable switch that drives β-cell mass (B). Antigen-specific tolerance converts effectors into Tregs (a flux that requires effectors to be present); anti-CD3 is lymphodepleting (it removes effectors and partly depletes Tregs). We score a cohort durable-control fraction and sweep the order and inter-drug interval of the two therapies. Results. The model reproduces the Foster antagonism and resolves it into a sequencing rule. Tolerance monotherapy controls 100% of the cohort; giving anti-CD3 simultaneously drops this to 59%, and anti-CD3-first to 41%, whereas tolerance-first fully rescues (100%). The antagonism arises because anti-CD3 removes the effector substrate that tolerance must convert into protective Tregs. The optimal protocol is tolerance-first or, if anti-CD3 must precede, a sufficient inter-drug gap (anti-CD3-first recovers 59%→100% as the gap grows to ~1.5 yr). The rule is robust: tolerance-first ≥ simultaneous in 171/171 viable parameter sets in the Treg-sparing regime, never worse. A second, falsifiable sub-prediction: the optimal order inverts if anti-CD3 strongly depletes Tregs. Conclusions. We predict that combining anti-CD3 with antigen-specific tolerance should be done tolerance-first, testable directly in a NOD sequencing experiment. Illustrative modeling hypothesis, conditional on the stated assumptions.
Type 1 diabetes results from autoimmune destruction of insulin-producing β-cells, with a long pre-symptomatic prodrome (stage 1–2) during which intervention can delay or prevent clinical (stage 3) disease. Anti-CD3 monoclonal antibody (teplizumab) delayed progression from stage 2 to stage 3 by a median of roughly two years in the TN10 trial — the first disease-modifying therapy approved for this window. A complementary strategy, antigen-specific tolerance (peptide, mRNA, or tolerogenic nanoparticle constructs), aims to re-educate autoreactive T cells into a regulatory phenotype rather than broadly debulk them. Because the two act by different mechanisms, combining them is an obvious next step, and several groups are advancing antigen-specific tolerance explicitly as a combination partner.
Against this momentum, Foster et al. (2025, NOD mice; ADA abstract 2136-LB) reported a surprising result: adding anti-CD3 reduced the efficacy of antigen-specific immunotherapy. An earlier NOD study (Stewart et al., 2020) had likewise found that antigen-specific microparticles plus anti-CD3 “fail to synergize.” The phenomenon is therefore reproducibly observed but unexplained, and the T1D “modeling” literature is statistical / machine-learning, with no mechanistic within-host model of the effector–Treg–β-cell system under these two therapies. We build the smallest such model that reproduces the antagonism and ask what it implies for how to combine the therapies.
Three states (time in years): β-cell mass B (fraction of healthy, ~ proportional to C-peptide),
autoreactive effector burden E, and antigen-specific Treg burden R. E and R form a
mutual-repression bistable switch (each self-promotes via a Hill term, each represses the other —
the canonical motif for immune cell-fate decisions; cf. Alexander & Wahl 2011), and β-cell mass grows
logistically and is killed in proportion to effector burden:
dE/dt = bE + VE·E^n/(K^n+E^n)/(1+(R/Ki)^n) − dE·E − u_a3(t)·E − u_tol(t)·E
dR/dt = bR + VR·R^n/(K^n+R^n)/(1+(E/Ki)^n) − dR·R + φ·u_tol(t)·E − ρ·u_a3(t)·R
dB/dt = ρB·B·(1−B) − κ·E·B
The switch has two stable basins: autoimmune (E high, R low → β-cells lost; B → 0.002) and
tolerant (E low, R high → β-cells preserved; B → 0.931). Late stage 2 sits in the autoimmune
basin. The two interventions deliver the same total drug; only order/overlap differs.
Antigen-specific tolerance converts effectors to Tregs (the flux φ·u_tol·E; it needs
effectors present to convert). Anti-CD3 is lymphodepleting: it removes effectors (u_a3·E) but
also depletes Tregs (ρ·u_a3·R), so on its own it transiently debulks E and the switch reverts.
Outcomes are reported as a cohort durable-control fraction (the bistable switch makes per-patient
outcomes binary, so the antagonism is read across a severity cohort). Equations, parameters, the
cohort sampling, and numerics are documented in analysis/METHODS.md; every headline number below is
re-derived and asserted by analysis/verify_claims.py (24/24 checks pass).
3.1 The model reproduces the antagonism (Fig. 1, Fig. 2). Across a severity cohort, durable control (β-cell mass above threshold at 5 years) was: untreated 0%, anti-CD3 monotherapy 0% (it delays but does not durably control), antigen-specific tolerance monotherapy 100%. Adding anti-CD3 simultaneously dropped control to 59% (a 41-point loss), reproducing the Foster antagonism. Mechanistically, anti-CD3 removes the effector pool that tolerance must convert into self-stabilizing Tregs, so the switch fails to flip.
3.2 The antagonism resolves into a sequencing rule (Fig. 2). Order matters: anti-CD3-first gave
only 41% durable control, whereas tolerance-first fully rescued at 100% — the ordering
tolerance-first ≥ tolerance-only > simultaneous > anti-CD3-first is the mechanism’s signature. The
falsifiable operational prediction is an optimal inter-drug interval (Fig. 3): tolerance-first is
flat at ~97–100% for any gap, while anti-CD3-first recovers monotonically with the gap (59% at gap 0
→ 100% at a ~1.5-year gap), i.e. if anti-CD3 must precede, one must wait for the effector pool to
recover before giving tolerance. Simultaneous dosing is the worst protocol.
3.3 Robustness, and a discovered two-channel caveat. Over a five-parameter grid, in the Treg-sparing regime (anti-CD3 spares Tregs, ρ<1 — matching teplizumab’s documented profile), tolerance-first ≥ simultaneous in 171/171 viable parameter sets (100%), never worse (strictly better in 26; the remainder saturate, so order is irrelevant there). Including strongly Treg-depleting anti-CD3 (ρ up to 1.1), the optimal order can invert (17/228 sets, all at the highest ρ). Two antagonism channels are at work: substrate-depletion (anti-CD3 deletes the effectors tolerance needs → favors tolerance-first) and Treg-destruction (anti-CD3 destroys freshly-built Tregs → favors anti-CD3-not-last). Which dominates is set by how Treg-depleting anti-CD3 is — a second, sharply falsifiable prediction.
3.4 Calibration (Fig. 4). A stage-2 cohort heterogeneous in effector severity and residual β-cell mass reproduces the untreated progression curve: median time-to-diagnosis 2.06 years (TN10 placebo ~2.0 yr) with ~45% progressed by 2 years (TrialNet stage-2 ~50%). The sequencing antagonism persists on this clinically-anchored cohort (tolerance-only 100% → simultaneous 65% → anti-CD3-first 35% → tolerance-first 97%).
3.5 A derived criterion that replicates the model (Fig. 5). Because each course (~2–4 weeks) is
near-impulsive relative to the year-scale switch, it integrates exactly to a linear map on (E, R):
tolerance gives E→T·E, R→R+φ(1−T)E and anti-CD3 gives E→A·E, R→A^ρ·R, with T=e^{−σ_tol·τ_tol}
and A=e^{−σ_a3·τ_a3}. Two closed-form results follow, both verified against the full ODE. (i) An
antagonism factor: co-administration retains only 𝒜 = Y_sim/Y_tol = A^ρ·σ_tol/(σ_tol+(1−ρ)σ_a3)
≈ 0.47 of the tolerogenic Treg yield — and this factorizes into exactly the two mechanisms found
numerically, A^ρ (Treg-destruction) × σ_tol/(σ_tol+(1−ρ)σ_a3) (substrate-competition). (ii) An
order-inversion law: tolerance-first nets ~A^ρ·φ(1−T)E₀ Tregs versus anti-CD3-first’s
~A·φ(1−T)E₀, so tolerance-first is optimal ⟺ A^ρ > A ⟺ ρ < 1, with crossover ρ* = 1. This
single inequality predicts the numerically-observed two-channel inversion: the ODE benefit of
going tolerance-first over simultaneous flips sign as ρ passes 1 (+41 points at ρ=0.9 → −6 points at
ρ=1.1). The criterion explains the whole result — why co-dosing antagonizes (𝒜<1), why tolerance-
first is best when anti-CD3 spares Tregs (ρ<1), and exactly where that advice reverses (ρ>1).
To our knowledge this is the first mechanistic within-host model of the anti-CD3 × antigen-specific-tolerance interaction, and the first to state a sequencing rule for it. It does not claim to discover the antagonism — the phenomenon was anticipated empirically (Stewart 2020; Foster 2025) — but to explain it and convert it into an actionable prediction. The bistable Treg/effector motif itself is established (Alexander & Wahl 2011); a sequencing-matters precedent exists in cancer immunotherapy with different drug classes (Messenheimer et al. 2017); neither models this combination or states this rule. The mechanism is intuitive: antigen-specific tolerance and anti-CD3 are not interchangeable “more immunosuppression” — tolerance needs the effector substrate that anti-CD3 removes, so the two cooperate only when tolerance acts first.
Falsifiable prediction. Combining the two therapies should be done tolerance-first, not simultaneously or anti-CD3-first; if anti-CD3 must precede, a sufficient inter-drug gap is required. Direct test: a NOD (or pre-clinical) sequencing experiment comparing (a) tolerance-first → anti-CD3, (b) simultaneous, (c) anti-CD3-first → tolerance, measuring diabetes incidence / banked C-peptide; the model predicts (a) > (b) > (c). A second prediction: if anti-CD3 is found to strongly deplete Tregs in vivo, the optimal order inverts toward anti-CD3-first — so the experiment should also measure Treg dynamics.
Limitations. This is an illustrative within-host model. (i) The antagonism magnitude is conditional on the Hill cooperativity n=2; at n≥3 the cross-repression sharpens and the antagonism can reverse — the robustness sweep did not vary n, so the result should be read for the moderate-cooperativity switch. (ii) “Tolerance-first = 100% protection” is specific to the 5-year evaluation horizon (it remains the best arm, but is not exactly 100%, at 6–10 yr). (iii) The model under-produces the teplizumab monotherapy delay magnitude (+0.68 yr vs TN10’s ~+2 yr) — the robust contribution is the sequencing antagonism, not the monotherapy-delay magnitude. (iv) Outcomes are cohort fractions because the underlying switch is bistable. (v) Parameters are illustrative beyond the calibrated progression timing, and the Foster result is a conference abstract. A self-audit (in the repository) added caveats (i)–(iii) and is the reason the framing is “a sequencing rule for a known antagonism,” not a discovery.
Figure 1. The mechanism — only tolerance-first builds a self-stabilizing Treg pool. Timecourses
of β-cell mass (B), effectors (E) and Tregs (R) for the untreated, simultaneous, anti-CD3→tolerance,
and tolerance→anti-CD3 arms. Tolerance-first is the only schedule in which the regulatory pool R is
established and held, flipping the switch to the tolerant basin.
analysis/t1d_mechanism.png
Figure 2. anti-CD3 antagonises antigen-specific tolerance, and order rescues it. Cohort
durable-control fraction (β-cell mass > 0.45 at 5 yr) by arm: tolerance-only 100%, simultaneous 59%,
anti-CD3-first 41%, tolerance-first 100%. 
analysis/t1d_cohort.png
Figure 3. The falsifiable prediction — an optimal inter-drug interval. Durable-control fraction
versus the inter-drug gap: tolerance-first is flat at ~97–100% for any gap, whereas anti-CD3-first
recovers monotonically (59%→100%) only as the gap grows to ~1.5 yr. 
analysis/t1d_gap.png
Figure 4. Calibration to the untreated progression curve. Stage-2 cohort (heterogeneous in
effector severity and residual β-cell mass): untreated median time-to-diagnosis 2.06 yr (TN10
placebo ~2.0 yr), ~45% progressed by 2 yr (TrialNet stage-2 ~50%). 
analysis/t1d_calibration.png
Figure 5. The derived order-inversion law. The closed-form A^ρ − A (whose sign decides the
optimal order) and the ODE benefit of tolerance-first over simultaneous dosing, both crossing zero
at the predicted crossover ρ* = 1. 
analysis/t1d_analytic.png
All code, figures, the literature corpus, the verification harness, and the audit trail are openly
available: https://github.com/sethc555/type1-diabetes-research (to be archived at Zenodo with a
citable DOI). Running cd analysis && python3 verify_claims.py re-derives every headline number
(24/24 checks pass).
Status: illustrative within-host modeling / hypothesis generation; not validated experimental or clinical findings, not medical advice. AI assistance: this work was developed with substantial help from an AI coding/analysis assistant (Anthropic Claude) for implementation, derivation, audit, and drafting, under the author’s direction; AI tools are not authors. Competing interests: none. Funding: none.
analysis/NOVELTY.md)