Skip to main content
CPT: Pharmacometrics & Systems Pharmacology logoLink to CPT: Pharmacometrics & Systems Pharmacology
. 2026 Nov 13;15(11):e70376. doi: 10.1002/psp4.70376

A PK‐Driven Quantitative Systems Pharmacology Model Predicts Cytokine Release Syndrome Severity Across T Cell‐Activating Therapies via a Locked Cytokine Amplification Network

Hajar Besbassi 1,2,3,✉
PMCID: PMC13762425  PMID: 43023728

ABSTRACT

Cytokine release syndrome (CRS) is a major dose‐limiting toxicity of T cell‐engaging immunotherapies. Existing CRS models are drug‐class‐specific; here we show that a single locked cytokine network captures CRS severity across mechanistically distinct drug classes, changing only drug‐specific upstream activation parameters. We developed a PK‐driven quantitative systems pharmacology (QSP) model linking drug exposure, T cell activation dynamics, and a macrophage‐amplified cytokine network to clinical CRS severity. The 17‐parameter downstream amplification network with macrophage‐gated STAT3 positive feedback was calibrated on blinatumomab, structurally refined using TGN1412 as a transparently disclosed development case, then locked and tested blind on OKT3. The locked network reproduced the clinically observed CRS severity ordering across all three antibody‐based T cell engager classes (bispecific, CD28 superagonist, and anti‐CD3 with activation‐induced cell death) without re‐fitting any shared parameter. The OKT3 blind prediction passed eight qualitative plausibility checks, and its peak TNF‐α fell within values measured in patients. Tocilizumab rescue simulation reproduced five clinically validated phenomena. A mechanistic parameter swap reversing the T cell exhaustion rate between OKT3 and TGN1412 reversed CRS severity in the expected direction. Local robustness analysis (926 of 5000 parameter sets accepted, 18.5%) and a 2D stability map over the two threshold‐setting parameters (0 of 900 wrong‐order combinations) confirmed that the cross‐drug severity ordering is a property of a feasible parameter region rather than a single tuned point. The locked model predicted three distinct dose–response shapes without re‐fitting. Findings should be interpreted as a mechanistic proof‐of‐concept; prospective clinical validation remains pending.

Keywords: cytokine release syndrome, model‐informed drug development, Pharmacometrics, quantitative systems pharmacology, T cell engagers

Study Highlights

  • What is the current knowledge on the topic?
    • ○
      Cytokine release syndrome (CRS) is a major dose‐limiting toxicity of T cell‐activating therapies. Existing models are largely drug‐specific and do not explain why agents with distinct mechanisms produce markedly different CRS severities.
  • What question did this study address?
    • ○
      Can a single locked downstream cytokine‐amplification network, changing only drug‐specific upstream T cell activation parameters, reproduce CRS severity across mechanistically distinct antibody‐based T cell‐activating therapies?
  • What does this study add to our knowledge?
    • ○
      A PK‐driven QSP model with a locked macrophage‐amplified cytokine network reproduces Grade 1–3 CRS across blinatumomab, TGN1412, and OKT3 without re‐fitting shared parameters. OKT3 was a blind, held‐out prediction, passing 8/8 qualitative checks, with peak TNF‐α inside the measured clinical range and peak IL‐6 inside its moderate‐CRS band; across 40 OKT3‐excluded calibrations, 27 of 30 feasible fits recovered OKT3 within this band. It also reproduces five tocilizumab rescue phenomena.
  • How might this change drug discovery, development, and/or therapeutics?
    • ○
      The architecture supports prospective CRS risk stratification for T cell‐activating agents from pharmacokinetics and upstream activation parameters alone, enabling model‐informed dose selection and earlier safety de‐risking.

1. Introduction

Cytokine release syndrome (CRS) is among the most clinically significant toxicities of T cell‐engaging immunotherapies. Bispecific T cell engagers such as blinatumomab cause CRS that is mostly low grade, with CRS of Grade 3 or higher in 4.9% of patients in a Phase 3 trial [1], while next‐generation bispecifics targeting BCMA (teclistamab, talquetamab) and CD20 (mosunetuzumab, glofitamab) extend the class with similar safety challenges. CAR‐T cell therapies can trigger severe CRS requiring intensive supportive care [2], and the management of CAR‐T‐associated CRS with tocilizumab and corticosteroids has become standard practice across the six FDA‐approved CAR‐T products. The TGN1412 CD28 superagonist trial in 2006 caused life‐threatening CRS in all six healthy volunteers within hours of first dosing [3], an event that fundamentally changed first‐in‐human trial design and remains the canonical example of catastrophic CRS in modern drug development. OKT3/muromonab‐CD3, a murine anti‐CD3 antibody historically used in transplant rejection, causes a systemic reaction (fever, chills, headache, gastrointestinal symptoms) that is invariably associated with the first injection and is characteristically self‐limiting, with cytokine release ceasing once circulating T cells have disappeared [4, 5]; here this self‐limitation is represented as activation‐induced cell death (AICD). Despite sharing a common downstream pathway, namely monocyte‐derived IL‐6 production [6], these agents produce CRS of vastly different clinical severity.

CRS now represents a primary dose‐limiting toxicity for the rapidly expanding class of T cell‐engaging immunotherapies. Severe CRS (Grade 3–4) requires intensive care management, prolongs hospitalization, and drives the dose adjustments and step‐dosing protocols used across this drug class. The disparity between the relatively predictable clinical CRS profile of approved bispecifics and the catastrophic outcome of the TGN1412 first‐in‐human trial points to a translational gap: existing frameworks cannot reliably distinguish between fundamentally different drug mechanisms before dosing humans. Empirical PK/PD models trained on existing clinical data extrapolate poorly to mechanistically novel agents, motivating mechanistic frameworks that encode the biology of T cell activation and downstream cytokine amplification rather than fitting drug‐specific dose–response curves.

Existing CRS models are drug‐class‐specific. Published models for CAR‐T CRS [2] and bispecific antibody CRS [7] capture the cytokine cascade for individual agents but do not address whether the same cytokine network architecture can explain severity differences across drug classes. QSP models of the tumor‐immune interface [8] focus on efficacy rather than safety. To our knowledge, no published model has demonstrated that a single locked mechanistic cytokine network, without changing any shared downstream parameters, can predict CRS severity for mechanistically distinct drugs.

The central challenge is that CRS severity depends on the interaction between drug‐specific T cell activation dynamics and shared downstream cytokine amplification. Blinatumomab activates a subset of T cells via CD3 engagement that requires CD19+ target cells, producing localized and limited activation. TGN1412 causes polyclonal T cell activation via CD28 superagonism, driving sustained expansion without the natural brake of AICD. OKT3 cross‐links CD3 to cause polyclonal activation followed by AICD, which limits T cell expansion and thereby caps the cytokine cascade. These three mechanisms should produce qualitatively different CRS severities through the same downstream IL‐6 amplification cascade, but this has not been demonstrated computationally with a single locked downstream network.

Here, a PK‐driven mechanistic QSP model demonstrates that a single locked downstream amplification network captures CRS severity differences across three mechanistically distinct antibody‐based T cell‐activating therapies, reproduces tocilizumab rescue, and produces three qualitatively distinct dose–response shapes (gradual, plateau, and switch‐like) without re‐fitting any shared parameter. The model was developed iteratively: calibrated on blinatumomab, structurally refined using TGN1412 as a transparently disclosed model development case (not a blind prediction), and then locked and tested on two independent held‐out challenges, namely a blind OKT3 CRS severity prediction and an independent tocilizumab‐rescue intervention validation. If a shared downstream cytokine network architecture is sufficient to explain CRS severity differences across mechanistically distinct drug classes, with only drug‐specific upstream T cell parameters changing, then prospective CRS risk assessment for novel agents could, in principle, be reduced to characterizing the drug's preclinical T cell activation phenotype, a parameter set routinely measured in early development.

2. Methods

2.1. Model Overview and Ethics

The CRS model was implemented in Python and comprised three coupled subsystems: drug pharmacokinetics, T cell state dynamics, and a five‐cytokine network, giving 10 dynamical state variables (Figure 1). All ODE systems were solved using the LSODA adaptive‐step integrator with relative tolerance 10−8 and absolute tolerance 10−10. Full equations for drug PK and IFN‐γ‐driven macrophage recruitment, the computational environment, and the reproducibility checks are provided in the Supplementary Methods.

FIGURE 1.

FIGURE 1

CRS model architecture. Drug concentration activates resting T cells via a Hill function. T cells progress through four states: Resting, activated (with proliferation, kprolif), cytokine‐producing (kdiff), and exhausted (kexhaust). Cytokine‐producing T cells release IFN‐γ and TNF‐α (kprod), which activate the macrophage pool. Activated macrophages produce IL‐6, IL‐1β, and anti‐inflammatory IL‐10. IL‐6 drives a STAT3 positive feedback loop (kfb, red arrow) gated by macrophage activation. IL‐10 exerts anti‐inflammatory negative feedback (flat‐head arrows), suppressing pro‐inflammatory cytokine production. All cytokines degrade at first‐order rates (kdeg). The 17 shared cytokine network parameters are locked after calibration on blinatumomab; only T cell activation parameters differ between drugs.

This study used only previously published, aggregate clinical data drawn from the cited sources; it involved no human participants, no identifiable human data, and no animals, and therefore required no institutional review board or animal care committee approval.

2.2. T Cell Activation and the Cytokine Network

Four T cell states were modeled with first‐order transitions: resting (Trest,0=1000 cells/μL), activated, cytokine‐producing, and exhausted. Drug‐driven activation followed a Hill function:

fact=CdrugnHEC50,actnH+CdrugnH (1)

Activated T cells proliferated (rate kprolif), differentiated into cytokine‐producing cells (kdiff), and exhausted (kexhaust). The exhaustion rate kexhaust was an upstream, drug‐specific mechanistic differentiator between drug classes, distinct from the shared downstream cytokine‐network parameters examined in the sensitivity analysis: high kexhaust represented activation‐induced cell death (AICD, OKT3), while low kexhaust represented sustained expansion (TGN1412). Drug‐specific values and their sources are given in Table 1; the OKT3 value represents the rapid loss of circulating CD3+ T cells after the first OKT3 injection (from 1370 to 15 cells/mm3 by the second injection; Abramowicz [5]), modeled as AICD.

TABLE 1.

Drug‐specific parameters for the three CRS drug classes.

Category Parameter Blinatumomab TGN1412 OKT3 Source
PK MW (Da) 54,100 148,000 150,000 a , d , e
Vcentral (L) 3.5 3.5 3.5 a , d , e
CL (L/h) 2.1 0.005 0.02 a , d , e
T cell EC50,act (nM) 1.0 1.0 2.0 b , f
nH (Hill) 1.5 3.0 1.5 c
kact (h−1) 0.08 0.50 0.30 f , g
kprolif (h−1) 0.03 0.10 0.03 g , h
kexhaust (h−1) 0.005 0.012 0.08 c , i
kdiff (h−1) 0.03 0.08 0.06 f , g
Tact,max (cells/μL) 5000 20,000 5000 j
Clinical dose 28 μg/day 0.1 mg/kg 5 mg Approved/trial
Route cIV k IV bolus IV bolus

Note: The 17 shared cytokine network parameters (Table S9) are identical across all drugs.

a

Blinatumomab PK. Clements [9] popPK: CL = 2.22 L/h, Vc = 5.98 L. Present CL = 2.1 L/h is within reported range; Vcentral = 3.5 L reflects a one‐compartment approximation.

b

In vivo EC50 for T cell activation in the nM range (Klinger [10]).

c

TGN1412 Hill exponent nH=3.0 reflects cooperative CD28 superagonist signaling (Stebbings [11]).

d

TGN1412 PK: no human PK published; CL = 0.005 L/h is consistent with typical IgG4 mAb clearance 0.003–0.013 L/h (Dirks [12]).

e

OKT3 PK: t1/2≈18 h (Hooks [13]). Present CL = 0.02 L/h is a known simplification; OKT3 CRS in the model is driven by AICD kinetics.

f

T cell activation and differentiation rates from QSP T cell engager models (Ma [14]).

g

Subsequent division time ≈9 h corresponds to kprolif≈0.077 h−1; first division 30–50 h (Hawkins [15]).

h

TGN1412 kprolif=0.10 h−1 (doubling time ∼7 h) reflects superagonist‐driven sustained expansion.

i

OKT3 kexhaust=0.08 h−1 represents AICD, reflecting the rapid loss of circulating CD3+ T cells after OKT3 (Abramowicz [5]); TGN1412 kexhaust=0.012 h−1 reflects sustained expansion.

j

Tact,max is a model‐assumed upper bound; normal lymphocyte range is 1000–4800 cells/μL. Values of 5000–20,000 cells/μL were assumed to reflect polyclonal activation capacity.

k

Blinatumomab is administered clinically as a continuous IV infusion over 28 days. All simulations reported here use a 7‐day infusion and read the cytokine profile over that window; the bolus agents are read over 72 h.

Cytokine‐producing T cells produced IFN‐γ and TNF‐α, which activated a macrophage pool through a saturating function (EC50,mac=48.6 pg/mL, one of the five calibrated parameters); activated macrophages produced IL‐6, IL‐1β, and anti‐inflammatory IL‐10. The central mechanistic feature was macrophage‐gated IL‐6 STAT3 positive feedback, in which the IL‐6 amplification rate was gated by the macrophage activation fraction fmac:

Jfeedback=kfb·maxIL6−IL6baseline0·IL6Km,fb+IL6·fmac (2)

with calibrated values kfb=0.528 h−1 and Km,fb=50 pg/mL (Table S9). The macrophage gating made the effective feedback rate drug‐dependent. At the simulated IL‐6 peak the macrophage activation fraction was 0.60 for blinatumomab, 0.87 for OKT3 and 0.98 for TGN1412, giving effective feedback rates of 0.32, 0.46 and 0.52 h−1 against an IL‐6 degradation rate of kdeg,IL6=0.519 h−1. The three agents therefore occupy a graded sequence of operating points approaching the amplification threshold: blinatumomab well below it with stable IL‐6 dynamics, OKT3 approaching it but capped by activation‐induced cell death, and TGN1412 at it, in the near‐threshold nonlinear amplification regime. IFN‐γ/TNF‐α‐driven macrophage recruitment (kmac,recruit=5.0) further amplified the effective macrophage pool (approximately 4.5‐fold for TGN1412), and IL‐10 provided anti‐inflammatory negative feedback (IC50,IL10=50 pg/mL). This macrophage‐gated formulation follows the finding that monocyte‐derived IL‐6 is the primary driver of CRS severity [6] and the abnormal macrophage activation reported in severe blinatumomab CRS [16].

2.3. CRS Grading

Because the model produces cytokine concentrations rather than clinical symptoms, a biomarker‐based surrogate grading was used. Peak IL‐6 thresholds were set for this study as surrogates for the clinical CRS grades [17]: Grade 1, IL‐6 >30 pg/mL; Grade 2, IL‐6 >200 pg/mL; Grade 3, IL‐6 >1000 pg/mL; Grade 4, IL‐6 >5000 pg/mL. These thresholds are pragmatic biomarker correlates, not clinical grades, and are not derived through formal optimization against grade‐labeled individual‐patient data.

2.4. Locked Versus Drug‐Specific Parameters and Calibration

The critical design principle was the separation of 17 locked shared cytokine network parameters (identical across all drugs; Table S9) from drug‐specific PK and T cell activation parameters (Table 1). The shared network was calibrated by constrained Nelder–Mead optimization against a range‐based objective comprising two soft penalties of identical form. Each penalty is zero when the simulated peak IL‐6 lies within the corresponding reference range and otherwise increases with the squared log‐fold distance to the nearest boundary; simulations exhibiting runaway amplification are rejected. Blinatumomab entered as the primary calibration agent, using an author‐defined mild‐CRS target range of 50–200 pg/mL, whereas TGN1412 entered as a transparently disclosed development case, using its published range 3000–100,000 pg/mL. Neither term imposes a point target. OKT3 was excluded entirely from the calibration objective and evaluated only after the network was locked. Because the OKT3 clinical studies did not measure IL‐6, OKT3 peak IL‐6 was compared with an author‐defined moderate‐CRS band of 200–2000 pg/mL, which lies between the blinatumomab and TGN1412 ranges. Only the five most influential cytokine network parameters (kIL6,fb, macpool, kprod,IL6,mac, kdeg,IL6, EC50,mac), identified by a prior one‐at‐a‐time sensitivity analysis, were optimized; the other 12 were held fixed. The macrophage‐activation EC50,mac is thus one of the five calibrated parameters, whereas the T cell activation EC50,act is taken from Klinger [10]. The five free parameters were initialized by manual adjustment so that simulated blinatumomab peak IL‐6 fell inside this target range; the 12 fixed parameters were set from the literature sources noted in the Table S9 footnotes. Because the range‐based objective is exactly zero anywhere inside the feasible region, and the initial values already satisfied it, the constrained optimization returned them unchanged. The calibration therefore establishes that the reported set lies in the feasible region rather than locating a unique optimum. The full objective‐function definition and the formal calibration procedure are detailed in the Supplementary Methods. Development and validation data are distinct: the shared network was developed using the blinatumomab IL‐6 target range and, for structural refinement, TGN1412 IL‐6 (Suntharalingam [3]), whereas OKT3 (Chatenoud [4]; Abramowicz [5]; Chatenoud [18]) was held out for blind validation and tocilizumab for independent intervention validation. Robustness was assessed by four analyses. In the one‐at‐a‐time sensitivity analysis each shared cytokine network parameter was perturbed by ±50% with all others held at their locked values, and the normalized sensitivity coefficient S=ΔY/Y0/Δp/p0 was computed with peak IL‐6 as the output; kIL6,fb was perturbed by ±20% instead, because larger perturbations drive the system into IL‐6 runaway (Figure S4). When one perturbation arm entered the runaway regime, that arm was reported as a regime transition and the finite one‐sided coefficient reported separately. One‐at‐a‐time perturbation does not capture parameter interactions. The remaining three analyses were a local accepted parameter ensemble (uniform ±30% prior on the five free parameters, 5000 samples, accepted when all antibody‐based drugs fall within their IL‐6 reference ranges); a 30×30 stability map over kIL6,fb and kdeg,IL6; and a profile‐likelihood analysis of the two threshold‐setting parameters. All bolus‐agent simulations (TGN1412 and OKT3) use a 72‐h readout window, and blinatumomab its 7‐day infusion.

2.5. Validation and Analysis

Following calibration and structural refinement on TGN1412, the 17 cytokine network parameters were locked and tested without changes on OKT3 (blind), tocilizumab rescue, dose–response sweeps, and a mechanistic parameter swap. Tocilizumab (anti‐IL‐6R) rescue was simulated by setting STAT3 feedback to zero and reducing IL‐6 degradation to 30% of normal at 13 intervention timepoints (0, 1, 2, 3, 4, 6, 8, 12, 18, 24, 36, 48, and 72 h), with all other equations unchanged. The parameter swap exchanged kexhaust between OKT3 (0.08 h−1) and TGN1412 (0.012 h−1). Illustrative virtual populations (N=100 per drug) were generated by sampling log‐normal inter‐individual variability (CV = 25%–30%) on 5 T cell activation parameters, macrophage pool size, and body weight; no acceptance criteria were applied, so population results reflect sensitivity to parameter variability rather than calibrated predictions.

3. Results

3.1. Cross‐Drug CRS Validation

As summarized in Figure 2, the locked network reproduced the clinically observed CRS severity ordering across the three antibody‐based T cell engagers using identical shared cytokine network parameters and only drug‐specific PK and T cell activation parameters (Table 1, Figure 3). Predicted CRS grades increased from blinatumomab (Grade 1, peak IL‐6 ∼100 pg/mL) through OKT3 (Grade 2, peak IL‐6 671 pg/mL) to TGN1412 (Grade 3, peak IL‐6 4349 pg/mL), a ∼43‐fold span in peak IL‐6, with each prediction within the IL‐6 reference range for the corresponding agent (Table S1). The severity ordering therefore emerged from differences in upstream PK and T cell activation alone, propagating through one shared downstream amplification network.

FIGURE 2.

FIGURE 2

Conceptual overview of the cross‐drug CRS mechanism. Left: drug‐specific upstream T cell activation by blinatumomab (CD19 × CD3 bispecific), OKT3 (anti‐CD3, self‐limited by activation‐induced cell death), and TGN1412 (CD28 superagonist, without the AICD brake). Centre: the shared downstream amplification network. Activated T cells produce IFN‐γ and TNF‐α, which activate macrophages; macrophages produce IL‐6, which reinforces its own production through macrophage‐gated STAT3 positive feedback. Whether the system remains controlled or enters strong amplification is set by the effective feedback rate (kIL6,fb·fmac) versus the IL‐6 clearance rate (kdeg,IL6). Bottom: the same locked network produces three operating points, blinatumomab below threshold (mild), OKT3 near threshold but capped by AICD (moderate), and TGN1412 in the strong‐amplification regime (severe).

FIGURE 3.

FIGURE 3

CRS cytokine dynamics across three drug classes from one locked model. IL‐6 (solid), TNF‐α (dashed), and IFN‐γ (dotted) time‐courses for blinatumomab (A, Grade 1), OKT3 (B, Grade 2), and TGN1412 (C, Grade 3). Note different y‐axis scales reflecting the ∼43‐fold IL‐6 difference between blinatumomab (100 pg/mL) and TGN1412 (4349 pg/mL). All three simulations used identical cytokine network parameters; only drug‐specific PK and T cell activation parameters differed (Table 1). Shaded regions denote the reference peak ranges for each cytokine (color‐matched to the corresponding trace); for OKT3, only the TNF‐α peak has measured clinical data in pg/mL (see Figure 4). All simulated peaks fall within their reference ranges except TGN1412 TNF‐α (558 pg/mL, below the 1000–3500 pg/mL range), the single structural underprediction discussed in Section 4.

TGN1412 was a model development case rather than a blind prediction: the macrophage‐gated STAT3 feedback structure was introduced specifically to address the initial model's failure to reproduce the TGN1412 cytokine profile (Supplementary Methods). The refined model predicted Grade 3 CRS for TGN1412 (peak IL‐6 = 4349 pg/mL, within the clinical 3000–100,000 pg/mL range from Suntharalingam [3]), and an illustrative population simulation predicted all virtual subjects in Grade 3 (67%) or Grade 4 (33%), consistent with the 6/6 clinical Grade 3–4 incidence. Two known discrepancies are noted: peak TNF‐α was underestimated (558 vs. clinical 1000–3500 pg/mL) and CRS onset was delayed (Table S2).

3.2. OKT3 Blind Prediction

OKT3/muromonab‐CD3 represents a fundamentally different T cell activation mechanism: anti‐CD3 cross‐linking causes polyclonal activation followed by AICD, captured in the model by high kexhaust=0.08 h−1 and low Tact,max=5000 cells/μL. The locked model predicted Grade 2 CRS for OKT3 at 5 mg IV (peak IL‐6 = 671 pg/mL). It passed 8/8 qualitative clinical plausibility checks (correct CRS grade, cytokine peaks above clinically observed thresholds, onset within 6 h, and severity ordering relative to TGN1412; Chatenoud [4], Abramowicz [5], Chatenoud [18]). Quantitatively, predicted peak TNF‐α (174 pg/mL) lies within the range of individual first‐dose peaks measured in 35 OKT3‐treated patients (90–3300 pg/mL [18]), at the low end of observed values, and predicted peak IL‐6 lies within the author‐defined moderate‐CRS band; IFN‐γ was reported only in activity units and could not be compared (Tables S3 and S8). The mechanistic basis was AICD: cytokine‐producing T cells peaked at 334 cells/μL at 21 h then declined to 183 cells/μL by 48 h, whereas TGN1412 rose continuously to 2790 cells/μL at 48 h, producing a ∼6.5‐fold IL‐6 difference between the two anti‐CD3 mechanisms (Figure S1). An illustrative population simulation predicted ∼96% of virtual patients in Grade 2 with ∼4% Grade 3 (Figure S2). Comparing the three virtual populations with the corresponding clinical reports: for OKT3 the model predicted CRS in 100% of virtual recipients, 96% of them Grade 2, against a systemic first‐dose reaction reported in almost all recipients [4, 5, 18]; for TGN1412 it predicted Grade 3–4 CRS in 100% of virtual patients (67% Grade 3, 33% Grade 4), consistent with severe reactions in all six clinical recipients [3]; and for blinatumomab it predicted CRS in 98% of virtual patients, predominantly Grade 1 and none of Grade 3 or higher, whereas CRS of Grade 3 or higher occurred in 4.9% of patients in the Phase 3 trial [1]. The first two are directionally consistent with the clinical observations. The blinatumomab discrepancy likely reflects, at least in part, the absence of interindividual variation in CD19‐positive target‐cell burden, a plausible determinant of blinatumomab CRS risk, from the present virtual‐population design, which samples T cell activation parameters, macrophage pool size and body weight, but not target‐cell burden. These comparisons should be interpreted cautiously, because the model‐derived cytokine severity categories are not identical to clinically adjudicated CRS outcomes. Critically, OKT3 was excluded from the calibration objective and from all parameter‐selection steps: the locked network, developed on blinatumomab and TGN1412 alone, predicts OKT3 peak IL‐6 = 671 pg/mL (Grade 2), TNF‐α = 174 pg/mL (within the measured clinical range), and IFN‐γ = 207 pg/mL. This held‐out prediction is robust by two independent tests. First, sampling the five free parameters (±30%) and accepting only sets in which blinatumomab and TGN1412 fall in their reference ranges (OKT3 excluded from acceptance), OKT3 lands within its moderate‐CRS IL‐6 band in 91.0%–92.3% of accepted sets across three independent seeds. Second, an OKT3‐excluded multi‐start calibration from 40 independently perturbed starting points reached the blinatumomab‐plus‐TGN1412 feasible region in 30 runs, of which 27 (90%) predicted OKT3 within this band (median 893, 5th–95th percentile 553–2283 pg/mL). OKT3 is therefore a blind, held‐out cross‐drug validation, recovered by a procedure that never uses OKT3 (Figure 4; Supplementary Methods).

FIGURE 4.

FIGURE 4

OKT3 is a blind, held‐out cross‐drug validation. OKT3 was excluded from the calibration objective and all parameter‐selection steps; the locked network was developed on blinatumomab and TGN1412 alone. (A) Predicted OKT3 cytokine time‐courses (IL‐6 solid, TNF‐α dashed, IFN‐γ dotted); shaded bands are comparison ranges, color‐matched to each trace. The peak‐and‐decline shape reflects activation‐induced cell death. (B) Predicted peak concentration (marker) versus clinical range (shaded bar) for each cytokine on a log scale: IL‐6 = 671 pg/mL (author‐defined moderate‐CRS band 200–2000, between the blinatumomab and TGN1412 ranges; IL‐6 was not measured in the OKT3 studies), TNF‐α = 174 pg/mL (band 100–800, inside the measured first‐dose range of 90–3300 pg/mL [18]), IFN‐γ = 207 pg/mL (author‐specified band 50–500; the OKT3 studies reported IFN‐γ only in activity units [4, 5]).

3.3. Tocilizumab Rescue

Without any change to the shared cytokine network, the locked model reproduced five clinically validated phenomena for tocilizumab rescue of TGN1412‐severity CRS (Figure 5; Table S5): breaking of the IL‐6 amplification loop, a paradoxical rise in raw IL‐6 under IL‐6R blockade [19], complete insensitivity of TNF‐α and IFN‐γ to tocilizumab, insufficiency of tocilizumab alone for TGN1412‐severity CRS [20], and a model‐predicted intervention‐timing threshold (∼24 h in the simulation to prevent a Grade 3 outcome). Post‐tocilizumab, raw IL‐6 rose (receptor‐mediated clearance blocked) while IL‐6 signaling fell to zero (STAT3 feedback eliminated), and TNF‐α (558 pg/mL) and IFN‐γ (1674 pg/mL) trajectories were identical with and without treatment, because T cell activation and IFN‐γ/TNF‐α production are IL‐6‐independent. This structural prediction matches ASTCT guidelines [21], which recommend adding corticosteroids for Grade ≥ 3 CRS.

FIGURE 5.

FIGURE 5

Tocilizumab rescue of TGN1412‐severity CRS. (A) IL‐6 time‐course: untreated (red), tocilizumab at 6 h (orange), tocilizumab at 24 h (green dashed). The IL‐6 amplification loop is broken by IL‐6R blockade. (B) TNF‐α and (C) IFN‐γ are completely unaffected by tocilizumab; trajectories are identical regardless of treatment, reflecting the IL‐6‐independent nature of T cell activation and macrophage inflammatory signaling. This emergent structural prediction matches ASTCT clinical guidelines: Tocilizumab alone is insufficient for severe CRS. Gray bands denote the published clinical peak ranges; IFN‐γ reaches its clinical range, while modeled TNF‐α (558 pg/mL) lies below the reported TGN1412 range.

3.4. Parameter Swap Test

Swapping the single exhaustion‐rate parameter kexhaust between OKT3 and TGN1412 reversed CRS severity in the expected direction (Figure 6). Removing AICD from OKT3 (kexhaust:0.08→0.012) increased severity from Grade 2 to Grade 3 (IL‐6: 671 → 1554 pg/mL); forcing AICD on TGN1412 (kexhaust:0.012→0.08) dropped peak IL‐6 by 46% (4349 → 2358 pg/mL, although both TGN1412 conditions remain Grade 3 under the locked parameterization). This mechanistically meaningful upstream parameter controlled whether a drug sat in the low‐to‐moderate or high‐severity regime, supporting a real biological distinction (AICD vs. sustained expansion) rather than arbitrary parameter tuning (Table S4).

FIGURE 6.

FIGURE 6

Parameter swap alters CRS severity in the expected direction. (A) Peak IL‐6 under four conditions: OKT3 with real AICD (kexhaust=0.08, Grade 2, 671 pg/mL), OKT3 without AICD (kexhaust=0.012, Grade 3, 1554 pg/mL), TGN1412 with real expansion (kexhaust=0.012, Grade 3, 4349 pg/mL), TGN1412 with forced AICD (kexhaust=0.08, Grade 3, 2358 pg/mL). Hatched bars indicate swapped parameters. (B) The underlying mechanism, shown directly from the model's cytokine‐producing T cell trajectories: under AICD (OKT3) the cytokine‐producing T cell pool peaks near 21 h and then declines, capping cytokine output, whereas sustained expansion (TGN1412) drives a continuously rising pool and runaway IL‐6. The swap moves OKT3 from Grade 2 to Grade 3 and drops TGN1412 peak IL‐6 by 46%.

3.5. Dose–Response Analysis

The same locked network produced three qualitatively distinct dose–response shapes, with no dose‐specific tuning (Figure 7). Blinatumomab was gradual: peak IL‐6 rose from 71 to 226 pg/mL across 0.001–0.1 mg, with Grade 1 predicted in 9 of 10 dose levels. TGN1412 was switch‐like: Grade 1 at doses ≤0.00066 mg/kg, jumping to Grade 3 (peak IL‐6 ∼1200 pg/mL) at 0.0017 mg/kg, reflecting cooperative CD28 signaling (nH=3.0) and near‐threshold STAT3 amplification. OKT3 was a plateau: peak IL‐6 ranged from 600 to 671 pg/mL across a 40‐fold dose range (0.5–20 mg), all Grade 2, reflecting the AICD ceiling. For context, the clinically administered doses are 0.028 mg for blinatumomab, 0.1 mg/kg for TGN1412 in the first‐in‐human trial (against 0.0001 mg/kg for the redesigned Theralizumab), and 5 mg for OKT3; these are marked on Figure 7. Because clinical cytokine measurements are not available across these simulated dose ranges, the curves are prospective model predictions rather than independently validated dose–response relationships, and the dose markers locate the clinical doses on the predicted curves rather than validating them.

FIGURE 7.

FIGURE 7

Three dose–response shapes from one locked model. (A) Blinatumomab: Gradual IL‐6 increase (71–226 pg/mL) across three orders of magnitude in dose; Grade 1 across nine of 10 dose levels, with only the highest dose (0.1 mg) crossing marginally into Grade 2. (B) TGN1412: switch‐like transition on log scale; Grade 1 below ~0.001 mg/kg, Grade 3 above. (C) OKT3: plateau from 600 to 671 pg/mL across a 40‐fold dose range; AICD caps severity at Grade 2 regardless of dose. Horizontal dashed lines are CRS grade thresholds; colored dotted lines mark the clinically administered dose of each agent: the approved blinatumomab dose (0.028 mg), the Theralizumab safe dose (0.0001 mg/kg) and TGN1412 trial dose (0.1 mg/kg), and the OKT3 clinical dose (5 mg). All panels use the same locked cytokine network parameters.

3.6. Robustness and Identifiability

One‐at‐a‐time sensitivity analysis on the calibration agent identified the macrophage pool size (macpool, S=+2.25), the macrophage‐activation threshold (EC50,mac, S=−1.47) and the macrophage‐driven IL‐6 production rate (kprod,IL6,mac, S=+1.23) as the dominant parameters. The coefficients are drug‐dependent because the system operates near a bifurcation: for the near‐threshold agents the IL‐6 STAT3 feedback gain dominates instead (S=+16.3 for OKT3), and halving IL‐6 clearance to 0.26 h−1 places it below the effective feedback rate of all three agents, crossing the amplification condition kIL6,fb·fmac>kdeg,IL6 so that peak IL‐6 diverges. That perturbation is therefore a regime transition rather than a local sensitivity, and only the forward arm of kdeg,IL6 yields a coefficient (Figure S4). Four complementary analyses confirmed that the cross‐drug architecture is a property of a feasible parameter region rather than a single tuned point. A local ABC‐style accepted ensemble (uniform ±30% prior on the five free cytokine parameters) accepted 926 of 5000 samples (18.5%) as simultaneously matching all three IL‐6 reference ranges, and across the accepted ensemble blinatumomab was Grade 1 in 100%, OKT3 spanned Grade 2–3 (62.2% Grade 2, 37.8% Grade 3; median IL‐6 885 pg/mL) and TGN1412 spanned Grade 3–4 (26.7% Grade 3, 73.3% Grade 4; median IL‐6 8999 pg/mL), preserving the cross‐drug severity ordering (Table S6). This three‐drug ensemble includes OKT3 in its acceptance rule and therefore characterizes the jointly feasible parameter region; it is distinct from, and complementary to, the OKT3‐excluded held‐out validation reported above, which accepts on blinatumomab and TGN1412 only. A 30×30 stability map over kIL6,fb and kdeg,IL6 gave 157 of 900 all‐in‐range points (17.4%, closely matching the independent ensemble) and 0 of 900 with wrong cross‐drug ordering (Figure S3). A multi‐cytokine acceptance test (7 cytokine‐by‐drug reference‐range constraints; 5000 samples) found 0 sets satisfying all 7 simultaneously: 6 of the 7 constraints were satisfiable, the single systematic failure being TGN1412 TNF‐α, a documented structural limitation. Profile likelihood analysis of the two threshold‐setting parameters showed complementary asymmetric profiles, indicating that they are practically identifiable as a ratio (the IL‐6 feedback effective rate kIL6,fb·fmac relative to the clearance rate kdeg,IL6) rather than individually, consistent with the bifurcation interpretation. A consolidated summary of all model‐evaluation analyses is provided in Table S7.

4. Discussion

This work demonstrates that a single 17‐parameter cytokine network, calibrated on blinatumomab and structurally refined on TGN1412, retrospectively captures CRS severity for a third drug class (OKT3) and a therapeutic intervention (tocilizumab) without re‐fitting the shared downstream network. The model also produces three qualitatively distinct dose–response shapes from the same locked parameters, and a single‐parameter swap reverses the CRS severity ordering between drug classes.

The three‐drug result addresses a specific gap in the CRS modeling literature. Published CRS QSP models for CAR‐T [2, 22] and bispecific antibodies [7] are drug‐class‐specific; they capture cytokine dynamics for one agent but do not address whether the same network applies across mechanistically different drugs. Here, one locked network, with only drug‐specific T cell activation parameters varying, captures severity spanning Grade 1 (blinatumomab) through Grade 2 (OKT3) to Grade 3 (TGN1412), a ∼43‐fold span in peak IL‐6 (100 to 4349 pg/mL). The phrase “without re‐fitting” refers to the 17 shared downstream cytokine parameters, which remained locked across all three drug classes and the tocilizumab intervention. Drug‐specific upstream T cell parameters (6–7 per drug) were set independently, several with direct literature support (nH=3.0 for TGN1412 cooperative signaling [11], kexhaust for the rapid T cell loss after OKT3 [5], IgG4 clearance from Dirks [12]). The claim therefore rests on architectural transferability of the shared amplification network, and the parameter swap provides additional evidence that severity ordering follows from a mechanistic distinction (AICD vs. sustained expansion) rather than from specific numerical values.

Two arguments constrain whether the macrophage‐gated feedback is necessary. An ablated model in which the feedback is not gated by macrophage activation fails at every constitutive gain tested. At kfb=0.4 h−1 it predicts TGN1412 as Grade 2 (peak IL‐6 = 794 pg/mL); at kfb=0.55 h−1 all three agents reach Grade 4 and the cross‐drug severity ordering is disrupted. Across a sweep of 10 constitutive gains from 0.10 to 0.80 h−1, none placed all three agents inside their IL‐6 reference ranges. Within the tested model architecture, these results support the need for a drug‐dependent effective feedback rate. The macrophage gating and recruitment terms are not free mathematical devices but encode the Norelli finding that monocyte‐derived IL‐6 is the primary driver of CRS severity [6], and both feedback layers have independent biological support.

The tocilizumab simulation tested model structure rather than parameters. The prediction that TNF‐α and IFN‐γ are completely unaffected by IL‐6R blockade emerges because these pathways are IL‐6‐independent; only the IL‐6 autocrine STAT3 loop is disrupted. This matches ASTCT guidelines [21], under which Grade ≥ 3 CRS requires added corticosteroids precisely because non‐IL‐6 inflammation persists.

Several limitations apply. TGN1412 is a model development case, not a blind prediction. TGN1412's modeled IL‐6 does not reach a fixed peak within the simulated window; the 72‐h value lies within the clinical range. The model is not prospectively validated; all three drugs are well‐studied historical cases. The model under‐predicts TGN1412 peak TNF‐α (558 vs. clinical geometric midpoint 1871 pg/mL, AFE = 3.35, approximately 1.8‐fold below the lower clinical bound of 1000 pg/mL), the single failing cytokine constraint, indicating that the bispecific‐calibrated network's macrophage TNF‐α pathway should be made drug‐class‐aware: making this production rate drug‐specific (0.30 for TGN1412 vs. the locked 0.15) brings the predicted TNF‐α peak into the clinical range (1104 pg/mL) while preserving Grade 3 IL‐6, a natural one‐parameter extension we deliberately did not adopt here so that the cross‐drug result rests on a single fully locked network; because CRS grade is anchored to peak IL‐6, this does not affect the severity ordering. The virtual populations sample T cell activation parameters, macrophage pool size and body weight, but not CD19‐positive target‐cell burden, which is a documented determinant of blinatumomab CRS risk; adding that axis is the natural next step for population‐level work. Each drug has 6–7 partially free T cell parameters, mitigated but not eliminated by the literature anchoring, the parameter swap, and the emergent dose–response shapes; full global identifiability (a biologically informed prior and variance‐based Sobol indices) remains future work. CRS onset is predicted later than clinical (3–4 h vs. 60–90 min for TGN1412), which a two‐compartment PK refinement would improve without altering steady‐state severity.

This work positions CRS modeling within a model‐informed drug development framework. The recently issued ICH M15 guideline on model‐informed drug development [23] formally recognizes mechanistic QSP safety models as valid evidence for regulatory decision‐making, and the cross‐drug architectural transferability shown here, in which only drug‐specific T cell activation parameters need specification for novel agents, aligns with the ICH M15 vision of platform models supporting first‐in‐human dose selection. In summary, the central contribution is architectural transferability of a shared CRS amplification network within the antibody‐based T cell engager class, with drug‐specific T cell activation parameters as the only varying inputs, preserved across a local feasible parameter region characterized by four independent robustness analyses. This is an in silico mechanistic proof‐of‐concept supporting prospective CRS risk stratification from a drug's preclinical activation phenotype, pending the prospective validation and global identifiability assessment outlined above.

Author Contributions

H.B. wrote the manuscript, designed the research, performed the research, and analyzed the data.

Funding

The author has nothing to report.

Conflicts of Interest

The author declares no conflicts of interest.

Supporting information

Table S1: Three‐drug CRS comparison from one locked cytokine network.

Table S2: TGN1412 comparison with Suntharalingam 2006 clinical data.

Table S3: OKT3 blind prediction: 8/8 validation checks passed.

Table S4: Parameter swap test: exchanging k exhaust between OKT3 and TGN1412.

Table S5: Tocilizumab rescue: five structural predictions matching clinical evidence. Tocilizumab is an independent intervention validation, not a blind prediction.

Table S6: Accepted parameter ensemble summary (5000 samples, 926 accepted). Median and 5th–95th percentile values across the accepted parameter sets, compared with the calibrated point.

Table S7: Consolidated model‐evaluation evidence supporting the cross‐drug CRS transferability claim across the three antibody‐based drug classes.

Figure S1: T cell mechanism determines CRS severity. (A) Cytokine‐producing T cell (T cyt) dynamics for blinatumomab (blue), OKT3 (green), and TGN1412 (red). OKT3 shows peak‐and‐decline (AICD), while TGN1412 shows sustained expansion. (B) Resulting IL‐6 dynamics (log scale). T cell expansion drives macrophage activation and IL‐6 amplification through the shared cytokine network.

Table S8: Quantitative comparison of predicted vs. observed cytokine peaks for the CRS module. Two acceptance criteria are applied: (i) whether the predicted peak falls within the reference range (measured clinical or author‐defined; see footnotes), and (ii) absolute fold‐error (AFE) relative to the geometric midpoint of that range (AFE = max(pred/G, G/pred) where G=low×high), with AFE ≤ 2 as the acceptance threshold.

Table S9: Locked cytokine network parameters. These 17 parameters are identical across all three drug classes and across the tocilizumab rescue simulation.

Figure S2: Illustrative CRS grade distributions across virtual populations (N = 100 per drug). Inter‐individual variability was sampled from log‐normal distributions (CV = 25%–30%) on T cell activation parameters (k act, k prolif, k exhaust, EC50,act, T act,max), macrophage pool size, and body weight (see Methods for assumptions and limitations). Blinatumomab: predominantly Grade 1 (~85%) with ~13% Grade 2 and ~2% sub‐clinical. OKT3: ~96% Grade 2 with ~4% Grade 3, reflecting the AICD ceiling on CRS severity. TGN1412: split between Grade 3 (~67%) and Grade 4 (~33%), qualitatively consistent with clinical Grade 3–4 in all six Suntharalingam 2006 subjects. These distributions reflect sensitivity to assumed parameter variability rather than formally calibrated population predictions.

Figure S3: Cross‐drug stability map for the two threshold‐setting parameters. (A) Regime classification across a 30 × 30 grid of (k IL6,fb, k deg,IL6). Muted teal: all three antibody‐based drugs predict peak IL‐6 within their reference ranges. Light gray: feasible but at least one drug outside its reference range. Muted rose‐red: at least one drug undergoes IL‐6 amplification runaway. Medium gray would indicate wrong cross‐drug severity ordering (never observed in the data). Dashed line: theoretical bifurcation diagonal k IL6,fb = k deg,IL6 (the boundary under maximal macrophage activation). White star: calibrated parameter point. Of 900 grid points, 157 (17.4%) fall in the “all in range” region; 0 produced wrong cross‐drug severity ordering. (B) Predicted TGN1412 peak IL‐6 across the same grid (log scale, single‐hue blue sequential colourmap; darker blue indicates higher peak IL‐6), with dark‐gray contour lines marking the 3000 pg/mL Grade 3 floor, the 5000 pg/mL Grade 4 threshold, and the 100,000 pg/mL upper clinical bound.

Figure S4: Tornado plot of one‐at‐a‐time sensitivity coefficients for the CRS model. Each bar shows the normalized sensitivity S = (ΔY/Y 0)/(Δp / p 0) with peak IL‐6 as the output, for blinatumomab (the calibration agent), computed at ±50% parameter perturbation from baseline; *k fb uses ±20% because larger perturbations push the system past the amplification threshold. †For k deg,IL6 the plotted value is the forward (+50%) arm only, and is therefore not comparable to the symmetric coefficients: halving the IL‐6 clearance rate to 0.26 h−1 places it below the effective feedback rate of every agent (0.32, 0.46, and 0.52 h−1), crossing the amplification condition k IL6,fb ·f mac >k deg,IL6, so the STAT3 loop becomes self‐sustaining and peak IL‐6 diverges. That arm is a regime transition rather than a local sensitivity, and is marked “runaway” on the plot. Parameters with |S| > 1 amplify the input change; parameters with |S| < 1 attenuate it. Macrophage‐derived IL‐6 production (macpool, k prod,IL6,mac) and the macrophage‐activation threshold (EC50,mac) dominate for blinatumomab. The coefficients are drug‐dependent because the system operates near a bifurcation: for the near‐threshold agents the feedback gain dominates instead (S = +16.3 for OKT3; for TGN1412 the upward arm diverges and only the backward coefficient, +4.0, is defined). One‐at‐a‐time perturbation does not capture parameter interactions. Computed by sensitivity_oat.py.

Table S10: Fixed model constants shared across all drugs (not fitted).

Supplementary Methods. Full drug pharmacokinetic and IFN‐γ‐driven macrophage‐recruitment equations; CRS model state equations; the formal constrained calibration objective function and procedure; the local accepted parameter ensemble; the multi‐cytokine acceptance test; the profile‐likelihood analysis; the cross‐drug stability‐map procedure; and the computational environment and reproducibility checks.

PSP4-15-e70376-s001.pdf (811.7KB, pdf)

Acknowledgments

The author thanks the Vaccine & Infectious Disease Institute (VAXINFECTIO), the Centre for Health Economics Research and Modelling Infectious Diseases (CHERMID), and the Antwerp Centre for Translational Immunology and Virology (ACTIV) at the University of Antwerp for providing the academic environment in which this work was conducted. FigureLab, an AI‐based figure‐generation tool, was used to prepare the schematic Figures 1 and 2; it was not used for the study design, model development, data analysis, interpretation, references, or the text of this article. The author reviewed and verified both figures for scientific accuracy.

Data Availability Statement

All validation data analyzed in this study were extracted from previously published sources cited in the text. The complete model equations, the macrophage‐gated STAT3 feedback and macrophage‐recruitment terms, the constrained calibration objective, and the locked 17‐parameter cytokine network are provided in full in Section S1 and Table S9, and are sufficient to reimplement the model. A reproducible implementation of the CRS model (the simulation engine, the locked cytokine‐network parameters, the drug‐specific parameter sets, and the calibration, validation, and robustness scripts), together with the digitized clinical target datasets and the analysis result files needed to reproduce every reported number, is openly available at Zenodo (DOI: https://doi.org/10.5281/zenodo.21243976). A preprint of this work is available on bioRxiv (DOI: https://doi.org/10.64898/2026.05.05.722920).

References

  • 1. Kantarjian H., Stein A., Gokbuget N., et al., “Blinatumomab Versus Chemotherapy for Advanced Acute Lymphoblastic Leukemia,” New England Journal of Medicine 376, no. 9 (2017): 836–847, 10.1056/NEJMoa1609783. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2. Teachey D. T., Lacey S. F., Shaw P. A., et al., “Identification of Predictive Biomarkers for Cytokine Release Syndrome After Chimeric Antigen Receptor T‐Cell Therapy for Acute Lymphoblastic Leukemia,” Cancer Discovery 6, no. 6 (2016): 664–679, 10.1158/2159-8290.CD-16-0040. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3. Suntharalingam G., Perry M. R., Ward S., et al., “Cytokine Storm in a Phase 1 Trial of the Anti‐CD28 Monoclonal Antibody TGN1412,” New England Journal of Medicine 355, no. 10 (2006): 1018–1028, 10.1056/NEJMoa063842. [DOI] [PubMed] [Google Scholar]
  • 4. Chatenoud L., Ferran C., Reuter A., et al., “Systemic Reaction to the Anti‐T‐Cell Monoclonal Antibody OKT3 in Relation to Serum Levels of Tumor Necrosis Factor and Interferon‐Gamma,” New England Journal of Medicine 320, no. 21 (1989): 1420–1421, 10.1056/NEJM198905253202117. [DOI] [PubMed] [Google Scholar]
  • 5. Abramowicz D., Schandene L., Goldman M., et al., “Release of Tumor Necrosis Factor, Interleukin‐2, and Gamma‐Interferon in Serum After Injection of OKT3 Monoclonal Antibody in Kidney Transplant Recipients,” Transplantation 47, no. 4 (1989): 606–608, 10.1097/00007890-198904000-00008. [DOI] [PubMed] [Google Scholar]
  • 6. Norelli M., Camisa B., Barbiera G., et al., “Monocyte‐Derived IL‐1 and IL‐6 Are Differentially Required for Cytokine‐Release Syndrome and Neurotoxicity due to CAR T Cells,” Nature Medicine 24, no. 6 (2018): 739–748, 10.1038/s41591-018-0036-4. [DOI] [PubMed] [Google Scholar]
  • 7. Weddell J., “Mechanistically Modeling Peripheral Cytokine Dynamics Following Bispecific Dosing in Solid Tumors,” CPT: Pharmacometrics & Systems Pharmacology 12, no. 11 (2023): 1726–1737, 10.1002/psp4.12928. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8. Sove R. J., Jafarnejad M., Zhao C., Wang H., Ma H., and Popel A. S., “QSP‐IO: A Quantitative Systems Pharmacology Toolbox for Mechanistic Multiscale Modeling for Immuno‐Oncology Applications,” CPT: Pharmacometrics & Systems Pharmacology 9, no. 9 (2020): 484–497, 10.1002/psp4.12546. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9. Clements J. D., Zhu M., Kuchimanchi M., Terminello B., and Doshi S., “Population Pharmacokinetics of Blinatumomab in Pediatric and Adult Patients With Hematological Malignancies,” Clinical Pharmacokinetics 59, no. 4 (2020): 463–474, 10.1007/s40262-019-00823-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10. Klinger M., Brandl C., Zugmaier G., et al., “Immunopharmacologic Response of Patients With B‐Lineage Acute Lymphoblastic Leukemia to Continuous Infusion of T Cell‐Engaging CD19/CD3‐Bispecific BiTE Antibody Blinatumomab,” Blood 119, no. 26 (2012): 6226–6233, 10.1182/blood-2012-01-400515. [DOI] [PubMed] [Google Scholar]
  • 11. Stebbings R., Findlay L., Edwards C., et al., “‘Cytokine Storm’ in the Phase I Trial of Monoclonal Antibody TGN1412: Better Understanding the Causes to Improve Preclinical Testing of Immunotherapeutics,” Journal of Immunology 179, no. 5 (2007): 3325–3331, 10.4049/jimmunol.179.5.3325. [DOI] [PubMed] [Google Scholar]
  • 12. Dirks N. L. and Meibohm B., “Population Pharmacokinetics of Therapeutic Monoclonal Antibodies,” Clinical Pharmacokinetics 49, no. 10 (2010): 633–659, 10.2165/11535960-000000000-00000. [DOI] [PubMed] [Google Scholar]
  • 13. Hooks M. A., Wade C. S., and W. J. Millikan, Jr. , “Muromonab CD‐3: A Review of Its Pharmacology, Pharmacokinetics, and Clinical Use in Transplantation,” Pharmacotherapy 11, no. 1 (1991): 26–37, 10.1002/j.1875-9114.1991.tb03595.x. [DOI] [PubMed] [Google Scholar]
  • 14. Ma H., Wang H., Sove R. J., et al., “A Quantitative Systems Pharmacology Model of T Cell Engager Applied to Solid Tumor,” AAPS Journal 22, no. 4 (2020): 85, 10.1208/s12248-020-00450-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15. Hawkins E. D., Turner M. L., Dowling M. R., van Gend C., and Hodgkin P. D., “A Model of Immune Regulation as a Consequence of Randomized Lymphocyte Division and Death Times,” Proceedings of the National Academy of Sciences of the United States of America 104, no. 12 (2007): 5032–5037, 10.1073/pnas.0700026104. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16. Teachey D. T., Rheingold S. R., Maude S. L., et al., “Cytokine Release Syndrome After Blinatumomab Treatment Related to Abnormal Macrophage Activation and Ameliorated With Cytokine‐Directed Therapy,” Blood 121, no. 26 (2013): 5154–5157, 10.1182/blood-2013-02-485623. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17. Lee D. W., Gardner R., Porter D. L., et al., “Current Concepts in the Diagnosis and Management of Cytokine Release Syndrome,” Blood 124, no. 2 (2014): 188–195, 10.1182/blood-2014-05-552729. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18. Chatenoud L., Ferran C., Legendre C., et al., “In Vivo Cell Activation Following OKT3 Administration. Systemic Cytokine Release and Modulation by Corticosteroids,” Transplantation 49, no. 4 (1990): 697–702, 10.1097/00007890-199004000-00009. [DOI] [PubMed] [Google Scholar]
  • 19. Nishimoto N., Terao K., Mima T., Nakahara H., Takagi N., and Kakehi T., “Mechanisms and Pathologic Significances in Increase in Serum Interleukin‐6 (IL‐6) and Soluble IL‐6 Receptor After Administration of an Anti‐IL‐6 Receptor Antibody, Tocilizumab, in Patients With Rheumatoid Arthritis and Castleman Disease,” Blood 112, no. 10 (2008): 3959–3964, 10.1182/blood-2008-05-155846. [DOI] [PubMed] [Google Scholar]
  • 20. Le R. Q., Li L., Yuan W., et al., “FDA Approval Summary: Tocilizumab for Treatment of Chimeric Antigen Receptor T Cell‐Induced Severe or Life‐Threatening Cytokine Release Syndrome,” Oncologist 23, no. 8 (2018): 943–947, 10.1634/theoncologist.2018-0028. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 21. Lee D. W., Santomasso B. D., Locke F. L., et al., “ASTCT Consensus Grading for Cytokine Release Syndrome and Neurologic Toxicity Associated With Immune Effector Cells,” Biology of Blood and Marrow Transplantation 25, no. 4 (2019): 625–638, 10.1016/j.bbmt.2018.12.758. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22. Singh A. P., Zheng X., Lin‐Schmidt X., et al., “Development of a Quantitative Relationship Between CAR‐Affinity, Antigen Abundance, Tumor Cell Depletion and CAR‐T Cell Expansion Using a Multiscale Systems PK‐PD Model,” MAbs 12, no. 1 (2020): 1688616, 10.1080/19420862.2019.1688616. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 23. International Council for Harmonisation (ICH) , “M15: General Principles for Model‐Informed Drug Development,” Step 4 Final Guideline (2026), https://database.ich.org/sites/default/files/ICH_M15_Step4_Final_Guideline_2026_0129.pdf.

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Table S1: Three‐drug CRS comparison from one locked cytokine network.

Table S2: TGN1412 comparison with Suntharalingam 2006 clinical data.

Table S3: OKT3 blind prediction: 8/8 validation checks passed.

Table S4: Parameter swap test: exchanging k exhaust between OKT3 and TGN1412.

Table S5: Tocilizumab rescue: five structural predictions matching clinical evidence. Tocilizumab is an independent intervention validation, not a blind prediction.

Table S6: Accepted parameter ensemble summary (5000 samples, 926 accepted). Median and 5th–95th percentile values across the accepted parameter sets, compared with the calibrated point.

Table S7: Consolidated model‐evaluation evidence supporting the cross‐drug CRS transferability claim across the three antibody‐based drug classes.

Figure S1: T cell mechanism determines CRS severity. (A) Cytokine‐producing T cell (T cyt) dynamics for blinatumomab (blue), OKT3 (green), and TGN1412 (red). OKT3 shows peak‐and‐decline (AICD), while TGN1412 shows sustained expansion. (B) Resulting IL‐6 dynamics (log scale). T cell expansion drives macrophage activation and IL‐6 amplification through the shared cytokine network.

Table S8: Quantitative comparison of predicted vs. observed cytokine peaks for the CRS module. Two acceptance criteria are applied: (i) whether the predicted peak falls within the reference range (measured clinical or author‐defined; see footnotes), and (ii) absolute fold‐error (AFE) relative to the geometric midpoint of that range (AFE = max(pred/G, G/pred) where G=low×high), with AFE ≤ 2 as the acceptance threshold.

Table S9: Locked cytokine network parameters. These 17 parameters are identical across all three drug classes and across the tocilizumab rescue simulation.

Figure S2: Illustrative CRS grade distributions across virtual populations (N = 100 per drug). Inter‐individual variability was sampled from log‐normal distributions (CV = 25%–30%) on T cell activation parameters (k act, k prolif, k exhaust, EC50,act, T act,max), macrophage pool size, and body weight (see Methods for assumptions and limitations). Blinatumomab: predominantly Grade 1 (~85%) with ~13% Grade 2 and ~2% sub‐clinical. OKT3: ~96% Grade 2 with ~4% Grade 3, reflecting the AICD ceiling on CRS severity. TGN1412: split between Grade 3 (~67%) and Grade 4 (~33%), qualitatively consistent with clinical Grade 3–4 in all six Suntharalingam 2006 subjects. These distributions reflect sensitivity to assumed parameter variability rather than formally calibrated population predictions.

Figure S3: Cross‐drug stability map for the two threshold‐setting parameters. (A) Regime classification across a 30 × 30 grid of (k IL6,fb, k deg,IL6). Muted teal: all three antibody‐based drugs predict peak IL‐6 within their reference ranges. Light gray: feasible but at least one drug outside its reference range. Muted rose‐red: at least one drug undergoes IL‐6 amplification runaway. Medium gray would indicate wrong cross‐drug severity ordering (never observed in the data). Dashed line: theoretical bifurcation diagonal k IL6,fb = k deg,IL6 (the boundary under maximal macrophage activation). White star: calibrated parameter point. Of 900 grid points, 157 (17.4%) fall in the “all in range” region; 0 produced wrong cross‐drug severity ordering. (B) Predicted TGN1412 peak IL‐6 across the same grid (log scale, single‐hue blue sequential colourmap; darker blue indicates higher peak IL‐6), with dark‐gray contour lines marking the 3000 pg/mL Grade 3 floor, the 5000 pg/mL Grade 4 threshold, and the 100,000 pg/mL upper clinical bound.

Figure S4: Tornado plot of one‐at‐a‐time sensitivity coefficients for the CRS model. Each bar shows the normalized sensitivity S = (ΔY/Y 0)/(Δp / p 0) with peak IL‐6 as the output, for blinatumomab (the calibration agent), computed at ±50% parameter perturbation from baseline; *k fb uses ±20% because larger perturbations push the system past the amplification threshold. †For k deg,IL6 the plotted value is the forward (+50%) arm only, and is therefore not comparable to the symmetric coefficients: halving the IL‐6 clearance rate to 0.26 h−1 places it below the effective feedback rate of every agent (0.32, 0.46, and 0.52 h−1), crossing the amplification condition k IL6,fb ·f mac >k deg,IL6, so the STAT3 loop becomes self‐sustaining and peak IL‐6 diverges. That arm is a regime transition rather than a local sensitivity, and is marked “runaway” on the plot. Parameters with |S| > 1 amplify the input change; parameters with |S| < 1 attenuate it. Macrophage‐derived IL‐6 production (macpool, k prod,IL6,mac) and the macrophage‐activation threshold (EC50,mac) dominate for blinatumomab. The coefficients are drug‐dependent because the system operates near a bifurcation: for the near‐threshold agents the feedback gain dominates instead (S = +16.3 for OKT3; for TGN1412 the upward arm diverges and only the backward coefficient, +4.0, is defined). One‐at‐a‐time perturbation does not capture parameter interactions. Computed by sensitivity_oat.py.

Table S10: Fixed model constants shared across all drugs (not fitted).

Supplementary Methods. Full drug pharmacokinetic and IFN‐γ‐driven macrophage‐recruitment equations; CRS model state equations; the formal constrained calibration objective function and procedure; the local accepted parameter ensemble; the multi‐cytokine acceptance test; the profile‐likelihood analysis; the cross‐drug stability‐map procedure; and the computational environment and reproducibility checks.

PSP4-15-e70376-s001.pdf (811.7KB, pdf)

Data Availability Statement

All validation data analyzed in this study were extracted from previously published sources cited in the text. The complete model equations, the macrophage‐gated STAT3 feedback and macrophage‐recruitment terms, the constrained calibration objective, and the locked 17‐parameter cytokine network are provided in full in Section S1 and Table S9, and are sufficient to reimplement the model. A reproducible implementation of the CRS model (the simulation engine, the locked cytokine‐network parameters, the drug‐specific parameter sets, and the calibration, validation, and robustness scripts), together with the digitized clinical target datasets and the analysis result files needed to reproduce every reported number, is openly available at Zenodo (DOI: https://doi.org/10.5281/zenodo.21243976). A preprint of this work is available on bioRxiv (DOI: https://doi.org/10.64898/2026.05.05.722920).


Articles from CPT: Pharmacometrics & Systems Pharmacology are provided here courtesy of Wiley

RESOURCES