Friberg's myelosuppression model, first published in 2002 in the Journal of Clinical Oncology, has become the most widely applied PK/PD framework for linking cytotoxic drug exposure to hematologic toxicity. At the time of this writing it has been applied to over 40 cytotoxic and targeted agents, from docetaxel to etoposide to ribociclib. Understanding what it actually models - and where it breaks down - is prerequisite knowledge for anyone using a dosing platform that generates toxicity probability estimates.
This article is a technical walkthrough, not a general introduction. Readers who are not familiar with compartmental PK modeling may want to read our article on population PK models and Bayesian estimation first.
The Transit Compartment Architecture
Friberg's model represents the bone marrow progenitor-to-circulating-cell pipeline as a series of transit compartments. The pipeline has five compartments: a proliferating progenitor compartment (called "prol"), three transit (maturation) compartments (transit1, transit2, transit3), and a peripheral blood compartment (circ) representing the observable circulating cell count.
The mathematical structure is: cells enter the progenitor compartment at a rate governed by a proliferation rate constant (kprol) that is self-regulated by a feedback mechanism. Cells move through each transit compartment at a maturation rate constant (ktr), assumed equal across all transit compartments in the original model. Cells exit the circulating compartment at a first-order elimination rate (kcirc).
At steady state (before any drug exposure), all compartment sizes are at their baseline values (circ0 for circulating, prol0 for the progenitor compartment), and the proliferation and elimination rates are in balance. The feedback term is what prevents the model from predicting runaway bone marrow expansion in the absence of drug: it reduces kprol proportionally to circ/circ0, so as the circulating count increases above baseline, proliferation slows. As count falls (in response to drug), proliferation accelerates.
How Drug Kills in the Model: The Emax Function
Drug effect in Friberg's model acts on the proliferating progenitor compartment. The drug's effect on kprol is modeled as a concentration-dependent inhibition: kprol x (1 - Emax x C / (EC50 + C)), where C is the drug concentration from the PK model, Emax is the maximum fractional inhibition, and EC50 is the drug concentration producing 50% of maximum inhibition.
This is a direct PK/PD linkage: the PK model's predicted concentration-time profile is fed into the myelosuppression model at each time step, and the resulting proliferation inhibition drives the compartment kinetics forward through time. The nadir timing and depth are emergent properties of the interaction between the PK profile and the transit compartment speeds (ktr).
ktr in the original model is derived from the mean transit time (MTT): ktr = (n+1)/MTT, where n is the number of transit compartments. MTT represents the biological time from committed progenitor to mature circulating cell, which is approximately 100-140 hours for neutrophils in adults. This biological grounding is what gives the model predictive credibility - the compartment transit rate is not a pure curve-fitting parameter but is anchored to a measurable physiological quantity.
The Feedback Parameter: Biology vs Curve Fitting
The feedback parameter (gamma) in the proliferation rate equation governs how aggressively the marrow compensates for drug-induced neutropenia. It appears in the feedback term as: kprol x (circ0/circ)^gamma. A gamma of 0.17 (the value in Friberg's original publication, estimated from docetaxel data) represents mild feedback - the marrow compensates somewhat but does not fully overcome drug-induced suppression. Higher gamma values produce more aggressive recovery; lower values produce deeper and more prolonged nadirs.
The problem is that gamma is estimated from population data and can vary substantially between drugs and populations. For agents that kill progenitors but spare the stromal niche (most cytotoxics), gamma behaves as Friberg described. For agents that damage the marrow stroma or the hematopoietic stem cell pool (certain alkylating agents, radiation conditioning), the feedback mechanism itself is impaired, and gamma will be lower than the population estimate. Using the published docetaxel gamma for busulfan conditioning in HSCT - where marrow ablation is the therapeutic goal rather than a side effect - will produce wildly inaccurate predictions.
Cycle-to-Cycle Degradation: Where the Model Gets Harder
In single-cycle predictions, Friberg's model performs well for most cytotoxics - typically within 10-20% of observed nadir timing and depth. In multi-cycle predictions, accuracy degrades systematically for two reasons.
First, the model assumes the same baseline circ0 at the start of each cycle. In reality, bone marrow reserve is progressively depleted with each cytotoxic cycle. A patient who enters cycle 1 with a circ0 of 5.0 x10^9/L neutrophils may effectively enter cycle 4 with a functional circ0 of 3.5 x10^9/L, even if their pre-cycle-4 measured count has recovered to that level. The recovery to pre-dose count does not mean marrow reserve is fully restored - it means the remaining progenitor pool has compensated to maintain the circulating count at the cost of reduced buffer for additional drug exposure.
Second, the feedback parameter estimated from early-cycle data may not represent the patient's late-cycle marrow biology. As cumulative drug exposure increases and marrow reserve declines, the effective gamma decreases - the feedback compensation mechanism weakens. Friberg's model, with a fixed gamma estimated from early cycles, will overpredict recovery and underpredict late-cycle nadir depth as a result.
Extensions to the Basic Framework
Several published extensions address specific limitations of the base model. The Quartino extension (2012) introduced a granulocyte colony-stimulating factor (G-CSF) effect module that models the pharmacodynamics of G-CSF support given between cycles. This is critical for any patient receiving prophylactic or therapeutic filgrastim, where the base Friberg model will drastically underpredict ANC recovery speed.
The Henningsson modification adds a growth factor feedback term that explicitly tracks progenitor pool size as a state variable, allowing cycle-to-cycle baseline erosion to be represented mechanistically rather than as a fixed-baseline approximation. This performs better in multi-cycle predictions but requires additional parameters estimated from multi-cycle data - which may not be available at the time clinical decisions need to be made.
For agents whose toxicity is predominantly on red cell series (erythroblasts rather than myeloid progenitors) - for example, certain ribonucleotide reductase inhibitors - modified transit compartment models with red cell maturation kinetics are required. Neutrophil and hemoglobin dynamics have different MTT values and different feedback sensitivities; using a neutrophil-calibrated model to predict red cell toxicity will produce inaccurate output.
What a Toxicity Probability Estimate Actually Means
When DoseMind generates a statement like "probability of grade 3/4 neutropenia: 42%," that number is derived from running the PK/PD myelosuppression model forward from the patient's current PK parameter estimate and applying a stochastic layer that propagates uncertainty from three sources: uncertainty in the individual's PK parameters (from the Bayesian posterior), uncertainty in the PD parameters (ktr, MTT, gamma - typically represented as population-level covariate relationships with residual unexplained variability), and measurement error in the observed CBC.
The 42% figure is a credible interval-integrated probability over all plausible parameter combinations. It is not a point prediction. The clinical interpretation is: for a patient with this PK profile and these CBC values going into this cycle, roughly 4 in 10 patients with similar characteristics have experienced grade 3/4 neutropenia. That calibration should be validated against your site's observed toxicity rates if you have sufficient data to do so.
Practical Implications for Protocol-Based Toxicity Monitoring
Integrating a myelosuppression model into clinical practice raises a question that model developers rarely address: at what probability threshold does the clinician change the dose? A 42% predicted probability of grade 3/4 neutropenia means the most likely outcome is that grade 3/4 neutropenia will not occur. Using that probability to trigger dose reduction will lead to dose reductions in the majority of patients who would not have had a DLT anyway.
The threshold decision depends on the clinical context. In Phase I dose escalation, where the therapeutic goal is characterizing the dose-toxicity relationship rather than maximizing individual efficacy, a lower threshold for preemptive dose reduction is appropriate. In Phase II where efficacy is the primary endpoint, unnecessarily reducing doses based on probabilistic toxicity predictions may sacrifice response rates in patients who would have tolerated the higher dose. Protocol language should specify the toxicity probability thresholds that trigger dose adjustment recommendations and whether those recommendations are mandatory or advisory.
Conclusion: The Model Is Only as Good as Its Calibration
Friberg's myelosuppression model is the best available framework for PK/PD-linked toxicity prediction in cytotoxic oncology. It is not a black box - its parameters have biological meaning, and understanding that meaning is essential for interpreting its output. The most important practical lesson from the literature is that single-cycle predictions are reliable when the model is well-calibrated to the drug and population; multi-cycle predictions require either a modified framework that tracks cumulative marrow depletion or explicit uncertainty acknowledgment that late-cycle nadir predictions carry higher error than early-cycle predictions.
A dosing platform that runs this model in real time, anchored to each patient's observed PK data, provides toxicity probability estimates that are meaningfully better than clinical intuition alone for pre-emptive dose management. Contact the DoseMind team at hello@dosemind.com to discuss how the myelosuppression model can be calibrated to your specific agents and patient population.