Hybrid Stochastic-Deterministic Models for Rare Cell Events

Why Combine Stochastic and Deterministic Dynamics?

A cell population can contain both abundant compartments and rare events. For example, a large susceptible population may change smoothly at the scale of interest, while the first few resistant cells, infected cells, or cells entering a new state occur as discrete events. A fully deterministic model can represent the large population efficiently, but it may erase the randomness in the timing or occurrence of those rare transitions. A fully stochastic model preserves that randomness, but simulating every event in a very large population can be computationally expensive.

A hybrid stochastic-deterministic model keeps selected states as integer counts and evolves other states as continuous quantities. The partition is a modeling assumption, not a claim that some biological processes are intrinsically random and others intrinsically deterministic. It is appropriate when the stochastic events are consequential at the scale of interest and the fluctuations omitted from the deterministic states are acceptably small.

The propensity \(a_j\) is an instantaneous event rate: conditional on the current state, the probability of one event of type \(j\) in a short interval of length \(dt\) is \(a_jdt+o(dt)\), while the probability of two or more such events is \(o(dt)\). The total propensity is \(a_0=\sum_j a_j\). The earlier material on cellular continuous-time Markov chains and exact simulation supplies the interpretation of propensities and discrete jumps; the hybrid extension is that propensities may evolve because \(Z\) follows a continuous trajectory between events.

Specifying the Partition

A useful partition is made at the level of states and reaction channels, with attention to what each channel does. A reaction whose occurrence represents a consequential rare cell transition should generally remain stochastic. A process involving a large population may be represented by a deterministic flux when its event count is large enough that individual fluctuations are negligible for the question being asked. A channel that consumes or produces both types of state needs special care: a discrete event can update a continuous concentration by a small amount, but a continuous flux cannot create a fractional individual in a state that is being treated as an integer count.

For a stochastic channel \(j\), let \(\nu_j\) be its change to \(X\), and let \(\zeta_j\) be any corresponding jump to \(Z\). At an event, the update is \((X,Z)\mapsto(X+\nu_j,Z+\zeta_j)\). Between events, \(X\) remains fixed and \(Z\) follows its differential equation. Deterministic channels contribute to \(f\), the continuous-state drift. This description also makes an important bookkeeping requirement visible: if a reaction moves material between represented states, the corresponding changes must preserve the intended mass or cell balance.

The criterion is therefore purpose-dependent. A population can be abundant in absolute terms yet require stochastic representation if the outcome depends on a small difference between competing event counts. Conversely, a rare species may be approximated deterministically in a narrowly defined setting if only its mean trajectory is needed and variability is demonstrably unimportant. The modeler should state which features are retained as random and which are intentionally averaged.

Event Probabilities Along a Deterministic Trajectory

Between stochastic events, \(X\) is fixed but \(Z(t)\) may change. Consequently, a channel’s propensity can vary with time even when the discrete state does not. Write \(a_0(t)=\sum_j a_j(X,Z(t),t)\) along the no-event trajectory. The cumulative hazard over an interval of length \(T\) is \(H(T)=\int_0^T a_0(u)\,du\). It determines the probability of no stochastic event during that interval.

This result follows from the conditional event-rate definition. If \(S(t)\) is the probability of no event by time \(t\), then over a sufficiently short interval \(dt\), survival is multiplied by \(1-a_0(t)dt+o(dt)\). Thus \(dS/dt=-a_0(t)S(t)\), with \(S(0)=1\). Separating variables and integrating gives \(\log S(T)=-\int_0^T a_0(u)\,du\), hence the displayed survival probability. If an event occurs at time \(t\), the probability that it is of type \(j\), conditional on an event at that instant, is \(a_j(t)/a_0(t)\) when \(a_0(t)>0\).

Worked Hybrid Model with Competing Rare Events

The next example shows how stochastic rare transitions can coexist with an abundant state that follows a deterministic equation. It also illustrates why one must check the time-variation assumption used in an event-probability calculation.

When a Population Can Become Deterministic

A hybrid model may keep a newly appearing or very small cell population stochastic, then represent it deterministically after it becomes sufficiently large. Such a switch is not automatically valid at a particular count. Its acceptability depends on the desired accuracy, the event rates, and whether outcomes such as extinction or threshold crossing remain important.

Diagnostics, Alternatives, and Limitations

Several checks help determine whether a hybrid specification is behaving as intended. First, inspect whether the chosen discrete states remain integer-valued and whether continuous states remain in their biologically valid ranges. Second, check the bookkeeping for every transition: cell gains, losses, and transfers should match the model’s stated conservation or balance relationships. Third, compare predictions under plausible alternative partitions or switch thresholds. If a small change in the partition changes the predicted event probability, extinction frequency, or time to a clinically meaningful threshold, the result may depend materially on the approximation.

A further check concerns time variation. If a deterministic driver changes substantially during an interval, a calculation based on a fixed propensity can misrepresent event timing and probability. Use the integrated hazard along the deterministic trajectory, and verify that numerical integration or simulation resolves the trajectory sufficiently. Likewise, a deterministic approximation should not silently allow negative populations or concentrations. Boundary handling must be part of the model, not an afterthought.

The main alternatives serve different purposes. A full stochastic simulation treats all modeled reactions as discrete events and is preferable when fluctuations throughout the system matter, but may require many event updates for very large populations. A fully deterministic model is efficient and can be suitable when only smooth average behavior is needed, but it cannot represent the chance that a rare event has not occurred or the random timing of its first occurrence. A hybrid model lies between these choices: it saves effort on abundant dynamics while retaining randomness for selected events. Its limitation is that it combines two representations and can therefore introduce errors at their interface, particularly through inconsistent transfers or abrupt switching.

A common mistake is to classify every low-abundance species as stochastic and every high-abundance species as deterministic without considering the reactions and predictions of interest. The correction is to assess event counts, coupling, boundaries, and the scientific question together. A large population can still have a consequential rare transition, while a small population’s fluctuations may be irrelevant to a narrowly defined average prediction. The partition and its rationale should be reported so that readers can judge which sources of variability the model retains.