PRISM: Probabilistic Model Checker

• PRISM System is a probabilistic symbolic model checker that encodes complex biological pathways using modular CTMCs and temporal logic.
• It enables quantitative verification of dynamic properties including transient, steady-state, and reward-based analyses for realistic pathway modeling.
• Application to the PDGF pathway demonstrates how PRISM analyzes gene perturbations and network kinetics, guiding hypotheses for targeted drug development.

PRISM: Probabilistic Symbolic Model Checker for Biological Pathways

The PRISM system is a probabilistic model checking toolset designed to enable the modeling, specification, and quantitative analysis of complex stochastic systems. In the context of systems biology—specifically the evaluation of signaling pathways such as the Platelet-Derived Growth Factor (PDGF) network—PRISM provides formal, modular, and scalable means for encoding the dynamic behavior of biochemical networks and systematically verifying time-dependent, steady-state, and reward-based quantitative properties via temporal logic (Yuan et al., 2011).

1. Modeling Formalism

PRISM encodes stochastic systems as collections of interacting modules, each defined by typed variables and guarded commands. Each module’s state is represented by the current valuation of its variables, and its behavior is governed by guarded commands of the form: $[a]\, g \rightarrow r : u;$ where:

• $a$: (optional) action label for synchronization.
• $g$: Boolean guard predicate over all variables.
• $r$: nonnegative real rate (for CTMCs), probability (for discrete models).
• $u$: variable update.

Modules are composed in parallel, synchronizing on identical action labels or progressing independently for unlabelled commands. The system’s overall stochastic semantics are specified by the chosen underlying model.

PRISM supports three stochastic model semantics:

• Discrete-Time Markov Chains (DTMCs)
• Continuous-Time Markov Chains (CTMCs)
• Markov Decision Processes (MDPs)

In the PDGF analysis, CTMCs are used: each state transition occurs with a rate, and sojourn times in states are exponentially distributed.

2. Pathway Model Construction

The PDGF pathway is encoded as a single-instance CTMC with 17 species (including receptors, kinases, phosphatases, adaptors, and crosstalk enzymes). Each species is a module with variables representing activity states (typically binary, except for multi-valued states such as PDGFR: inactive/active/degraded).

Example: The PDGFR receptor module

module PDGFR PDGFR : [0..2] init 0; // 0=inactive, 1=active, 2=degraded [] PDGFR=0 -> k_on : (PDGFR'=1); [bkoff1] PDGFR=1 -> k_off1 : (PDGFR'=0); [bkoff2] PDGFR=1 -> k_off2 : (PDGFR'=0); [bkubi] PDGFR=1 -> k_ubi : (PDGFR'=0); [] PDGFR=1 -> k_deg : (PDGFR'=2); [] PDGFR=2 -> k_rec : (PDGFR'=0); endmodule

Rate parameters $k$ are normalized for computational tractability.

The global CTMC’s transition rate matrix $R(s,s')$ aggregates enabled commands and synchronized reactions. Each reaction, e.g., ligand binding or kinase activation, corresponds to a command with appropriate guards and rates.

Reward structure: PRISM's rewards enable quantification of accumulated quantities—“time spent active,” molecular concentrations—via reward-augmented logic and analysis.

3. Property Specification and Model Checking

Properties are specified in Continuous Stochastic Logic (CSL):

• State Probability:

$P_{=?}[F^{[t,t]} \phi]$

(Probability that property $\phi$ will eventually be true at time $t$.)

• Steady-State:

$S_{=?}[\phi]$

(Long-run probability of being in a state satisfying $\phi$.)

• Reward-based:

$R_{=?}[C^{\le t}]$

(Expected accumulated reward up to time $t$.)

Example queries:

• Transient activation probability for $m$ at time $t$:

$P_{=?}[F^{[t,t]}(m=1)]$

• Expected active-time for $m$ up to $t$:

$R_{=?}[\text{mActive}~C^{\le t}]$

• Steady-state activity:

$S_{=?}[m=1]$

PRISM supports model-checking for complex reachability, safety, and liveness properties, with direct applications to kinetic pathway dynamics, genetic "mutant" in silico knockouts, and reward accumulation metrics.

4. Quantitative and Mutant Analysis Results

For the wild-type PDGF network:

• CTMC size: $589,824$ states and $7,145,472$ transitions.
• Transient activation: PDGFR exhibits a peak at $t \approx 3$ with $P \approx 0.33$, then relaxes to steady-state $S_{=?}[PDGFR=1] \approx 0.22$.
• Mutant studies (e.g., SHP2-, PI3K-, cCbl-) demonstrate altered peak height and timing in transient activation, with up to 50% increase for cCbl knockout.
• Reward-based analyses: expected active-time transitions to linear scaling in post-transients.

Steady-state probabilities:

Molecule	Probability
PDGFR	0.22
SHP2	0.45
PI3K	0.62
Ras	0.72
MEK12	0.77
PIP3	0.72
Akt	0.82

Mutant-specific effects:

• Removing Ras→PI3K feedback lowers PI3K activation peak by $\sim$25%.
• Removing Akt→cRaf feedback increases Ras steady-state activity by $\sim$15%.

Knockout scatter plot analysis identifies reactions that modulate both proliferation (MEK12) and survival (Akt), guiding potential drug-targeting hypotheses.

All properties are computed using PRISM’s specialized engines (uniformization for transient, linear solvers for steady-state, value iteration for rewards), and large models (up to $10^6$ states) are tractable within a few minutes.

5. System Features and Computational Aspects

• Modularity: Each species as a module allows straightforward extension and perturbation (e.g., mutants, feedback removal).
• Synchronization: Action labels enable precise modeling of molecular interactions (enzyme-substrate, feedback, degradation).
• Statistical Robustness: 10,000+ queries per mutant efficiently executed; highly sparse transition matrices and binary/ternary variables mitigate state-explosion.
• Model Management: Both graphical UI and batch/CLI scripts, plain-text models, and automated parameter sweeps.
• Scalability: The largest assessed models (wild type and mutants) complete full evaluation on commodity hardware (single-core, few GB memory) in under five minutes.

6. Limitations and Future Directions

• Parameter uncertainty: All reaction rates are set relative, not absolute. The integration of Bayesian statistical model checking and parametric model checking is proposed for future work to match in vitro data and handle measurement noise.
• Approximation gaps: Discrepancies between ODE simulations and CTMCs arise in low-copy-number regimes, with PRISM’s discrete stochastic approach yielding differences in peak levels and equilibrium values.
• Model scalability: State-space growth is controlled through the use of binary/ternary variables, but expansion to heterogeneous copy numbers or more detailed multi-state models could pose challenges.

Future enhancements are expected to leverage advanced parametric and statistical model checking for systematic robustness analysis, along with application to larger, more complex pathway networks.

---

In summary, the PRISM model checker enables rigorous quantitative verification of biological pathways via guarded-command modular CTMCs, the expressivity of temporal logic, and efficient numerical analysis. Its application to the PDGF signaling pathway demonstrates both the feasibility and the analytic power of formal stochastic verification in elucidating biomolecular network dynamics, effects of gene perturbations, and candidate intervention strategies (Yuan et al., 2011).