Sensitivity Analysis

Core Idea

Sensitivity analysis determines how changes in model parameters or inputs affect model outputs, identifying which parameters most strongly influence the system's behavior. Local sensitivity analysis computes partial derivatives of outputs with respect to individual parameters near a specific operating point. Global sensitivity analysis (Sobol indices, Morris screening) explores the full parameter space, accounting for interactions between parameters and nonlinear effects. In metabolic systems, metabolic control analysis (MCA) formalizes sensitivity as flux control coefficients and elasticity coefficients. Sensitivity analysis guides experimental design (measure the parameters that matter most), identifies drug targets (the most sensitive nodes in disease networks), and reveals which model predictions are robust versus parameter-dependent.

Explainer

Every systems biology model, whether an ODE model of signaling dynamics or an FBA model of metabolism, depends on parameters whose values are uncertain. Sensitivity analysis asks: which of these uncertain parameters actually matter for the model's predictions? If the model output barely changes when a parameter varies over its plausible range, that parameter can be fixed at a nominal value without loss. If the output changes dramatically, that parameter is a priority for experimental measurement — and a potential point of biological control or therapeutic intervention.

Local sensitivity analysis is the simplest approach: compute the partial derivative of the output with respect to each parameter at the current operating point. For an ODE model, this can be done analytically (solving the sensitivity equations alongside the state equations) or numerically (perturbing each parameter by a small amount and observing the output change). The result is a sensitivity coefficient for each parameter that quantifies its local influence. In metabolic systems, this concept is formalized as metabolic control analysis (MCA), where the flux control coefficient C_i^J measures how much a fractional change in enzyme i's activity changes the flux J. The summation theorem (all flux control coefficients sum to 1) reveals that metabolic control is shared among enzymes, demolishing the concept of a single rate-limiting step.

Global sensitivity analysis goes further by exploring the entire plausible parameter range. Sobol indices decompose the total variance of the model output into contributions from individual parameters (first-order indices) and from interactions between parameters (higher-order indices). A parameter with a high first-order Sobol index drives substantial output uncertainty on its own; a parameter with a high total-order index (including interactions) may not matter individually but strongly modulates the effect of other parameters. Morris screening provides a cheaper alternative: it samples the parameter space with a design that estimates, for each parameter, the mean and variance of its elementary effect — classifying parameters as negligible (small mean, small variance), linearly important (large mean, small variance), or nonlinearly important or interacting (large variance).

The practical payoff of sensitivity analysis is threefold. For experimental design, it prioritizes which parameters to measure: invest limited experimental resources in the parameters that most influence model predictions. For drug target identification, the most sensitive nodes in a disease network model are the most promising therapeutic targets — perturbations at these nodes produce the largest phenotypic effects. For model robustness assessment, outputs that are insensitive to most parameters are reliable predictions even with uncertain parameter values, while outputs that are highly sensitive to poorly constrained parameters should be interpreted cautiously. Sensitivity analysis transforms a model from a black box into a transparent tool that communicates not just what it predicts, but how confident those predictions are and what drives the uncertainty.