Global Sensitivity Analysis and Drug Target Ranking

Local sensitivity analysis (Equation sensitivity) reveals parameter importance near a specific operating point. Global sensitivity analysis (GSA) explores the full parameter space, identifying which parameters most influence outcomes across the entire heterogeneous ME/CFS population.

1 Sobol Index Framework

Variance-based Sobol indices decompose the variance of a model output \(Y = g(\mathbf{\theta})\) into contributions from individual parameters and their interactions (Sobol’ 2001):

\[ S_i = \frac{\text{Var}_{\theta_i}[\mathbb{E}_{\mathbf{\theta}_{~ i}}(Y | \theta_i)]}{\text{Var}(Y)} \tag{1}\]

\[ S_{T_i} = 1 - \frac{\text{Var}_{\mathbf{\theta}_{~ i}}[\mathbb{E}_{\theta_i}(Y | \mathbf{\theta}_{~ i})]}{\text{Var}(Y)} \tag{2}\]

where \(S_i\) is the first-order index (main effect of parameter \(\theta_i\)) and \(S_{T_i}\) is the total-order index (including all interactions). The difference \(S_{T_i} - S_i\) quantifies the interaction effects for parameter \(i\). For the 64-variable ME/CFS model (Chapter Integrated Multi-System Models), applying GSA with the symptom severity composite as the output \(Y\) identifies the druggable parameters with greatest population-level impact.

Predicted ranking by total Sobol index \(S_{T_i}\) for overall symptom burden:

  • \(\alpha_\text{CI}\) (Complex I activity): \(S_T \approx 0.22\)—the single most influential parameter across the population, consistent with central role of energy deficit
  • \(P_0\) (BBB permeability): \(S_T \approx 0.14\)—high interaction effects (\(S_T - S_1 \approx 0.09\)) because BBB gating amplifies peripheral inflammation centrally
  • \(k_\text{exh}\) (immune exhaustion rate): \(S_T \approx 0.12\)
  • \(n_F\) (HPA feedback sensitivity): \(S_T \approx 0.10\)
  • \(K_\text{MC}\) (mast cell activation threshold, Equation mast cell dynamics): \(S_T \approx 0.08\)
  • \(\beta_\text{epoxy}\) (microclot formation rate, Equation coagulation dynamics): \(S_T \approx 0.07\)
  • BH4 synthesis rate (\(k_\text{BH4syn}\), Equation bh4 dynamics): \(S_T \approx 0.05\)
  • \(\mathcal{G}_\text{set}\) (gut motility set point, Equation motility setpoint): \(S_T \approx 0.04\)—low main effect but significant interaction with \(K_\text{MC}\) and vagal tone \(V\), reflecting the mast cell–motility–SIBO feedback loop (Section Gut–Brain–Immune Axis)
NoteModel Insight: Interaction-Dominant Parameters as Combination Targets

GSA uniquely reveals that some parameters have large interaction effects (\(S_{T_i} >> S_i\)) but small main effects (\(S_i\) alone). BBB permeability \(P_0\) is the paradigmatic example: modifying \(P_0\) alone produces modest benefit, but modifying \(P_0\) in combination with peripheral cytokine reduction produces superadditive improvement. This identifies BBB-stabilizing agents (e.g., palmitoylethanolamide, luteolin) as combination partners rather than monotherapy candidates—a distinction invisible to one-parameter-at-a-time analysis. Similarly, the high interaction index of \(K_\text{MC}\) suggests that mast cell stabilizers gain most of their therapeutic value in combination with other interventions, explaining the clinical observation that cromolyn sodium shows variable efficacy as monotherapy.

2 Subtype-Stratified Sensitivity

The four disease attractors identified in Section Bifurcation Analysis and Disease Subtypes define distinct subtypes with different sensitivity profiles. Performing GSA within each attractor basin reveals subtype-specific drug targets:

  • Metabolic-dominant (Attractor B): \(\alpha_\text{CI}\) dominates (\(S_T \approx 0.35\)); mitochondrial-targeted therapies (CoQ10, NAD⁺ precursors, methylene blue) predicted most effective
  • Immune-dominant (Attractor C): \(k_\text{exh}\) and cytokine production rates dominate; immunomodulators (LDN, rintatolimod, daratumumab per Equation daratumumab) predicted most effective
  • Autonomic-dominant: baroreflex gain \(G_\text{baro}\) and blood volume \(V_\text{blood}\) dominate; volume expansion and autonomic agents (fludrocortisone, midodrine, pyridostigmine) predicted most effective. GI motility \(\mathcal{G}_\text{set}\) (Equation motility setpoint) shows elevated sensitivity in this subtype (\(S_T \approx 0.09\)) because vagal impairment propagates to SIBO-mediated immune activation and malabsorption; prokinetics and SIBO eradication are predicted as high-value co-interventions
  • Severe/locked (Attractor D): epigenetic modification rates (Equation epigenetic evolution) dominate (\(S_T \approx 0.20\)); epigenetic modifiers or high-intensity combination therapy required
NoteModel Insight: Subtype-Specific Target Inversion

The model predicts target inversion between subtypes: a parameter that is the top target in one subtype may be irrelevant in another. For example, \(\alpha_\text{CI}\) has \(S_T \approx 0.35\) in the metabolic-dominant attractor but only \(S_T \approx 0.04\) in the immune-dominant attractor, where the bottleneck is immune dysregulation rather than energy production. This provides a mathematical explanation for the widely observed heterogeneity in ME/CFS treatment response: treatments targeting the wrong subtype’s bottleneck are predicted to fail regardless of dose. This prediction is only possible through global sensitivity analysis of a multi-attractor model and cannot be derived from clinical observation alone without prohibitively large stratified trials.

References

Sobol’, Ilya M. 2001. “Global Sensitivity Indices for Nonlinear Mathematical Models and Their Monte Carlo Estimates.” Mathematics and Computers in Simulation 55 (1–3): 271–80. https://doi.org/10.1016/S0378-4754(00)00270-6.