A Physics-Constrained Marcus-Theory Optimization Framework Restores Parameter Identifiability in Deep-Pocket Oncogenic Drug Kinetics: A KRAS G12C and STAT3 Case Study

Background: More than 90% of drug candidates fail preclinically, partly because empirical dose-response models such as the Hill equation ignore microenvironmental effects inside the deep binding pockets of targets such as mutant KRAS G12C and STAT3. Bulk water is hypothesized to raise the reorganization energy (lambda_ext). Objective: To develop a physics-constrained framework, the Cluchagues-Marcus Framework, that links local dielectric permittivity to ligand-pocket kinetics, and to determine which constraints restore parameter identifiability. Methods: Non-adiabatic Marcus kinetics were coupled to a Hill-type equation derived from a steady-state scheme, and water exclusion was modeled as a collapse of epsilon_local from 80.4 to 2.0. Pocket radii and published reference constants were used for two pockets (KRAS G12C, PDB 6UT0; STAT3 SH2, PDB 6NJS). Three propositions were proved and checked with synthetic data and a global sensitivity analysis, and the identifiability result was numerically stress-tested on real napabucasin viability curves from 701 cancer cell lines. Results: Water exclusion lowered lambda_ext by 90%. At dG = -0.10 eV, k_C rose 8.0-fold (KRAS) and 3.6-fold (STAT3). Acceleration occurs if and only if |dG| < sqrt(lambda_s x lambda_u). The model has exactly two non-identifiable parameter directions: fixing nu_eff removes one and anchoring K_d removes the other. In real curves, fits were unchanged by the K_d anchor (SSE change < 10^-7 %) while the fitted k_off,q varied by 3.5 log units, as predicted. Conclusions: A dissociation-constant anchor makes the framework identifiable. The real-data stress test is not biological validation of the dielectric mechanism, which requires experiments.