A mechanistic computational cell-based potency assay for botulinum neurotoxin A enables transparent potency estimation and cross-assay harmonization.

Cell-based potency assays have replaced the historical LD50 test for botulinum neurotoxin type A (BoNT/A), yet commercial platforms remain proprietary, limiting transparent cross-product comparison. We developed a mechanistically anchored Computational Cell-Based Potency Assay calibrated solely from published validation data of the onabotulinumtoxinA SiMa CBPA. The toxin-cell interaction model reproduced the validated analytical envelope: simulated EC50 values clustered within 3.9-4.4 U/mL (mean 4.11), the median slope was 1.08, and signal limits aligned with the reported dynamic range. Within-run variability corresponded to 6.8-7.2 % CV and between-run variability to 8.9-9.4 % CV. A characteristic variance bowl emerged, with minimal dispersion near nominal potency and increasing uncertainty toward assay limits. To evaluate transferability, the intracellular mechanistic layer was held constant while the observation model was re-parameterized using published characteristics of the abobotulinumtoxinA BoCell reporter assay. Across simulated lots spanning 50-200 % potency, EC50 distributions and slope ranges matched published behaviour, with potency RMSE of 3.7-4.8 % for the SiMa simulation and 4.9-6.1 % for the reporter simulation. Cross-assay harmonization showed strong agreement: Deming slope 1.03 (95 % CI 1.00-1.06), Passing-Bablok slope 1.04, Bland-Altman bias 1.8 %, and concordance correlation coefficient 0.947. Incorporation of an exploratory Cell2Sentence embedding reduced RMSE by 15-25 % and revealed interpretable variance contributions at lot and run levels. This mechanistically structured, cross-assay validated framework provides a transparent in silico platform for assay optimisation and robustness analysis.