Transformer-Based PK/PD Modeling for Monoclonal Antibody Dosing from Sparse Concentration Data Download PDF

Monoclonal antibodies are central to modern biologic therapy for inflammatory, oncologic, and immune-mediated diseases, yet their pharmacokinetics are often nonlinear, highly variable between patients, and challenging to individualize when only limited concentration measurements are available. Traditional population PK/PD modeling and Bayesian forecasting rely on predefined structural models and assumptions regarding clearance, distribution, and variability, which can be fragile when data consist of sparse or irregularly timed therapeutic drug monitoring samples. To address these limitations, this MDL article proposes a transformer-based sequence model for individualized monoclonal antibody dosing, designed to infer patient-specific exposure patterns from sparse concentration data, dosing history, and clinical covariates without requiring an explicit compartmental model. The model employs a transformer encoder to process variable-length sequences of time, concentration, dose, and covariate tuples, using self-attention to capture relationships among prior doses, measured concentrations, elapsed time, and patient factors, thereby generating predicted future concentration trajectories and dose recommendations aimed at a target exposure window. Conceptually, this approach enables individualized concentration forecasts and dose suggestions even with minimal concentration data, complementing Bayesian forecasting by learning flexible temporal patterns that are difficult to specify parametrically. By extending model-informed precision dosing to data-sparse biologic treatment settings, a transformer-based framework could enhance the clinical utility of therapeutic drug monitoring for monoclonal antibodies while maintaining the need for prospective validation and clinical oversight