An AI-agent-orchestrated grey-box Transformer framework for sparse pharmacokinetic curve reconstruction and pharmacometric model initialization

Abstract

Clinical pharmacokinetic (PK) modelling is constrained by sparse sampling, limited general-isability of single-drug models, and labour-intensive workflows, making it difficult to infer complete drug exposure from limited concentration observations. We present the Pharmacokinetic Foundation Model (PKFM), a grey-box Transformer framework pre-trained across 32 drugs that reconstructs concentration-time profiles from sparse concentration observations, dosing events, molecular descriptors, and physiological covariates while preserving output interpretability. In representative oral PK curves, three sparse input points recovered the principal absorption-elimination trajectory, achieving coefficient of determination (R2) = 0.992 for Midazolam oral and R2 = 0.990 for Verapamil oral. Using reconstructed curves in NONMEM (nonlinear mixed-effects modelling) improved covariance stability and individual prediction accuracy. Contrastive-learning embeddings supported Top-10 physiologically based pharmacokinetic (PBPK) candidate retrieval, with 75.6% of observations within the 2-fold range. A pharmacometrics-informed AI Agent (PM Agent) outperformed general-purpose programming tools in stability and pairwise win rate on a standardised modelling benchmark, with each run requiring human pharmaco-metrician confirmation before downstream use. These results support cross-drug pre-trained PK models as an information-completion layer for sparse PK evidence and a structured scaffold for the modelling workflow; clinical or regulatory use requires prospective validation, broader external benchmarking, and independent expert assessment.