A RISKmilk framework for predicting drug and chemical exposure in human milk during breastfeeding

Medication decisions during breastfeeding are often made with limited quantitative evidence, partly because lactating women remain underrepresented in pharmacokinetic studies. Current assessments rely largely on sparse milk concentration data, empirical milk-to-plasma ratios, and fragmented case reports, providing limited capacity for systematic, quantitative, and mechanistic evaluation of lactational drug exposure and infant safety. We constructed a structure-driven, AI-enabled physiologically based pharmacokinetic (PBPK) framework, RISKmilk, to predict drug and chemical transfer into milk from molecular structure and support large-scale exposure prioritization during breastfeeding.

A mechanistic plasma-milk transfer model was used to derive the milk-to-plasma partition coefficient, K_p,milk, and the permeability-surface area product, PS. Molecular fingerprints and physicochemical descriptors were then used to train machine-learning models, with support vector regression selected for downstream implementation. AI-predicted K_p,milk and PS were incorporated into a lactation-extended whole-body PBPK model with time-resolved postpartum physiology. Our model performance was evaluated against observed clinical M/P ratios and compared with a conventional equation-based lactation PBPK benchmark. The validated framework was then applied to marketed drugs, herbal medicines, and food additives. For marketed drugs with predicted M/P ≥ 1, average and maximum relative infant dose (RID) were further calculated.

The RISKmilk framework substantially improved clinical M/P prediction compared with the conventional equation-based benchmark. The proportion of predictions within twofold error increased from 29.6% to 66.4%, with corresponding improvements within threefold error from 45.6% to 77.6% and within fivefold error from 57.6% to 84.0%. Sensitivity analysis identified maternal systemic clearance, blood-to-plasma ratio, and K_p,milk as the dominant determinants of milk exposure. In large-scale screening, most marketed drugs were predicted to enter milk to some extent, and 14 exceeded the M/P ≥ 1 enrichment threshold. Among these 14 drugs, four exceeded the 10% threshold under the maximum RID scenario. Together with reported clinical adverse-event data, these findings indicate that lactation safety assessment should integrate M/P, RID, and drug pharmacology rather than rely on M/P alone. Screening of herbal constituents and food additives further showed that the framework can be applied beyond conventional pharmaceuticals to identify high-risk compounds for lactation exposure assessment.

The structure-driven, RISKmilk framework was established to improve clinical M/P prediction and supports more refined prioritization of lactational drug exposure and infant safety. By integrating PBPK-derived milk exposure, relative infant dose, and pharmacological context, this approach offers a scalable digital framework for lactation pharmacology when direct clinical evidence is sparse.