RISKmilk Combines Machine Learning and PBPK to Predict Drug Transfer into Milk
A structure-based framework improved milk-to-plasma ratio predictions against a conventional lactation model. Its screening results separate predicted milk enrichment from relative infant dose and pharmacological risk.
Estimating Transfer Parameters from Molecular Structure
Researchers at the University of Macau, Guangzhou Medical University and the University of California, San Francisco developed RISKmilk to estimate drug transfer into human milk when clinical concentration data are sparse.
The framework uses molecular fingerprints and physicochemical descriptors to predict a milk-to-plasma partition coefficient and a permeability–surface-area parameter. Support vector regression was selected for implementation, and the predicted inputs were incorporated into a whole-body physiologically based pharmacokinetic (PBPK) model that represents changes in postpartum physiology.
Comparing Predictions with Clinical Ratios
Against observed clinical milk-to-plasma ratios, 66.4% of RISKmilk predictions fell within twofold error, compared with 29.6% for a conventional equation-based lactation PBPK benchmark. The corresponding proportions within threefold error were 77.6% and 45.6%; within fivefold error, they were 84.0% and 57.6%.
Sensitivity analysis identified maternal systemic clearance, the blood-to-plasma ratio and milk partitioning as major determinants of milk exposure. In a marketed-drug screen, 14 compounds reached the model’s milk-to-plasma enrichment threshold of at least 1. Four of these exceeded a relative infant dose of 10% in the maximum-exposure scenario.
The authors interpret milk concentration, infant dose and drug pharmacology together when prioritizing compounds for further assessment. Their framework also screened herbal constituents and food additives, while the drug-development application concerns predicted medicine exposure during lactation.