Machine learning-based prediction of drug lactation risk: Bridging molecular features and breastfeeding safety

Clinical Snapshot

Key Findings

• P Value: Not reported

• Effect Size: AUC = 0.80 for overall classification, AUC = 0.85 for extreme categories (L1 vs L5)

• Primary Outcome: Lactation risk category prediction accuracy

• Nnt Or Sensitivity: MCC = 0.52 overall, MCC = 0.62 for L1 vs L5 discrimination

• Confidence Interval: Not reported

Clinical Application

High - web-based platform requires only molecular structure input, no complex clinical data Highly relevant for Australian practice - complements TGA pregnancy/lactation categories and supports RACGP guidelines for medication safety during breastfeeding. Could integrate with PBS prescribing decisions. Healthcare providers prescribing medications to breastfeeding mothers, pharmaceutical researchers developing new drugs

Abstract

BACKGROUND: Breastfeeding medication safety assessment presents critical challenges due to limited clinical evidence and ethical constraints on lactation studies. While existing computational models focus on predicting pharmacokinetic surrogates such as milk-to-plasma ratios, these endpoints lack direct clinical actionability for therapeutic decision-making. METHODS: We developed LRCpredictor, a computational framework for predicting drug lactation risk based on Dr. Thomas W. Hale's evidence-based Lactation Risk Categories (LRC) system. A dataset of 391 drugs (179 high-risk L4/L5, 212 low-risk L1/L2) was characterized using three complementary molecular representations. Systematic feature engineering with four selection algorithms identified optimal feature subsets. Five ensemble machine learning algorithms were evaluated through cross-validation and independent testing. Multi-level SHAP analysis and structural alert mining were employed to elucidate molecular determinants of lactation risk. RESULTS: The optimal Gradient Boosting Decision Tree (GBDT) model using 35 selected features achieved robust cross-validated performance (AUC = 0.80, MCC = 0.52) and demonstrated notably superior discrimination for extreme risk categories (L1 vs. L5: AUC = 0.85, MCC = 0.62). Analysis revealed lactation risk is governed by the complex interplay of electronic properties, structural topology, electrotopological characteristics, polarizability and drug-likeness attributes. Matched molecular pair analysis demonstrated how specific structural modifications translate into interpretable risk changes. We further identified 18 structural alerts, including five exclusive to high-risk drugs, enabling rapid toxicophore recognition during risk assessment and drug design. CONCLUSIONS: LRCpredictor provides the first interpretable framework for clinically actionable lactation risk prediction with mechanistic molecular insights. The freely accessible web platform (https://lrcpredictor.streamlit.app/) with integrated visual explanations supports evidence-based decision-making for healthcare providers and drug design optimization.

References

1. 1.Liu, P., Huang, S., Xie, X., Chen, J., Wu, S., Huang, L., & Huang, X. (2026). Machine learning-based prediction of drug lactation risk: Bridging molecular features and breastfeeding safety. European Journal of Medicinal Chemistry, 118794. https://doi.org/10.1016/j.ejmech.2026.118794