Multitask learning for predicting pulmonary absorption of chemicals.

Area of Science:

• Pharmacology and Toxicology
• Computational Chemistry
• In Silico Modeling

Background:

• Pulmonary absorption is critical for drug delivery and chemical exposure assessment.
• Animal-free models like cell cultures and computational approaches are used, but ex vivo isolated perfused lung (IPL) models offer higher clinical relevance.
• Current IPL experiments are resource-intensive, limiting large-scale screening of inhaled substances.

Purpose of the Study:

• To develop a novel machine learning (ML) method for predicting the absorption rate constant (ka_IPL) in isolated perfused lungs (IPL).
• To leverage multitask learning to extract shared permeability knowledge from Caco-2 and Calu-3 cell permeability data alongside IPL absorption.
• To identify key physicochemical descriptors influencing pulmonary absorption.

Main Methods:

• An extratrees-based multitask learning approach was employed to predict ka_IPL.
• Simultaneous learning of three tasks: Caco-2 cell permeability, Calu-3 cell permeability, and IPL absorption rate.
• Identification and utilization of seven informative physicochemical descriptors.

Main Results:

• The developed ML model demonstrated good predictive performance.
• A high correlation (r = 0.84) was observed between predicted and observed ka_IPL values in an independent test dataset.
• Seven key physicochemical descriptors were identified as informative for pulmonary absorption prediction.

Conclusions:

• The novel multitask learning model effectively predicts pulmonary absorption rates using IPL data.
• The model shows potential as a valuable tool for screening inhaled drugs and toxicants, supporting the reduction of animal testing.
• Case studies on inhalation drugs and respiratory sensitizers highlight the practical applicability of the developed in silico method.