Graph Machine Learning Can Estimate Drug Concentrations in Whole Blood from Forensic Screening Results.

Area of Science:

• Forensic Toxicology
• Computational Chemistry
• Machine Learning

Background:

• Liquid Chromatography-High-Resolution Mass Spectrometry (LC-HRMS) is vital for broad-scope screening in forensic toxicology.
• Quantifying novel or rare compounds identified via LC-HRMS is challenging due to the cost and time associated with acquiring reference materials.
• Accurate quantification is crucial for determining the relevance of a compound to a toxicological case.

Purpose of the Study:

• To develop and evaluate a machine learning (ML) framework for estimating drug concentrations in whole blood.
• To utilize molecular structure information and LC-HRMS signals for drug concentration estimation.
• To provide a rapid, semiquantitative method for prioritizing compounds exceeding toxic thresholds.

Main Methods:

• Trained and evaluated various ML models, including Random Forests and Graph Neural Networks (GNNs), using a dataset of 191 drugs spiked into whole blood.
• Leveraged atomic features and global molecular properties within the GNN architecture.
• Validated the GNN model's performance using cross-validation and an external benchmark dataset of ionization efficiencies (logIE).

Main Results:

• The Graph Neural Network (GNN) model demonstrated superior accuracy compared to standard ML models.
• The GNN achieved accurate concentration estimates (within 50-200% of true values) for 79% of molecules.
• The GNN model outperformed the state-of-the-art on an external dataset for ionization efficiencies.

Conclusions:

• Graph-based machine learning offers a feasible approach for estimating drug concentrations in whole blood without requiring reference materials.
• This ML framework serves as a practical tool to support toxicological decision-making, especially for emerging or rare compounds.
• The open-source GNN model and publicly available dataset facilitate further research and application in forensic toxicology.