Detective-QSAR

Detective-QSAR predicts physicochemical properties and toxicities from mass spectrometry (MS) spectra and gas chromatography-mass spectrometry (GC-MS) retention indices to enable QSAR assessment without explicit chemical structures.

Key Features:

• Input descriptors: Uses analytical descriptors derived from MS spectra and retention indices obtained via GC-MS.
• Structure-independent QSAR: Implements QSAR predictions that do not require explicit chemical structures or traditional molecular descriptors.
• Machine learning algorithm: Employs XGBoost as the machine learning method for model development.
• Predicted endpoints: Predicts log K_o-w, molecular weight, melting point, boiling point, vapor pressure, water solubility, and LD50 values in rats and mice.
• Performance metrics: Reported root-mean-square errors (RMSEs) are 0.97 (log K_o-w), 0.052 (molecular weight), 51 (melting point), 23 (boiling point), 0.74 (vapor pressure), 1.1 (water solubility), 0.74 (LD50 in rats), and 0.6 (LD50 in mice).
• Spectral deconvolution: Performs spectral deconvolution to analyze unknown-structured chemicals in complex samples.
• Validation data: Validated using chemical standard mixtures and contaminated oil samples.
• Addressing database limitations: Designed to mitigate limitations of traditional QSAR when structural data are missing in large databases such as PubChem.

Scientific Applications:

• Environmental monitoring: Enables assessment of unknown compounds in environmental samples using MS and GC-MS data.
• Chemical safety assessment: Supports evaluation of physicochemical properties and toxicities for chemical safety and risk assessment.
• Contaminated-sample analysis: Applies spectral deconvolution and prediction to contaminated oil samples and other complex mixtures.
• Screening of unidentified chemicals: Facilitates research and screening of chemicals lacking structural information in databases and sample collections such as PubChem.

Methodology:

Analytical descriptors from MS spectra and GC-MS retention indices are used as input to XGBoost models, and spectral deconvolution is applied for analysis of unknown-structured chemicals.