DeepReI

DeepReI predicts gas chromatographic retention indices from molecular structures using deep learning to support analyte identification in non-targeted gas chromatography.

Key Features:

• Predictive Model: Employs a deep learning model to compute retention indices in silico from suggested molecular structures.
• Input Format: Utilizes the Simplified Molecular Input Entry System (SMILES) as its molecular input format.
• Model Architecture: Implements 2D-convolutional layers with batch normalization, max pooling, dropout, and three residual connections.
• Performance Metrics: Reports median absolute errors of 16.4 retention index units on validation and 16.0 units on test sets, with a median percentage error ≤ 0.81% across datasets.
• Stationary Phase Specificity: Architecture and training are tailored to capture molecular features relevant to gas chromatography on semi-standard non-polar stationary phases.

Scientific Applications:

• Non-targeted gas chromatographic analysis: Predicts retention indices for compounds absent from existing retention index libraries to aid identification in non-targeted workflows.
• Analyte identification and interpretation: Facilitates more accurate analyte identification and improves the reliability of chromatographic data interpretation through predicted retention indices.

Methodology:

Trains a deep learning model on existing retention index data using SMILES inputs; model architecture includes 2D-convolutional layers, batch normalization, max pooling, dropout, and three residual connections, and is tailored to capture molecular features relevant to gas chromatography on semi-standard non-polar stationary phases.