AllCCS2

AllCCS2 predicts collision cross-section (CCS) values for small molecules to support ion mobility–mass spectrometry analyses such as metabolomics, lipidomics, and exposome studies.

Key Features:

• Enhanced prediction accuracy: Trained on 10,384 experimental CCS records and 7,713 unified values, achieving median relative errors (MedRE) of 0.31%, 0.72%, and 1.64% on training, validation, and testing datasets, respectively.
• Advanced neural network architecture: Integrates mass spectrometry features, molecular descriptors, and graph features extracted using a graph convolutional network (GCN) within a neural network model.
• Broad instrument compatibility: Demonstrates compatibility with Drift Tube IMS (DTIMS), Traveling Wave IMS (TWIMS), and Trapped Ion Mobility Spectrometry (TIMS) platforms.
• Comprehensive uncertainty analysis: Assesses prediction uncertainty by analyzing structural similarity to training data and model prediction variation, linking higher similarity and lower variability to reduced relative errors.

Scientific Applications:

• Small-molecule identification and characterization: Provides CCS predictions to assist identification and structural characterization of small molecules in IM-MS datasets.
• Metabolomics, lipidomics, and exposome studies: Supports CCS-based annotation and confidence assessment in these IM-MS workflows.

Methodology:

Integrates 10,384 experimental CCS records and 7,713 unified values to train a neural network combining mass spectrometry features, molecular descriptors, and graph features extracted via a graph convolutional network (GCN), with analysis of structural similarity and model prediction variation for uncertainty assessment.