MS2DeepScore

MS2DeepScore predicts structural similarity between chemical compounds from MS/MS fragmentation spectra using deep learning.

Key Features:

• Siamese Neural Network Architecture: Employs a Siamese neural network trained to predict molecular structural similarity scores (Tanimoto scores) from pairs of mass spectrometry spectra, enabling direct spectrum-to-spectrum comparisons.
• Training Dataset: Trained on an extensive dataset comprising over 100,000 mass spectra from approximately 15,000 unique known compounds.
• Prediction Accuracy and Uncertainty Estimation: Evaluated on 3,600 spectra from 500 unseen compounds with a reported RMSE of ~0.15 for Tanimoto score predictions and using Monte-Carlo Dropout to sample model variants and estimate prediction uncertainty, allowing selection of predictions with lower RMSE (~0.1).
• Performance Comparison: Outperforms traditional spectral similarity measures for retrieving chemically related compound pairs from large datasets.
• Spectral Embeddings and Clustering: Generates chemically meaningful mass spectral embeddings that facilitate clustering of large numbers of spectra.

Scientific Applications:

• Spectral Library Matching: Improves matching of experimental spectra to reference libraries by predicting structural similarity from MS/MS spectra.
• Metabolomics Data Processing: Provides reliable similarity metrics for large-scale metabolomics spectral analysis.
• Chemical Structure Elucidation: Assists in inferring chemical structural relationships from fragmentation spectra via predicted Tanimoto similarities.

Methodology:

Uses a Siamese neural network trained to predict Tanimoto similarity scores from pairs of MS/MS spectra using a training set of >100,000 spectra from ~15,000 compounds; evaluated on 3,600 spectra from 500 unseen compounds (RMSE ≈0.15); applies Monte-Carlo Dropout to sample model variants and estimate prediction uncertainty; produces spectral embeddings for clustering.