FIDDLE: a deep learning method for chemical formulas prediction from tandem mass spectra

Area of Science:

• Computational chemistry
• Mass spectrometry
• Bioinformatics

Background:

• Molecular identification via tandem mass spectrometry is crucial for small molecule analysis.
• Current computational formula identification methods face challenges in accuracy, speed, and scalability, especially for larger molecules, hindering high-throughput workflows.

Purpose of the Study:

• To develop and evaluate FIDDLE (Formula IDentification by Deep LEarning), a novel deep learning-based method for accelerating and improving molecular formula identification.
• To assess FIDDLE's performance against state-of-the-art methods on diverse mass spectrometry datasets.

Main Methods:

• Developed FIDDLE, a deep learning model trained on over 38,000 molecules and 1 million MS/MS spectra from Q-TOF and Orbitrap instruments.
• Evaluated FIDDLE's speed and accuracy in formula identification.
• Compared FIDDLE's performance against top-down (SIRIUS) and bottom-up (BUDDY) computational approaches.

Main Results:

• FIDDLE accelerates formula identification by over 10-fold compared to existing methods.
• Achieved top-1 accuracy of 88.3% and top-5 accuracy of 93.6% on training data.
• Outperformed SIRIUS and BUDDY by over 10% in accuracy.
• On external metabolomics datasets, FIDDLE achieved top-5 accuracies of 75.1% (positive ion mode) and 66.2% (negative ion mode), improving to 80.0% and 73.8% when combined with SIRIUS and BUDDY.

Conclusions:

• FIDDLE offers a significant advancement in computational molecular formula identification.
• The deep learning approach enhances both the speed and accuracy of high-throughput mass spectrometry-based analyses.
• FIDDLE shows promise for improving metabolomics research and other small molecule identification workflows.