Pre-trained Maldi Transformers improve MALDI-TOF MS-based prediction.

Area of Science:

• Computational Biology
• Analytical Chemistry
• Machine Learning

Background:

• Matrix-assisted laser desorption/ionization time-of-flight mass spectrometry (MALDI-TOF MS) is the standard for microbial species identification.
• Machine learning models for MALDI-TOF MS are underdeveloped due to limited open data.
• Increasing data complexity necessitates advanced, specialized models for MALDI-TOF MS.

Purpose of the Study:

• To develop a machine learning model tailored for MALDI-TOF MS data.
• To introduce a novel self-supervised pre-training method for mass spectral data.
• To enhance predictive performance on downstream tasks like species identification and antimicrobial resistance prediction.

Main Methods:

• Adaptation of the transformer architecture for MALDI-TOF mass spectral data, named Maldi Transformer.
• Development of a novel self-supervised pre-training technique involving peak shuffling and peak discrimination.
• Extensive benchmarking to validate model efficacy on various prediction tasks.

Main Results:

• Maldi Transformer demonstrates state-of-the-art or competitive performance on downstream prediction tasks.
• The proposed self-supervised pre-training method is effective for mass spectral data.
• Leveraging noisy spectra identification improves overall predictive performance.

Conclusions:

• Maldi Transformer offers a powerful, specialized deep learning approach for MALDI-TOF MS data analysis.
• The novel pre-training strategy advances machine learning applications in mass spectrometry.
• The model shows promise for improving clinical microbiology diagnostics and antimicrobial resistance surveillance.