LSM1-MS2: A Foundation Model for MS/MS, Encompassing Chemical Property Predictions, Search and de novo Generation

We present LSM1-MS2, a pre-trained self-supervised foundation model designed for tandem mass spectrometry (MS/MS) utilizing a transformer architecture with custom tokenization for masked MS2 peak reconstruction. Our model is fine-tuned on smaller, labeled datasets for tasks such as compound property prediction, spectral matching, and de novo molecular generation. LSM1-MS2 demonstrates superior performance compared to traditional supervised models, achieving high accuracy with minimal labeled data. It outperforms conventional methods in database lookups and molecular query retrievals and shows promising results in the opening field of de novo molecular generation. The model's efficiency in spectral lookup tasks, with significantly reduced evaluation times, underscores its potential for large-scale applications. Our findings highlight the transformative capability of self-supervised pre-training in enhancing the predictive power of models for mass spectrometry, particularly in data-limited scenarios. The success of LSM1-MS2 in property prediction, database spectral lookup, and molecular generation paves the way for its application in metabolomics and drug discovery, facilitating robust and scalable analysis with reduced data requirements.