Why Machine Learning Fails At Mass Spectrometry For Small Molecules
Why machine learning fails at mass spectrometry for small molecules
Current situation
Most models are currently predicting a molecular fingerprint, then querying a molecular db, like pubchem, to find candidates. Useful for identifying compounds that have never been identified in MS, as the model can map any spectra to a fingerprint. Training occurs with commonly used datasets like NPLIB1 or MassSpecGym. The neural network workflow goes: 1) MS collected (m/z) 2) encoded 3) then embedded 4) run through neural network 5) predicted fingerprint obtained 6) fingerprint lookup in database 7) evaluation of database candidates Models evaluated, includig DreaMS and MIST performed poorly. For scaffolding splits, nearest neighbor outperforms MISt, and is near the performance of DreaMS.
Inability to generalize across experimental conditions
Current models fail to model changes in conditions impact spectra.
Inability to capture peak intensity
Models struggle with distinguishing similar m/s distributions (could this be overfit?). And intensity information is largely ignored.
Inability to generalize to new chemical formulas
Examples where the models fell short were also largely due to molecular fragments being present that were not present in the training data.