Machine learning and AI in MS metabolomics

Terms you will encounter

Predictive machine learning

Learns a relationship from examples, such as molecular structure to fragmentation spectrum or MS/MS to chemical class. Evaluate the specific prediction task and model applicability.

Spectral representation learning

Learns numerical representations of spectra for tasks such as similarity search. DreaMS uses self-supervised learning; a similar representation is not proof of identical structure.

Generative models

Generate candidates, such as molecular structures or fragmentation graphs. Generation does not establish that a proposed molecule is present in the sample.

Foundation models

Broadly pretrained models adapted to multiple downstream tasks. This describes training and reuse, not a guarantee of reliability for every instrument or sample type.

Agentic AI

A model selects or coordinates analysis tools over multiple steps. Inspect the actual tools, parameters, intermediate evidence and human review points. An agent can call both learned models and conventional software.

Workflow automation

Runs processing steps automatically. Automation alone does not imply learned inference or an AI agent. Everything Bagel is currently documented here as automated feature finding; use within an agentic workflow does not change that classification.

These descriptions overlap. They describe how a method works, while the catalogue’s scientific functions describe what it helps you do.

Browse tools tagged with machine learning →

Method papers

• DreaMS — self-supervised spectral representations
• ICEBERG — neural fragmentation-graph generation
• MSNovelist — de novo structure generation