PUMA

PUMA models pathway activity likelihoods and assigns probabilistic metabolite annotations to interpret untargeted metabolomics data.

Key Features:

• Probabilistic Modeling Framework: Employs a generative model that integrates metabolomics measurements with a biological network and uses stochastic sampling to compute posterior probability distributions.
• Pathway Activity Prediction: Defines a pathway as active when the likelihood of generating the observed measurements exceeds a user-defined threshold, enabling probabilistic pathway activity calls.
• Metabolite Annotation: Produces probabilistic annotations that assign multiple putative chemical identities, including isomer candidates, extending beyond spectral database lookups.
• Validation and Performance: Validated on synthetic datasets that simulate cellular processes and on an experimentally validated dataset of 50 compounds, reporting precision of 0.833, recall of 0.676, and outperforming pathway enrichment analysis by an average of 8%.

Scientific Applications:

• Metabolic profiling: Interprets untargeted metabolomics measurements to profile metabolite-level changes.
• Pathway analysis: Identifies pathways with high likelihoods of generating observed metabolomic signatures.
• Disease mechanism investigation: Supports studies that link metabolite and pathway changes to disease processes.
• Drug discovery: Aids identification of pathway-level and metabolite-level targets relevant to pharmacology.
• Systems biology: Integrates metabolomics with network context to elucidate biochemical network behavior.
• Personalized medicine: Enables probabilistic interpretation of individual sample metabolomes for personalized insights.

Methodology:

Uses a generative probabilistic model that integrates metabolomics measurements with a sample-specific biological network, applies stochastic sampling to compute posterior probability distributions, defines pathway activity by likelihood exceeding a user-defined threshold, generates probabilistic metabolite annotations including multiple isomer candidates beyond spectral database lookups, and was validated using synthetic datasets and an experimentally validated 50-compound dataset.