MelonnPan

Coming soon

Runner not wired yet. R-bridge (MelonnPan::melonnpan.predict()) required before this module can execute.

MelonnPan predicts metabolite levels from microbial community composition by fitting an elastic-net regression per metabolite against microbiome features on a reference cohort, then applying the trained model to a new sample's microbiome. The prediction is microbiome-only — no metabolomics required at inference time.

Comparator positioning

MelonnPan is the inference-from-statistics counterpart to GeMMA's inference-from-mechanism. Both output predicted metabolite profiles from microbiome composition, but:

• MelonnPan learns a regression model from a training cohort.
• GeMMA derives predictions from community-weighted GSMM pathway capacity — no training, reconstruction-first.

Paired with observed metabolomics, both can be validated by predicted vs measured scatter (per-metabolite R² / Spearman ρ). A study where GeMMA matches observed better than MelonnPan argues for mechanism-first framing; the reverse argues for training-data dependence.

Required inputs

• Microbial feature abundance table (preprocessed matrix).
• A published MelonnPan reference bundle (the Borenstein group distributes several; HMP, IBD, etc.).
• Target metabolite namespace — HMDB to align with GeMMA's synthesis chain and with the platform's global metabolite vocabulary.

Planned outputs

• Predicted metabolite table (rows × samples).
• Per-sample predicted-vs-observed scatter (R² + Spearman ρ).
• Per-metabolite heatmap across samples.
• Optional leave-one-out retrain when group sizes permit.