Prediction QC dashboard
A model-internal quality check on the metabolite predictions made by the microbiome reconstructions. It flags predictions a model makes but cannot mechanistically support — likely artifacts of the reconstruction(e.g. orphan or gap-filled exchanges).
What predictions are analysed?
Every metabolite-exchange prediction made by a microbiome reconstruction — each row is one organism → metabolite → direction triple, where direction is whether the organism is predicted to produce (secrete) or consume (take up) that metabolite. These come from constraint-based (FBA) simulation of two independent, curated reconstruction resources: AGORA2 and APOLLO (legacy AGORA1 is excluded). The header shows how many were scored.
How quality is judged — and how it is not
The verdict is assessed inside each reconstruction: does the model actually contain a metabolic reaction (ideally gene-associated) that produces or consumes the metabolite, or is it only moving it across the cell boundary with a transport/exchange reaction? It is the same engine you see when you open a single prediction in the prediction investigator.
We do not compare APOLLO against AGORA2 to judge quality. They are independent resources covering different organisms and were not built to validate each other, so whether one predicts something the other does not is not evidence of error. (Whether the other resource also predicts it is shown in the investigator as context only.)
The five verdicts
- Well supportedConfirmed by curated lab evidence, or backed by a gene-associated internal metabolic reaction.
- Plausible — partially supportedAn internal metabolic reaction involves the metabolite, but it is not gene-associated.
- Inconclusive — needs reviewNot enough internal evidence to call either way.
- Likely artifact — unsupported exchangeThe model exchanges the metabolite with no internal metabolic reaction to support it (see below).
- Not predicted by this reconstructionThis reconstruction does not actually make the prediction.
How is an “artifact” defined?
“Likely artifact — unsupported exchange” means the reconstruction is predicted to secrete or take up the metabolite, yet it has no internal metabolic reaction that produces or consumes it — only a transport/exchange reaction (or none at all). The exchange has no metabolic basis in the model, which is a model-internal inconsistency, typically an orphan or gap-filled exchange. This is judged purely from the reconstruction's own reactions, independent of the other resource. Curated experimental evidence overrides it (then it is a model gap, not an artifact). It is a flag for review, not proof of error — open any row in the investigator for the full reasoning.
Predicted / associative — VMH is a research knowledge resource, not a medical device.
Loading…
Worst-offending metabolites
Metabolites with the highest share of artifact verdicts. n = predictions for that metabolite, % artifact = how many were flagged. Click a metabolite to filter the table below.
| Metabolite | n | artifact | % artifact ▼ | avg score |
|---|
Worst-offending clades (phylum)
Bacterial phyla whose predictions are most often flagged as unsupported exchanges. Click a phylum to filter the table below.
| Phylum | n | artifact | % artifact ▼ |
|---|
Predictions
One row per prediction. Dir: → = produces (secretes), ← = consumes (takes up). Score: higher = more supported, negative = likely artifact. Click a column header to sort; click investigate ↗ to see the full evidence dossier and reasoning for that prediction.
| Organism | Metabolite | Dir | Source | Phylum | Score ▲ | Verdict | Conf | |
|---|---|---|---|---|---|---|---|---|
| No predictions match. | ||||||||
Curator flag queue (high-confidence artifacts)
A worklist of the most confident artifact calls for a curator to review: open it in the investigator, then confirm artifact (records it as a known false positive) or dismiss (not an artifact). Acting on a row removes it from the queue.
Queue empty.
Source: Supplementary Table of Heinken A, Hulshof TO, Nap B, … Thiele I. A genome-scale metabolic reconstruction resource of 247,092 diverse human microbes… Cell Systems 2025;16(2):101196. PMID: 39947184.