MERIT-ML — Metabolomics Workbench Readiness Assessment

Machine Learning Readiness for Tabular Metabolomics Data focused on Metabolomics Workbench.

MERIT-ML web app

Scoring ParametersDefault MERIT-ML

Change thresholds, then apply to recalculate the same study with updated scores, gates, bands, and tooltip text.

Default scope: MERIT-ML scores supervised classification and feature-selection readiness. Small-n designs such as triplicate time-course, cell-culture, or 13C-tracing experiments may be scientifically valid for their original aim, but they remain limited for reliable supervised ML training, validation, and feature selection.

Band labels

Diplomatic display names; internal cached labels are not edited.

ML-ready cutoffdefault 85

Minimum displayed core score for the strongest band before gate ceilings; higher values make the top label harder to reach.

Caveat cutoffdefault 70

Minimum displayed core score for ML-ready with caveats; raising it moves borderline studies into lower-use bands.

Exploratory cutoffdefault 50

Minimum displayed core score for exploratory ML use; below this, model training is treated as strongly class-support limited.

Structural and gates

Controls the feasibility gates that can cap the final band.

Preferred ML-eligible samplesdefault 20

Minimum ML-eligible sample count for the sample-size gate to pass. MERIT-ML defaults are for supervised classification / feature-selection reuse; triplicate time-course or isotope-tracing designs may be valid experimentally but remain underpowered for reliable ML validation.

Sample fail-below thresholddefault 10

ML-eligible sample count below this is treated as fail for the sample-size gate. This does not judge original-study validity; it flags that very small ML-eligible sample sets cannot support reliable supervised classifier training or feature selection.

Minimum class targetdefault 5

Smallest class size required for the class-support gate and label-suitability metric to pass; protects cross-validation from under-filled classes.

Class warning floordefault 3

Smallest class size needed to avoid a fail class-support gate; below this, minority-class learning is treated as infeasible for supervised ML reuse.

Missingness pass %default 50%

Median sample-level missingness at or below this passes the missingness gate; lower missingness means fewer imputation-driven signals.

Missingness fail %default 80%

Median sample-level missingness above this is treated as fail for the missingness gate; high missingness can dominate model behavior.

Label structure

Controls class-balance and group-support status/scoring.

Class balance pass scoredefault 40

Minimum smallest/largest class ratio for pass status; stricter balance reduces majority-class dominance during training.

Strong class supportdefault 20

Smallest class size that receives the full group-support score; supports stable folds and minority-class estimates.

Moderate class supportdefault 10

Smallest class size for the intermediate group-support score; indicates training is possible but validation is less stable.

Weak class supportdefault 5

Smallest class size for the weak group-support score; below this, classes are too sparse for dependable learning.

Entropy pass scoredefault 70

Minimum normalized class-label entropy for pass status; higher entropy means samples are more evenly distributed across labels.

ML task readiness

Controls the p/n score mapping for feature-to-sample ratio.

p/n low-riskdefault 10

Feature-to-sample ratio at or below this gets full score; low p/n reduces overfitting pressure.

p/n moderate-riskdefault 50

Feature-to-sample ratio at or below this gets an intermediate-high score; regularization is likely needed.

p/n high-riskdefault 200

Feature-to-sample ratio at or below this gets a reduced score; high dimensionality makes model selection fragile.

p/n tail denominatordefault 1000

Controls how sharply extremely high p/n ratios are penalized; larger values make ultra-high-dimensional datasets less harshly penalized.

Analytical QC

Controls status thresholds where cached raw distributions support safe v2 recalculation.

Missingness score passdefault 85

Minimum sample-missingness score for pass status; higher cutoffs demand more complete samples before training.

Class gap warning %default 10%

Marks a warning signal when missingness differs this much between classes; class-specific missingness can leak label information.

Outlier score passdefault 90

Minimum outlier-burden score for pass status; stricter values flag studies where unusual samples may drive the model.

Correlation score passdefault 85

Minimum redundancy score for pass status; stricter values flag feature blocks that can overweight duplicated signals.

Feature burden warning %default 10%

Marks a warning signal when too many features exceed the high-missingness threshold; high feature dropout increases imputation burden.

Annotation

Controls annotation/interoperability status thresholds.

General annotation passdefault 70

Minimum annotation/interoperability score for pass status; higher values demand clearer biological interpretation of model features.

Redundancy passdefault 85

Minimum feature-redundancy score for pass status; stricter values flag duplicated or repeated feature labels.

Unknown feature max %default 20%

Maximum allowed unknown-feature fraction; unknown features can still train models but weaken interpretation and reuse.

Bulk MERIT-ML Analysis

Build a study set from Find Similar Studies, save per-study matrix/threshold edits, then run one sortable report.

Selected studies: 0

No studies selected yet. Use Add to bulk in Find Similar Studies.

Tip: open a study, edit matrix labels or thresholds, then click Save current edits before moving to another study.

Run a MERIT-ML Assessment

Evaluate the machine-learning readiness of tabular metabolomics datasets from Metabolomics Workbench.

Workflow Ready

Enter a Metabolomics Workbench accession ID and run the pipeline. The report will display all readiness dimensions, a Readiness Score radar chart, per-source data availability, and per-metric recommendations.

Run a MERIT-ML Assessment

Repository & Accession

Workflow Ready