Core concepts

MPDR

MPDR means Microbiota profile data representation.

An MPDR defines how a microbiome profile is represented before a learner is fitted. In a standard microbiome sweep this combines a taxonomic resolution with an abundance transformation, optionally including training-fitted filtering such as prevalence filtering.

Examples include:

genus + identity
genus + joint CLR
class-order + presence/absence
species + relative abundance + prevalence filter

Any operation that learns parameters from observations is fitted on the appropriate training partition.

MPMA

MPMA means Microbiota profile modelling algorithm.

An MPMA combines an MPDR with a learner. If three MPDRs are evaluated with four learners, the sweep contains twelve MPMA candidates before exclusions or modality-specific expansions.

MPMA-B

MPMA-B is the selected base microbiota profile modelling algorithm.

For nested evaluation, selection is based on inner validation evidence within each outer training partition. The corresponding outer test partition remains held out for performance estimation.

MPMA-E

MPMA-E is the selected ensemble microbiota profile modelling algorithm.

Candidate membership, ensemble strategy, and aggregation are selected from training-side or inner out-of-fold evidence. Outer-test performance is used for evaluation, not for choosing the ensemble.

OOF

Out-of-fold (OOF) means that a prediction or explanation for an observation is generated by a model that was not fitted on that observation.

OOF predictions underpin performance estimation and are also used by the held-out explainability workflow.

LODO

Leave-one-dataset-out (LODO) evaluation holds out one dataset or cohort at a time. It is useful when the primary question concerns generalization across studies, sites, or cohorts rather than random within-cohort resampling.

Selection versus evaluation

The distinction is central:

inner validation → choose representation / learner / ensemble
outer test       → estimate generalization performance

A configuration can contain many candidate models, but the held-out outer observations must not participate in choosing among them.