Measurement Classification: Phenotypes vs. Features#

PhenoTypic measures many columns per colony. This page explains how to apply each one — which numbers you can report directly as a biological result, and which are best used together as inputs to classification or clustering — without needing the underlying math.

Two questions place every measurement#

  • Interpretability — does the number name a biological thing (a diameter, a pigment, biomass), or is it a mathematical descriptor (a texture value, a colour coordinate)?

  • Analytical role — do you use it as a result (quantify an effect), or as a feature (feed many of them into a classifier/clustering)?

Four kinds, then three tiers#

Every column is first one of four kinds:

  • Identity / design factors — the variables you analyse against (metadata, locators). Not outcomes.

  • Quality — gates whether to trust a row/plate. Never a biological claim.

  • Primary measurement — the measured signal. These get a tier (below).

  • Derived / model output — computed from primary measurements; classified by how they were derived.

Primary measurements fall on a three-tier spectrum:

Tier

Name

What it is

How to apply it

1

Direct phenotype

A real biological quantity with units/meaning (size, intensity/opacity)

Report a single value as a result; compare across conditions; dose–response.

2

Descriptive trait

A named, interpretable form/colour property, usually unitless (shape descriptors, Lab/HSV colour, radial/zone structure)

Interpret the direction of change against a control; also good clustering input.

3

Discriminative feature

A mathematical fingerprint with no single biological meaning (texture, XYZ/xy/composition colour)

Don’t read one value; use the whole block together for classification/clustering.

The trust contract#

The tier is a promise about what a single number licenses:

  • Tier 1 — pre-validated for direct biological claims; safe to report alone.

  • Tier 2 — interpret directionally, anchored to a control.

  • Tier 3 — make no single-value biological claim; its job is discrimination, judged by how well groups separate.

Derived outputs inherit by how they were made#

A model fit on a primary phenotype is classified by its transformation:

  • Parameterization (e.g. logistic/softplus growth: growth rate, lag, carrying capacity) → same tier as the input phenotype. Colony size and growth rate are interchangeable fitness proxies, so growth parameters are Tier 1.

  • Normalization (e.g. edge correction) → the input’s tier, cleaned.

  • Fit diagnostics (R², RMSE, optimizer state, regularization knobs) → Quality.

See also: Measurement metrics and their biological meaning for per-metric detail, and the measurement reference for the Use/Tier badge on every column.