Distance Metrics

Once images are converted to embeddings, we measure mathematical distance between them to quantify phenotypic change.

• Low distance = similar to the reference (e.g. no treatment effect).
• High distance = different from the reference (e.g. strong phenotypic change).

Common methods

• Euclidean vs cosine: Euclidean measures absolute straight-line distance; cosine measures the angle between vectors (often better for high-dimensional embeddings because it ignores magnitude).
• Centroid vs all-to-all: compare each image to the average (centroid) of the reference group (fast and stable), or to every individual reference image (more sensitive but slower).

For a complete guide, see Common workflows.