# Similarity Backends

`psma` currently supports three ways to build the similarity matrix.

## RDKit Morgan Tanimoto

This backend uses SMILES strings or fingerprints to build Morgan
fingerprints and then computes pairwise Tanimoto similarity.

Use it when:

- you have SMILES
- you want a standard cheminformatics baseline

## Embedding cosine

This backend uses dense vector embeddings and computes cosine
similarity.

Use it when:

- you already have learned molecular embeddings
- you want to compare representation-learning workflows

## Imported triples

This backend reconstructs the full similarity matrix from a sparse
triples table containing:

- first molecule id
- second molecule id
- similarity score

Use it when:

- the similarities were computed elsewhere
- you want the package to consume a precomputed similarity definition

## Choosing between them

Start with the backend that matches the representation you already have.
The downstream PSMA surface workflow is the same once the similarity
matrix exists.