# Input Data

The package accepts a main dataframe plus, for one backend, an optional
triples dataframe.

## Main dataframe

Always required:

- endpoint column selected by `y_col`

Optional but commonly used:

- `mol_id`
- SMILES column
- fingerprint column
- embedding column

## Backend-specific requirements

### `rdkit_morgan_tanimoto`

Requires either:

- a SMILES column
- or a fingerprint column

### `embedding_cosine`

Requires:

- an embedding column containing vector-like values

### `nams_triples_import`

Requires:

- the main dataframe
- a triples dataframe or CSV with pairwise similarity rows

## Label semantics

The package thresholds `y_col` into binary labels using:

- `label_threshold`
- `label_direction`

For already-binary endpoints, a threshold such as `0.5` with
`label_direction="ge"` preserves the existing `0/1` semantics.