# PSMA Workflow

The package implements a probabilistic molecular activity surface
workflow in a sequence of small, testable steps.

## High-level flow

1. Validate input data and backend-specific columns.
2. Build a similarity matrix.
3. Split compounds into train and test sets.
4. Transform similarities into distances.
5. Embed the training compounds into a 2D reference space.
6. Threshold the endpoint into binary labels.
7. Fit class-conditional KDE models.
8. Evaluate density surfaces on a regular grid.
9. Convert densities plus priors into a posterior surface.
10. Project test compounds into the learned space.
11. Score test points and compute metrics.

## Why the workflow is structured this way

The package mirrors the paper’s workflow while keeping each stage
independently testable. That makes it easier to:

- verify parity with the original method
- harden failure handling
- expose a clear public API

## Key constraints

- the package assumes a 2D reference space
- the workflow needs both classes present in the training set for KDE
- the projection step uses a pseudoinverse-based mapping

These constraints are important to understand when interpreting results.