# ailab-bio/RetroTAC

Caruana-weighted ensemble of 27 per-fold PROTAC
synthesizability surrogate models (GNN:13, MLP:7, XGB:7), selected by greedy
forward ensemble selection over 5x5-CV fold models -- see
`scripts/models/evaluation.py`'s `EnsembleSelector.greedy_selection` /
`run_ensemble_strategies` in the [RetroTAC](https://github.com/ribesstefano/RetroTAC) repo.

- method: `caruana`
- rmse: 0.1320

## Usage

```python
from retrotac.models.ensemble import RetroTAC

model = RetroTAC.from_pretrained("ailab-bio/RetroTAC")
predictions = model.predict(["CCO", "c1ccccc1"])
```