Quick examples

Run these commands from the repository root.

Project a molecule

uv run python scripts/examples/projection.py --smiles "COc1ccc(-c2ccnc(Nc3ccccc3)n2)cc1"

The script samples 64 pathways, ranks their products by ECFP4 Tanimoto similarity, and prints the top 10 as YAML. The first run downloads the default model and chemical space.

Useful options:

Option	Default	Purpose
`--num-samples`	`64`	Number of pathways sampled per target
`--top`	`10`	Maximum number of products printed
`--device`	`cuda`	PyTorch device, such as `cuda` or `cpu`
`--config`	Default released model YAML	Model and chemical-space configuration

To render pathways, install Graphviz and add an output directory:

uv run python scripts/examples/projection.py \
    --smiles "COc1ccc(-c2ccnc(Nc3ccccc3)n2)cc1" \
    --draw-output-dir ./draw

Images are written as synthesis_<rank>_sim<similarity>.png.

Projection example

Run the molecular sampler

The current sampler uses a genetic algorithm over ECFP4 fingerprints. The shipped example maximizes RDKit QED:

uv run python scripts/examples/sampling.py --out-fig ./sampling.png

It initializes a population, runs 20 generations, and prints the best and mean fitness at each step. --out-fig is optional and requires Graphviz. Use --device cpu only when CUDA is unavailable.

To optimize another objective, see Sampling molecules using your scoring function.