GRASSY-Net

A novel representation-first approach to molecular graph generation.

GRASSY-Net — Usage: data preparation → learnable scattering → GRASSY training → latent visualization

Prereqs

• Python 3.8+
• numpy, pandas, scikit-learn (or umap-learn), matplotlib
• repository scripts: train_learnable_scattering.py, train_grassy.py (expecting the file arguments below) Datasets helper scripts

1. Convert CSV to .npy and compute relevant properties of the molecules.

   • datasets/csv_to_npy.py — convert CSV (columns: "smiles", prop1, prop2, ...) → single container .npy (dict with "smiles", "props", "prop_names") or optional separate smiles/props .npy
     • example:

       python datasets/csv_to_npy.py --input molecules.csv --output molecules.npy
       

     • output: molecules.npy (load later with np.load(..., allow_pickle=True).item() if saved as a dict)

2. Compute per-property statistics

   • datasets/compute_tranche_statistics.py — compute per-property mean/std/min/max given the molecules.npy file computed from step 1.
     • example:

       python datasets/compute_tranche_statistics.py
       

     • outputs: molecules_stats.npy

Notes

• Use these scripts to keep preprocessing reproducible; they replace the inline CSV->npy and stats code snippets above.
• Preserve molecule ordering across all steps (CSV → scattering → GRASSY). If you filter/reorder, save an index mapping file.
• When loading molecules.npy that contains a dict, use np.load(..., allow_pickle=True).item().
• If your filenames differ, adjust the CLI arguments accordingly.

3. Load data into train_learnable_scattering

• Expected inputs: molecules.npy and statistics file molecules_stats.npy
• Uses the dataloader defined in datasets/load_ZINC_tranche.py
• CLI example:

python train_learnable_scattering.py

• Loading inside a script:

dataset = ZINCDataset(f'datasets/fields_1.npy', prop_stat_dict=f'datasets/fields_1_stats.npy', include_ki=False)
# pass smiles/props to dataset creation used by train_learnable_scattering

• Output: trained weights + scattering moments file, e.g. scripts/trained_models/molecules.npy

4. Load scattering moments into train_grassy

• Ensure scattering moments align with the same molecule order used for properties.
• Expected inputs: molecules.npy, statistics file molecules_stats.npy, and saved scattering model scripts/trained_models/molecules.npy
• CLI example:

python train_grassy.py

• In-code example:

full_dataset = ZINCDataset(f'datasets/molecules.npy', prop_stat_dict=f'datasets/molecules_stats.npy', transform=Scattering(scatter_model_name=f'scripts/trained_models/molecules.npy'))

• Output: GRASSY model + latent embeddings file, e.g. outputs/grassy/embeddings.npy (shape: N x D)

5. Visualize latent embeddings colored by a property and save visualize_latent_embeddings.npy

• Load embeddings and pick a property (by name or index).

Notes & tips

• Keep molecule order consistent across all steps (CSV → scattering → GRASSY). If any data filtering/reordering occurs, store an index mapping file.
• Use prop_mean.npy/prop_std.npy to normalize properties before training.
• Verify shapes at each step (N molecules) and check for NaNs.
• If your training scripts accept .npz or separate .npy files, adapt the CLI accordingly.

This sequence produces:

• molecules.npy (raw data)
• property_stats.json / prop_mean.npy / prop_std.npy
• outputs/learnable_scattering/scattering_moments.npy
• outputs/grassy/embeddings.npy
• visualize_latent_embeddings.npy (2D projection for plotting)
• corresponding visualization PNG(s)