See if we could expand the BACE dataset from MoleculeNet-BACE, using synthethic data augmentation.

An $\beta$-CVAE was constructed. The CVAE parts was verry loosley based on work by Jaechang, Lim (link to github). Link to the paper (Molecular generative model based on conditional variational autoencoder for de novo molecular design).

The project is divided into three distinct, but chained parts (which has its own more detailed READMEs). The three parts are:

1. Pre-processing:

   1. Loads, cleans and canoicalize the data
   2. Splits the data into five CV-folds.
   3. Avoids data leakage

2. $\beta $-CVAE:

   1. Points towards the data from pre-processing
   2. Trains a $\beta$-CVAE, for each of the five folds.
   3. Generates (for each of the five CV-iterations) n-valid, molecules of which does not share scaffolds with the validation or test-sets.
      1. Also analyses the generated molecules and produce plots describing everything from t-SNE of the Morgan fingerprints, to the label head residual errors.
      2. Computes V.U.N metrics, and calculates the external Diversity between the generated molecules and the train dataset.

3. Downstream models

   1. MPNN/GAT (used synonomusly):
      • Training: three diffrent mixes of synthetic versus BACE data. (synthetic data matched with each CV-iteration). The mixes where: 0%,33%,67%. The 0% is a baseline Where the percentage describes what share of the data is synthetic vs BACE data. Each iteration has around 1000 BACE datasamples for training, and around 260 for validation.
      • The MPNN can also be pre-trained on the synthetic data (for each of the three datasets, 0%,33%,67%). The 0% case means that the model sees no Syntheitc data (BASELINE) Then fine-tuned on only natrual data of each set.
   2. RF (Random forrest):
      • Trains on the mixes, and does a parameter sweep

• Conda
• Preferably a machine with CUDA support for faster training

From the project root:

If you have a CUDA 12.1 compatible NVIDIA GPU (typical Windows/Linux setup), use:

conda env create -f environment.yml
conda activate rdkit_draw

If you are on a modern Mac (Apple Silicon, e.g. M1/M2/M3), use the Apple Silicon environment file:

conda env create -f environment.macos-arm64.yml
conda activate rdkit_draw_mac

Notes for Mac:

• Training will run on CPU or Apple MPS, not CUDA.
• The root environment.yml is Windows/CUDA-pinned and includes a Windows-specific prefix, so it is not portable to macOS as-is.

If environment creation fails because of a machine-specific prefix: in environment.yml, remove the prefix line and run the command again.

If activation fails in PowerShell:

conda init powershell

Then restart terminal and activate again.

Optional update command (if environment already exists):

conda env update -f environment.yml --prune

Quick package check:

python -c "import torch, rdkit, sklearn, pandas, matplotlib; print('ENV_OK')"

Run the three parts in this order:

1. Pre-processing
2. beta-CVAE generation
3. Downstream MPNN/GAT training + holdout evaluation

All commands below are run from the project root.

This creates scaffold-split folds and the held-out test set.

python src/original_preprocessing.py

Outputs:

• data/heldout_datasets/heldout_testset.csv
• data/combination_1300_molecules_and_0_%_synthetic/original_fold_0.csv ... original_fold_4.csv

This trains the CV pipeline over folds and generates synthetic molecules.

Note: in vae/fold_pipeline_config.example.yaml, both train.enabled and sampling.enabled are false by default. Set them to true before running generation.

python vae/run_fold_pipeline.py --config vae/fold_pipeline_config.example.yaml

Fast smoke run (single fold):

python vae/run_fold_pipeline.py --config vae/fold_pipeline_config.example.yaml --only-fold 0

Expected synthetic outputs are written under:

• vae/fold_pipeline_outputs/cv_iteration_0 ... cv_iteration_4

After beta-CVAE generation, create the mixed training folders:

python src/synthetic_preprocessing.py

Outputs:

• data/combination_1950_molecules_and_33_%_synthetic
• data/combination_3900_molecules_and_67_%_synthetic

Train downstream GAT/MPNN models on 0%, 33%, and 67% datasets (5-fold CV each):

python src/GAT_model/gat_predictor.py

Main artifacts:

• src/GAT_model/0%/checkpoints, src/GAT_model/33%/checkpoints, src/GAT_model/67%/checkpoints
• src/GAT_model/0%/MPNN_cv_results_0%.csv (and matching files for 33%/67%)

Evaluate saved checkpoints on heldout_testset.csv:

python src/GAT_model/run_gat.py

Evaluate only one dataset (example 0%):

python src/GAT_model/run_gat.py --folder 0%

Outputs:

• results/pretrain/gat_results_heldout.csv or results/no_pretrain/gat_results_heldout.csv
• results/pretrain/GAT_predictions_heldout_set.csv or results/no_pretrain/GAT_predictions_heldout_set.csv

• Pre-processing and dataset files are regenerated by running the scripts above.
• VAE analysis plots are regenerated as part of the fold pipeline when analysis is enabled in the config.
• Downstream GAT/MPNN metrics and held-out predictions are regenerated by rerunning gat_predictor.py and run_gat.py.