Script recipes

Run from the skill root after installing the selected model's dependencies. These external-data training commands are illustrative; the review records the actual offline synthetic tests and dependency limits.

Solubility

python scripts/predict_solubility.py --epochs 50
python scripts/predict_solubility.py --data measured.csv \
  --smiles-col smiles --target-col logS --epochs 50 --predict CCO c1ccccc1

The benchmark path loads ESOL; the custom path requires complete finite continuous targets. Both use 2048-bit circular fingerprints and a Torch multitask regressor, train-only target normalization, and original-unit metrics/predictions. Custom units are whatever the input column encodes; the script does not convert them.

Graph network

python scripts/graph_neural_network.py --model dmpnn --dataset bbbp --epochs 20
python scripts/graph_neural_network.py --model attentivefp --data molecules.csv \
  --targets activity --task-type regression --epochs 50

DMPNN needs torch-geometric in stable 2.8.0. The other four graph choices need a compatible DGL and DGL-LifeSci stack. The same model-specific featurizer is used for benchmark and custom input. Classification requires binary observed labels; empty splits fail early and class-deficient holdout tasks get undefined AUC, not a made-up value. Regression benchmark transforms are inverted for evaluation.

Pretrained encoder

python scripts/transfer_learning.py --model chemberta --dataset bbbp --epochs 10
python scripts/transfer_learning.py --model chemberta --data measured.csv \
  --target logS --task-type regression --model-id ./local-chemberta --local-files-only

HF training supports one fully observed, unweighted task. Tox21 is rejected for this route because the bundled stable HF wrapper does not implement masked multitask loss. Strings, tokenizer and sequence-classification model objects are kept separate. Classification logits are converted with softmax before scoring. Pin the Hub revision for reproducibility; a new task head needs supervised training.

For MoLFormer, use --model molformer --trust-remote-code --revision <reviewed-commit> with the current IBM repository. Its current card targets Transformers 5; compat-v4 is the documented branch for Transformers 4: resolve it to the reviewed 40-character commit SHA before passing it to this script. Moving branch names do not pin code. Review/pin both code and weights; the script does not enable repository code implicitly.

For GROVER, provide --model grover --checkpoint encoder.pt --grover-config architecture.json. The JSON must match the checkpoint encoder, for example {"hidden_size":128, "num_attn_heads":4,"depth":1} only when those were its training settings. The checkpoint must contain DeepChem's embedding state dictionary; the script loads it strictly with weights_only=True and initializes a fresh task head. Original external GROVER checkpoints require a separately verified conversion; they are not assumed to match DeepChem's layout. Fixed 151/165 graph feature widths and 2048 additional fingerprint features must match the intended pipeline.