scripts

Reproduction Scripts

One bash script per (dataset, model, learning_type) combination, named:

<dataset>-<model>-<learning_type>.sh

Where learning_type is one of:

code	meaning
al	Active learning (informative selection: --select_method explorative)
pl	Passive learning (random selection: --select_method random)
smal	Short-term Memory Active Learning (AL with --forget_method and --n_forget)

Each run script loops over the seeds, error rates, and (for SMAL) forget methods used in the manuscript, calling molalkit_run once per combination. The grouped SIMPD scripts use one script per learning type and loop over all committed datasets/CHEMBL*.csv files.

Prerequisites

Install MolALKit v1.2.0 following the instructions in the main README. The molalkit_run CLI must be on your PATH, and python -c "import molalkit" must succeed (used by some scripts to locate packaged dataset CSVs).

DMPNN scripts additionally require a MolALKit environment with Chemprop support and a CUDA-capable GPU for the manuscript-scale runs.

MolFormer scripts additionally require the IBM MolFormer pretrained checkpoint file. Download the checkpoint from https://github.com/IBM/molformer and pass its path with MOLFORMER_CKPT:

MOLFORMER_CKPT=/path/to/N-Step-Checkpoint_3_30000.ckpt bash scripts/pgp_broccatelli-molformer-al.sh

Running

From the repository root:

bash scripts/ames-rf_morgan-al.sh

Outputs are written under ./results-initial/<active_learning|passive_learning|smal>/.... Override the destination with the OUT_ROOT environment variable:

OUT_ROOT=/path/to/output bash scripts/ames-rf_morgan-smal.sh

DMPNN and MolFormer pgp active-learning scripts default to model-specific output roots, ./results-dmpnn/active_learning and ./results-molformer/active_learning, respectively. They also accept OUT_ROOT.

SIMPD scripts default to ./results-simpd/<active_learning|passive_learning|smal>. Each SIMPD dataset CSV contains a split column; the scripts preserve this split by creating temporary MolALKit input files with SMILES and Y columns, using the train rows for active learning and the test rows for validation. Override the dataset location with DATA_ROOT if needed:

DATA_ROOT=/path/to/simpd_csvs OUT_ROOT=/path/to/output bash scripts/simpd-rf_morgan-al.sh

The SIMPD SMAL script reads per-dataset f_min_train_size and max_iter values from scripts/simpd-smal-params.csv. Override with FMIN_TABLE only if you have regenerated those manuscript parameters.

Stratified-shuffle label-error scripts default to ./results-stratified_shuffle/<active_learning|smal>:

bash scripts/stratified_shuffle-rf_morgan-al.sh
bash scripts/stratified_shuffle-rf_morgan-smal.sh

The stratified-shuffle SMAL script reads f_min_train_size and max_iter from scripts/stratified_shuffle-smal-params.csv.

RF parameter-set scripts default to ./results-rf-parameter-sets. They use the custom RandomForest/Morgan configs in scripts/rf_parameter_configs/; the active-learning script writes manifest.csv, and the SMAL script writes smal_manifest.csv for scripts/build_processed_results.py.

bash scripts/rf_parameter_sets-rf_morgan-al.sh
bash scripts/rf_parameter_sets-rf_morgan-smal.sh

The RF parameter-set SMAL script reads f_min_train_size and max_iter from scripts/rf_parameter_sets-smal-params.csv.

The scripts are intentionally simple sequential loops — they do not submit SLURM jobs, parallelize, or skip already-completed runs. For HPC use, wrap each molalkit_run invocation in your own job scheduler.

Currently included experiments

Model	Datasets	Learning types
RandomForest/Morgan	ames, CYP2D6_Veith, CYP3A4_Veith, MDR1_MDCK_classification2, PAMPA_NCATS, pgp_broccatelli	al, pl, smal
RandomForest/Morgan	SIMPD ChEMBL datasets (CHEMBL*.csv)	al, pl, smal
RandomForest/Morgan	ames, CYP2D6_Veith, CYP3A4_Veith, MDR1_MDCK_classification2, PAMPA_NCATS, pgp_broccatelli with stratified-shuffle label errors	al, smal
RandomForest/Morgan RF parameter sweep	MDR1_MDCK_classification2, PAMPA_NCATS, pgp_broccatelli	al, smal
DMPNN	pgp_broccatelli	al
MolFormer	pgp_broccatelli	al