Command-Line Interfaces

The project exposes user-facing commands through python -m src.cli.<name> and, for training, also through python -m src.training.train.

Feature generation: src.cli.process_features

python -m src.cli.process_features data/example/raw/psychlight_a.csv \
    --output-dir data/example/processed/psychlight_a

Current outputs:

• psychlight_a_descriptors.csv
• psychlight_a_mordred.csv
• psychlight_a_morgan.csv
• optional psychlight_a_dpk.csv
• optional psychlight_a_dpk.smi
• optional psychlight_a_deeppk_extra.csv

Defaults from the current CLI help:

• --output-dir: data/processed
• --fingerprint-radius: 3
• --fingerprint-bits: 1024
• --deeppk-pred-type: admet
• --deeppk-poll-interval: 60
• --deeppk-timeout: 3600

Data exploration: src.cli.explore

python -m src.cli.explore \
    --dimensionality-reduction data/example/processed/psychlight_a/psychlight_a_descriptors.csv \
    --target-column Class \
    --output-dir data/example/exploration

Current CLI default:

• --output-dir: data/example/exploration

If you pass no action flags, _ensure_actions enables the full exploration bundle:

• class distribution
• missingness
• PCA summary
• feature importance
• feature distributions
• t-SNE
• UMAP
• variance threshold export
• mutual information
• binary activity
• outlier report

See the dedicated Data Exploration page for artifact details.

Training: src.cli.train and src.training.train

python -m src.cli.train --config configs/train.yaml \
    --datasets psychlight_a_descriptors \
    --models logisticregression

Current CLI flags:

• --config
• --models
• --datasets
• --skip-existing

Important current behavior:

• the CLI default path is configs/train.yaml
• the CLI prints a short completion summary, not YAML
• train_from_config also writes a pipeline-steps overview CSV before training when dataset/model rows are available

Evaluation: src.cli.evaluate

python -m src.cli.evaluate \
    --model models/example/trained_models/.../logisticregression_model_<hash>_<timestamp>.joblib \
    --test-csv data/example/processed/psychlight_a/psychlight_a_descriptors.csv \
    --target-col Class

Current CLI default:

• --output-dir: models/evaluation

Behavior:

• accepts either a single .joblib file or a directory of .joblib files
• writes one YAML metrics file per evaluated model
• when given a directory, also writes a summary CSV
• aligns feature names using the same sanitization logic as training

Prediction: src.cli.predict

python -m src.cli.predict \
    --model models/example/trained_models/.../logisticregression_model_<hash>_<timestamp>.joblib \
    --input-csv data/example/processed/psychlight_a/psychlight_a_descriptors.csv \
    --target-col Class

Behavior:

• appends prediction for a single model
• appends <model_name>_prediction for a directory of models
• ignores --target-col if that column is present, so labeled CSVs can be reused
• defaults the output path to <input>_predictions.csv

Quick validation commands

conda run -n lig-cls python -m src.cli.process_features --help
conda run -n lig-cls python -m src.cli.explore --help
conda run -n lig-cls python -m src.cli.train --help
conda run -n lig-cls python -m src.cli.evaluate --help
conda run -n lig-cls python -m src.cli.predict --help

Common failure modes

• Missing RDKit usually means the wrong Python environment is active.
• Missing feature columns in evaluate or predict are reported using sanitized feature names plus hints back to the original headers.
• DeepPK timeouts can be handled with --deeppk-timeout and --deeppk-poll-interval.
• Training resume behavior depends on models/tracking/training_runs.csv, not on the configured log directory.