Bambu

Bambu (BioAssays Model Builder), is a simple tool to generate QSAR models based on PubChem BioAssays datasets. It relies on RDKit and the FLAML AutoML framework and provides utilitaries for downloading and preprocessing datasets, as well training and running the predictive models.

Try it!

Try Bambu on this Google Colab Notebook ^^

Installing

Installing as a PyPI package using pip:

$ pip install bambu-qsar

Note: RDKit must be installed separately.

Intalling as an environment using conda on Linux:

$ git clone https://github.com/omixlab/bambu-v2
$ cd bambu-qsar
$ make setup PLATFORM=linux
$ conda activate bambu-qsar

Running it with Docker

$ docker run -ti omixlab/bambu-qsar:latest

Downloading a PubChem BioAssays data

Downloads a PubChem BioAssays data and save in a CSV file, containing the InchI representation and the label indicating molecules that were found to be active or inactive against a given target.

$ bambu-download \
	--pubchem-assay-id 29 \
	--pubchem-InchI-chunksize 100 \
	--output 29_raw.csv

The generated output contains the columns pubchem_molecule_id (Substance ID or Compound ID, depending on the option selected during download), InChI and activity. Only the fields InchI and activity are used in futher steps.

Computing descriptors or fingerprints

Computes molecule descriptors or Morgan fingerprints for a given datasets produced by bambu-download (or following the same format). The output also contains a train and test subsets, whose sizes are defined based on the --train-test-split-percent argument. The argument --resample might be used to perform a random undersampling in the dataset, as most HTS datasets are heavily umbalanced. The path passed to --output is used as template to generate the train and test file. In this case, 29_preprocess_train.csv and 29_preprocess_test.csv respectively.

$ bambu-preprocess \
	--input 29_raw.csv \
	--output 29_preprocess.csv \
	--output-preprocessor 29_preprocessor.pickle \
	--feature-type morgan-2048 \
	--train-test-split 0.75 \
	--undersample

Train

Trains a classification model using the FLAML AutoML framework based on the bambu-preprocess output datasets. The user may adjust most of the flaml.automl.AutoML parameters using the command line arguments. In this case we are using an Extra Trees Classifier.

$ bambu-train \
	--input-train 29_preprocess_train.csv \
	--output 29_model.pickle \
	--time-budget 3600 \
	--estimators extra_tree

A list of all available estimators can be accessed using the command bambu-train --list-estimators. Currently, only rf (Random Forest) and extra_tree are available.

Validation

An y-randomization validation can be performed using the command bambu-validate, which will compute accuracy, recall, precision, f1-score and ROC AUC training with the original training dataset and by validating with the test one, and furtherly randomizing the training labels and several times (--randomizations). For each randomization, classification metrics are computed again and significancy value (p-value) is computed based on the z-score-normalized metrics.

$ bambu-validate \
	--input-train 29_preprocess_train.csv \
	--input-test 29_preprocess_test.csv \
	--model 29_model.pickle \
	--output 29_model.validation.json \
	--randomizations 100

Predict

Receives an inputs, preprocess it using a preprocess object (generated using bambu-preprocess) and then runs a classification model (generated using bambu-train). Results are saved in a CSV file.

$ bambu-predict \
	--input pubchem_compounds.sdf \
	--preprocessor 29_preprocessor.pickle \
	--model 29_model.pickle \
	--output 29_predictions.csv

Contact

Feel free to open issues or pull requests! You may also contact us by email.

Dr. Frederico Schmitt Kremer, PhD. E-mail: fred.s.kremer@gmail.com.