QSARcons - smart search for consensus of QSAR models

The motivation behind this project is that there are many available chemical descriptors and machine learning methods, and usually, it is not obvious which combination to prefer for modelling the target property of the molecule. Therefore, the idea is just to build multiple (>100) simple individual QSAR models with diverse descriptors and algorithms, and then a genetic algorithm to smartly optimize the subset of models delivering the best performance on the validation dataset.

Motivation

1. Simple design: QSARcons focuses on simplicity of use. The default pipeline just requires training and test data.

2. Traditional QSAR: QSARcons includes a wide range of traditional molecular descriptors and machine learning algorithms, providing a transparent baseline for comparison with more advanced approaches like deep learning-based or complex QSAR workflows.

3. Universal workflow - QSARcons can be applied to any type of chemical property modelling.

Overview

QSARcons provides a two-layer workflow.

1. Model generation

Build multiple QSAR models (>100) using 2D chemical descriptors and traditional machine learning algorithms. The individual model building pipeline is kept simple, without advanced data preprocessing. Optional in-house stepwise hyperparameter optimization is available for all ML methods.

2. Consensus search

Identify the optimal subset of QSAR models using several search strategies:

• Random search

• Systematic search

• Genetic search

Installation

pip install qsarcons

QSARcons benchmarking

QSARcons can be easily benchmarked against alternative approaches. For that, just call the default pipeline function below. Input data are dataframes where the first column is molecule SMILES and the second column is molecule property (regression or binary classification).

from datasets import load_dataset
from qsarcons.meta import ConsensusModel

train_df = load_dataset("openadmet/openadmet-expansionrx-challenge-data", split="train").to_pandas()
test_df = load_dataset("openadmet/openadmet-expansionrx-challenge-data", split="test").to_pandas()

prop_name = "Caco-2 Permeability Efflux"
train_df = train_df[["SMILES", prop_name]].dropna()
test_df = test_df[["SMILES", prop_name]].dropna()

output_folder = f"{prop_name}_qsarcons"

model = ConsensusModel(hopt=False, output_folder=output_folder, verbose=True)
test_df_pred = model.run_predict(train_df, test_df)

print(model.best_cons)

Colab

See an example in QSARcons pipeline .

QSARcons Basic vs. QSARcons Pro

The QSARcons idea is that diverse and strong individual models can be combined to even stronger consensus. Currently, two versions are under development:

• QSARcons Basic: includes RDKit descriptors + scikit-learn ML methods

• QSARcons Pro: will include QSARcons Basic + other workflows for building individual models (e.g. chemprop and QSARmil) to combine traditional and advanced modelling approaches into stronger consensuses