Multiclass Synthetic Accessibility Prediction.

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Machine learning

Background:

• Predicting synthetic accessibility of molecules is crucial in drug discovery.
• Existing binary classification models face challenges due to data imbalances and fixed thresholds.
• Machine learning is increasingly applied to predict synthetic ease or difficulty.

Purpose of the Study:

• To develop a novel multiclass classification approach for predicting the minimum synthetic steps required for a molecule.
• To address limitations of binary classification methods in synthetic accessibility prediction.
• To introduce fuzzy evaluation metrics for more realistic performance assessment.

Main Methods:

• A multiclass fold-ensembled classification approach was developed.
• Base models were trained on multiple stratified subsampled folds to mitigate class imbalance.
• Probability or voting aggregation strategies were used.
• Fuzzy evaluation metrics were proposed to account for prediction tolerances.

Main Results:

• The model demonstrated effectiveness in multiclass synthetic accessibility prediction on benchmark datasets.
• The proposed method outperformed six existing models in binary synthetic accessibility prediction.
• The fold-ensembling strategy successfully mitigated class imbalance issues.

Conclusions:

• The novel multiclass approach offers a more nuanced and accurate prediction of synthetic accessibility.
• Fuzzy evaluation metrics provide a more practical assessment of model performance.
• This work advances the application of machine learning in optimizing drug discovery pipelines.