Deep Learning-Based Imbalanced Data Classification for Drug Discovery.

Area of Science:

• Computational chemistry and cheminformatics
• Machine learning in drug discovery
• Bioassay data analysis

Background:

• Drug discovery is costly and time-consuming, with early stages focusing on identifying and optimizing drug-like compounds.
• High-throughput screening (HTS) is a conventional method for detecting active compounds, generating vast datasets.
• The PubChem repository offers millions of HTS bioassays, valuable for machine learning but often suffer from imbalanced data.

Purpose of the Study:

• To investigate the classification performance of deep neural networks (DNNs) on imbalanced compound datasets from HTS.
• To evaluate the effectiveness of various data balancing techniques in mitigating performance issues caused by data imbalance.
• To assess the impact of the degree of data imbalance on DNN classification performance.

Main Methods:

• Utilized five confirmatory HTS bioassays from the PubChem repository.
• Applied one undersampling and three oversampling methods for data balancing.
• Employed a fully connected, two-hidden-layer DNN for classifying active and inactive molecules.
• Evaluated performance using balanced accuracy, precision, recall, F1 score, Matthews correlation coefficient, and AUC.

Main Results:

• Data balancing methods demonstrated a degree of mitigation for the negative effects of imbalanced data on DNN performance.
• The extent of data imbalance was found to negatively correlate with the overall classification performance of the network.
• Specific balancing methods showed varying degrees of success in improving classification metrics for imbalanced datasets.

Conclusions:

• Data balancing techniques are crucial for improving machine learning model performance on imbalanced HTS bioassay data.
• While balancing methods help, significant data imbalance remains a challenge in accurately classifying active compounds.
• Further research is needed to optimize balancing strategies for complex, imbalanced datasets in drug discovery.