Virtual screening of bioassay data.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Challenges in virtual screening of bioassay data include limited access to curated datasets, a high rate of false positives in primary screening, and significant class imbalance (few active compounds vs. many inactive ones).
• Pharmaceutical data accessibility is restricted, with PubChem data lacking curation and cross-referencing between primary and confirmatory assays.
• Analysis of false positives in primary screening is hampered by poor data cross-referencing, though identified cases show an average of 64% false positives in High-Throughput Primary screening.

Purpose of the Study:

• To discuss the key challenges in virtual screening of bioassay data.
• To evaluate the performance of various Weka cost-sensitive classifiers (Naive Bayes, SVM, C4.5, Random Forest) on diverse bioassay datasets.
• To investigate the impact of classifier choice and cost matrix settings on classification performance.

Main Methods:

• Application of cost-sensitive classification algorithms available in the Weka software suite.
• Utilizing Naive Bayes, Support Vector Machine (SVM), C4.5 decision tree, and Random Forest classifiers.
• Testing classifiers on multiple bioassay datasets with varying characteristics.

Main Results:

• Weka's Support Vector Machine (SVM) and C4.5 decision tree implementations demonstrated relatively strong performance in cost-sensitive classification tasks.
• The optimal configuration of the Weka cost matrix is dependent on the specific base classifier being used, rather than solely on the ratio of class imbalance.
• The high percentage of false positives in primary screening raises questions about the suitability of such data for virtual screening.

Conclusions:

• Enhanced accessibility of curated pharmaceutical screening data is crucial for both industry and academia.
• Virtual screening can benefit drug discovery by reducing the search space and improving primary screening processes through analysis of false positives.
• Care must be taken when applying Weka's cost-sensitive classifiers; using generic misclassification costs based on class ratios is not recommended for comparing different classifiers on the same dataset.