Application of Bioactivity Profile-Based Fingerprints for Building Machine Learning Models.

Area of Science:

• Computational chemistry
• Cheminformatics
• Drug discovery

Background:

• High-throughput screening (HTS) generates vast biochemical and cell-based assay data, offering opportunities for data mining and drug repurposing.
• High-throughput fingerprints (HTSFPs) represent molecular bioactivity profiles, applicable to virtual screening, iterative screening, and target deconvolution.
• Previous studies primarily focused on predicting single HTS assay outcomes, with limited exploration of modeling compound activity across multiple protein targets.

Purpose of the Study:

• To compare the performance of in-house HTSFPs combined with multitask deep learning against single-task support vector machines for predicting compound activity across a panel of targets.
• To evaluate the effectiveness of HTSFPs in hit identification and scaffold hopping potential.
• To investigate the impact of high-throughput screening false positives and negatives on model performance.

Main Methods:

• Development and application of HTSFP-based models using multitask deep learning and single-task support vector machines.
• Comparison of HTSFP models against models built using Extended Connectivity Fingerprints (ECFPs).
• Analysis of model performance in terms of hit identification and scaffold hopping potential, considering the influence of screening data quality.

Main Results:

• HTSFPs combined with multitask deep learning and support vector machines showed comparable performance in predicting compound activity.
• The two fingerprinting methods (HTSFPs and ECFPs) identified diverse hits with minimal overlap, indicating descriptor orthogonality.
• Combining ECFPs with HTSFPs in predictive models significantly increased scaffold hopping potential.

Conclusions:

• Bioactivity profile-based descriptors (HTSFPs) are complementary to structural descriptors (ECFPs) for drug discovery.
• Integrating HTSFPs into predictive modeling enhances the ability to discover novel chemical scaffolds.
• The study highlights the utility of HTSFP and multitask learning for exploring large-scale screening data and improving hit identification and scaffold hopping in drug discovery pipelines.