Integrating Incompatible Assay Data Sets with Deep Preference Learning.

Area of Science:

• Computational Chemistry
• Bioinformatics
• Machine Learning

Background:

• Vast amounts of bioactivity assay data exist in public databases.
• Integrating diverse datasets for quantitative structure-activity relationship (QSAR) studies is challenging due to experimental variability.

Purpose of the Study:

• To present an efficient deep-learning approach, Deep Preference Data Integration (DPDI), for integrating heterogeneous bioactivity assay data.
• To improve the accuracy and reliability of QSAR models by leveraging diverse data sources.

Main Methods:

• DPDI utilizes a neural network trained with a surrogate variable to reconcile outcome variables from different assay types.
• The training process maximizes the consistency of the surrogate variable's induced order with the provided datasets.

Main Results:

• DPDI successfully integrated 2959 molecules across 129 assay datasets for predicting factor Xa inhibitor efficacy.
• Data integration using DPDI significantly improved prediction accuracy in both interpolation and extrapolation tasks.

Conclusions:

• DPDI is an effective deep learning tool for integrating diverse bioactivity data for QSAR studies.
• The method demonstrates robust performance, enhancing predictive capabilities in drug discovery and development.