BatchDTA: implicit batch alignment enhances deep learning-based drug-target affinity estimation.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Drug-target affinity (DTA) estimation is crucial for identifying potential drug candidates.
• Deep neural networks (DNNs) are efficient for DTA estimation but are negatively impacted by batch effects.
• Batch effects cause data deviation, limiting DNN training to small, clean datasets and hindering precise estimations.

Purpose of the Study:

• To develop a batch-sensitive training framework, BatchDTA, to mitigate the impact of batch effects on DNN models for DTA estimation.
• To enhance the accuracy and robustness of DNN models in DTA prediction.

Main Methods:

• Designed BatchDTA, a framework that implicitly aligns multiple batches by learning compound orders within batches.
• Applied BatchDTA to four mainstream DNN models for DTA estimation.
• Evaluated BatchDTA on BindingDB, Davis, and KIBA datasets.

Main Results:

• BatchDTA significantly enhanced the ability and robustness of DNN models across multiple DTA datasets.
• The average concordance index showed a relative improvement of 4.0% for DNN models trained with BatchDTA.
• Case studies confirmed BatchDTA's ability to learn compound ranking orders from diverse batches and improve fused data performance.

Conclusions:

• BatchDTA effectively alleviates batch effects in DNN-based DTA estimation.
• The framework improves the precision and consistency of DTA predictions, aiding drug discovery.
• BatchDTA offers a valuable approach for training robust DNN models on heterogeneous DTA data.