A meta learning and task adaptive approach for drug target affinity prediction.

Area of Science:

• Computational chemistry
• Pharmacology
• Artificial intelligence in drug discovery

Background:

• Accurate drug-target affinity (DTA) prediction is crucial for efficient drug discovery.
• Current deep learning DTA models face challenges with limited data and poor generalization.
• Addressing low-data scenarios is essential for advancing DTA prediction.

Purpose of the Study:

• To introduce AdaMBind, a novel DTA prediction model utilizing a meta-learning framework.
• To design an adaptive task module within AdaMBind for effective low-data DTA prediction.
• To enhance training efficiency and robustness using a dynamic "easy-to-hard" task scheduling mechanism.

Main Methods:

• Developed AdaMBind, a meta-learning based DTA prediction model.
• Implemented an adaptive task module and dynamic "easy-to-hard" task scheduling.
• Evaluated model performance on three benchmark datasets, focusing on few-shot learning conditions.

Main Results:

• AdaMBind outperformed 8 baseline models in predicting affinity for unseen targets, especially in few-shot scenarios.
• The model successfully identified high-affinity compounds for ESR and TP53 under stringent data constraints.
• AdaMBind identified potent FLT3 inhibitors for acute myeloid leukemia, validated by preliminary experimental assays.

Conclusions:

• AdaMBind offers a robust and effective framework for few-shot drug-target affinity prediction.
• The meta-learning approach with adaptive tasks significantly improves DTA prediction accuracy in data-limited settings.
• AdaMBind demonstrates practical utility in virtual screening and drug discovery for diseases like acute myeloid leukemia.