Abstract
Abstract
Advances in AI-driven drug-target interaction (DTI) prediction methods require extensive training data, which is often unavailable for many target proteins. This limitation prevents the development of effective computational models for predicting interactions between drug candidates and these understudied proteins, posing a significant challenge in drug discovery.
To address this issue, we investigate the use of deep transfer learning for predicting interactions between drug candidates and proteins with limited training data. Our approach involves first training a deep neural network classifier on a large, generalized source training dataset and then reusing this pre-trained network for re-training or fine-tuning on a smaller, specialized target training dataset. Through systematic evaluation, we show that transfer learning significantly improves performance when the target dataset contains fewer than 100 compounds.
Building on this, we propose a novel self-supervised learning framework that uses contrastive learning to generate rich graph embeddings without the need for explicit labels. Our method employs a dual local-global strategy integrating detailed local structural attributes with high-level molecular features. This approach allows the model to learn detailed and transferable graph representations, proving effective for various downstream tasks such as DTI and molecular property predictions.
In conclusion, this study highlights the value of combining transfer learning and self-supervised learning with contrastive learning for DTI prediction, especially in scenarios with limited training data. Our framework provides robust and transferable models capable of generating rich graph embeddings, facilitating accurate predictions of drug-target interactions and supporting the discovery of novel therapeutics for understudied proteins.