New AI Model 'GPS-DTI' Navigates Uncharted Territory in Drug Discovery

Accurate prediction of drug-target interactions (DTIs) is the cornerstone of modern drug discovery. However, computational methods often stumble when facing 'generalisation' issues—failing to predict how new, unseen drugs will behave with specific protein targets. To bridge this gap, scientists have introduced GPS-DTI, a sophisticated deep learning framework designed to capture the intricate geometry of molecular structures.

The system utilises a graph neural network combined with a multi-head attention mechanism to model the structural nuances of drug molecules. On the biological side, it processes protein representations using a pre-trained Evolutionary Scale Model (ESM-2) refined by convolutional neural networks. A specialised cross-attention module then integrates these drug and protein features, highlighting biologically meaningful interactions.

comprehensive benchmarking shows that GPS-DTI outperforms existing methods in both in-domain and cross-domain tasks. It has achieved state-of-the-art performance in drug-target affinity and demonstrated impressive adaptability when tested on an independent COVID-19 dataset. Furthermore, the model provides interpretable visualisations of cross-attention maps, allowing researchers to see exactly which molecular parts are interacting. This combination of high accuracy and transparency suggests GPS-DTI could be a powerful engine for real-world pharmaceutical applications.