Protein-ligand interactions play a central role in many biological processes ranging from cell signaling to disease treatment. However, traditional drug discovery struggles with the immense number of potential drug molecules and the limited availability of detailed protein-ligand complex data. This bottleneck hinders the development of highly selective drugs with strong binding affinity to their target proteins. Recent research has opened an exciting new avenue by demonstrating the potential to predict the most probable orientations and positions of chemical groups solely from the local backbone conformation and the identity of the interacting amino acid. This eliminates the need for scarce protein-ligand complex data, offering a significant advantage in drug design.
Our research builds upon this promising avenue and takes a significant step forward by extending the concept to the broader protein micro-environment. We propose formulating the identification of functional hotspots within these micro-environments as a semantic segmentation task. Our method harnesses a convolutional neural network (CNN) trained on a meticulously curated database, encompassing a diverse array of protein micro-environments surrounding various functional groups. To assess the efficacy of our approach, we evaluated the predicted hotspots based on their pharmacological relevance and compared them to existing protein-ligand complexes. This evaluation demonstrates the significant potential of our deep learning approach to expedite the drug discovery process by prioritizing promising target sites for further investigation.

Co-authors: Thomas Lemmin