Target-based drug discovery through contrastive learning and latent optimization

The development of novel compounds targeting proteins of interest is an important step in the process of drug discovery to cure diseases. Several studies have focused on generating target-specific molecules based on protein sequences alone. However, these models consider only interacting protein-ligand pairs, overlooking the informative value of non-interacting pairs. In this thesis, we introduce ConGen, a novel contrastive learning-based approach for targeted drug generation that effectively utilizes both interacting and non-interacting pairs. ConGen consists of two stages: contrastive learning and latent optimization. In the contrastive learning stage, ConGen learns to map protein and molecule representations into a shared space, clustering interacting molecules near their target proteins and non-interacting molecules farther away. In the latent optimization stage, ConGen uses this space to identify representations that are close to the representation of the target protein. This approach enables the generation of molecules that are specifically designed for a particular target by efficiently utilizing the arrangements in the shared latent space. Different contrastive learning loss functions are compared and the best performing functions are used to benchmark the model against other studies. The effectiveness of ConGen's contrastive learning and latent optimization stages are demonstrated using an ablation study. Moreover, the architecture of ConGen is easily adapted for the drug-target interaction prediction task, where it achieves results slightly better than the baseline. Lastly, a web service has been developed to deploy target-based drug generative models, enabling researchers to create and evaluate new drug molecules efficiently.