ConGen: Targeted Molecule Generation Through Contrastive Learning and Latent Optimization

The discovery of novel compounds for protein targets is an important step in the drug discovery process. Target-specific molecules can be designed based on protein structures, but molecule prediction based on protein sequences alone remains a challenge. We hypothesize that including information on noninteracting pairs can improve the specificity and generalizability of molecule prediction models. To this end, we introduce ConGen, a sequence-conditioned framework for targeted molecule generation from protein sequence alone. To our knowledge, it is the first model that combines contrastive protein-molecule representation learning with interacting and noninteracting molecules and latent optimization (LO) in a shared pretrained space, addressing the limitations of interacting-only training. ConGen follows two stages: (i) contrastive learning (CL) maps proteins and molecules into a joint space where interacting pairs are pulled together and noninteracting pairs are pushed apart and (ii) LO starts from random molecule latents, moves them toward the target protein's embedding, and decodes a local neighborhood to propose candidates. This approach enables the generation of molecules specifically designed for a given target by efficiently exploiting the structure of the shared latent space. Our results show that ConGen achieves performance comparable to EncDecLM, a state-of-the-art sequence-based model for targeted drug generation. An ablation study further demonstrates the contributions of both the CL and LO stages.