ConGen: Targeted Molecule Generation Through Contrastive Learning and Latent Optimization.

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Designing target-specific molecules from protein structures is established, but prediction from protein sequences alone is challenging.
• Existing models often focus solely on interacting molecule pairs, potentially limiting specificity and generalizability.

Purpose of the Study:

• To introduce ConGen, a sequence-conditioned framework for targeted molecule generation from protein sequence.
• To improve molecule prediction models by including information on noninteracting pairs.

Main Methods:

• ConGen utilizes a two-stage approach: contrastive learning (CL) and latent optimization (LO) in a shared pretrained space.
• CL maps proteins and molecules into a joint space, attracting interacting pairs and repelling noninteracting ones.
• LO refines random molecule latents towards the target protein's embedding for candidate generation.

Main Results:

• ConGen demonstrates performance comparable to state-of-the-art sequence-based models like EncDecLM for targeted drug generation.
• An ablation study confirmed the significant contributions of both the CL and LO stages to ConGen's effectiveness.

Conclusions:

• ConGen represents a novel approach for targeted molecule generation using only protein sequences.
• The inclusion of noninteracting molecule information and the combined CL/LO strategy enhance model specificity and generalizability in drug discovery.