CoDrug: A Text-Driven Molecular Virtual Screening and Multiproperty Optimization Framework via Multimodal Language...

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Bioinformatics

Background:

• Traditional molecular screening is computationally expensive and relies on complete 3D protein structures.
• Lack of reliable 3D structures hinders drug design.
• Existing methods face limitations in speed and accessibility.

Purpose of the Study:

• To introduce CoDrug, a multimodal fusion framework for AI-assisted drug discovery.
• To overcome limitations of traditional methods by integrating text and structure data.
• To enable text-driven virtual screening and molecular optimization.

Main Methods:

• CoDrug uses multimodal fusion combining text (SciBERT, ChemFormer) and structural (ESM, ChemFormer) data.
• Contrastive learning aligns textual and structural embeddings in a shared latent space.
• Two fusion strategies: text-protein sequence and text-compound structure.

Main Results:

• CoDrug achieves competitive or superior performance compared to state-of-the-art baselines.
• The framework excels in zero-shot settings and when 3D structural data is incomplete.
• Demonstrates strong generalization and representation expressiveness.

Conclusions:

• CoDrug offers a scalable and flexible paradigm for early-stage drug discovery.
• The natural language interface lowers the technical barrier for chemists.
• Accelerates the integration of language-driven hypotheses with structure-guided molecular design.