SFU researchers develop AI that designs and synthesizes real-world drug candidates

SFU Researchers Develop AI That Designs Drugs and Maps Their Synthesis

A team from Simon Fraser University's School of Computing Science has introduced a new artificial intelligence framework that could speed up drug discovery by simultaneously designing drug molecules and outlining how to synthesize them. Their study, titled "Compositional Flows for 3D Molecule and Synthesis Pathway Co-design," presents CGFlow, an AI method that integrates 3D molecular modeling with practical chemical synthesis strategies.

This work addresses a major challenge in pharmaceutical research: many AI-generated molecules fit disease targets well but can't be realistically synthesized in the lab. Without a viable synthesis pathway, these promising molecules remain theoretical, causing delays and wasted effort.

How CGFlow Advances Molecular Design

CGFlow takes a stepwise approach to molecule construction, assembling compounds piece by piece. This method allows the AI to understand how each added fragment affects the molecule's 3D structure and function. It combines two key processes:

• Compositional Flow: Models the sequence of chemical reactions that form the molecule’s synthesis pathway using Generative Flow Networks (GFlowNets), focusing on high-value molecular structures.
• State Flow: Refines continuous molecular properties like atomic positions through temporal learning, ensuring each fragment fits correctly within the 3D structure.

Introducing 3DSynthFlow: Practical Drug Design

Building on CGFlow, 3DSynthFlow targets drug design against specific proteins, often related to diseases. Unlike traditional models that only focus on molecular structure or binding affinity, 3DSynthFlow co-designs both the molecule’s binding pose and its synthetic pathway. This dual focus is vital for producing molecules that can be realistically manufactured.

The framework applies industry-standard reaction rules and limits synthesis to two steps, balancing complexity and practicality. Its performance is notable:

• Superior Binding: Achieved state-of-the-art binding affinity across 15 protein targets on the LIT-PCBA benchmark.
• Efficiency: Sampled viable candidates 5.8 times faster than earlier 2D synthesis-based models, discovering molecules with more diverse protein-ligand interactions.
• High Synthesizability: Reached a 62.2% synthesis success rate on the CrossDocked benchmark, significantly outperforming similar models with comparable binding strength.
• Validity: All generated molecules were 100% chemically valid, indicating reliability in both design and synthesis potential.

Implications for Drug Development

This approach integrates synthesizability into early molecular design, creating a clearer path from AI-generated molecules to real-world drugs. Several research groups have already adopted the framework for early cancer drug discovery projects, demonstrating its practical value.

Though current versions focus on simpler synthesis steps and omit complex reactions like ring formation, the team plans to expand the system’s chemical action space and building block library. The goal is to tackle more intricate drug designs while maintaining practical manufacturability.

SFU’s work highlights an important step toward faster, more cost-effective drug development pipelines and offers a promising tool for therapeutic discovery.

Contact Information for SFU Experts

• Tony Shen, PhD student, Computing Science | tony_shen@sfu.ca
• Martin Ester, Professor, Computing Science | martin_ester@sfu.ca

For those interested in AI applications in drug discovery and molecular design, exploring advanced AI courses can provide valuable skills. Visit Complete AI Training for resources on AI in science and research.