PGFS++: Molecular Property Improvement under Synthesis and Diversity Constraints

A team of researchers has developed PGFS++, an advanced AI method designed to improve molecular properties such as drug-likeness and binding affinity in early-stage drug discovery. Unlike previous approaches, PGFS++ incorporates synthesis constraints directly into the optimization process, ensuring that generated molecules are not only effective but also feasible to produce in a lab. The method builds on the Policy Gradient for Forward Synthesis (PGFS) framework but addresses a critical limitation: PGFS relies on reactant embedding prediction, which indirectly guides reactant selection and can reduce learning effectiveness. PGFS++ improves this by explicitly representing reaction templates and second reactants, enabling more direct and efficient optimization. The work highlights the importance of balancing theoretical molecular improvements with practical synthesis considerations, a challenge that has long hindered the translation of AI-driven discoveries into real-world pharmaceutical applications.

The research demonstrates that PGFS++ achieves better performance in generating synthesizable molecules without sacrificing property improvements. This is particularly significant for drug discovery, where the gap between computational predictions and experimental feasibility often leads to wasted resources. By integrating synthesis constraints into the AI model, the method reduces the likelihood of generating molecules that, while theoretically promising, cannot be practically synthesized. The team’s findings suggest that future drug discovery pipelines could benefit from adopting such synthesis-aware AI approaches, potentially accelerating the development of new therapeutics.