SynFrag: Synthetic Accessibility Predictor Based on Fragment Assembly Generation in Drug Discovery.

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Artificial Intelligence

Background:

• AI-driven molecular generation faces a "generation-synthesis gap," where most designed molecules are not experimentally synthesizable.
• Current synthetic accessibility (SA) methods, including CASP tools and ML models, have limitations in computational cost, chemical logic, or performance consistency.
• This gap hinders the application of AI-assisted drug design (AIDD).

Purpose of the Study:

• To develop an efficient and accurate SA prediction model to bridge the generation-synthesis gap in AIDD.
• To enable rapid, large-scale screening of computationally generated molecules for synthetic feasibility.
• To provide an interpretable tool for assessing synthetic difficulty in drug discovery workflows.

Main Methods:

• Developed SynFrag, a novel SA prediction model utilizing fragment assembly autoregressive generation.
• Employed self-supervised pretraining on millions of unlabeled molecules to learn dynamic fragment assembly patterns.
• Incorporated attention mechanisms to identify key reactive sites relevant to synthesis.

Main Results:

• SynFrag demonstrates consistent and high performance across diverse chemical spaces, including public benchmarks, clinical drugs, and AI-generated molecules.
• The model achieves subsecond prediction times, making it suitable for high-throughput screening.
• Attention mechanisms provide interpretability by highlighting critical reactive sites influencing synthetic difficulty.

Conclusions:

• SynFrag effectively addresses the generation-synthesis gap by providing computationally efficient and accurate SA predictions.
• The model enhances AIDD by enabling rapid filtering of synthesizable molecules, accelerating drug discovery pipelines.
• SynFrag offers a valuable, interpretable tool for both large-scale screening and detailed synthetic accessibility assessment.