FFLOM: A Flow-Based Autoregressive Model for Fragment-to-Lead Optimization.

Area of Science:

• Computational Chemistry
• Artificial Intelligence in Drug Discovery
• Molecular Modeling

Background:

• Deep generative models show promise for fragment-based drug design (FBDD).
• Current models struggle with generating molecules with specific properties, particularly with limited data.
• Designing linkers and R-groups is crucial for optimizing drug candidates.

Purpose of the Study:

• To introduce FFLOM, a novel flow-based autoregressive model for linker and R-group design.
• To evaluate FFLOM's performance in generating molecules with desired properties.
• To assess FFLOM's applicability in practical drug design scenarios.

Main Methods:

• Development of a flow-based autoregressive model named FFLOM.
• Large-scale benchmark evaluation using ZINC, CASF, and PDBbind datasets.
• Assessment of molecular generation metrics: validity, uniqueness, novelty, and recovery.

Main Results:

• FFLOM achieved state-of-the-art performance across benchmark datasets.
• The model demonstrated high molecule recovery rates (over 92% on PDBbind).
• FFLOM successfully reconstructed ground-truth compounds and generated novel fragments with potentially higher binding affinity.

Conclusions:

• FFLOM is a powerful tool for fragment-based drug design, particularly in low-data regimes.
• The model shows significant potential in practical applications like fragment linking and PROTAC design.
• FFLOM enables the generation of novel molecular fragments, advancing drug discovery efforts.