PharmacoForge: pharmacophore generation with diffusion models.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery

Background:

• Structure-based drug design (SBDD) utilizes machine learning (ML) for virtual screening and de novo design.
• Current ML methods face limitations in screening speed and the synthetic accessibility of generated molecules.

Purpose of the Study:

• Introduce PharmacoForge, a diffusion model for generating 3D pharmacophores conditioned on protein pockets.
• Improve virtual screening by generating valid, commercially available ligand queries.
• Address limitations of existing ML-based drug design strategies.

Main Methods:

• Developed PharmacoForge, a conditional diffusion model for 3D pharmacophore generation.
• Evaluated PharmacoForge against automated pharmacophore generation using the LIT-PCBA benchmark.
• Assessed generated pharmacophores via retrospective screening on the DUD-E dataset.
• Compared PharmacoForge-derived ligands against de novo generated ligands using docking and strain energy analysis.

Main Results:

• PharmacoForge outperformed existing pharmacophore generation methods on the LIT-PCBA benchmark.
• Ligands identified through PharmacoForge pharmacophores showed comparable performance to de novo designed ligands in docking studies.
• Molecules generated via PharmacoForge exhibited lower strain energies than de novo designed molecules.

Conclusions:

• PharmacoForge offers an effective approach for generating 3D pharmacophores, enhancing structure-based drug design.
• The model generates valid and synthetically accessible molecular queries, overcoming limitations of current generative models.
• PharmacoForge represents a significant advancement in ML-driven drug discovery, improving efficiency and reliability.