MolPIF: a parameter interpolation flow model for molecule generation.

Area of Science:

• Computational chemistry
• Molecular modeling
• Drug discovery

Background:

• Deep generative models show promise for structure-based drug design (SBDD).
• A key challenge is unifying continuous atomic coordinates with discrete atom types.
• Existing methods struggle with heterogeneous molecular data, limiting generative model performance.

Purpose of the Study:

• To introduce MolPIF, a novel parameter interpolation flow mechanism.
• To unify the generation of continuous and discrete molecular variables in SBDD.
• To enhance the capture of atomic contexts through geometry-enhanced learning.

Main Methods:

• MolPIF interpolates between distributions in parameter space, not sample space.
• Theoretically recovers Wasserstein-2 optimal transport for continuous coordinates.
• Establishes Fisher-Rao geodesics for discrete atom types, incorporating geometry-enhanced learning.

Main Results:

• MolPIF outperforms existing methods on the CrossDocked2020 dataset.
• Demonstrates superior binding affinity, chemical validity, and geometric fidelity.
• Shows strong performance in chemical space coverage and lead optimization.

Conclusions:

• MolPIF establishes a robust paradigm for SBDD by unifying molecular variable generation.
• The method offers flexibility in prior distribution selection.
• MolPIF advances generative modeling for protein-ligand complex design.