PubChem3D: Shape compatibility filtering using molecular shape quadrupoles.

Area of Science:

• Computational Chemistry
• Cheminformatics
• Drug Discovery

Background:

• PubChem's 3-D neighboring identification relies on computationally intensive shape overlap calculations.
• Existing pre-filters based on molecular volume remove ~65% of conformer pairs but can be improved.
• The study explores using additional shape descriptors to enhance filtering efficiency.

Purpose of the Study:

• To investigate if shape descriptors like length, width, and height can predict shape incompatibility between compound conformers.
• To develop and evaluate new filtering methods for PubChem's 3-D neighboring process.
• To reduce computational costs associated with identifying 3-D similar conformer pairs.

Main Methods:

• Computed shape descriptors (volume, shape quadrupoles representing length, width, height) for billions of 3-D conformer pairs.
• Binned descriptor values and analyzed frequencies to create similarity maps.
• Tested the efficiency of these shape descriptor filters on large datasets, including real-world chemical data.

Main Results:

• Developed shape descriptor maps (Qx, Qy, Qz) that effectively filter incompatible conformer pairs with high efficiency (40.4% true negatives at shape Tanimoto ≥ 0.8).
• Monopole volume-based quadrupole filters (Qxyz) showed the best performance.
• Real-world application reduced PubChem 3-D neighboring computation costs by 24-38%.

Conclusions:

• Basic shape descriptors (size, length, width, height) are effective for identifying shape-incompatible conformer pairs.
• Implementing these descriptors as filters significantly improves the efficiency of PubChem's 3-D neighboring computation.
• Throughput increased by 31%, from 154,000 to 202,000 conformer pairs per second per CPU core.