Utilizing Low-Dimensional Molecular Embeddings for Rapid Chemical Similarity Search.

Area of Science:

• Computational Chemistry
• Cheminformatics
• Drug Discovery

Background:

• Nearest neighbor searching is crucial in chemistry, especially for drug discovery.
• Current brute-force methods are computationally expensive for large chemical databases.
• Existing advancements often rely on hardware or dataset-specific solutions lacking generalizability.

Purpose of the Study:

• To evaluate if low-dimensional chemical embeddings combined with k-d trees can enable fast nearest neighbor queries.
• To assess the performance of this approach on standard chemical similarity search benchmarks.
• To explore dimensionality reductions of chemical embeddings and a novel learned embedding (SmallSA).

Main Methods:

• Utilizing low-dimensional chemical embeddings.
• Implementing k-d tree data structures for efficient searching.
• Comparing different dimensionality reduction techniques and the SmallSA embedding.

Main Results:

• Searches on over one billion chemicals completed in under a second on a single CPU core.
• Achieved a speedup of five orders of magnitude compared to brute-force methods.
• The SmallSA embedding demonstrated competitive performance on chemical similarity benchmarks.

Conclusions:

• A combination of low-dimensional embeddings and k-d trees offers a highly efficient solution for chemical similarity searching.
• This framework significantly accelerates large-scale chemical database queries.
• The developed approach maintains high performance on standard benchmarks, advancing computational chemistry tools.