Utilizing Low-Dimensional Molecular Embeddings for Rapid Chemical Similarity Search

, , , , In the field of chemistry, nearest neighbor-based similarity searching is a crucial task, particularly in applications like drug discovery. Traditional approaches to this task often rely on brute-force methods, which can be computationally expensive and time-consuming due to the vast size of modern chemical databases. Previous advancements in computational techniques for this task have been limited in their generalizability and efficiency. In this study, we explore the potential of combining low-dimensional chemical embeddings with a k-d tree data structure to achieve rapid nearest neighbor queries while maintaining high performance on standard chemical similarity benchmarks. Our approach addresses challenges faced by existing tools such as ChemMine and FPSim2, which utilize partitioning schemes based on Baldi bounds but are limited by binary embeddings and unequal partition sizes. To overcome these limitations, we propose SmallSA, a learned embedding generated using the SALSA framework trained on 1.4 million chemical compounds from the Enamine catalog. We compare SmallSA with commonly used Extended-Connectivity Fingerprints (ECFPs) and Extended-Connectivity Count Vectors (ECFCs) as well as their low-dimensional versions. Our results demonstrate that our approach significantly accelerates searches on large chemical databases, executing queries on over one billion chemicals in less than a second on a single CPU core – five orders of magnitude faster than brute-force methods. Furthermore, SmallSA achieves competitive performance on standard chemical similarity benchmarks compared to traditional fingerprinting methods like ECFPs and ECFCs. This highlights the potential of leveraging low-dimensional embeddings and k-d trees for efficient chemical similarity searching in drug discovery and other applications within the field of chemistry.

Introduction

Chemical similarity searching is a fundamental task in the field of chemistry, particularly in applications such as drug discovery. The goal of this task is to identify compounds with similar chemical structures and properties, which can aid in the development of new drugs or materials. Traditional approaches to chemical similarity searching rely on brute-force methods, which involve comparing each compound in a database to a query compound. However, with the ever-increasing size of modern chemical databases, these methods can be computationally expensive and time-consuming.

In recent years, there have been advancements in computational techniques for chemical similarity searching. These include partitioning schemes based on Baldi bounds used by tools like ChemMine and FPSim2. However, these methods are limited by binary embeddings and unequal partition sizes, leading to suboptimal performance.

To address these limitations, researchers have explored the use of low-dimensional chemical embeddings combined with data structures such as k-d trees for efficient nearest neighbor queries. In this study, we propose SmallSA – a learned embedding generated using the SALSA framework trained on 1.4 million compounds from the Enamine catalog – as an alternative approach for rapid nearest neighbor searches while maintaining high performance on standard chemical similarity benchmarks.

Methodology

The first step in our methodology was to train SmallSA using SALSA (Structure-Agnostic Latent Semantic Analysis), a deep learning-based method that learns low-dimensional representations of molecules without any prior knowledge about their structure or function. This allows for more generalizable embeddings compared to traditional fingerprinting methods like ECFPs (Extended-Connectivity Fingerprints) and ECFCs (Extended-Connectivity Count Vectors).

Next, we evaluated SmallSA's performance on standard chemical similarity benchmarks including Maximum Unbiased Validation (MUV), PubChem BioAssay AID1706 dataset and ChEMBL20 datasets. We also compared its performance with ECFPs and ECFCs, as well as their low-dimensional versions.

Finally, we tested the efficiency of SmallSA for nearest neighbor queries on large chemical databases. We compared its performance with brute-force methods and traditional fingerprinting methods like ECFPs and ECFCs.

Results

Our results demonstrate that SmallSA significantly accelerates searches on large chemical databases. It can execute queries on over one billion chemicals in less than a second on a single CPU core – five orders of magnitude faster than brute-force methods. Furthermore, SmallSA achieves competitive performance on standard chemical similarity benchmarks compared to traditional fingerprinting methods like ECFPs and ECFCs.

In terms of efficiency, our approach outperforms existing tools such as ChemMine and FPSim2 by utilizing k-d trees instead of partitioning schemes based on Baldi bounds. This allows for more equal partition sizes and eliminates the need for binary embeddings, resulting in improved performance.

Comparison with Traditional Fingerprinting Methods

We also compared SmallSA's performance with traditional fingerprinting methods like ECFPs and ECFCs. Our results show that while these methods have been widely used in chemical similarity searching, they are limited by their fixed-length representations which may not capture all relevant information about a molecule's structure or function. In contrast, SmallSA's learned embedding is able to capture more complex relationships between compounds due to its low-dimensional nature.

Generalizability

One major advantage of using SALSA-based embeddings is their generalizability across different datasets. This was demonstrated through our evaluation on various benchmark datasets where SmallSA achieved competitive or even better performance compared to other methods without any dataset-specific tuning or parameter optimization.

Conclusion

In conclusion, our study highlights the potential of combining low-dimensional embeddings generated using deep learning techniques with data structures such as k-d trees for efficient chemical similarity searching. Our approach, SmallSA, outperforms traditional fingerprinting methods in terms of efficiency and generalizability while maintaining competitive performance on standard benchmarks. This has significant implications for applications such as drug discovery, where rapid identification of similar compounds can aid in the development of new drugs or materials. Further research in this area could lead to even more advanced techniques for chemical similarity searching and other tasks within the field of chemistry.