Scalable Similarity Search for Molecular Descriptors

TL;DR

This paper introduces SITAd, a novel efficient indexing method for similarity search in large chemical compound databases represented by molecular descriptors, significantly improving search performance.

Contribution

The paper presents SITAd, a new succinct data structure-based index for integer vector similarity search, extending binary-vector methods to molecular descriptors.

Findings

SITAd outperforms existing methods in large-scale experiments

The index is both time- and space-efficient

Effective for databases with over 40 million compounds

Abstract

Similarity search over chemical compound databases is a fundamental task in the discovery and design of novel drug-like molecules. Such databases often encode molecules as non-negative integer vectors, called molecular descriptors, which represent rich information on various molecular properties. While there exist efficient indexing structures for searching databases of binary vectors, solutions for more general integer vectors are in their infancy. In this paper we present a time- and space- efficient index for the problem that we call the succinct intervals-splitting tree algorithm for molecular descriptors (SITAd). Our approach extends efficient methods for binary-vector databases, and uses ideas from succinct data structures. Our experiments, on a large database of over 40 million compounds, show SITAd significantly outperforms alternative approaches in practice.