In the early stages of pharmaceutical research, screening compound collections is a crucial step for identifying biologically active molecules. Historically, the challenge lay in assembling chemical libraries that were as large as possible, in order to maximize the chances of discovering drug candidates. However, the rise of "on-demand" virtual libraries, which can contain billions or even trillions of theoretical compounds, raises practical difficulties: costs of acquisition, storage and maintenance, but also the heterogeneous quality of the molecules and the growing environmental footprint of the IT sector. As a result, the diversity, novelty and "drug-like" character of compounds are emerging as more relevant criteria than the mere size of the collection. It is within this context, and more specifically that of efficiency-conscious academic research projects, that the present work is situated.
The study describes a collaboration between computational and synthetic chemists aimed at designing chemical libraries suited to limited resources. Cheminformatics tools were developed to analyze an initial set of approximately 10,000 compounds and to derive two complementary libraries from it. The first, named eIMS, comprises 578 diverse compounds available in stock and immediately usable for screening. The second, vIMS, constitutes its virtual counterpart: it is built by adding various functional groups to the molecular scaffolds of eIMS, and then filtering out reactive or unusual structures to ensure the synthetic accessibility of the generated molecules.
Methodologically, the generation of new molecules relies on decomposing the central scaffolds into substitution points, to which R groups—categorized by the medicinal chemists according to their chemical nature (alkyls, amino acids, aryl and benzyl groups, carbonyls, heterocyclics, sulfonyls and small substituents)—are attached. The resulting combinations were filtered using simple molecular descriptors and rules excluding undesirable substructures, prior to three-dimensional conformer generation and standardization. The diversity and quality of the libraries were assessed using several approaches: ring system recovery curves, molecular shape representation, projection of Bemis-Murcko scaffolds into chemical space using the UMAP algorithm, enumeration of potential three- and four-point pharmacophores, and calculation of a synthetic accessibility score.
The analyses show that the vIMS library offers a broader range of binding motifs and improved drug-like characteristics compared to eIMS. Its novelty was highlighted through comparison with the virtual chemical space of commercial suppliers, notably Enamine's REAL space. This work thus proposes a pragmatic cheminformatics approach enabling the construction, from an existing collection, of both a diverse, directly screenable library and its synthetically accessible virtual extension.