The systematic exploration of Chemical Compound Space (CCS) has become a cornerstone of modern chem- istry, driven by the rapid growth of machine learning (ML) approaches. The emergence of large-scale chemical datasets such as GDB17, (1) QCML, (2) or OMol25, (3) to name a few, has opened unprecedented opportunities to uncover structure–property relationships and design novel molecules across chemistry, materials science, and drug discovery. Yet, this abundance also poses challenges: many datasets lack transparent generation protocols, internal consistency, or comprehensive information about their chemical coverage, a nontrivial issue given the multidimensional nature of chemical space. Beyond elemental composition and topology, meaningful coverage must also capture additional factors such as conformational diversity (e.g., equilibrium vs. reactive geometries) and the inclusion of uni- and multimolecular systems, among others. As a result, redundancies and biases often arise, obscuring chemical logic and hindering the assessment of model transferability and performance, ultimately limiting data-driven discovery. In this work, we introduce ChemDBX, a computational platform for the systematic construction, indexing, and exploration of molecular databases. ChemDBX imports existing datasets in XYZ or HDF5 format, auto- matically indexes and deduplicates entries, and characterizes their elemental composition, molecular formula, and aggregation degree. It performs hierarchical structural classification and conformational clustering to map regions of CCS and evaluate dataset diversity. Beyond indexing, ChemDBX computes local geometrical and chemical descriptors-including functional groups and atomic environments-providing a multiscale view that links molecules to their structural building blocks. For supramolecular systems, it applies graph-based decom- position to identify components and supports user-defined functional classifications. Finally, ChemDBX integrates molecular search tools, interactive analysis and visualization of molecular and quantum-mechanical properties, and modules for tailor-made database generation through both top-down and bottom-up strategies. Equipped with a graphical user interface (GUI) and a command-line interface (CLI), this unified framework enables robust and transparent exploration of CCS, while supporting the informed creation of novel databases aligned with active learning protocols.
Dr. Miguel Gallegos