χ-SOM: emergent self-organizing maps for large chemical space visualization
Abstract
The exploration and visualization of chemical space using dimensionality reduction (DR) methods are fundamental to cheminformatics, as they enable the exploitation of the similar property principle. However, most methods, including t-SNE and UMAP, struggle to preserve valuable distance and density properties from ambient space in the latent space representation. The emergent self-organizing map (ESOM), an often-ignored extension of the self-organizing map (SOM) algorithm, demonstrates projection behavior that circumvents the disadvantages of these methods. As a suitable implementation of this algorithm for large cheminformatic datasets is not available, we present χ-SOM, the ChemInformatics (ChI)-SOM library, a high-performance framework for training emergent self-organizing maps on large molecular datasets and for interactively exploring the resulting projections. χ-SOM integrates an optimized implementation of the SOM algorithm with CPU and CUDA backends, an on-disk HDF5-based datastore with PyTorch DataLoader support for efficient parallel access to millions of fingerprints, and a graphical user interface tailored to chemical structure visualization and property overlays. We compared our implementation with t-SNE and UMAP on synthetic and chemical data benchmarks that probe density and inter‑cluster distance preservation. On the synthetic benchmarks, ESOM retained both properties where t‑SNE and UMAP did not. On the fingerprint-based equivalents, all three methods performed comparably. χ-SOM was also benchmarked against established SOM libraries across varying dataset sizes and fingerprint lengths, demonstrating competitive or superior performance. As a final proof-of-concept, we successfully trained an ESOM of the ChEMBL33 database (1.7 million compounds; 5.1 million neurons), to our knowledge the largest SOM reported to date, and show that local neighborhood relationships on the map align well with medicinal chemists’ intuition of chemical similarity. χ-SOM thus provides a practical, extensible tool for large scale exploratory analysis of chemical space.
Related
- Efficient visualization of chemical space
- Lost in Chemical Space: Local Open-Source Large Language Models Display Model-Dependent Adherence to Reference Data in Molecular Generation
- SmileyLlama: modifying large language models for directed chemical space exploration
- A New Perspective in Chemical Probe Discovery: Expanding Virtual Screening into Metal Complex Chemical Space
- IQARIS: An Atlas of Interacting Quantum Atoms across Chemical Space
BibTeX
@article{kaminski2026χsom,
title = {{χ-SOM: emergent self-organizing maps for large chemical space visualization}},
author = {Johannes Kaminski and Oliver Koch},
journal = {Journal of Cheminformatics},
year = {2026},
month = oct,
doi = {10.1186/s13321-026-01317-z},
url = {https://link.springer.com/article/10.1186/s13321-026-01317-z},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - χ-SOM: emergent self-organizing maps for large chemical space visualization
AU - Johannes Kaminski
AU - Oliver Koch
PY - 2026
DA - 2026/10/18
JO - Journal of Cheminformatics
DO - 10.1186/s13321-026-01317-z
UR - https://link.springer.com/article/10.1186/s13321-026-01317-z
AB - The exploration and visualization of chemical space using dimensionality reduction (DR) methods are fundamental to cheminformatics, as they enable the exploitation of the similar property principle. However, most methods, including t-SNE and UMAP, struggle to preserve valuable distance and density properties from ambient space in the latent space representation. The emergent self-organizing map (ESOM), an often-ignored extension of the self-organizing map (SOM) algorithm, demonstrates projection behavior that circumvents the disadvantages of these methods. As a suitable implementation of this algorithm for large cheminformatic datasets is not available, we present χ-SOM, the ChemInformatics (ChI)-SOM library, a high-performance framework for training emergent self-organizing maps on large molecular datasets and for interactively exploring the resulting projections. χ-SOM integrates an optimized implementation of the SOM algorithm with CPU and CUDA backends, an on-disk HDF5-based datastore with PyTorch DataLoader support for efficient parallel access to millions of fingerprints, and a graphical user interface tailored to chemical structure visualization and property overlays. We compared our implementation with t-SNE and UMAP on synthetic and chemical data benchmarks that probe density and inter‑cluster distance preservation. On the synthetic benchmarks, ESOM retained both properties where t‑SNE and UMAP did not. On the fingerprint-based equivalents, all three methods performed comparably. χ-SOM was also benchmarked against established SOM libraries across varying dataset sizes and fingerprint lengths, demonstrating competitive or superior performance. As a final proof-of-concept, we successfully trained an ESOM of the ChEMBL33 database (1.7 million compounds; 5.1 million neurons), to our knowledge the largest SOM reported to date, and show that local neighborhood relationships on the map align well with medicinal chemists’ intuition of chemical similarity. χ-SOM thus provides a practical, extensible tool for large scale exploratory analysis of chemical space.
SN - 1758-2946
ER -