We're excited to announce a new feature on ChEMBL compound pages: an interactive 3D viewer showing a low-energy, quantum-mechanically informed conformer for each compound, wherever one is available. This is part of a collaboration with the Isayev Lab at Carnegie Mellon University who have developed ChEMBL3D.
Until now, ChEMBL compound pages have shown molecules as flat 2D structures. A 2D depiction is great for seeing connectivity, but it says nothing about a molecule's actual shape in three dimensions, and shape is often what determines whether a compound fits a binding pocket, how it interacts with a target, or how its stereochemistry affects potency. With this release, you can rotate, zoom, and explore a real 3D conformer directly on the compound page, view it alongside the familiar 2D structure, and download the coordinates as an SDF file.
What's in the current delivery
For the compound pages you'll see live now, the ChEMBL3D team matched the available ChEMBL3D and newly-generated LoQI structures (see details below) to ChEMBL v37 identifiers, providing a single lowest-energy, validated conformer per compound. Calculations for additional ChEMBL v37 molecules are ongoing on their side, and we'll continue to roll out updated deliveries as coverage expands, so if a compound doesn't have a 3D structure yet, it may well get one soon.
It's worth being clear about what these structures are: they are computationally generated, low-energy conformers of biologically-likely protonation states, not experimentally determined structures (for example, from X-ray crystallography). They represent a plausible, energetically favorable 3D shape for the isolated molecule, rather than a structure observed in a specific bound or crystalline state. Bringing experimentally derived structures into ChEMBL is something we're looking at for the future, and we'll share more as those plans develop.
Try it out
Not every ChEMBL compound has a 3D conformer available, but where one exists, you'll find a "View 3D" button underneath the 2D structure on the compound page. Clicking it opens the 3D conformer rendered with Mol*, where you can freely rotate the structure and zoom in or out using your mouse scroll wheel or by dragging with two fingers on a trackpad, just as you would in most other 3D viewers. If you lose your bearings, you can always reset the zoom, re-orient the axes, or reset the axes back to their default position. You can also switch back to the 2D view, or download the conformer as an SDF file for use in your own tools.
3D visualization tool with Imatinib (CHEMBL941),Where the data comes from
The conformers powering this feature come from ChEMBL3D, a large-scale dataset of quantum-mechanically accurate molecular geometries developed by Filipp Nikitin, et al., described in their preprint "Scalable Low-Energy Molecular Conformer Generation with Quantum Mechanical Accuracy" (ChemRxiv, July 2026). The group generously shared their data with the ChEMBL team, together with structures generated by their new generative model, LoQI, so that we could bring 3D structures to as many ChEMBL compounds as possible.
How the conformers are generated
Broadly, the pipeline behind ChEMBL3D works like this:
- Starting point: around 2.1 million molecules from the ChEMBL database were taken as the source set.
- Protonation and stereochemistry enumeration: using OpenEye's toolkit, ionized and de-ionized protomers were generated for each molecule, and any missing stereocenters were enumerated so that all reasonable stereoisomers were represented.
- Initial 3D structures: OpenEye's Omega Classic tool was used to generate a first set of candidate 3D conformers for each molecule (up to 200 per molecule).
- Quantum-level refinement: each candidate conformer was then geometry-optimized using AIMNet2, a neural network potential trained to reproduce ωB97M-D3/def2-TZVPP density functional theory energies with implicit water solvation. This step pushes the initial force-field-like geometries toward near-quantum-mechanical accuracy without the enormous cost of running DFT on every structure directly.
- Filtering and deduplication: duplicate conformers were removed, and only geometries within 6 kcal/mol of the lowest-energy structure found for each molecule were retained.
The result is a dataset of roughly 280 million optimized conformers spanning about 1.9 million unique ChEMBL molecules, several orders of magnitude larger than earlier 3D conformer resources such as GEOM.
On top of this dataset, the Isayev group also trained LoQI (Low-energy QM-Informed conformer generator), a stereochemistry-aware diffusion model that learns to generate low-energy 3D structures directly, without relying on the traditional heuristic conformer search used by classical tools. In their benchmarks, LoQI recovered near-minimum-energy conformers substantially more often than classical generators, while correctly preserving R/S and E/Z stereochemistry, and performed well even on difficult cases such as macrocycles and taxol.
We thank Filipp Nikitin, Olexandr Isayev, and the rest of the ChEMBL3D/LoQI team for sharing their data and continuing to work with us to expand coverage. For the full technical details behind the dataset and model, see their preprint on ChemRxiv.
Comments