Loading…
Ligand identification in cryoEM and X-ray density maps, classifying a density blob into one of 219 ligand groups from its 3D point cloud shape.
No providers recorded yet. Browse all providers
When a structural biologist sees an unexplained blob of density in an X-ray or cryoEM map, deciding which small molecule it represents is a manual, error-prone judgment. Buffer components, cryoprotectants, and cofactors all produce plausible-looking shapes, and the literature contains ligands modelled into density that could not support them. Automated help has existed for X-ray crystallography — iterative fitting and feature-engineered machine learning over blob descriptors — but nothing comparable existed for cryoEM, where Coulomb potential maps are processed differently by different laboratories and no standard density threshold applies.
The ChimeraX Ligand Recognizer, from the Institute of Computing Science at Poznań University of Technology with collaborators at the University of Colorado Boulder, addresses that gap with an end-to-end deep network. Rather than computing descriptors of a density fragment, it thresholds the map, resamples the fragment onto a 0.2 Å voxel grid, converts it to a 3D point cloud, and classifies the cloud directly. Three point-cloud architectures were evaluated — RiConv++, TransLoc3D, and MinkLoc3Dv2 — and the deployed model is the MinkLoc3Dv2 variant, a sparse-convolutional network with a U-Net-like feature pyramid, efficient channel attention, and a NetVLAD global descriptor replacing the original pooling layer.
It is a closed-set classifier, which matters for how it should be used: the output is a probability distribution over 219 ligand groups, and a component outside that vocabulary cannot be recognized. The work was published in Bioinformatics at the end of 2024.
Training data came from a Protein Data Bank snapshot yielding 957,855 ligand blobs extracted from Fo-Fc maps at a 2.8σ isosurface, filtered on resolution, real-space correlation, and occupancy down to 696,887. Because several ligands are indistinguishable by density alone, chemically similar ligands were clustered by atom count, ring count, connectivity, chirality, and atomic numbers; the 218 groups with at least 100 instances were kept and everything else pooled into a single rare class, giving 219 output classes. For cryoEM, 6,103 EM-derived models at 4.0 Å resolution or better were paired with their EMDB maps, and difference maps against ligand-free models yielded 34,671 blobs labelled with the same groups.
On a stratified X-ray test set of 208,896 ligands, MinkLoc3Dv2 reached 0.660 accuracy and 0.946 top-10 accuracy with a mean correct-prediction rank of 3.627, matching the feature-engineered CheckMyBlob baseline on accuracy (0.672) while improving top-10 and mean rank; the other two point-cloud architectures underperformed both. On cryoEM ligands under three-fold cross-validation, training on cryoEM alone gave 0.639 accuracy, while mixed X-ray and cryoEM training gave 0.691 accuracy, 0.974 top-10 accuracy, and a mean rank of 2.589. Inference takes 0.103 seconds per ligand against 3.991 for the descriptor-based baseline on the same CPU core.
The direct users are structural biologists interpreting maps during model building, particularly in structure-guided drug design, where a misassigned ligand propagates into every downstream conclusion. Fragment screening campaigns benefit from the throughput: hundreds of datasets can be triaged before a human looks at any of them. The blob datasets are archived on Zenodo, reusable for method development and not only for inference.
This is the first deep learning approach to recognize ligands in cryoEM maps and the first to work from raw density in either modality without hand-designed features. The cross-modality result is the more transferable finding: X-ray density and cryoEM potential differ physically but are similar enough in shape at good resolution that one supplies useful supervision for the other. The paper is candid about what stands in the way of more. Automatic thresholding is the weak link — the authors show a case where a poorly chosen threshold turns a correct heme identification into an incorrect one — cryoEM ligand quality metrics are immature, and the closed 219-group vocabulary bounds what the model can ever say. The Bioinformatics article is open access under CC BY 4.0, but no license is attached to the code repository or the released weights.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.