Loading…
Graph neural network that predicts magnesium ion binding sites on RNA structures, powering SAXS-based validation of RNA solution conformations.
No providers recorded yet. Browse all providers
IonNet is a graph neural network that predicts Mg²⁺ ion binding sites on RNA structures, from Edan Patt and Dina Schneidman-Duhovny at the Hebrew University of Jerusalem with Michal Hammel and Scott Classen at Lawrence Berkeley National Laboratory. It was published in Biophysical Journal in 2025.
RNA is highly charged, and divalent cations — magnesium above all — are essential to its folding and charge neutralization. Structure prediction methods do not place them, and experimental structure determination frequently misses them. That matters when a predicted RNA model is checked against a small-angle X-ray scattering profile measured in solution: magnesium has roughly twice the X-ray scattering length density of water, so its absence biases the computed profile. Existing ion predictors were thin — MetalionRNA relies on a statistical potential fitted on about 100 structures, and 3D convolutional approaches are slow, not rotation-invariant, and burdened by sparse voxels.
IonNet takes a graph approach instead. A probe placed on the RNA surface is classified as either a magnesium ion or a water molecule from the graph of RNA atoms in its neighborhood, which is rotation-invariant by construction and far cheaper than a voxel grid. The model is the learned component of SCOPER (Solution Conformation Predictor for RNA), the surrounding pipeline that combines IonNet's ion placements with kinematics-based conformational sampling and multi-state fitting against experimental SAXS data. IonNet occupies the same niche for RNA that Metalorian and SuperMetal occupy for protein metal-site prediction.
Training used roughly 1,000 PDB structures containing about 41,000 Mg²⁺ ions as positive examples, with water molecules and their atomic neighborhoods as negatives. Four-fold cross-validation gives a mean AUROC of 0.89. At inference, surface probes are generated with Connolly's surface method, each probe's 8 Å neighborhood is passed to the network, and the highest-confidence predictions are clustered. On the P4P6 domain, IonNet identifies probes near 11 of the 12 experimentally observed ions, flagging 84 of roughly 5,500 probes as magnesium; at a strict 0.9 confidence cutoff instead of 0.5, 6 of the 12 observed ions are correctly predicted from 30 predicted positions. The full SCOPER pipeline was benchmarked against 14 experimental SEC-SAXS datasets collected at SIBYLS beamline 12.3.1, three of them for RNAs with known crystal structures. It generates 1,000 conformations with KGSRNA, which perturbs geometry while preserving base pairing, predicts ions for each, selects the subset minimizing χ² by branch-and-bound, and applies MultiFoXS where one conformer cannot fit within the noise. For P4P6 the best four-conformation model reached χ² = 1.17 against 1.48 for a single conformation.
The audience is structural biologists validating RNA models against solution scattering. Given a starting structure from crystallography or a predictor such as DeepFoldRNA, RNAComposer or AlphaFold 3, SCOPER answers whether that model is consistent with the measured SAXS profile once ions and conformational plasticity are accounted for, and returns a corrected atomistic model including the ions. IonNet alone is useful wherever magnesium positions are needed on an RNA model — setting up molecular dynamics, interpreting a density map, or reasoning about a folding intermediate.
IonNet supplies a missing piece in the RNA structure workflow: predicted models have improved rapidly while the ions that stabilize them remain absent, and comparing such models to solution data without ions is systematically biased. The authors are candid about the boundaries. SCOPER is a validation tool, not a structure-determination method — its output depends heavily on the accuracy of the initial model. Fitting ion placement, conformational plasticity and hydration-layer parameters at once risks overfitting the SAXS data, so they fix the excluded volume and hydration parameters to 1.0 and recommend a high confidence threshold on IonNet predictions. Training labels carry noise, since Mg²⁺ is isoelectronic with water and Na⁺ and can be misassigned in deposited structures, and the model supports magnesium only. The code and checkpoint are archived on Zenodo with a citable DOI alongside the GitHub repository, though the repository itself carries no license file, so reuse terms are undefined. Seven of the fourteen SEC-SAXS datasets were collected under a proprietary agreement and are withheld from the SIMPLE SCATTERING deposition.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.