CryoLigATE

Cryo-electron microscopy is a mainstay of structure-based drug discovery, yet the density around a bound ligand is routinely worse resolved than the surrounding protein scaffold — in one β-galactosidase reconstruction the global density reached 1.5 Å while the ligand density resolved only to 3.0–3.5 Å. Deep-learning post-processing has improved overall map quality, but tools such as DeepEMhancer and EMReady (continued in EMReady2) were trained predominantly on protein density and learn helices, sheets, and loops. Applied to a binding pocket they generalize poorly to the more irregular chemical space of ligands, and have been shown to degrade the model-to-map Q-score there.

CryoLigATE, from Nandan Haloi, Rebecca J. Howard, and Erik Lindahl at Stockholm University and SciLifeLab, is built for that specific gap. It takes an experimental cryo-EM map plus a preliminary atomic model, automatically crops the sub-volume around a nominated ligand, and returns a sharpened local density suitable for confident model building.

Its position differs deliberately from co-folding predictors such as AlphaFold 3 and Boltz-2, which infer a complex from sequence and a SMILES string. CryoLigATE never receives a chemical identifier; it refines what the microscope measured, so its output stays anchored to experimental signal rather than to a learned prior over binding modes. The enhanced volume then feeds interpretation by hand or by automated model builders like CryoACE.

#Key Features

• Ligand-first training signal: The loss penalizes voxels inside the ground-truth ligand mask 30-fold more heavily than protein and solvent, concentrating capacity where the map is weakest.
• No chemical input at inference: The supplied coordinates only localize the region; the network sees just the local density and a protein occupancy mask, limiting user-induced bias.
• Single-command pipeline: cryoligate-infer handles extraction, resampling, and inference with no manual box curation, taking about a second per complex on a desktop RTX 3070.
• Tolerant of imperfect coordinates: Rotating the input ligand by 90°, shifting it 3 Å, or replacing it with a single carbon atom still recovered the expected density.
• Chemically broad coverage: Training spans drug-like molecules, lipids, steroids, carbohydrates, and nucleotides, not one scaffold family.

#Technical Details

The network is a 3D adaptation of the Swin-Conv UNet (SCUNet): three downsampling and three upsampling stages joined by additive skip connections, built from ConvTransBlocks that split channels between a local branch of two 3×3×3 convolutions with Filter Response Normalization and a global branch of 3D windowed multi-head self-attention over 4×4×4 voxel windows. Regular and shifted (2×2×2 offset) partitioning alternate, with a learned 3D relative position bias. Input and output are single-channel 64×64×64 grids resampled to 0.5 Å isotropic voxels — a 32 Å cube centered on the ligand centroid.

Training used 6,511 unique protein-ligand pairs from PDB and EMDB entries at 4.0 Å or better, capped at 50 representatives per ligand species; drug-like molecules make up 59.3%, lipids 12.2%, and nucleotides 9.7%, and membrane proteins dominate the targets (22.1% receptors, 10.1% transporters). Ground truth is a map forward-simulated from the deposited coordinates, and splits of 5,195/667/649 were drawn by clustering structures at 40% sequence identity to prevent leakage. The composite loss combines spatially weighted MSE (0.30), a concordance correlation coefficient on the ligand region (0.40), and a Dice plus Tversky shape term (0.30), trained for 200 epochs with AdamW on A100 GPUs. On the 649-complex held-out set, cases whose deposited ligand Q-score fell below 0.5 (n = 81) improved significantly, while the 558 well-resolved cases showed no significant degradation. Examples include NAD⁺ in MapSPARTA (0.19 → 0.41) and siponimod at the sphingosine-1-phosphate receptor (0.22 → 0.46).

#Applications

The beneficiaries are structural biologists and computational chemists interpreting consensus cryo-EM maps of drug targets, where marginal pocket density stalls ligand placement. The membrane-protein-weighted training set suits GPCRs, ion channels, and transporters, and the method handles lipidic and carbohydrate ligands that protein-focused sharpening ignores. It slots between reconstruction and model building: run it on a pocket, then use the enhanced volume to confirm a binding mode, resolve a functional group, or restore continuity through a flexible tail.

#Impact

CryoLigATE targets a bottleneck whole-map enhancement has not addressed, and its MIT license, PyPI packaging, and released cryoligate_v2.0.0.pth checkpoint make it usable without retraining. The authors are explicit about the risks: a negative control aimed at a region with no experimental density produced an amorphous blob, so the method must be restricted to refining pre-existing putative ligand density and never used to probe an empty pocket. It also treats the ligand as a single static entity, leaving conformational heterogeneity and partial occupancy unmodeled, and the supply of high-resolution deposited complexes caps how far the training strategy can scale. The work is a preprint, and evaluation rests on the authors' own held-out test set rather than a third-party benchmark; they recommend treating AI-enhanced maps as interpretive aids while final refinement stays tethered to the original density.