Loading…
A 3D UNETR++ transformer that enhances experimental cryo-EM density maps, improving map-model agreement and automated de novo structure modeling.
No providers recorded yet. Browse all providers
A cryo-EM reconstruction is rarely of uniform quality: the same map can be crisp through a rigid core and smeared into noise at a flexible periphery, so a single global B-factor correction either over-sharpens one region or leaves the other unreadable. Local sharpening adapts to that variation, but each method carries a prerequisite the experimenter usually lacks — LocScale needs an atomic model already in hand, and LocSpiral and LocalDeblur need a mask separating macromolecule from background.
CryoTEN, from Joel Selvaraj, Liguo Wang and Jianlin Cheng at the University of Missouri and Brookhaven National Laboratory, replaces that prerequisite with a learned map-to-map transform. It is a 3D UNETR++ style transformer trained once on 1,295 deposited experimental maps whose targets are simulated maps computed from the matching PDB structures with a reference Gaussian function. The network learns what an experimental map would look like if the deposited atomic model were ground truth, and applies that transform to any new map at inference — no mask, no prior model, nothing re-fit per map.
What distinguishes it from the other learned map enhancers is efficiency rather than peak accuracy. Efficient paired attention, inherited from UNETR++, shares the query and key projections between the spatial and channel attention branches. Enhancement averages 1.66 minutes and 3.41 GB of GPU memory per map, against 19.65 minutes for EMReady — since succeeded by EMReady2 — 43.27 for DeepEMhancer and 340.41 for EM-GAN, and CryoTEN ranks second to EMReady in map quality. It comes from the same lab as Cryo2Struct2, which tackles the adjacent problem of building the atomic model itself.
The 29.5-million-parameter network is four encoder-decoder pairs with U-Net skip connections. Each encoder pairs a downsampling convolution with three transformer layers carrying efficient paired attention; each decoder pairs a transposed convolution with three more. A parallel residual convolution branch reads the raw input block and joins the last decoder's output before a final upsampling convolution. Training data came from RCSB PDB structures built from single-particle maps at 2-7 Å, filtered on map-model cross-correlation and clustered by MMseqs2 at 30% sequence identity, giving 1,521 non-redundant map-structure pairs split 1,295 / 76 / 150 into training, validation and test. Maps were resampled to a 1 Å grid and cut into blocks of 64 voxels a side, randomly cropped to 48. Training ran 827 epochs on four NVIDIA A40 GPUs under a masked mean squared error loss.
On the 150-map test set, map-model FSC@0.143 improves from 3.55 Å to 2.48 Å and FSC@0.5 from 6.38 Å to 3.79 Å, with 99.33% of maps improving on the former. CC_box rises from 0.7231 to 0.8512 and CC_peaks from 0.6439 to 0.7480, while CC_mask falls slightly to 0.7685 and the Q-score improves marginally to 0.5425. Structures built automatically from the enhanced maps are better: phenix.map_to_model residue coverage rises from 61.87% to 70.74% and sequence match from 34.37% to 37.38%, MAINMAST coverage from 72.31% to 80.0%.
The intended use is a preprocessing step before automated de novo model building, where map interpretability rather than fidelity is what matters. Accepting both post-processed primary maps and raw half maps, it drops in after reconstruction, and its runtime makes it practical across a whole dataset. The authors are explicit about the boundary: the network modifies density values directly rather than amplitudes in Fourier space, so enhanced maps should not be deposited in the EMDB or used to compute map resolution.
CryoTEN was posted to bioRxiv in September 2024 and published in Bioinformatics in February 2025. Its contribution is less a new ceiling on map quality than a demonstration that a well-chosen attention mechanism delivers near-best enhancement at an order of magnitude less compute, which makes it usable inside a routine pipeline rather than as a special-case rescue tool. It is candid about where it falls short: a slightly reduced CC_mask, marginal Q-score gains, and residual risk to water and ligand density. Source code is MIT licensed and the trained checkpoint is deposited on Zenodo under CC-BY 4.0.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.