Using spIsoNet to address the preferred-orientation problem in cryoEM reconstructions

Abstract

Cryogenic electron microscopy (cryoEM) is now routinely used for high-resolution structure determination of biological macromolecules. However, many biological specimens exhibit varying degrees of preferred orientation on cryoEM grids, resulting in uneven sampling of three-dimensional Fourier space. This orientation bias produces anisotropic reconstruction artifacts and, in severe cases, can exacerbate particle misalignment during iterative refinement, thereby limiting the success rate of near-atomic resolution cryoEM structure determination. This protocol provides a practical guide for applying spIsoNet, a self-supervised deep-learning method, to mitigate preferred-orientation issues in cryoEM reconstructions. We describe two complementary workflows: (1) map Anisotropy Correction to correct anisotropic artifacts of cryoEM maps and (2) particle Misalignment Correction, which integrates spIsoNet with RELION external reconstruction to improve particle-pose estimation. We demonstrate these workflows using two influenza hemagglutinin (HA) trimer datasets representing moderate and severe degrees of preferred-orientation bias. The protocol includes installation instructions, parameter-selection guidance, quality-control checkpoints and troubleshooting advice, and can typically be completed in ~7 hours on a workstation equipped with four NVIDIA A100 GPUs. Together, these workflows provide step-by-step guidance for using the open-source spIsoNet software to mitigate the preferred-orientation problem directly from experimental data.

Introduction

The preferred-orientation problem in cryoEM

Single-particle cryogenic electron microscopy (cryoEM) is widely used in structural biology to resolve high-resolution structures of biological macromolecules in near-native states. Uniformly distributed particle orientations in vitrified ice are critical for complete Fourier-space sampling and for achieving isotropic resolution in single-particle cryoEM reconstructions^{1, 2}. However, during cryoEM sample preparation, specimens often adsorb preferentially to the air-water interface (AWI) or the support-film interface, leading to preferred orientations. Such specimen adsorption can bias particle orientations (i.e., “preferred orientation”) and may also promote partial dissociation or even denaturation of biological macromolecules³.

Preferred orientation causes missing or under-sampled regions in Fourier space, leading to direction-dependent resolution and anisotropic reconstruction artifacts. In real-space cryoEM maps, this anisotropy can be apparent as density that appears well defined from one viewing direction but streaky, elongated or poorly resolved from orthogonal directions. Typical manifestations include overall density elongation, distorted secondary-structure features, and discontinuous side-chain or nucleobase densities⁴. When an anisotropic map is used as the reference during iterative particle refinement, pose-estimation errors may accumulate, causing the refinement to converge to a suboptimal reconstruction or even fail^{5, 6}. Therefore, a general, specimen-independent method remains highly desirable for routine high-resolution cryoEM structure determination and downstream structure-guided studies.

Methods available to assess, alleviate or mitigate the problem

Several computational tools are available for assessing preferred orientation and its consequences in cryoEM reconstructions. Map-based metrics evaluate anisotropy directly from the reconstructed map. For example, the three-dimensional Fourier shell correlation (3DFSC) metric estimates directional variations in resolution and reports map anisotropy using 3DFSC sphericity⁵; more recently, Fourier shell occupancy (FSO) and the Bingham test (BT) have been introduced as complementary descriptors of resolution anisotropy in cryoEM maps⁷. The FSO curve is evaluated across increasing spatial-frequency shells, with values close to 1 indicating near-uniform isotropy and lower values indicating increasing anisotropy. BT complements FSO by testing whether the reliable directions are uniformly distributed on the sphere, helping distinguish directional anisotropy from a more uniform loss of high-resolution signal⁷. Orientation-distribution-based methods, such as cryoEF, instead assess preferred orientation from assigned particle Euler angles. CryoEF computes an orientation-distribution efficiency metric, E_od, which characterizes how well a given angular distribution can provide uniform Fourier-space coverage and isotropic resolution. Lower E_od values indicate that non-uniform angular sampling is more likely to produce direction-dependent resolution loss and anisotropic features in the reconstruction⁸. These tools are useful for diagnosis, but additional correction strategies are needed when the goal is to improve reconstructions affected by the underlying orientation bias.

Toward this end, earlier efforts focused on alleviating the preferred orientation problem through experimental strategies. They include chemical modification of specimens⁹, the addition of surfactants to reduce particle adsorption at the air–water interface (AWI)^10–13; and physical approaches such as the use of cryoEM grids coated with continuous support films^{3, 14–18}, time-resolved vitrification devices^19–21, data collection from thicker ice regions, and tilted-stage imaging^{5, 22, 23}. These approaches can be effective in individual cases, but they are often time-consuming and require additional trial-and-error sample screening for optimization. They may be accompanied by trade-offs, including increased background signal, altered particle behavior at newly introduced support-water interfaces, exacerbated beam-induced motion, or larger defocus gradients in tilted datasets. Computational methods therefore provide a complementary route, particularly for datasets that have already been collected or for samples for which experimental re-optimization is difficult.

Recent efforts have led to the development of computational methods that mitigate reconstruction artifacts associated with preferred orientation. Our group developed spIsoNet⁴ that provides an end-to-end, self-supervised deep learning method for processing preferred-orientation-limited single particle cryoEM and cryogenic electron tomography (cryoET) subtomogram averaging (STA) datasets. The method is implemented through two workflows to overcome the preferred orientation problem: (1) in the map Anisotropy Correction workflow, users start from half-maps, a solvent mask and an anisotropy descriptor, such as a 3DFSC volume, to generate corrected half-maps for downstream map post-processing and model building; and (2) in the particle Misalignment Correction workflow, spIsoNet is integrated with RELION external reconstruction so that spIsoNet-corrected half-maps can guide particle pose estimation during iterative refinement. Similar to the idea of the Misalignment Correction workflow of spIsoNet, cryoPROS²⁴ tackles the preferred orientation problem by overcoming mis-alignment with evenly sampled templates called “auxiliary particles”. It uses a self-supervised deep generative model to generate such auxiliary particles with more balanced orientation distributions and co-refines them with experimental particles to improve pose estimation in single-particle cryoEM²⁴. Along a similar principle of recovering missing information, AR-Decon²⁵ uses deconvolution of a point spread function that captures the preferred orientation of the sample to improve map anisotropy without using a deep learning framework. Two other computational tools, DeepEMhancer²⁶ and EMReady²⁷ are supervised deep-learning map post-processing tools trained on experimental cryoEM maps paired with model-informed target densities, and are subsequently applied to enhance the visual interpretability of cryoEM maps of interest. In addition, the signal-to-noise ratio iterative reconstruction method (SIRM) aims to mitigate preferred-orientation induced map distortions by removing misaligned particles, followed by dual-space-constrained iterative reconstruction guided by the signal-to-noise ratio of Fourier components²⁸.

Since the publication of the original spIsoNet paper in 2025⁴, the method has attracted widespread interest, as reflected by its growing number of citations. Nonetheless, many users have expressed interest in a more accessible illustration of its practical implementation in a representative use case. Here, we provide a detailed operational protocol for two spIsoNet workflows, each covering input preparation, parameter selection, quality-control checkpoints, troubleshooting and downstream validation.

Applications

The spIsoNet package comprises two modules: (1) map Anisotropy Correction and (2) Anisotropy Correction-powered particle Misalignment Correction. These modules can be applied independently or sequentially, depending on whether the primary failure mode is anisotropic map distortion, particle pose estimation errors, or both. When both map anisotropy and pose estimation errors are evident, we recommend applying particle Misalignment Correction first, followed by map Anisotropy Correction to further improve map isotropy and interpretability.

SpIsoNet can be applied to both single particle analysis (SPA) in cryoEM and subtomogram averaging (STA) in cryoET. In SPA, spIsoNet has been used to improve map isotropy and particle-pose estimation in datasets affected by preferred orientation. A representative example is the previously intractable untilted influenza hemagglutinin (HA) trimer dataset EMPIAR-10096, which yielded a near-atomic resolution reconstruction after spIsoNet particle Misalignment Correction. In STA, spIsoNet has also been shown to improve reconstructions when non-uniform particle orientations compromise map quality⁴. Beyond these benchmark datasets, several cryoEM and cryoET studies have used spIsoNet to improve map quality and interpretability in preferred-orientation biased datasets^29–33.

Limitations

spIsoNet does not generate high-resolution information beyond what is supported by the experimental cryoEM data. The Misalignment Correction workflow is also more computationally demanding than standard RELION refinement because it is embedded in an external-reconstruction loop, in which the spIsoNet model is retrained at each RELION refinement iteration. For cryoEM datasets with extremely limited angular coverage, such as those dominated by one or a few preferred views with substantial under-sampling in Fourier space, the ability of spIsoNet to improve map isotropy may be limited. In such cases, the situation can be mitigated by combining spIsoNet with a straight-forward experimental strategy to alleviate the severe preferred-orientation problem, such as using a single tilt angle within the 20–30° range⁵, using cryoEM grids coated with a continuous support film³, or adding a mild detergent to the sample¹⁰.

Overview of the spIsoNet pipeline

This protocol presents practical workflows established through testing across multiple cryoEM datasets (Fig. 1 and Table 1). Procedure 1, map Anisotropy Correction, takes as input (i) two cryoEM half-maps (Fig. 2a-b), (ii) a solvent mask (Fig. 2c-d), and (iii) a descriptor of directional-resolution anisotropy, for which we used a 3DFSC volume (Fig. 2e-f). A U-Net-style model is trained in a self-supervised manner to generate anisotropy-corrected half-maps that can be used for downstream post-processing, visualization, and atomic-model building. Procedure 2, particle Misalignment Correction, incorporates Anisotropy Correction into an iterative RELION auto-refine workflow. After each refinement iteration, the half-maps are corrected using spIsoNet and then low-pass filtered to the current gold-standard FSC resolution before being used as reference maps for the next refinement iteration. The final output is a STAR file with updated particle Euler angles, together with a pair of RELION reconstructed half-maps suitable for downstream postprocessing, visualization and model building. These half-maps can be further subjected to an additional round of Procedure 1 for further map Anisotropy Correction (Fig. 1).

Fig. 1 |. Workflows for cryoEM data processing with spIsoNet.

Two complementary procedures are shown: Procedure 1, Map Anisotropy Correction (orange), and Procedure 2, Particle Misalignment Correction (blue). In Procedure 1, two independent half-maps are prepared and a 3DFSC volume is computed in spIsoNet. spIsoNet is then trained in a self-supervised manner using the half-maps and the 3DFSC volume to finally produce anisotropy-corrected half-maps, which are subsequently subjected to standard map post-processing. In Procedure 2, raw particle stacks and associated metadata file are prepared, and an external reconstruction environment is configured in RELION so that spIsoNet can be called during iterative refinement. Each refinement iteration uses spIsoNet-based Anisotropy Correction to regularize the reference maps, improving particle orientation assignment and finally yielding misalignment-corrected half maps for downstream map post-processing. The optional dashed connection indicates that misalignment-corrected half-maps from Procedure 2 can be used as input for a further round of Anisotropy Correction using Procedure 1. Together, the two procedures mitigate two major consequences of preferred orientation: particle misalignment and map anisotropy.

Table 1 |.

SpIsoNet workflows

Workflow	Required inputs	Main tool	Main outputs	Use case
Map Anisotropy Correction	Two half-maps, a solvent mask and a 3DFSC volume	spisonet.py reconstruct	Anisotropy-corrected half-maps, a trained model, and the loss curve	Maps affected by directional resolution anisotropy
Particle Misalignment Correction	A Particle stack, particle metadata file with CTF information, a reference map, ans a solvent mask	RELION external reconstruction with spIsoNet	Updated particle poses and RELION generated half-maps	Reconstruction artifacts affected by the particle misalignment

Fig. 2 |. Input files required for spIsoNet.

a, Representative cryoEM half-map. b, Orthogonal slice through the cryoEM half-map. c, Representative cryoEM half-map overlaid with the solvent mask. d, Orthogonal slice through the solvent mask. e, Representative 3DFSC volume. f, Orthogonal slice through the 3DFSC volume.

Together, these workflows provide a guided computational strategy for addressing two common consequences of preferred orientation: direction-dependent map anisotropy and particle pose-estimation errors during iterative refinement. We expect that the approach described here will be broadly applicable to the processing of diverse cryoEM datasets.

Prerequisite for using the protocol

This protocol is intended for cryoEM projects in which preferred orientation impedes structure determination. spIsoNet supports multi-GPU acceleration and parallelized processing to improve computational efficiency. Because spIsoNet is command-line based, users should be comfortable working in a Unix/Linux environment and should have basic experience with cryoEM data processing, particularly RELION-based refinement for the Misalignment Correction workflow. Users can also refer to the associated spIsoNet tutorial, provided in Supplementary Information 1 and available through the spIsoNet GitHub repository (https://github.com/IsoNet-cryoET/spIsoNet/tree/main/tutorial), for a more detailed explanation of the algorithms underlying each step, as well as instructions for running the workflows.

Materials

Equipment and setup

Hardware

The minimum computational requirements for running this protocol include a Linux-based workstation or computing cluster equipped with at least one NVIDIA GPU with at least 12 GB of GPU memory, at least 64 CPU cores, 64 GB of random-access memory (RAM), and 250 GB of available disk space for storing raw data and output files. These requirements are comparable to those of conventional cryoEM reconstruction software.

Software

spIsoNet version 1.0, distributed under the MIT license, is required. We recommend installing spIsoNet in an Anaconda or Miniconda environment. Detailed installation instructions are maintained at the spIsoNet GitHub repository (https://github.com/IsoNet-cryoET/spIsoNet) and are also provided in Supplementary Information 1.

RELION version 3.1 or 4.0 is required and is available from http://www2.mrc-lmb.cam.ac.uk/relion or https://github.com/3dem/relion.

IMOD version 4.11 or later is required and is available from http://bio3d.colorado.edu/imod/.

UCSF ChimeraX version 1.6.0 or later is required and is available from https://www.cgl.ucsf.edu/chimerax/download.html.

Datasets

EMD-8731 cryoEM half-maps can be downloaded from https://zenodo.org/records/20783378 or through the EMDB web interface at https://www.ebi.ac.uk/emdb/EMD-8731/

The EMPIAR-10096 subset particle stack can be downloaded from https://zenodo.org/records/20783378. The full particle stack can be downloaded through the EMPIAR web interface at https://www.ebi.ac.uk/empiar/EMPIAR-10096/.

The reference map and associated particle metadata file, in STAR format, used for spIsoNet particle Misalignment Correction of EMPIAR-10096 can be downloaded from https://zenodo.org/records/20783378.

All expected results and tutorial data for this protocol can be downloaded from https://zenodo.org/records/20783378.

Procedure

CRITICAL

This protocol provides step-by-step procedures for applying spIsoNet through two modules: Procedure 1, map Anisotropy Correction using cryoEM half-maps only, and Procedure 2, particle Misalignment Correction by integrating spIsoNet into RELION external reconstruction. Map Anisotropy Correction can be used as a standalone workflow, whereas particle Misalignment Correction, requires RELION external reconstruction. We encourage users to first run Procedure 1 as a quick test to evaluate the improvement of their own cryoEM maps after map Anisotropy correction. However, particle alignment may still be affected by map distortion caused by preferred orientation. Procedure 2 is therefore useful for improving particle alignment. These two procedures can be used either independently or in tandem; for example, users may first perform particle Misalignment Correction and then apply map Anisotropy Correction to address two major ramifications of the preferred-orientation problem.

User should run all commands in a terminal shell within the project directory. Users should modify and optimize key parameters according to their own projects. We provide individual commands with specific parameters, discuss the expected results, and describe potential issues and troubleshooting strategies. Novice users are encouraged to follow the steps sequentially, inspect the output from each step, and compare their results with those described here. A more comprehensive tutorial is provided in Supplementary Information 1 and is also available at https://github.com/IsoNet-cryoET/spIsoNet/tree/main/tutorial.

Procedure 1| Map Anisotropy Correction ● Timing ~40 min

Preparation: arrangement of input files and directories ● Timing ~1 min

• 1Connect to the workstation on which spIsoNet is installed and all required files have been downloaded. We assume that the base directory ./spIsonet_anisotropy/EMDB-8731/ contains the downloaded EMDB-8731 half-maps (emd_8731_half_map_1.mrc and emd_8731_half_map_2.mrc), the solvent mask (emd_8731_msk_1.mrc), and the target map (emd_8731.mrc) with anisotropic resolution that will be improved in this procedure. Most commands in this procedure are run from this base directory. Ensure that the spIsoNet conda environment is activated, as this is required for all spIsoNet commands. Enter the following commands at the terminal:

cd /path/to/base/directory
conda activate spisonet
spisonet.py --help

The final lines provide a list of available spIsoNet commands. To learn more about a specific spIsoNet command, simply enter:

spIsoNet.py [command] --help

PAUSE POINT The two half-maps and the solvent mask have opposite handedness relative to emd_8731.mrc. However, this does not affect spIsoNet training. Users may choose to flip the handedness of the half-maps and solvent mask to match emd_8731.mrc. In this protocol, we instead flip the final post-processed map after spIsoNet map Anisotropy Correction using the ChimeraX command vop flip #vol_num, where vol_num should be replaced with the actual volume number.

? TROUBLESHOOTING

3DFSC computation ● Timing ~2 min

• 2Calculate the 3DFSC volume. The default output file name is FSC3D.mrc, which describes the Fourier shell correlation (FSC) in different directions (Fig. 2e). This step does not require GPU acceleration. Multiple CPU cores can be used for parallelization by specifying --ncpus. The resulting FSC3D.mrc file will be used as the aniso_file in the subsequent spIsoNet reconstruction step. The inputs for 3DFSC calculation are two unfiltered half-maps and a solvent mask. Users can also set the resolution limit for 3DFSC calculation by specifying --limit_res; for the example dataset used here, this is set to --limit_res 3.5. By default, this value is set to the overall resolution of the cryoEM map on the basis of the gold-standard FSC criterion of 0.143. Run the following command:

spisonet.py fsc3d emd_8731_half_map_1.mrc
emd_8731_half_map_2.mrc emd_8731_msk_1.mrc --ncpus 16
--limit_res 3.5

PAUSE POINT Regions with low intensity values in an anisotropic 3DFSC volume (Fig. 2f) indicate under-sampled regions in Fourier space, whereas an isotropic map will generate an isotropic 3DFSC volume. Therefore, we use 3DFSC as a proxy for directional resolution and map anisotropy. Applying an anisotropic 3DFSC volume as a filter to a map in Fourier space approximates a cryoEM map with uneven orientation sampling.

? TROUBLESHOOTING

Training the spIsoNet network ● Timing ~35 min, depending on the hardware configuration, the number of GPUs used, and the number of training epochs

• 3Training the spIsoNet network for map Anisotropy Correction from scratch. A description of all available training parameters can be displayed by running spisonet.py reconstruct --help. The following parameters are critical and may need to be adjusted for different cryoEM map datasets.

Critical Parameters:

--alpha %% the relative weighting between the data-consistency loss and the rotation-equivariance loss. The default value is 1. A larger value places greater weight on the rotation-equivariance loss and generally produces smoother maps.

--beta %% the relative weighting between denoising and missing-information recovery. A larger value places greater weight on denoising.

--limit_res %% the resolution limit for spIsoNet training during map Anisotropy Correction. By default, this value is set to the overall resolution of the cryoEM map as determined by the gold-standard FSC criterion.

--epochs %% the number of epochs used to train the neural network. The default value is 30.

Run the following command:

spisonet.py reconstruct emd_8731_half_map_1.mrc
emd_8731_half_map_2.mrc --aniso_file FSC3D.mrc --mask
emd_8731_msk_1.mrc --limit_res 3.5 --epochs 30 --alpha
1 --beta 0.5 --output_dir isonet_maps --gpuID 0,1,2,3
--acc_batches 2

PAUSE POINT This step creates a folder to store the output files from spIsoNet correction. The anisotropy-corrected maps are saved as correctedXXX.mrc files in the output folder. The trained neural network model, XXX.pt, and the loss curve, lossXXX.png, are also saved in this folder.

? TROUBLESHOOTING

• 4Continue training from a pretrained model or perform direct prediction. If a job is interrupted, or if users want to restart or extend training, training can be resumed from a saved model by specifying --pretrained_model. When starting from a pretrained model, users may also adjust the number of epochs. For example, if the pretrained model was saved after 30 epochs, setting –-epochs 20 continues training for 20 additional epochs, equivalent to 50 total epochs from scratch. Run the following command:

spisonet.py reconstruct emd_8731_half_map_1.mrc
emd_8731_half_map_2.mrc --aniso_file FSC3D.mrc --mask
emd_8731_msk_1.mrc --limit_res 3.5 --epochs 20 --alpha
1 --beta 0.5 --output_dir spisonet_continued_maps --
gpuID 0,1,2,3 --acc_batches 2 --pretrained_model
spisonet_maps/XXX.pt

• 5Users can also set –epochs to 0 together with --pretrained_model to perform prediction only, without further training. Run the following command:

spisonet.py reconstruct emd_8731_half_map_1.mrc
emd_8731_half_map_2.mrc --aniso_file FSC3D.mrc --mask
emd_8731_msk_1.mrc --limit_res 3.5 --epochs 0 –-alpha
1 --beta 0.5 --output_dir
spisonet_directprediction_maps --gpuID 0,1,2,3 --
acc_batches 2 --pretrained_model spisonet_maps/XXX.pt

? TROUBLESHOOTING

Map post-processing ● Timing ~2 min

• 6Perform RELION post-processing. Post-processing of the anisotropy-corrected half-maps is not implemented in spIsoNet. In this protocol, we use relion_postprocess in RELION for B-factor estimation and map sharpening. Run the following command:

relion_postprocess --i
corrected_emd_8731_half_map_1.mrc --i2
corrected_emd_8731_half_map_2.mrc --mask
emd_8731_msk_1.mrc --auto_bfac

Alternatively, users can define their own B-factor and low-pass filtered values by running:

relion_postprocess --i
corrected_emd_8731_half_map_1.mrc --i2
corrected_emd_8731_half_map_2.mrc --mask mask.mrc --
adhoc_bfac <user defined bfactor> --low_pass <user
defined resolution>

The post-processed map has different handedness relative to the deposited EMDB map, emd_8731.mrc. Users can flip the post-processed map using the ChimeraX command vop flip #vol_num, where vol_num should be replaced with the actual volume number.

PAUSE POINT Noise2Noise-based denoising is used by default during the spIsoNet Anistropy Correction process and produces cleaner maps. Denoised spIsoNet output maps are intended for visualization and model interpretation. Users can inspect the loss curve, run spIsoNet correction with different numbers of epochs, and perform map post-processing with the same B-factor. The resulting maps can then be opened in IMOD, allowing users to rapidly load, inspect, and compare map quality across multiple outputs. This comparison can help determine the optimal result. For the example dataset used here, the 30-epoch result is satisfactory, and further training to 50 epochs does not produce substantial differences in map quality (Extended Data Fig. 1).

? TROUBLESHOOTING

Procedure 2| Particle Misalignment Correction ●Timing ~6h

Preparation: arrangement of particle-stack input files and directories ●Timing ~3 min

• 7Misalignment Correction requires the installation of spIsoNet and RELION version 3.1 or 4.0. In this procedure, we assume that the base directory contains the downloaded subset of EMPIAR-10096 available at https://zenodo.org/records/20783378. This subset includes the particle metadata file with CTF information (job025_tutorial.star), the particle stack (job025_tutorial.mrcs), the mask file (mask.mrc), and a low-resolution reference map at 10 Å resolution (HA_reference.mrc), located in ./spIsonet_misalignment/EMPIAR-10096/.

The majority of the commands in this procedure are run within this base directory. Ensure that the spIsoNet conda environment is activated, as it is required for all spIsoNet commands. Enter the following commands at the terminal:

cd /path/to/base/directory
conda activate spisonet
spisonet.py --help

? TROUBLESHOOTING

Set the RELION external reconstruction environment ●Timing ~2 min

• 8Set the RELION_EXTERNAL_RECONSTRUCT_EXECUTABLE environment variable to point to relion_wrapper.py. Enter the following commands at the terminal:

export RELION_EXTERNAL_RECONSTRUCT_EXECUTABLE=“python
   <path to spIsoNet>/spIsoNet/bin/relion_wrapper.py”
export CONDA_ENV=“spisonet”
export PATH=<path to spIsoNet>/spIsoNet/bin:$PATH
export PYTHONPATH=<path to spIsoNet>:$PYTHONPATH

• 9Execute the relion_wrapper.py script during iterative RELION refinement. To call relion_wrapper.py during relion_refine, add the argument --external_reconstruct either to the command line or to the Additional arguments field under the Running tab in RELION GUI (Extended Data Fig. 2). To use spIsoNet Misalignment Correction, the RELION refinement command should include the following arguments:

1. --external_reconstruct
2. --solvent_mask
3. --solvent_correct_fsc
4. --keep_lowres (optional)

PAUSE POINT In cryoEM datasets with severe preferred orientation, standard RELION refinement may fail to achieve reliable particle alignment during iterative refinement. In such cases, --keep_lowres is necessary to preserve low-resolution information from a user-provided reliable reference throughout subsequent RELION iterative alignment. The low-resolution cutoff to be retained is defined by the --ini_high parameter in RELION. The low-resolution reference map can be obtained using various approaches, such as reconstruction from a small particle dataset collected with stage tilt or subtomogram averaging from tomography datasets.

? TROUBLESHOOTING

Anisotropy correction-powered particle Misalignment Correction ●Timing ~6 h, depending on the hardware configuration, the number of GPUs used, and the number of training epochs

• 10Re-estimate particle poses in RELION using spIsoNet. In this protocol, RELION version 4.0 is used. Start a new RELION refinement from the baseline run directory and enable external reconstruction. SpIsoNet trains and applies anisotropy correction internally during each iteration of RELION refinement. This step is expected to require substantially longer runtime than standard RELION refinement. For the EMPIAR-10096 subset used in this protocol, the estimated processing time is approximately 6 h using four NVIDIA A100 GPUs. Additional Linux environment variables can be set as described below to tune the key parameters of spIsoNet Misalignment Correction.

Critical Parameters:

export CUDA_VISIBLE_DEVICES=”0,1,2,3” %% Specifies which GPU(s) to use. By default, spIsoNet uses all available GPUs.

export ISONET_ALPHA=1 %% the relative weighting between the data-consistency loss and the rotation-equivariance loss. The default value is 1. A larger value places greater weight on the rotation-equivariance loss and generally produces smoother maps.

export ISONET_BETA=0.5 %% the relative weighting between denoising and missing-information recovery. A larger value places greater weight on denoising.

export ISONET_START_HEALPIX=3 %% Defines the angular sampling step at which spIsoNet training starts.

export ISONET_START_RESOLUTION=15 %% Defines the refinement resolution at which spIsoNet training starts.

export ISONET_RETRAIN_EACH_ITER=True %% Defines whether the network is retrained from scratch at each RELION refinement iteration. This is typically set to True.

export ISONET_EPOCHS=5 %% Defines the number of epochs used to train the neural network.

export ISONET_KEEP_LOWRES=False %% Defines whether low-resolution information from a correct reference map is retained during particle alignment. This parameter can be overridden by the --keep_lowres parameter in the RELION command line or GUI.

export ISONET_WHITENING=True %% Defines whether whitening is performed before running spIsoNet.

export ISONET_WHITENING_LOW=10 %% Defines the starting resolution for whitening.

export ISONET_FSC_WEIGHTING=True %%Defines whether FSC weighting is performed before running spIsoNet. This is typically set True.

Run the following command:

mpirun -np 5 `which relion_refine_mpì --o
Refine3D/job001/run --auto_refine --
split_random_halves --i job025_tutorial.star --ref
HA_reference.mrc --firstiter_cc --ini_high 10 --
dont_combine_weights_via_disc --preread_images --pool
30--pad 2 --ctf --particle_diameter 170 --
flatten_solvent --zero_mask --solvent_mask mask.mrc --
solvent_correct_fsc --oversampling 1 --healpix_order 2 --
auto_local_healpix_order 5 --offset_range 5 --
offset_step 2 --sym C3 --low_resol_join_halves 40 --
norm --scale --j 4 --gpu ““ --external_reconstruct --
keep_lowres --pipeline_control Refine3D/job001/

PAUSE POINT For this particle dataset, which exhibits severe preferred orientation, we retained low-resolution information to 10Å from the reference by setting --keep_lowres, thereby improving the reliability of particle alignment.

? TROUBLESHOOTING

• 11Misalignment correction generates intermediate anisotropy-corrected half-maps at each iteration of RELION refinement for particle pose estimation, These maps are saved as correctedXXX.mrc files in the refinement output directory. Users can also inspect the RELION-generated half-maps from each iteration (run_itXXX_half1_unfil_backup.mrc), which are reconstructed based on the updated particle Euler angles, to evaluate the effect of misalignment correction (Extended Data Fig. 3). After the refinement is complete, RELION generates the final half-maps (run_half1_class001_unfil.mrc) and (run_half2_class001_unfil.mrc) and the refined particle metadata file (run_data.star).

PAUSE POINT The 3DFSC volume, the trained neural network model (XXX.pt), and a plot showing the training loss (lossXXX.png) are also generated for each iteration.

? TROUBLESHOOTING

Map post-processing

• 12Perform post-processing of the final half-maps generated by RELION. In this procedure, relion_postprocess is used for map sharpening. Enter the following command:

relion_postprocess --i run_half1_class001_unfil.mrc --
i2 run_half2_class001_unfil.mrc --mask mask.mrc --
auto_bfac

Alternatively, users can define their own B-factor and low-pass filtered resolution by running:

relion_postprocess --i run_half1_class001_unfil.mrc --
i2 run_half2_class001_unfil.mrc --mask mask.mrc --
adhoc_bfac <user defined bfactor> --low_pass <user
defined resolution>

PAUSE POINT Because the spIsoNet misalignment-corrected half-maps are independently reconstructed in RELION using the refined particle poses in run_data.star, their independence is preserved, allowing them to be used for gold-standard FSC-based overall resolution estimation.

CRITICAL

Users can optionally use these misalignment-corrected half-maps for further map Anisotropy Correction following Procedure 1, thereby addressing both major consequences of preferred orientation.

Troubleshooting

User questions and feedback were collected through the spIsoNet Google Group: https://groups.google.com/u/0/g/spisonet. Troubleshooting information is provided in Table 2.

Table 2 |.

Troubleshooting table

Step	Problem	Possible reason	Possible solution
All	SpIsoNet command fails to run	SpIsoNet may not be installed correctly, or the spIsoNet environment may not be activated	Reinstall spIsoNet following the recommended installation methods in the spIsoNet installation guide and activate the spIsoNet environment
1	Half-map input files are not recognized	Incorrect half-map filename format	Recommend RELION- or CryoSPARC-style half-map names and explicitly provide the full paths to the required half-maps
2	3DFSC calculation fails	Missing or incorrectly specified input files	Provide half-maps and the solvent mask from RELION 3D refinement or CryoSPARC homogeneous/non-uniform refinement
	3DFSC calculation is slow	Insufficient CPU resources	Increase CPU cores using the -- ncpus option
3	CUDA out-of-memory error occurs during spisoNet training	Batch size or map box size exceeds available GPU memory	Reduce batch size, use additional GPUs, crop or bin input maps, or increase --acc_batches to reduce GPU memory usage
	Corrected map contains excessive noise	The alpha and beta parameters are suboptimal, the limit_res value is too high	Optimize the alpha and beta parameters. For example, increasing beta can increase the denoising effect. Use a more conservative --limit_res value
4	Continued training from a pretrained model fails	Missing required input files or training parameters	Provide the half-maps, the solvent mask, 3DFSC volume, pretrained model, and corresponding training parameters
5	Direct prediction using a pretrained model fails	Missing required input files or prediction parameters	Provide the half-maps, the solvent mask, 3DFSC volume, resolution limit, and pretrained model
6	Anisotropy-corrected half-maps show no significant improveme nt in map features	SpisoNet outputs corrected half-maps, not a final sharpened map	Perform RELION post-processing with the corrected half-maps and solvent mask. If automatic B-factor sharpening is not optimal, manually optimize the B-factor during postprocessing
	Model building and EMDB deposition for the anisotropy-corrected map	Users may be unsure whether the spisoNet anisotropy-corrected map is suitable for atomic model building and EMDB deposition	After map postprocessing, the corrected map can facilitate model building and improve interpretability. For EMDB deposition, report the original half-maps and FSC as the primary data; include the anisotropy-corrected map as an additional map
7	Compatibility with RELION version 5.0 and later versions	RELION 5.0 and later versions require a separate conda environment from spIsoNet	Use RELION 3.1 or 4.0 instead. If using RELION 5.0 or later version, keep RELION and spIsoNet in separate conda environments. Set the external reconstruction executable to the full path of the spIsoNet Python executable followed by relion_wrapper.py. In the RELION GUI, disable Blush regularization in RELION 5.0 when using the spIsoNet external reconstruction workflow
8	RELION does not call the external reconstruction wrapper	The external reconstruction executable is not set correctly, the wrapper path is incorrect, or the wrapper is not executable	Check that RELION_EXTERNAL_RECONST RUCT_EXECUTABLE is set correctly. Ensure that the wrapper script is executable and points to the correct spIsoNet environment. Test the setup with a short RELION run before starting a full refinement
9–11	spIsoNet Particle Misalignment correction fails	relion_wrapper.py cannot locate the required run_itXXX_halfX_unfil.mrc map files during external reconstruction	Verify that the corresponding *_unfil.mrc files exist before the wrapper tries to copy them. And add -- solvent_correct_fsc when required
	relion_wr apper.py reports mask_file = None and then fails	The spIsoNet particle Misalignment correction workflow requires a mask.	Provide a solvent mask in the RELION job and ensure that it is written to the optimization set. Rerun the RELION job with the mask specified before invoking spIsoNet reconstruction
9–11	The choice of low-resolution reference is inappropriate	Using an incorrect Alphafold-predicted model as the reference may introduce reference bias into the spIsoNet-corrected maps	Recommend choosing a low-resolution reference reconstructed directly from tilted cryoEM data or cryoET subtomogram averaging. Use --ref_resolution to limit high-resolution information contributed from the reference and thus minimize reference bias. If using an Alphafold model as the reference, interpret the results cautiously by comparing them with the uncorrected workflow, and confirming that improved features are supported by experimental density
12	Model building and EMDB deposition for the spIsoNet misalignmen-corrected map	Users may be uncertain whether the spIsoNet misalignment-corrected map is suitable for atomic model building or EMDB deposition	The spIsoNet misalignment-corrected half-maps are reconstructed in RELION using the updated particle Euler angles. After map post-processing, the sharpened map can be used for atomic model building and EMDB deposition

Timing

The time required to run this protocol depends on the available hardware. We provide approximate timings in each section of the protocol based on our hardware described in Materials. For users applying this protocol to their own datasets, runtime is mainly determined by the duration of spIsoNet model training and the size of the particle images, as these steps are the most demanding in terms of both time and computational resources. Instructions for how to use spIsoNet are also available at https://github.com/IsoNet-cryoET/spIsoNet/tree/main/tutorial.

Anticipated results

We illustrate the anticipated results of this protocol using two test datasets: (i) deposited half-maps reconstructed from a tilted cryoEM influenza hemagglutinin (HA) trimer dataset for Procedure 1 and (ii) an untilted cryoEM influenza HA trimer particle dataset for Procedure 2.

In the first test dataset, the deposited cryoEM map, EMD-8731, reconstructed using standard RELION Refinement, shows relatively poor map quality due to preferred orientation, as reflected by rod-like α-helical densities with weak or missing side-chain densities (Fig. 3a-b). We downloaded the EMD-8731 half-maps and solvent mask and applied spIsoNet map Anisotropic Correction to the half-maps, followed by map post-processing. Compared with the deposited emd_8731.mrc map, the anisotropy-corrected map exhibits a more clearly resolved α-helical pitch and improved side-chain densities (Fig. 3c-d).

Fig. 3 |. SpIsoNet map Anisotropy Correction of the cryoEM map reconstructed from the tilted influenza HA trimer dataset (EMD-8731).

a, CryoEM map of the HA trimer obtained by standard RELION Refinement. b, Representative α-helical density of the HA trimer fitted with the fitted atomic model. Residue numbers are indicated. c, CryoEM map of the HA trimer after spIsoNet map Anisotropy Correction. d, Representative α-helical density of the HA trimer with the atomic model fitted after spIsoNet correction. Residue numbers are indicated.

In the second test dataset, a representative cryoEM micrograph and RELION 2D classification results show that the particle dataset consists predominantly of top views, with rare side views observed in the 2D class averages (Fig. 4a-b). This HA trimer particle dataset therefore exhibits severe preferred orientation, resulting in a highly distorted cryoEM map with no interpretable secondary-structure features when processed using the standard RELION Refinement pipelines (Fig. 4c). We applied spIsoNet Misalignment Correction during RELION 3D auto-refinement. Particle orientations were first re-estimated through a global orientation search followed by a local orientation search and map post-processing, yielding an approximately 3.5 Å reconstruction (Fig. 4d–f, Extended Data Fig. 3). Typical structural features, including α-helices, β-sheets, and side-chain densities, are clearly visible in this map.

Fig. 4 |. SpIsoNet particle Misalignment Correction of the untilted HA trimer particle images (EMPIAR-10096).

a, Representative cryoEM micrograph of the untilted influenza HA trimer dataset. b, Representative 2D class averages computed from the EMPIAR-10096 particle subset. Oblique and side views are indicated by white stars. c, CryoEM map of the HA trimer obtained by standard RELION Refinement. d, Gold-standard FSC curve of the HA trimer map after Misalignment Correction. e, CryoEM map of the HA trimer after spIsoNet particle Misalignment Correction. f, Model fitting and representative density regions of the HA trimer after spIsoNet particle Misalignment Correction, shown with the fitted atomic model. Residue numbers are indicated.

Overall, by following this protocol, users should be able to reproducibly generate the expected outputs, including:

1. trained spIsoNet models for prediction;

2. a spIsoNet training-loss curve plot reflecting model training performance;

3. spIsoNet anisotropy-corrected half-maps generated using Procedure 1.

4. spIsoNet particle misalignment-corrected half-maps generated through RELION refinement using Procedure 2.

Extended Data

Extended Data Fig. 1 |. Training process for spIsoNet map Anisotropy Correction.

a, Loss curve during spIsoNet training. b, Orthogonal slice through cryoEM maps from epoch 0, 30 and 50. All map were sharpened using the same parameters.

Extended Data Fig. 2 |. Configuration of spIsoNet particle Misalignment Correction in the RELION 4 graphical user interface (GUI).

Add the --external_reconstruct parameter and, optionally, the --keep_lowres parameter.

Extended Data Fig. 3 |. Iterative refinement process for spIsoNet particle Misalignment Correction.

Orthogonal slices through cryoEM half-maps reconstructed in RELION at iterations 0, 3, 6, 10, 16, and 22.

Data availability

All expected results presented in this protocol are available at https://zenodo.org/records/20783378.

Code availability

The spIsoNet is available at the GitHub under MIT license (https://github.com/IsoNet-cryoET/spIsoNet).