PartiNet is a dynamic adaptive neural network for high-performance particle picking in cryo-electron microscopy

cryo-EM, particle picking, AI, machine learning, single particle analysis 18 19 20 21 22 23 24 25 26 27 28 29 30 31 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint

32 Accurate, efficient and autonomous particle picking is a major bottleneck in high-resolution cryo-electron 33 microscopy (cryo-EM). We introduce PartiNet, an Artificial Intelligence (AI)-based particle picker with 34 size-agnostic detection and pre-trained models that eliminate manual parameter specification and dataset-35 specific training, pioneering dynamic neural network inference for single particle cryo-EM pipeline. 36 Unlike static architectures, PartiNet employs a dynamic framework that adjusts network complexity in 37 real-time based on perceived micrograph quality. This adaptive approach accelerates inference up to 7-38 fold compared to existing tools without sacrificing particle selection quality. Training on diverse protein 39 datasets showed that PartiNet improves particle yields, enhances sampling of rare orientations, and is 40 compatible with on-the-fly workflows. Comprehensive evaluation on benchmark datasets and validation 41 on a new dataset of the chromatin remodeler MORC2 demonstrates superior precision and recall, with the 42 ability to detect heterogeneous protein species, leading to more complete structural models and 43 consistently higher-resolution reconstructions. 44

45 Three-dimensional (3D) reconstructions of protein molecules from cryo-EM data enable the generation of 46 atomically detailed chemical models that underpin studies of normal and pathological biological function 47 and guide therapeutic development1. These reconstructions are obtained by advanced image processing 48

that combine large numbers of two-dimensional (2D) images containing individual protein 49 “particles” extracted from electron micrographs, often achieving resolutions better than ~4 Å, where 50 secondary structure becomes unambiguous1,2. Cryo-EM micrographs are inherently noisy, low-dose 51 projections of ensembles of protein molecules, making accurate identification of individual particles a 52 critical early step2. During particle picking, candidate particles are detected, extracted, and subsequently 53 aligned and averaged to generate a 3D reconstruction3,4. This step is particularly challenging due to low 54 signal-to-noise ratios, conformational heterogeneity, contaminating features, and the scale of modern 55 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint cryo-EM datasets. Consequently, particle detection remains a major bottleneck affecting both the 56 accuracy and throughput of cryo-EM structure determination5,6. 57 Classical particle picking relied on heuristic approaches like template matching and feature-based 58 detection. In template matching, users manually select ‘representative’ particles as templates, which are 59 searched across micrographs dataset using the Cross Correlation Function (CCF)7,8, with correlation peaks 60 indicating candidate particles. Feature-based methods instead apply image filters to detect regions of rapid 61 intensity change, most commonly using the Laplacian of Gaussian (LoG) filter9–11. However, LoG 62 performance is sensitive to filter parameters and often struggles with low contrast or overlapping 63 particles. Both CCF- and LoG-based methods are also prone to false positives from ice contaminants 64 (frost), which can produce intense signal peaks due to high electron beam opacity11. These limitations 65 prompted the development of AI-based methods for particle detection. 66 Particle detection can be framed as an AI object-detection problem, in which each particle coordinate 67 (object localisation) and boundary of each particle (object segmentation) must be determined accurately 68 to extract the particle for subsequent 3D reconstruction. A number of AI-based particle picking programs 69 have been developed, including DeepPicker12, APPLE13, Topaz14, crYOLO15, PIXER16, CASSPER17, 70 CryoTransformer18 and CryoSegNet6, with Topaz and crYOLO among the most widely used. Topaz is a 71 Convolution Neural Network (CNN)-based picker that frames the task as a positive-unlabeled (PU) 72 learning problem14. This approach requires prior identification of a small number of positive particle 73 regions19,20 . crYOLO adapts the YOLO9000 object detection framework21; it uses image tiling to divide 74 images into grids, and predicts the presence of particle centres within each grid15. 75 Whilst AI methods have improved particle picking accuracy and throughput, important limitations 76 remain. Firstly, current methods struggle to detect particles in challenging micrographs with complex 77 noise profiles, contamination, ice-thickness gradients or sample heterogeneity5,6. Secondly, all current 78

employ developer-defined static architectures, applying the same network capacity to all images 79 regardless of complexity22,23. This one-size-fits-all approach is computationally inefficient for 80 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint straightforward images and may be inadequate for challenging ones. Finally, with the exception of 81 CryoTransformer and CryoSegNet, many models have been trained on relatively small datasets (<1000 82 micrographs), which can limit generalisability and often necessitates de novo training or fine-tuning of 83 models for new data. 84 Two recent developments in machine learning and cryo-EM provide opportunities to overcome these 85 limitations. The first is the advancement of dynamic architecture frameworks in machine learning, which 86 allow model architectures to adapt during inference and detection, promising improved adaptability to 87 varying image complexity22–24. To date, no dynamic AI particle picking algorithms have been developed 88 for cryo-EM25. The second is the creation of CryoPPP26 - a comprehensive, curated, dataset of particle 89 coordinates from EMPIAR specifically for machine learning applications. CryoPPP has already 90 accelerated the development of new architectures, with CryoSegNet and CryoTransformer matching or 91 outperforming Topaz and crYOLO in terms of final map resolution6,18. 92 Here, we introduce PartiNet, a novel particle picking method based on the DynamicDet neural network 93 architecture24. We trained this framework on the CryoPPP dataset with custom denoiser pre-processing 94 stages and demonstrate the efficacy of the dynamic architectures for protein structure determination. 95 Through comprehensive benchmarking on test datasets and validation on full-scale experimental datasets 96 including the chromatin remodeler MORC2, we show that PartiNet matches or exceeds the quality of 97 particles extracted by current AI pickers while processing micrographs up to 7 times faster. 98 99

100 PartiNet is a dynamic particle picker 101 PartiNet employs a dynamic architecture that adapts network complexity in response to micrograph 102 difficulty, distinguishing it from existing, static particle-picking methods24 (Fig 1a). The system 103 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint comprises two detectors and an adaptive router that analyses extracted features to classify micrographs 104 relatively as ‘easy’ or ‘hard’. Based on this assessment, images are processed either by a single detector 105 for straightforward cases or two cascaded detectors for perceived challenging micrographs, dynamically 106 scaling computational resources to match image requirements in situ (Fig 1b). Our implementation 107 utilises dual YOLOv7 detectors with a shallow convolutional neural network adaptive router20,24. PartiNet 108 differs from the existing YOLO network for particle picking by crYOLO in two key respects: it leverages 109 the newer YOLOv7 architecture rather than a customised YOLO9000 implementation, and it dynamically 110 adjusts network depth based on the input micrograph instead of relying on a single static model. Together, 111 these design choices enable PartiNet to flexibly balance accuracy and computational efficiency across 112 heterogeneous cryo-EM datasets. 113 114 PartiNet enables turnkey particle picking without size specification or training 115 PartiNet enables truly out-of-box particle picking through a simple, single-command interface that 116 requires no parameter specification, model training, or manual intervention. Unlike existing methods that 117 require users to specify expected particle size, train on their own datasets, or manually adjust detection 118 parameters, PartiNet automatically detects particles across all sizes using pre-trained models. The 119 software accepts raw motion-corrected micrographs as input and outputs particle coordinates with 120 confidence scores, integrating seamlessly into automated cryo-EM workflows. This is achieved through a 121 modified Python27 implementation of DynamicDet24, built with PyTorch28, expanded to read motion-122 corrected MRC micrographs and output particle coordinates in STAR file format compatible with popular 123 processing packages (Fig 1). We have provided a custom Wiener filter-based denoiser algorithm, based 124 on the work of CryoSegNet and CryoTransformer6,18, with a novel multiprocessing layer allowing for 125 efficient denoising of large datasets. We have also provided scripts for preparing custom datasets for 126 training PartiNet on user-generated datasets and for finetuning of the current model. PartiNet is designed 127 for command line use and supports high performance computing (HPC) environments, enabling scalable 128 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint multi-GPU/CPU execution via job schedulers such as Slurm. Typical PartiNet workflows comprises three 129 automatable steps: micrograph denoising, particle detection, and coordinate filtering with conversion to 130 STAR format for downstream analysis (Fig 1). The ability to perform size-agnostic particle picking, 131 unlike crYOLO and Topaz, represents a key feature of PartiNet and enables seamless integration into 132 fully automated, unsupervised cryo-EM image processing pipelines. 133 134 PartiNet training on comprehensive dataset enables generalisation without user-specific fine-tuning 135 PartiNet was trained using the CryoPPP dataset, a curated benchmark comprising ~10,000 labelled 136 micrographs from 34 cryo-EM protein datasets (Supplementary Table 1). These datasets span diverse 137 protein sizes, symmetries and imaging conditions, and include gold standard particle coordinates 138 manually curated by the authors of CryoPPP29, enabling PartiNet to generalise to new datasets without 139 requiring user-specific training or fine-tuning - a significant advantage over methods that require 140 retraining for optimal performance. We used seven, randomly selected datasets as test sets for the 141 comparison of crYOLO, Topaz and PartiNet performance. For the remaining datasets, 80% of 142 micrographs from each dataset were allocated randomly for training, and the remaining 20% were used 143 for validation. The training, validation and test sets comprised 6224, 1563 and 1879 micrographs, 144 respectively. Since each of the CryoPPP datasets selected contained only 300 micrographs per set, we 145 also prepared 5 full size datasets for particle picking and reconstruction: MORC2 bound with the 146 H3K9me3 peptide (in-house), rabbit muscle aldolase previously tested with Topaz (EMPIAR-10215)14, 147 TcdA1 originally published with crYOLO (EMPIAR-10089)15, and two heterogeneous samples: Ankyrin-148 1 complex (EMPIAR-11043)30 and MlaCD complex (EMPIAR-12531)31. This design enabled systematic 149 assessment of performance across diverse sample sizes and levels of heterogeneity. 150 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint PartiNet preprocesses micrographs for enhanced picking 151 An obvious first step in particle picking from noisy cryo-EM data is some form of denoising. During 152 PartiNet testing, we observed qualitative improvements for precision and recall on denoised micrograph 153 datasets. To investigate this, we compared PartiNet's performance across four preprocessing conditions: 154 raw motion-corrected micrographs, and motion-corrected micrographs processed with three different 155 denoising methods: two popular deep learning-based denoisers Janni32 and Topaz33 (both using 156 theNoise2Noise framework34), and a heuristic Wiener filter -based algorithm introduced in CryoSegNet6. 157 Training was conducted for 100 epochs using CryoPPP datasets across all four conditions. We evaluated 158 the pre-trained models provided by the Topaz and JANNI developers to assess their out-of-the-box 159 performance, rather than training custom denoiser models for each optical condition in our dataset. We 160 acknowledge that users may see improved performance with Topaz and JANNI denoisers using trained 161 sub-models33. 162 For each training condition, we calculated and plotted the Mean Average Precision @ 50% confidence, 163 Precision, and Recall (Supplementary Fig 1a). We observed that PartiNet-CryoSegNet denoise scored the 164 highest across these metrics, followed closely by raw motion-corrected micrographs and then Topaz-165 denoised, with JANNI-denoised data lagging substantially across all metrics. 166 We suspected that JANNI and Topaz had sub-par performance in our testing due to the shared underlying 167 Noise2Noise algorithm adversely affecting the quality of the particles in the micrographs. We plotted the 168 particle coordinates generated by PartiNet for all 4 regimes on the same micrograph selected from the 169 Influenza haemagglutinin trimer dataset (EMPIAR-10093; Supplementary Fig 1b-d). With no denoising 170 applied, PartiNet was able to identify some particles but struggled with the complex noise of the image, as 171 well as differentiating individual particles near contamination (Supplementary Fig 1b). We observed that 172 both Topaz and JANNI suppressed high frequency information across this micrograph and seemed to be 173 especially susceptible to this behaviour when electron-dense contamination was present in the 174 micrograph. Particles close to contamination were “flattened”, preventing effective delineation of closely 175 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint packed particles, both during denoising and subsequent picking with PartiNet (Supplementary Fig 1c-d). 176 Conversely, the PartiNet implementation of CryoSegNet denoise retained discrete particles present in the 177 micrographs, even when the particles were closely packed to the boundary of signal-dense contamination 178 (Supplementary Fig 1e). 179 Given the efficacy of the CryoSegNet denoiser in conjunction with PartiNet, we integrated it directly into 180 our software. The original denoiser was a single-threaded process, significantly bottlenecking denoising 181 of large datasets. We developed a multiprocessing pipeline for this denoiser using the concurrent.futures 182 module in Python, allowing for asynchronous execution of batch denoising for micrographs. Our 183 implementation is especially suited for efficient use of HPC resources using job schedulers. Denoising 184 1000 micrographs on a single node with 32 CPU cores had a CPU efficiency of 94.47% for 01:41:20 185 (hh:mm:ss) walltime, thus demonstrating that denoising can be performed efficiently at scale without 186 introducing a preprocessing bottleneck. 187 During training, we augmented each micrograph before passing it through PartiNet. Augmentations 188 comprised a random combination of image transformations, each of which had a specified probability to 189 be applied to every micrograph during each epoch (Supplementary Table 2, Supplementary Fig 2). Across 190 training epochs, micrographs were randomly augmented for each pass. This served to prevent model 191 overfitting and increased the feature space of the training data for PartiNet to learn 35. During inference, 192 PartiNet employs test-time augmentation (TTA) to improve detection robustness and accuracy 20. Each 193 micrograph is processed at multiple scales (100%, 83%, 67%) and with horizontal flipping to increase the 194 feature space available for picking. The predictions from all augmented versions are then transformed 195 back to the original coordinate system and combined. This allows PartiNet to aggregate features identified 196 at different resolutions and orientations, increasing detection confidence through multiple perspectives of 197 the same micrograph. 198 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint PartiNet implements a refined detector configuration 199 PartiNet was designed with YOLOv7 networks for each detector. We prepared 6 different configurations 200 of YOLOv720, with differing numbers of model parameters, input sizes, reported inference speed in image 201 frames per second, and reported performance on the benchmarking object detection dataset COCO in 202 mAP@50% (Supplementary Table 3)36 . We trained each of these configurations for 100 epochs on 203 CryoPPP training data denoised with PartiNet-CryoSegNet. We measured the mAP@50%, Precision and 204 Recall, and observed that for all metrics, YOLOv7-W6 had the highest scores after 100 epochs of training 205 (Supplementary Fig 3). YOLOv7, YOLOv7-X and YOLOv7-E6E achieved similar performance metrics, 206 ranking as the second-best performing group, withYOLOv7-E6 scoring the lowest overall. In our hands, 207 YOLOv7-D6 failed to converge, and would crash repeatedly during testing, and so was discontinued. 208 Based on these results, we opted to use YOLOv7-W6 for all subsequent testing. 209 210 Benchmarking PartiNet against established AI pickers shows improved performance 211 To compare the performance of PartiNet against Topaz, and crYOLO, we picked particles and 212 reconstructed density maps on the 7 test datasets from cryoPPP (Supplementary Table 4). We evaluated 213 the performance of PartiNet against gold-standard networks crYOLO and Topaz in terms of number of 214 particles picked and the final resolution of the 3D maps. We mirrored a test workflow previously outlined 215 in CryoSegNet and CryoTransformer to fairly compare the networks for all datasets6,18 (Supplementary 216 Fig 4). We report the number of particles picked by each of Topaz, crYOLO and PartiNet, the number of 217 particles remaining after "Select 2D" for reconstruction, and the final global resolution of the 3D map 218 reconstruction with these selected particles (Fig 2). 219 For six of the seven test sets, PartiNet identified more particles prior to "Select 2D" filtering than the 220 other AI methods, with crYOLO and Topaz alternating as the second-best performer. The exception was 221 EMPIAR-10017, where crYOLO identified 3,000 more particles than PartiNet, with Topaz picking less 222 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint particles (16,000) than both PartiNet and crYOLO. PartiNet excelled with smaller, lower molecular 223 weight proteins (EMPIAR-10532 and EMPIAR-11056), where it substantially outperformed both 224 crYOLO and Topaz in particle counts. This likely reflects PartiNet's more aggressive identification 225 strategy when dealing with smaller proteins that are inherently difficult to detect. A similar pattern was 226 observed for EMPIAR-10093, despite this protein's high molecular weight of 779 kDa. We attribute this 227 to the challenging non-globular shape of NOMPC ion channel proteins, which complicates particle 228 identification14. After “Select 2D” filtering, PartiNet consistently retained the most particles across all test 229 sets except EMPIAR-10017, followed by crYOLO and then Topaz. Most importantly, PartiNet achieved 230 the best final resolution for all datasets tested (Supplementary Fig 5). For EMPIAR-10017, this superior 231 resolution was achieved despite PartiNet having fewer retained particles than crYOLO, suggesting that 232 PartiNet identified higher-quality particles with better sampling of protein views. In fact, for all CryoPPP 233 test sets, PartiNet demonstrated stronger sampling of protein views, with alignments of particles covering 234 a broader range of Euler sphere when compared to crYOLO and Topaz (Supplementary Fig 6). 235 We extended our comparison of PartiNet, crYOLO and Topaz with two further datasets: TcdA1 236 (EMPIAR-10089)15 and rabbit muscle aldolase (EMPIAR-10215)14, published with crYOLO and Topaz 237

papers14,15, respectively. We first evaluated the computational efficiency of PartiNet and crYOLO 238 by measuring picking speed in micrographs per second using per-micrograph inference timing (Fig 3a). 239 Topaz was excluded from this speed comparison as it supports only single GPU processing, while both 240 crYOLO and PartiNet support parallel processing with up to 4 GPUs under identical hardware conditions. 241 PartiNet demonstrated substantially faster processing speeds than crYOLO across both datasets, with 242 performance improvements of up to seven-fold on the rabbit muscle aldolase dataset (EMPIAR-10215) 243 and a peak inference speed of 8 micrographs/second for TcdA1. 244 Next, we applied a standardised workflow for processing TcdA1 to compare PartiNet, Topaz and 245 crYOLO objectively (Supplementary Fig 7). We performed 2D classification on PartiNet-selected 246 particles to assess particle quality through visual inspection of 2D class averages (Fig 3b). The resulting 247 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint averages showed coherent, well-centred representations of TcdA1 spanning multiple orientations, with 248 clearly resolved secondary structure features. Subsequent 3D reconstruction using PartiNet particles 249 yielded a high-quality density map with well-defined secondary structure and no observable anisotropy 250 that would result from particle orientation bias (Fig 3c). For comparative analysis, we reconstructed maps 251 of TcdA1 with crYOLO and Topaz coordinates and plotted Fourier Shell Correlation (FSC) curves for 252 each density map (Fig 3d). At an FSC cutoff of 0.143, PartiNet had the highest global resolution of 3.0 Å, 253 followed by Topaz and crYOLO with 3.2 Å. These resolutions represent improvements over the 3.4 Å 254 structure published originally for EMPIAR-10089, likely reflecting differences in reconstruction 255 algorithms between SPHIRE and CryoSPARC. Finally, we compared PartiNet’s performance in terms of 256 particle counts before and after cleaning by 2D classification (Fig 3e). PartiNet identified more particles 257 (15,551) than Topaz or crYOLO (13,380 and 8,282, respectively). After filtering, PartiNet again retained 258 the most particles (10,016) compared to crYOLO (9,292) and Topaz (8,258), revealing that PartiNet 259 coordinates were well ranked. Interestingly, crYOLO retained effectively all particles picked, suggesting 260 that the publicly available model for crYOLO may have been trained on this dataset prior to publication. 261 To complete analysis of TcdA1, we performed de novo, automated model building with the reconstructed 262 maps from PartiNet, crYOLO and Topaz particles using ModelAngelo37 (Supplementary Fig 8). All three 263 models demonstrated excellent sequence coverage and high ModelAngelo confidence scores for the 264 majority of residues. Mean confidence scores were almost identical: PartiNet’s model had the highest 265 mean confidence of 99.4%, followed by crYOLO with 99.2% and then Topaz with 98.9%. We also 266 calculated the Root Mean Square Deviation (RMSD) in Å between a single monomer of the 267 ModelAngelo prediction and the crystal structure of TcdA1 (PDB 4O9Y). PartiNet had the lowest RMSD 268 of 0.67, followed by Topaz with 0.71 and then crYOLO with 0.78. All values were < 1.0 Å for most of 269 the monomer sequence, indicating that resolutions of 3.0-3.2 Å are sufficient for accurate model building 270 of TcdA1 in this dataset. We completed our assessment of the models by calculating Q-scores, an 271 independent metric for map-model fitness, defined as the measure of an atom's resolvability within a 272 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint cryo-EM map. PartiNet had the highest Q-score with 0.68, followed by Topaz and crYOLO with 0.65, 273 indicating PartiNet’s reconstruction led to the highest quality map built with ModelAngelo. 274 We repeated this analysis on rabbit muscle aldolase, again with a standardised workflow of picking with 275 PartiNet, Topaz or crYOLO (Supplementary Fig 9). Subsequent 3D reconstruction of rabbit muscle 276 aldolase with PartiNet showed a high-quality map with clear secondary structure, with no observable 277 orientation bias or anisotropy (Fig 3g). At an FSC cutoff of 0.143, PartiNet had the highest global 278 resolution at 2.8 Å, followed by Topaz with 3.0 Å and crYOLO with 3.1 Å (Fig 3h). We report the 279 number of picked particles and those retained after filtering (Fig 3i). crYOLO picked the highest number 280 of particles (1,621,138) compared to PartiNet (364,185) and Topaz (61,822); however, after filtering, 281 PartiNet retained the most particles (212,176) followed by crYOLO (57,639) and Topaz (23,078). The 282 large discrepancy between crYOLO’s picked and retained particles may be due to the size of the protein 283 or the high density of particles in micrographs. 284 We again used ModelAngelo to validate the reconstructed maps of rabbit muscle aldolase (Supplementary 285 Fig 10). Whilst PartiNet and crYOLO’s models had comparable sequence coverage, ModelAngelo was 286 unable to build most of the sequence into the Topaz map. We inspected the map of rabbit muscle aldolase 287 built with Topaz particles and observed strong anisotropy and poor reconstruction of secondary structure, 288 suggesting that even though the map had a global resolution of 3.0 Å, it was not sufficient to build an 289 appropriate model. We reported the confidence of residue predictions, as RMSD, between the 290 ModelAngelo predictions and the crystal structure of rabbit muscle aldolase (PDB 6ALD), and map-291 model fit with Q-scores. PartiNet had the highest mean confidence of residue predictions at 99.7% 292 followed by crYOLO with 99.1% and Topaz with 95.3%. PartiNet had the lowest mean RMSD of 0.45, 293 followed by crYOLO with 0.53 and Topaz with 5.3. PartiNet had the highest mean Q-score of 0.68, 294 followed by crYOLO with 0.66 and Topaz with 0.63. PartiNet consistently outperformed both crYOLO 295 and Topaz across all metrics in terms of models built from particles picked. 296 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint Together, these results demonstrate that PartiNet consistently delivers superior particle selection, faster 297 inference, and higher-quality reconstructions and atomic models across diverse datasets, establishing 298 improved accuracy/efficiency trade-offs relative to existing AI-based particle pickers. 299 Reconstructions from PartiNet particles enable comprehensive mapping of protein sequences 300 After confirming PartiNet’s performance on the 7 test datasets from CryoPPP, rabbit muscle aldolase and 301 TcdA1, we extended our analysis to an unpublished dataset for the chromatin remodeler MORC238 bound 302 to the H3K9me3 peptide (Supplementary Fig 11) and compared performance against Topaz and crYOLO 303 (Fig 4). First, we compared the picking speed of PartiNet against crYOLO for the dataset and measured 304 PartiNet to be 5-fold faster (Fig 4a). Next, we again performed a standardised workflow for processing 305 MORC2 with PartiNet, Topaz and crYOLO picks (Supplementary Fig 12). We plotted 2D averages for 306 MORC2 particles picked with PartiNet, which showed convergence to high quality, coherent 2D classes, 307 showing secondary structure and evidence of strong sampling across multiple views of MORC2 (Fig 4b). 308 Next, we plotted the FSC curves for each map and observed that the 3D map reconstructed with PartiNet 309 picks had a global resolution of 2.3 Å, compared to 2.5 Å and 2.7 Å with crYOLO and Topaz, 310 respectively (Fig 4c). To assess the quality of the 3D map reconstruction from PartiNet, Topaz and 311 crYOLO picks, we took the final reconstruction of each map and estimated the resolution of each voxel 312 (rather than a global resolution) with the "Local Resolution Estimation" function in CryoSPARC (Fig 4d). 313 As expected, the central globular structure of all MORC2 maps were resolved best due to this being 314 comparatively rigid and electron dense. Conversely, the two double coil (CC) domain arms extending 315 from the central body of the MORC2 were not resolved. This was expected, as these domains are highly 316 flexible because they are responsible for the DNA compaction activity of MORC2, and the presence and 317 activation of these domains are evidenced in other works where 3D variability analysis showed this 318 movement38. The PartiNet map’s resolution was isotropic, whereas crYOLO’s and Topaz’s maps 319 displayed local resolution loss on the fringes of the protein, and in the case of Topaz, within the central 320 structure of MORC2 itself. Additionally, when viewed from the bottom, Topaz’s map exhibited 321 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint anisotropy in the form of “smearing” between the flat front faces, indicating poor sampling of the top and 322 bottom views of MORC2 during particle picking (Fig 4d, bottom panel). We extended our analysis to 323 measure the quality of sampling across all views of MORC2. We calculated cFAR and SCF values for 324 each map with the "Orientation Diagnostics" in CryoSPARC, with higher cFAR and SCF scores 325 corresponding to stronger sampling across all views of the proteins39,40. This value is effectively an 326 estimation of "orientation bias" in the map, where certain views of the protein are heavily sampled and 327 rarer views of the protein are undersampled. This occurs due to the tendency of many proteins to partition 328 to the air-water interface of the protein solution directly prior to vitrification, and represents a major 329 bottleneck for data analysis. PartiNet had the highest cFAR score of 0.27, followed by crYOLO with 0.11 330 and the Topaz with 0.03. Interestingly, PartiNet had a SCF score of 0.77 compared to crYOLO’s 0.78, 331 whereas Topaz has a substantially different SCF of 0.65. These discrepancies reveal interesting details 332 about the particles picked by PartiNet and crYOLO. cFAR is calculated by measuring the correlations of 333 half-maps in specific viewing angles in Fourier space39, meaning that both the alignment and quality of 334 signal in that specific alignment is necessary for a good score. Conversely, SCF is calculated by 335 quantifying how the alignments of particles cover the Euler sphere without quantifying signal quality 336 from particles40. Given that PartiNet map had the highest global resolution and cFAR score, it can be 337 concluded that PartiNet particles contributed meaningful particle signal information from alignments, 338 with good coverage over the viewing angles of the protein. crYOLO, on the other hand, had comparable 339 numbers of particles spanning the Euler sphere but may have contained more poor particles or particles 340 that contributed poorly to the assigned alignment. LowTopaz scores for cFAR, SCF and number of 341 particles indicate that these particles were of lesser quality and exhibited strong orientation bias, with poor 342 contribution to signal in many alignments. 343 We completed our analysis of MORC2 by using a reconstructed 3D map from PartiNet to build a 344 structural model of MORC2 using ModelAngelo37. We superimposed this model with a published X-ray 345 crystallography structure41 (PDB-5OF9, Fig 4e). We observed strong congruence between these models, 346 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint especially in the central structure of the protein, with the notable absence of the two flexible coiled-coil 347 (CC1) domains in our model. This is a limitation of cryo-EM, as highly flexible protein domains are 348 difficult to resolve to high resolution with 3D cryo-EM maps without further advanced processing tools42. 349 We plotted the per-residue confidence, RMSD with the crystal structure and Q-score for model-map fit 350 (Supplementary Fig 13a-c). In vitro and in vivo, MORC2 exists as a homodimer38,41. We calculated 351 RMSD for each monomer of each map, and found that our reconstructed map and the published map 352 deviated within 1 Å of each other for the whole protein, except for unstructured tails and double CC 353 domain arms, which are highly flexible (Supplementary Fig 13d).We observed a segment of the MORC2 354 protein that was mapped in our cryo-EM model but was not present in the published crystal model (Fig 355 4f). In the crystal structure, eight amino acids between F510 and Y519 were absent; in contrast, the 356 MORC2 cryo-EM density map allowed us to build all but one residue. Consequently, we accurately 357 determined the structural context of seven amino acids that were unresolved in the published model. 358 Collectively, these analyses show that PartiNet enables faster particle picking, improved orientation 359 sampling, and higher-quality reconstructions on previously unseen datasets, facilitating accurate model 360 building and recovery of structural features that were unresolved in prior studies. 361 362 PartiNet enables size-agnostic, single-run identification of multiple species 363 PartiNet's size-agnostic detection - which requires no manual specification of particle dimensions - 364 enables simultaneous identification of multiple species in heterogeneous datasets. This eliminates the 365 need for multiple picking runs with different size parameters, a common requirement for existing methods 366 when analyzing samples containing particles of varying sizes. We investigated this capability using two 367 published datasets containing multiple molecular species We first evaluated this on EMPIAR-11043, 368 containing the Ankyrin-1 complex in a micelle, and the free ‘Band 3’ protein (Fig 5)30. Processing this 369 dataset proved to be quite involved for the original authors, as identification and processing of the two 370 species required 4 different rounds of particle picking (including manual picking and training a Topaz 371 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint model) in conjunction with iterative 2D and 3D classification before model refinement. Using PartiNet, 372 we were able to demonstrate a high-speed, single picking step which was able to identify both species 373 using our general model (Fig 5a-b), whilst simultaneously avoiding multiple rounds of particle picking 374 and classification (Supplementary Fig 14). To verify whether both species were picked, we plotted the 375 box size of each PartiNet prediction against the confidence associated with each confidence as a 2D 376 histogram. Bivariate and univariate kernel density estimations identified two distinct populations of box 377 sizes in PartiNet coordinates, corresponding approximately to the expected sizes of Ankyrin-1 and Band 3 378 (Fig 5c). From these coordinates, we were able to import, extract and refine these particlesubsets, 379 resulting in 3.1 Å and 3.2 Å maps of the Ankyrin-1 complex and Band 3, respectively (Fig 5d-f). 380 We completed our validation of PartiNet on EMPIAR-12531 containing the MlaCD complex with two 381 species comprising 1:6 and 2:6 stoichiometries of MlaC:MlaD (Fig 6, Supplementary Fig 15)31. PartiNet 382 was able to pick 2 times faster than crYOLO on this dataset (Fig 6a). PartiNet also was able to contribute 383 more particles to final reconstruction of each species than the template picking used in the original 384 publication (Fig 6b). PartiNet picked 1,802,428 particles, of which 211,021 were used for 1:6 385 reconstruction, and 203,014 for 2:6. Conversely, in the original publication, 519,770 particles were picked 386 with template picking and 97,460 particles were used for 1:6 reconstruction and 58,259 for 2:6 387 reconstruction. Using PartiNet picks, we were able to calculate reconstructions with 1:6 species at 3.7 Å 388 and 2:6 species at 3.4 Å compared to 4.4 Å for each species published previously31 (Fig 6c). Smearing of 389 signal was seen in 3D reconstructions of both complexes, suggesting suboptimal “stacking” of the protein 390 complexes in the sample. This necessitated masking for only the complex in final refinements 391 (Supplementary Fig 15). Subsequent reconstructions of both 1:6 and 2:6 species showed well-defined and 392 isotropic secondary structure (Fig 6e-h), with well-resolved interfaces between subunits in the 2:6 393 complex. Particle distribution plots confirmed distributed sampling across most views for both species, 394 with excellent angular coverage (Fig 6f-i). With the reconstruction of the two species, we were able to 395 confirm the observation from the original authors that the binding of the MlaC subunit to the MlaD 396 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint multimer breaks the hexameric symmetry of the internal pore, with the MlaD monomer contracting 397 slightly at the centre. We also observed a density in the central pore of the 2:6 species reconstruction; 398 however, we were unable to determine at this resolution if this was the presence of lipid or an artifact 399 from refinement. 400 Taken together, the results from Ankyrin-1 and MlaC/D complexes show that PartiNet can recover 401 multiple molecular species from heterogeneous datasets using a single picking run, improving both 402 throughput and reconstruction quality in multi-component samples. 403 404

405 PartiNet is a dynamic neural network architecture for particle picking in cryo-EM that addresses critical 406 bottlenecks in protein structure determination. By implementing adaptive inference pathways that adjust 407 computational complexity based on micrograph difficulty, PartiNet achieves substantial improvements in 408 both speed and accuracy compared to existing methods. Our comprehensive validation demonstrates 409 consistent improvements in particle identification, reconstruction quality, and computational efficiency 410 across diverse datasets spanning different protein classes, sizes, and imaging conditions. 411 PartiNet is a dynamic particle picker for cryo-EM 412 PartiNet is the first dynamic architecture applied to cryo-EM particle picking. Unlike static networks that 413 process all micrographs through fixed detection layers 6,12,14–16,18, PartiNet's adaptive router adjusts 414 network depth in situ based on learned image features. This addresses a fundamental limitation: the 415 inability to modulate computational resources according to image difficulty. Critically, the router learns to 416 identify challenging features - complex noise, contamination, variable defocus, support film—without 417 supervision, capturing imaging and sample-related complexity without manual labeling. This modularity 418 enables future detector upgrades without architectural redesign and opens possibilities for adaptive 419 approaches in other cryo-EM steps including 3D classification, refinement, model building. 420 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint PartiNet enables truly automated on-the-fly processing workflows 421 Consistent resolution improvements of up to 1.0 Å suggest PartiNet's selection criteria better align with 422 reconstruction requirements, reflecting diverse CryoPPP training and dynamic resource allocation for 423 high-quality particle detection. Superior ModelAngelo metrics confirm that improvements translate to 424 more reliable atomic models37. 425 Critically, PartiNet's combination of speed (2–7× faster) and size-agnostic picking—without prior 426 parameter specification—enables truly automated workflows. Existing methods require manual size input, 427 and they struggle with heterogeneous samples, necessitating user intervention. PartiNet's YOLO 428 architecture20 simultaneously determines presence, location, and dimensions, handling heterogeneous 429 samples (Ankyrin-1/Band 3, MlaCD) in single passes versus multiple rounds with manual intervention. 430 This transforms cryo-EM practice: real-time particle detection with quality metrics provides immediate 431 assessment of sample quality, grid selection, and defocus optimisation during acquisition rather than days 432 later. Integration with existing on-the-fly platforms - Warp/M43, CryoSPARC Live44, and RELION3 - 433 would enable direct sample assessment during imaging. Microscopists can identify failing samples or 434 optimal parameters without interrupting collection, substantially reducing experimental iteration cycles 435 and microscope time waste, which is particularly valuable for high-throughput facilities processing 436 diverse samples. 437

and considerations for practical deployment 438 Several limitations warrant consideration. First, extremely unusual morphologies (elongated filaments, 439 <50 kDa proteins, extreme aspect ratios) may benefit from fine-tuning, requiring GPU resources (≥8 GB 440 VRAM) and PyTorch familiarity. Second, speed advantages require multi-GPU setups (our benchmarks: 441 4× A100 GPUs); single-GPU users may see only modest improvements. Memory requirements (~64 GB) 442 may limit older hardware. Third, severe micrograph quality issues (motion artifacts, thick ice, damage, 443 contamination) challenge all AI pickers25 - no computational approach substitutes for high-quality sample 444 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint preparation. Fourth, while PartiNet identifies multiple species by size, closely-sized particles with similar 445 contrast require downstream 2D/3D classification. Fifth, PartiNet uses YOLOv7 (2022); newer 446 architectures like YOLOv1045 and Detection Transformer46 may offer improvements, though the modular 447 framework facilitates future upgrades. 448 In conclusion, PartiNet demonstrates that dynamic architectures can simultaneously improve speed, 449 accuracy, and robustness in cryo-EM particle picking. Several lessons emerge from this work: diverse 450 training data is essential for broad applicability, modular architectures allow future improvements, 451 combining AI with domain-specific methods (like appropriate denoising) yields better results than AI 452 alone, and open-source distribution accelerates community adoption. 453 As AI integration into cryo-EM accelerates, approaches balancing automation with flexibility and 454 providing interpretable outputs (confidence scores, difficulty classifications) will be essential. The 455 dynamic architecture paradigm provides a template for future developments that adapt intelligently to 456 biological complexity. 457 Online Methods 458 Dynamic architecture inference and training 459 PartiNet is built on a dynamic, deep learning architecture called DynamicDet24 and trained on EM 460 micrographs to detect and localise protein particles. Dynamic architectures can adjust the network 461 architectures in response to different imaging and sample conditions, making it possible to train models 462 efficiently, and also to detect particles with increased speed and accuracy22,23. 463 Static object detection algorithms (hereafter called detectors) contain a backbone, neck and head19,45. 464 During model inference, the backbone performs the bulk of the feature extraction operations on the input 465 image, the neck pools and aggregates these features, and the head performs the final detection by 466 generating bounding boxes or providing pixel coordinates. This architecture is common across all detector 467 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint classes. DynamicDet differs from other object detection algorithms because, instead of passing images 468 sequentially through the backbone, neck, and head of a single detector, it evaluates the intermediate 469 outputs at each stage to adjust the network’s size for processing the input image (Fig 1). 470 471 Assume for a static Detector 𝐷1 which has backbone 𝐵1, neck and head 𝐻1 such that for a given input 472 image x: 473 𝑦 = 𝐷1(𝑥) = 𝐻1(𝐵1(𝑥)) 474 where 𝑦 is the output (coordinates/bounding boxes/contour) of a static object detection network 𝐷1 on 475 input image 𝑥. 476 Instead of a static network, DynamicDet performs the following operations on input 𝑥: 477 𝐹1 = 𝐵1(𝑥) 478 where 𝐹1 are the multiscale feature maps extracted by the first backbone 𝐵1. These feature maps are fed to 479 an adaptive router 𝑅 which determines a difficulty score 𝜙 of the image: 480 𝜙 = 𝑅(𝐹1), 𝜙 ∈ (0,1) 481 If an image is easy (based on a learned threshold from training, see below) then processing is completed 482 by the first detector: 483 𝑦 = 𝐻1(𝐹1) 484 However, if the image is ‘hard’, then these features 𝐹1 are passed along with the input image 𝑥 to the 485 second detector such that: 486 𝐹2 = 𝐵2(𝑥, 𝐹1) 𝑦 = 𝐷2(𝐹2) 487 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint In this way, DynamicDet can dynamically route images to a complex or simpler network based on 488 perceived difficulty of the input images. Currently, PartiNet uses two identical YOLOv7 detector 489 backbones20; however there is scope to change these detectors, highlighting the modularity of the 490 DynamicDet framework. 491 492 Total training of a model for PartiNet requires two steps: training of the detectors, then training of the 493 adaptive router (Supplementary Fig 16). To train the adaptive router, the model weights of the dual 494 detector are frozen, leaving only the adaptive router to be trained to determine if a micrograph is easy or 495 hard 24. This is accomplished with an adversarial loss comparison. A micrograph is passed through a 496 single detector and the total loss ℒ1 is calculated. Then, the micrograph is passed through both detectors 497 and total loss ℒ2 is calculated. If the difference between the two ℒ2 − ℒ1 is low, there is no advantage in 498 using two detectors for the micrograph, and the adaptive router learns this micrograph as easy. If ℒ2 − ℒ1 499 is high then there is a distinct advantage in using the second detector and the micrograph is labelled as 500 easy. Importantly, this stage is unsupervised: the training dataset does not need to be labelled individually 501 as easy or hard (for example, by a skilled human). Instead, only the micrograph is required, removing bias 502 during training of the adaptive router. The adaptive router is able to encode important optical 503 considerations from the image acquisition into its training regime without the need for explicit labelling. 504 This allows for in situ dynamic routing of images based on the difficulty of input micrographs. 505 Calculation of Precision, Recall and mAP@50% 506 In PartiNet, the Intersection over Union (IoU) is used for assessing the overlap between predicted and 507 ground truth bounding boxes. The IoU was calculated as: 508 IoU = |𝐵pred ∩ 𝐵gt| |𝐵pred ∪ 𝐵gt| 509 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint where 𝐵pred represents the predicted bounding box and 𝐵gt represents the ground truth bounding box. The 510 confidence score associated with each predicted bounding box represents the model’s confidence of 511 prediction in both the protein particle’s presence and the accuracy of its localisation. PartiNet also 512 calculates Precision and Recall during training: 513 Precision = 𝑇𝑃 𝑇𝑃 + 𝐹𝑃 514 Recall = 𝑇𝑃 𝑇𝑃 + 𝐹𝑁 515 Where: 516 • 𝑇𝑃 (True Positives): Correctly detected objects with IoU ≥ 0.5 517 • 𝐹𝑃 (False Positives): Detected objects with no corresponding ground truth 518 • 𝐹𝑁 (False Negatives): Ground truth objects not detected 519 The mean Average Precision at 50% IoU threshold (mAP@50%) is calculated as: 520 mAP@50 = 1 𝑁𝑐 ∑ AP𝑐 𝑁𝑐 𝑐=1 521 where: 522 • 𝑁𝑐: Total number of classes (in the case of PartiNet 𝑁𝑐 = 1) 523 • AP𝑐: Average Precision for each class, computed by integrating the precision-recall curve 524 The precision-recall curve was generated by varying the confidence threshold and calculating precision 525 and recall at each point. The area under this curve represents the Average Precision for a given class45. 526 Bounding box coordinates are provided with an associated confidence in PartiNet to represent a 527 probabilistic assessment combining a) the probability that protein particle is present in the box, and b) the 528 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint correct localisation of the box in the micrograph. Confidence in the correct class assignment is only 529 relevant for multi-class inference strategies, which PartiNet does not use. Confidence can be calculated as 530 Confidence = 𝑃(Object Exists) × Localisation Accuracy 531 Confidences can be plotted as a histogram to give an initial assessment of the performance of PartiNet. A 532 confidence threshold is specified: bounding boxes with a confidence above this threshold are retained and 533 those below are discarded. PartiNet defaults to a threshold of 0.3 which provides a balance between 534 retaining the majority of particles identified whilst rejecting most junk and spurious picking. 535 CryoPPP Image Processing 536 The CryoPPP test datasets were processed in CryoSPARC v4.6.2 (Supplementary Fig 4). We utilised our 537 trained PartiNet model with a confidence threshold of 0.3, crYOLO’s publicly available model with the 538 "PhosaurusNet" architecture 15 at the default confidence threshold of 0.3, and Topaz’s publicly available 539 model with the "ResNet16 (64 units)" integrated directly in CryoSPARC4,14 with default parameters to 540 pick particles. We picked particles on all 7 datasets denoised with our integrated denoiser and imported 541 the coordinates alongside the motion-corrected micrographs. We performed CTF estimation on the 542 micrographs to correct for microscope aberrations, then extracted protein particles with an appropriate 543 box size (1.5x largest particle diameter) and performed 2D classification with 50 classes. 2D averages 544 were selected corresponding to protein particles with CryoSPARC’s interactive "Select 2D" function. 545 These selected particle stacks were used for "ab initio Reconstruction", to coarsely reconstruct an initial 546 3D map at low resolution without a reference. This coarse reconstruction was then refined using 547 CryoSPARC’s "Homogeneous Refinement" with C1 symmetry for each dataset leading to the final map 548 for evaluation. 549 TcdA1 image processing and model building 550 TcdA1 (EMPIAR-10089) was processed in CryoSPARC v4.6.2. PartiNet, Topaz and crYOLO were 551 compared objectively using a standardised workflow (Supplementary Fig 7). 97 movies were imported, 552 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint motion-corrected and CTF estimated. Picking was performed using default parameters of each AI picker: 553 the general trained PartiNet model with a confidence threshold of 0.3, crYOLO’s publicly available 554 model with the "PhosaurusNet" architecture15at the default confidence threshold of 0.3, and Topaz’s 555 publicly available model with the "ResNet16 (64 units)" integrated directly in CryoSPARC 4,14 with 556 default parameters to pick particles. Particles were picked on all 97 micrographs denoised with PartiNet’s 557 integrated denoiser. Particles were extracted with a box size of 384 pixels. Extracted particles were 558 filtered in 3D with multi-class “ab initio Reconstruction”, and then subjected to iterative “Heterogeneous 559 Refinement” and “Homogeneous Refinement” of the best class using C5 symmetry. After filtering 560 particles in 3D, “Non-Uniform Refinement” was applied with Global CTF Refinement (Tilt, Trefoil, 561 Spherical Aberration, Tetrafoil and Anisotropic Magnification), and optimising per-particle defocus and 562 scale. Reconstruction was completed with “Reference Based Motion Correction” and “Homogeneous 563 Refinement” of motion-corrected particles with C5 symmetry. 564 Initial model building of TcdA1 was performed with the reconstructed maps from PartiNet, crYOLO and 565 Topaz picked particles using ModelAngelo v1.0.137, an automated AI model-building pipeline 566 (Supplementary Fig 8). Each map along with the FASTA sequence of TcdA1 (UniProt Q9RN43) was 567 input into ModelAngelo and a model prediction was performed. Per-residue confidence scores for each 568 residue were plotted. Root mean square deviation (RMSD) values for a single monomer against the 569 crystal structure of TcdA1 (PDB 4O9Y) were calculated in ChimeraX v1.10.1 using the MatchMaker 570 utility with default parameters and then plotted. The Q-score of each residue was calculated and plotted 571 with ChimeraX using the Q-score plugin47. 572 Rabbit muscle aldolase image processing and model building 573 Rabbit muscle aldolase (EMPIAR-10215) was processed in CryoSPARC v4.6.2. PartiNet, Topaz and 574 crYOLO were compared objectively using a standardised workflow (Supplementary Fig 9). 1,021 movies 575 were imported, motion-corrected and CTF estimated. Particles were picked with default parameters for 576 PartiNet, crYOLO and Topaz as outlined previously. Particles were picked on all 1,021 micrographs 577 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint denoised with PartiNet’s integrated denoiser. Particles were extracted with a box size of 256 pixels. 578 Following extraction of PartiNet particles, “2D Classification” was performed with 50 classes. Due to 579 preferred orientation, the ‘four wing’ view of rabbit muscle aldolase was under-represented. To overcome 580 this, the classes were rebalanced after “Select 2D” with 10 superclasses and a rebalance factor of 0.8 581 (resulting in ~ 50% of particles being temporarily discarded), and then initial maps were generated with 582 multiclass “ab initio Reconstruction” with C2 symmetry. The initialised maps were inspected and the map 583 which had all four wings present was selected. With this map, “Homogeneous Refinement” with D2 584 symmetry was performed and included the particles that were discarded during class rebalancing. 585 Processing was completed with “Non-Uniform Refinement” with Global CTF Refinement (Tilt, Trefoil, 586 Spherical Aberration, Tetrafoil and Anisotropic Magnification), and optimising per-particle defocus and 587 scale with D2 symmetry applied. 588 Initial model building of rabbit muscle aldolase was performed with ModelAngelo. Each map along with 589 the FASTA sequence of rabbit muscle aldolase (UniProt P00883) was input into ModelAngelo and a 590 model prediction was performed. Per-residue confidence scores for each residue were plotted. The RMSD 591 values for a single monomer against the crystal structure of rabbit muscle aldolase (PDB 6ALD) were 592 calculated in ChimeraX v1.10.1 using the MatchMaker utility with default parameters, and then plotted. 593 The Q-score of each residue was calculated and plotted with ChimeraX using the Q-score plugin47 594 (Supplementary Fig 10). 595 Purification of MORC2 596 MORC21-603 (residues 1-603) was cloned into pFastbac construct with 6xHis at the N-terminus and was 597 purified as described previously38. Briefly, MORC2 construct was transformed into EMBacY cells to 598 generate a bacmid. Bacmid DNA was transfected into Sf9 cells using FuGENE transfection reagent 599 (Promega) and virus was passaged twice in the same cell line before large-scale infection. Sf9 cells were a 600 kind gift from the Glukhova lab at WEHI (negative for mycoplasma, identity not independently 601 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint authenticated by us). For large-scale expression, Sf9 cells at a density of 2–2.5 × 106 cells per mL were 602 infected with 1% (v/v) second passage (P2) virus and incubated at 27 °C for 50–60 h, 220 rpm. 603 The cells were harvested by centrifugation, resuspended in lysis buffer containing 50 mM HEPES, pH 604 8.0, 300 mM NaCl, 1 mM TCEP, 5% glycerol, 10 mM imidazole, 10 U per ml benzonase solution 605 (Sigma), 1× cOmplete EDTA-free protease inhibitors (Roche) and 5 mM benzamidine hydrochloride. 606 Cells were lysed by sonication. Clarified cell lysate was incubated with cOmplete His-tag purification 607 resin (Merck) for 1 h followed by wash with lysis buffer. MORC2 proteins were eluted in elution buffer 608 (50 mM HEPES pH 8.0, 300 mM NaCl, 1 mM TCEP, 5% glycerol and 250 mM imidazole). Further 609 purification was performed via size exclusion chromatography using the Superose 6 Increase 10/300 GL 610 column (Cytiva) in SEC buffer (50 mM HEPES pH 8.0, 300 mM NaCl, 1 mM TCEP). 611 Surface Plasmon Resonance (SPR) 612 SPR binding studies of MORC21-603 to H3K9me3 were performed using a Biacore S200 Instrument 613 (Cytiva). Biotinylated H3K9me3 peptide (Active Motif) was diluted to 5 µg/mL in SPR running buffer 614 (10 mM HEPES pH 7.4, 300 mM NaCl, 3 mM EDTA and 0.05% (v/v) Tween-20) to a final 615 immobilisation level of 200-220 response units (RU) on the Streptavidin chip (Cytiva). A blank 616 activation/deactivation was used for the reference surface. DNA binding studies were performed at 20 °C 617 in SPR running buffer. MORC2 proteins were diluted to 1000 nM stock in SPR running buffer and 618 prepared as an 8-point concentration series (2-fold serial dilution, 7-500 nM). Samples were injected in a 619 multi-cycle run (flow rate 30 µL/min, contact time of 60 s, dissociation 120 s) with regeneration with 0.5 620 M EDTA buffer. Sensorgrams were double referenced, and steady-state binding data were fitted using a 621 1:1 binding model using Biacore S200 Evaluation Software (Cytiva, v1.1). Representative sensorgrams 622 and fitted dissociation constant (KD) values are depicted as mean ± SEM (n=3 independent experiments). 623 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint MORC2 cryo-EM sample preparation and data collection 624 A 0.5 mg/mL MORC2 (residues 1-603) protein solution was incubated with 1 mg/mL biotinylated 625 H3K9me3 peptide (Active Motif, Catalogue number 81047, peptide sequence = ARTKQTAR-Kme3- 626 STGGKAPRKQLA - GGYK (Biotin) - NH2) and 2.5 mM AMP-PNP for 1 h on ice. The MORC2-627 H3K9me3 sample was applied to UltrAuFoil R1.2/1.3 gold grids (Quantifoil). The grids were glow-628 discharged for 120 s before application of 4 µL sample onto the grid. The sample was subsequently 629 blotted for 3.5 s (Blot force -3) and vitrified by plunging into liquid ethane using a Vitrobot Mark IV 630 (ThermoFisher) operated at 4 ºC and 100% humidity. Cryo-EM data were acquired on a Titan Krios 631 transmission electron microscope (ThermoFisher) operated at 300 keV, equipped with a K3 direct 632 electron detector (Gatan, Pleasanton) with a pixel size of 0.82 Å/pixel and electron dose of 60 e/Å2. 633 MORC2 Image Processing and model building 634 MORC2 data was processed in CryoSPARC v4.6.2. PartiNet, Topaz and crYOLO were compared using a 635 standardised workflow (Supplementary Fig 12). 6,148 movies were imported, motion corrected and CTF 636 estimated. Particles were picked with default parameters for PartiNet, crYOLO and Topaz as outlined 637 previously. Particles were picked on all 6,148 micrographs denoised with PartiNet’s integrated denoiser. 638 Particles were extracted with a box size of 256 pixels. Classes were selected from a single round of 2D 639 classification. Maps were initialised for each with “ab-initio Reconstruction”, and “Homogeneous 640 Refinement” with C2 symmetry was performed. A final round of “Non-Uniform Refinement” with 641 Global CTF Refinement (Tilt, Trefoil, Spherical Aberration, Tetrafoil and Anisotropic Magnification), 642 and optimising per-particle defocus and scale with D2 symmetry was applied. 643 Ankyrin-1 and Band 3 Image Processing 644 EMPIAR-11043 was processed in CryoSPARC v4.7.0 (Supplementary Fig 14). 14,926 movies were 645 imported, motion-corrected and CTF estimated. Particles were picked with PartiNet on all 14,926 646 micrographs denoised with the integrated denoiser. 3,614,613 particle coordinates were imported into 647 CryoSPARC, and particles were extracted with a box size of 600 pixels, Fourier downsampled to 150 648 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint pixels. Particle stacks were cleaned by 2 rounds of 2D classification (each with 100 classes), with classes 649 selected showing clear averages of both species. Another 2 rounds of 2D classification with 50 classes 650 were generated, with classes selected showing the Ankryin-1 complex in micelle, resulting in 190,334 651 particles. Ankyrin-1 particles were then re-extracted with a box size of 600 pixels with aligned shifts from 652 2D. The full resolution particles were initialised with single class “ab initio Reconstruction” and then 653 “Homogenous Refinement” in C1. Finally, “Non-Uniform Refinement” was applied with global CTF 654 correction (Tilt and Trefoil) and minimising over per-particle defocus and scale. 655 In a parallel processing pathway, 2 rounds of 2D classification with 50 classes were done and classes 656 showing Band 3 protein were identified, resulting in 177,900 particles. Band 3 particles were then re-657 extracted with a box size of 320 pixels with aligned shifts from 2D. The full resolution particles were 658 initialised with single class “ab initio Reconstruction” and then “Homogenous Refinement” in C2. To 659 complete processing of Band 3, “Non-Uniform Refinement” was applied with global CTF correction 660 (Tilt, Trefoil, Spherical Aberration, Tetrafoil, Anisotropic Magnification) and minimising over per-661 particle defocus and scale with C2 symmetry. 662 MlaCD Image Processing 663 EMPIAR-12531 was processed in CryoSPARC v4.7.0 (Supplementary Fig 14). 9,046 movies were 664 imported, motion-corrected and CTF estimated. Particles were picked with PartiNet on all 9,046 665 micrographs denoised with the integrated denoiser. 1,802,428 particle coordinates were imported into 666 CryoSPARC, and particles were extracted with a box size of 350 pixels and Fourier downsampled to 144 667 pixels. Two rounds of 2D classification with 150 classes were used to filter junk particles, resulting in 668 1,535,431 particles. These particles were initialised with “ab initio Reconstruction” with 4 classes. 669 Theclasses were then heterogeneously refined. The two classes representing the 1:6 and 2:6 were selected 670 for “Non-Uniform Refinement” with C1 and C2 symmetry applied, respectively. The “Heterogeneous 671 Refinement” and “Non-Uniform Refinement” step was repeated 10 times, to effectively filter particles in 672 3D. The particles corresponding to each 3D class of 1:6 and 2:6 species were then extracted 673 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint independently at 350 pixels, with updated 2D and 3D aligned shifts. For the 1:6 species, particles were 674 reconstructed with alignments and then “Non-Uniform Refinement” was applied with C1 symmetry, 675 minimising over particle scale. The particles were then corrected for global CTF aberrations (Tilt, Trefoil, 676 Tetrafoil, Spherical Aberration and Anisotropic Magnification) over 3 iterations. The CTF corrected 677 particles were then used for another “Non-Uniform Refinement” with C1 symmetry. Because some 678 particles were stacked on top of each other in 2D projections, spurious density was observed outside the 679 main refinement of the complex, affecting particle alignment and FSC. A mask was generated with 680 “Volume Tools”, with a lowpass filter = 10 Å, threshold = 0.104, dilation radius = 20 pixels, soft-padding 681 = 25 pixels. This mask was used for a final “Local Refinement”. 682 In a parallel processing pathway, the 2:6 species particles were reconstructed with alignments as above 683 and then “Non-Uniform Refinement” was applied with C2 symmetry, minimising over particle scale. The 684 particles were then corrected for global CTF aberrations (Tilt, Trefoil, Tetrafoil, Spherical Aberration and 685 Anisotropic Magnification) over 2 iterations. The CTF-corrected particles were then used for another 686 “Non-Uniform Refinement” with C2 symmetry and optimising per-particle defocus. Again, the spurious 687 density from suboptimal particle stacking was observed. A mask was generated with “Volume Tools”, 688 with a lowpass filter = 10 Å, threshold = 0.121, dilation radius = 20 pixels, soft-padding = 24 pixels. This 689 mask was used for a final “Local Refinement” with C2 symmetry and optimising per-particle defocus. 690 Map visualisation and data plotting 691 Visualisation of EM maps and atomic coordinates for analysis and figures were done in UCSF ChimeraX 692 v1.1048. Charts for analysis and figures were generated in RStudio v 2025.09.0 with R v4.5.1. All figure 693 layouts and exports were done in Affinity Designer v2.6.0 694 695 696 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint Data availability 697 CryoEM maps have been deposited in the EM Data Bank with the following accession codes: EMD-698 68600 (MORC2H3K9-bound). Atomic coordinates have been deposited in the Protein Data Bank with the 699 accession codes 22QM (MORC2H3K9-bound). The raw micrographs, particle coordinates from PartiNet, 700 Topaz and crYOLO for MORC2H3K9-bound have been deposited in EMPIAR the following accession codes 701 (EMPIAR-13226). All other cryo-EM maps for other datasets are accessible from 702 https://doi.org/10.57967/hf/7618. 703 Code availability 704 Source code for PartiNet is publicly available on GitHub at https://github.com/WEHI-705 ResearchComputing/PartiNet. PartiNet is licensed under the MIT License. The model weights can be 706 found here: https://huggingface.co/MihinP/PartiNet. The documentation can be found here: https://wehi-707 researchcomputing.github.io/PartiNet/. 708

709 We acknowledge use of transmission electron microscopes at the Monash University Ramaciotti Centre 710 for Cryo-Electron Microscopy and Ian Homes Imaging Centre, Bio21. We thank the WEHI Cryo-EM 711 Facility, the WEHI Research Computing Platform and Milton high-performance computing facility, the 712 University of Melbourne Spartan high-performance computing facility and the Monash University 713 MASSIVE high-performance computing facility for providing facilities and support. We thank Nicholas 714 Kirk, Alisa Glukhova and Peter Czabotar for their comments on the manuscript. This work was initially 715 supported by WEHI’s New Medicines and Advance Technology funds to AL and SS. MP is supported by 716 Research Training Program (RTP) Scholarship from Faculty of Engineering and Technology, University 717 of Melbourne and Graeme Clark Institute for Medical Engineering Top-Up scholarship. WT is supported 718 by an NHMRC Investigator Grant (GNT 2026635). JDB received support from an Australian Research 719 Council Future Fellowship (FT220100319) funded by the Australian Government. AL is supported in part 720 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint by funds from the estate of Akos and Marjorie Talon. SS is supported by funds from WEHI, the estate of 721 Akos and Marjorie Talon, The University of Melbourne Attraction and Retention Funds, the NHMRC 722 Investigator grant (GNT 2016827), the Australian Research Council Discovery Project grant 723 (DP250100450), the US Department of Defence Rare Cancer Research Concept Award (HT9425-24-1-724 0922) and the US Department of Defence Lung Cancer Research Program Award (HT94252510699). 725 Author contributions 726 MP, OJ, MA wrote the initial scripts for PartiNet, with MP developing the final version. WT purified 727 MORC2, prepared cryoEM samples and performed SPR experiments. HV collected cryoEM data. EY 728 contributed to program parallelisation, debugging, and packaging of PartiNet. JI prepared training data. 729 MP and SS conceived the project. JI, JDB, AL and SS supervised the project. AL and SS acquired the 730 funding. MP, JDB, AL and SS analysed the data and wrote the manuscript with contributions from all 731 authors. 732 Competing interests 733 The authors declare no competing interests 734 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint Main figures 735 736 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint Figure 1. PartiNet has a dynamic architecture 737 a. PartiNet employs a dynamic architecture with dual detectors and an adaptive router. Micrographs are 738 denoised and processed through Detector 1's backbone to generate feature maps. The adaptive router 739 assigns difficulty scores to these feature maps, directing easy micrographs back through Detector 1's neck 740 and head, while difficult micrographs are concatenated with the original image and processed through 741 Detector 2. A post-processing module removes duplicate picks and converts YOLO coordinates to STAR 742 format. b. The adaptive router differentiates micrographs based on imaging conditions without 743 supervision. (I-II) Two micrographs from EMPIAR-10017 showing PartiNet's classification of the 744 micrograph containing support film (yellow arrow) as more difficult than the one without support film. 745 Fringing artifacts from aggressive motion correction are visible at image edges. Scale bar, 100 nm. (III-746 IV) Two micrographs from EMPIAR-10089 demonstrating PartiNet's identification of the lower defocus 747 micrograph as more difficult due to reduced (phase) contrast. The difficult micrograph is at -1.07 µm 748 defocus versus -2.07 µm for the easy micrograph. Scale bar, 120 nm. 749 750 751 752 753 754 755 756 757 758 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint 759 Figure 2. PartiNet outperforms popular particle pickers on small datasets 760 Seven datasets from CryoPPP were randomly selected to benchmark PartiNet against crYOLO and 761 Topaz. The table summarises key dataset characteristics including EMPIAR ID, molecular weight, 762 micrograph count, defocus range, and final map resolution, with the highest resolution result underlined 763 for each protein. The accompanying heatmap shows the number of protein particles retained after 2D 764 classification ("Select 2D") for each method, with darker blue indicating higher particle numbers used for 765 reconstruction. 766 767 768 769 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint 770 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint Figure 3. PartiNet outperforms other particle pickers on EMPIAR-10089 and EMPIAR-10215 771 a. Bar chart comparing picking speed of PartiNet (black) and crYOLO (red) on TcdA1 (EMPIAR-10089) 772 and rabbit muscle aldolase (EMPIAR-10215), reported as micrographs per second. Approximate X-fold 773 increase in speed with PartiNet is indicated for each dataset. b. Selected 2D class averages of TcdA1 774 particles picked with PartiNet. c. Cryo-EM map reconstruction of TcdA1 from PartiNet particles picked 775 on EMPIAR-10089. d. FSC curves for final reconstructions of TcdA1 from particles picked with 776 crYOLO (teal), Topaz (purple) and PartiNet (yellow) with resolution annotated at FSC = 0.143 cutoff. e. 777 Bar chart comparing numbers of particles picked (pink) and used for final reconstruction (purple) of 778 TcdA1 for PartiNet, crYOLO, and Topaz. f. Selected 2D class averages of rabbit muscle aldolase 779 particles picked with PartiNet. c. Cryo-EM map reconstruction of rabbit muscle aldolase from PartiNet 780 particles picked on EMPIAR-10215. d. FSC curves for final reconstructions of rabbit muscle aldolase 781 from particles picked with crYOLO (teal), Topaz (purple) and PartiNet (yellow) with resolution annotated 782 at FSC = 0.143 cutoff. e. Bar chart comparing numbers of particles picked (dark teal) and used for final 783 reconstruction (light teal) of rabbit muscle aldolase for PartiNet, crYOLO, and Topaz. 784 785 786 787 788 789 790 791 792 793 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint 794 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint Figure 4. PartiNet picks allow high resolution reconstructions for model building 795 a. Bar chart comparing picking speed of PartiNet (black) and crYOLO (red) on MORC2. Approximate 5-796 fold increase in speed with PartiNet is indicated. b. Selected 2D class averages of MORC2 particles 797 picked with PartiNet. c. FSC curves for final reconstructions of MORC2 from particles picked with 798 crYOLO (teal), Topaz (purple) and PartiNet (yellow) with resolution annotated at FSC = 0.143 cutoff d. 799 MORC2 maps were reconstructed using coordinates from PartiNet, Topaz, and crYOLO particle picking 800 algorithms. Local resolution was estimated in CryoSPARC and visualised on the reconstructions in 801 ChimeraX (v.1.10.1), where teal voxels indicate lower resolution and maroon indicates higher resolution. 802 A central cross-section of each map is shown. The global resolution, final particle count, conical FSC area 803 ratio (cFAR), and Sampling Compensation Factor (SCF) are shown for each map. e. The crystal structure 804 of MORC2 (PDB 5OF9) was superimposed with the ModelAngelo-generated model from the PartiNet 805 map after alignment in ChimeraX. The blue boxed region is shown in detail in f. highlighting differences 806 in modeling of previously unresolved residues between the crystal structure and ModelAngelo prediction 807 of MORC2. 808 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint 809 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint Figure 5. PartiNet can identify multiple species in a dataset 810 a. Bar chart comparing picking speed of PartiNet (black) and crYOLO (red) on EMPIAR-11043. 811 Approximate 2-fold increase in speed with PartiNet is annotated b. Simplified workflow for processing 812 EMPIAR-11043 to reconstruct Ankyrin-1 and Band 3 proteins. The published workflow required multiple 813 rounds of particle picking (manual, template-based, and Topaz with trained models) to identify multiple 814 species, whereas PartiNet identifies heterogeneous proteins in a single step. Full workflow details are 815 available in Supplementary Fig 14 c. Bivariate histogram showing the relationship between box size and 816 confidence for PartiNet picks on EMPIAR-11043. Particle counts are hexagonally binned with linear box 817 size and logarithmic confidence scaling. Bin color intensity (lighter blue = higher counts) is scaled 818 logarithmically. Only picks with > 10% confidence and > 300 pixel box size are shown to exclude low-819 quality, noisy detections. Red and yellow contours show bivariate kernel density peaks, with 820 corresponding marginal density plots (right) revealing two distinct box size populations at 620 and 403 821 pixels, corresponding to Ankyrin-1 complex and Band 3 proteins, respectively. d-e. Representative 2D 822 class averages and consensus refinement maps for (d) Ankyrin-1 complex and (e) Free Band 3 proteins 823 identified by PartiNet. Final particle counts for each reconstruction are indicated. f. FSC curves for final 824 reconstructions of Ankyrin-1 (yellow) and Band 3 (pink) proteins with resolution annotated at FSC = 825 0.143 cutoff. 826 827 828 829 830 831 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint 832 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint Figure 6. PartiNet improves the resolution of recently published maps 833 a. Bar chart comparing picking speed of PartiNet (black) and crYOLO (red) on EMPIAR-12531. 834 Approximate 2-fold increase in speed with PartiNet is indicated. b. Bar chart comparing number of 835 particles used for final reconstruction of the MlaCD complex (both 1:6 and 2:6 stoichiometry species) 836 between published workflow (light blue) and with PartiNet picking (dark blue). c. FSC curves for final 837 reconstructions of 1:6 stoichiometry (light blue) and 2:6 stoichiometry (dark blue) species of MlaCD with 838 resolution annotated at FSC = 0.143 cutoff. d. Representative 2D class averages of 1:6 stoichiometry 839 MlaCD from particles picked with PartiNet. e. Cryo-EM map reconstruction of 1:6 stoichiometry MlaCD 840 from PartiNet particles. f. 2D heatmap of particle distributions for 1:6 stoichiometry reconstruction 841 (yellow indicating high particle counts and dark blue low particle counts). g. Representative 2D class 842 averages of 2:6 stoichiometry MlaCD from particles picked with PartiNet. h. Cryo-EM map 843 reconstruction of 2:6 stoichiometry MlaCD from PartiNet particles. i. 2D heatmap of particle distributions 844 for 2:6 stoichiometry reconstruction (yellow indicating high particle counts; dark blue indicating low 845 particle counts). 846 847 848 849 850 851 852 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint

853 1. Baker, T. S., Olson, N. H. & Fuller, S. D. Adding the Third Dimension to Virus Life Cycles: Three-854 Dimensional Reconstruction of Icosahedral Viruses from Cryo-Electron Micrographs. Microbiology 855 and Molecular Biology Reviews 63, 862–922 (1999). 856 2. Thompson, R. F., Iadanza, M. G., Hesketh, E. L., Rawson, S. & Ranson, N. A. Collection, pre-857 processing and on-the-fly analysis of data for high-resolution, single-particle cryo-electron 858 microscopy. Nat Protoc 14, 100–118 (2019). 859 3. Scheres, S. H. W. RELION: Implementation of a Bayesian approach to cryo-EM structure 860 determination. Journal of Structural Biology 180, 519–530 (2012). 861 4. Punjani, A., Rubinstein, J. L., Fleet, D. J. & Brubaker, M. A. cryoSPARC: algorithms for rapid 862 unsupervised cryo-EM structure determination. Nat Methods 14, 290–296 (2017). 863 5. Gyawali, R., Dhakal, A., Wang, L. & Cheng, J. Accurate cryo-EM protein particle picking by 864 integrating the foundational AI image segmentation model and specialized U-Net. 865 2023.10.02.560572 Preprint at https://doi.org/10.1101/2023.10.02.560572 (2024). 866 6. Gyawali, R., Dhakal, A., Wang, L. & Cheng, J. CryoSegNet: accurate cryo-EM protein particle 867 picking by integrating the foundational AI image segmentation model and attention-gated U-Net. 868 Briefings in Bioinformatics 25, bbae282 (2024). 869 7. Roseman, A. M. Particle finding in electron micrographs using a fast local correlation algorithm. 870 Ultramicroscopy 94, 225–236 (2003). 871 8. Roseman, A. M. FindEM—a fast, efficient program for automatic selection of particles from electron 872 micrographs. Journal of Structural Biology 145, 91–99 (2004). 873 9. Woolford, D., Hankamer, B. & Ericksson, G. The Laplacian of Gaussian and arbitrary z-crossings 874 approach applied to automated single particle reconstruction. Journal of Structural Biology 159, 122–875 134 (2007). 876 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint 10. Voss, N. R., Yoshioka, C. K., Radermacher, M., Potter, C. S. & Carragher, B. DoG Picker and 877 TiltPicker: Software tools to facilitate particle selection in single particle electron microscopy. 878 Journal of Structural Biology 166, 205–213 (2009). 879 11. Zivanov, J. et al. New tools for automated high-resolution cryo-EM structure determination in 880 RELION-3. eLife 7, e42166 (2018). 881 12. Wang, F. et al. DeepPicker: A deep learning approach for fully automated particle picking in cryo-882 EM. Journal of Structural Biology 195, 325–336 (2016). 883 13. Heimowitz, A., Andén, J. & Singer, A. APPLE picker: Automatic particle picking, a low-effort cryo-884 EM framework. Journal of Structural Biology 204, 215–227 (2018). 885 14. Bepler, T. et al. Positive-unlabeled convolutional neural networks for particle picking in cryo-886 electron micrographs. Nat Methods 16, 1153–1160 (2019). 887 15. Wagner, T. et al. SPHIRE-crYOLO is a fast and accurate fully automated particle picker for cryo-888 EM. Commun Biol 2, 1–13 (2019). 889 16. Zhang, J. et al. PIXER: an automated particle-selection method based on segmentation using a deep 890 neural network. BMC Bioinformatics 20, 41 (2019). 891 17. George, B. et al. CASSPER is a semantic segmentation-based particle picking algorithm for single-892 particle cryo-electron microscopy. Commun Biol 4, 1–12 (2021). 893 18. Dhakal, A., Gyawali, R., Wang, L. & Cheng, J. CryoTransformer: a transformer model for picking 894 protein particles from cryo-EM micrographs. Bioinformatics 40, btae109 (2024). 895 19. Ronneberger, O., Fischer, P. & Brox, T. U-Net: Convolutional Networks for Biomedical Image 896 Segmentation. Preprint at https://doi.org/10.48550/arXiv.1505.04597 (2015). 897 20. Wang, C.-Y., Bochkovskiy, A. & Liao, H.-Y. M. YOLOv7: Trainable bag-of-freebies sets new state-898 of-the-art for real-time object detectors. Preprint at https://doi.org/10.48550/arXiv.2207.02696 899 (2022). 900 21. Redmon, J. & Farhadi, A. YOLO9000: Better, Faster, Stronger. in 2017 IEEE Conference on 901 Computer Vision and Pattern Recognition (CVPR) 6517–6525 (2017). doi:10.1109/CVPR.2017.690. 902 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint 22. Lin, T.-Y. et al. Feature Pyramid Networks for Object Detection. in 2017 IEEE Conference on 903 Computer Vision and Pattern Recognition (CVPR) 936–944 (2017). doi:10.1109/CVPR.2017.106. 904 23. Liang, T. et al. CBNet: A Composite Backbone Network Architecture for Object Detection. IEEE 905 Transactions on Image Processing 31, 6893–6906 (2022). 906 24. Lin, Z., Wang, Y., Zhang, J. & Chu, X. DynamicDet: A Unified Dynamic Architecture for Object 907 Detection. Preprint at https://doi.org/10.48550/arXiv.2304.05552 (2023). 908 25. Dhakal, A., Gyawali, R., Wang, L. & Cheng, J. Artificial Intelligence in Cryo-EM Protein Particle 909 Picking: The Hope, Hype, and Hurdles. Preprint at https://doi.org/10.20944/preprints202408.1936.v1 910 (2024). 911 26. Dhakal, A., Gyawali, R., Wang, L. & Cheng, J. A large expert-curated cryo-EM image dataset for 912 machine learning protein particle picking. Sci Data 10, 392 (2023). 913 27. van Rossum, G. Python tutorial. (1995). 914 28. Ansel, J. et al. PyTorch 2: Faster Machine Learning Through Dynamic Python Bytecode 915 Transformation and Graph Compilation. in Proceedings of the 29th ACM International Conference 916 on Architectural Support for Programming Languages and Operating Systems, Volume 2 vol. 2 929–917 947 (Association for Computing Machinery, New York, NY, USA, 2024). 918 29. Ashwin Dhakal, Rajan Gyawali, Liguo Wang, & Jianlin Cheng. CryoPPP: A Large Expert-Labelled 919 Cryo-EM Image Dataset for Machine Learning Protein Particle Picking. 920 https://doi.org/10.1101/2023.02.21.529443 (2023) doi:10.1101/2023.02.21.529443. 921 30. Vallese, F. et al. Architecture of the human erythrocyte ankyrin-1 complex. Nat Struct Mol Biol 29, 922 706–718 (2022). 923 31. Wotherspoon, P. et al. Structure of the MlaC-MlaD complex reveals molecular basis of periplasmic 924 phospholipid transport. Nat Commun 15, 6394 (2024). 925 32. Wagner, T. & Raunser, S. The evolution of SPHIRE-crYOLO particle picking and its application in 926 automated cryo-EM processing workflows. Communications Biology 3, 61 (2020). 927 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint 33. Bepler, T., Kelley, K., Noble, A. J. & Berger, B. Topaz-Denoise: general deep denoising models for 928 cryoEM and cryoET. Nat Commun 11, 5208 (2020). 929 34. Lehtinen, J. et al. Noise2Noise: Learning Image Restoration without Clean Data. Preprint at 930 https://doi.org/10.48550/arXiv.1803.04189 (2018). 931 35. Shorten, C. & Khoshgoftaar, T. M. A survey on Image Data Augmentation for Deep Learning. 932 Journal of Big Data 6, 60 (2019). 933 36. Lin, T.-Y. et al. Microsoft COCO: Common Objects in Context. Preprint at 934 https://doi.org/10.48550/arXiv.1405.0312 (2015). 935 37. Jamali, K. et al. Automated model building and protein identification in cryo-EM maps. Nature 628, 936 450–457 (2024). 937 38. Tan, W. et al. MORC2 is a phosphorylation-dependent DNA compaction machine. Nat Commun 16, 938 5606 (2025). 939 39. Tan, Y. Z. et al. Addressing preferred specimen orientation in single-particle cryo-EM through 940 tilting. Nature Methods 14, 793–796 (2017). 941 40. Baldwin, P. R. & Lyumkis, D. Non-uniformity of projection distributions attenuates resolution in 942 Cryo-EM. Progress in Biophysics and Molecular Biology 150, 160–183 (2020). 943 41. Douse, C. H. et al. Neuropathic MORC2 mutations perturb GHKL ATPase dimerization dynamics 944 and epigenetic silencing by multiple structural mechanisms. Nature Communications 9, 651 (2018). 945 42. Tan, W. et al. MORC2 phosphorylation fine tunes its DNA compaction activity. 2024.06.27.600912 946 Preprint at https://doi.org/10.1101/2024.06.27.600912 (2024). 947 43. Tegunov, D. & Cramer, P. Real-time cryo-electron microscopy data preprocessing with Warp. Nat 948

16, 1146–1152 (2019). 949 44. Punjani, A. Algorithmic advances in single particle cryo-EM data processing using CryoSPARC. 950 Microscopy and Microanalysis 26, 2322–2323 (2020). 951 45. Wang, A. et al. YOLOv10: Real-Time End-to-End Object Detection. arXiv.org 952 https://arxiv.org/abs/2405.14458v1 (2024). 953 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint 46. Carion, N. et al. End-to-End Object Detection with Transformers. Preprint at 954 https://doi.org/10.48550/arXiv.2005.12872 (2020). 955 47. Pintilie, G. et al. Measurement of atom resolvability in cryo-EM maps with Q-scores. Nat Methods 956 17, 328–334 (2020). 957 48. Goddard, T. D. et al. UCSF ChimeraX: Meeting modern challenges in visualization and analysis. 958 Protein Sci 27, 14–25 (2018). 959 960 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint