{"paper_id":"28db014c-c8e8-489a-a3d3-469d2db8d941","body_text":"PartiNet is a dynamic adaptive neural network for high-performance particle picking in cryo-electron 1 \nmicroscopy 2 \n 3 \nMihin Perera1,2, Winnie Tan1,3,4,7, Edward Yang1,7, Ojasvi Jain1,8, Mansi Aggarwal1,8, Hariprasad 4 \nVenugopal5, Julie Iskander1, Joseph D. Berry2,9, Andrew Leis1,3,9, Shabih Shakeel1,4,6,9* 5 \n 6 \n1 WEHI, 1G Royal Parade, Parkville, VIC, 3052, Australia 7 \n2 Department of Chemical Engineering, The University of Melbourne, Parkville, VIC, 3010, Australia 8 \n3 Department of Medical Biology, The University of Melbourne, Parkville, VIC, 3052, Australia 9 \n4 ARC Centre for Cryo-electron Microscopy of Membrane Proteins, Bio21 Molecular Science and 10 \nBiotechnology Institute, University of Melbourne, Parkville, VIC, 3052, Australia 11 \n5 Monash University, Clayton, VIC 3168, Australia 12 \n6 Department of Biochemistry and Pharmacology, The University of Melbourne, Parkville, VIC, 3010, 13 \nAustralia 14 \n7,8 Equal contributions 15 \n9 Senior authors 16 \n*Correspondence to shakeel.s@wehi.edu.au 17 \nKeywords: cryo-EM, particle picking, AI, machine learning, single particle analysis 18 \n 19 \n 20 \n 21 \n 22 \n 23 \n 24 \n 25 \n 26 \n 27 \n 28 \n 29 \n 30 \n 31 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\nAbstract 32 \nAccurate, efficient and autonomous particle picking is a major bottleneck in high-resolution cryo-electron 33 \nmicroscopy (cryo-EM). We introduce PartiNet, an Artificial Intelligence (AI)-based particle picker with 34 \nsize-agnostic detection and pre-trained models that eliminate manual parameter specification and dataset-35 \nspecific training, pioneering dynamic neural network inference for single particle cryo-EM pipeline. 36 \nUnlike static architectures, PartiNet employs a dynamic framework that adjusts network complexity in 37 \nreal-time based on perceived micrograph quality. This adaptive approach accelerates inference up to 7-38 \nfold compared to existing tools without sacrificing particle selection quality. Training on diverse protein 39 \ndatasets showed that PartiNet improves particle yields, enhances sampling of rare orientations, and is 40 \ncompatible with on-the-fly workflows. Comprehensive evaluation on benchmark datasets and validation 41 \non a new dataset of the chromatin remodeler MORC2 demonstrates superior precision and recall, with the 42 \nability to detect heterogeneous protein species, leading to more complete structural models and 43 \nconsistently higher-resolution reconstructions.   44 \nIntroduction 45 \nThree-dimensional (3D) reconstructions of protein molecules from cryo-EM data enable the generation of 46 \natomically detailed chemical models that underpin studies of normal and pathological biological function 47 \nand guide therapeutic development1. These reconstructions are obtained by advanced image processing 48 \nmethods that combine large numbers of two-dimensional (2D) images containing individual protein 49 \n“particles” extracted from electron micrographs, often achieving resolutions better than ~4 Å, where 50 \nsecondary structure becomes unambiguous1,2. Cryo-EM micrographs are inherently noisy, low-dose 51 \nprojections of ensembles of protein molecules, making accurate identification of individual particles a 52 \ncritical early step2. During particle picking, candidate particles are detected, extracted, and subsequently 53 \naligned and averaged to generate a 3D reconstruction3,4. This step is particularly challenging due to low 54 \nsignal-to-noise ratios, conformational heterogeneity, contaminating features, and the scale of modern 55 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\ncryo-EM datasets. Consequently, particle detection remains a major bottleneck affecting both the 56 \naccuracy and throughput of cryo-EM structure determination5,6.  57 \nClassical particle picking relied on heuristic approaches like template matching and feature-based 58 \ndetection. In template matching, users manually select ‘representative’ particles as templates, which are 59 \nsearched across micrographs dataset using the Cross Correlation Function (CCF)7,8, with correlation peaks 60 \nindicating candidate particles. Feature-based methods instead apply image filters to detect regions of rapid 61 \nintensity change, most commonly using the Laplacian of Gaussian (LoG) filter9–11. However, LoG 62 \nperformance is sensitive to filter parameters and often struggles with low contrast or overlapping 63 \nparticles. Both CCF- and LoG-based methods are also prone to false positives from ice contaminants 64 \n(frost), which can produce intense signal peaks due to high electron beam opacity11. These limitations 65 \nprompted the development of AI-based methods for particle detection. 66 \nParticle detection can be framed as an AI object-detection problem, in which each particle coordinate 67 \n(object localisation) and boundary of each particle (object segmentation) must be determined accurately 68 \nto extract the particle for subsequent 3D reconstruction. A number of AI-based particle picking programs 69 \nhave been developed, including DeepPicker12, APPLE13, Topaz14, crYOLO15, PIXER16, CASSPER17, 70 \nCryoTransformer18 and CryoSegNet6, with Topaz and crYOLO among the most widely used. Topaz is a 71 \nConvolution Neural Network (CNN)-based picker that frames the task as a positive-unlabeled (PU) 72 \nlearning problem14. This approach requires prior identification of a small number of positive particle 73 \nregions19,20 . crYOLO adapts the YOLO9000 object detection framework21; it uses image tiling to divide 74 \nimages into grids, and predicts the presence of particle centres within each grid15. 75 \nWhilst AI methods have improved particle picking accuracy and throughput, important limitations 76 \nremain. Firstly, current methods struggle to detect particles in challenging micrographs with complex 77 \nnoise profiles, contamination, ice-thickness gradients or sample heterogeneity5,6. Secondly, all current 78 \nmethods employ developer-defined static architectures, applying the same network capacity to all images 79 \nregardless of complexity22,23. This one-size-fits-all approach is computationally inefficient for 80 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\nstraightforward images and may be inadequate for challenging ones. Finally, with the exception of 81 \nCryoTransformer and CryoSegNet, many models have been trained on relatively small datasets (<1000 82 \nmicrographs), which can limit generalisability and often necessitates de novo training or fine-tuning of 83 \nmodels for new data. 84 \nTwo recent developments in machine learning and cryo-EM provide opportunities to overcome these 85 \nlimitations. The first is the advancement of dynamic architecture frameworks in machine learning, which 86 \nallow model architectures to adapt during inference and detection, promising improved adaptability to 87 \nvarying image complexity22–24. To date, no dynamic AI particle picking algorithms have been developed 88 \nfor cryo-EM25. The second is the creation of CryoPPP26 - a comprehensive, curated, dataset of particle 89 \ncoordinates from EMPIAR specifically for machine learning applications. CryoPPP has already 90 \naccelerated the development of new architectures, with CryoSegNet and CryoTransformer matching or 91 \noutperforming Topaz and crYOLO in terms of final map resolution6,18. 92 \nHere, we introduce PartiNet, a novel particle picking method based on the DynamicDet neural network 93 \narchitecture24. We trained this framework on the CryoPPP dataset with custom denoiser pre-processing 94 \nstages and demonstrate the efficacy of the dynamic architectures for protein structure determination. 95 \nThrough comprehensive benchmarking on test datasets and validation on full-scale experimental datasets 96 \nincluding the chromatin remodeler MORC2, we show that PartiNet matches or exceeds the quality of 97 \nparticles extracted by current AI pickers while processing micrographs up to 7 times faster. 98 \n 99 \nResults 100 \nPartiNet is a dynamic particle picker 101 \nPartiNet employs a dynamic architecture that adapts network complexity in response to micrograph 102 \ndifficulty, distinguishing it from existing, static particle-picking methods24 (Fig 1a). The system 103 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\ncomprises two detectors and an adaptive router that analyses extracted features to classify micrographs 104 \nrelatively as ‘easy’ or ‘hard’. Based on this assessment, images are processed either by a single detector 105 \nfor straightforward cases or two cascaded detectors for perceived challenging micrographs, dynamically 106 \nscaling computational resources to match image requirements in situ (Fig 1b). Our implementation 107 \nutilises dual YOLOv7 detectors with a shallow convolutional neural network adaptive router20,24. PartiNet 108 \ndiffers from the existing YOLO network for particle picking by crYOLO in two key respects: it leverages 109 \nthe newer YOLOv7 architecture rather than a customised YOLO9000 implementation, and it dynamically 110 \nadjusts network depth based on the input micrograph instead of relying on a single static model. Together, 111 \nthese design choices enable PartiNet to flexibly balance accuracy and computational efficiency across 112 \nheterogeneous cryo-EM datasets. 113 \n 114 \nPartiNet enables turnkey particle picking without size specification or training 115 \nPartiNet enables truly out-of-box particle picking through a simple, single-command interface that 116 \nrequires no parameter specification, model training, or manual intervention. Unlike existing methods that 117 \nrequire users to specify expected particle size, train on their own datasets, or manually adjust detection 118 \nparameters, PartiNet automatically detects particles across all sizes using pre-trained models. The 119 \nsoftware accepts raw motion-corrected micrographs as input and outputs particle coordinates with 120 \nconfidence scores, integrating seamlessly into automated cryo-EM workflows. This is achieved through a 121 \nmodified Python27 implementation of DynamicDet24, built with PyTorch28, expanded to read motion-122 \ncorrected MRC micrographs and output particle coordinates in STAR file format compatible with popular 123 \nprocessing packages (Fig 1). We have provided a custom Wiener filter-based denoiser algorithm, based 124 \non the work of CryoSegNet and CryoTransformer6,18, with a novel multiprocessing layer allowing for 125 \nefficient denoising of large datasets. We have also provided scripts for preparing custom datasets for 126 \ntraining PartiNet on user-generated datasets and for finetuning of the current model. PartiNet is designed 127 \nfor command line use and supports high performance computing (HPC) environments, enabling scalable 128 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\nmulti-GPU/CPU execution via job schedulers such as Slurm. Typical PartiNet workflows comprises three 129 \nautomatable steps: micrograph denoising, particle detection, and coordinate filtering with conversion to 130 \nSTAR format for downstream analysis (Fig 1). The ability to perform size-agnostic particle picking, 131 \nunlike crYOLO and Topaz, represents a key feature of PartiNet and enables seamless integration into 132 \nfully automated, unsupervised cryo-EM image processing pipelines.   133 \n 134 \nPartiNet training on comprehensive dataset enables generalisation without user-specific fine-tuning 135 \nPartiNet was trained using the CryoPPP dataset, a curated benchmark comprising ~10,000 labelled 136 \nmicrographs from 34 cryo-EM protein datasets (Supplementary Table 1). These datasets span diverse 137 \nprotein sizes, symmetries and imaging conditions, and  include gold standard particle coordinates 138 \nmanually curated by the authors of CryoPPP29,  enabling PartiNet to generalise to new datasets without 139 \nrequiring user-specific training or fine-tuning - a significant advantage over methods that require 140 \nretraining for optimal performance. We used seven, randomly selected datasets as test sets for the 141 \ncomparison of crYOLO, Topaz and PartiNet performance. For the remaining datasets, 80% of 142 \nmicrographs from each dataset were allocated randomly for training, and the remaining 20% were used 143 \nfor validation. The training, validation and test sets comprised 6224, 1563 and 1879 micrographs, 144 \nrespectively. Since each of the CryoPPP datasets selected contained only 300 micrographs per set, we 145 \nalso prepared 5 full size datasets for particle picking and reconstruction: MORC2 bound with the 146 \nH3K9me3 peptide (in-house), rabbit muscle aldolase previously tested with Topaz (EMPIAR-10215)14, 147 \nTcdA1 originally published with crYOLO (EMPIAR-10089)15, and two heterogeneous samples: Ankyrin-148 \n1 complex (EMPIAR-11043)30 and MlaCD complex (EMPIAR-12531)31. This design enabled systematic 149 \nassessment of performance across diverse sample sizes and levels of heterogeneity. 150 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\nPartiNet preprocesses micrographs for enhanced picking 151 \nAn obvious first step in particle picking from noisy cryo-EM data is some form of denoising. During 152 \nPartiNet testing, we observed qualitative improvements for precision and recall on denoised micrograph 153 \ndatasets. To investigate this, we compared PartiNet's performance across four preprocessing conditions: 154 \nraw motion-corrected micrographs, and motion-corrected micrographs processed with three different 155 \ndenoising methods: two popular deep learning-based denoisers Janni32 and Topaz33 (both using 156 \ntheNoise2Noise framework34), and a heuristic Wiener filter -based algorithm introduced in CryoSegNet6.  157 \nTraining was conducted for 100 epochs using CryoPPP datasets across all four conditions. We evaluated 158 \nthe pre-trained models provided by the Topaz and JANNI developers to assess their out-of-the-box 159 \nperformance, rather than training custom denoiser models for each optical condition in our dataset. We 160 \nacknowledge that users may see improved performance with Topaz and JANNI denoisers using trained 161 \nsub-models33.  162 \nFor each training condition, we calculated and plotted the Mean Average Precision @ 50% confidence, 163 \nPrecision, and Recall (Supplementary Fig 1a). We observed that PartiNet-CryoSegNet denoise scored the 164 \nhighest across these metrics, followed closely by raw motion-corrected micrographs and then Topaz-165 \ndenoised, with JANNI-denoised data lagging substantially across all metrics.  166 \nWe suspected that JANNI and Topaz had sub-par performance in our testing due to the shared underlying 167 \nNoise2Noise algorithm adversely affecting the quality of the particles in the micrographs. We plotted the 168 \nparticle coordinates generated by PartiNet for all 4 regimes on the same micrograph selected from the 169 \nInfluenza haemagglutinin trimer dataset (EMPIAR-10093; Supplementary Fig 1b-d). With no denoising 170 \napplied, PartiNet was able to identify some particles but struggled with the complex noise of the image, as 171 \nwell as differentiating individual particles near contamination (Supplementary Fig 1b). We observed that 172 \nboth Topaz and JANNI suppressed high frequency information across this micrograph and seemed to be 173 \nespecially susceptible to this behaviour when electron-dense contamination was present in the 174 \nmicrograph. Particles close to contamination were “flattened”, preventing effective delineation of closely 175 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\npacked particles, both during denoising and subsequent picking with PartiNet (Supplementary Fig 1c-d). 176 \nConversely, the PartiNet implementation of CryoSegNet denoise retained discrete particles present in the 177 \nmicrographs, even when the particles were closely packed to the boundary of signal-dense contamination 178 \n(Supplementary Fig 1e). 179 \nGiven the efficacy of the CryoSegNet denoiser in conjunction with PartiNet, we integrated it directly into 180 \nour software. The original denoiser was a single-threaded process, significantly bottlenecking denoising 181 \nof large datasets. We developed a multiprocessing pipeline for this denoiser using the concurrent.futures 182 \nmodule in Python, allowing for asynchronous execution of batch denoising for micrographs. Our 183 \nimplementation is especially suited for efficient use of HPC resources using job schedulers. Denoising 184 \n1000 micrographs on a single node with 32 CPU cores had a CPU efficiency of 94.47% for 01:41:20 185 \n(hh:mm:ss) walltime, thus demonstrating that denoising can be performed efficiently at scale without 186 \nintroducing a preprocessing bottleneck. 187 \nDuring training, we augmented each micrograph before passing it through PartiNet. Augmentations 188 \ncomprised a random combination of image transformations, each of which had a specified probability to 189 \nbe applied to every micrograph during each epoch (Supplementary Table 2, Supplementary Fig 2). Across 190 \ntraining epochs, micrographs were randomly augmented for each pass. This served to prevent model 191 \noverfitting and increased the feature space of the training data for PartiNet to learn 35. During inference, 192 \nPartiNet employs test-time augmentation (TTA) to improve detection robustness and accuracy 20. Each 193 \nmicrograph is processed at multiple scales (100%, 83%, 67%) and with horizontal flipping to increase the 194 \nfeature space available for picking. The predictions from all augmented versions are then transformed 195 \nback to the original coordinate system and combined. This allows PartiNet to aggregate features identified 196 \nat different resolutions and orientations, increasing detection confidence through multiple perspectives of 197 \nthe same micrograph. 198 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\nPartiNet implements a refined detector configuration 199 \nPartiNet was designed with YOLOv7 networks for each detector. We prepared 6 different configurations 200 \nof YOLOv720, with differing numbers of model parameters, input sizes, reported inference speed in image 201 \nframes per second, and reported performance on the benchmarking object detection dataset COCO  in 202 \nmAP@50% (Supplementary Table 3)36 . We trained each of these configurations for 100 epochs on 203 \nCryoPPP training data denoised with PartiNet-CryoSegNet. We measured the mAP@50%, Precision and 204 \nRecall, and observed that for all metrics, YOLOv7-W6 had the highest scores after 100 epochs of training 205 \n(Supplementary Fig 3). YOLOv7, YOLOv7-X and YOLOv7-E6E achieved similar performance metrics, 206 \nranking as the second-best performing group, withYOLOv7-E6 scoring the lowest overall. In our hands, 207 \nYOLOv7-D6 failed to converge, and would crash repeatedly during testing, and so was discontinued. 208 \nBased on these results, we opted to use YOLOv7-W6 for all subsequent testing. 209 \n 210 \nBenchmarking PartiNet against established AI pickers shows improved performance 211 \nTo compare the performance of PartiNet against Topaz, and crYOLO, we picked particles and 212 \nreconstructed density maps on the 7 test datasets from cryoPPP (Supplementary Table 4). We evaluated 213 \nthe performance of PartiNet against gold-standard networks crYOLO and Topaz in terms of number of 214 \nparticles picked and the final resolution of the 3D maps. We mirrored a test workflow previously outlined 215 \nin CryoSegNet and CryoTransformer to fairly compare the networks for all datasets6,18 (Supplementary 216 \nFig 4). We report the number of particles picked by each of Topaz, crYOLO and PartiNet, the number of 217 \nparticles remaining after \"Select 2D\" for reconstruction, and the final global resolution of the 3D map 218 \nreconstruction with these selected particles (Fig 2).  219 \nFor six of the seven test sets, PartiNet identified more particles prior to \"Select 2D\" filtering than the 220 \nother AI methods, with crYOLO and Topaz alternating as the second-best performer. The exception was 221 \nEMPIAR-10017, where crYOLO identified 3,000 more particles than PartiNet, with Topaz picking less 222 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\nparticles (16,000) than both PartiNet and crYOLO. PartiNet excelled with smaller, lower molecular 223 \nweight proteins (EMPIAR-10532 and EMPIAR-11056), where it substantially outperformed both 224 \ncrYOLO and Topaz in particle counts. This likely reflects PartiNet's more aggressive identification 225 \nstrategy when dealing with smaller proteins that are inherently difficult to detect. A similar pattern was 226 \nobserved for EMPIAR-10093, despite this protein's high molecular weight of 779 kDa. We attribute this 227 \nto the challenging non-globular shape of NOMPC ion channel proteins, which complicates particle 228 \nidentification14. After “Select 2D” filtering, PartiNet consistently retained the most particles across all test 229 \nsets except EMPIAR-10017, followed by crYOLO and then Topaz. Most importantly, PartiNet achieved 230 \nthe best final resolution for all datasets tested (Supplementary Fig 5). For EMPIAR-10017, this superior 231 \nresolution was achieved despite PartiNet having fewer retained particles than crYOLO, suggesting that 232 \nPartiNet identified higher-quality particles with better sampling of protein views. In fact, for all CryoPPP 233 \ntest sets, PartiNet demonstrated stronger sampling of protein views, with alignments of particles covering 234 \na broader range of Euler sphere when compared to crYOLO and Topaz (Supplementary Fig 6).  235 \nWe extended our comparison of PartiNet, crYOLO and Topaz with two further datasets: TcdA1 236 \n(EMPIAR-10089)15 and rabbit muscle aldolase (EMPIAR-10215)14, published with crYOLO and Topaz 237 \nmethod papers14,15, respectively. We first evaluated the computational efficiency of PartiNet and crYOLO 238 \nby measuring picking speed in micrographs per second using per-micrograph inference timing (Fig 3a). 239 \nTopaz was excluded from this speed comparison as it supports only single GPU processing, while both 240 \ncrYOLO and PartiNet support parallel processing with up to 4 GPUs under identical hardware conditions. 241 \nPartiNet demonstrated substantially faster processing speeds than crYOLO across both datasets, with 242 \nperformance improvements of up to seven-fold on the rabbit muscle aldolase dataset (EMPIAR-10215) 243 \nand a peak inference speed of 8 micrographs/second for TcdA1.  244 \nNext, we applied a standardised workflow for processing TcdA1 to compare PartiNet, Topaz and 245 \ncrYOLO objectively (Supplementary Fig 7). We performed 2D classification on PartiNet-selected 246 \nparticles to assess particle quality through visual inspection of 2D class averages (Fig 3b). The resulting 247 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\naverages showed coherent, well-centred representations of TcdA1 spanning multiple orientations, with 248 \nclearly resolved secondary structure features. Subsequent 3D reconstruction using PartiNet particles 249 \nyielded a high-quality density map with well-defined secondary structure and no observable anisotropy 250 \nthat would result from particle orientation bias (Fig 3c). For comparative analysis, we reconstructed maps 251 \nof TcdA1 with crYOLO and Topaz coordinates and plotted Fourier Shell Correlation (FSC) curves for 252 \neach density map (Fig 3d). At an FSC cutoff of 0.143, PartiNet had the highest global resolution of 3.0 Å, 253 \nfollowed by Topaz and crYOLO with 3.2 Å. These resolutions represent improvements over the 3.4 Å 254 \nstructure published originally for EMPIAR-10089, likely reflecting differences in reconstruction 255 \nalgorithms between SPHIRE and CryoSPARC. Finally, we compared PartiNet’s performance in terms of 256 \nparticle counts before and after cleaning by 2D classification (Fig 3e). PartiNet identified more particles 257 \n(15,551) than Topaz or crYOLO (13,380 and 8,282, respectively). After filtering, PartiNet again retained 258 \nthe most particles (10,016) compared to crYOLO (9,292) and Topaz (8,258), revealing that PartiNet 259 \ncoordinates were well ranked. Interestingly, crYOLO retained effectively all particles picked, suggesting 260 \nthat the publicly available model for crYOLO may have been trained on this dataset prior to publication.  261 \nTo complete analysis of TcdA1, we performed de novo, automated model building with the reconstructed 262 \nmaps from PartiNet, crYOLO and Topaz particles using ModelAngelo37 (Supplementary Fig 8). All three 263 \nmodels demonstrated excellent sequence coverage and high ModelAngelo confidence scores for the 264 \nmajority of residues. Mean confidence scores were almost identical: PartiNet’s model had the highest 265 \nmean confidence of 99.4%, followed by crYOLO with 99.2% and then Topaz with 98.9%. We also 266 \ncalculated the Root Mean Square Deviation (RMSD) in Å between a single monomer of the 267 \nModelAngelo prediction and the crystal structure of TcdA1 (PDB 4O9Y). PartiNet had the lowest RMSD 268 \nof 0.67, followed by Topaz with 0.71 and then crYOLO with 0.78. All values were < 1.0 Å for most of 269 \nthe monomer sequence, indicating that resolutions of 3.0-3.2 Å are sufficient for accurate model building 270 \nof TcdA1 in this dataset. We completed our assessment of the models by calculating Q-scores, an 271 \nindependent metric for map-model fitness, defined as the measure of an atom's resolvability within a 272 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\ncryo-EM map. PartiNet had the highest Q-score with 0.68, followed by Topaz and crYOLO with 0.65, 273 \nindicating PartiNet’s reconstruction led to the highest quality map built with ModelAngelo.  274 \nWe repeated this analysis on rabbit muscle aldolase, again with a standardised workflow of picking with 275 \nPartiNet, Topaz or crYOLO (Supplementary Fig 9). Subsequent 3D reconstruction of rabbit muscle 276 \naldolase with PartiNet showed a high-quality map with clear secondary structure, with no observable 277 \norientation bias or anisotropy (Fig 3g). At an FSC cutoff of 0.143, PartiNet had the highest global 278 \nresolution at 2.8 Å, followed by Topaz with 3.0 Å and crYOLO with 3.1 Å (Fig 3h). We report the 279 \nnumber of picked particles and those retained after filtering (Fig 3i). crYOLO picked the highest number 280 \nof particles (1,621,138) compared to PartiNet (364,185) and Topaz (61,822); however, after filtering, 281 \nPartiNet retained the most particles (212,176) followed by crYOLO (57,639) and Topaz (23,078). The 282 \nlarge discrepancy between crYOLO’s picked and retained particles may be due to the size of the protein 283 \nor the high density of particles in micrographs.  284 \nWe again used ModelAngelo to validate the reconstructed maps of rabbit muscle aldolase (Supplementary 285 \nFig 10). Whilst PartiNet and crYOLO’s models had comparable sequence coverage, ModelAngelo was 286 \nunable to build most of the sequence into the Topaz map. We inspected the map of rabbit muscle aldolase 287 \nbuilt with Topaz particles and observed strong anisotropy and poor reconstruction of secondary structure, 288 \nsuggesting that even though the map had a global resolution of 3.0 Å, it was not sufficient to build an 289 \nappropriate model. We reported the confidence of residue predictions, as RMSD, between the 290 \nModelAngelo predictions and the crystal structure of rabbit muscle aldolase (PDB 6ALD), and map-291 \nmodel fit with Q-scores.  PartiNet had the highest mean confidence of residue predictions at 99.7% 292 \nfollowed by crYOLO with 99.1% and Topaz with 95.3%. PartiNet had the lowest mean RMSD of 0.45, 293 \nfollowed by crYOLO with 0.53 and Topaz with 5.3. PartiNet had the highest mean Q-score of 0.68, 294 \nfollowed by crYOLO with 0.66 and Topaz with 0.63. PartiNet consistently outperformed both crYOLO 295 \nand Topaz across all metrics in terms of models built from particles picked. 296 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\nTogether, these results demonstrate that PartiNet consistently delivers superior particle selection, faster 297 \ninference, and higher-quality reconstructions and atomic models across diverse datasets, establishing 298 \nimproved accuracy/efficiency trade-offs relative to existing AI-based particle pickers. 299 \nReconstructions from PartiNet particles enable comprehensive mapping of protein sequences 300 \nAfter confirming PartiNet’s performance on the 7 test datasets from CryoPPP, rabbit muscle aldolase and 301 \nTcdA1, we extended our analysis to an unpublished dataset for the chromatin remodeler MORC238 bound 302 \nto the H3K9me3 peptide (Supplementary Fig 11) and compared performance against Topaz and crYOLO 303 \n(Fig 4). First, we compared the picking speed of PartiNet against crYOLO for the dataset and measured 304 \nPartiNet to be 5-fold faster (Fig 4a). Next, we again performed a standardised workflow for processing 305 \nMORC2 with PartiNet, Topaz and crYOLO picks (Supplementary Fig 12). We plotted 2D averages for 306 \nMORC2 particles picked with PartiNet, which showed convergence to high quality, coherent 2D classes, 307 \nshowing secondary structure and evidence of strong sampling across multiple views of MORC2 (Fig 4b). 308 \nNext, we plotted the FSC curves for each map and observed that the 3D map reconstructed with PartiNet 309 \npicks had a global resolution of 2.3 Å, compared to 2.5 Å and 2.7 Å with crYOLO and Topaz, 310 \nrespectively (Fig 4c). To assess the quality of the 3D map reconstruction from PartiNet, Topaz and 311 \ncrYOLO picks, we took the final reconstruction of each map and estimated the resolution of each voxel 312 \n(rather than a global resolution) with the \"Local Resolution Estimation\" function in CryoSPARC (Fig 4d). 313 \nAs expected, the central globular structure of all MORC2 maps were resolved best due to this being 314 \ncomparatively rigid and electron dense. Conversely, the two double coil (CC) domain arms extending 315 \nfrom the central body of the MORC2 were not resolved. This was expected, as these domains are highly 316 \nflexible because they are responsible for the DNA compaction activity of MORC2, and the presence and 317 \nactivation of these domains are evidenced in other works where 3D variability analysis showed this 318 \nmovement38. The PartiNet map’s resolution was isotropic, whereas crYOLO’s and Topaz’s maps 319 \ndisplayed local resolution loss on the fringes of the protein, and in the case of Topaz, within the central 320 \nstructure of MORC2 itself. Additionally, when viewed from the bottom, Topaz’s map exhibited 321 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\nanisotropy in the form of “smearing” between the flat front faces, indicating poor sampling of the top and 322 \nbottom views of MORC2 during particle picking (Fig 4d, bottom panel). We extended our analysis to 323 \nmeasure the quality of sampling across all views of MORC2. We calculated cFAR and SCF values for 324 \neach map with the \"Orientation Diagnostics\" in CryoSPARC, with higher cFAR and SCF scores 325 \ncorresponding to stronger sampling across all views of the proteins39,40. This value is effectively an 326 \nestimation of \"orientation bias\" in the map, where certain views of the protein are heavily sampled and 327 \nrarer views of the protein are undersampled. This occurs due to the tendency of many proteins to partition 328 \nto the air-water interface of the protein solution directly prior to vitrification, and represents a major 329 \nbottleneck for data analysis. PartiNet had the highest cFAR score of 0.27, followed by crYOLO with 0.11 330 \nand the Topaz with 0.03. Interestingly, PartiNet had a SCF score of 0.77 compared to crYOLO’s 0.78, 331 \nwhereas Topaz has a substantially different SCF of 0.65. These discrepancies reveal interesting details 332 \nabout the particles picked by PartiNet and crYOLO. cFAR is calculated by measuring the correlations of 333 \nhalf-maps in specific viewing angles in Fourier space39, meaning that both the alignment and quality of 334 \nsignal in that specific alignment is necessary for a good score. Conversely, SCF is calculated by 335 \nquantifying how the alignments of particles cover the Euler sphere without quantifying signal quality 336 \nfrom particles40. Given that PartiNet map had the highest global resolution and cFAR score, it can be 337 \nconcluded that PartiNet particles contributed meaningful particle signal information from alignments, 338 \nwith good coverage over the viewing angles of the protein. crYOLO, on the other hand, had comparable 339 \nnumbers of particles spanning the Euler sphere but may have contained more poor particles or particles 340 \nthat contributed poorly to the assigned alignment. LowTopaz scores for cFAR, SCF and number of 341 \nparticles indicate that these particles were of lesser quality and exhibited strong orientation bias, with poor 342 \ncontribution to signal in many alignments.  343 \nWe completed our analysis of MORC2 by using a reconstructed 3D map from PartiNet to build a 344 \nstructural model of MORC2 using ModelAngelo37. We superimposed this model with a published X-ray 345 \ncrystallography structure41 (PDB-5OF9, Fig 4e). We observed strong congruence between these models, 346 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\nespecially in the central structure of the protein, with the notable absence of the two flexible coiled-coil 347 \n(CC1) domains in our model. This is a limitation of cryo-EM, as highly flexible protein domains are 348 \ndifficult to resolve to high resolution with 3D cryo-EM maps without further advanced processing tools42. 349 \nWe plotted the per-residue confidence, RMSD with the crystal structure and Q-score for model-map fit 350 \n(Supplementary Fig 13a-c). In vitro and in vivo, MORC2 exists as a homodimer38,41. We calculated 351 \nRMSD for each monomer of each map, and found that our reconstructed map and the published map 352 \ndeviated within 1 Å of each other for the whole protein, except for unstructured tails and double CC 353 \ndomain arms, which are highly flexible (Supplementary Fig 13d).We observed a segment of the MORC2 354 \nprotein that was mapped in our cryo-EM model but was not present in the published crystal model (Fig 355 \n4f). In the crystal structure, eight amino acids between F510 and Y519 were absent; in contrast, the 356 \nMORC2 cryo-EM density map allowed us to build all but one residue. Consequently, we accurately 357 \ndetermined the structural context of seven amino acids that were unresolved in the published model.  358 \nCollectively, these analyses show that PartiNet enables faster particle picking, improved orientation 359 \nsampling, and higher-quality reconstructions on previously unseen datasets, facilitating accurate model 360 \nbuilding and recovery of structural features that were unresolved in prior studies. 361 \n 362 \nPartiNet enables size-agnostic, single-run identification of multiple species  363 \nPartiNet's size-agnostic detection - which requires no manual specification of particle dimensions - 364 \nenables simultaneous identification of multiple species in heterogeneous datasets. This eliminates the 365 \nneed for multiple picking runs with different size parameters, a common requirement for existing methods 366 \nwhen analyzing samples containing particles of varying sizes. We investigated this capability using two 367 \npublished datasets containing multiple molecular species We first evaluated this on EMPIAR-11043, 368 \ncontaining the Ankyrin-1 complex in a micelle, and the free ‘Band 3’ protein (Fig 5)30. Processing this 369 \ndataset proved to be quite involved for the original authors, as identification and processing of the two 370 \nspecies required 4 different rounds of particle picking (including manual picking and training a Topaz 371 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\nmodel) in conjunction with iterative 2D and 3D classification before model refinement. Using PartiNet, 372 \nwe were able to demonstrate a high-speed, single picking step which was able to identify both species 373 \nusing our general model (Fig 5a-b), whilst simultaneously avoiding multiple rounds of particle picking 374 \nand classification (Supplementary Fig 14). To verify whether both species were picked, we plotted the 375 \nbox size of each PartiNet prediction against the confidence associated with each confidence as a 2D 376 \nhistogram. Bivariate and univariate kernel density estimations identified two distinct populations of box 377 \nsizes in PartiNet coordinates, corresponding approximately to the expected sizes of Ankyrin-1 and Band 3 378 \n(Fig 5c). From these coordinates, we were able to import, extract and refine these particlesubsets, 379 \nresulting in 3.1 Å and 3.2 Å maps of the Ankyrin-1 complex and Band 3, respectively (Fig 5d-f). 380 \nWe completed our validation of PartiNet on EMPIAR-12531 containing the MlaCD complex with two 381 \nspecies comprising 1:6 and 2:6 stoichiometries of MlaC:MlaD (Fig 6, Supplementary Fig 15)31. PartiNet 382 \nwas able to pick 2 times faster than crYOLO on this dataset (Fig 6a). PartiNet also was able to contribute 383 \nmore particles to final reconstruction of each species than the template picking used in the original 384 \npublication (Fig 6b). PartiNet picked 1,802,428 particles, of which 211,021 were used for 1:6 385 \nreconstruction, and 203,014 for 2:6. Conversely, in the original publication, 519,770 particles were picked 386 \nwith template picking and 97,460 particles were used for 1:6 reconstruction and 58,259 for 2:6 387 \nreconstruction. Using PartiNet picks, we were able to calculate reconstructions with 1:6 species at 3.7 Å 388 \nand 2:6 species at 3.4 Å compared to 4.4 Å for each species published previously31 (Fig 6c). Smearing of 389 \nsignal was seen in 3D reconstructions of both complexes, suggesting suboptimal “stacking” of the protein 390 \ncomplexes in the sample. This necessitated masking for only the complex in final refinements 391 \n(Supplementary Fig 15). Subsequent reconstructions of both 1:6 and 2:6 species showed well-defined and 392 \nisotropic secondary structure (Fig 6e-h), with well-resolved interfaces between subunits in the 2:6 393 \ncomplex. Particle distribution plots confirmed distributed sampling across most views for both species, 394 \nwith excellent angular coverage (Fig 6f-i). With the reconstruction of the two species, we were able to 395 \nconfirm the observation from the original authors that the binding of the MlaC subunit to the MlaD 396 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\nmultimer breaks the hexameric symmetry of the internal pore, with the MlaD monomer contracting 397 \nslightly at the centre. We also observed a density in the central pore of the 2:6 species reconstruction; 398 \nhowever, we were unable to determine at this resolution if this was the presence of lipid or an artifact 399 \nfrom refinement. 400 \nTaken together, the results from Ankyrin-1 and MlaC/D complexes show that PartiNet can recover 401 \nmultiple molecular species from heterogeneous datasets using a single picking run, improving both 402 \nthroughput and reconstruction quality in multi-component samples. 403 \n 404 \nDiscussion 405 \nPartiNet is a dynamic neural network architecture for particle picking in cryo-EM that addresses critical 406 \nbottlenecks in protein structure determination. By implementing adaptive inference pathways that adjust 407 \ncomputational complexity based on micrograph difficulty, PartiNet achieves substantial improvements in 408 \nboth speed and accuracy compared to existing methods. Our comprehensive validation demonstrates 409 \nconsistent improvements in particle identification, reconstruction quality, and computational efficiency 410 \nacross diverse datasets spanning different protein classes, sizes, and imaging conditions. 411 \nPartiNet is a dynamic particle picker for cryo-EM 412 \nPartiNet is the first dynamic architecture applied to cryo-EM particle picking. Unlike static networks that 413 \nprocess all micrographs through fixed detection layers 6,12,14–16,18, PartiNet's adaptive router adjusts 414 \nnetwork depth in situ based on learned image features. This addresses a fundamental limitation: the 415 \ninability to modulate computational resources according to image difficulty. Critically, the router learns to 416 \nidentify challenging features - complex noise, contamination, variable defocus, support film—without 417 \nsupervision, capturing imaging and sample-related complexity without manual labeling. This modularity 418 \nenables future detector upgrades without architectural redesign and opens possibilities for adaptive 419 \napproaches in other cryo-EM steps including 3D classification, refinement, model building. 420 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\nPartiNet enables truly automated on-the-fly processing workflows 421 \nConsistent resolution improvements of up to 1.0 Å suggest PartiNet's selection criteria better align with 422 \nreconstruction requirements, reflecting diverse CryoPPP training and dynamic resource allocation for 423 \nhigh-quality particle detection. Superior ModelAngelo metrics confirm that improvements translate to 424 \nmore reliable atomic models37. 425 \nCritically, PartiNet's combination of speed (2–7× faster) and size-agnostic picking—without prior 426 \nparameter specification—enables truly automated workflows. Existing methods require manual size input, 427 \nand they struggle with heterogeneous samples, necessitating user intervention. PartiNet's YOLO 428 \narchitecture20 simultaneously determines presence, location, and dimensions, handling heterogeneous 429 \nsamples (Ankyrin-1/Band 3, MlaCD) in single passes versus multiple rounds with manual intervention. 430 \nThis transforms cryo-EM practice: real-time particle detection with quality metrics provides immediate 431 \nassessment of sample quality, grid selection, and defocus optimisation during acquisition rather than days 432 \nlater. Integration with existing on-the-fly platforms - Warp/M43, CryoSPARC Live44, and RELION3 - 433 \nwould enable direct sample assessment during imaging. Microscopists can identify failing samples or 434 \noptimal parameters without interrupting collection, substantially reducing experimental iteration cycles 435 \nand microscope time waste, which is particularly valuable for high-throughput facilities processing 436 \ndiverse samples. 437 \nLimitations and considerations for practical deployment 438 \nSeveral limitations warrant consideration. First, extremely unusual morphologies (elongated filaments, 439 \n<50 kDa proteins, extreme aspect ratios) may benefit from fine-tuning, requiring GPU resources (≥8 GB 440 \nVRAM) and PyTorch familiarity. Second, speed advantages require multi-GPU setups (our benchmarks: 441 \n4× A100 GPUs); single-GPU users may see only modest improvements. Memory requirements (~64 GB) 442 \nmay limit older hardware. Third, severe micrograph quality issues (motion artifacts, thick ice, damage, 443 \ncontamination) challenge all AI pickers25 - no computational approach substitutes for high-quality sample 444 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\npreparation. Fourth, while PartiNet identifies multiple species by size, closely-sized particles with similar 445 \ncontrast require downstream 2D/3D classification. Fifth, PartiNet uses YOLOv7 (2022); newer 446 \narchitectures like YOLOv1045 and Detection Transformer46 may offer improvements, though the modular 447 \nframework facilitates future upgrades. 448 \nIn conclusion, PartiNet demonstrates that dynamic architectures can simultaneously improve speed, 449 \naccuracy, and robustness in cryo-EM particle picking. Several lessons emerge from this work: diverse 450 \ntraining data is essential for broad applicability, modular architectures allow future improvements, 451 \ncombining AI with domain-specific methods (like appropriate denoising) yields better results than AI 452 \nalone, and open-source distribution accelerates community adoption. 453 \nAs AI integration into cryo-EM accelerates, approaches balancing automation with flexibility and 454 \nproviding interpretable outputs (confidence scores, difficulty classifications) will be essential. The 455 \ndynamic architecture paradigm provides a template for future developments that adapt intelligently to 456 \nbiological complexity.  457 \nOnline Methods 458 \nDynamic architecture inference and training 459 \nPartiNet is built on a dynamic, deep learning architecture called DynamicDet24 and trained on EM 460 \nmicrographs to detect and localise protein particles. Dynamic architectures can adjust the network 461 \narchitectures in response to different imaging and sample conditions, making it possible to train models 462 \nefficiently, and also to detect particles with increased speed and accuracy22,23. 463 \nStatic object detection algorithms (hereafter called detectors) contain a backbone, neck and head19,45. 464 \nDuring model inference, the backbone performs the bulk of the feature extraction operations on the input 465 \nimage, the neck pools and aggregates these features, and the head performs the final detection by 466 \ngenerating bounding boxes or providing pixel coordinates. This architecture is common across all detector 467 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\nclasses. DynamicDet differs from other object detection algorithms because, instead of passing images 468 \nsequentially through the backbone, neck, and head of a single detector, it evaluates the intermediate 469 \noutputs at each stage to adjust the network’s size for processing the input image (Fig 1). 470 \n 471 \nAssume for a static Detector 𝐷1 which has backbone 𝐵1, neck and head 𝐻1 such that for a given input 472 \nimage x: 473 \n𝑦 = 𝐷1(𝑥) = 𝐻1(𝐵1(𝑥)) 474 \nwhere 𝑦 is the output (coordinates/bounding boxes/contour) of a static object detection network 𝐷1 on 475 \ninput image 𝑥. 476 \nInstead of a static network, DynamicDet performs the following operations on input 𝑥: 477 \n𝐹1 = 𝐵1(𝑥) 478 \nwhere 𝐹1 are the multiscale feature maps extracted by the first backbone 𝐵1. These feature maps are fed to 479 \nan adaptive router 𝑅 which determines a difficulty score 𝜙 of the image: 480 \n𝜙 = 𝑅(𝐹1), 𝜙 ∈ (0,1) 481 \nIf an image is easy (based on a learned threshold from training, see below) then processing is completed 482 \nby the first detector: 483 \n𝑦 = 𝐻1(𝐹1) 484 \nHowever, if the image is ‘hard’, then these features 𝐹1 are passed along with the input image 𝑥 to the 485 \nsecond detector such that: 486 \n𝐹2 = 𝐵2(𝑥, 𝐹1)\n𝑦 = 𝐷2(𝐹2)  487 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\nIn this way, DynamicDet can dynamically route images to a complex or simpler network based on 488 \nperceived difficulty of the input images. Currently, PartiNet uses two identical YOLOv7 detector 489 \nbackbones20; however there is scope to change these detectors, highlighting the modularity of the 490 \nDynamicDet framework. 491 \n 492 \nTotal training of a model for PartiNet requires two steps: training of the detectors, then training of the 493 \nadaptive router (Supplementary Fig 16). To train the adaptive router, the model weights of the dual 494 \ndetector are frozen, leaving only the adaptive router to be trained to determine if a micrograph is easy or 495 \nhard 24. This is accomplished with an adversarial loss comparison. A micrograph is passed through a 496 \nsingle detector and the total loss ℒ1 is calculated. Then, the micrograph is passed through both detectors 497 \nand total loss ℒ2 is calculated. If the difference between the two ℒ2 − ℒ1 is low, there is no advantage in 498 \nusing two detectors for the micrograph, and the adaptive router learns this micrograph as easy. If ℒ2 − ℒ1 499 \nis high then there is a distinct advantage in using the second detector and the micrograph is labelled as 500 \neasy. Importantly, this stage is unsupervised: the training dataset does not need to be labelled individually 501 \nas easy or hard (for example, by a skilled human). Instead, only the micrograph is required, removing bias 502 \nduring training of the adaptive router. The adaptive router is able to encode important optical 503 \nconsiderations from the image acquisition into its training regime without the need for explicit labelling. 504 \nThis allows for in situ dynamic routing of images based on the difficulty of input micrographs. 505 \nCalculation of Precision, Recall and mAP@50% 506 \nIn PartiNet, the Intersection over Union (IoU) is used for assessing the overlap between predicted and 507 \nground truth bounding boxes. The IoU was calculated as: 508 \nIoU =\n|𝐵pred ∩ 𝐵gt|\n|𝐵pred ∪ 𝐵gt| 509 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\nwhere 𝐵pred represents the predicted bounding box and 𝐵gt represents the ground truth bounding box. The 510 \nconfidence score associated with each predicted bounding box represents the model’s confidence of 511 \nprediction in both the protein particle’s presence and the accuracy of its localisation. PartiNet also 512 \ncalculates Precision and Recall during training: 513 \nPrecision = 𝑇𝑃\n𝑇𝑃 + 𝐹𝑃 514 \nRecall = 𝑇𝑃\n𝑇𝑃 + 𝐹𝑁 515 \nWhere: 516 \n• 𝑇𝑃 (True Positives): Correctly detected objects with IoU ≥ 0.5 517 \n• 𝐹𝑃 (False Positives): Detected objects with no corresponding ground truth 518 \n• 𝐹𝑁 (False Negatives): Ground truth objects not detected 519 \nThe mean Average Precision at 50% IoU threshold (mAP@50%) is calculated as: 520 \nmAP@50 = 1\n𝑁𝑐\n∑ AP𝑐\n𝑁𝑐\n𝑐=1\n 521 \nwhere: 522 \n• 𝑁𝑐: Total number of classes (in the case of PartiNet 𝑁𝑐 = 1) 523 \n• AP𝑐: Average Precision for each class, computed by integrating the precision-recall curve 524 \nThe precision-recall curve was generated by varying the confidence threshold and calculating precision 525 \nand recall at each point. The area under this curve represents the Average Precision for a given class45. 526 \nBounding box coordinates are provided with an associated confidence in PartiNet to represent a 527 \nprobabilistic assessment combining a) the probability that protein particle is present in the box, and b) the 528 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\ncorrect localisation of the box in the micrograph.  Confidence in the correct class assignment is only 529 \nrelevant for multi-class inference strategies, which PartiNet does not use. Confidence can be calculated as 530 \nConfidence = 𝑃(Object Exists) × Localisation Accuracy 531 \nConfidences can be plotted as a histogram to give an initial assessment of the performance of PartiNet. A 532 \nconfidence threshold is specified: bounding boxes with a confidence above this threshold are retained and 533 \nthose below are discarded. PartiNet defaults to a threshold of 0.3 which provides a balance between 534 \nretaining the majority of particles identified whilst rejecting most junk and spurious picking. 535 \nCryoPPP Image Processing 536 \nThe CryoPPP test datasets were processed in CryoSPARC v4.6.2 (Supplementary Fig 4). We utilised our 537 \ntrained PartiNet model with a confidence threshold of 0.3, crYOLO’s publicly available model with the 538 \n\"PhosaurusNet\" architecture 15 at the default confidence threshold of 0.3, and Topaz’s publicly available 539 \nmodel with the \"ResNet16 (64 units)\" integrated directly in CryoSPARC4,14 with default parameters to 540 \npick particles. We picked particles on all 7 datasets denoised with our integrated denoiser and imported 541 \nthe coordinates alongside the motion-corrected micrographs. We performed CTF estimation on the 542 \nmicrographs to correct for microscope aberrations, then extracted protein particles with an appropriate 543 \nbox size (1.5x largest particle diameter) and performed 2D classification with 50 classes. 2D averages 544 \nwere selected corresponding to protein particles with CryoSPARC’s interactive \"Select 2D\" function. 545 \nThese selected particle stacks were used for \"ab initio Reconstruction\", to coarsely reconstruct an initial 546 \n3D map at low resolution without a reference. This coarse reconstruction was then refined using 547 \nCryoSPARC’s \"Homogeneous Refinement\" with C1 symmetry for each dataset leading to the final map 548 \nfor evaluation. 549 \nTcdA1 image processing and model building 550 \nTcdA1 (EMPIAR-10089) was processed in CryoSPARC v4.6.2. PartiNet, Topaz and crYOLO were 551 \ncompared objectively using a standardised workflow (Supplementary Fig 7). 97 movies were imported, 552 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\nmotion-corrected and CTF estimated. Picking was performed using default parameters of each AI picker:  553 \nthe general trained PartiNet model with a confidence threshold of 0.3, crYOLO’s publicly available 554 \nmodel with the \"PhosaurusNet\" architecture15at the default confidence threshold of 0.3, and Topaz’s 555 \npublicly available model with the \"ResNet16 (64 units)\" integrated directly in CryoSPARC 4,14 with 556 \ndefault parameters to pick particles. Particles were picked on all 97 micrographs denoised with PartiNet’s 557 \nintegrated denoiser. Particles were extracted with a box size of 384 pixels. Extracted particles were 558 \nfiltered in 3D with multi-class “ab initio Reconstruction”, and then subjected to iterative “Heterogeneous 559 \nRefinement” and “Homogeneous Refinement” of the best class using C5 symmetry. After filtering 560 \nparticles in 3D, “Non-Uniform Refinement” was applied with Global CTF Refinement (Tilt, Trefoil, 561 \nSpherical Aberration, Tetrafoil and Anisotropic Magnification), and optimising per-particle defocus and 562 \nscale. Reconstruction was completed with “Reference Based Motion Correction” and “Homogeneous 563 \nRefinement” of motion-corrected particles with C5 symmetry.  564 \nInitial model building of TcdA1 was performed with the reconstructed maps from PartiNet, crYOLO and 565 \nTopaz picked particles using ModelAngelo v1.0.137, an automated AI model-building pipeline 566 \n(Supplementary Fig 8). Each map along with the FASTA sequence of TcdA1 (UniProt Q9RN43) was 567 \ninput into ModelAngelo and a model prediction was performed. Per-residue confidence scores for each 568 \nresidue were plotted. Root mean square deviation (RMSD) values for a single monomer against the 569 \ncrystal structure of TcdA1 (PDB 4O9Y) were calculated in ChimeraX v1.10.1 using the MatchMaker 570 \nutility with default parameters and then plotted. The Q-score of each residue was calculated and plotted 571 \nwith ChimeraX using the Q-score plugin47.  572 \nRabbit muscle aldolase image processing and model building 573 \nRabbit muscle aldolase (EMPIAR-10215) was processed in CryoSPARC v4.6.2. PartiNet, Topaz and 574 \ncrYOLO were compared objectively using a standardised workflow (Supplementary Fig 9). 1,021 movies 575 \nwere imported, motion-corrected and CTF estimated. Particles were picked with default parameters for 576 \nPartiNet, crYOLO and Topaz as outlined previously. Particles were picked on all 1,021 micrographs 577 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\ndenoised with PartiNet’s integrated denoiser. Particles were extracted with a box size of 256 pixels. 578 \nFollowing extraction of PartiNet particles, “2D Classification” was performed with 50 classes. Due to 579 \npreferred orientation, the ‘four wing’ view of rabbit muscle aldolase was under-represented. To overcome 580 \nthis, the classes were rebalanced after “Select 2D” with 10 superclasses and a rebalance factor of 0.8 581 \n(resulting in ~ 50% of particles being temporarily discarded), and then initial maps were generated with 582 \nmulticlass “ab initio Reconstruction” with C2 symmetry. The initialised maps were inspected and the map 583 \nwhich had all four wings present was selected. With this map, “Homogeneous Refinement” with D2 584 \nsymmetry was performed and included the particles that were discarded during class rebalancing. 585 \nProcessing was completed with “Non-Uniform Refinement” with Global CTF Refinement (Tilt, Trefoil, 586 \nSpherical Aberration, Tetrafoil and Anisotropic Magnification), and optimising per-particle defocus and 587 \nscale with D2 symmetry applied. 588 \nInitial model building of rabbit muscle aldolase was performed with ModelAngelo. Each map along with 589 \nthe FASTA sequence of rabbit muscle aldolase (UniProt P00883) was input into ModelAngelo and a 590 \nmodel prediction was performed. Per-residue confidence scores for each residue were plotted. The RMSD 591 \nvalues for a single monomer against the crystal structure of rabbit muscle aldolase (PDB 6ALD) were 592 \ncalculated in ChimeraX v1.10.1 using the MatchMaker utility with default parameters, and then plotted. 593 \nThe Q-score of each residue was calculated and plotted with ChimeraX using the Q-score plugin47 594 \n(Supplementary Fig 10). 595 \nPurification of MORC2 596 \nMORC21-603 (residues 1-603) was cloned into pFastbac construct with 6xHis at the N-terminus and was 597 \npurified as described previously38. Briefly, MORC2 construct was transformed into EMBacY cells to 598 \ngenerate a bacmid. Bacmid DNA was transfected into Sf9 cells using FuGENE transfection reagent 599 \n(Promega) and virus was passaged twice in the same cell line before large-scale infection. Sf9 cells were a 600 \nkind gift from the Glukhova lab at WEHI (negative for mycoplasma, identity not independently 601 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\nauthenticated by us). For large-scale expression, Sf9 cells at a density of 2–2.5 × 106 cells per mL were 602 \ninfected with 1% (v/v) second passage (P2) virus and incubated at 27 °C for 50–60 h, 220 rpm.  603 \nThe cells were harvested by centrifugation, resuspended in lysis buffer containing 50 mM HEPES, pH 604 \n8.0, 300 mM NaCl, 1 mM TCEP, 5% glycerol, 10 mM imidazole, 10 U per ml benzonase solution 605 \n(Sigma), 1× cOmplete EDTA-free protease inhibitors (Roche) and 5 mM benzamidine hydrochloride. 606 \nCells were lysed by sonication. Clarified cell lysate was incubated with cOmplete His-tag purification 607 \nresin (Merck) for 1 h followed by wash with lysis buffer. MORC2 proteins were eluted in elution buffer 608 \n(50 mM HEPES pH 8.0, 300 mM NaCl, 1 mM TCEP, 5% glycerol and 250 mM imidazole). Further 609 \npurification was performed via size exclusion chromatography using the Superose 6 Increase 10/300 GL 610 \ncolumn (Cytiva) in SEC buffer (50 mM HEPES pH 8.0, 300 mM NaCl, 1 mM TCEP). 611 \nSurface Plasmon Resonance (SPR) 612 \nSPR binding studies of MORC21-603 to H3K9me3 were performed using a Biacore S200 Instrument 613 \n(Cytiva). Biotinylated H3K9me3 peptide (Active Motif) was diluted to 5 µg/mL in SPR running buffer 614 \n(10 mM HEPES pH 7.4, 300 mM NaCl, 3 mM EDTA and 0.05% (v/v) Tween-20) to a final 615 \nimmobilisation level of 200-220 response units (RU) on the Streptavidin chip (Cytiva).  A blank 616 \nactivation/deactivation was used for the reference surface. DNA binding studies were performed at 20 °C 617 \nin SPR running buffer. MORC2 proteins were diluted to 1000 nM stock in SPR running buffer and 618 \nprepared as an 8-point concentration series (2-fold serial dilution, 7-500 nM). Samples were injected in a 619 \nmulti-cycle run (flow rate 30 µL/min, contact time of 60 s, dissociation 120 s) with regeneration with 0.5 620 \nM EDTA buffer. Sensorgrams were double referenced, and steady-state binding data were fitted using a 621 \n1:1 binding model using Biacore S200 Evaluation Software (Cytiva, v1.1). Representative sensorgrams 622 \nand fitted dissociation constant (KD) values are depicted as mean ± SEM (n=3 independent experiments). 623 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\nMORC2 cryo-EM sample preparation and data collection 624 \nA 0.5 mg/mL MORC2 (residues 1-603) protein solution was incubated with 1 mg/mL biotinylated 625 \nH3K9me3 peptide (Active Motif, Catalogue number 81047, peptide sequence = ARTKQTAR-Kme3- 626 \nSTGGKAPRKQLA - GGYK (Biotin) - NH2) and 2.5 mM AMP-PNP for 1 h on ice. The MORC2-627 \nH3K9me3 sample was applied to UltrAuFoil R1.2/1.3 gold grids (Quantifoil). The grids were glow-628 \ndischarged for 120 s before application of 4 µL sample onto the grid. The sample was subsequently 629 \nblotted for 3.5 s (Blot force -3) and vitrified by plunging into liquid ethane using a Vitrobot Mark IV 630 \n(ThermoFisher) operated at 4 ºC and 100% humidity. Cryo-EM data were acquired on a Titan Krios 631 \ntransmission electron microscope (ThermoFisher) operated at 300 keV, equipped with a K3 direct 632 \nelectron detector (Gatan, Pleasanton) with a pixel size of 0.82 Å/pixel and electron dose of 60 e/Å2. 633 \nMORC2 Image Processing and model building 634 \nMORC2 data was processed in CryoSPARC v4.6.2. PartiNet, Topaz and crYOLO were compared using a 635 \nstandardised workflow (Supplementary Fig 12). 6,148 movies were imported, motion corrected and CTF 636 \nestimated. Particles were picked with default parameters for PartiNet, crYOLO and Topaz as outlined 637 \npreviously. Particles were picked on all 6,148 micrographs denoised with PartiNet’s integrated denoiser. 638 \nParticles were extracted with a box size of 256 pixels. Classes were selected from a single round of 2D 639 \nclassification. Maps were initialised for each with “ab-initio Reconstruction”, and “Homogeneous 640 \nRefinement” with C2 symmetry was performed.  A final round of “Non-Uniform Refinement” with 641 \nGlobal CTF Refinement (Tilt, Trefoil, Spherical Aberration, Tetrafoil and Anisotropic Magnification), 642 \nand optimising per-particle defocus and scale with D2 symmetry was applied. 643 \nAnkyrin-1 and Band 3 Image Processing 644 \nEMPIAR-11043 was processed in CryoSPARC v4.7.0 (Supplementary Fig 14). 14,926 movies were 645 \nimported, motion-corrected and CTF estimated. Particles were picked with PartiNet on all 14,926 646 \nmicrographs denoised with the integrated denoiser. 3,614,613 particle coordinates were imported into 647 \nCryoSPARC, and particles were extracted with a box size of 600 pixels, Fourier downsampled to 150 648 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\npixels. Particle stacks were cleaned by 2 rounds of 2D classification (each with 100 classes), with classes 649 \nselected showing clear averages of both species. Another 2 rounds of 2D classification with 50 classes 650 \nwere generated, with classes selected showing the Ankryin-1 complex in micelle, resulting in 190,334 651 \nparticles. Ankyrin-1 particles were then re-extracted with a box size of 600 pixels with aligned shifts from 652 \n2D. The full resolution particles were initialised with single class “ab initio Reconstruction” and then 653 \n“Homogenous Refinement” in C1. Finally, “Non-Uniform Refinement” was applied with global CTF 654 \ncorrection (Tilt and Trefoil) and minimising over per-particle defocus and scale. 655 \nIn a parallel processing pathway, 2 rounds of 2D classification with 50 classes were done and classes 656 \nshowing Band 3 protein were identified, resulting in 177,900 particles. Band 3 particles were then re-657 \nextracted with a box size of 320 pixels with aligned shifts from 2D. The full resolution particles were 658 \ninitialised with single class “ab initio Reconstruction” and then “Homogenous Refinement” in C2. To 659 \ncomplete processing of Band 3, “Non-Uniform Refinement” was applied with global CTF correction 660 \n(Tilt, Trefoil, Spherical Aberration, Tetrafoil, Anisotropic Magnification) and minimising over per-661 \nparticle defocus and scale with C2 symmetry. 662 \nMlaCD Image Processing 663 \nEMPIAR-12531 was processed in CryoSPARC v4.7.0 (Supplementary Fig 14). 9,046 movies were 664 \nimported, motion-corrected and CTF estimated. Particles were picked with PartiNet on all 9,046 665 \nmicrographs denoised with the integrated denoiser. 1,802,428 particle coordinates were imported into 666 \nCryoSPARC, and particles were extracted with a box size of 350 pixels and Fourier downsampled to 144 667 \npixels. Two rounds of 2D classification with 150 classes were used to filter junk particles, resulting in 668 \n1,535,431 particles. These particles were initialised with “ab initio Reconstruction” with 4 classes. 669 \nTheclasses were then heterogeneously refined. The two classes representing the 1:6 and 2:6 were selected 670 \nfor “Non-Uniform Refinement” with C1 and C2 symmetry applied, respectively. The “Heterogeneous 671 \nRefinement” and “Non-Uniform Refinement” step was repeated 10 times, to effectively filter particles in 672 \n3D. The particles corresponding to each 3D class of 1:6 and 2:6 species were then extracted 673 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\nindependently at 350 pixels, with updated 2D and 3D aligned shifts. For the 1:6 species, particles were 674 \nreconstructed with alignments and then “Non-Uniform Refinement” was applied with C1 symmetry, 675 \nminimising over particle scale. The particles were then corrected for global CTF aberrations (Tilt, Trefoil, 676 \nTetrafoil, Spherical Aberration and Anisotropic Magnification) over 3 iterations. The CTF corrected 677 \nparticles were then used for another “Non-Uniform Refinement” with C1 symmetry. Because some 678 \nparticles were stacked on top of each other in 2D projections, spurious density was observed outside the 679 \nmain refinement of the complex, affecting particle alignment and FSC. A mask was generated with 680 \n“Volume Tools”, with a lowpass filter = 10 Å, threshold = 0.104, dilation radius = 20 pixels, soft-padding 681 \n= 25 pixels. This mask was used for a final “Local Refinement”. 682 \nIn a parallel processing pathway, the 2:6 species particles were reconstructed with alignments as above 683 \nand then “Non-Uniform Refinement” was applied with C2 symmetry, minimising over particle scale. The 684 \nparticles were then corrected for global CTF aberrations (Tilt, Trefoil, Tetrafoil, Spherical Aberration and 685 \nAnisotropic Magnification) over 2 iterations. The CTF-corrected particles were then used for another 686 \n“Non-Uniform Refinement” with C2 symmetry and optimising per-particle defocus. Again, the spurious 687 \ndensity from suboptimal particle stacking was observed. A mask was generated with “Volume Tools”, 688 \nwith a lowpass filter = 10 Å, threshold = 0.121, dilation radius = 20 pixels, soft-padding = 24 pixels. This 689 \nmask was used for a final “Local Refinement” with C2 symmetry and optimising per-particle defocus. 690 \nMap visualisation and data plotting 691 \nVisualisation of EM maps and atomic coordinates for analysis and figures were done in UCSF ChimeraX 692 \nv1.1048. Charts for analysis and figures were generated in RStudio v 2025.09.0 with R v4.5.1. All figure 693 \nlayouts and exports were done in Affinity Designer v2.6.0  694 \n 695 \n 696 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\nData availability 697 \nCryoEM maps have been deposited in the EM Data Bank with the following accession codes: EMD-698 \n68600 (MORC2H3K9-bound). Atomic coordinates have been deposited in the Protein Data Bank with the 699 \naccession codes 22QM (MORC2H3K9-bound). The raw micrographs, particle coordinates from PartiNet, 700 \nTopaz and crYOLO for MORC2H3K9-bound have been deposited in EMPIAR the following accession codes 701 \n(EMPIAR-13226). All other cryo-EM maps for other datasets are accessible from  702 \nhttps://doi.org/10.57967/hf/7618. 703 \nCode availability 704 \nSource code for PartiNet is publicly available on GitHub at https://github.com/WEHI-705 \nResearchComputing/PartiNet. PartiNet is licensed under the MIT License. The model weights can be 706 \nfound here: https://huggingface.co/MihinP/PartiNet. The documentation can be found here: https://wehi-707 \nresearchcomputing.github.io/PartiNet/.  708 \nAcknowledgements 709 \nWe acknowledge use of transmission electron microscopes at the Monash University Ramaciotti Centre 710 \nfor Cryo-Electron Microscopy and Ian Homes Imaging Centre, Bio21. We thank the WEHI Cryo-EM 711 \nFacility, the WEHI Research Computing Platform and Milton high-performance computing facility, the 712 \nUniversity of Melbourne Spartan high-performance computing facility and the Monash University 713 \nMASSIVE high-performance computing facility for providing facilities and support. We thank Nicholas 714 \nKirk, Alisa Glukhova and Peter Czabotar for their comments on the manuscript. This work was initially 715 \nsupported by WEHI’s New Medicines and Advance Technology funds to AL and SS. MP is supported by 716 \nResearch Training Program (RTP) Scholarship from Faculty of Engineering and Technology, University 717 \nof Melbourne and Graeme Clark Institute for Medical Engineering Top-Up scholarship. WT is supported 718 \nby an NHMRC Investigator Grant (GNT 2026635). JDB received support from an Australian Research 719 \nCouncil Future Fellowship (FT220100319) funded by the Australian Government. AL is supported in part 720 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\nby funds from the estate of Akos and Marjorie Talon. SS is supported by funds from WEHI, the estate of 721 \nAkos and Marjorie Talon, The University of Melbourne Attraction and Retention Funds, the NHMRC 722 \nInvestigator grant (GNT 2016827), the Australian Research Council Discovery Project grant 723 \n(DP250100450), the US Department of Defence Rare Cancer Research Concept Award (HT9425-24-1-724 \n0922) and the US Department of Defence Lung Cancer Research Program Award (HT94252510699). 725 \nAuthor contributions 726 \nMP, OJ, MA wrote the initial scripts for PartiNet, with MP developing the final version. WT purified 727 \nMORC2, prepared cryoEM samples and performed SPR experiments. HV collected cryoEM data. EY 728 \ncontributed to program parallelisation, debugging, and packaging of PartiNet. JI prepared training data. 729 \nMP and SS conceived the project. JI, JDB, AL and SS supervised the project. AL and SS acquired the 730 \nfunding. MP, JDB, AL and SS analysed the data and wrote the manuscript with contributions from all 731 \nauthors. 732 \nCompeting interests 733 \nThe authors declare no competing interests 734 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\nMain figures 735 \n 736 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\nFigure 1. PartiNet has a dynamic architecture 737 \na. PartiNet employs a dynamic architecture with dual detectors and an adaptive router. Micrographs are 738 \ndenoised and processed through Detector 1's backbone to generate feature maps. The adaptive router 739 \nassigns difficulty scores to these feature maps, directing easy micrographs back through Detector 1's neck 740 \nand head, while difficult micrographs are concatenated with the original image and processed through 741 \nDetector 2. A post-processing module removes duplicate picks and converts YOLO coordinates to STAR 742 \nformat. b. The adaptive router differentiates micrographs based on imaging conditions without 743 \nsupervision. (I-II) Two micrographs from EMPIAR-10017 showing PartiNet's classification of the 744 \nmicrograph containing support film (yellow arrow) as more difficult than the one without support film. 745 \nFringing artifacts from aggressive motion correction are visible at image edges. Scale bar, 100 nm. (III-746 \nIV) Two micrographs from EMPIAR-10089 demonstrating PartiNet's identification of the lower defocus 747 \nmicrograph as more difficult due to reduced (phase) contrast. The difficult micrograph is at -1.07 µm 748 \ndefocus versus -2.07 µm for the easy micrograph. Scale bar, 120 nm. 749 \n 750 \n 751 \n 752 \n 753 \n 754 \n 755 \n 756 \n 757 \n 758 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\n 759 \nFigure 2. PartiNet outperforms popular particle pickers on small datasets 760 \nSeven datasets from CryoPPP were randomly selected to benchmark PartiNet against crYOLO and 761 \nTopaz. The table summarises key dataset characteristics including EMPIAR ID, molecular weight, 762 \nmicrograph count, defocus range, and final map resolution, with the highest resolution result underlined 763 \nfor each protein. The accompanying heatmap shows the number of protein particles retained after 2D 764 \nclassification (\"Select 2D\") for each method, with darker blue indicating higher particle numbers used for 765 \nreconstruction. 766 \n 767 \n 768 \n 769 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\n 770 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\nFigure 3. PartiNet outperforms other particle pickers on EMPIAR-10089 and EMPIAR-10215 771 \na. Bar chart comparing picking speed of PartiNet (black) and crYOLO (red) on TcdA1 (EMPIAR-10089) 772 \nand rabbit muscle aldolase (EMPIAR-10215), reported as micrographs per second. Approximate X-fold 773 \nincrease in speed with PartiNet is indicated for each dataset. b. Selected 2D class averages of TcdA1 774 \nparticles picked with PartiNet. c. Cryo-EM map reconstruction of TcdA1 from PartiNet particles picked 775 \non EMPIAR-10089. d. FSC curves for final reconstructions of TcdA1 from particles picked with 776 \ncrYOLO (teal), Topaz (purple) and PartiNet (yellow) with resolution annotated at FSC = 0.143 cutoff. e. 777 \nBar chart comparing numbers of particles picked (pink) and used for final reconstruction (purple) of 778 \nTcdA1 for PartiNet, crYOLO, and Topaz. f. Selected 2D class averages of rabbit muscle aldolase 779 \nparticles picked with PartiNet. c. Cryo-EM map reconstruction of rabbit muscle aldolase from PartiNet 780 \nparticles picked on EMPIAR-10215. d. FSC curves for final reconstructions of rabbit muscle aldolase 781 \nfrom particles picked with crYOLO (teal), Topaz (purple) and PartiNet (yellow) with resolution annotated 782 \nat FSC = 0.143 cutoff. e. Bar chart comparing numbers of particles picked (dark teal) and used for final 783 \nreconstruction (light teal) of rabbit muscle aldolase for PartiNet, crYOLO, and Topaz. 784 \n 785 \n 786 \n 787 \n 788 \n 789 \n 790 \n 791 \n 792 \n 793 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\n 794 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\nFigure 4. PartiNet picks allow high resolution reconstructions for model building 795 \na. Bar chart comparing picking speed of PartiNet (black) and crYOLO (red) on MORC2. Approximate 5-796 \nfold increase in speed with PartiNet is indicated. b. Selected 2D class averages of MORC2 particles 797 \npicked with PartiNet. c. FSC curves for final reconstructions of MORC2 from particles picked with 798 \ncrYOLO (teal), Topaz (purple) and PartiNet (yellow) with resolution annotated at FSC = 0.143 cutoff d. 799 \nMORC2 maps were reconstructed using coordinates from PartiNet, Topaz, and crYOLO particle picking 800 \nalgorithms. Local resolution was estimated in CryoSPARC and visualised on the reconstructions in 801 \nChimeraX (v.1.10.1), where teal voxels indicate lower resolution and maroon indicates higher resolution. 802 \nA central cross-section of each map is shown. The global resolution, final particle count, conical FSC area 803 \nratio (cFAR), and Sampling Compensation Factor (SCF) are shown for each map. e. The crystal structure 804 \nof MORC2 (PDB 5OF9) was superimposed with the ModelAngelo-generated model from the PartiNet 805 \nmap after alignment in ChimeraX. The blue boxed region is shown in detail in f. highlighting differences 806 \nin modeling of previously unresolved residues between the crystal structure and ModelAngelo prediction 807 \nof MORC2. 808 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\n 809 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\nFigure 5. PartiNet can identify multiple species in a dataset 810 \na. Bar chart comparing picking speed of PartiNet (black) and crYOLO (red) on EMPIAR-11043. 811 \nApproximate 2-fold increase in speed with PartiNet is annotated b. Simplified workflow for processing 812 \nEMPIAR-11043 to reconstruct Ankyrin-1 and Band 3 proteins. The published workflow required multiple 813 \nrounds of particle picking (manual, template-based, and Topaz with trained models) to identify multiple 814 \nspecies, whereas PartiNet identifies heterogeneous proteins in a single step. Full workflow details are 815 \navailable in Supplementary Fig 14 c. Bivariate histogram showing the relationship between box size and 816 \nconfidence for PartiNet picks on EMPIAR-11043. Particle counts are hexagonally binned with linear box 817 \nsize and logarithmic confidence scaling. Bin color intensity (lighter blue = higher counts) is scaled 818 \nlogarithmically. Only picks with > 10% confidence and > 300 pixel box size are shown to exclude low-819 \nquality, noisy detections. Red and yellow contours show bivariate kernel density peaks, with 820 \ncorresponding marginal density plots (right) revealing two distinct box size populations at 620 and 403 821 \npixels, corresponding to Ankyrin-1 complex and Band 3 proteins, respectively. d-e. Representative 2D 822 \nclass averages and consensus refinement maps for (d) Ankyrin-1 complex and (e) Free Band 3 proteins 823 \nidentified by PartiNet. Final particle counts for each reconstruction are indicated. f. FSC curves for final 824 \nreconstructions of Ankyrin-1 (yellow) and Band 3 (pink) proteins with resolution annotated at FSC = 825 \n0.143 cutoff. 826 \n 827 \n 828 \n 829 \n 830 \n 831 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\n 832 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\nFigure 6. PartiNet improves the resolution of recently published maps 833 \na. Bar chart comparing picking speed of PartiNet (black) and crYOLO (red) on EMPIAR-12531. 834 \nApproximate 2-fold increase in speed with PartiNet is indicated. b. Bar chart comparing number of 835 \nparticles used for final reconstruction of the MlaCD complex (both 1:6 and 2:6 stoichiometry species) 836 \nbetween published workflow (light blue) and with PartiNet picking (dark blue). c. FSC curves for final 837 \nreconstructions of 1:6 stoichiometry (light blue) and 2:6 stoichiometry (dark blue) species of MlaCD with 838 \nresolution annotated at FSC = 0.143 cutoff. d. Representative 2D class averages of 1:6 stoichiometry 839 \nMlaCD from particles picked with PartiNet. e. Cryo-EM map reconstruction of 1:6 stoichiometry MlaCD 840 \nfrom PartiNet particles. f. 2D heatmap of particle distributions for 1:6 stoichiometry reconstruction 841 \n(yellow indicating high particle counts and dark blue low particle counts). g. Representative 2D class 842 \naverages of 2:6 stoichiometry MlaCD from particles picked with PartiNet. h. Cryo-EM map 843 \nreconstruction of 2:6 stoichiometry MlaCD from PartiNet particles. i. 2D heatmap of particle distributions 844 \nfor 2:6 stoichiometry reconstruction (yellow indicating high particle counts; dark blue indicating low 845 \nparticle counts). 846 \n 847 \n 848 \n 849 \n 850 \n 851 \n 852 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\nReferences 853 \n1. Baker, T. S., Olson, N. H. & Fuller, S. D. Adding the Third Dimension to Virus Life Cycles: Three-854 \nDimensional Reconstruction of Icosahedral Viruses from Cryo-Electron Micrographs. Microbiology 855 \nand Molecular Biology Reviews 63, 862–922 (1999). 856 \n2. Thompson, R. F., Iadanza, M. G., Hesketh, E. L., Rawson, S. & Ranson, N. A. Collection, pre-857 \nprocessing and on-the-fly analysis of data for high-resolution, single-particle cryo-electron 858 \nmicroscopy. Nat Protoc 14, 100–118 (2019). 859 \n3. Scheres, S. H. W. RELION: Implementation of a Bayesian approach to cryo-EM structure 860 \ndetermination. Journal of Structural Biology 180, 519–530 (2012). 861 \n4. Punjani, A., Rubinstein, J. L., Fleet, D. J. & Brubaker, M. A. cryoSPARC: algorithms for rapid 862 \nunsupervised cryo-EM structure determination. Nat Methods 14, 290–296 (2017). 863 \n5. Gyawali, R., Dhakal, A., Wang, L. & Cheng, J. Accurate cryo-EM protein particle picking by 864 \nintegrating the foundational AI image segmentation model and specialized U-Net. 865 \n2023.10.02.560572 Preprint at https://doi.org/10.1101/2023.10.02.560572 (2024). 866 \n6. Gyawali, R., Dhakal, A., Wang, L. & Cheng, J. CryoSegNet: accurate cryo-EM protein particle 867 \npicking by integrating the foundational AI image segmentation model and attention-gated U-Net. 868 \nBriefings in Bioinformatics 25, bbae282 (2024). 869 \n7. Roseman, A. M. Particle finding in electron micrographs using a fast local correlation algorithm. 870 \nUltramicroscopy 94, 225–236 (2003). 871 \n8. Roseman, A. M. FindEM—a fast, efficient program for automatic selection of particles from electron 872 \nmicrographs. Journal of Structural Biology 145, 91–99 (2004). 873 \n9. Woolford, D., Hankamer, B. & Ericksson, G. The Laplacian of Gaussian and arbitrary z-crossings 874 \napproach applied to automated single particle reconstruction. Journal of Structural Biology 159, 122–875 \n134 (2007). 876 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\n10. Voss, N. R., Yoshioka, C. K., Radermacher, M., Potter, C. S. & Carragher, B. DoG Picker and 877 \nTiltPicker: Software tools to facilitate particle selection in single particle electron microscopy. 878 \nJournal of Structural Biology 166, 205–213 (2009). 879 \n11. Zivanov, J. et al. New tools for automated high-resolution cryo-EM structure determination in 880 \nRELION-3. eLife 7, e42166 (2018). 881 \n12. Wang, F. et al. DeepPicker: A deep learning approach for fully automated particle picking in cryo-882 \nEM. Journal of Structural Biology 195, 325–336 (2016). 883 \n13. Heimowitz, A., Andén, J. & Singer, A. APPLE picker: Automatic particle picking, a low-effort cryo-884 \nEM framework. Journal of Structural Biology 204, 215–227 (2018). 885 \n14. Bepler, T. et al. Positive-unlabeled convolutional neural networks for particle picking in cryo-886 \nelectron micrographs. Nat Methods 16, 1153–1160 (2019). 887 \n15. Wagner, T. et al. SPHIRE-crYOLO is a fast and accurate fully automated particle picker for cryo-888 \nEM. Commun Biol 2, 1–13 (2019). 889 \n16. Zhang, J. et al. PIXER: an automated particle-selection method based on segmentation using a deep 890 \nneural network. BMC Bioinformatics 20, 41 (2019). 891 \n17. George, B. et al. CASSPER is a semantic segmentation-based particle picking algorithm for single-892 \nparticle cryo-electron microscopy. Commun Biol 4, 1–12 (2021). 893 \n18. Dhakal, A., Gyawali, R., Wang, L. & Cheng, J. CryoTransformer: a transformer model for picking 894 \nprotein particles from cryo-EM micrographs. Bioinformatics 40, btae109 (2024). 895 \n19. Ronneberger, O., Fischer, P. & Brox, T. U-Net: Convolutional Networks for Biomedical Image 896 \nSegmentation. Preprint at https://doi.org/10.48550/arXiv.1505.04597 (2015). 897 \n20. Wang, C.-Y., Bochkovskiy, A. & Liao, H.-Y. M. YOLOv7: Trainable bag-of-freebies sets new state-898 \nof-the-art for real-time object detectors. Preprint at https://doi.org/10.48550/arXiv.2207.02696 899 \n(2022). 900 \n21. Redmon, J. & Farhadi, A. YOLO9000: Better, Faster, Stronger. in 2017 IEEE Conference on 901 \nComputer Vision and Pattern Recognition (CVPR) 6517–6525 (2017). doi:10.1109/CVPR.2017.690. 902 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\n22. Lin, T.-Y. et al. Feature Pyramid Networks for Object Detection. in 2017 IEEE Conference on 903 \nComputer Vision and Pattern Recognition (CVPR) 936–944 (2017). doi:10.1109/CVPR.2017.106. 904 \n23. Liang, T. et al. CBNet: A Composite Backbone Network Architecture for Object Detection. IEEE 905 \nTransactions on Image Processing 31, 6893–6906 (2022). 906 \n24. Lin, Z., Wang, Y., Zhang, J. & Chu, X. DynamicDet: A Unified Dynamic Architecture for Object 907 \nDetection. Preprint at https://doi.org/10.48550/arXiv.2304.05552 (2023). 908 \n25. Dhakal, A., Gyawali, R., Wang, L. & Cheng, J. Artificial Intelligence in Cryo-EM Protein Particle 909 \nPicking: The Hope, Hype, and Hurdles. Preprint at https://doi.org/10.20944/preprints202408.1936.v1 910 \n(2024). 911 \n26. Dhakal, A., Gyawali, R., Wang, L. & Cheng, J. A large expert-curated cryo-EM image dataset for 912 \nmachine learning protein particle picking. Sci Data 10, 392 (2023). 913 \n27. van Rossum, G. Python tutorial. (1995). 914 \n28. Ansel, J. et al. PyTorch 2: Faster Machine Learning Through Dynamic Python Bytecode 915 \nTransformation and Graph Compilation. in Proceedings of the 29th ACM International Conference 916 \non Architectural Support for Programming Languages and Operating Systems, Volume 2 vol. 2 929–917 \n947 (Association for Computing Machinery, New York, NY, USA, 2024). 918 \n29. Ashwin Dhakal, Rajan Gyawali, Liguo Wang, & Jianlin Cheng. CryoPPP: A Large Expert-Labelled 919 \nCryo-EM Image Dataset for Machine Learning Protein Particle Picking. 920 \nhttps://doi.org/10.1101/2023.02.21.529443 (2023) doi:10.1101/2023.02.21.529443. 921 \n30. Vallese, F. et al. Architecture of the human erythrocyte ankyrin-1 complex. Nat Struct Mol Biol 29, 922 \n706–718 (2022). 923 \n31. Wotherspoon, P. et al. Structure of the MlaC-MlaD complex reveals molecular basis of periplasmic 924 \nphospholipid transport. Nat Commun 15, 6394 (2024). 925 \n32. Wagner, T. & Raunser, S. The evolution of SPHIRE-crYOLO particle picking and its application in 926 \nautomated cryo-EM processing workflows. Communications Biology 3, 61 (2020). 927 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\n33. Bepler, T., Kelley, K., Noble, A. J. & Berger, B. Topaz-Denoise: general deep denoising models for 928 \ncryoEM and cryoET. Nat Commun 11, 5208 (2020). 929 \n34. Lehtinen, J. et al. Noise2Noise: Learning Image Restoration without Clean Data. Preprint at 930 \nhttps://doi.org/10.48550/arXiv.1803.04189 (2018). 931 \n35. Shorten, C. & Khoshgoftaar, T. M. A survey on Image Data Augmentation for Deep Learning. 932 \nJournal of Big Data 6, 60 (2019). 933 \n36. Lin, T.-Y. et al. Microsoft COCO: Common Objects in Context. Preprint at 934 \nhttps://doi.org/10.48550/arXiv.1405.0312 (2015). 935 \n37. Jamali, K. et al. Automated model building and protein identification in cryo-EM maps. Nature 628, 936 \n450–457 (2024). 937 \n38. Tan, W. et al. MORC2 is a phosphorylation-dependent DNA compaction machine. Nat Commun 16, 938 \n5606 (2025). 939 \n39. Tan, Y. Z. et al. Addressing preferred specimen orientation in single-particle cryo-EM through 940 \ntilting. Nature Methods 14, 793–796 (2017). 941 \n40. Baldwin, P. R. & Lyumkis, D. Non-uniformity of projection distributions attenuates resolution in 942 \nCryo-EM. Progress in Biophysics and Molecular Biology 150, 160–183 (2020). 943 \n41. Douse, C. H. et al. Neuropathic MORC2 mutations perturb GHKL ATPase dimerization dynamics 944 \nand epigenetic silencing by multiple structural mechanisms. Nature Communications 9, 651 (2018). 945 \n42. Tan, W. et al. MORC2 phosphorylation fine tunes its DNA compaction activity. 2024.06.27.600912 946 \nPreprint at https://doi.org/10.1101/2024.06.27.600912 (2024). 947 \n43. Tegunov, D. & Cramer, P. Real-time cryo-electron microscopy data preprocessing with Warp. Nat 948 \nMethods 16, 1146–1152 (2019). 949 \n44. Punjani, A. Algorithmic advances in single particle cryo-EM data processing using CryoSPARC. 950 \nMicroscopy and Microanalysis 26, 2322–2323 (2020). 951 \n45. Wang, A. et al. YOLOv10: Real-Time End-to-End Object Detection. arXiv.org 952 \nhttps://arxiv.org/abs/2405.14458v1 (2024). 953 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint \n\n46. Carion, N. et al. End-to-End Object Detection with Transformers. Preprint at 954 \nhttps://doi.org/10.48550/arXiv.2005.12872 (2020). 955 \n47. Pintilie, G. et al. Measurement of atom resolvability in cryo-EM maps with Q-scores. Nat Methods 956 \n17, 328–334 (2020). 957 \n48. Goddard, T. D. et al. UCSF ChimeraX: Meeting modern challenges in visualization and analysis. 958 \nProtein Sci 27, 14–25 (2018). 959 \n 960 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted January 23, 2026. ; https://doi.org/10.64898/2026.01.23.700950doi: bioRxiv preprint","source_license":"CC-BY-4.0","license_restricted":false}