Template Learning: Deep Learning with Domain Randomization for Particle Picking in Cryo-Electron Tomography

10 Cryo-electron tomography (cryo -ET) enables the three -dimensional visua lization of 11 biomolecules and cellular components in their near-native state. Particle picking, a crucial 12 step in cryo-ET data analysis, is traditionally performed by template matching—a method 13 utilizing cross -correlations with available biomolecular templa tes. Despite the 14 effectiveness of recent deep learning -based particle picking approaches, their 15 dependence on initial data annotation datasets for supervised training remains a 16 significant limitation. Here, we propose a technique that combines the accuracy of deep 17 learning particle identification with the convenience of the model training on biomolecular 18 templates enabled through a tailored domain randomization approach. Our technique, 19 named Template Learning, automates the simulation of training datasets, incorporating 20 considerations for molecular crowding, structural variabilities, and data acquisition 21 variations. This reduces or even eliminates the dependence of supervised deep learning 22 on annotated experimental datasets. We demonstrate that models traine d on simulated 23 datasets, optionally fine -tuned on experimental datasets, outperform those exclusively 24 trained on experimental datasets. Also, we illustrate that Template Learning used as an 25 alternative to template matching, can offer higher precision and b etter orientational 26 isotropy, especially for picking small non-spherical particles. Template Learning software 27 is open-source, Python-based, and GPU and CPU parallelized. 28 29

30 Cryo-ET; Particle Picking; Deep Learning; Domain Randomization; Physics-based 31 Simulations; Cell Crowding; Molecular Mechanics 32 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 2

33 Recent technological advances in cryo -electron tomography (cryo -ET) expanded our 34 capabilities to explore the three-dimensional (3D) cell architecture at the molecular scale. 35 On one side, the automation of vitreous cryo-lamellae milling significantly enhanced the 36 throughput of tomographic sample preparation 1–3. On the other side, improvements in 37 the hardware of transmission electron microscopes (TEM), especially new generations of 38 direct electron detectors, enabled the routine acquisition of tomograms with the resolution 39 that allows identification of biomolecules and determination of their structure and its 40 variations, directly in their functional environment 4–8. 41 The cryo-ET pipeline begins with the vitrification of the sample by plunge or high-pressure 42 freezing. In the case of bulky samples such as eukary otic cells and tissues, this step is 43 followed by thinning using cryo -focused ion beam milling or cryo -sectioning. A series of 44 2D images (micrographs) is then obtained by rotating the sample in cryo -TEM. The 45 resulting images, referred to as a tilt series, a re computationally reconstructed into a 3D 46 volume called a tomogram 9,10. 47 Although cryo-ET allows the acquisition of high -resolution data from cellular samples, it 48 faces serious limitations. Existing sample preparation and cryo-TEM setups limit the tilting 49 range, typically within ±60 degrees. This angular limitation leads to a missing wedge 50 region in Fourier space resulting in severe real -space data anisotropies commonly 51 referred to as missing wedg e artifacts 11–14. In addition, the sensitivity of biological 52 samples to radiation damage requires image acquisition with a limited electron dose, 53 resulting in a compromised signal-to-noise ratio (SNR). 54 Locating copies of biomolecules in cryo -electron tomograms, known as particle picking, 55 is important for understanding biomolecular distribution in cells and constitutes an initial 56 stage in obtaining higher resolution 3D structures via aver aging and classification 15,16. 57 However, manual particle picking is labor-intensive, further complicated by the challenges 58 posed by the missing wedge and low SNR. 59 A well -established method to automate particle picking using the similarity with an 60 external template via cross-correlation is called template matching 17,18. In recent years, 61 several d eep learning-based methods have emerged and were shown capable of 62 outperforming template matching, for some studies, in speed and accuracy 19–24. 63 A specialized track of the SHape REtrieval Contest (SHREC) was launched to compare 64 various deep learning methods for cryo-ET particle picking 25. These evaluations were 65 conducted on simulated datasets, where tomograms with ground truth annota tions are 66 available and are used for training. However, on experimental datasets, supervised deep 67 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 3 learning methods will require a training dataset, usually obtained via manual template 68 matching-assisted picking. 69 Open databases for 3D structural data of biomolecules, e.g., the Protein Data Bank (PDB) 70 26, give access to structures that can serve as templates for template matching. However, 71 despite the remarkable advancements in the physical modeling of cr yo-ET data 27–30, 72 templates were not used for supervised training of deep learning models on customized 73 particle picking 20,25,31. Some methods found a use for simulated datasets in training deep 74 learning-based general -purpose particle picking (e.g. cryoYOLO 32) or unsupervised 75 structural mining (e.g., TomoTwin 31). However, these methods face several limitations, 76 including molecular crowding, structural variations, or computational complexity 22,23,31,33. 77 Using simulations derived from previously defined or inferred structures, such as those 78 obtained from in vitro or in silico studies, to train models to annotate similar structures in 79 cryo-electron tomograms would be highly advantageous over the tim e-consuming initial 80 annotation directly from in situ data. This raises the question of why such methodologies 81 are not more widespread. The literature points to a challenge known as the "synthetic-to-82 real domain gap" 34 where deep learning models trained on simulations can only operate 83 in the synthetic domain. This domain is characterized by attributes like texture, 84 illumination, noise, background, and other factors, which are notably different from those 85 in the real-world domain 34. 86 Domain randomization is a state -of-the-art approach for addressing the domain gap 87 during synthetic dataset generation, specifically in the context of training models for object 88 classification and segmentation. Domain randomization is based on simulating no n-89 realistic scenarios that were proven powerful to supervised deep learning of models 90 capable of generalizing to real -world data 35,36. It hinges on three principles: 1) utilizing 91 models of target objects with randomized shape and pose variations; 2) incorporating a 92 mix of diverse objects termed “distractors”, positioned in the background near the target 93 objects; and 3) integrating a broad spectrum of random rendering options. 94 In cryo -ET literature , addressing the synthetic -to-real domain gap for deep learning 95 training has been largely overlooked. Nevertheless, a method called CryoShift 37 has 96 utilized domain adaptation and randomization techniques for subtomogram classification, 97 primarily focusing on mapping the synthetic data domain to resemble the real data during 98 classifier training. Yet, this method was limited to handling pre-isolated subtomograms of 99 molecules of interest and was not design ed to process entire tomograms for particle 100 pickling or segmentation. 101 Here, we introduce “Template Learning”, a domain randomization -based strategy for 102 generating cryo -ET simulated tomograms, incorporating considerations for molecular 103 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 4 crowding, structural variabilities, and data acquisition variations, with which we can learn 104 deep learning models to achieve state -of-the-art performance in particle picking within 105 cryo-ET experimental tomograms. In this paper, we detail our approach and 106 systematically benchmark it against alternative methods, utilizing a recently published, 107 exhaustively annotated in situ cryo-ET dataset. We also apply Template Learning to 108 nucleosome picking within a densely populated cryo -ET sample of isolated mitotic 109 chromosomes. We illustrat e that models trained on simulated datasets, optionally fine -110 tuned on experimental datasets, outperform those exclusively trained on experimental 111 datasets. Also, we illustrate that Template Learning used as an alternative to template 112 matching, can offer hi gher precision and better orientational isotropy, especially for 113 picking small non-spherical particles. 114

115 Customized simulations to train deep learning models on particle 116 picking 117 Template Learning is a streamlined pipeline for simulating cryo -ET synt hetic data for 118 training deep learning models on particle picking. The fundamental goal of Template 119 Learning is to train deep learning models in three aspects: 120 1. Identifying targeted particles across diverse variations. 121 2. Differentiating target particles from other structures, especially in crowded 122 environments. 123 3. Expanding the capabilities of the models to handle the inherent variations present 124 in cryo-ET experimental data. 125 The practical application of these aspects is achi eved through our proposed pipeline, 126 illustrated in Fig.1 and further elaborated in the subsequent sections. 127 Accounting for the structural and orientational variabilities 128 To enable trained deep learning models to effectively identify various variations of t he 129 target biomolecule, we utilize two following strategies: 130 1. We use multiple templates of one target biomolecule to account for compositional 131 or significant conformational variabilities of the target particles. 132 2. We generate random flexible variations of eac h template using Normal Mode 133 Analysis (NMA—a method for fast molecular mechanics simulations, see Methods 134 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 5 for details) 38. The templates with their flexible variations are then simulated at 135 uniformly random orientations to account for orientational variations. 136 Templates in this study are typically atomic structures. However, since the availability of 137 atomic structures may vary for different target biomolecules, using lower resolution cryo-138 EM density maps (volumes) is also possible and is explored in a subsequent section. 139 Accounting for the crowded environment 140 The existing literature on domain randomization for object recognition in natural images 141 simulates the target objects within a background containin g a library of other random 142 objects 39. These additional objects, distractors, obscure or cast shadows on segments 143 of the target objects. In this study, we selected 100 dissimilar protein assemblies, w ith 144 diverse molecular weights ranging from 30 kDa to 1 MDa to serve as elements within the 145

(a comprehensive list of these proteins is available in Extended Data Fig.1). 146 To achieve a similar effect of distractors for cryo-ET data, i.e., overlapping at the level of 147 the tilt series and the missing wedge artifacts, these proteins must be positioned in close 148 proximity to the templates during data modeling. 149 Recent works 27,40 that involve simu lating cryo-ET data for practical applications (e.g., 150 deep learning) utilize an iterative brute -force random placement algorithm, involving the 151 rotation of a duplicate of a molecule in each iteration to find a suitable non -overlapping 152 position within the sample. Nonetheless, random placement leads to unstructured empty 153 spaces between the molecules that prevent achieving highly dense samples. 154 Two other approaches simulate molecular crowding of cryo -ET data 41,42. The first 155 represents molecules as spheres, based on calculating the minimum bounding sphere 156 for each molecule, and simulates crowding by optimizing a sphere -packing problem. 157 However, this approach is limited to generating high crowding only for spherical-shaped 158 molecules. The second builds on the first by including an additional step of molecular 159 dynamics simulations to enhance packing density. However, routine application of this 160 approach requires significant computational resources. 161 To achieve high molecular crowding while maintaining generality and efficiency, we drew 162 inspiration for a simplified approach from a recent concept on packing generic 3D objects 163 43, a solution we term the "Te tris algorithm." Our algorithm, illustrated in Fig.2, operates 164 on an intuitive principle: it places molecules iteratively, with each iteration positioning a 165 new molecule at a uniform random orientation as close as possible to those already 166 placed (the user only needs to choose the minimum distance). For more details on the 167 principles of the Tetris algorithm, see Methods. 168 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 6 Accounting for the real-world domain variance 169 Simulations can partially replicate the domain of real -world data; however, variations in 170 the real world can be unpredictable, and simulations simplify some complex phenomena 171 for practical reasons. Thanks to domain randomization 34,36, simulating an exact match of 172 the real world is not necessary for training deep learning models capable of generalizing 173 to real-world data. Instead, domain randomization focuses on training these models on 174 diverse simulated scenarios to minimize the influence of the domain on their capabilities. 175 In this work, we employ a cryo-ET physics-based simulator called Parakeet 28. Parakeet, 176 and its MULTEM backend 44, provide control over numerous parameters for simulating 177 cryo-ET hardware and the behavior of biological samples during data recording. While 178 many parameters related to the electron beam, lenses, detector, sa mple, and data 179 acquisition strategy can be controlled, it might not be necessary to randomize every 180 available parameter. On one hand, the combinations required to cover all conditions grow 181 exponentially with the number of variables. On the other hand, some variables might have 182 undesirable effects on the overall simulation. Here, we propose varying a few essential 183 parameters that significantly influence the output simulation: the electron dose, defocus, 184 tilting range, tilting step, and ice density, while giv ing other parameters default values. 185 The electron dose is crucial for varying the SNR and the irradiation effects on the sample. 186 The defocus plays a key role in controlling the Contrast Transfer Function (CTF) 45. 187 Varying the tilting range and step may not replicate experimental data acquisition 188 strategies, however, it can expose the deep learning model to variations that increase its 189 robustness, for example, to artifacts and imperfections resulting from tilt series alignment 190 and reconstruction. Lastly, different ice densities can simulate the variability of sample 191 thickness and solvent composition. A combination of three values for the defocus and two 192 values for each of the other variables results in 48 combinations that we found sufficient 193 for efficient domain randomization. 194 Benchmarking Template Learning for picking ribosomes in situ 195 In a recent study 20, the first exhaustively annotated in situ cryo-ET dataset was published 196 (EMPIAR-10988). Here, we use this dataset to evaluate the proposed Template Learning 197 thoroughly. 198 In the following, we generate, based on the Template Learning workflow, a simulated 199 dataset to train DeepFinder 19 for ribosome annotation. We compare the performance of 200 this model (i.e., DeepFinder trained solely on simulations) to the contemporary 201 techniques, namely, DeePiCt 20 and DeepFinder (trained solely on annotated 202 experimental data), and template matching. 203 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 7 We organized our experiments based on two principles. First, we benchmark the 204 complete method that integrates the aforementioned concepts of Template Learning. 205 Then, we conducted experiments where specific elements from the data simulation 206 pipeline were intentionally omitted (e.g., the structural variations of the target template, 207 the crowding, etc.) to assess their impact on performance and to offer guidelines for future 208 users. Second, acknowledging that the size of a simulated dataset is user -defined 209 (generally in deep learning, the more training data, and the more diversity, the better), we 210 ensured the feasibility of routine use of the method by limiting its time re quirement to a 211 maximum of two days on a single GPU (within one day on a typical computing node with 212 4 GPUs, see Software requirements and availability section for details). 213 The typical Template Learning workflow 214 To establish a Template Learning workflow, w e proceeded with the following steps. We 215 selected 6 eukaryotic ribosome PDB structures (PDB IDs: 4UG0, 4V6X, 6Y0G, 6Y2L, 216 6Z6M, and 7CPU). The 6 selected structures differ slightly in composition or 217 conformation. By applying NMA, we generated 25 flexible va riations for each structure 218 (refer to the NMA section in Methods for details). While generating synthetic data, we 219 employed the aforementioned general distractors (a full list of distractors is given in 220 Extended Data Fig.1). To simulate a crowded environme nt, we generated volumes from 221 the template structures (i.e., the 6 ribosome PDBs and their flexible variations) and 222 distractors with a sampling rate of 16 Å using Eman2 46 software e2pdb2mrc. We used 223 the generated volumes as input to populate 48 crowded volumetric distributions of size 224 192 x 192 x 64 voxel3 using the Tetris algorithm, with ribosome templates appearing at a 225 frequency of one for every 5 distractors to produce a balanced volumetric density ratio 226 balance between distractors and templates in the simulated tomograms (see Tetris 227 algorithm section in Methods for details). Subsequently, we fed the sample information 228 (i.e., atomic structures of templates and distractors with the positions and orie ntations of 229 crowding generated using the Tetris algorithm) to Parakeet 28 to simulate 48 tilt -series 230 using a dose symmetric tilting scheme using the parameters listed in Extended Data 231 Tab.1. All other parameters adhered to default values for Volta Phase Plates (VPP) 232 simulations within the Parakeet software. Subsequently, we binned the simulated tilt 233 series to a similar sampling rate as the analyzed dataset (13 Å/pixel) and reconstructed 234 them in IMOD 47 using Weighted Back Projection. We used the tomograms and their 235 corresponding ground -truth template coordinates (around 6,500 simulated ribosomes) 236 and segmentations (i.e., volumes of the same size as t he tomograms where template 237 instances appear in white over black background, illustrated in Extended Data Fig.2) to 238 train DeepFinder 19 using its default parameters settings (the list of parameters is given 239 in Extended Data Tab.2). 240 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 8 We applied the DeepFinder simulations -trained model to the 10 VPP tomograms of 241 EMPIAR-10988 dataset without any preprocessing, resulting in segmentation maps, i.e., 242 ribosome segmentations based on the decisions of the model. Subsequently, we 243 employed the MeanShift algorithm offered by the DeepFinder software, using a clustering 244 radius of 10 voxels, to extract both the coordinates and size of these segments (count of 245 the number of voxels of each segment), with and without the a pplication of a mask 246 (specifically, a cytosol mask sourced from the dataset, utilized to eliminate false positives 247 outside the mask). We compare the extracted coordinates to the expert -validated 248 annotations provided by the dataset at different levels of se gment sizes used as a 249 threshold (i.e., segments smaller than the threshold were removed). For this comparison, 250 we maintained the criterion that two annotations —one from the output of the deep 251 learning model and the other from the expert annotations—were considered to target the 252 same particle if their spatial distance was within 10 voxels (a value consistent with the 253 methodology of a previous study analyzed the same dataset 20). An example of a 254 tomogram, its corresponding segmentation map, and the coordinates set at a threshold 255 that removed smaller objects compared to expert annotations are presented in Fig.3. To 256 evaluate the performance of the method, we computed the overall Re call and Precision 257 with and without masking (refer to the Methods section for more details on the 258 assessment metrics). The point where Precision and Recall curves intersect was utilized 259 to determine the F 1 score per tomogram for the dataset (shown in Fig.4 A). The results, 260 further illustrated in Fig.4I -J, reveal that Template Learning used to train DeepFinder 261 exclusively on simulations, achieved state -of-the-art performance, outperforming 262 template matching and previous deep learning methods trained solely on annotations 263 derived from the same experimental dataset based on the median F1 score (compared to 264 values reported in 20). 265 To reflect on these results, it is essential to highlight that the previously reported findings 266 for DeePiCt and DeepFinder were based on models trained with 8 out of 10 fully 267 annotated tomograms, containing approximately 20,000 ribosome annotations 20. Hence, 268 previous evaluations were conducted on 2 out of 10 tomograms, for annotating the 269 remaining 5,000 ribosomes, utilizing three cross-validation splits (in each split, a random 270 set of 8 tomograms was used for training and the remaining 2 tomograms for 271 benchmarking). Additionally, each of the two models underwent a different training 272 strategy to achieve optimal performance. In particular, in the case of DeepFinder, the 273 training encompassed two classes —ribosomes and Fatty Acid Synthase (FAS) —as 274 training solely on ribosomes led to suboptimal performance. On the other hand, DeePiCt 275 was exclusively trained on ribosomes, as unlike DeepFinder, combining ribosomes and 276 FAS during training led to suboptimal performance. 277 In contrast, the DeepFinder model trained solely on the simulations gene rated from the 278 Template Learning workflow, not fine -tuned on any annotated experimental data, was 279 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 9 benchmarked on the complete dataset (i.e., all 10 tomograms). Hence, despite the 280 relatively modest increase in the F 1 score for the Template Learning method ( 0.85, 281 compared to the previous best of 0.83), it stands as the first demonstration that deep 282 learning can be trained effectively for picking a target structure, starting from only prior 283 templates, and domain -randomized simulations. Also, a model of the sam e network 284 architecture (DeepFinder) trained only on simulations outperforms its counterpart trained 285 only on annotated experimental datasets for cryo-ET particle picking. 286 Template Learning variations 287 We introduced several novel concepts in the Template Learning data simulation workflow, 288 demonstrating that, when combined, they achieve state -of-the-art performance for 289 supervised deep learning in annotating ribosomes within in situ cryo-ET tomograms. To 290 study the significance of each proposed concept and explore the potential for integrating 291 various modifications, we conducted additional experiments, which we have organized 292 into the following four categories: 293 Incorporating multiple atomic template structures and flexible variations 294 In three experiments, we progressively reduced the structural variability of input templates 295 while keeping the remaining parameters of the method unchanged. Our motivation behind 296 this series of experiments is to guide potential users regarding the outcomes of employing 297 multiple template structures and integrating flexible variations. These experiments can be 298 particularly useful to future studies, recognizing that, for certain biomolecules, the 299 presence of multiple structures and the feasibility of simulating molecular mechanics may 300 vary. 301 In one experiment, we retained from the templates a single structure and its flexible 302 variations. In a second experiment, we retained the six structures while excluding the 303 flexible variations. Fin ally, in a third experiment, we retained only a single structure, 304 eliminating other structures and all flexible variations. The trained deep learning models 305 resulting from the Template Learning workflows of these three experiments were used to 306 analyze the dataset as presented previously. 307 Fig.4(B-D, I) depicts the Recall-Precision curves and F1 scores for the three experiments. 308 Notably, employing a single structure with flexible variations yielded comparable results 309 to employing multiple structures without f lexible variations. We can infer from these 310 findings that introducing artificial flexible variations can serve as an effective 311 compensatory mechanism for the absence of multiple PDBs. This becomes particularly 312 relevant in studies where only a limited number of atomic structures, or even just a single 313 structure, is publicly accessible. Both approaches (i.e., a single PDB with flexibility and 314 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 10 multiple PDBs without flexibility) showed a slight adverse impact on the median F1 score 315 of the aforementioned typical template learning workflow, of approximately 5%. However, 316 they still demonstrated comparable performance to previously reported results from 317 supervised deep learning on annotated experimental data (compare Fig.4I/B -C and 318 Fig.4J). 319 In contrast, employing a single template without any flexible variations yielded a stricter 320 model characterized by higher Precision and lower Recall —translating to fewer false 321 positives at the cost of more false negatives (see Extended Data Fig.3 for the output 322 segmentation maps). Overall, the use of a single PDB template led to a relative reduction 323 in the median F 1 score by approximately 10%. Nevertheless, it maintains a significant 324 advantage over traditional template matching (compare Fig.4I/D and Fig.4J/Template 325 matching). 326 Incorporating distractors 327 In contrast to our proposed Template Learning method, previous studies did not 328 emphasize the incorporation of a background combining a comprehensive array of 329 dissimilar distractors. In 48, no distractors were employed in the data simulation to train 330 deep learning models on subtomogram segmentation. Two recent studies 27,40 utilized a 331 small set of objects, serving as analogs to distractors in the simulation of data for training 332 deep neural networks in regression tasks, including signal restoration and segmentation. 333 These objects included gold fiducials, actin bundles, vesicles, and randomly placed small 334 spheres that were relatively sparsely distributed in the volume. Our strategy is principally 335 different; it utilizes other real biomolecules as distractors for a simple approximation of 336 the realistic crowding of biomolecular environments. 337 To investigate the contribution of emplo ying as many as 100 dissimilar distractors to 338 annotate only a single target structure on the state -of-the-art performance of Template 339 Learning, we conducted two additional experiments on distractors while keeping other 340

parameters unchanged. In one e xperiment, we removed all distractors from the 341 simulated tomograms, following a methodology similar to that outlined in 48. Fig.4(E, I) 342 depicts the Recall -Precision curves and F 1 scores for this experi ment, revealing that 343 removing distractors from the simulated data results in hallucinations in the output 344 segmentations leading to a high rate of false positives (refer to Extended Data Fig.3). 345 In a subsequent experiment, we explored if there are benefits of using many distractors 346 compared to a limited number, similar to previous approaches 27,40. Subsequently, we 347 reintroduced three distractors—small (PDB 1S3X), medium (PDB 5A20), and large (PDB 348 6UP6) proteins —while maintaining the previously established volumetric density ratio 349 balance between distractors and templates in the simulated tomograms. Fig.4(F, I) 350 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 11 depicts the Recall-Precision curves and F1 scores for this experiment. These results show 351 that limiting the variability of distractors had a serious adverse impact on the picking 352 Precision. 353 These experiments suggest that exposing the deep learning model to a diverse range of 354 unwanted structures (i.e., distractors) alongside the target structures (i.e., templates) is 355 essential for the best performance. 356 Simulating crowding 357 To investigate the contribution of simulating high crowding, (i.e., using the Tetris algorithm 358 introduced in this work) to the state -of-the-art performance of Template Learning, we 359 aimed to design an experiment involving crowding reduction. Consequently, we continued 360 our investigation of the crowding reduction impact on Template Learning by adapting the 361 Tetris algorithm to systematically generate crowding levels aiming for 25 -35% crowding 362 relative to the original settings while keeping all other parameters unchanged (to see how 363 crowding can be controlled in the Tetris algorithm, see the corresponding Methods 364 section). 365 Fig.4(G, I) depicts the Recall -Precision curves and F 1 scores for this experiment, 366 revealing that the deep learning model trained on a simulated dataset with less crowding 367 resulted in a 20-25% decrease in the median F1 score. 368 Using a volumetric template 369 To the best of our knowledge, this work represents the f irst effort to train deep learning 370 models to utilize fully atomic structures and physics-based simulations as alternatives to 371 volume templates in traditional cryo -ET template matching. An advantage of employing 372 fully atomic structures is the utilization of rapid molecular mechanics simulations using 373 NMA. Particularly, our strategy involves coarse-graining the structures before NMA, and 374 interpolating the resulting motions to the original fully -atomic model (refer to the NMA 375 section in Methods for further details). 376 However, using volumetric templates, in the context of particle picking, can be 377 advantageous in specific scenarios, particularly when an initial structure can be derived 378 from the dataset. For instance, a subtomogram average at low resolution derived from a 379 partially annotated dataset can serve as an initial template to annotate more particles in 380 the dataset, enabling the generation of a higher-resolution average. 381 Consequently, we explored a method to incorporate a volume as a template in the 382 physics-based simulation, rather than relying solely on atomic structures. Existing 383 literature demonstrates the feasibility of generating pseudoatomic models from volumes, 384 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 12 primarily by estimating the volume using a set of 3D Gaussian functions 49,50. The mean 385 and standard deviations of these 3D Gaussians are calculated in a way that the 386 pseudoatomic model, when converted back to a volume (i.e., via summation of the 387 contribution of each 3D Gaussian to each voxel), estimates the volume with low error 388 (e.g., 5%). However, advanced cryo -ET physics-based simulators (e.g., Parakeet, used 389 in this work), do make use of the characteristics of these 3D Gaussians in the simulations, 390 instead, they use electron -atom interactions to model the image formation in the TEM. 391 Meaning, pseudoatomic structures based on previous works do not generate the 392 expected contrast when used in such physics-based simulators. 393 Hence, we devised a two -stage algorithm for real -time conversion of volumes into 394 pseudoatomic models that allows tuning the contrast by substituting some pseudoatoms 395 with actual phosphorus atoms (see Extended Data Fig.4 for illustration and Methods for 396 details). 397 To proceed with investigating the effect of using a volume as a template for ribosome 398 annotations, we employed the atomic structure of a ribosome (PDB 4UG0) to generate a 399 volume at 2 nm resolution using the Eman2 46 pdb2mrc software. It is noteworthy that in 400 the study 20 where the dataset (EMPIAR -10988) was published, a sub -nanometer 401 subtomogram average was obtained. Here, we explored a hypothetical scenario where a 402 2 nm resolution initial average is obtai ned by averaging a partially annotated subset of 403 the original dataset and investigated if the Template Learning workflow can employ this 404 structure for annotating more ribosomes. Subsequently, we transformed this volume into 405 a pseudoatomic structure, follow ing the steps outlined in the corresponding Methods 406 section. As part of this conversion process, we introduced a substitution, replacing one -407 third of the pseudoatoms with phosphorus atoms to approximate the contrast of 408 ribosomes observed in experimental data (Extended Data Fig.4). 409 We employed this pseudoatomic model to generate simulations using the Template 410 Learning workflow, by replacing the atomic structures that were previously used as input, 411 keeping the remaining parameters unchanged. 412 Fig.4(H, I) depic ts the Recall -Precision curves and F 1 scores for this experiment, 413 indicating that the deep learning model trained on a simulated dataset with the mentioned 414 pseudoatomic structure exhibits a lower Precision and a higher Recall compared to 415 training with an actual atomic structure. We interpret these findings suggesting that lower 416 resolution template structure (2 nm instead of atomic resolution) removed certain details 417 from the simulation, resulting in a deep learning model that is more permissive but less 418 precise in its predictions. Nonetheless, the F 1 score of this experiment shows a 27% 419 increase relative to traditional template matching, despite that both methods used a 420 similar prior (i.e., a single volume ribosome template). 421 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 13 Adapting Template Learning to different domains 422 In the preceding section, we benchmarked Template Learning for annotating ribosomes 423 in a close-to-focus VPP cryo-ET dataset in situ. Hence, the previous Template Learning 424 workflow for ribosome annotation involved simulating data using VPP c lose-to-focus, 425 consistent with the data analyzed. 426 In addition to VPP data, the dataset (EMPIAR -10988) contains an additional set of 10 427 tomograms acquired with defocus (DEF tomograms) and without VPP. In the original 428 study 20, the DeePiCt deep learning model was trained on the experimental VPP 429 tomograms and subsequently evaluated on DEF tomograms pre -processed with 430 spectrum matching (SM). SM enhanced the contrast of the DEF dataset to resemble that 431 of the VPP dataset, achieved by extracting a target spectrum from a VPP tomogram and 432 applying it to the DEF tomograms (see Extended Data Fig.5). This approach was referred 433 to as the DeePiCt "cross-domain" experiment. 434 We conducted two experiments to explore the impact of the data domain on Template 435 Learning. In the first experiment, we set up a Template Learning workflow targeting the 436 DEF domain. We accomplished this by generating a simulated dataset using a procedure 437 identical to the one outlined in the preceding section except for excluding VPP simulations 438 and employing a different range of defocus values (refer to Extended Data Tab.1 for 439 details). This simulated dataset was used to train and benchmark DeepFinder on 440 annotating ribosomes within the DEF tomograms, without undergoing SM preprocessing. 441 In the second experiment, we utilized the DeepFinder model trained on Template 442 Learning VPP simulations to annotate ribosomes within the DEF dataset that had 443 undergone SM preprocessing. 444 Fig.5 depicts the Recall -Precision curves and F 1 scores for the two aforementioned 445 experiments, compared to the previously reported result of the DeePiCt cross -domain 446 experiment. Both of our experiments yield comparable outcomes, outperforming the 447 DeePiCt cross-domain experiment by more than 10% on the median F 1 score. Notably, 448 the Template Learning-trained model for DEF data exhibited a higher Precision curve and 449 ultimately achieved a higher F 1 score, particularly when no mask (cytosol mask) was 450 applied. 451 The first experiment shows that the Template Learning workflow can be directly adapted 452 to the DEF domain avoiding the need for data preprocessing such as SM. The second 453 experiment shows that if a model trained on the Template Learning simulations of the 454 VPP domain exists, it can be reused in the cross-domain (i.e., the DEF domain) after SM, 455 and achieves a similar performance without repeating the entire Template Learning 456 workflow. 457 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 14 Fine-tuning a deep learning model trained on Template Learning 458 simulations using experimental data annotations 459 The literature on domain randomization indicates that models pre-trained on simulations 460 and subsequently fine-tuned on real-world data outperform models trained exclusively on 461 real-world data 35. This approach can be particularly valuable to cryo -ET particle picking 462 studies that involve challenging -to-annotate molecules, or when training on domain 463 randomization simulations alone does not yield satisfactory results. 464 In the same dataset used for benchmar king Template Learning on ribosomes in situ 465 (EMPIAR-10988), another molecule, FAS, is annotated. Previous attempts based on 466 template matching and supervised deep learning on experimental datasets have reported 467 challenges in annotating FAS, attributed to it s distinctive shell -like structure, where 468 particles have lower SNR compared to ribosomes 20. 469 In the following, we present the benchmarking of Template Learning combined with 470 DeepFinder, initially train ed exclusively on a simulated dataset, on FAS annotation in 471 EMPIAR-10988 tomograms. Subsequently, we gradually introduce fine -tuning based on 472 the experimental data annotations to assess whether performance can be improved. 473 To establish a Template Learning workflow for FAS, we utilized two PDB structures (PDB 474 IDs: 4V59 and 6QL5). By using NMA, we generated 80 flexible variations for each 475 structure (refer to the NMA section in the Methods for details). We kept the remaining 476 parameters for simulated data generation consistent with those aforementioned for the 477 ribosome study, except for training the models on sphere segmentations due to the shell-478 like structure of FAS. 479 The F1 score results of the simulation-trained DeepFinder model for FAS annotations are 480 presented in Fig.7. The results show that on the VPP dataset, the DeepFinder model 481 trained on the Template Learning workflow outperformed that trained exclusively on 482 experimental data. The results on the DEF dataset gave an F 1 score ranging up to 22%, 483 outperforming previous attempts that failed to pick any particle, based on the results 484 reported in 20. However, these results can still benefit from improvements compared to 485 previously obtained results with DeePiCt trained on annotated experimental data. 486 We proceeded to fine-tune the simulations-trained DeepFinder model progressively using 487 annotations from experimental data. In the following experiment, we fine -tuned it using 488 annotations from 2 VPP tomograms containing approximately 150 FAS annotations. We 489 kept all the DeepFinder training parameters to default (listed in Extended Data Tab.2), 490 except for reducing the number of steps and epochs to 10 steps for 10 epochs (in place 491 of 100 steps for 100 epochs) to avoid overfitting (since the number of training data 492 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 15 examples is low). We performed three cross -validation experiments (i.e., in each 493 experiment, the data is randomly split into 2 tomograms for training and 8 for 494 benchmarking). The results presented in Fig.6 show a significant increase in the median 495 F1 score compared to training on simulated data only. 496 In a subsequent experiment, we fine -tuned the simulation-trained DeepFinder model on 497 8 VPP tomograms containing approximately 600 FAS annotations and inferred the model 498 on the remaining 2 tomograms, using 3 cross -validation splits. Again, we kept all the 499 DeepFinder training parameters to default, except for the number of steps and epochs, 500 keeping 10 steps for 30 epochs (again, to p revent overfitting). The results presented in 501 Fig.6 show a significant increase in the median F 1 score, outperforming previous 502 supervised deep learning methods trained solely on experimental data. 503 Finally, we performed a cross -domain experiment, where we f ine-tuned the simulation -504 trained DeepFinder model on the VPP dataset, and applied this model to annotate FAS 505 in the DEF dataset after SM preprocessing. Consistently with previous findings, the 506

presented in Fig.6 show that training on simulations an d fine -tuning on 507 experimental data outperforms training only on experimental data. 508 Template Learning offers higher Precision and better orientational 509 isotropy than template matching on nucleosome picking 510 In cryo-ET data processing, localizing a target biomolecule in a new dataset is a common 511 objective. The application of supervised deep learning to this task requires an initial 512 training dataset of particles of interest extracted directly from the new data. Such a dataset 513 is usually in the order of thousands of particles in different orientations, where more 514 particles are needed for better performance. Manually annotating these particles is time-515 consuming, making template matching followed by manual elimination of obvious false 516 positives and further curation through subtomogram averaging and classification the most 517 common approach 19,20. The principle of Template learning offers a time and computing-518 efficient alternative to this complex procedure by annotating new data directly in a single 519 step using a model trained on synthetic data. 520 In this section, we explore the efficiency of Template Learning for nucleosome annotation 521 within a new cryo-ET dataset of partially decondensed mitotic chromosomes in vitro (refer 522 to the Method section for sample preparation details) and compare it with template 523 matching-based annotation routine. 524 A tomogram central slice of our data is shown in Fig.8A. Despite this dataset being 525 isolated chromosomes in vitro , it contains a n abundance of structures other than 526 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 16 nucleosomes that can be false positives which are: DNA linkers, gold nanoparticles, 527 percoll, and the non-histone components of chromatin. 528 Template Learning for picking nucleosomes 529 To establish a Template Learning workfl ow, we utilized six nucleosome template 530 structures with PDB IDs: 2PYO, 7KBE, 7PEX, 7PEY, 7XZY, and 7Y00. By applying NMA, 531 we generated 100 flexible variations for each structure (more details can be found in the 532 NMA section of Methods). The remaining parameters of the Template Learning workflow 533 were configured similarly to those used for the workflow established for ribosome and 534 FAS annotation, with a few notable exceptions explained below. 535 Firstly, recognizing the smaller size of nucleosomes compared to ribosomes, we adjusted 536 the frequency of appearance of nucleosome templates to one nucleosome for every two 537 distractors (in contrast to one ribosome for every five distractors). This adjustment aimed 538 to maintain a balanced volumetric density ratio between dis tractors and templates in the 539 simulated tomograms. 540 Secondly, recognizing that nucleosomes have fewer atoms than ribosomes, we observed 541 increased speed in the execution of the physics -based simulator (i.e., Parakeet). This 542 allowed for the generation of larg er tomograms for training, all within the same runtime 543 for the ribosome study (see Software requirements section for details). In particular, the 544 tomogram size generated in the Tetris algorithm here was set to 256x256x64 (for a pixel 545 size of 16 Å), in contrast to 192x192x64 used for ribosome simulations. 546 Lastly, we adjusted the pixel size of the simulated data to 8 Å through binning, closely 547 approximating the pixel size of our experimental tomogram, which had been previously 548 binned four times before annotation (the unbinned pixel size was 2.075 Å). 549 Consequently, we trained DeepFinder on the resulting Template Learning simulations 550 and applied it to segment our data. The corresponding segmentation map (score map), 551 is shown in Fig.7B. Upon visual inspection, it is evident that that model has high Precision 552 in localizing nucleosomes. Subsequently, we employed the MeanShift algorithm from the 553 DeepFinder software to extract the corresponding annotations for the nucleosomes, 554 utilizing a clustering radius of 6 pixels , approximately equivalent to the nucleosome's 555 radius at this pixel size. We excluded particles close to the edges of the tomogram, the 556 air-water interface, and the sample-carbon interface. 557 Following this process, we obtained 18k annotated particles and pr oceeded with the 558 reference-free subtomogram averaging procedure using Relion V4.0 15 by generating an 559 initial model from the data, followed by two stages of refinement at binning 4 and binning 560 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 17 2, resul ting in resolving the nucleosome structure at 12.8 Å (at 0.143 FSC threshold) 561 resolution shown in Fig.7B. 562 The angular distribution of the averaged particles seems isotropic, showcasing that the 563 Template Learning method is not biased to certain orientations. 564 To further verify the Precision of nucleosome identification, we classified the particles into 565 10 and 50 classes (presented in Extended Data Fig.6). All class average results showed 566 well-recognizable nucleosome shapes with DNA gyres, with the major diff erences being 567 in the heterogeneity of the DNA entry-exit segments. We also performed a local resolution 568 assessment using ResMap 51 (presented in Extended Data Fig.7) and observed that 569 resolution has ranged from 12-20 Å, with the highest resolution in the DNA close to the 570 core histones, and the lowest resolution around the DNA entry-exit segments. 571 Our averaging and classification results show that our Template Learning workflow 572 generated precise nucleoso me annotations of uniform orientations, facilitating high -573 throughput subtomogram averaging without the need for multiple rounds of classification 574 to eliminate false-positive particles. 575 Template matching for picking nucleosomes 576 We applied template matching starting from a nucleosome template structure generated 577 from the PDB 2PYO using the default procedure in PyTom 18. We then applied the 578 template matching procedure to our new data using an angular sampling of 7° increment 579 (45,123 rotations). 580 The score map is shown in Fig.7C. The visual inspection of the score map shows some 581 high signal for some nucleosomes also identified by the Template Learning procedure 582 (Fig.7A), but also obvious false positive signal for other objects (an example of high false 583 signal for the Percoll particles is shown at the center of the zoomed-in image in Fig.7C). 584 To have a meaningful and fair comparison with the Template Learning procedure, we 585 extracted from the same tomographic region the 18k particles showing the highest cross-586 correlation scores. Unlike the results of Template Learning, a brief visual inspection of the 587 template matching peaks showed 50 obvious false positives corresponding to the 10 nm 588 gold and Percoll particles. The false positives were removed, and data were processed 589 Relion V4.0 15. The reference -free initial model was generated, followed by a stage of 590 refinement at binning 4. Unlike Template Learning, the refinement did not result in a 591 significant improvement of the resolution of the initial model (judged by the shape and the 592 FSC), indicating the presence of a significant ratio of non -nucleosome particles among 593 the annotations. In agreement with this assumption, classification, and alignment into 10 594 classes at the same binning resulted in only 2 nucleosome-like classes that sum to 10.25k 595 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 18 picks (57% of the original picks). The particles of these 2 cla sses were selected for a 596 further refinement process at binning 4 reached Nyquist resolution, and the further 597 refinement at binning 2 led to 12.9 Å resolution which is comparable to that achieved with 598 Template Learning. 599 Importantly, the angular distribution of nucleosomes annotated by template matching 600 showed a significant imbalance towards side views compared to round top views (Fig.7C). 601 Our simulations (Extended Data Fig.8) showed that due to the cylindrical shape of 602 nucleosomes and the missing wedge probl em of cryo -ET data, the constrained cross -603 correlation between a template and the particles is a function of the orientation, where 604 side views show higher cross-correlation than top and oblique views. This problem results 605 in extracting only side views of nucleosomes at an adequate Precision. 606 The orientation bias of template matching hampers its efficiency for particle detection and 607 is prone to resolution loss in subtomogram averaging. In our case, the resolution of the 608 nucleosome was not affected significantly, because its pseudo -symmetric cyl indrical 609 shape allowed the complete 360° orientation range of side views to have cross-correlation 610 sufficient for detection (Fig.7C). However, this constraint may pose a more significant 611 challenge when annotating particles with asymmetric non -spherical shapes. Our results 612 demonstrate that Template Learning is capable of overcoming this limitation. 613

614 In this study, we introduced Template Learning, a novel approach that blends the 615 simplicity of Template Matching, requiring only template structures as input, with the 616 advanced capabilities of Deep Learning, thereby reducing the dependence on extensive, 617 labor-intensive annotated datasets for supervised training. Our method effectively tackles 618 key challenges in cryo -ET particle picking, such as structural variability, cellular 619 crowdedness, and experimental data variance, through a refined implementation of 620 domain randomization to create simulated data for deep learning training. The efficiency 621 of Template Learning arises from its focus on essential aspects , enabling the well -622 established DeepFinder model, known for its particle annotation capabilities, to train 623 without the need for highly realistic cellular environment simulations. This approach 624 simplifies the process for users, allowing them to generate effective simulations that train 625 DeepFinder in annotating specific biomolecules in cryo -ET data, using only templates in 626 a streamlined yet customizable simulation process. Template learning can utilize any 627 atomic models, including those generated by artificia l intelligence -based tools, as 628 templates. We also confirm the feasibility of relevant template generation from low -629 resolution cryo-EM densities. 630 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 19 In this paper, we employed a comprehensively annotated in situ cryo-ET dataset for 631 ribosomes and FAS, covering defocused cryo -ET and VPP, to validate Template 632 Learning's efficacy and versatility. We demonstrated how to enhance DeepFinder's 633 supervised training by initially training on simulations, and then fine -tuning on 634 experimental data. Notably, DeepFinder, when trained solely on simulations, surpassed 635 previously established training on experimental annotations for ribosomes. Furthermore, 636 the combined approach of pre -training on simulations and subsequent fine -tuning on 637 experimental data showed improved performance for FAS. 638 We evaluated the efficacy of our method for the identification of a known target molecule 639 in tomographic data without any prior annotation, focusing on the case of localization of 640 nucleosomes in isolated chromosomes in vitro . Template Learning o utperformed the 641 conventional template matching and classification routine typically employed for this type 642 of task both in terms of annotation Precision and orientational isotropy. Notably, all 643 annotations generated by Template Learning could be directly u sed for achieving high -644 resolution subtomogram averaging without requiring any manual or classification -based 645 curation. 646 In future work, we envision adapting our framework to locate structures bound to other 647 biomolecules such as membrane proteins. Another ch allenge will be to reduce 648 computational issues as physics-based simulations are still resource-intensive. 649 Conclusively, we believe Template Learning's versatile and straightforward framework 650 marks it as a timely and potent tool for a broad spectrum of cry o-ET studies, ready to 651 make significant impacts in the field. 652 Software requirements and availability 653 Template Learning code is freeware, open source, GPU and CPU optimized, and fully 654 implemented in Python (https://github.com/MohamadHarastani/TemplateLearni ng). It 655 provides all the steps that are required to reproduce the results presented in this article 656 and can be straightforwardly applied to other studies. It provides wrappers for ProDy, 657 Eman2 and Parakeet that will be installed during the installation. The software was tested 658 on our workstations (Dell Precision 5820, Intel(R) Xeon(R) W-2145 CPU @ 3.70GHz, 96 659 GB DDR4 RAM, 2 X NVIDIA RTX A6000 or NVIDIA Quadro RTX 8000). The time required 660 to perform the Template Learning workflow ranged from 1 to 2 days (dep ending on the 661 number of used GPUs). Installation and user guides are available on GitHub. 662 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 20 Data availability 663 The tilt series, tomogram, coordinates, and other metadata necessary to reproduce the 664 nucleosome subtomogram averaging are deposited on EMPIAR via the accession code 665 EMPIAR-11969. The subtomogram averages resulting from the Template Learning and 666 template matching workflows are available on EMDB via the accession codes EMD -667 19823 and EMD-19825 respectively. Other benchmarking datasets employed in this work 668 are publicly available. 669 Author contributions 670 Investigation, conceptualization, and methodology were done by MH and ME with 671 contributions from all authors. Software programming and computational experiments 672 were done by MH. Sample preparation and data acquisition were done by GP under the 673 supervision of ME. Subtomogram averaging was done by MH and GP under the 674 supervision of ME. Validation and results interpretation were done by all authors. Writing 675 the original draft was done by MH with input from GP. Review and editing were done by 676 MH, ME, and CK. Funding acquisition was done by ME and CK. Project administration 677 was done by ME. 678 Conflicts of Interest 679 The authors declare no conflict of interest. 680 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 21 Figures 681 Fig.1: Template Learning workflow for generating s imulations used for supervised deep learning used for cryo-ET picking of target particles. Template structures are used as input, commonly differing by some structural variations of the target biomolecule (we mean by the target biomolecule, the structure that the deep learning model is desired to learn to annotate in cryo-ET tomograms, in the figure it is the nucleosome). The input templates are augmented with flexible variations based on fast molecular mechanics simulations (see text for details). A gene ral set of other proteins, different from the input templates, are used as “distractors” when generating synthetic data. The templates, along with their flexible variations, and distractors are placed at random orientations in close proximity using the “Te tris algorithm” — an algorithm proposed in this work to enable fast simulation of high molecular crowding. A physics-based simulator (Parakeet 28) is used to generate synthetic cryo-ET data using a set of different parameters that were chosen to produce domain randomization at the lowest cost, which are the electron dose, tilt range, tilt step, ice density, and defocus. The resulting output comprises tomograms and ground truth segmentations and coordinates corresponding to the templates. The generated data is subsequently utilized to train deep learning models (e.g., DeepFinder 19) for experimental tomogram segmentation and particle picking. The PDB I Ds used for templates in this figure are (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 22 2PYO, 7KBE, 7PEX, 7PEY, 7XZY, and 7Y00, and for distractors are 3QM1, 7NIU, and 6UP6. The display of structures was performed using ChimeraX 52 and IMOD 47 software. 682 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 23 683 Fig.2: The Tetris algorithm for generating crowded 3D samples. (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 24 A The flowchart of the algorithm, demonstrated in 2D for simplicity. The molecules that are desired to populate the sample are used as input. In the initial iteration, a molecule is positioned at the center of an expanded volumetric sample, forming the curr ent output. In subsequent iterations, a new molecule undergoes binarization and dilation. By multiplying the binarized version with a large positive value and subtracting the result from the dilated version, an insertion shell is generated (referred to as 'in-shell'). This insertion shell is then correlated with a binarized version of the current output volume, resulting in a correlation map (referred to as 'c-map'). Positive values in the c-map (white voxels) represent viable positions for adding the new m olecule without intersection, whereas negative and zero values (black and gray voxels, respectively) represent positions where intersection with existing molecules occurs or where the distance from other molecules is too large. If the correlation map is not entirely negative, the molecule is added to the current output at the index of the maximum entry, ensuring it avoids intersections while maintaining the highest compactness. B An example of the Tetris algorithm input and output at each iteration, alterna ting between placing templates (in green) and distractors (in red). The numbers in B correspond to the output at different iterations of the algorithm in A. C An example of a Tetris algorithm output in 3D. 684 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 25 685 Fig.3: Template Learning trains DeepFinder to single molecule precision on segmenting and annotating ribosomes in situ. Left: Display of a 2D central slice of a VPP tomogram. Middle: Segmentation of ribosome from the Template Learning workflow shown in black an d white, with a 50% transparent overlay of the mask. Right: resulting annotations within the mask, obtained by converting segmentations to annotations using the MeanShift algorithm (provided by the DeepFinder 19 software) using a clustering radius of 10 voxels, compared to expert annotations. Comparison includes True Positives (TP), False Positives (FP), and False Negatives (FN) relative to expert annotations. The tomogram, mask and expert annotations are sourced from the EMPIAR-10988 dataset. 686 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 26 687 Fig.4: DeepFinder trained on Template Learning simulations only outperforms previous techniques for ribosome annotations in cryo-electron tomograms. Performance benchmarking and comparative analysis of the Te mplate Learning on ribosome annotation on the S. pombe dataset (EMPIAR -10988, 10 VPP tomograms). (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 27 A-H Performance measures for 8 variations of Template Learning settings. The curves depict the overall Precision and Recall against the volume of the segmented region (horizontal axis). Precision is evaluated with and without masking (±mask). Boxplots show the F1 score per tomogram at the threshold corresponding to the highest overall F1 score (often coinciding with the intersection of the Precision and Recall c urves). I Table listing the differences between the experiments in A-H comparing their results based on their median F 1 scores (bold value is the highest among proposed and previous techniques). J Table of previous results with their reported performance based on the findings in 20. Comparison of segmentations resulting from the different experiments on an example tomogram are shown in Extended Data Fig.3. 688 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 28 689 Fig.5: Template Learning does not necessitate data preprocessing for DEF tomograms. Performance benchmarking and comparative analysis of Template Learning workflow applied to ribosome annotations within the S. pombe dataset (EMPIAR-10988, 10 DEF tomograms). A-B The curves depict the overall Precision and Recall against the volume of the segmented region (horizontal axis). Precision is evaluated with and without cytosol masking (±mask). Boxplots show the F 1 score per tomogram at the threshold corresponding to the highest overall F 1 score (coinciding with the intersection of the Precision and Recall curves). C On the top, a table listing the differences between the experiments in A-B, comparing their results based on their median F 1 scores. On the bottom, previously reported performance based on the findings in 20. 690 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 29 691 Fig.6: DeepFinder can be pre-trained on Template Learning simulations and fine- tuned on experimental data annotations. Performance benchmarking and comparative analysis of Template Learning applied to FAS annotation within the S. pombe dataset (EMPIAR -10988, 10 DEF tomograms). The results show the performance of the Template Learning workflow used for training DeepFinder on simulations, followed by different levels of fine -tuning on experimental data, compared to results based on the findings in 20. 692 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 30 693 Fig.7: Template Learning outperforms traditional template matching in Precision and orientational isotropy in annotating nucleosomes. A Central slice and zoomed -in 10 -slice average of a cryo -ET tomogram of partially decondensed mitotic chromosomes in vitro . B-C The results of applying Template Learning (6 PDB structures as input, other parameters kept default) and traditional template matching (using PyTom with a nucleosome template) for annotating nucleosomes within the tomogram in A, followed by reference -free subtomogram averaging in Relion V4. B Top: score map result ing from training DeepFinder on (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 31 nucleosome annotations using simulated datasets only from the proposed Template Learning workflow. Below: results of subtomogram averaging using 18k annotations extracted using 6 voxel clustering radius of the MeanShift algo rithm (provided by the DeepFinder software). The results show initial model generation, followed by two stages of auto-refinement, and the angular distribution of averaged particles. The refinement has converged without particle curation and the angular di stribution of averaged particles is isotropic. C Top: score map resulting from traditional template matching. Below: results of subtomogram averaging using the top 18k local maxima from the score map. The first step was an initial model generation, followe d by a stage of refinement. The first stage of refinement did not improve the resolution of the initial model, therefore, it was followed by simultaneous classification and alignment into 10 classes. The particles corresponding to nucleosome -like classes ( 57% of the original 18k particles) were used for two stages of refinement. The results show a slightly lower resolution average than the one obtained using Template Learning and show an anisotropic angular distribution. 694 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 32

695 Normal Mode Analysis (NMA) 696 Our study employed NMA on various atomic structures, prioritizing computational 697 efficiency over precise molecular mechanics simulation, as the primary focus was on 698 training the network on variations of the target biomolecule rath er than precisely 699 predicting its conformations. This emphasis on expanding the domain of the synthetic 700 data aimed to capture a broad range of variations, not solely constrained to those 701 witnessed in experimental data. 702 To manage computational demands, we adopted a coarse-grained modeling strategy 53, 703 focusing on Carbon Alpha and Phosphorus atoms for NMA and subsequently expanding 704 (interpolating) the modes to the complete atomic structure. Typically, a su bstantial 705 number (three times the number of atoms in the atomic structure) of normal modes are 706 calculated simultaneously, which can be ordered based on their frequency. Low -707 frequency normal modes are particularly useful as they represent global movements, 708 while high -frequency normal modes depict local movements. Prior studies have 709 consistently highlighted the significance of these low-frequency, high-collective modes in 710 capturing experimentally observed conformational variabilities 54,55. 711 The user of NMA retains control over the selection of modes and the amplitude of 712 deformation to generate flexible variations, ensuring a balance between simulating 713 variability and maintaining structural integrity. As a guideline, we chose the first 20 low -714 frequency normal modes for generating random flexible variations. Setting the amplitude 715 range to 150 for ribosomes and 100 for nucleosomes resulted in Root Mean Square 716 Deviation (RMSD) values of approximately 1 Å a nd 2 Å from the initial structures, 717 respectively. Adjusting this parameter incurs minimal computational costs (a few minutes) 718 by generating variations and assessing the RMSD from the initial structure, for instance 719 using ChimeraX 52. 720 The computational time required for the calculation of NMA varies depending on the size 721 of the input atomic structure. Specifically, for an atomic structure of 200 kDa, such as a 722 nucleosome, the calculation typically takes approximately 1 minute. In contrast, for larger 723 structures like a ribosome weighing around 4.5 MDa, the calculation time extends to about 724 1 hour when processed on a single CPU (benchmarked on Intel(R) Xeon(R) W-2145 CPU 725 @ 3.70GHz). 726 In this work, all NMA c alculations were performed using ProDy 56, a widely recognized 727 open-source Python package for molecular mechanics simulations. 728 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 33 Tetris algorithm for generating highly crowded samples 729 The Tetris algorithm (flowchart shown in Fig.2) utilizes input volumes of the biomolecules, 730 essentially volumetric versions of the PDB templates and distractors, to generate a 731 densely packed sample comprising randomly rotated biomolecule copies. It 732 systematically places one randomly rotated molecule at a time in the closest proximity to 733 previously positioned molecules. The algorithm terminates when a new copy cannot be 734 placed in the sample. 735 In the initial step, the Tetris algorithm places a molecule at the center of a larger volumetric 736 sample, the size of which is set by the user, establishing the current Tetris output. 737 Subsequent iterations involve randomly rotating a new molecule, followed by its 738 conversion to binary form after applying low -pass filtering and thresholding. W hile the 739 values for low -pass filtering and the volume threshold are empirical and subject to 740 variation, a standard deviation sigma of 2 for a Gaussian low -pass filter, and a volume 741 threshold of 100 have generally proven effective in our trials (note that t he volumes are 742 generated from PDBs using Eman2 46 e2pdb2mrc software). This copy (molecule) is then 743 diluted using a ball -like structural element of user -defined radius. Multiplying the output 744 of the binarized volume by a large positive number (integer infinity in programming) and 745 subtracting the result from the dilated volume, results in an "insertion shell", which serves 746 as a spatial guide for placing the next molecule. Notably, the radius of the struc tural 747 element provides direct control over the compactness of the sample. 748 The correlation between this insertion shell and a binary version of the current output (the 749 binary version is smoothed and thresholded the same way as above) generates a 750 correlation map. White voxels in the correlation map represent potential locations fo r 751 adding the copy, ensuring the desired intermolecular distance. In contrast, black and gray 752 voxels represent locations where the copy would intersect with previous copies or be too 753 distant from other molecules, respectively. If the correlation map is not entirely negative 754 (stopping criteria), the current copy is added at the index of the maximum entry, 755 corresponding to the location closest to the sample within the desired distance. 756 During Tetris sample generation, the oriented coordinates used for placing molecules in 757 the volumetric sample are preserved. To generate cryo-ET simulations of this sample, the 758 oriented coordinates are passed to the physical-based simulator (Parakeet). 759 Generating volumetric samples using the Tetris algorithm with molecules at a r esolution 760 of 32 Å and a voxel size of 16 Å 3, as implemented in our current software, achieves a 761 placement rate of approximately 1000 molecules per minute. The creation of a typical 762 Tetris, with dimensions of 3072 x 3072 x 1024 Å3, requires approximately 5 to 6 minutes 763 on a single CPU (benchmarked on Intel(R) Xeon(R) W -2145 CPU @ 3.70GHz) and can 764 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 34 accommodate around 5000 molecules. Consequently, the generation of 48 Tetris 765 samples necessary for the tests performed in this study takes approximately 4.5 hours on 766 a single workstation. 767 Performance assessment metrics of particle picking 768 In evaluating how well our deep learning models pick particles, we used the commonly 769 employed metrics in particle localization studies, namely: the Precision, Recall, and F 1 770 score. Precision measures the percentage of correct picks out of all the particles selected 771 by the model (True Positives / Positives). Recall measures how many of the actual 772 particles were correctly identified by the model (True Positives / Ground Truth). F1 score, 773 being the harmonic mean of Precision and Recall, provides a comprehensive measure of 774 overall model performance. 775 Converting volumes to pseudoatoms with contrast tuning 776 To convert a volume to a pseudoatomic structure that allows physics-based simulations, 777 we devised a two-stage algorithm. In the initial stage, we binarize the volume, similar to 778 the process of generating tight 3D masks in cryo-ET and cryo-EM studies. Subsequently, 779 we assign a pseudoatom to the position of every non-zero voxel, resulting in a point cloud 780 as output. We parse this point cloud into a PDB format, representing each point (i.e., 781 pseudoatom) as a density (DENS line in PDB format). Then, the resulting file is ready to 782 be integrated into the physics -based simulations. However, project ions of this structure 783 lack the desired contrast compared to those observed in experimental data. This 784 discrepancy arises from the inherent modeling of electron -atom interactions in physics -785 based simulators. To address this contrast issue, in the second st age, we substitute a 786 fraction of pseudoatoms with actual atoms, such as phosphorus (e.g., replacing one out 787 of three pseudoatoms with phosphorus atoms). This substitution is not aimed at 788 replicating the precise chemical composition of the structure (a dire ction that may be 789 explored in future research) but rather to empirically enhance contrast. We determine this 790 conversion ratio (i.e., pseudoatoms to phosphorus) by projecting the pseudoatomic model 791 in 2D and comparing it to the experimental data targeted for analysis. 792 Cryo-ET of partially decondensed mitotic chromosomes 793 Chromosome Purification 794 Chicken DT40 cells were maintained at a density of 8-10 x 105 /ml at 39 ℃ and 5% CO2. 795 The cells were synchronized to mitosis using 0.5 µg/ml of nocodazole for 13 hr resulting 796 in a mitotic index of 70 - 90%. The mitotic chromosome isolation was performed using the 797 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 35 classic polyamine-EDTA buffer-based method optimized for DT40 57. The synchronized 798 cells were harvested by centrifuging the culture at 1600 g for 5 mins at 4°C. The cells 799 were swollen at room temperature for 5 minutes in a low salt buffer in the presence of 800 polyamines and then lysed using a dounce homogenizer. The lysate was then overlaid 801 on a step sucrose gradient (15%, 60%, and 80% [w/v]) centrifugation. The 60 -80% 802 interface was recovered and sedimented on a self -forming percoll gradient in the 803 presence of polyamine. The band containing the chromosomes was recovered and 804 washed to remove the excess of percoll. The isolated chromosomes are stored in Tris -805 HCl buffer (pH 7.5) containing polyamines in 60% glycerol. 806 Sample preparation and data acquisition 807 The DT40 chromosomes were vitrified in glow-discharged (operating at 30% power with 808 a gas mixture of 80% Argon:20% Oxygen) 200 copper mesh, Quantifoil Multi Holes grids 809 (Quantifoil Micro Tools GmbH, Germany). 810 The isolated mitotic chromosomes were exposed to low ionic strength buf fer by 100-fold 811 dilution in an aqueous solution containing TEEN buffer (1 mM Triethanolamine:HCl pH 812 8.5, 0.2 mM Na-EDTA pH 9 and 10 mM NaCl) for 1 hour at 4 ℃. This resulted in swelling 813 of the mitotic chromosomes which were then carefully spun down onto the surface of the 814 grid. 10 nm BSA Gold Tracer (EMS, Hatfield, PA, USA) was added to the grid at a 5:1 815 ratio (chromosome:gold). The grid was blotted from the carbon side using a teflon sheet 816 and the metal side using a blotting paper respectively. The grid was blotted for 25 seconds 817 with a blot force of 10 and flash -frozen into liquid ethane using Vitrobot Mark IV (FEI) at 818 4 ℃ and 100 % humidity. 819 The tilt series was recorded on Titan Krios G3 (Thermo Fisher Scientific) equipped with 820 a Quantum energy filter wi th a slit width 20eV for higher magnifications and Gatan K2 821 detector using SerialEM Version 4.1.0 beta 58. The images were acquired at a pixel size 822 of 2.075 Å/pixel at 2 –4.0 µm defocus at a dose rate o f 3.2 e −/Å2 per image fractionated 823 over 10 frames. A dose symmetric tilt scheme 59 was used with a 3 -degree increment 824 step and the tilt range was set to ±60 degree using Serial EM. 825 Data processing 826 Individual movies are motion -corrected and averaged to form a tilt series using 827 MotionCor2 60. Gctf was used to perform CTF estimation 61. The tilt series were aligned 828 and processed with IMOD47. The position of the grid in each image is aligned by tracing 829 10 nm gold particles as fiducials throughout the tilt series. The aligned tilt series were 830 reconstructed into 3D tomograms with weighted back -projection with simultaneous 831 iterative reconstruction technique (SIRT)-like filter after 4-fold binning. 832 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 36 After picking nucleosomes with our Template Learning trained DeepFinder model, we 833 performed a reference-free subtomogram averaging following the procedure suggested 834 in Relion V4. 835 836 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 37 Extended Data 837 Extended Data Fig.1: The list of distractors. Template Le arning employs a list of 100 dissimilar protein assemblies, termed distractors, used for domain -randomized cryo-ET data simulations. In this figure, the (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 38 distractors are displayed with corresponding PDB IDs and molecular weights. Display of these structures was done using ChimeraX 52. 838 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 39 839 Extended Data Fig.2: Example of synthetic data based on the Template Learning workflow for supervised deep learning on ribosome annotation. Left: the positions for a crowded sample of size 3072x3072x1024 Å 3 determined using the Tetris algorithm for ribosomes (in green) and distractors (in red), which are fed to Parakeet (cryo-ET physics-based simulator). Right: a central slice of simulated data and corresponding ribosome segmentations using VPP close-to-focus parameters. 840 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 40 Extended Data Fig.3: Segmentation maps resulting from different Template Learning variations for ribosome annotations. Central slice of a VPP tomogram from EMPIAR-10988 with its expert-validated (ground truth) ribosome segmentation and the output segmentation maps using the different Template Learning variations presented in Fig.4. 841 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 41 Extended Data Fig.4: Template Learning can use volumes as templates via volume-to-pseudoatoms conversion. Volume to pseudoatomic model conversion algorithm to produce structures that can be contrast-tuned for usage in physics-based cryo-ET simulators (Parakeet). The method starts by binarizing the volume, a process akin to creating a tight mask through low - pass filtering and thresholding. All non-zero voxels of the binarized volume are parsed into "pseudoatoms", represented in PDB file format utilizing "DENS" entries as atom types. The generated pseudoatomic structure can be used directly for cryo -ET simulations in Parakeet, but it does not necessarily produce a similar contrast to experimental data as when using fully atomic structures. Therefore, the method allows replacing gradually a portion of the pseudoatoms with phosphorus atoms to emulate more contrasted simulated projections. The optimal proportion is determined by the similarity in contrast, comparing the resulting projections to experimental data. The resultant pseudoatomic model is used for further Template Learning simulations. 842 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 42 843 Extended Data Fig.5: Spectrum matching applied to enhance the contrast of an input tomogram obtained with defocused cryo-ET without VPP. A target spectrum is extracted from a tomogram obtained with VPP, and is applied to enhance the contrast of another tomogram obtained with defocus cryo-ET without VPP, via spectrum matching as proposed in 20. Input and target tomograms are sources from EMPIAR-10988. (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 43 844 Extended Data Fig.6: Two experiments of post -alignment classification for nucleosome particles annotated with Template Matching workflow. A Classification into 10 classes. B Classification into 50 classes. Both classification experiments were done in Relion V4 15. (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 44 845 Extended Data Fig.7: Analysis of the nucleosome subtomogram average obtained via Template Learning annotations. A The nucleosome average compared to an available structure (PDB ID: 2CV5). Left: the subtomogram average. Middle: atomic structure docked in the subtomogram average displayed at 50% transparency. Right: atomic structure. B Angular distribution of the parti cles contributing to the global average. C Analysis of the resolution of the average. Left: Fourier Shell Correlation curves. Right: local resolution analysis using ResMap 51. (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 45 846 Extended Data Fig.8: Template matching for nucleosome picking can only extract certain views at an adequate Precision. In this work, we utilized template matching and Template Learning techniques to annotate nucleosomes in cryo -electron tomograms. The a ngular distribution obtained through averaging of template matching -based annotated particles was found (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 46 anisotropic. Leveraging simulations, we explain this phenomenon. A Constrained cross-correlation (CCC) of nucleosome template and simulated nucleosome subtomograms as a function of the angular distribution. Left: CCC histogram of a nucleosome template and ~800 simulated nucleosome subtomograms. The simulated subtomograms were obtained using ±60 degree tilt range with a 2 -degree step, 2 microns defocus, and 120 e-/Å dose symmetric scheme using Parakeet. The template was used with the ground truth angular assignment before calculating the constrained cross-correlation (CCC). Right: angular distribution colored based on the CCC values showing that side views have higher CCC compared to top views. The nucleosome atomic structure (PDB 2PYO) was oriented similarly to the volume template shown in the middle. B CCC histogram as a result of template matching using PyTom on an experimental tomogram. C Illustration ba sed on combining the results of A and B showing that thresholding based on the CCC value can extract only side views at an adequate precision while lowering the threshold increases the number of extracted top views but at the expense of extracting many fal se positives. D Angular distribution of template matching picked particles at different thresholds in A showing that the particles with the highest CCC scores correspond to nucleosome side views. 847 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 47 Extended Data Tab.1: Cryo-ET simulation parameters in Template Learning. 848 The parameters used for domain randomization simulations in this article were based on 849 a combination of the parameters listed in this table. 850 Parameter Values Defocus for close-to-focus VPP [microns] 0, -0.5, -1 Defocus without VPP [microns] -2.5, -3.25, -4 Electron dose [e−/Å2] 75, 150 Tilting range [deg] [-40, 40], [-60, 60] Tilting step [deg] 2, 4 Ice density [g/cm3] 0.9, 1.1 851 Extended Data Tab.2: Parameters used for training DeepFinder on Template 852 Learning simulations. 853 Default hyperparameters used to train DeepFinder for the different particle picking 854 experiments. 855 Parameter Values Number of epochs 100 Number of steps 100 Shifting for data augmentation [vox] 13 Patch size [vox3] 48 Batch size 25 856 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 48

857 1. Schaffer, M. et al. Cryo-focused Ion Beam Sample Preparation for Imaging Vitreous 858 Cells by Cryo-electron Tomography. Bio Protoc 5, (2015). 859 2. Wagner, F. R. et al. Preparing samples from whole cells using focused-ion-beam 860 milling for cryo-electron tomography. Nat. Protoc. 15, 2041–2070 (2020). 861 3. Tacke, S. et al. A streamlined workflow for automated cryo focused ion beam 862 milling. J. Struct. Biol. 213, 107743 (2021). 863 4. Obr, M. et al. Exploring high-resolution cryo-ET and subtomogram averaging 864 capabilities of contemporary DEDs. J. Struct. Biol. 214, 107852 (2022). 865 5. Danev, R. & Baumeister, W. Expanding the boundaries of cryo-EM with phase 866 plates. Curr. Opin. Struct. Biol. 46, 87–94 (2017). 867 6. Okamoto, K. et al. Structural variability and complexity of the giant Pithovirus 868 sibericum particle revealed by high-voltage electron cryo-tomography and energy-869 filtered electron cryo-microscopy. Sci. Rep. 7, 13291 (2017). 870 7. McIntosh, R., Nicastro, D. & Mastronarde, D. New views of cells in 3D: an 871

to electron tomography. Trends Cell Biol. 15, 43–51 (2005). 872 8. Harastani, M., Eltsov, M., Leforestier, A. & Jonic, S. TomoFlow: Analysis of 873 Continuous Conformational Variability of Macromolecules in Cryogenic 874 Subtomograms based on 3D Dense Optical Flow. J. Mol. Biol. 434, 167381 (2022). 875 9. Zheng, S. et al. AreTomo: An integrated software package for automated marker-876 free, motion-corrected cryo-electron tomographic alignment and reconstruction. J 877 Struct Biol X 6, 100068 (2022). 878 10. Radermacher, M. Weighted Back-projection Methods. in Electron Tomography: 879 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 49

for Three-Dimensional Visualization of Structures in the Cell (ed. Frank, 880 J.) 245–273 (Springer New York, New York, NY, 2006). 881 11. Boisset, N. et al. Overabundant single-particle electron microscope views induce a 882 three-dimensional reconstruction artifact. Ultramicroscopy 74, 201–207 (1998). 883 12. Zhai, X. et al. LoTToR: An Algorithm for Missing-Wedge Correction of the Low-Tilt 884 Tomographic 3D Reconstruction of a Single-Molecule Structure. Sci. Rep. 10, 885 10489 (2020). 886 13. Moebel, E. & Kervrann, C. A Monte Carlo framework for missing wedge restoration 887 and noise removal in cryo-electron tomography. J Struct Biol X 4, 100013 (2020). 888 14. Guesdon, A., Blestel, S., Kervrann, C. & Chrétien, D. Single versus dual-axis cryo-889 electron tomography of microtubules assembled in vitro: limits and perspectives. J. 890 Struct. Biol. 181, 169–178 (2013). 891 15. Bharat, T. A. M. & Scheres, S. H. W. Resolving macromolecular structures from 892 electron cryo-tomography data using subtomogram averaging in RELION. Nat. 893 Protoc. 11, 2054–2065 (2016). 894 16. Wan, W. & Briggs, J. A. G. Cryo-Electron Tomography and Subtomogram 895 Averaging. Methods Enzymol. 579, 329–367 (2016). 896 17. Best, C., Nickell, S. & Baumeister, W. Localization of protein complexes by pattern 897 recognition. Methods Cell Biol. 79, 615–638 (2007). 898 18. Hrabe, T. et al. PyTom: a python-based toolbox for localization of macromolecules 899 in cryo-electron tomograms and subtomogram analysis. J. Struct. Biol. 178, 177–900 188 (2012). 901 19. Moebel, E. et al. Deep learning improves macromolecule identification in 3D 902 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 50 cellular cryo-electron tomograms. Nat. Methods 18, 1386–1394 (2021). 903 20. de Teresa-Trueba, I. et al. Convolutional networks for supervised mining of 904 molecular patterns within cellular context. Nat. Methods 20, 284–294 (2023). 905 21. Hao, Y. et al. VP-Detector: A 3D multi-scale dense convolutional neural network for 906 macromolecule localization and classification in cryo-electron tomograms. Comput. 907

Programs Biomed. 221, 106871 (2022). 908 22. Genthe, E. et al. PickYOLO: Fast deep learning particle detector for annotation of 909 cryo electron tomograms. J. Struct. Biol. 215, 107990 (2023). 910 23. Zeng, X. et al. High-throughput cryo-ET structural pattern mining by unsupervised 911 deep iterative subtomogram clustering. Proc. Natl. Acad. Sci. U. S. A. 120, 912 e2213149120 (2023). 913 24. Martinez-Sanchez, A. et al. Template-free detection and classification of 914 membrane-bound complexes in cryo-electron tomograms. Nat. Methods 17, 209–915 216 (2020). 916 25. Gubins, I. et al. SHREC 2020: Classification in cryo-electron tomograms. Comput. 917 Graph. 91, 279–289 (2020). 918 26. Berman, H. M. et al. The Protein Data Bank. Nucleic Acids Res. 28, 235–242 919 (2000). 920 27. Martinez-Sanchez, A., Jasnin, M., Phelippeau, H. & Lamm, L. Simulating the 921 cellular context in synthetic datasets for cryo-electron tomography. bioRxiv 922 2023.05.26.542411 (2023) doi:10.1101/2023.05.26.542411. 923 28. Parkhurst, J. M. et al. Parakeet: a digital twin software pipeline to assess the impact 924 of experimental parameters on tomographic reconstructions for cryo-electron 925 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 51 tomography. Open Biol. 11, 210160 (2021). 926 29. Rodríguez de Francisco, B., Bezault, A., Xu, X.-P., Hanein, D. & Volkmann, N. 927 MEPSi: A tool for simulating tomograms of membrane-embedded proteins. J. 928 Struct. Biol. 214, 107921 (2022). 929 30. Himes, B. & Grigorieff, N. Cryo-TEM simulations of amorphous radiation-sensitive 930 samples using multislice wave propagation. IUCrJ 8, 943–953 (2021). 931 31. Rice, G. et al. TomoTwin: generalized 3D localization of macromolecules in cryo-932 electron tomograms with structural data mining. Nat. Methods 20, 871–880 (2023). 933 32. Wagner, T. et al. SPHIRE-crYOLO is a fast and accurate fully automated particle 934 picker for cryo-EM. Commun Biol 2, 218 (2019). 935 33. Moebel, E. & Kervrann, C. Towards unsupervised classification of macromolecular 936 complexes in cryo electron tomography: Challenges and opportunities. Comput. 937

Programs Biomed. 225, 107017 (2022). 938 34. de Melo, C. M. et al. Next-generation deep learning based on simulators and 939 synthetic data. Trends Cogn. Sci. 26, 174–187 (2022). 940 35. Prakash, A. et al. Structured Domain Randomization: Bridging the Reality Gap by 941 Context-Aware Synthetic Data. in 2019 International Conference on Robotics and 942 Automation (ICRA) 7249–7255 (2019). 943 36. Tobin, J. et al. Domain randomization for transferring deep neural networks from 944 simulation to the real world. in 2017 IEEE/RSJ International Conference on 945 Intelligent Robots and Systems (IROS) 23–30 (2017). 946 37. Bandyopadhyay, H. et al. Cryo-shift: reducing domain shift in cryo-electron 947 subtomograms with unsupervised domain adaptation and randomization. 948 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 52 Bioinformatics 38, 977–984 (2022). 949 38. Bahar, I., Lezon, T. R., Bakan, A. & Shrivastava, I. H. Normal mode analysis of 950 biomolecular structures: functional mechanisms of membrane proteins. Chem. Rev. 951 110, 1463–1497 (2010). 952 39. Valtchev, S. Z. & Wu, J. Domain randomization for neural network classification. J 953 Big Data 8, 94 (2021). 954 40. Purnell, C. et al. Rapid Synthesis of Cryo-ET Data for Training Deep Learning 955 Models. bioRxiv (2023) doi:10.1101/2023.04.28.538636. 956 41. Pei, L., Xu, M., Frazier, Z. & Alber, F. Simulating cryo electron tomograms of 957 crowded cell cytoplasm for assessment of automated particle picking. BMC 958 Bioinformatics 17, 405 (2016). 959 42. Liu, S. et al. A unified framework for packing deformable and non-deformable 960 subcellular structures in crowded cryo-electron tomogram simulation. BMC 961 Bioinformatics 21, 399 (2020). 962 43. Cui, Q., Rong, V., Chen, D. & Matusik, W. Dense, Interlocking-Free and Scalable 963 Spectral Packing of Generic 3D Objects. ACM Trans. Graph. 42, 1–14 (2023). 964 44. Lobato, I. & Van Dyck, D. MULTEM: A new multislice program to perform accurate 965 and fast electron diffraction and imaging simulations using Graphics Processing 966 Units with CUDA. Ultramicroscopy 156, 9–17 (2015). 967 45. Glaeser, R. M. Invited review article: Methods for imaging weak-phase objects in 968 electron microscopy. Rev. Sci. Instrum. 84, 111101 (2013). 969 46. Tang, G. et al. EMAN2: an extensible image processing suite for electron 970 microscopy. J. Struct. Biol. 157, 38–46 (2007). 971 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 53 47. Kremer, J. R., Mastronarde, D. N. & McIntosh, J. R. Computer visualization of 972 three-dimensional image data using IMOD. J. Struct. Biol. 116, 71–76 (1996). 973 48. Che, C., Xian, Z., Zeng, X., Gao, X. & Xu, M. Domain randomization for 974 macromolecule structure classification and segmentation in electron cyro-975 tomograms. in 2019 IEEE International Conference on Bioinformatics and 976 Biomedicine (BIBM) (IEEE, 2019). doi:10.1109/bibm47256.2019.8983110. 977 49. Jonić, S. & Sorzano, C. Ó. S. Coarse-Graining of Volumes for Modeling of 978 Structure and Dynamics in Electron Microscopy: Algorithm to Automatically Control 979 Accuracy of Approximation. IEEE J. Sel. Top. Signal Process. 10, 161–173 (2016). 980 50. Chen, M. & Ludtke, S. J. Deep learning-based mixed-dimensional Gaussian 981 mixture model for characterizing variability in cryo-EM. Nat. Methods 18, 930–936 982 (2021). 983 51. Kucukelbir, A., Sigworth, F. J. & Tagare, H. D. Quantifying the local resolution of 984 cryo-EM density maps. Nat. Methods 11, 63–65 (2014). 985 52. Goddard, T. D. et al. UCSF ChimeraX: Meeting modern challenges in visualization 986 and analysis. Protein Sci. 27, 14–25 (2018). 987 53. Doruker, P., Jernigan, R. L. & Bahar, I. Dynamics of large proteins through 988 hierarchical levels of coarse-grained structures. J. Comput. Chem. 23, 119–127 989 (2002). 990 54. Ma, J. Usefulness and limitations of normal mode analysis in modeling dynamics of 991 biomolecular complexes. Structure 13, 373–380 (2005). 992 55. Harastani, M., Eltsov, M., Leforestier, A. & Jonic, S. HEMNMA-3D: Cryo Electron 993 Tomography Method Based on Normal Mode Analysis to Study Continuous 994 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint Harastani et al., Template Learning 54 Conformational Variability of Macromolecular Complexes. Front Mol Biosci 8, 995 663121 (2021). 996 56. Bakan, A., Meireles, L. M. & Bahar, I. ProDy: protein dynamics inferred from theory 997 and experiments. Bioinformatics 27, 1575–1577 (2011). 998 57. Lewis, C. D. & Laemmli, U. K. Higher order metaphase chromosome structure: 999 evidence for metalloprotein interactions. Cell 29, 171–181 (1982). 1000 58. Mastronarde, D. N. Automated electron microscope tomography using robust 1001 prediction of specimen movements. J. Struct. Biol. 152, 36–51 (2005). 1002 59. Hagen, W. J. H., Wan, W. & Briggs, J. A. G. Implementation of a cryo-electron 1003 tomography tilt-scheme optimized for high resolution subtomogram averaging. J. 1004 Struct. Biol. 197, 191–198 (2017). 1005 60. Zheng, S. Q. et al. MotionCor2: anisotropic correction of beam-induced motion for 1006 improved cryo-electron microscopy. Nat. Methods 14, 331–332 (2017). 1007 61. Zhang, K. Gctf: Real-time CTF determination and correction. J. Struct. Biol. 193, 1–1008 12 (2016). 1009 1010 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted March 20, 2024. ; https://doi.org/10.1101/2024.03.20.585905doi: bioRxiv preprint