Cryo-ET has revolutionized the visualization of macromolecules within their cellular context. Advances in sample preparation techniques now enable routine creation of thin lamellae for imaging. Cryo-ET, coupled with STA, provides unprecedented detail of macromolecular complexes in situ, particularly when combined with structure prediction tools like AlphaFold2. However, a major bottleneck is the accurate localization of macromolecules within the tomograms. The 3D nature of the data, low signal-to-noise ratios, and cellular crowding pose significant challenges. Existing deep learning-based tools often rely on 3D-Unet architectures but lack generalizability—requiring manual annotation and model retraining for each protein. This severely limits their applicability. Template matching, while simpler, offers inferior accuracy. To overcome these limitations, the authors propose TomoTwin, a generalized model that learns a representation of 3D molecular shapes to differentiate between macromolecules based on their structure, eliminating the need for extensive manual annotation and retraining.

The paper reviews existing deep learning-based particle picking methods for cryo-ET, highlighting their limitations in generalizability and the need for extensive manual annotation. It also discusses template matching methods, noting their lower accuracy compared to deep learning approaches. The authors emphasize the need for a method that combines the accuracy of deep learning with the usability and efficiency of template matching. The application of deep metric learning in 2D single-particle cryo-EM is mentioned, demonstrating the potential of this approach for generalizable particle picking. The paper also touches upon existing software for template matching in cryo-ET, such as EMAN2, Dynamo, and PyTom.

TomoTwin employs deep metric learning to create a generalized particle picking model. The model is trained on a diverse set of simulated tomograms generated using a custom pipeline involving TEM Simulator, IMOD, and a set of open-source scripts called 'tem-simulator-scripts'. The training data consists of subvolumes from 120 structurally dissimilar proteins, ranging in size from 30 kDa to 2.7 MDa, plus membranes, noise, and fiducials. The model architecture is a 3D convolutional neural network (CNN) consisting of five convolutional blocks followed by a head network, which transforms each 37 × 37 × 37 subvolume into a 32-length feature vector (embedding). A triplet loss function, utilizing cosine similarity as the distance metric and online semi-hard triplet mining, is used to train the model. Data augmentation techniques like rotation, dropout, translation, and noise addition are employed to enhance generalization. Hyperparameter optimization is conducted using Optuna to identify optimal training parameters. The model offers two workflows for macromolecule localization: a reference-based workflow using a single molecule as a target and a de novo clustering workflow identifying macromolecular structures on a 2D manifold (using UMAP). The software includes functions for embedding, mapping, locating, and picking particles, providing interactive filtering capabilities.

TomoTwin demonstrated high accuracy in particle picking on both simulated and experimental data. On simulated validation tomograms, the median F1 score was 0.88, ranging from 0.76 to 0.98 across all 120 proteins. The model generalized well to unseen proteins, achieving a median F1 score of 0.82 on a generalization tomogram containing proteins not present in the training set. The picking accuracy improved with increased training set size, showing a logarithmic increase in generalization accuracy. The model performed robustly even in densely packed tomograms. On experimental datasets containing mixtures of apoferritin, RhsA, and TcdA1, TomoTwin accurately identified each protein with minimal confounding from the crowded environment. In a cellular tomogram of *Mycoplasma pneumoniae*, TomoTwin successfully picked 70S ribosomes, yielding high-quality 3D classes after subtomogram averaging. Comparison with EMAN2 (a template matching method) revealed TomoTwin's superior picking accuracy and consistency across a wide range of protein sizes. When applied to a tomogram of *C. reinhardtii* pyrenoid, TomoTwin achieved comparable or superior RuBisCO reconstruction resolution to a previously published supervised deep learning approach, but with significantly reduced processing time. The de novo clustering workflow enabled the identification of 70S ribosomes and a potential bacterial RNA polymerase in *Yersinia entomophaga* tomograms, demonstrating its potential for exploring understudied regions of the proteome. Analysis of simulated tomograms with varying tilt range and total dose showed a decrease in performance with reduced tilt range and dose, but overall demonstrated adaptability.

TomoTwin addresses the critical need for a highly accurate and user-friendly particle picking tool in cryo-ET. Its ability to generalize to new proteins without retraining, combined with its high accuracy on both simulated and experimental datasets, significantly improves the efficiency and accessibility of cryo-ET analysis. The superior performance compared to template matching and the reduced processing time compared to supervised deep learning methods make it a valuable tool for high-throughput analysis. The de novo clustering workflow further expands the scope of cryo-ET by enabling the exploration of less-studied proteins. While the method shows high promise, future improvements could include the ability to pick membrane proteins and filaments, and improved performance on lower-abundance proteins.

TomoTwin presents a significant advance in cryo-ET particle picking. It combines the accuracy of deep learning with exceptional usability, eliminating the need for manual annotation and retraining. The two workflows (reference-based and clustering-based) cater to different experimental needs. TomoTwin facilitates high-throughput analysis and opens avenues for exploring understudied aspects of the proteome. Future work could focus on improving the model's performance on membrane proteins and filaments, as well as enhancing its ability to handle low-abundance proteins.

TomoTwin's current limitations include its inability to reliably pick membrane proteins or filaments. The effectiveness of the clustering workflow depends on the protein's abundance in the tomogram. The model's training at a 10 Å pixel size currently prevents differentiation between multiple conformations of the same protein at the particle picking level. Accurate picking of proteins smaller than approximately 150 kDa is currently challenging.