ProPicker

Cryogenic electron tomography (cryo-ET) images cells in a near-native, frozen-hydrated state, producing 3D volumes in which individual macromolecular complexes are visible but buried under severe noise, missing-wedge artifacts, and a crowded cellular background. Converting a tomogram into structural information depends on particle picking: locating every instance of one target complex so that the copies can be extracted and averaged. Established pickers either rely on template matching or on supervised segmentation networks trained separately for each particle, so every new target demands a fresh round of manual annotation and training.

ProPicker reframes particle picking as promptable 3D segmentation. Developed by Simon Wiedemann and Reinhard Heckel at the Technical University of Munich with Zalan Fabian and Mahdi Soltanolkotabi at the University of Southern California, it was posted to bioRxiv in March 2025 and published in the Journal of Structural Biology in 2026. A single pretrained network is conditioned on a prompt — an embedding of a reference subtomogram showing the particle of interest — and emits a segmentation map for that particle alone. Swapping the prompt swaps the target, with no retraining.

The design borrows the promptable-segmentation idea from 2D natural-image models such as the Segment Anything Model and CLIPSeg, and adapts it to the noise regime and volumetric geometry of cryo-ET. Within the cryo-ET catalog it complements representation-learning approaches like CryoSiam and organelle-scale segmentation models such as CryoViT: ProPicker's contribution is a fixed checkpoint that becomes a picker for an arbitrary complex through a single prompt, and that can optionally be adapted into a particle-specific picker when a small amount of labelled data is available.

#Key Features

• Prompt-conditioned detection: A reference subtomogram of the target complex is encoded into a compact embedding that steers the segmentation decoder, so one set of weights serves many particles.
• Zero-shot generalization: The pretrained model detects a range of particles that were not part of its training set, using only a single prompt.
• Optional fine-tuning: When limited annotated data exists, fine-tuning the prompted model into a particle-specific picker outperforms state-of-the-art particle-specific methods trained on the same limited data.
• Fast inference: Prompted picking runs up to an order of magnitude faster than competing prompt-based pickers, processing a tomogram in minutes.
• Released code and checkpoint: The implementation ships with tutorial notebooks for prompt-based picking and for fine-tuning, plus scripts to retrain from scratch.

#Technical Details

ProPicker is a 3D residual U-Net built on the encoder-decoder used in DeepETPicker, with feature maps of 108, 216, 432, and 864 channels, coordinate convolutions, and image-pyramid inputs. The prompt is a 32-dimensional embedding produced by the contrastively trained embedding network from TomoTwin, and it is injected into every decoder residual block by feature-wise linear modulation (FiLM), so the prompt reshapes the decoder's activations rather than being concatenated as an extra channel. The output is a single-channel sigmoid occupancy map, from which particle centres are recovered by clustering.

Training used simulated tomograms: 70 volumes drawn from ten rounds of the TomoTwin simulation set, plus eight SHREC 2021 volumes, with held-out tomograms from each round reserved for validation. Subtomograms of 64³ voxels were extracted at a stride of 32, with up to eight particle classes sampled per tomogram and a prompt drawn at random for each particle. The network was optimized with binary cross-entropy loss and Adam at a learning rate of 1e-2 for 10 epochs, using flip and in-plane rotation augmentation. Evaluation covers both simulated benchmarks and real cellular tomograms from EMPIAR-10988.

#Applications

The model targets the bottleneck between tomogram reconstruction and subtomogram averaging in structural cell biology. A researcher who has identified a complex of interest can supply one example subtomogram and immediately obtain candidate picks across a tomogram series, without annotating a training set or building a template. This is most valuable for exploratory work — surveying which complexes are present in a cellular volume, or picking a target for which no dedicated picker exists — and for laboratories with a handful of annotated tomograms, where fine-tuning the pretrained checkpoint is more data-efficient than training a picker from scratch.

#Impact

ProPicker demonstrates that the prompt-conditioned segmentation paradigm transfers from 2D natural images to noisy 3D cryo-ET volumes, and that a checkpoint trained purely on simulated data generalizes to experimental tomograms. Its practical limits are worth stating plainly: pretraining relies on simulated tomograms rather than experimental ones, prompted accuracy is characterized as close to or on par with — rather than clearly ahead of — specialized pickers, and the largest gains appear in the low-data fine-tuning regime. The code is public on GitHub with worked notebooks, though the repository carries no license file and the pretrained weights are distributed through a file-sharing link rather than a model registry, which constrains reuse and redistribution.