Skip to content

Trained Models

CryoSiam is distributed with a collection of pretrained models that enable out-of-the-box denoising, segmentation, particle identification, and subtomogram embedding generation for cryo-electron tomography (cryo-ET) data.

All pretrained models are hosted on Hugging Face and can be downloaded individually as .ckpt files. Using pretrained weights is strongly recommended for inference and for fine-tuning on new datasets, as it significantly reduces training time and improves convergence.

Model repository:
All models listed below are available at
https://huggingface.co/frosinastojanovska/cryosiam_v1.0

Overview of available models

Model file Task Used in module Description
cryosiam_denoising.ckpt Denoising Denoising Self-supervised denoising model trained on simulated WBP tomograms to reduce noise while preserving structural details.
cryosiam_lamella.ckpt Lamella prediction Semantic segmentation Binary segmentation model for identifying lamella regions and suppressing false positives outside the lamella.
cryosiam_semantic_segmentation.ckpt Semantic segmentation Semantic segmentation Multi-class voxel-wise semantic segmentation of biological structures in denoised tomograms.
cryosiam_ribosome_segmentation.ckpt Semantic segmentation Semantic segmentation Specialized semantic model trained for ribosome segmentation.
cryosiam_semantic_myco_candidates.ckpt Particle identification Particle identification Semantic model trained to predict candidate macromolecular complexes for particle picking.
cryosiam_instance.ckpt Instance segmentation Instance segmentation Instance segmentation model for separating individual macromolecular complexes.
cryosiam_instance_convex_hull.ckpt Instance segmentation Instance segmentation Variant of the instance model using convex-hull–based instance representations.
dense_simsiam_pretrained.ckpt Self-supervised pretraining Semantic training Dense SimSiam backbone pretrained for initializing semantic segmentation training.
simsiam_embeds_denoised_convex_hull.ckpt Subtomogram embeddings Subtomogram embeddings Embedding model using convex-hull masking (masking_type: 1), recommended default for instance-based embeddings.
simsiam_embeds_denoised_no_masking.ckpt Subtomogram embeddings Subtomogram embeddings Embedding model without masking (masking_type: 0), required for center-based embeddings.
simsiam_embeds_denoised_strict.ckpt Subtomogram embeddings Subtomogram embeddings Embedding model with strict instance masking (masking_type: 2) for isolating internal object structure.

Choosing the right model

Denoising

Use cryosiam_denoising.ckpt to preprocess raw WBP tomograms before any downstream task.

Semantic segmentation

  • Use cryosiam_semantic_segmentation.ckpt for general multi-class segmentation.
  • Use cryosiam_ribosome_segmentation.ckpt for ribosome-focused workflows.
  • Always combine with cryosiam_lamella.ckpt for lamella masking when available.

Particle identification

Use cryosiam_semantic_myco_candidates.ckpt to predict candidate regions and extract particle centers.

Instance segmentation

Use cryosiam_instance.ckpt for general instance segmentation. The convex hull variant can be used for alternative instance representations.

Subtomogram embeddings

Select the embedding model based on the masking strategy:

  • masking_type: 0 → simsiam_embeds_denoised_no_masking.ckpt (required for center-based embeddings)
  • masking_type: 1 → simsiam_embeds_denoised_convex_hull.ckpt (recommended default)
  • masking_type: 2 → simsiam_embeds_denoised_strict.ckpt

Important:
When using simsiam_embeddings_from_centers_predict, you must use the no-masking embedding model and provide a patch size close to the expected particle size to avoid embedding background signal.

Using pretrained models

After downloading a model checkpoint, reference it in your configuration file:

trained_model: /path/to/model.ckpt

CryoSiam will automatically load the weights and configure the network accordingly.

Fine-tuning and reproducibility

  • All models can be fine-tuned on custom datasets using the semantic segmentation training pipeline.
  • For reproducibility, record:
    • model filename
    • CryoSiam version
    • configuration file used for inference or training

Citation

If you use CryoSiam pretrained models in your work, please cite:

Stojanovska et al.
CryoSiam: self-supervised representation learning for automated cryo-ET analysis
bioRxiv (2025)

Contributing trained models

If you have trained your own semantic segmentation model using CryoSiam and would like to make it available to the community, you are very welcome to contribute it.

Please contact Frosina Stojanovska with: - a short description of the model (task, classes, data type), - the CryoSiam version used, - and a link to the trained checkpoint.

After review, the model can be added to the official Hugging Face repository and listed on this page.