Skip to content

Workflows

CryoSiam workflows are modular pipelines designed to process cryo-electron tomography (cryo-ET) data step by step.
Each workflow consists of one or more modules, each addressing a specific task such as denoising or segmentation.

You can use individual modules independently or combine them into complete analysis pipelines depending on your scientific question.

For methodological details, please refer to the CryoSiam preprint.

How to choose a workflow

If you are new to CryoSiam, consider the following guidelines:

  • Use Workflow 1 for general denoising and semantic segmentation (or particle identification) of tomograms
  • Use Workflow 1b to train a new semantic segmentation model
  • Use Workflow 2 when individual instances of structures need to be separated
  • Use Workflow 3 for downstream analysis of subtomograms such as clustering or similarity comparisons

Workflow 1

Denoising → Semantic Segmentation / Particle Identification

This workflow is suitable for interpreting cellular context or detecting specific structures in cryo-ET data.

  1. Denoising
    Reduce noise in the raw tomogram while preserving structural features.

  2. Semantic Segmentation
    Assign each voxel to a biological class (e.g., membranes, filaments, complexes).

Alternative:
Particle Identification can be used directly after denoising to locate specific particles of interest.

See detailed documentation:

Workflow 1b

Denoising → Semantic Segmentation (Training)

This workflow is used to train a semantic segmentation model on cryo-ET data using CryoSiam’s self-supervised pretrained weights and user-provided ground truth data.

It is intended for users who want to adapt CryoSiam to new datasets or biological contexts.

  1. Denoising
    Preprocess raw tomograms to improve signal quality.

  2. Semantic Segmentation (Training)
    Train (fine-tune) a semantic segmentation model using the self-supervised trained weights as starting point.

After training, the resulting model can be used for inference as described in Workflow 1.

See detailed documentation:

Workflow 2

Denoising → Instance Segmentation

This workflow focuses on separating individual particles.

  1. Denoising
    Improve signal quality for reliable downstream processing.

  2. Instance Segmentation
    Identify and separate individual instances of particles.

See detailed documentation:

Workflow 3

Denoising → (Optional) Instance Segmentation → Subtomogram Embeddings

This workflow enables quantitative downstream analysis of extracted subtomograms.

  1. Denoising
    Preprocess tomograms for improved structural information.

  2. Instance Segmentation (optional)
    Detect and separate candidate particles.

Alternative:
Instance segmentation can be skipped if particle centers are provided; see Subtomogram Embeddings for details.

  1. Subtomogram Embeddings
    Represent particle subtomograms as feature vectors for clustering, comparison, or further analysis.

See detailed documentation:

Trained Models

Pre-trained models available for CryoSiam workflows are described here:

Running a workflow

Each CryoSiam module is configured using a YAML configuration file that defines inputs, outputs, and model parameters.

To run a module, use:

cryosiam <module name> --config_file configs/<module>.yaml

Detailed explanations of configuration options are provided on the documentation page of each module.

Running CryoSiam at Scale (Slurm / HPC)

CryoSiam prediction commands run on one GPU per process. To efficiently process large datasets on HPC systems, predictions should be parallelized across tomograms using Slurm job submission.

Instead of processing all tomograms in a single command, CryoSiam allows processing one tomogram per job via the --filename argument.

This enables:

  • Full utilization of multi-GPU clusters
  • Independent job retries
  • Clean scaling to hundreds of tomograms

For more details see: Scaling to large datasets on HPC