Automated pipeline for detecting and orienting microtubule doublets in cryo-electron tomograms of axonemes, with correct polarity assignment for downstream subtomogram averaging in RELION.
Cilia and flagella are built around a highly conserved structure called the axoneme, composed of nine microtubule doublets arranged radially around a central pair. Each doublet consists of an A-tubule and a B-tubule fused together, and the proteins that decorate them repeat at defined periodicities along the axoneme (8, 16, 24, 48, or 96 nm).
To resolve these proteins by cryo-electron tomography (cryo-ET) and subtomogram averaging, it is necessary to:
- Precisely locate each doublet in 3D space
- Sample along its length at regular intervals
- Assign the correct polarity (all doublets pointing in the same direction)
- Export particle positions and orientations for direct import into RELION
Currently, this process is performed manually using IMOD and custom scripts — a bottleneck that limits throughput and introduces inter-user variability, particularly in polarity assignment.
- Automate the detection and tracing of all 9 microtubule doublets in a tomogram
- Correctly assign polarity to each doublet (tip-to-base orientation)
- Sample along each doublet and compute the 3D orientation angles at each position
- Export a RELION-compatible particle star file ready for subtomogram averaging
- Reduce manual intervention to zero for standard axoneme datasets
The pipeline is split into two main stages: interactive label generation and automated 3D segmentation.
Raw tomogram
│
▼
┌─────────────────────┐
│ nnInteractive │ ← sparse user clicks → dense 3D labels
│ (label generation) │
└────────┬────────────┘
│ training masks
▼
┌─────────────────────┐
│ 3D U-Net │ ← learns to segment doublets from labelled data
│ (segmentation) │
└────────┬────────────┘
│ binary segmentation mask
▼
┌─────────────────────┐
│ Post-processing │ ← skeletonisation, polarity assignment,
│ & export │ sampling, angle computation
└────────┬────────────┘
│
▼
RELION star file
nnInteractive is used to generate high-quality 3D training labels from a small number of user interactions (point clicks or scribbles), without requiring manual segmentation of every voxel.
A Biapy 3D U-Net is trained on the nnInteractive-generated labels to produce a fully automated segmentation model.
- Input: raw or deconvolved tomogram (
.mrc) - Output: multi-class segmentation mask — one label per doublet (1–9) + background
- The network learns the appearance of doublets across different tomograms, noise levels, and orientations
- Once trained, inference runs without any user input
From the segmented mask, the following steps are applied automatically:
- Skeletonisation — reduce each segmented doublet to a 1D centreline in 3D
- Polarity assignment — use the known 9-fold symmetry of the axoneme cross-section and the characteristic A/B tubule geometry to assign a consistent tip-to-base orientation to each doublet
- Sampling — place particle positions at regular intervals along each centreline (configurable spacing, e.g. every 8 nm)
- Angle computation — compute the three Euler angles (rot, tilt, psi) at each position to align the doublet to a common reference frame
- RELION export — write a
.starfile with particle coordinates, angles, and tomogram metadata directly importable into RELION 4/5
On the left, a cross-section showing 9 doublets arranged in circle around central pair of microtubules (here the tip of the cilium is pointing into the page). On the rigth, the image of data along the length of the axoneme/microtubule doublets (longitudinal section along the axoneme).
AI4LIFE.mov
Figure 1 | Automated microtubule doublet labelling from cryo-ET tomograms using nnInteractive. (Top) A user provides sparse point clicks on a tomogram cross-section; nnInteractive propagates these into dense 3D segmentation masks covering the microtubule doublets of the axoneme. Labels are generated in a fraction of the time required for manual annotation and serve as training data for the downstream 3D U-Net segmentation model. (Bottom) Longitudinal slice through the tomogram showing four individually colour-coded microtubule doublets (red, teal, yellow, and light green) traced along their full length. Each colour corresponds to a distinct doublet identity, demonstrating that the labelling correctly separates adjacent doublets running in parallel — a critical prerequisite for polarity assignment and downstream subtomogram averaging in RELION.
Figure 2 | Cross-sectional segmentation of axoneme microtubule doublets by the 3D U-Net. Tomogram cross-section showing the predicted segmentation mask (green) overlaid on the raw cryo-ET data. The 3D U-Net correctly identifies the nine microtubule doublets arranged in the characteristic ring geometry of the axoneme, as well as the central pair of microtubules. The asymmetric distribution of segmented doublets around the ring reflects the inherent structural polarity of the axoneme — a key feature that the model learns to capture and that enables downstream polarity assignment without manual intervention.
