Cilia SubQ, a modular suite of pipelines for automated analysis of primary cilia and ciliary subdomains

The paper introduces Cilia SubQ, a modular, machine-learning-based pipeline suite for ZEISS arivis Pro that enables rapid, high-throughput, and reproducible automated segmentation and quantification of primary cilia and their subdomains, significantly reducing analysis time and user bias.

The Gist

Imagine your cells are like tiny, bustling cities. Protruding from the surface of most of these cities is a single, hair-like antenna called a primary cilium. Think of this antenna as the cell's "smartphone" or "weather station." It reaches out to catch signals from the outside world (like growth factors or chemical cues) and translates them into instructions for the cell to build, repair, or move.

However, this antenna isn't just a simple stick. It has very specific neighborhoods:

• The Base (Basal Body): The root where the antenna is planted.
• The Gate (Transition Zone): A security checkpoint that decides what gets in and out.
• The Tip: The very end where important messages are processed.

The Problem:
For scientists, studying these antennas is a nightmare. They are incredibly small (microscopic), and looking at them under a microscope is like trying to count grains of sand on a beach while wearing thick gloves. To understand diseases related to these antennas (called ciliopathies), researchers have to manually trace, measure, and count thousands of these tiny structures. It takes forever, is boring, and humans make mistakes (we get tired, or we see what we want to see).

The Solution: Cilia SubQ
This paper introduces Cilia SubQ, a new "digital toolkit" that acts like a super-smart, tireless robot assistant for scientists. It's a suite of automated pipelines (a set of step-by-step instructions) designed to do all the hard work of analyzing these cellular antennas.

Here is how it works, using some fun analogies:

1. The "Eye" that Learns: Cilia.AI

The heart of the toolkit is a machine-learning model called Cilia.AI.

• The Analogy: Imagine teaching a dog to find a specific type of ball in a giant park. At first, the dog might confuse a rock for a ball. But if you show it thousands of pictures of the ball and say, "Yes, that's it," and "No, that's a rock," the dog eventually becomes an expert.
• In the Paper: The scientists trained Cilia.AI on thousands of images of cells. They taught it to recognize the primary cilium (the antenna) even when it's bent, broken, or glowing dimly. Once trained, this AI can scan a whole batch of microscope images in minutes and find the antennas with 90%+ accuracy, doing in seconds what would take a human hours.

2. The "Neighborhood Watch": SubQ Pipelines

Once the AI finds the antenna, the toolkit breaks it down into its specific neighborhoods using different "pipelines" (specialized routines):

• SubQ_BB_DC (The Root Finder): This routine looks at the base of the antenna. It distinguishes between the "mother" centriole (the one holding the antenna) and the "daughter" centriole (the younger sibling next to it).
  • Analogy: It's like a detective who can look at a tree and tell you exactly which root is the main trunk and which is a smaller side root, even if they are tangled together.
• SubQ_TZ (The Gatekeeper): This routine zooms in on the "Transition Zone," the security gate. It counts the specific proteins that act as the gate.
  • Analogy: It's like a toll booth operator who automatically counts every car passing through a specific gate, ignoring the cars parked in the parking lot nearby.
• SubQ_CT (The Tip Tracker): This looks at the very end of the antenna.
  • Analogy: It's like a lighthouse keeper who only cares about the light at the very top of the tower, ignoring the rest of the building.

3. The "Traffic Cam": Kymographs

The toolkit also includes a way to watch things move inside the antenna.

• The Analogy: Imagine taking a long-exposure photo of a highway at night. You don't see the cars; you see long streaks of light showing where the traffic is going. This is called a kymograph.
• In the Paper: The toolkit takes video of proteins zooming up and down the antenna (like a train on a track) and turns it into a clear map showing the speed and direction of the traffic. This helps scientists see if the "delivery trucks" inside the cell are moving correctly.

Why is this a Big Deal?

• Speed: The paper says this toolkit makes the analysis 8 times faster. What used to take a human a whole day of staring at a screen now takes a few minutes of computer time.
• Fairness: Humans get tired and might accidentally measure things differently on Tuesday than on Friday. The robot is consistent every single time.
• Accessibility: The scientists didn't just keep this to themselves. They put all the "recipes" (the code and instructions) on a public website (Open Science Framework) so any scientist with a microscope can download it and start using it immediately.

In Summary:
Cilia SubQ is like giving scientists a pair of super-vision glasses and a team of robot assistants. Instead of spending years manually counting tiny cellular antennas, they can now let the AI do the heavy lifting, allowing them to focus on the big picture: understanding how these antennas work and how to fix them when they break in diseases.

Technical Summary

1. Problem Statement

Primary cilia are antenna-like organelles critical for embryonic development and tissue homeostasis. Their dysfunction leads to "ciliopathies," a broad class of developmental disorders. Analyzing these structures is challenging due to:

• Scale: Primary cilia are small (a few microns), and their subdomains (basal body, transition zone, ciliary tip) are submicron (<1 µm).
• Complexity: Manual quantification of ciliary morphology, length, and protein distribution is labor-intensive, repetitive, and prone to user bias.
• Limitations of Existing Tools: While tools exist for general cilia detection (e.g., CiliaQ, detectCilia), they lack robust, automated capabilities for analyzing specific ciliary subdomains (basal body, transition zone, tip) and intraflagellar transport (IFT) dynamics in a unified, high-throughput workflow.

2. Methodology

The authors developed Cilia SubQ, a modular suite of pipelines built on ZEISS arivis Pro software, integrated with a custom machine-learning model named Cilia.AI.

A. Cilia.AI (The Core Detection Model)

• Training: Trained using the ZEISS arivis Cloud on 8-bit maximum intensity projection (MIP) images from diverse datasets, including wild-type and mutant Mouse Embryonic Fibroblasts (MEFs), Retinal Pigment Epithelial (RPE-1) cells, human skin fibroblasts, and mouse brain tissue sections.
• Data Diversity: The training set included cells with normal, short (siRNA knockdown), and long (knockout) cilia, as well as cells with activated Sonic Hedgehog (SHH) signaling, to ensure robustness against morphological variations.
• Optimization: The model underwent 35 rounds of training (32 versions), annotating 1,819 cilia and 3,229 background regions.
• Post-Processing: To minimize false positives, the pipeline includes:
  1. Object-feature filtering: Excluding objects below a specific size threshold.
  2. Proximity filtering: Restricting selections to objects near the basal body.

B. Modular Pipelines

The suite consists of specific pipelines for different analytical needs:

1. SubQ_BB_DC (Basal Body & Daughter Centriole):
   • Uses Cilia.AI to locate the cilium, then identifies the nearest centrosomal objects (using markers like $\gamma$-tubulin or Centrin-2).
   • Distinguishes the Basal Body (mother centriole) from the Daughter Centriole based on proximity to the cilium.
   • Adaptable parameters (diameter, probability threshold) allow it to work with closely apposed centrioles in RPE-1 cells.
2. SubQ_TZ (Transition Zone):
   • Detects transition zone markers (e.g., AHI1, TMEM67) using a "Blob Finder."
   • Applies a multi-step filtering strategy: expanding the pericentriolar material (PCM) mask, intersecting it with the cilium, and retaining only signals touching the "transition zone area" and closest to the basal body.
3. SubQ_CT (Ciliary Tip):
   • Identifies tip markers (e.g., KIF7, IFT81).
   • Uses a "parent-child" relationship to associate tips with cilia.
   • Implements iterative distance-based filtering to remove proximal non-tip signals, ensuring only the distal tip is quantified.
4. SubQ_Length:
   • Exports Cilia.AI segmented cilia as masked images.
   • Uses a custom FIJI (ImageJ) script to skeletonize the cilia and calculate length, handling branching artifacts.
5. SubQ_Kymo (Trafficking):
   • A semi-manual workflow for live-cell imaging.
   • Involves cropping, rotating, and annotating cilia in time-lapse videos, followed by a Python script to generate kymographs for visualizing IFT trafficking (anterograde/retrograde).

3. Key Contributions

• First Integrated Suite for Subdomains: Unlike previous tools focused on whole-cilium metrics, Cilia SubQ provides a unified framework for quantifying the basal body, transition zone, and ciliary tip simultaneously.
• Machine Learning Integration: The Cilia.AI model achieves high detection accuracy (90–97%) across diverse cell lines and tissue sections, significantly reducing the need for manual tracing.
• Modularity and Flexibility: The pipelines are designed to be adaptable. Users can adjust parameters (e.g., blob finder sensitivity) to accommodate different markers (e.g., switching from $\gamma$-tubulin to Centrin-2) or staining conditions.
• Open Resources: All pipeline files (.zpipeline), trained models (.czann), export scripts, and video tutorials are publicly available via the Open Science Framework (OSF).

4. Results

• Detection Accuracy:
  • Cilia: Cilia.AI achieved >90% detection success across MEFs, RPE-1, and human fibroblasts, comparable to manual annotation. It successfully detected cilia in mouse brain tissue (69% raw, 87% with partial correction).
  • Basal Body: The SubQ_BB_DC pipeline correctly identified 92.6% of basal bodies in MEFs and 82.4% in RPE-1 cells (using Centrin-2) without manual correction.
  • Transition Zone: The SubQ_TZ pipeline achieved 99.5% accuracy for AHI1 and 98.75% for TMEM67 after minimal correction.
  • Ciliary Tip: The SubQ_CT pipeline achieved ~82% accuracy for KIF7 and 83% for IFT81 tip detection.
• Quantitative Reliability: Fluorescence intensity and projected area measurements generated by the pipelines showed no statistically significant difference compared to expert manual quantification.
• Efficiency: The suite reduced hands-on analysis time by approximately 8-fold compared to manual methods.
• Biological Validation: The tool successfully recapitulated known biological phenomena, such as the SAG-dependent reduction of GPR161 levels in cilia, validating its use for signaling studies.

5. Significance

• High-Throughput Ciliopathy Research: Cilia SubQ enables the rapid, unbiased analysis of large datasets, which is essential for studying rare ciliopathies and screening genetic or pharmacological interventions.
• Standardization: By providing a standardized, traceable workflow (using SIS files to store parameters and corrections), the tool reduces inter-operator variability and bias.
• Versatility: The framework bridges the gap between 2D cell culture and complex in vivo tissue analysis, offering a pathway to study cilia in brain sections and other tissues.
• Future-Proofing: The modular design allows researchers to easily integrate new markers or adapt the pipelines for 3D/4D imaging as the technology evolves, filling a critical gap in current ciliary analysis toolkits.

In summary, Cilia SubQ represents a major advancement in cell biology imaging, transforming the analysis of primary cilia from a slow, manual bottleneck into a rapid, automated, and highly precise process.