Using image classifiers to predict CMT2A disease-relevant mitochondrial motility phenotypes in iPSC motor neurons

This paper presents a vision transformer-based classification framework that utilizes kymographs to more accurately predict CMT2A disease phenotypes in iPSC motor neurons compared to traditional summary statistics, thereby enabling scalable computational screening for mitochondrial trafficking defects.

The Gist

The Big Picture: A Traffic Jam in the Cell's Highway

Imagine your body is a giant city, and inside every cell, there are tiny delivery trucks called mitochondria. These trucks carry energy to keep the cell running. They travel along long roads called axons (which are like the cell's nerve fibers).

In a healthy person, these trucks move smoothly, stopping only when necessary. But in a disease called Charcot-Marie-Tooth type 2A (CMT2A), a genetic glitch causes these trucks to get confused. They might stop too much, jitter in place, or move erratically. This leads to the nerve cells dying, causing muscle weakness and trouble walking.

The scientists in this paper wanted to find a way to spot this "traffic jam" automatically, so they could test new medicines to fix it.

---

The Problem: Trying to Count Cars with a Stopwatch

Previously, scientists tried to study this disease by watching videos of these trucks and manually counting them or measuring their speed.

• The old way: Imagine trying to count every single car on a busy highway by watching a video and writing down "Car 1: 50mph, Car 2: 48mph." It's slow, tedious, and if you miss one car or misread the speedometer, your whole study is wrong.
• The paper's finding: When the researchers tried to use simple math (like "average speed" or "how often they stop") to tell the difference between healthy and sick cells, it didn't work well. The sick cells didn't always move slower; sometimes they moved faster but in a weird, jittery way that simple math couldn't catch.

The Solution: The "Smart Camera" (AI)

Instead of counting cars, the researchers decided to let a super-smart AI camera look at the whole picture.

1. Turning Video into a Map (Kymographs):
   First, they took the long, boring videos of the trucks moving and turned them into a special kind of map called a kymograph.

   • Analogy: Imagine taking a photo of a highway, but instead of showing the cars at one moment, the photo stretches out over time. The horizontal line is the road, and the vertical line is time. A truck moving looks like a diagonal line; a truck stopped looks like a straight vertical line. It turns a 3D movie into a 2D picture.

2. The AI Detective (Vision Transformer):
   They fed these maps into a powerful AI called a Vision Transformer (ViT). Think of this AI as a detective who has seen millions of pictures of traffic. It doesn't just look at one car; it looks at the whole pattern of the road.

   • The AI learned to look at the texture of the lines on the map. It realized that sick cells have a specific "fingerprint" of movement that simple math misses.

3. The Result:
   The AI was incredibly good at telling the difference. It could look at a map and say, "This is a healthy cell," or "This is a sick cell," with over 85-90% accuracy. Even better, it worked on new data it had never seen before, proving it actually learned the disease, not just memorized the pictures.

The Discovery: The "Jitter" Phenomenon

The coolest part of the paper is what the AI "saw" that humans missed.

The researchers asked the AI: "What exactly are you looking at to make that decision?"

The AI pointed to a specific type of movement: Jitter.

• Healthy Trucks: When they stop, they sit perfectly still.
• Sick Trucks (CMT2A): When they stop, they vibrate or shake back and forth slightly, like a phone on "vibrate" mode.

The AI realized that this tiny, back-and-forth "jitter" in stationary trucks was the biggest sign of the disease. Previous methods were too slow to see this tiny vibration, but the AI saw it clearly.

Why This Matters

This is a game-changer for finding cures.

• Before: Testing new drugs was like trying to find a needle in a haystack by looking at one needle at a time. It was too slow.
• Now: With this AI, scientists can screen thousands of potential drugs quickly. They can treat a cell with a drug, take a picture, run it through the AI, and instantly know: "Did the drug stop the jitter? Yes! This drug might work."

Summary

The paper is about teaching a computer to look at the "traffic patterns" inside nerve cells. Instead of counting cars, the computer learned to recognize the unique "shaking" pattern of sick cells. This allows scientists to find cures much faster than ever before.

Technical Summary

1. Problem Statement

Charcot-Marie-Tooth disease type 2A (CMT2A) is a hereditary neuropathy caused by mutations in the MFN2 gene, leading to dysregulated mitochondrial trafficking along axons. While induced pluripotent stem cell (iPSC) derived motor neurons offer a scalable model for drug screening, a critical bottleneck exists: the lack of robust, scalable computational methods to quantify mitochondrial motility phenotypes.

Traditional approaches rely on:

• Manual tracing: Labor-intensive and not scalable for high-throughput screening.
• Automated track segmentation (e.g., KymoButler): While faster, these methods extract summary statistics (speed, pausing, direction changes) that often fail to robustly distinguish between healthy (Wild Type, WT) and diseased (R364W mutant) genotypes due to segmentation errors and the loss of subtle dynamic features (e.g., mitochondrial "jitter" or texture).

The authors aim to develop a computational framework that bypasses the limitations of manual feature extraction by directly analyzing raw kymograph images to predict disease status and interpret underlying biological mechanisms.

2. Methodology

Data Generation and Preprocessing

• Model System: The study used human iPSC-derived motor neurons with two genotypes: MFN2 Wild Type (WT) and MFN2 R364W (heterozygous mutant).
• Imaging: Time-lapse fluorescence microscopy (2-second intervals for 3 minutes) captured axons (green channel) and mitochondria (red channel).
• Kymograph Construction:
  • Axons were segmented from the green channel to create masks.
  • Mitochondrial intensity along these axon segments was extracted over time to generate kymographs (2D images where the x-axis is axonal position and the y-axis is time).
  • Filtering: Only axon segments >62µm were retained to ensure sufficient data and reduce noise.

Baseline Approach: Automated Track Segmentation

• The authors first applied KymoButler to automatically segment mitochondrial tracks.
• They calculated standard metrics: motility percentage (thresholded at 0.1 µm/s), speed, pause duration, direction changes, and acceleration.
• Result: These summary statistics showed significant overlap between WT and R364W distributions and failed to accurately predict genotype when tested on held-out biological replicates.

Proposed Approach: Vision Transformer (ViT) Classification

• Architecture: The authors utilized a Vision Transformer (ViT) based on the DINOv3 backbone (specifically dinov3_vitl16_pretrain_sat463m), which was pre-trained on large-scale diverse image datasets.
• Input Processing:
  • Kymographs were divided into overlapping 272×272 pixel tiles.
  • Tiles were upsampled via bicubic interpolation.
  • Low-variance tiles (likely empty background) were filtered out.
• Training Strategy:
  • Encoder: The DINOv3 backbone was frozen (weights fixed) to leverage pre-trained features and prevent overfitting on the small dataset.
  • Classifier: A trainable linear classification head (logistic regression) was attached to the frozen encoder to predict the genotype (WT vs. R364W) for each tile.
  • Data Split: Biological Replicate 1 (BR1) was used for training and testing (80/20 split), while Biological Replicate 2 (BR2) was held out entirely as a rigorous test set to ensure generalizability across experimental batches.

Interpretability and Analysis

• Patch Embeddings: The authors analyzed the 1,024-dimensional patch embeddings generated by the ViT.
• Clustering: Principal Component Analysis (PCA) reduced dimensionality, followed by K-means clustering (K=5) to group patches with similar visual and dynamic properties.
• Validation: Cluster definitions were cross-referenced with KymoButler track metrics (velocity, acceleration) and visual inspection.
• Spatial/Temporal Analysis: The authors computed neighbor enrichment matrices to analyze the spatial and temporal configurations of patch clusters (e.g., how often a "moving" patch is adjacent to a "stationary" patch).

3. Key Contributions

1. Demonstration of Feature Limitation: The study provides empirical evidence that traditional summary statistics (speed, pausing) derived from automated track segmentation are insufficient for distinguishing MFN2 R364W genotypes, highlighting the need for holistic image analysis.
2. ViT-Based Phenotyping: Development of a robust, high-throughput image classifier that achieves >87% accuracy in distinguishing WT from R364W kymographs, even on a completely independent biological replicate (BR2).
3. Discovery of Novel Phenotypes: Through ViT interpretability, the authors identified a previously unquantified disease phenotype: "jitter" in stationary mitochondria.
   • C0 Cluster: Stationary mitochondria with small-amplitude back-and-forth displacement (jitter), enriched in R364W.
   • C1 Cluster: Stationary mitochondria that are relatively still, enriched in WT.
4. Spatiotemporal Dynamics: The analysis revealed that R364W axons exhibit distinct spatiotemporal patterns, such as increased fusion/fission events between moving and jittery stationary mitochondria, which are not captured by standard track segmentation.

4. Key Results

• Classification Performance:
  • Tile-Level Accuracy: ~87–88% on both test and holdout sets.
  • Well-Level Accuracy: 100% on the holdout set (BR2), indicating the model successfully aggregates tile-level predictions to correctly identify the genotype of the entire sample.
  • Robustness: The model maintained high performance when the training/holdout sets were swapped (BR2 trained, BR1 held out), confirming it learned biological signals rather than technical batch artifacts.
• Failure of Traditional Metrics:
  • KymoButler-derived motility percentages and speeds showed no significant difference between genotypes in some replicates and contradictory trends in others.
  • A logistic regression classifier trained on motility percentages failed to predict the holdout set (F1 score = 0.67).
• Biological Insights from Clusters:
  • Stationary Jitter (C0): R364W mitochondria showed significantly higher mean absolute acceleration and direction changes even when classified as "stationary" (speed < 0.1 µm/s). This "jitter" is likely related to defective mitochondrial tethering or fusion/fission dynamics.
  • Spatial Organization: R364W axons showed a depletion of "moving mitochondria adjacent to moving mitochondria" (C4-C4) compared to WT, suggesting disrupted coordinated transport.
  • Temporal Stability: Stationary jitter patches (C0) were less temporally stable in R364W axons, frequently switching positions or states compared to WT.

5. Significance and Implications

• Scalable Drug Screening: This approach enables the automated, high-throughput scoring of disease phenotypes in iPSC models without the need for manual annotation or perfect track segmentation. This is crucial for screening large libraries of potential therapeutics for CMT2A.
• Beyond Summary Statistics: The study demonstrates that deep learning models can capture subtle, high-dimensional features (like texture, jitter, and complex spatiotemporal patterns) that are invisible to traditional statistical summaries.
• New Biological Understanding: The identification of "jitter" as a specific R364W phenotype provides a new mechanistic hypothesis: the mutation may cause mitochondria to be tethered loosely or undergo rapid, futile fission/fusion cycles, leading to micro-movements even when they appear stationary.
• Generalizability: The framework is adaptable. While trained on a specific protocol, the architecture can be retrained on other iPSC lines, differentiation protocols, or even other organelle trafficking diseases, provided the kymograph format is maintained.

In conclusion, this paper establishes a new paradigm for analyzing mitochondrial dynamics, shifting from error-prone manual/semi-automated feature extraction to robust, interpretable deep learning models that reveal novel disease mechanisms.