Abstract
C. elegans is an established model organism for studying genetic and drug effects on ageing, many of which are conserved in humans. It is also an important model for basic research, and C. elegans pathologies is a new emerging field. Here we develop a proof-of-principal convolutional neural network-based platform to segment C. elegans and extract features that might be useful for lifespan prediction. We use a dataset of 734 worms tracked throughout their lifespan and classify worms into long-lived and short-lived. We designed WormNet - a convolutional neural network (CNN) to predict the worm lifespan class based on young adult images (day 1 – day 3 old adults) and showed that WormNet, as well as, InceptionV3 CNN can successfully classify lifespan. Based on U-Net architecture we develop HydraNet CNNs which allow segmenting worms accurately into anterior, mid-body and posterior parts. We combine HydraNet segmentation, WormNet prediction and the class activation map approach to determine the segments most important for lifespan classification. Such a tandem segmentation-classification approach shows posterior part of the worm might be more important for classifying long-lived worms. Our approach can be useful for the acceleration of anti-ageing drug discovery and for studying C. elegans pathologies.
Introduction
C. elegans is a cornerstone of aging research because many of its genetic pathways are shared with humans. However, measuring how long a worm lives or how well it moves is traditionally a slow, manual process. While recent machine learning models can predict a worm's age or lifespan, they are often "black boxes"—they might give a correct prediction, but they don't tell researchers why or where they are looking on the worm's body to make that call.
The authors sought to bridge this gap by creating a system that not only predicts longevity and health (motility) but also identifies the specific body parts (anterior, mid-body, or posterior) that contain the most important biological markers for those traits.
Methods
The researchers implemented a two-part "tandem" approach consisting of segmentation and classification modules.
Segmentation (HydraNet)
Based on the popular U-Net architecture, the authors created "HydraNet." Unlike standard models that try to label everything at once, HydraNet uses a single "body" (encoder) with multiple "heads" (decoders). Each head is dedicated to identifying one specific part: the head (anterior), the mid-section, or the tail (posterior). This prevents the model from getting confused where body parts overlap, a common issue in multi-class segmentation.
Classification (WormNet)
They designed a shallow Convolutional Neural Network (CNN) called WormNet to classify worms into "short-lived" vs. "long-lived" and "low-motility" vs. "high-motility" based on images taken early in their adult life (Days 1–3).
The "Tandem" Link
To interpret the results, they used Class Activation Maps (CAMs). These are essentially heatmaps that show which pixels the classifier focused on. By overlaying these heatmaps onto the segments identified by HydraNet, they could statistically calculate which body part was most "important" to the model’s decision.
Datasets
- Primary Dataset: 734 worms tracked longitudinally throughout their entire lifespans, with images captured every 3.5 hours.
- Annotation Dataset: 130 images were manually annotated by humans to create ground-truth masks for the head, mid-body, and tail to train the HydraNet.
Results
Segmentation and Classification Performance
Standard multi-class segmentation often struggles with boundaries. By giving each body part its own dedicated decoder "head," HydraNet achieved a near-perfect Jaccard index of 0.97, significantly outperforming standard multi-class U-Net approaches. For classification, the model could distinguish between long-lived and short-lived worms using images from just the third day of adulthood, a stage where worms often look identical to the human eye.
Key Findings
- Movement Prediction: On Day 3 of adulthood, the model predicted whether a worm would be a "high mover" or "low mover" throughout its life with 88% accuracy.
- Lifespan Prediction: The model achieved 72% accuracy in classifying worms as long-lived vs. short-lived using Day 3 images.
- The Tail Tells the Tale: The most striking finding was that for long-lived worms, the model consistently focused on the posterior (tail) region. Statistical analysis showed that the anterior (head) part was significantly less important for predicting longevity than the mid-body and posterior parts. This suggests that early signs of healthy aging are most visible in the rear of the worm.
Discussion
While the study points to the tail as a predictor, it doesn't yet identify the specific biological feature (e.g., a specific organ like the gut or a change in tissue texture) the AI is seeing. Much of the initial work was done on low-resolution (96x96) images to save computing power; while findings were validated with high-res InceptionV3 models, more high-resolution data could reveal finer morphological details.
With a sample size of 734 worms, deep learning models are prone to overfitting. Future work utilizing larger, automated datasets from "lifespan machines" will be necessary to refine these "biomarkers of aging" and accelerate anti-ageing drug discovery.
Summary updated