Development of a microbiome based health score for non-invasive monitoring of farmed Atlantic salmon (Salmo salar)

This study demonstrates that 16S rRNA sequencing of non-invasive urogenital and skin swabs can accurately differentiate between healthy and lesioned Atlantic salmon, enabling the development of novel microbiome-based health scores for proactive disease management in aquaculture.

The Gist

Imagine a massive underwater city where thousands of Atlantic salmon live together. For years, farmers have been like security guards who only notice trouble when they see a fish swimming strangely, looking pale, or having a visible sore. By the time they see these signs, the "disease" has already taken hold, and the whole city is at risk.

This paper introduces a new, smarter way to be a security guard: listening to the city's background noise before the alarm even rings.

Here is the story of how the researchers did it, explained simply:

1. The "Garden" Inside the Fish

Think of every salmon as a walking garden. Inside their guts and on their skin, there are trillions of tiny invisible plants and animals (bacteria, viruses, fungi). This is called the microbiome.

• Healthy Fish: Their garden is like a well-tended park. It's diverse, balanced, and the plants are happy.
• Sick Fish: Their garden is in chaos. Some bad weeds are taking over, and the good plants are struggling.

The problem? To check the garden inside the fish (the gut), you usually have to kill the fish. To check the gills, you have to grab the fish, which stresses them out. Farmers need a way to check the garden without hurting the fish.

2. The "Proxy" Trick: Peeking Through the Window

The researchers had a brilliant idea: What if the outside of the fish tells us what's happening on the inside?

They tested a hypothesis that sounds like a detective's theory:

• The Skin is the Window to the Gills: The skin and the gills are both constantly touching the same water. The researchers found that the "garden" on the skin looks almost exactly like the "garden" on the gills. So, swabbing the skin is like looking through a window to see the gills.
• The "Back Door" is the Window to the Gut: The fish's urogenital opening (a small hole near the tail) is right next to the exit of the digestive tract. The researchers found that swabbing this tiny hole gives them a perfect snapshot of the gut microbiome.

The Analogy: Imagine you want to know if a house has a leaky pipe. Instead of tearing down the walls (killing the fish), you just smell the air coming out of the drain (swabbing the urogenital pore). If it smells like mold, you know there's a problem inside, even if the walls look fine.

3. The "Health Score" Calculator

Once they collected these "window" samples (skin and urogenital swabs), they didn't just look at them with a microscope. They used Artificial Intelligence (Machine Learning) to act like a super-smart translator.

• The Old Way: "Is this fish sick? Yes or No?" (Binary, reactive).
• The New Way: The AI looks at the tiny bacteria patterns and gives the fish a Health Score from 0 to 1.
  • 0.0: Perfectly healthy.
  • 0.5: Something is starting to go wrong (the "Transition Zone").
  • 1.0: The fish is definitely sick.

This is like a car's "Check Engine" light. In the past, the light only came on when the engine was smoking. This new system is like a dashboard that tells you, "Hey, your fuel efficiency is dropping slightly," allowing you to fix it before the engine breaks.

4. What They Found

• The "Anna Karenina" Effect: The researchers noticed that healthy fish all looked very similar to each other (like a happy family). But sick fish were all different from each other; their internal gardens were chaotic and unpredictable. This chaos is a clear sign of trouble.
• The Results: The AI models were incredibly accurate. By just swabbing the skin or the urogenital pore, the computer could predict if a fish was sick with over 90% accuracy.
• The "Transition Zone": The most exciting part is that the score can catch fish that are not yet visibly sick but are starting to get sick. This gives farmers a head start to treat the water or change the feed before a full-blown disease outbreak happens.

Why This Matters

This is a game-changer for sustainable farming.

1. No Killing: You don't have to sacrifice valuable fish to check their health.
2. No Stress: You don't have to grab and handle the fish, which makes them less likely to get sick from stress.
3. Proactive, Not Reactive: Instead of waiting for a disease to spread and kill half the farm, farmers can see the warning signs early and stop it.

In a nutshell: This paper teaches us that by listening to the tiny, invisible "background noise" of bacteria on a fish's skin and tail, we can predict its health before it even shows a single symptom. It turns fish farming from a game of "wait and see" into a game of "predict and prevent."

Technical Summary

1. Problem Statement

Atlantic salmon aquaculture faces significant sustainability challenges due to infectious diseases (e.g., Piscirickettsia salmonis, Tenacibaculum spp.), environmental stressors, and economic constraints. Current health monitoring relies on:

• Reactive Visual Inspection: Identifying clinical signs (lesions, pale gills) only after disease is advanced.
• AI/Computer Vision: Monitoring behavior and feeding, which detects stress responses rather than early physiological disruptions.
• Invasive Sampling: Traditional microbiome analysis requires lethal sampling (gut) or stressful handling (gills), making longitudinal monitoring of valuable stock impractical and ethically challenging.

There is a critical need for proactive, non-invasive biomarkers that can detect health deviations before clinical symptoms manifest.

2. Methodology

Study Design & Sampling

• Cohort: 171 Atlantic salmon from a single commercial marine facility, classified by on-site veterinarians into two groups: Healthy ($n=85$) and Lesioned ($n=86$).
• Sampling Sites: Four anatomical niches were sampled to test the hypothesis that non-invasive sites proxy internal health:
  1. Gills (Invasive/Reference for branchial health)
  2. Skin (Non-invasive proxy for gills)
  3. Mucosa/Intestine (Invasive/Reference for gut health)
  4. Urogenital Pore (Non-invasive proxy for gut)
• Processing: Samples were preserved, DNA extracted (ZymoBIOMICS kit), and 16S rRNA gene V3-V4 regions sequenced (Illumina NextSeq 1000).

Bioinformatics & Statistical Analysis

• Pipeline: DADA2 for ASV inference, strict contamination filtering (removing ASVs in negative controls >1%), and rarefaction to 20,000 reads.
• Diversity Analysis:
  • Alpha Diversity: Shannon and Simpson indices (no significant differences found between groups).
  • Beta Diversity: PERMANOVA using Aitchison distance on CLR-transformed data.
• Proxy Validation:
  • Procrustes Analysis: To compare PCA ordinations of paired sites (Skin vs. Gill; Urogenital vs. Mucosa).
  • Mantel Tests: To quantify distance matrix correlations between paired sites.
• Machine Learning (ML) Framework:
  • Algorithms Tested: Random Forest, XGBoost, LightGBM, SVM, Logistic Regression, KNN, Gaussian Naive Bayes.
  • Validation: Nested cross-validation with group blocking by sea cage to prevent data leakage.
  • Model Selection: Multi-Criteria Decision Analysis (MCDA) weighting AUC, Accuracy, MCC, Brier Score, and Kappa.
  • Preprocessing: Total Sum Scaling (TSS) for tree-based models; CLR transformation + z-score + PCA for linear/distance-based models.

Health Score Development

• Two quantitative indices were formulated based on the top-performing Logistic Regression models:
  1. Gut Health Score (GHS): Derived from urogenital pore samples.
  2. Skin Health Score (SHS): Derived from skin samples.
• Definition: Both scores represent the predicted probability ($0$ to $1$) of a fish belonging to the "lesioned" cohort.

3. Key Results

Microbial Composition & Diversity

• Alpha Diversity: No significant differences in Shannon or Simpson indices between healthy and lesioned fish in any tissue. This suggests disease involves microbial replacement rather than a collapse in richness/evenness.
• Beta Diversity: Significant differences in community composition based on clinical status and the interaction between tissue and status.
• Dispersion: Lesioned fish showed significantly higher dispersion (heterogeneity) in gill microbiomes ($p=0.029$), supporting the Anna Karenina Principle (dysbiotic communities are more variable).

Proxy Validation (Non-invasive vs. Invasive)

Strong statistical coupling was confirmed between external and internal sites:

• Skin vs. Gill: Procrustes $r = 0.842$ ($p=0.001$); Mantel $r = 0.52$ ($p=0.001$).
• Urogenital vs. Mucosa: Procrustes $r = 0.701$ ($p=0.001$); Mantel $r = 0.54$ ($p=0.001$).
• Conclusion: Skin swabs accurately reflect gill health, and urogenital swabs accurately reflect gut health.

Machine Learning Performance

• Best Model: Logistic Regression outperformed all other algorithms for both sites.
• Performance Metrics:
  • Skin Model: AUC = 0.97, Accuracy = 0.87.
  • Urogenital Model: AUC = 0.87, Accuracy = 0.78.
• Discriminative Power: The models successfully separated healthy and lesioned cohorts, with a "transition zone" identified (scores 0.4–0.6) representing fish with divergent microbiomes but no overt clinical signs.

4. Key Contributions

1. Validation of Non-Invasive Proxies: The study provides robust statistical evidence that skin and urogenital swabs can serve as reliable, non-invasive surrogates for gill and gut microbiome profiling, respectively.
2. Development of Salmon Microbiome Health Scores (SMHS): The authors created the first formalized, quantitative, multi-parameter health indices for farmed fish. These scores move beyond binary classification to provide a continuous metric of physiological resilience.
3. Shift to Proactive Management: By demonstrating that microbiome shifts precede or accompany clinical signs without requiring lethal sampling, the study enables high-frequency, real-time health monitoring.
4. Methodological Rigor: The use of group-blocked cross-validation and MCDA for model selection ensures the results are robust against overfitting and cage-specific biases.

5. Significance & Implications

• Animal Welfare: Eliminates the need for lethal gut sampling and reduces stress associated with gill handling, aligning with sustainable aquaculture practices.
• Early Warning System: The continuous nature of the health scores allows farmers to detect "transition states" (microbial dysbiosis) before visible lesions appear, enabling early intervention (e.g., feed adjustment, water quality management).
• Precision Aquaculture: Complements existing AI/camera systems by adding a biological layer of data, facilitating a shift from reactive disease treatment to proactive health management.
• Future Applications: The framework lays the groundwork for microbiome engineering (probiotics), antibiotic reduction, and longitudinal studies to validate pre-symptomatic detection capabilities across diverse farms and production stages.

Limitations Noted: The study is cross-sectional (single time point) and from a single facility. The "pre-symptomatic" capability requires future longitudinal validation, and thresholds may need recalibration for different environments or genetic stocks.