NeMO: a flexible R package for nested multi-species occupancy modelling and eDNA study optimization

The paper introduces NeMO, a flexible R package that utilizes Bayesian multi-species occupancy models to address detection uncertainty in eDNA metabarcoding by accounting for nested sampling structures, estimating key detection parameters, and optimizing survey design for effective biodiversity monitoring.

The Gist

Imagine you are trying to take a census of all the fish living in a giant, flowing river like the Rhône in France. In the old days, scientists had to go out with nets or electric shocks to catch fish and count them. It was hard work, often missed small or shy fish, and could be harmful to the animals.

Now, we have a high-tech tool called eDNA (environmental DNA). Think of it like finding a "scent trail" or "hair follicle" left behind by the fish in the water. By taking a cup of water, filtering it, and looking at the genetic material inside, we can tell which fish were there without ever seeing them.

The Problem: The "Ghost" in the Machine
Here is the catch: Just because you don't find a fish's DNA in your cup of water, doesn't mean the fish isn't there. Maybe the water was too muddy, maybe the DNA washed away, or maybe the lab test just missed it. This is called a false negative.

If scientists treat a "miss" as a "true absence," they might think a rare fish is gone forever when it's actually just hiding. It's like looking for a friend in a crowded room, not seeing them, and assuming they aren't there, when they are actually just standing behind a pillar.

The Solution: NeMO (The Detective's Toolkit)
The authors of this paper created a new computer program called NeMO (Nested eDNA Metabarcoding Occupancy). You can think of NeMO as a super-smart detective that doesn't just count what it sees; it calculates the probability of what it might have missed.

Here is how NeMO works, using a simple analogy:

1. The Three-Layer Cake (The Nested Design)

Imagine you are trying to find a specific cookie in a bakery.

• Layer 1 (The Site): You pick a bakery (a spot in the river). Is the cookie even in this bakery? (This is Occupancy).
• Layer 2 (The Sample): You buy a box of cookies from that bakery. Did you actually grab the box containing the cookie, or did you grab an empty one? (This is Collection).
• Layer 3 (The PCR Replicate): You open the box and look at the cookies. Did you spot the specific cookie, or did it get lost in the pile? (This is Amplification).

NeMO understands that you have to get lucky at all three layers to find the fish. If you fail at any step, you get a "zero," but NeMO knows that zero doesn't always mean "not there."

2. The "Read Count" (The Volume of Evidence)

Sometimes, the lab doesn't just say "Yes/No." It gives you a number: "We found 50 strands of DNA for this fish."
NeMO is smart enough to use these numbers. If you find 50 strands, it's a strong signal. If you find 1 strand, it's a weak signal. NeMO weighs this evidence to decide how likely it is that the fish is actually present.

3. The "Resource Calculator" (The Budget Planner)

One of the coolest features of NeMO is that it acts like a travel agent for scientists. Before they even go to the river, NeMO can tell them:

• "To be 95% sure you find this shy fish, you need to take 3 cups of water instead of 1."
• "You need to run the lab test 20 times instead of 5."
• "You need to look at 5,000 DNA strands instead of 500."

This helps scientists save money and time. Instead of guessing, they know exactly how much effort is needed to find the "ghost" fish.

What Did They Find?

The team tested NeMO on fish in the Rhône River. They discovered some fascinating things:

• Rarity vs. Elusiveness: Some fish are rare (they live in very few places), but when they are there, they are easy to find (conspicuous). Other fish are common (they live everywhere), but they are very hard to detect (elusive). NeMO can tell the difference between a fish that is actually gone and a fish that is just good at hiding.
• Upstream vs. Downstream: The model successfully figured out which fish prefer the cold, mountain headwaters and which prefer the warm, lower parts of the river, even when the data was messy.
• The "Missed" Fish: They found that the original study design (2 cups of water, 12 lab tests) was good enough for most fish, but for a few very tricky species, they would have needed way more effort to be sure they weren't missing them.

Why Does This Matter?

Conservation is all about knowing what is there and what is gone. If we think a species is extinct because we missed it once, we might stop protecting its habitat. If we think a species is safe when it's actually disappearing, we might not act in time.

NeMO is like a safety net. It ensures that when we say a fish is "absent," we are actually sure it's gone, and not just that our net was too small or our test was too weak. It turns a "maybe" into a "statistically confident answer," helping us protect biodiversity more effectively.

Technical Summary

1. Problem Statement

Environmental DNA (eDNA) metabarcoding has become a standard tool for biodiversity monitoring, offering a non-invasive method to detect a wide range of species. However, a critical limitation of eDNA surveys is imperfect detection. Non-detections are frequently misinterpreted as true absences, leading to false negatives. This issue is exacerbated by the nested structure of eDNA workflows, which involves multiple levels of replication:

1. Site level: Multiple field samples collected per site.
2. Sample level: Multiple PCR replicates per sample.
3. Sequencing level: High-throughput sequencing (HTS) generating read counts.

Existing occupancy models often fail to account for this full hierarchy or lack user-friendly software that integrates sequence read counts (quantitative data) alongside presence/absence data. Furthermore, there is a lack of tools to rigorously quantify detection uncertainty and optimize resource allocation (e.g., determining the minimum number of samples, PCR replicates, and sequencing depth required to reliably detect species).

2. Methodology

The authors developed NeMO (Nested eDNA Metabarcoding Occupancy), a user-friendly R package that implements a Bayesian hierarchical multi-species occupancy model using Markov Chain Monte Carlo (MCMC) sampling via JAGS.

A. The Statistical Framework

The model explicitly captures the nested design (Site $i$ $\rightarrow$ Sample $j$ $\rightarrow$ PCR Replicate $k$) and estimates four key parameters for each species $n$:

1. Occupancy Probability ($\psi$): The probability that a species' eDNA is present at a site.
2. Collection Probability ($\theta$): The conditional probability of collecting eDNA in a field sample, given presence.
3. Amplification Probability ($p$): The conditional probability of amplifying eDNA in a PCR replicate, given collection.
4. Expected Read Count ($\phi$): The expected number of sequencing reads generated, modeled using a Negative Binomial distribution to account for overdispersion in HTS data.

The model treats species as conditionally independent at the ecological level but allows parameters to covary through a multivariate normal community-level prior, enabling "borrowing of strength" across the community. It supports both presence/absence data and sequence read count data.

B. Modeling Protocols

NeMO offers three distinct modeling protocols via the Nemodel() function to accommodate different experimental designs:

1. PCR_rep_seq_read: The most complex model. Includes eDNA samples, individually indexed PCR replicates, and sequence read counts.
2. seq_read: Omits PCR replicate indexing (pools replicates) but retains sequence read counts.
3. PCR_rep: Focuses on individually indexed PCR replicates but ignores sequence read counts (presence/absence only).

C. Key Functions

• covarray(): Standardizes environmental and methodological covariates for integration into the model.
• min_resources(): Calculates the minimum number of samples ($J_{min}$), PCR replicates ($K_{min}$), and sequencing depth ($M_{min}$) required to achieve a target detection probability (e.g., 95%).
• WAIC(): Computes the Watanabe-Akaike Information Criterion for model comparison.

3. Key Contributions

• Software Development: The release of NeMO, the first R package to fully integrate nested multi-species occupancy modeling with sequence read counts and flexible PCR replication schemes.
• Methodological Advancement: Unlike previous tools (e.g., occumb, eDNAPlus), NeMO explicitly models the PCR replication step and allows for model comparison (WAIC) in a multi-species context.
• Study Design Optimization: It provides a statistical framework to calculate the minimum technical resources required to detect species, moving beyond post-hoc analysis to proactive study design.
• Distinction between Rarity and Elusiveness: The framework allows researchers to distinguish between species that are rare (low occupancy) and those that are elusive (low detectability despite presence).

4. Results

The authors applied NeMO to a fish biodiversity dataset from the Rhône River (France), comprising 80 sites, 2 samples per site, and 12 PCR replicates per sample, covering 100 fish species.

• Model Performance: All three protocols yielded similar performance (AUC $\approx$ 0.8), but the PCR_rep protocol was significantly faster (2 hours vs. 10 days for the full model).
• Occupancy Gradients: The model successfully identified spatial gradients. For example, species like Thymallus thymallus showed positive associations with upstream sites (distance to sea), while diadromous species like Alosa spp. were restricted to downstream sites.
• Correlation Analysis:
  • Strong positive correlations were found between detection parameters ($\theta, p, \phi$), indicating that species with high eDNA abundance are easier to collect, amplify, and sequence.
  • Moderate correlations existed between occupancy ($\psi$) and detection parameters, suggesting that rarity (low $\psi$) and elusiveness (low detection probability) are distinct concepts.
  • Example: Alosa spp. was rare (low occupancy) but conspicuous (high detectability), whereas Scardinius erythrophthalmus was common (high occupancy) but elusive (low detectability).
• Resource Optimization:
  • The original study design (2 samples, 12 PCR replicates) was sufficient for 63% of species to reach 95% detection probability.
  • For 18 species, additional effort was needed. For extreme cases like Leucaspius delineatus, achieving 95% detection would require >1,000 PCR replicates, highlighting the impracticality of the current design for certain targets.
  • Sequencing depth was generally sufficient, with only minor deficits for specific species.

5. Significance

• Improved Reliability: NeMO addresses the critical issue of false negatives in eDNA studies by quantifying detection uncertainty, leading to more accurate biodiversity assessments and conservation decisions.
• Cost-Effectiveness: By calculating minimum resource requirements, the tool helps stakeholders optimize budgets, preventing wasted resources on over-sampling or under-sampling.
• Interpretation of Non-Detections: The framework provides statistical support for filtering low-detection amplicons (e.g., distinguishing true absence from methodological failure) and guides the interpretation of "rare" vs. "elusive" species.
• Future Directions: The authors note that while NeMO handles false negatives well, future iterations could integrate false positive modeling and hydrological transport effects (DNA movement downstream) to further refine inference in lotic systems.

In summary, NeMO bridges the gap between complex hierarchical eDNA data and accessible statistical analysis, providing a robust, flexible tool for ecologists to design efficient surveys and interpret biodiversity data with greater rigor.