Accuracy of occurrence and abundance estimates from insect metabarcoding

This study demonstrates that while both mild lysis and homogenization introduce distinct taxon-specific biases in insect occurrence data, combining homogenization with biological spike-ins enables reasonably accurate abundance estimates, advancing the potential of DNA metabarcoding for robust quantitative biodiversity monitoring.

The Gist

Imagine you have a giant bowl of mixed-up fruit salad, but instead of fruit, it's filled with thousands of tiny insects from a forest. Your goal is to answer two questions:

1. Who is in the bowl? (Which species are there?)
2. How many of each are there? (Is there one beetle or a thousand?)

For a long time, scientists have used a high-tech method called DNA metabarcoding to answer these questions. It's like taking a tiny sip of the "insect soup," reading the genetic code of everything in it, and identifying the guests. But there's a big problem: the method isn't perfect. Sometimes it misses the small guests, and sometimes it thinks there are more of a certain bug than there actually are.

This paper is like a massive "quality control test" to figure out how to get the most accurate results possible. The researchers tested different ways of making the "soup" and different ways of measuring the ingredients.

Here is the breakdown of their findings, using some everyday analogies:

1. The Two Ways to Make the Soup: "Steeping" vs. "Blending"

The researchers compared two main methods for getting DNA out of the bugs:

• Method A: The "Tea Bag" (Mild Lysis/Non-destructive).
  Imagine putting the insects in a cup of hot water (a chemical solution) and letting them sit. The DNA leaks out like tea from a tea bag, but the bugs stay whole.

  • The Good News: This is great for finding tiny, soft bugs (like tiny flies or parasitic wasps) because they dissolve easily. Plus, you can still look at the bugs under a microscope later if you want to study them.
  • The Bad News: It misses the tough, armored bugs (like beetles or ants with hard shells). Their DNA doesn't leak out well, so they might be invisible in the "tea."

• Method B: The "Smoothie" (Homogenization/Destructive).
  Imagine throwing the bugs into a blender and turning them into a complete mush.

  • The Good News: This gets DNA out of everything, especially the tough, armored bugs. It gives a very accurate count of how many of each bug are in the bowl.
  • The Bad News: You destroy the bugs. You can't look at them later, and it's harder to find the tiny ones because they get lost in the "smoothie" of DNA from the big bugs.

The Verdict: Neither method is perfect. "Tea" is better for finding rare, tiny species and saving the bugs. "Smoothie" is better for counting the tough ones accurately. The best strategy? Do both! Use "Tea" for most samples to save the bugs, and "Smoothie" for a few samples to get accurate counts.

2. The "Calibration Weights" (Spike-ins)

One of the biggest headaches in this science is that reading the DNA doesn't tell you the number of bugs. A big bug might release 100 times more DNA than a small bug, making it look like there are 100 big bugs when there's actually only one.

To fix this, the researchers added "Calibration Weights" (called spike-ins) to every bowl.

• Biological Weights: They added a known number of specific, foreign insects (like 5 fruit flies and 5 crickets) to every sample before processing.
• Synthetic Weights: They added tiny, artificial DNA strands that don't exist in nature.

The Analogy: Imagine you are weighing apples on a scale that is broken and fluctuates wildly. If you put a known 1kg weight on the scale first, you can see how much the scale is off and adjust your apple weight accordingly.

The Discovery:

• The Biological Weights (real bugs) were the champions. Because they went through the whole process (the "tea" or "smoothie" making), they told the scientists exactly how much DNA was lost or gained during the process.
• The Synthetic Weights (fake DNA) were okay, but they were added after the bugs were processed. They couldn't tell the scientists about the messiness of the "blending" or "steeping" steps.
• Result: Using real bugs as weights made the abundance estimates (counting the bugs) much more accurate.

3. The "Crystal Ball" (Bayesian Modeling)

Even with the weights, counting is still tricky. The researchers built a sophisticated computer model (a "Crystal Ball") that uses math to predict the number of bugs based on the DNA reads and the calibration weights.

• The Result: This model was surprisingly good. For about 73% of the bug species, the model guessed the number of individuals within just one of the actual number.
  • Example: If there were actually 5 beetles, the model guessed 4, 5, or 6. That's a huge improvement over just guessing based on raw DNA numbers!

The Big Takeaway

This paper is a roadmap for the future of insect monitoring. It tells us:

1. Don't just destroy everything: Keep some bugs intact ("Tea method") to find the tiny, rare ones and preserve them for the future.
2. Use real "weights": Add real insects to your samples to calibrate the machine. Don't rely on fake DNA alone.
3. Math helps: Use smart computer models to turn messy DNA data into accurate bug counts.

By combining these methods, scientists can finally get a clear, accurate picture of the insect world, helping us understand if insect populations are crashing or thriving, which is vital for our planet's health.

Technical Summary

1. Problem Statement

DNA metabarcoding has become a standard tool for assessing insect biodiversity, yet significant gaps remain regarding its quantitative accuracy. Two primary challenges hinder its adoption for rigorous biomonitoring:

• Methodological Bias: There is an ongoing debate between mild lysis (non-destructive, preserving specimens) and homogenization (destructive, "insect soup"). It is unclear how these methods differentially affect the recovery of species based on body size, sclerotization, and taxonomy.
• Quantitative Limitations: While metabarcoding provides presence/absence data, converting sequence read counts into accurate species abundance (specimen numbers) is notoriously difficult due to biological variation (body size, DNA content) and technical noise (PCR bias, extraction efficiency).
• Calibration Uncertainty: The efficacy of spike-ins (internal standards) in correcting these biases is not fully understood. Specifically, it is unknown whether biological spike-ins (foreign insect specimens) or synthetic spike-ins (artificial DNA fragments) are more effective, or if they can be combined for better results.

2. Methodology

The study utilized a large-scale, real-world dataset from the Insect Biome Atlas (IBA) project in Sweden, comprising 4,749 weekly Malaise trap samples.

• Experimental Design:

  • Dual Processing: All 4,749 samples underwent mild lysis. A subset of 856 samples was subsequently homogenized, allowing for a direct, paired comparison of the two methods.
  • Spike-in Integration:
    • Biological Spike-ins: Six insect species (Drosophila spp., Gryllodes sigillatus, Gryllus bimaculatus, Shelfordella lateralis) were added to every sample prior to lysis.
    • Synthetic Spike-ins: Two plasmids with artificial DNA targets were added to the homogenized subset prior to DNA extraction.
  • Ground Truth Benchmark: 15 samples were subjected to individual barcoding. Every specimen in these samples was sorted, individually barcoded, and sequenced. This provided a "gold standard" dataset of known species composition and exact specimen counts to validate metabarcoding results.

• Sequencing & Bioinformatics:

  • Targeted the COI gene (418 bp) using BF3-BR2 primers.
  • Sequenced on Illumina NovaSeq 6000.
  • Processed using the HAPP pipeline (denoising, chimera removal, taxonomic assignment against a custom BOLD database).

• Statistical Modeling:

  • Linear Calibration: Used linear models to test how well biological and synthetic spike-ins could reduce variance in read counts across samples.
  • Hierarchical Bayesian Modeling: Developed a TreePPL model to estimate specimen abundance ($n$) from read counts ($r$). The model treated reads as Gamma-distributed variables dependent on specimen count, PCR factors, and species-specific DNA release parameters ($k$ and $\theta$).
  • Priors: Tested both uninformative priors (assuming equal probability of counts) and informative priors (derived from the geometric distribution of specimen counts observed in the reference data).

3. Key Contributions

• First Large-Scale Real-World Benchmark: Unlike previous studies relying on mock communities, this study validates metabarcoding accuracy against thousands of real field samples with known ground truth (via individual barcoding).
• Quantitative Abundance Framework: Demonstrates a robust Bayesian workflow that translates raw read counts into probabilistic abundance estimates with quantified uncertainty.
• Spike-in Efficacy Analysis: Provides empirical evidence comparing biological vs. synthetic spike-ins, establishing that biological spike-ins are superior for capturing pre-extraction variation.
• Taxon-Specific Bias Mapping: Systematically maps how lysis vs. homogenization biases recovery across different arthropod orders and families.

4. Key Results

A. Occurrence Accuracy (Lysis vs. Homogenization)

• Overlap: 84% of taxa were shared between treatments.
• Lysis Advantages: Mild lysis recovered significantly more small, soft-bodied taxa (e.g., Cecidomyiidae, Chironomidae, small parasitoid wasps) that were often missed in homogenates. It preserves specimens for future study.
• Homogenization Advantages: Homogenization yielded higher read counts for large, heavily sclerotized, or hairy taxa (e.g., large Diptera, Coleoptera, Formicidae, Ichneumonidae).
• Conclusion: Neither method is universally superior; they have complementary taxonomic biases.

B. Abundance Estimation & Spike-in Calibration

• Raw Data Variance: Raw read counts varied wildly (orders of magnitude) even for identical specimen numbers, making uncalibrated abundance estimates unreliable.
• Biological vs. Synthetic Spike-ins:
  • Biological spike-ins were highly effective at reducing variance, particularly in homogenates.
  • Synthetic spike-ins provided minimal additional benefit over biological ones alone.
  • Conclusion: Biological spike-ins capture the majority of variation (including extraction efficiency and lysis conditions) that synthetic spike-ins miss. Using both did not significantly outperform biological spike-ins alone.
• Calibration Performance: Using 3–4 biological spike-ins was sufficient to maximize accuracy; adding more yielded diminishing returns.

C. Bayesian Abundance Prediction

• Accuracy: When applying the hierarchical Bayesian model with an informative prior (based on the geometric distribution of insect counts in Malaise traps), the model achieved:
  • Spot-on estimates (error = 0) for 51.7% of species occurrences.
  • High accuracy (error $\le$ 1 individual) for 72.9% of occurrences.
• Model Sensitivity: Models that allowed for species-specific DNA release parameters ($k$) performed significantly better than those assuming uniform behavior across spike-ins.
• Lysis vs. Homogenate for Abundance: Abundance estimates were significantly more accurate for homogenates than for lysates. Lysate read counts correlated poorly with specimen numbers for many species, likely due to variable DNA leakage from intact exoskeletons.

5. Significance and Recommendations

This study provides a critical step toward "quantitative metabarcoding," moving the field from qualitative presence/absence surveys to reliable abundance monitoring.

Practical Recommendations for Future Monitoring:

1. Hybrid Workflow: For large-scale monitoring, process the majority of samples via mild lysis to preserve specimens and detect soft-bodied taxa. Homogenize a subset (e.g., 25%) to obtain robust abundance data and recover sclerotized taxa.
2. Spike-in Strategy: If abundance estimation is required, biological spike-ins are essential and should be added as early as possible (ideally in the field or immediately upon lab arrival). Synthetic spike-ins are not necessary if biological ones are used.
3. Calibration: Use at least 3–4 biological spike-in species representing a range of body sizes and morphologies.
4. Data Analysis: Employ hierarchical Bayesian models with informative priors and species-specific calibration curves to translate read counts into specimen estimates.

Conclusion:
The paper demonstrates that while metabarcoding has inherent biases, these can be systematically corrected. By combining mild lysis for occurrence data, homogenization for abundance data, and biological spike-ins for calibration, researchers can achieve robust, quantitative biodiversity monitoring capable of detecting subtle ecological changes.