eDNAPlus: A unifying modelling framework for DNA-based biodiversity monitoring

Alex Diana, Eleni Matechou, Jim E. Griffin, Douglas W. Yu, Mingjie Luo, Marie I. Tosa, Alex Bush, R. E. Griffiths

TL;DR

eDNAPlus is a new Bayesian hierarchical framework designed to turn messy DNA metabarcoding "read counts" into reliable biodiversity maps. By explicitly modeling the noise and errors that occur at every step—from collecting samples in the forest to sequencing DNA in the lab—it allows researchers to estimate species biomass changes and correlations with unprecedented accuracy.

What Problem Does This Paper Address?

Ecologists are increasingly using "metabarcoding"—sequencing DNA from environmental samples like soil, water, or bulk insect traps—to monitor biodiversity. While this method is fast and can identify thousands of species at once, the data it produces is notoriously "noisy."

There are three main hurdles:

  1. False Negatives/Positives: A species might be at a site but missed during sampling (false negative), or DNA might contaminate a sample where the species isn't actually present (false positive).
  2. Species Effects: Some species shed more DNA than others, and some DNA sequences amplify more easily in the lab, making "read counts" a poor direct measure of absolute abundance.
  3. Complex Correlations: Species aren't distributed randomly; they interact with each other and their environment in ways that standard models often ignore.

Existing approaches often throw away valuable information by converting counts into simple "present/absent" data or fail to account for the multi-stage nature of the errors.

Methods

The authors propose eDNAPlus, a three-level hierarchical crossed-effects model:

  • Level 1 (Biomass Availability): Models the latent (hidden) amount of DNA at a site based on environmental covariates and spatial/species correlations.
  • Level 2 (Biomass Collection): Models the stochastic process of actually catching that DNA in a physical sample (e.g., a bottle of water or an insect trap).
  • Level 3 (Biomass Analysis): Models the laboratory process, including PCR amplification and sequencing, which introduces further noise and "pipeline effects."

Technical Innovations: * Inference: They use Markov chain Monte Carlo (MCMC) with Laplace approximations. * Optimization: To handle the computational burden of thousands of latent variables, they implemented a re-parameterization scheme for crossed-effects models to improve "mixing" (efficiency) and an adaptive approach for updating variables to slash computation time. * Priors: They use a "Graphical Horseshoe" prior to infer the complex web of correlations between different species.

Datasets

  1. Simulated Data: Extensive simulations were used to test how the number of sites, samples, and PCR replicates affect the model's accuracy.
  2. HJA Malaise-trap Dataset: A real-world case study involving 121 Malaise-trap samples (mass-trapped insects) from 89 sites in the H. J. Andrews Experimental Forest, Oregon. The study focused on 50 arthropod species and two environmental covariates: elevation and distance-to-road.

Key Insights and Ideas

  • Within-Species vs. Between-Species: The model acknowledges that we cannot compare absolute biomass between different species (because their DNA "signatures" are too different). However, we can accurately track how a single species' biomass changes across different sites or over time.
  • The Power of Replication: A major "aha moment" from their study design analysis is that collecting more physical samples in the field is significantly more beneficial than running more PCR replicates in the lab. If you have a limited budget, go back to the field.
  • Spike-ins: Adding known quantities of synthetic DNA ("spike-ins") to samples helps the model calibrate lab noise. The authors found that just one or two spike-ins provide most of the benefit; adding more yields diminishing returns.
  • Environmental Drivers: In the Oregon case study, elevation was a much stronger predictor of insect biomass than distance-to-road, likely due to its pervasive effect on temperature and vegetation.

Key Results

  • Accuracy: The model successfully distinguished between high- and low-biomass sites, with error rates (Brier scores) dropping sharply as field replication increased.
  • Validation: In the case study, the probability of "Stage 2" (lab) errors was found to be near zero for most species, validating the high quality of modern laboratory processing when paired with robust statistical modeling.
  • Conservation Mapping: The framework produced high-resolution maps of "biomass-weighted species richness," allowing researchers to identify and rank specific areas of the forest by their conservation value.

Limitations and Future Work

  • Scalability: While the optimizations help, modeling "microbiomes" with tens of thousands of species remains a computational challenge. The authors suggest variational Bayes or low-dimensional representations as potential future fixes.
  • Bioinformatics Integration: Currently, the model starts with an "OTU table" (pre-processed data). Future versions could potentially integrate the raw sequencing data directly to better handle genetic diversity within species.
  • Absence vs. Zero: Because the model treats biomass on a continuous scale, it doesn't strictly distinguish between a species being "truly absent" and its biomass being "practically zero." This is a subtle but important distinction for certain ecological theories.

Summary updated

Abstract

DNA-based biodiversity surveys involve collecting physical samples from survey sites and assaying the contents in the laboratory to detect species via their diagnostic DNA sequences. DNA-based surveys are increasingly being adopted for biodiversity monitoring. The most commonly employed method is metabarcoding, which combines PCR with high-throughput DNA sequencing to amplify and then read `DNA barcode' sequences. This process generates count data indicating the number of times each DNA barcode was read. However, DNA-based data are noisy and error-prone, with several sources of variation. In this paper, we present a unifying modelling framework for DNA-based data allowing for all key sources of variation and error in the data-generating process. The model can estimate within-species biomass changes across sites and link those changes to environmental covariates, while accounting for species and sites correlation. Inference is performed using MCMC, where we employ Gibbs or Metropolis-Hastings updates with Laplace approximations. We also implement a re-parameterisation scheme, appropriate for crossed-effects models, leading to improved mixing, and an adaptive approach for updating latent variables, reducing computation time. We discuss study design and present theoretical and simulation results to guide decisions on replication at different stages and on the use of quality control methods. We demonstrate the new framework on a dataset of Malaise-trap samples. We quantify the effects of elevation and distance-to-road on each species, infer species correlations, and produce maps identifying areas of high biodiversity, which can be used to rank areas by conservation value. We estimate the level of noise between sites and within sample replicates, and the probabilities of error at the PCR stage, which are close to zero for most species considered, validating the employed laboratory processing.
BibTeX
@article{diana2024ednaplus,
  author = {Alex Diana and Eleni Matechou and Jim E. Griffin and Douglas W. Yu and Mingjie Luo and Marie I. Tosa and Alex Bush and R. E. Griffiths},
  title = {{eDNAPlus: A unifying modelling framework for DNA-based biodiversity monitoring}},
  journal = {Journal of the American Statistical Association},
  year = {2024},
  doi = {10.1080/01621459.2024.2412362},
}
Figure