DNA metabarcoding improves the specificity and practicality of machine learning classification in bulk invertebrate samples

Monday, November 14, 2022

11:06 AM – 11:18 AM PT

Location: Vancouver Convention Centre, Meeting Room 221/222

Presenting Author(s)

Jarrett D. Blair

PhD Candidate
The University of British Columbia
Vancouver, British Columbia, Canada

Co-Author(s)

MW

Michael Weiser

The University of Oklahoma
Tucson, Arizona
Kd

Kirsten de Beurs

The University of Oklahoma
Norman, Oklahoma
MK

Michael Kaspari

The University of Oklahoma
Norman, Oklahoma
CS

Cameron Siler

The University of Oklahoma
Norman, Oklahoma
Katie Marshall

Assistant Professor
The University of British Columbia
Vancouver, British Columbia, Canada

There is growing evidence that invertebrate populations are declining at an alarming rate. However, these declines may not be uniform across taxa and geography, so we need more data to make intelligent conservation decisions. Machine learning classification is a potential solution that has shown promise as an efficient tool for quickly measuring invertebrate abundance, but species-level classifications become increasingly difficult as taxonomic scope increases. Conversely, DNA metabarcoding has shown it can reliably detect species presence/absence but struggles when measuring abundance. In this study, we outline a novel combination of machine learning and DNA metabarcoding approaches to increase accuracy, specificity, and practicality of image classification in bulk invertebrate samples. We used >50,000 terrestrial invertebrates spanning three phyla (Annelida, Arthropoda, and Mollusca) collected by the National Ecological Observatory Network (NEON) in pitfall traps to train our machine learning classification models. Additionally, we extracted extracellular DNA from the ethanol each sample was stored in for metabarcoding. This gave us an approximate species list of all species present in each sample that we could compare against our machine learning classifications. Doing this greatly increased the average taxonomic specificity of our model and allowed for morphospecies clustering of the entire dataset. The primary advantage of our approach is its scalability. Rather than attempting the virtually impossible task of training a model (or models) to classify all invertebrates down to species, generalist models can be trained at higher levels while still making species-level classifications with the assistance of metabarcoding.