Assessing the potential of bee-collected pollen sequence data to train machine learning models for geolocation of sample origin

Pollen is a robust and widespread substance that captures a historical snapshot of a specific time and place, and it can be used to track movements through space by examining the pollen deposited on various objects. Palynology, the study of pollen, is used across fields such as conservation, natural history, and forensics, where it is particularly useful for tracing the origin and movement of objects. However, pollen has remained underutilized due to the difficulty of distinguishing many pollen taxa beyond the family level and limited pollen reference material to support location predictions. With recent developments in pollen DNA metabarcoding these issues have been rectified, but much of the available pollen data are primarily from wind-pollinated species, which are widespread and less informative of specific sample locations. Bee-collected pollen presents an untapped resource in training predictive models to geolocate sample origin. Here we compiled bee-collected pollen DNA sequence relative abundance data from three projects in the western U.S. and assessed the accuracy of supervised machine learning models to predict the location of sample origin based solely on pollen assemblage, without the need of incorporating additional data. Random Forest and k-Nearest Neighbors models yielded high accuracy across all projects. We also found that models trained on taxonomically clustered pollen assigned sequence variants (ASVs) performed slightly better than those trained on raw sequence data, but the difference was minor, indicating that models trained on raw sequence data can reliably predict location and avoid the time-consuming taxonomic assignment process. Our results demonstrate the utility of repurposing bee-collected pollen for geolocation and provide a framework for employing supervised machine learning in future geolocation efforts. Highlights: Bee-collected pollen metabarcoding data was used to accurately predict sample originRandom Forest and k-Nearest Neighbors algorithms were most accurate with lowest errorTaxonomically-classified and raw DNA sequence data training sets performed comparably.