Unsupervised Machine Learning for Species Delimitation, Integrative Taxonomy, and Biodiversity Conservation

Abstract

Integrative taxonomy combining data from multiple axes of biologically relevant variation is a major recent goal of systematics. Ideally, such taxonomies would be backed by similarly integrative species-delimitation analyses. Yet, most current methods rely solely or primarily on molecular data, with other layers often incorporated only in a post hoc qualitative or comparative manner. A major limitation is the difficulty of deriving and implementing quantitative parametric models linking different datasets in a unified ecological and evolutionary framework. Machine Learning methods offer flexibility in this arena by learning high-dimensional associations between observations (e.g., individual specimens) across a wide array of input features (e.g., genetics, geography, environment, and phenotype) to delineate statistical clusters. Here, I implement an unsupervised method using Self-Organizing (or 'Kohonen') Maps (SOMs). Recent extensions called SuperSOMs can integrate an arbitrary number of layers, each of which exerts independent influence on the two-dimensional output clustering via empirically estimated weights. These output clusters can then be delimited into K significant units that are interpreted as species or other entities. I show an empirical example in Desmognathus salamanders with layers representing alleles, space, climate, and traits. Simulations reveal that the SOM/SuperSOM approach can detect K=1, does not over-split, reflects contributions from all layers with signal, and does not allow layer size (e.g., large genetic matrices) to overwhelm other datasets, desirable properties addressing major concerns from previous methods. Finally, I suggest that these and similar methods could integrate conservation-relevant layers such as population trends and human encroachment to delimit management units from an explicitly quantitative framework grounded in the ecology and evolution of species limits and boundaries.

Integrating Genomics, Collections, and Community Science to Reveal Speciation in a Variable Monitor Lizard (Varanus tristis)

“…We used the SuperSOM approach to infer species limits based on the simultaneous analysis of multiple types of data. Unlike the SOM analyses presented above for the molecular data, SuperSOM can reduce dimensionality based on multiple input layers (Wehrens and Buydens 2007; Pyron 2023). Weighing of the layers is independent of scale or the amount of data in each layer.…”

“…Our SuperSOM analysis was based on molecular, phenotypic, spatial, and environmental data. We implemented the approach of Pyron (2023) and used the scripts accompanying that publication, which rely primarily on the ‘kohonen’ R package. Since the effects of missing data are poorly known (Pyron 2023), we conducted this analysis on 23 individuals for which all types of data are available.…”

“…We implemented the approach of Pyron (2023) and used the scripts accompanying that publication, which rely primarily on the ‘kohonen’ R package. Since the effects of missing data are poorly known (Pyron 2023), we conducted this analysis on 23 individuals for which all types of data are available. This reduced sampling still sufficiently covers the geographic range of V .…”

“…We specified 4 x 4 cells for the output grid and performed 100 replicate analyses, each consisting of 100 iterations with a learning rate α (0.5, 0.1). We follow Pyron (2023) and use the term “species coefficients” to denote the frequency of assignment of each individual to a particular cluster across the replicates.…”

“…In that way, they can provide valuable insight into the drivers of genotypic and phenotypic divergence. Thus, the development of approaches that can integrate multiple sources of evidence to inform taxonomy (e.g., Solís-Lemus et al 2015; Pyron 2023) is critical.…”