Using natural history to guide supervised machine learning for cryptic species delimitation with genetic data

Area of Science:

• Evolutionary Biology
• Genetics
• Bioinformatics

Background:

• Species delimitation is complicated by diverse biological traits and genetic processes.
• Genetic methods, like the multispecies coalescent, can overestimate species numbers, especially in structured or cryptic taxa.
• This genetic oversplitting contrasts with lumping tendencies from other evidence types, creating a delimitation conundrum.

Purpose of the Study:

• To address the challenge of species delimitation in taxa prone to genetic oversplitting.
• To showcase the species delimitation conundrum using the low-dispersal harvester mite, Theromaster brunneus.
• To introduce and validate a novel supervised machine learning approach for cryptic species delimitation.

Main Methods:

• Integration of morphological, mitochondrial, and sub-genomic (double-digest RADSeq, ultraconserved elements) data.
• Application of multispecies coalescent models and a supervised machine learning approach.
• Creation of a custom training dataset from a well-studied lineage to inform delimitation in Theromaster brunneus.

Main Results:

• High discordance observed between different data types and analyses in inferring species numbers for Theromaster brunneus.
• Evidence suggests multispecies coalescent approaches tend to oversplit species in this taxon.
• The supervised machine learning approach demonstrated effectiveness in delimiting cryptic species.

Conclusions:

• The developed supervised machine learning method offers a powerful tool for resolving cryptic species, especially in taxa with complex genetic structures.
• This approach leverages known biological characteristics to inform delimitation in poorly understood taxa.
• The method is broadly applicable across the tree of life for more biologically informed species delimitation.