Assessing bivalve phylogeny using Deep Learning and Computer Vision approaches

Abstract

Phylogenetic analyses using morphological data currently require hand-crafted character matrices, limiting the number of taxa that can be included. Here I explore how Deep Learning and Computer Vision approaches typically applied to image classification tasks, may be used to infer phylogenetic relationships among bivalves. A convolutional neural network (CNN) was trained on thousands of images showing species of 75 bivalve families. The predictions of the CNN on a large number of bivalve images are then interpreted as an indication of how similar these bivalves are to each other, are averaged by the families to which the species belonged, and visualized in a cluster diagram. In this cluster diagram, significantly more families clustered with members of their subclasses than expected by chance, confirming the feasibility of the approach. To address the issue of convergent evolution, two further CNNs were trained, on the same images but grouped by the orders and subclasses to which the species belonged. Combining predictions for the same images but on different taxonomic levels improved the inferred phylogenetic relationships also of families that the CNNs had not been trained on. Finally, this combined tree is merged with five published phylogenetic trees into a supertree, representing the largest single phylogeny of the Bivalvia to date, encompassing 128 families, including six exclusively fossil families and nine extant families for which presently no molecular data are available. Issues inherent to the approach and suggestions for future directions are discussed.