Deep generative learning helps extract ecological meaning from the complex three dimensional shapes of bird bills

Abstract

Data on the three dimensional shape of organismal morphology is becoming increasingly common. The availability of this data forms part of a new revolution in high-throughput phenomics that promises to help understand ecological and evolutionary processes that influence phenotypes at unprecedented scales. However, in order to meet the potential of this revolution we need new data analysis tools to deal with the complexity and heterogeneity of large-scale phenotypic data such as that represented by three dimensional shapes. In this study we explore the potential of techniques derived from Artificial Intelligence research to help organise and extract meaning from complex 3D data. Specifically, we train a deep representational learning method known as DeepSDF, a type of deep generative model, on a dataset of 3D scans of the bills of 2,020 bird species. The model is designed to learn a continuous vector representation of 3D shapes, along with a ‘decoder’ function, that allows the transformation from this vector space to the original 3d morphological space. We find that that the latent vector representations learned by the approach have morphological coherence in the sense that movement along various directions in the latent space form smooth continuous transformations in 3D morphological space, and that particular directions are associated with discernible morphological meaning (such as elongation, flattening, etc.). Learned latent vectors also have ecological meaning as shown by their ability to predict the trophic niche of the bird each bill belongs to with a high degree of accuracy. Unlike existing 3D morphometric techniques, this method has very little requirements for human supervised tasks such as landmark placement, increasing it accessibility to labs with fewer labour resources. It has fewer strong assumptions than alternative dimension reduction techniques such as PCA. The computational requirements for training the model, while substantial, is still within the reasonable reach of most researchers, with a ~2000 shape model taking just over 2 days to train on only a single current generation consumer-level GPU. Once trained, 3D morphology predictions can be made from latent vectors very computationally cheaply. Lastly, we present an in development R package with an implementation of the model, with the goal of making model training accessible to any researcher with access to data and GPU computing resources.