Abstract Delimitation of morphotypes in cryptic lineages remains challenging because classical linear statistics and visual inspection may fail to capture complex non-linear phenotypic variations. To address this, we introduced an unsupervised machine learning framework that couples a Graph Autoencoders (GAE) with community detection algorithms to map high-dimensional morphospaces. Unlike many traditional models that treat biological specimens as independent points in feature space, our approach represents multidimensional morphological data as an interconnected network. Using a k-nearest neighbor graph, we captured local topological relationships and embedded them into an optimized latent space. As an empirical case study, we applied this pipeline to 484 specimens of the morphologically conservative Neotropical pitviper genus Porthidium, using 21 linear and pholidosis (scale counts) traits. The GAE revealed a morphological structure that showed limited concordance with current taxonomic boundaries, identifying 12 distinct morphotypes with high structural modularity (Q = 0.6973) but low agreement with taxonomy (normalized mutual information, NMI = 0.2812). This framework provides an objective and scalable approach for exploring phenotypic structure in complex dataset. Because it does not require prior taxonomic assignments, it may be particularly useful for investigating cryptic lineages where morphological boundaries are ambiguous. More broadly, this approach can serve as a complementary exploratory step in integrative taxonomy and evolutionary studies.
Patron-Rivero et al. (Fri,) studied this question.