Graph autoencoders and community detection algorithms to improve polymorphic identification

Area of Science:

• Evolutionary Biology
• Computational Biology
• Taxonomy

Background:

• Delimiting morphotypes in cryptic lineages is difficult due to limitations of traditional statistical methods in capturing complex phenotypic variations.
• Existing models often treat biological specimens as independent data points, failing to account for intricate relationships within morphological data.

Purpose of the Study:

• To introduce an unsupervised machine learning framework for mapping high-dimensional morphospaces.
• To address the challenge of identifying morphotypes in cryptic lineages by analyzing complex, non-linear phenotypic variations.

Main Methods:

• Developed a framework coupling Graph Autoencoders (GAE) with community detection algorithms.
• Represented morphological data as a network using a k-nearest neighbor graph to capture topological relationships.
• Applied the pipeline to 484 Neotropical pitviper specimens (Porthidium) using 21 linear and pholidosis traits.

Main Results:

• The GAE identified 12 distinct morphotypes with high structural modularity (Q = 0.6973).
• The discovered morphotypes showed limited concordance with current taxonomic boundaries (NMI = 0.2812).
• The unsupervised approach revealed a complex morphological structure not apparent through traditional methods.

Conclusions:

• The proposed GAE framework offers an objective and scalable method for exploring phenotypic structure in complex datasets.
• This approach is particularly valuable for investigating cryptic lineages with ambiguous morphological boundaries.
• The framework serves as a complementary tool for integrative taxonomy and evolutionary studies.