Extending GroupStruct2: a Bayesian and machine-learning framework for testing taxonomic hypotheses using morphometric...

Despite considerable advances in statistical methods, taxonomic delimitation using morphometric data (morphometric delimitation) has not significantly progressed beyond the use of simple summary statistics or univariate tests to quantify differences among predefined operational taxonomic units (OTUs). These methods typically rely on visual inspection of graphs or p-value thresholds to determine if character means are statistically different. Tiburtini et al. (2025) introduced a conceptually different approach for morphometric delimitation using Bayesian model-testing and Gaussian Mixture Models (GMM). This approach can infer morphological clusters with or without a priori OTU groupings and jointly evaluates the fit of alternate taxonomic hypotheses to the data, providing a probabilistic, model-based framework that moves beyond traditional significance testing. Additionally, a machine-learning method was proposed to identify diagnostic characters based on a Random Forest classification algorithm. Initially developed for plant morphometrics, we adapted Tiburtini et al.'s approach for any morphometric dataset and integrated it into GroupStruct2, a Shiny R-based application with a full graphical user interface that also includes conventional statistical methods (e.g. univariate/multivariate tests, PCA, DAPC, MFA). We demonstrate that a more robust, nuanced, and comprehensive perspective on morphological variation and character diagnoses can be achieved using GroupStruct2's integrative workflow that combines classical statistical analyses with Bayesian GMM and machine-learning methods. The integration of frequentist and Bayesian methods within a user-friendly graphical interface democratizes access to robust statistical analyses and enables researchers to adopt quantitative rigor in taxonomic studies.