Deep learning reveals ecological decoupling and rate heterogeneity in Hymenoptera wing evolution.

Quantifying biomorphological evolution and integrating phenotypic data into large-scale ecological and evolutionary studies are often limited by the scalability, subjectivity, and poor interpretability of existing methods. We introduce HAIKA (Hierarchical Attention-based Interpretable K-dimensional Autoencoder), a self-supervised residual channel-attention autoencoder that extracts continuous morphological representations from unlabeled standardized images and highlights the image regions that contribute to those representations. Applied to insect forewings across broad pterygote lineages, HAIKA outperformed most baseline deep learning models and conventional descriptors in clustering, dimensionality reduction, and morphological embedding. When applied to Hymenoptera (i.e., ants, bees, and wasps), the HAIKA-derived continuous characters improved morphology-based phylogenetic reconstruction and illuminated the evolutionary dynamics of wing shape and venation. Our results reveal that wing morphological evolution in Hymenoptera proceeds at decoupled rates across shape and venation modules and is not consistently aligned with key ecological transitions such as parasitoidism or eusociality. HAIKA provides a scalable, image-based route for phenotypic analyses, broadly applicable across diverse taxa and to both 2D and 3D datasets.