Supplementary Data for "Deep learning reveals ecological decoupling and rate heterogeneity in Hymenoptera wing evolution"

This repository contains the complete raw data, computational matrices, and codes utilized in the study "Deep learning reveals ecological decoupling and rate heterogeneity in Hymenoptera wing evolution". The files are organized as follows: S1: Raw image dataset of insect forewings. The complete, unprocessed image collection utilized throughout the analytical framework. The dataset is organized into folders, with filenames corresponding to their taxonomic classification. S2: Raw model-extracted matrices used in Case Studies 1–3. Contains the raw output matrices for Case Studies 1–3. These files include grouping metadata to facilitate visualization. They are provided to allow direct reproduction of the study's results or to serve as a basis for future analyses. S3: Phylogenetic matrix and results for Hymenoptera in Case study 2. Comprises the phylogenetic matrices (encompassing discrete, continuous, and combined datasets) and analytical results for Case Study 2. The discrete feature matrix and topological constraints for all Hymenoptera species align with the protocols established by Ronquist et al. (2012). Reference: Ronquist, F. et al. (2012). A Total-Evidence Approach to Dating with Fossils, Applied to the Early Radiation of the Hymenoptera. Systematic Biology, 61(6), 973–999. https://doi.org/10.1093/sysbio/sys058 S4: Raw model-extracted matrix of Primate 3D dataset. A dataset comprising 331 data entries across six skeletal classes (Cuneiform, Mandible, Patella, Scapula, Vertebra, and Femur) derived from 3D models. Volume data are from Almécija et al. (2024). Reference: Almécija, S., Pugh, K. D., Anaya, A., Smith, C. M., Simmons, N. B., Voss, R. S., ... & Catalano, S. A. (2024). Primate Phenotypes: A Multi-Institution Collection of 3D Morphological Data Housed in MorphoSource. Scientific data, 11(1), 1391. S5: Raw model-extracted matrix of standard geometric 3D dataset. Contains data for five classes of standard volumetric structures (Sphere, Cube, Ellipsoid, Cylinder, Torus), with 200 structures per class. S6: The code, sample files, and user guide for the HAIKA model. All files are packaged into a single compressed file, containing:1. README.mdProvides an overview of the HAIKA project, including its purpose, methodology, and basic instructions for use.2. requirements.txtLists all Python libraries and dependencies required to run the HAIKA pipeline.3. preprocessing.pyHandles image preprocessing tasks such as resizing, format conversion, file organization, and dataset statistics generation.4. main.pyThe main execution script. It trains the deep learning autoencoder, extracts latent features from images, performs clustering, and generates visualizations.5. user_guide.docxA detailed user manual explaining how to install, configure, and operate the HAIKA pipeline.6. heatmap/ sample imagesStores 100 visualization outputs, including heatmaps that show neural network attention regions.7. segmentation/Includes scripts and outputs for automatic wing segmentation and morphological measurement extraction.8. segmentation/segmentation_res/Stores 30 visualization outputs, including segmentatioin mask that show the segmentation results.9. segmentation/outline_res/ sample imagesStores 30 visualization outputs, including outline images that show the outline extraction results.10. segmentation/seg_main.pyRuns the segmentation pipeline using the Segment Anything Model (SAM) to isolate wing regions.11. segmentation/outline_extract_and_calculate.pyExtracts wing outlines from segmentation masks and calculates morphological metrics such as area and aspect ratio.12. checkpoints/Stores saved model weights and training checkpoints from the autoencoder.13. 3d_expansion/Contains experimental scripts demonstrating extension of the HAIKA framework to 3D shape data.14. 3d_expansion/main.pyThe main scripts apply 3D convolutional filters to extend the HAIKA framework to three-dimensional shape data.