AI reads insect wings to test evolution’s family tree. The drawings still begin with human expertise

The network of veins in an insect's wing preserves clues about evolutionary history, but converting those delicate patterns into comparable measurements has long demanded specialist labour. Scientists publishing in Science Advances on 16 October describe a self-supervised model called HAIKA that learns a numerical representation from standardised drawings of insect forewings. They then used those representations to help reconstruct relationships among Hymenoptera, the group containing bees, wasps and ants. The result offers a useful example of AI producing data for a scientific argument, rather than merely classifying a photograph.[1]

The central problem is how to describe a shape without asking a human to mark every landmark in advance. Taxonomists can recognise a useful vein junction or wing outline, but expert decisions about which characters count are difficult to repeat over thousands of species. A model that learns continuous measures from images could make those comparisons more scalable, provided that its attention lands on biologically meaningful regions and its outputs agree with independent evolutionary evidence.[1]

A neural model trained without species answers

HAIKA stands for Hierarchical Attention-based Interpretable K-dimensional Autoencoder. It was trained on more than a thousand wing images that had been standardised as line drawings, a process that removes much irrelevant background information. An autoencoder compresses an input into a smaller numerical representation and learns to reconstruct the original image. The researchers did not need to supply the correct taxonomic identity as a label for every training example; the reconstruction task itself supplied a learning signal.[1]

The attention mechanisms indicate which parts of the wing influence the model's representation. The researchers report that these regions concentrated on outlines, vein intersections and other morphological structures familiar to specialists. That is encouraging because an otherwise accurate embedding might rely on accidental scanner marks, image resolution or empty space rather than biological form. An attention map alone does not prove causal scientific validity, but it gives investigators something concrete to inspect.[1]

Numbers that test the method

On a dataset of 610 wings from major insect groups, a classifier built from the learned representations assigned them to the relevant group with reported accuracy of 85.6%. A more conservative leave-one-out evaluation produced 84.1%. For 405 species within Hymenoptera, the respective figures were 83.9% and 78.7%. These values are task-specific taxonomic grouping results, not a promise of 85% accuracy for identifying any insect from a casual phone photo.[1]

The more interesting comparison concerns an evolutionary family tree. Using conventional hand-selected morphological characters, a standard reconstruction recovered 11 of 23 deep nodes used as a reference. When the analysis included HAIKA's image-derived features, it recovered 19 of the 23. The combined analysis also recovered a major grouping in the hymenopteran tree that the conventional-only version missed. This is not proof that every branch is now settled; it shows that one additional representation of morphology supplied information useful to a specified reconstruction method.[1]

Shape and veins tell partly different stories

The team also investigated whether insect wings changed in tandem with major ecological transitions, such as parasitoidism, pollen collection and social living. The study describes different rates of evolution for overall wing shape and venation, with shape changing about 1.7 times as rapidly in its fitted analysis. Changes were not consistently aligned with major ecological transitions. That result challenges a simple narrative that every large change in lifestyle automatically generates an equally large change in the venation pattern.[1]

Those inferences depend on the chosen evolutionary tree, sampling, the standardisation of drawings and the statistical method. An apparent difference in rates can be sensitive to which species and fossils are included. Researchers will need to see whether the pattern holds when less carefully prepared imagery, alternative family trees or more three-dimensional morphological data are used.[1]

Why museums may care

Historical collections contain material that cannot be sequenced easily, especially fossils. Morphological characters therefore remain essential even in an age of genetic phylogeny. A reproducible, source-code-backed method for extracting wing features could let researchers revisit older specimens and challenge interpretations that depended on a limited set of manually selected landmarks. It could also make disagreements easier to investigate because the underlying numerical representations and attention regions can be compared.[1]

The limitation is practical as well as scientific. The input to this study was mostly a curated drawing, not a muddy, partly hidden wing in an uncontrolled natural image. Standardisation and expert selection remain part of the pipeline. HAIKA shows a credible route for bringing machine-learning measurements into systematic biology; it does not replace the taxonomist whose knowledge supplies the biological questions and checks the result.[1]

Report an error or suggest a correction