Shared Blueprint Helps AI Predict Aphid Protein Structures

Researchers at the Stowers Institute for Medical Research and the University of Pittsburgh have successfully leveraged artificial intelligence to decode the structures of rapidly evolving aphid proteins, known as BICYCLE proteins, which enable these insects to hijack plant development and form protective galls. The findings, published September 2, 2026, in the Proceedings of the National Academy of Sciences, establish a novel framework for applying deep learning to molecular systems that evolve too quickly for conventional database matching. Aphid saliva contains hundreds of these proteins, named for their repeating cysteine motifs. Because they mutate at an extreme rate, traditional sequence homology searches fail to identify their biological family. To bypass this limitation, the research team, led by Investigator David Stern and structural biologists under Angela Gronenborn, initially crystallized two BICYCLE proteins and determined their three-dimensional shapes using X-ray diffraction. The structures revealed a saposin-like fold, a motif common in other biological systems but previously unrecorded in this context. When the team initially input the protein sequences into AlphaFold2, the advanced deep learning model failed to predict the correct architecture. This failure proved instrumental, revealing that AlphaFold2 relies heavily on evolutionary context rather than pure machine learning. To supply the missing data, researchers collected aphid samples across Virginia, West Virginia, and Japan, sequenced their genomes, and integrated the resulting phylogenetic information into the AI pipeline. With this customized evolutionary framework, AlphaFold2 accurately reconstructed the known crystalline structures. Building on this successful methodology, the researchers generated approximately 2,400 high-confidence structural predictions spanning seven aphid species. The analysis uncovered a consistent underlying saposin-like blueprint that is continuously duplicated, reoriented, and chemically modified. Despite this structural conservation, the protein surfaces exhibit extreme chemical diversity, lacking any single conserved feature across the family. This variation suggests that BICYCLE proteins operate through multiple overlapping mechanisms to manipulate plant cells while evading host immune recognition. The study demonstrates that supplementing AI models with targeted evolutionary data can unlock the structural analysis of fast-evolving molecular arsenals. This approach extends beyond entomology, offering a scalable pathway for investigating proteins involved in host-parasite dynamics, immune evasion, and agricultural pest resistance. By mapping the molecular tactics insects use to override plant genomes, the research lays the groundwork for future interventions in crop protection and provides a new methodological standard for structural biology amid rapid biological evolution.