Integrative transcriptomic and machine learning approaches reveal candidate genes for silk and venom production in <scp>B</scp> ibionomorpha ( <scp>D</scp> iptera)

Insects display adaptive diversity in their secretory systems, which mediate ecological interactions ranging from defence to adhesion. Within Diptera, the co-occurrence of silk and venom represents a rare functional combination. Here, we integrate multi-species transcriptomics, coexpression network analysis and interpretable deep learning to investigate the physiological and evolutionary architecture underlying these traits. Using de novo assemblies from 11 Bibionomorpha species, we identified thousands of orthogroups and modelled their expression across species with Weighted Gene Coexpression Network Analysis and categorical neural classifiers. The models achieved near-perfect discrimination of venom and silk phenotypes (AUC >0.95), enabling functional inference at the orthogroup level. Venom-associated genes were enriched in proteolytic and inhibitory domains (trypsin-like, WAP, Kunitz, serpin), while silk-related genes exhibited chitin-binding CBM14/peritrophin motifs typical of adhesive matrices, supporting their physiological specialisation. AlphaFold3 structural predictions confirmed these domain-level signatures and revealed conserved disulfide-stabilized folds, supporting extracellular secretion. The minimal overlap between silk- and venom-associated modules indicates independent evolutionary origins of these secretory systems, exemplifying adaptive molecular convergence. Our findings provide molecular, structural and physiological insights into the adaptive evolution of secretory systems in Diptera, offering new perspectives for understanding the evolution of insect physiological innovations.