Protein family membership governs exosite predictability across the structural proteome.
No hay miniatura disponible
Fecha
Autores
Título de la revista
ISSN de la revista
Título del volumen
Editor
Resumen
Descripción
Exosites, defined as protein surface regions that mediate macromolecular recognition at sites distinct from catalytic centers, represent emerging targets for selective drug design, yet their structural diversity has precluded systematic computational identification. Here we demonstrate that exosite prediction performance varies substantially across protein families, ranging from Matthews correlation coefficient (MCC) of 0.47 for coagulation factors to 0.14 for kinases. Using ExositeDB, we developed STINGExoFind, a gradient boosting framework leveraging 87 structural descriptors from the STINGRDB2 database, and evaluated 180 proteins under leave-one-protein-out cross-validation (LOPO-CV). Coagulation proteases achieved 50% success rates at the MCC ≥ 0.5 threshold, whereas kinases and caspases remained largely unpredictable. Ten structures spanning six families exceeded MCC ≥ 0.7, including MAPK/ERK2 (MCC = 0.86) within the otherwise challenging kinase family, indicating that high-confidence predictions remain achievable for specific proteins even in poorly-performing families. These results establish exosite prediction as a family-specific rather than universal challenge: computational approaches can meaningfully guide experimental validation for coagulation factors and similarly consistent protein families, while structurally diverse families require experimental characterization. STINGExoFind is provided as a community resource to support future method development and exosite-targeting drug discovery.
Palabras clave
Aprendizado de máquina, Estrutura proteica, Aumento de gradiente, Descritores de nanoambiente, Descoberta de fármacos, Exosite prediction, Machine learning, Gradient boosting, Nanoenvironment descriptors, Drug discovery, Protein structure