Machine learning-driven identification of DLL3 as a molecular target and development of a DLL3-binding cyclic peptide for glioblastoma.

Background: Glioblastoma (GBM) remains the most aggressive primary brain tumor and is associated with a dismal prognosis despite maximal multimodal therapy. Targeted radionuclide therapy requires a tumor-selective, cell-surface-accessible molecular target that is highly expressed in GBM but minimally present in normal brain tissue. Methods: We integrated transcriptomic profiling, differential expression analysis, LASSO regression, gradient boosting, SHAP interpretation, structural modeling, molecular docking, molecular dynamics simulation, peptide synthesis, and fluorescence-based validation to identify a GBM-associated membrane target and develop a corresponding cyclic peptide ligand. Public GBM and normal brain transcriptomic datasets were obtained from TCGA, GTEx, and CGGA. Candidate genes were prioritized using machine learning and structural criteria, followed by systematic peptide design and computational refinement. Results: DLL3 was identified as a top candidate target with high discriminative power and prognostic relevance in GBM. A cyclic peptide candidate, IMP-3, displayed favorable predicted binding to the DLL3 extracellular domain and stable interaction during molecular dynamics simulation. The probe-labeled derivative MPA-IMP-3 selectively accumulated in GBM cells and showed significantly stronger fluorescence signals than in normal astrocytes, supporting DLL3-associated targeting specificity. Conclusion: We identified DLL3 as a promising GBM-associated membrane target and developed a cyclic peptide ligand with selective binding potential. This integrated AI-driven workflow provides a rational framework for the development of targeted theranostic strategies in GBM.