针对AML的结构性表面蛋白靶点

Background: Nearly all current antibody- or CAR-T-directed therapies in AML (e.g., CD33) target antigens that are also expressed (often at lower levels) on healthy hematopoietic stem/progenitor cells (HSPCs), causing dose-limiting myeloablation or prolonged cytopenias. Driven by this critical need for tumor-specific targets, we developed a deep learning approach that combines structural proteomics with multi-modal biological network integration and high-throughput protein complex structure prediction to predict potential targets based on predicted surface localization, changes in protein-protein interactions (PPIs), degree of conformational change, and estimated potential therapeutic relevance. Methods: Protein tagging reagents were used to quantify changes in amino acid surface accessibility and solvent accessibility (SASA) via quantitative LC-MS/MS. Using this innovative approach, changes in the global structural surfaceome of NOMO-1 AML cells treated with all-trans retinoic acid (atRA) were compared with those of vehicle-treated cells. Peptides with significant SASA changes (FDR: q < 0.05; log 2 fold-change > +/-1) were filtered using a surfaceome localization score, and conformational ensembles generated for this protein list were scored by a custom GNN-based structure encoder fine-tuned on surface proteomics data. Further, to identify PPIs with interface SASA changes, an all-by-all list of protein pairs was first fed to a graph attention-based module integrating multiple data modalities to retain high-confidence interactions, which were subsequently passed to AlphaFold-Multimer for complex structure prediction, and final target PPIs were predicted by a scoring function based on interface SASA. Results: The global structural analysis identified 20,933 modified peptides from 3,540 proteins, of which 2,627 peptides from 1,089 proteins exhibited significant changes in SASA. These proteins and the corresponding 1089×1089 candidate PPIs were analyzed by our AI/ML pipeline to generate a final ranked list of potential structural targets, with the top candidates demonstrating high predicted affinity and specificity against AML. The data set achieved in the current study has successfully identified new surface targets in response to atRA treatment that are independent of atRA-RARA differentiation. Conclusion: Our structural proteomics platform significantly broadens the potential druggable space in AML by identifying novel structural surface targets. The data set achieved in the current study has successfully identified new surface targets in response to multiple atRA-induced mechanisms. The top-ranked structural targets from this study are being evaluated as potential antibody-drug conjugates (ADCs)-based therapies for AML.