Bridging multi-omics to routine precision oncology: Biologically informed knowledge transfer for precision pCR prediction in neoadjuvant immunotherapy.

e12569 Background: Immune Checkpoint Inhibitors (ICI) significantly improve pCR rates, yet optimal patient selection is impeded, particularly in resource-limited settings, by reliance on costly omics data unavailable in standard practice. To develop the iM-KT (immune Multi-modal Knowledge Transfer) framework, a biologically informed model designed to distill complex immune landscapes into accessible routine data, enabling precise pCR prediction relying solely on accessible routine inputs, achieving robust predictive performance without the need for costly omics profiling. Material and. Methods: iM-KT was pre-trained on the I-SPY 2 landscape (n = 979), integrating transcriptomic, proteomic, MRI, and clinical data. We employed a cross-modal knowledge transfer paradigm to distill high-dimensional omics insights into standard inputs (MRI, clinical factors). Specifically, a Teacher network, trained on matched multi-omics, supervised a Student network restricted to MRI and clinical variables (age, HER2, HR). To ensure biological alignment, the Student was optimized to reconstruct immune-pathway gene expression. The pCR predictor was then trained on the NACT+ICI subset (n = 69) and validated in an independent real-world cohort (n = 59). Performance endpoints included pCR accuracy, Distant Recurrence-Free Survival (DRFS), and biological validation via Gene Set Enrichment Analysis (GSEA). Results: In the independent test cohort receiving NACT+ICI, iM-KT outperformed receptor status in predicting pCR (AUC 0.76 vs. 0.65, p = .004). Prognostically, iM-KT effectively stratified risk, distinguishing predicted responders from non-responders with significant separation (HR 0.18; 95% CI, 0.04-0.88; p = .034). Within the surgical non-pCR subpopulation (n = 27), the model distinguished a favorable "near-pCR" subset with a clinically relevant reduction in recurrence risk (HR 0.34, recurrence 16.7% vs. 33.3%), implying the potential to de-escalate adjuvant therapy for biologically responsive patients. Biological validation confirmed predictions were underpinned by Primary immunodeficiency (NES = 2.70; p < .001) and Th1/Th2 cell differentiation (NES = 2.69; p < .001), capturing complex immune dynamics that standard receptor status fails to resolve. Conclusions: iM-KT demonstrates that accurate response prediction is achievable in clinical practice without relying on high-cost omics data. By successfully transferring intrinsic immune mechanisms into routine diagnostics, the framework captures subtle biological responses missed by standard pathology, offering a scalable strategy to guide personalized therapeutic escalation or de-escalation in the neoadjuvant immunotherapy setting.