Multimodal spatiotemporal atlas to reveal CCDC3 ⁺ CAFs as drivers of liver metastasis in colorectal cancer.

25

Background: Colorectal cancer liver metastasis (CRLM) exhibits pronounced heterogeneity and dynamic evolution of the immune microenvironment. However, most studies remain limited to static snapshots, hindering identification of key regulatory axes and actionable nodes across multimodal and temporal scales. Methods: We integrated single-cell sequencing with multi-omic analyses to derive a chromosomal instability (CIN) index. To map spatiotemporal progression, we built a multimodal atlas of CRLM by integrating 10× Visium spatial transcriptomics, MALDI-MSI spatial metabolomics, and Olink spatial proteomics, alongside multi-timepoint sampling in humanized PDX models. Positional encoding and adversarial domain adaptation corrected modality- and batch-specific effects, achieving subcellular-scale alignment and a high-fidelity spatiotemporal feature matrix. Graph neural network–based analyses quantified spatial gradients and interactions among CCDC3⁺ CAFs, Tregs, and CD8⁺ T cells. To decode high-dimensional spatiotemporal signals, we developed a physics-informed deep dynamic model (SpaTemNet-PINN) that integrates spatial topology with temporal dependencies and embeds diffusion–reaction kinetic constraints; key model-predicted nodes were validated by CRISPRi. Finally, reinforcement learning modeled optimal dosing schedules for combined CCDC3/CDT1 blockade with PD-1 inhibition, and a spatial immune scoring system was established for patient stratification and longitudinal monitoring. Results: We identified CAF-derived CCDC3 as a central driver linking stromal remodeling to CIN, demonstrating that CCDC3 engages CXCR3 on tumor cells to activate STAT3 phosphorylation and induce CDT1 transcription, establishing a CCDC3/CXCR3/STAT3/CDT1 axis that promotes proliferation, metastasis, and CIN. The spatiotemporal atlas resolved an immune-excluded niche marked by coordinated spatial gradients of CCDC3⁺ CAFs and Tregs alongside CD8⁺ T-cell depletion. SpaTemNet-PINN quantitatively captured how CCDC3 gradients perturb PCNA–CDT1 homeostasis and shape CIN trajectories, with CRISPRi validating key regulators. CIN-associated DAMPs–NETs shielding structures were spatially linked to impaired CD8⁺ T-cell infiltration. Reinforcement learning–guided simulations optimized dual CCDC3/CDT1 blockade with PD-1 inhibition and informed a spatial immune score for patient stratification and dynamic response monitoring. Conclusions: This study establishes an integrated framework—from mechanistic dissection to spatially informed therapeutic intervention—providing systematic evidence for understanding the spatiotemporal nature of immune exclusion in CRLM and advancing precision therapeutic strategies.