A neural network model delivers a highly prognostic protein signature in cancer stem cells that identifies relapse in stage III colorectal cancer patients

Abstract Background Stage III colorectal cancer poses a significant threat of metastasis development, as tumour resection and adjuvant chemotherapy do not guarantee prolonged disease-free survival. Objective The spatial, quantitative, and qualitative characteristics of various cell types within tumour tissues could be key to developing accurate prognostic AI models. Design Tissue microarrays created from primary tumour tissues collected during surgical resection from a cohort of 493 stage III colorectal cancer (CRC) patients were analysed for 61 protein markers at the single-cell level using multiplexed immunofluorescence imaging via the Cell DIVE™ platform. Subsequent cell-type classification enabled quantitative cell-type analyses, co-localisation neighbourhood assessments, and cell-type-specific protein signature discoveries that distinguish between early and late/non-recurring patient samples. Results This study identifies a stem cell protein profile that drives tumour relapse. A deep neural network (DNN) model, based on a stem cell protein signature composed of BAX, MLKL, FLIP, GLUT1, and CDX2, provided accurate prognosis for stage III CRC patients in both discovery and validation cohorts and in an independent validation cohort. Nodal count-based metric further increased prognosis accuracy. Our study also revealed distinct spatial arrangements of immune, endothelial, and stem cells that were linked to early tumour recurrence. Conclusion Our findings propose a clinically promising prognostic tool based on a five-protein stem cell signature. These markers not only predict chemotherapy resistance in cancer stem cells but also suggest potential therapeutic strategies such as combinatorial treatments incorporating small molecule inhibitors targeting FLIP and GLUT1. Key messages What is already known on this topic More than 20% of stage III colorectal cancer patients will experience early tumour recurrence within the first 3 years post treatment that includes surgery and adjuvant 5-FU based chemotherapy treatment. Several studies pointed towards involvements of number of cell type specific spatial neighbourhoods in tumour progression where some immune tumour microenvironment promoting angiogenesis and intravasation events, some may provide immunosuppression. Cancer stem cells could be responsible for metastatic tumour spread, early recurrence and chemoresistance. What this study adds Spatial single cell quantitative multiplex profiling of 45 cancer hallmark proteins and 15 cell identity markers in 493 stage III CRC patients tissue samples demonstrated significant differences in cellular proximity neighbourhoods, cell type specific abundance and expression between the early and late recurrence samples. We discover that macrophages show spatial association with the blood vessels in early recurrence samples. Moreover, we observed conglomeration of B cells and macrophages with Tregulatory, Thelper and Tcytotoxic cells in association with early recurrences. We showed that stromal abundance of Tregulatory, Thelper, Tcytotoxic cells and monocytes are significantly in late, and no recurrence samples compared to early recurrence samples. The most differential expression profile that differentiates late and no recurrence samples from the early recurrence samples is related to the stem cell population. Particularly, we found overexpression of GLUT1, FLIP and downregulation of BAX, BAK, MLKL and CDX2 proteins in the cancer stem cell of early recurrence samples. We built a neural network based on the cancer stem cell protein signature (BAX, MLKL, FLIP, GLUT1 and CDX2 proteins) that delivers a high-performance prognostic classifier. How this study might affect research, practice or policy Our results propose a clinically promising prognostic tool based on a five-protein stem cell signature that outperforms existing clinical and proposed transcriptomic based signatures for separation between risk groups. Moreover, our five-protein signature markers not only predict stem cell chemotherapy resistance and therefore tumour recurrence but also suggest potential therapeutic strategies. For instance, this approach could guide combinatorial treatments at high risk of chemoresistance, such as incorporating small molecule inhibitors targeting FLIP (currently in discovery phase) and GLUT1 (already under preclinical trial evaluation).