Machine Learning Reveals the Association Between Gene Expression and Immune Infiltration in Colorectal Cancer: A Comprehensive Study From Single‐Cell to Survival Analysis

Colorectal cancer (CRC) is one of the most common causes of cancer mortality globally. Analysis of immune cell infiltration patterns in the tumour microenvironment (TME) is critical to treatment outcomes, but the molecular mechanisms which regulate this process are still poorly understood. We uniquely applied machine learning to single-cell RNA sequencing analysis to unravel the complex interaction between gene expression profiles and immune cell infiltration in CRC. We present a new computational framework that integrates different machine learning methods to analyse single-cell RNA sequencing data from CRC patients. The system leverages unsupervised clustering, survival, and gene-set enrichment analyses to pinpoint principal molecular signatures. CIBERSORT & ESTIMATE were used for immune cell quantification, whereas UMAP and t-SNE were used for high-dimensional data visualisation and pattern discovery. Our analyses uncovered gene expression signatures that closely associated with immune cell infiltration patterns in CRC. Using unsupervised clustering, we discovered two novel molecular subtypes that displayed markedly different outcomes (p = 0.049). We identified CD19, MAP2, CALB2 and TGFB2 as key biomarkers involved in immune modulation. However, gene enrichment analysis of these subgroups revealed new biological pathways involving the immune response. Our proposed models showed strong predictive capabilities verified by ROC curve analysis. Using single-cell analysis to identify previously uncharacterized interactions between specific immune cell populations and tumour cells, thereby uncovering novel immune evasion mechanisms and potential immunotherapy targets within the TME. Our results uncover novel candidate biomarkers for response to immunotherapy prediction and highlight molecular profiles that could support guided treatment approaches. The predictive models derived at present have the potential to be implemented in clinical practice for decision-making in CRC management.