Macrophage Gene Signatures Predict Colorectal Cancer Prognosis, Machine Learning Study Finds

Macrophages, the scavenger cells of the immune system, have long been recognized as double agents inside tumours. Depending on the molecular cues they receive, they can rally anti-cancer defences or, more insidiously, feed tumour growth, sculpt blood vessels and shield the malignancy from immune attack. In colorectal cancer, one of the most common and lethal malignancies worldwide, these tumour-associated macrophages, or TAMs, are abundant players in the tumour microenvironment. Yet pinning down exactly which macrophage behaviours matter for a patient’s survival has remained a stubborn challenge. A new study published in Cancer Cell International by a team of colorectal surgeons and researchers at the First Affiliated Hospital of Fujian Medical University in Fuzhou, China, now offers a high-resolution answer, combining bulk and single-cell RNA sequencing with a battery of machine learning algorithms to isolate macrophage gene signatures that track with prognosis.

The central difficulty the researchers set out to solve is a familiar one in genomics. Bulk RNA sequencing, which measures the average gene activity across millions of cells in a tissue sample, is cheap and widely available but blurs the contributions of individual cell types. Single-cell RNA sequencing, by contrast, resolves gene expression cell by cell, revealing the dazzling diversity of immune cells infiltrating a tumour, but it is expensive and applied to relatively few samples. The team, led by Ye Wang, Zhicheng Zhuang and Tengyi Peng, who contributed equally to the work, downloaded large volumes of both bulk and single-cell CRC data from multiple public databases and set about integrating the two scales of information into a single, coherent picture of the colorectal tumour microenvironment.

The linchpin of their integrative strategy is an algorithm called Scissor. Rather than treating bulk and single-cell datasets as separate worlds, Scissor correlates the two, using bulk tumour samples annotated with clinical outcomes and genomic data to tag individual cells in the single-cell data as either associated with a poor phenotype or a favourable one. Cells linked to adverse outcomes are designated Scissor-positive, while those linked to better outcomes are Scissor-negative. When the researchers applied this framework to macrophages infiltrating colorectal tumours, a striking pattern emerged: the Scissor-positive subgroup of TAMs, the macrophages whose expression patterns mirrored the molecular profiles of tumours from patients with poor outcomes, was strongly associated with worse prognosis. In other words, not all macrophages are equal, and a specific transcriptional state of these cells carries a measurable survival penalty.

Having split TAMs into these prognostically distinct subtypes, the team moved to identify the characteristic genes that define the dangerous macrophage state. By comparing the gene expression profiles of Scissor-positive and Scissor-negative macrophages, and cross-referencing with clinical follow-up data, they pinpointed a set of TAM subset genes closely tied to patient survival. Among these, one gene stood out: SPP1, which encodes osteopontin, a secreted protein with well-documented roles in cell adhesion, migration and immune modulation. To probe what SPP1 actually does in colorectal cancer cells, the researchers turned to the laboratory bench, performing colony formation assays to measure proliferative capacity, scratch assays to gauge cell migration, and invasion assays to test the cells’ ability to penetrate surrounding matrix. These functional experiments provided direct evidence that SPP1 is not merely a passive biomarker but an active participant in the malignant behaviour of colorectal cancer cells.

The next phase of the study tackled a problem that has increasingly dogged the field of computational oncology: with dozens of machine learning algorithms available, each with its own biases and assumptions, how do you decide which model best captures the biology? The team deployed multiple machine learning techniques in parallel and compared their outputs, ultimately converging on nine TAM signature genes associated with colorectal cancer progression. This consensus-driven approach, letting several algorithms vote rather than trusting a single method, is designed to produce gene sets that are robust rather than artefacts of one particular modelling framework. The nine-gene signature then served as the backbone for a risk stratification scheme, in which each patient with colorectal cancer could be assigned a risk score based on the expression levels of these macrophage-derived genes in their tumour.

The clinical payoff of this risk score is potentially substantial. Colorectal cancer management currently leans heavily on tumour stage, nodal status and a handful of established markers, but outcomes among patients with apparently similar disease can diverge dramatically. A macrophage-based signature offers a window into the immunological character of an individual’s tumour, information that conventional staging cannot capture. Patients flagged as high risk by the TAM signature might, in principle, be candidates for more intensive surveillance or adjuvant therapy, while those in the low-risk group could potentially be spared unnecessary treatment. The study’s authors suggest that this TAM-related prognostic model could become an important tool in the management of patients with colorectal cancer, complementing rather than replacing existing clinical criteria.

To make the signature usable at the bedside, the researchers went one step further and linked the TAM signature genes to clinicopathological features, the standard clinical variables such as patient characteristics and disease measures that oncologists already collect. From this integration they constructed a nomogram, a graphical calculating tool long favoured in oncology for translating complex multivariable models into a single, individualized probability estimate. By plotting a given patient’s values on the nomogram’s axes and summing the corresponding points, a clinician can obtain a personalized assessment of prognosis that blends the molecular intelligence of the macrophage signature with the familiar anchors of clinical practice. It is a concrete step toward turning single-cell atlas data, which can feel abstract and remote, into a decision-support instrument.

The study also speaks to a broader shift in how cancer immunology is being conducted. The tumour microenvironment has moved from the margins of cancer research to its centre, driven in large part by the success of immunotherapies that unleash immune cells against tumours. But immunotherapy responses are notoriously uneven, and macrophages are increasingly implicated as one reason why: in many tumours, TAMs actively suppress T-cell activity and promote tissue remodeling that favours invasion and metastasis. By dissecting TAM heterogeneity at single-cell resolution and tying specific macrophage states to patient survival, the Fujian team’s work adds to a growing body of evidence that the immune contexture of a tumour, not just its genetic mutations, shapes its clinical trajectory. The identification of SPP1 as a functional driver further suggests that macrophage-related pathways could be druggable targets in their own right.

Caveats remain, as they do for any study built primarily on computational analysis of public datasets. Gene expression signatures derived from retrospective cohorts must be validated in independent, prospectively collected patient populations before they can inform treatment decisions, and the biological mechanisms linking the nine-gene signature to tumour progression will require deeper experimental dissection beyond the SPP1 functional assays reported here. The authors themselves frame the work as an analysis of the potential association between TAMs and colorectal cancer prognosis, a foundation on which future studies can build. Still, the methodological template, fusing bulk and single-cell transcriptomics through the Scissor algorithm, stress-testing candidate gene sets with multiple machine learning approaches, and validating key players in the wet lab, represents a rigorous and reproducible pipeline that other cancer types could readily adopt.

For patients with colorectal cancer, the promise is a future in which a routine genomic readout of their tumour reveals not only which mutations it carries but what kind of immune neighbourhood it inhabits, and in which that information translates into a personalized risk estimate and a tailored treatment plan. For researchers, the study delivers a curated set of macrophage signature genes and a functional lead in SPP1 to interrogate further. The work was supported by the National Natural Science Foundation of China, the Fujian Province Finance Project and talent programs from the First Affiliated Hospital of Fujian Medical University, and it appears as an open-access publication, making the full technical detail available to clinicians and computational biologists alike. As single-cell technologies mature and machine learning methods are tamed by consensus approaches, the macrophages lurking inside colorectal tumours are giving up their secrets, one transcript at a time.

Subject of Research: Tumour-associated macrophage gene signatures and prognostic modelling in colorectal cancer using bulk and single-cell RNA sequencing and machine learning

Article Title: Bulk and single-cell RNA-sequencing analyses along with machine learning methods identify macrophage signatures in colorectal cancer

Article References: Wang, Y., Zhuang, Z., Peng, T., Pan, Z., Zhang, Q., Zheng, S., Liu, X., Zhuang, J., Guan, G., Li, S., & Chen, B. (2026). Bulk and single-cell RNA-sequencing analyses along with machine learning methods identify macrophage signatures in colorectal cancer. Cancer Cell International. https://doi.org/10.1186/s12935-026-04445-6

Image Credits: AI Generated

DOI: 10.1186/s12935-026-04445-6

Keywords: colorectal cancer, tumour-associated macrophages, single-cell RNA sequencing, bulk RNA sequencing, machine learning, Scissor algorithm, tumour microenvironment, SPP1, prognostic signature, nomogram, gene expression profiling, transcriptomics