A machine learning-derived immune-related prognostic model identifies PLXNA3 as a functional risk gene in colorectal cancer

Heng Lyu, Song Wang, Guodong Guo, Wen‐chang Lin, Chenshen Huang, Hong Chen, Chao Xu, Li Liu, Qi Huang, Fangqin XueFrontiers in Immunology 2025 · DOI

TL;DR

Researchers used a "stacked ensemble" of over 100 machine learning algorithms to identify PLXNA3 as a critical risk gene in colorectal cancer (CRC). The study reveals that PLXNA3 not only predicts poor patient survival but also drives "immune exclusion," effectively creating a shield that keeps cancer-fighting T cells away from the tumor.

What Problem Does This Paper Address?

Colorectal cancer remains a leading cause of cancer-related deaths globally. A major hurdle in treatment is interpatient heterogeneity: two patients with the same "stage" of cancer can have vastly different outcomes because their tumors are driven by different genetic and immunological profiles.

Current clinical indicators often fail to capture the complexity of the tumor microenvironment (TME). There is an urgent need for robust biomarkers that can predict prognosis and identify patients who might not respond to standard therapies or immunotherapy. The authors sought to move beyond simple statistical models to find "master regulator" genes that influence both tumor growth and immune evasion.

Methods

The authors employed a "brute-force" meets "high-refinement" computational approach:

  • Stacked Ensemble Learning: Instead of picking one algorithm, they integrated over 100 machine learning combinations (including LASSO, CoxBoost, and StepCox). They selected the model with the highest concordance index (C-index) to ensure the most stable prognostic predictions.
  • Multi-Omics Integration: The study didn't stop at computer code. They validated their findings using bulk transcriptomics, spatial transcriptomics (mapping gene expression to specific physical locations in a tissue slice), and single-cell RNA sequencing (scRNA-seq).
  • Functional Validation: They used DepMap (CRISPR-based gene dependency data) to see if cancer cells actually "depend" on PLXNA3 for survival and performed Immunohistochemistry (IHC) on 22 clinical samples from Fujian Provincial Hospital to confirm protein-level expression.
  • Drug Repurposing: Using the Connectivity Map (CMap) and drug sensitivity databases (PRISM, GDSC), they screened thousands of compounds to find those capable of reversing the PLXNA3 transcriptional signature.

Datasets

The study leveraged several massive public and private repositories:

  • TCGA-COAD: The Cancer Genome Atlas (Colon Adenocarcinoma) for primary training.
  • GSE39582: A Gene Expression Omnibus (GEO) dataset for external validation.
  • GTEx: Used as a baseline for gene expression in healthy tissues.
  • Sparkle & TISCH2: Databases for spatial and single-cell transcriptomics, respectively.
  • Human Protein Atlas (HPA): For protein-level validation across various organs.

Key Insights and Ideas

  • PLXNA3 as a "Cold" Tumor Driver: The most striking finding was the negative correlation between PLXNA3 and immune infiltration. Spatial transcriptomics showed PLXNA3 is highly expressed in malignant epithelial regions but almost entirely absent in areas rich with CD8+ T cells and plasma cells. This suggests the gene helps the tumor create an "immune-excluded" environment.
  • The "Aha" Moment in the TME: While PLXNA3 was originally known for its role in axon guidance (neural development), this paper highlights its "moonlighting" role in cancer. It appears to suppress interferon-gamma signaling and DNA repair pathways, making the tumor more aggressive and harder for the immune system to recognize.
  • Regulatory Hub: The authors identified specific transcription factors (like MZF1 and OGT) and microRNAs (like hsa-miR-221-3p) that act as upstream "volume knobs" for PLXNA3, providing multiple potential points of therapeutic intervention.

Key Results

  • Prognostic Power: The machine learning-derived risk score was an independent predictor of survival. In a multivariate analysis, the risk score yielded a Hazard Ratio (HR) of 2.2446 ($p < 0.001$), outperforming several traditional clinical metrics.
  • Predictive Accuracy: The ensemble model achieved a C-index of 0.645 and maintained consistent performance across 1-, 3-, and 5-year survival predictions (AUCs ~0.68–0.72).
  • Clinical Confirmation: IHC staining confirmed that PLXNA3 protein levels are significantly higher in tumor tissues compared to adjacent normal tissues ($p = 0.0005$) and showed a trend of increasing with advanced nodal (N) stage.
  • Therapeutic Leads: The study identified imatinib, fasudil, and MS-275 (an HDAC inhibitor) as potential drugs that could reverse the harmful effects of PLXNA3 overexpression.

Limitations and Future Work

The authors acknowledge several caveats:

  1. Retrospective Data: The study relies heavily on public, retrospective datasets which can introduce batch effects and clinical heterogeneity.
  2. Immunotherapy Gap: Due to a lack of large-scale CRC-specific immunotherapy cohorts with available genomic data, the authors couldn't directly test if PLXNA3 levels predict a patient's response to PD-1/PD-L1 inhibitors.
  3. Need for In Vivo Testing: While the computational and IHC evidence is strong, prospective clinical trials and animal models are needed to prove that inhibiting PLXNA3 (or its pathways) actually slows tumor growth or restores immune cell infiltration in a living system.

Summary updated

Abstract

This study identifies <i>PLXNA3</i> as a novel immune-related biomarker in colorectal cancer and elucidates its multifaceted role in tumor progression, immune evasion, and therapeutic resistance. These findings provide a foundation for incorporating <i>PLXNA3</i> into precision oncology frameworks for gastrointestinal malignancies.
BibTeX
@inproceedings{lyu2025machine,
  author = {Heng Lyu and Song Wang and Guodong Guo and Wen‐chang Lin and Chenshen Huang and Hong Chen and Chao Xu and Li Liu and Qi Huang and Fangqin Xue},
  title = {{A machine learning-derived immune-related prognostic model identifies PLXNA3 as a functional risk gene in colorectal cancer}},
  booktitle = {Frontiers in Immunology},
  year = {2025},
  doi = {10.3389/fimmu.2025.1653794},
}
Figure