---

The mathematical oncology community has long been interested in understanding and modeling how tumor tissue content—the tumor microenvironment (TME)—affects clinical outcomes. These efforts often rely on observations that certain cell types correlate with clinical outcomes. However, several nuances must be considered when interpreting these findings.

In the following sections, I discuss key points I’ve encountered during the past three years of studying the role of the breast tumor microenvironment in clinical outcomes and writing two papers on this topic.

Clinical Outcomes: A Complex Landscape

Much of what we know about the role of cell types comes from clinical trial results. However, clinical trials vary in their design, objectives, and endpoints. As a result, their findings don’t always measure the same outcomes, and the differences can be dramatic.

Consider two types of clinical outcomes:

1. Response to neoadjuvant chemotherapy (NAC) that is the treatment given before surgery, aiming to shrink the tumor.
2. Relapse-free survival (RFS): The time from surgery to tumor relapse.

One might assume that a cell type positively associated with one outcome would also associate positively with the other, but this is not always the case. Tumors with better response rates to NAC do not necessarily have longer RFS. Consequently, a cell type can show opposite associations with response to NAC and RFS. In [2], we demonstrate that macrophages, CD8+ T cells, and some cancer cell types exhibit such behavior.

The complexity deepens when comparing RFS and overall survival (OS), two highly correlated outcomes that are still distinct. Exploring these associations can yield differing results.

Cancer Subtypes: A Factor in Variability

Each cancer type comprises distinct subtypes with unique behaviors. Associations between cell types and outcomes can change—or even reverse—across subtypes, requiring caution when generalizing results. For example, in [1], we found that dendritic cells (DCs) and natural killer T cells (NKT cells) exhibit opposite associations with NAC response across different tumor subtypes.

Measurement Challenges

Outcome definitions can vary significantly. For instance in NAC, achieving a "pathological complete response" (pCR, pathologically defined as complete elimination of cancer cells in the breast and auxiliary lymph nodes in the case of breast cancer) might not align with single-cell RNA sequencing (scRNA-seq) results, which could identify cancerous cells in tissue deemed cancer-free by pathology.

Treatment-Driven Effects

Clinical outcomes often depend on specific treatments. Findings may reflect the effects of the treatment rather than underlying tumor biology. As such, generalizing results from one clinical trial requires additional validation.

Defining Cell Types: Challenges in Annotation

Our ability to identify cell types has advanced significantly with technologies like scRNA-seq. However, challenges persist:

1. Annotation Variability: Cell type annotations depend on the technology used and can vary across studies. Some cell types that are easily identified in scRNA-seq data may be difficult to distinguish in other datasets. Even within the same technology, annotations can differ across datasets, and some cell types are annotated with greater confidence than others.
2. Spatial Context: The functions of cells often depend on their spatial location. The same cell type might show opposite associations with outcomes based on its location within the tumor or microenvironment. In [1], we show that the distance of DCs from epithelial cells affects their association with response to NAC. Similarly, in [2], macrophage distance from a specific group of epithelial cells reverses their association with RFS.

Correlation vs. Causation: A Fundamental Challenge

As modelers, we strive to incorporate causal relationships into our work. However, clinical trial results often reveal correlations rather than causality. Establishing causal links requires more comprehensive analyses to identify the functional aspects of observed findings, which can then be integrated into models.

Conclusion

With the growing availability of data, we have new opportunities to develop models that bridge gaps, connect datasets, and uncover actionable insights. By addressing the nuances of clinical outcomes, cell type definitions, and causal relationships, we can pave the way for more personalized and effective cancer treatments.

References

1. Azimzade, Y., Haugen, M. H., Tekpli, X., Steen, C. B., Fleischer, T., Kilburn, D., Ma, H., Egeland, E. V., Mills, G., Engebraaten, O., Kristensen, V. N., Frigessi, A., & Köhn-Luque, A. (2023). Explainable Machine Learning Reveals the Role of the Breast Tumor Microenvironment in Neoadjuvant Chemotherapy Outcome. Cold Spring Harbor Laboratory. doi: 10.1101/2023.09.07.556655
2. Azimzade, Y., Haugen, M. H., Kristensen, V. N., Frigessi, A., & Köhn-Luque, A. (2024). Integrated multiomics analysis unveils how macrophages drive immune suppression in breast tumors and affect clinical outcomes. Cold Spring Harbor Laboratory. doi: 10.1101/2024.11.09.622776