The PD-L1 biomarker problem. In NSCLC, PD-L1 expression on tumor cells is the most widely used biomarker for selecting patients for immunotherapy. However, PD-L1 has well-recognized limitations: its expression varies in different regions of the same tumor, changes over time as the tumor evolves, and results differ depending on which antibody and staining protocol is used. These factors explain why many patients with high PD-L1 do not respond, and some with low PD-L1 do.
The spatial dimension of immune response. The immune cells within a tumor do not act independently - their effectiveness depends on where they are located relative to tumor cells, which immune cell types they are adjacent to, and how these spatial relationships change during treatment. A T cell sitting next to a tumor cell acts very differently from one isolated in surrounding connective tissue. Capturing this spatial architecture requires methods that preserve and analyze the position of each cell within the tissue.
Multiplexed imaging as a solution. Multiplexed fluorescence imaging simultaneously marks seven different proteins on a single tissue slice, allowing researchers to identify and locate multiple cell types - tumor cells, effector T cells, regulatory T cells, checkpoint markers - within the same image. This approach generates data on both cell identity and spatial position that cannot be obtained from bulk molecular profiling or single-protein staining.
Study objective. This study analyzed pre- and during-treatment multiplexed images from nine NSCLC patients who had already failed standard immunotherapy and were then treated with a combination of an HDAC inhibitor (vorinostat) and a PD-1 inhibitor (pembrolizumab). The goal was to discover whether spatial immune patterns in pre-treatment biopsies could predict which patients would develop stable disease versus progressive disease, and whether these patterns could serve as companion biomarkers alongside PD-L1.
Patient cohort and imaging. Ninety-two multiplexed field-of-view images (FoVs) were collected from nine Stage 4 NSCLC patients enrolled in a Phase I clinical trial (NCT02638090) at Moffitt Cancer Center. Five patients achieved stable disease (SD) for at least 24 weeks; four had progressive disease (PD) before 24 weeks. Each image was stained with seven markers: DAPI (nuclei), CD3 (T cells), CD8 (effector T cells), FoxP3 (regulatory T cells), PD-1 (exhaustion marker), PD-L1 (checkpoint ligand), and Pan-cytokeratin/PanCK (epithelial and tumor cells).
Cell segmentation approach. A U-Net convolutional neural network was trained to identify and segment individual cell nuclei within each image. Using Gaussian Mixture Model clustering on the marker expression of each segmented cell, the pipeline automatically identified heterogeneous cell subpopulations without requiring a pre-defined list of cell types. Tumor-rich regions were computationally outlined using convex hulls around cell clusters with high PanCK expression, allowing separate analysis of tumor core versus tumor border versus surrounding stroma.
Quadrat approach. Complementary to single-cell analysis, each image was divided into 100 micrometer by 100 micrometer quadrats - fixed-size tiles that capture the aggregate marker density within a multicellular region. This spatial ecology approach, borrowed from landscape ecology, allows analysis of how marker communities are distributed and associated across the tissue without requiring precise cell-by-cell segmentation. Channel bleed between fluorescence signals was minimized through mutual exclusivity assignments and intensity subtraction between adjacent fluorophore channels.
Cellular and quadrat neighborhoods. For each segmented cell, a cellular neighborhood (CN) was defined by averaging the marker expression of its ten nearest spatial neighbors, encoding not just what that cell expresses but what cell types surround it. Similarly, quadrat neighborhoods (QNs) were identified by clustering adjacent quadrats using the Leiden algorithm. Species association networks (SANs), estimated using Gaussian copula graphical lasso, revealed which markers are directly co-localized versus those that appear associated only due to indirect effects.
Pre-treatment ecology already predicts outcome. Non-spatial analysis using positive pixel counts from quadrats showed that PD tumors had significantly different marker compositions from SD tumors even before treatment began (PERMANOVA p-value = 0.007). This means the immune environment that determines whether a patient will respond or progress is already established at diagnosis, not created by the treatment itself.
Characteristic cellular neighborhoods. Spatial clustering identified twelve distinct cellular neighborhoods across all images. Progressive disease patients were characterized by neighborhoods dominated by PanCK+FoxP3+PD-1+ cells - tumor cells surrounded by exhausted T cells (PD-1+) and immunosuppressive regulatory T cells (FoxP3+). Stable disease patients showed opposing neighborhoods dominated by PanCK+PD-L1+ with CD3+CD8+ - tumor cells adjacent to active cytotoxic T cells, a pattern associated with functional anti-tumor immunity.
Species association networks confirm distinct architectures. SANs revealed that in pre-treatment PD biopsies, FoxP3 was positively associated with PD-1 and PanCK, indicating clustering of regulatory T cells near tumor cells. In SD biopsies, CD8 was positively associated with PD-L1 and PanCK, indicating cytotoxic T cell proximity to tumor cells. These network-level differences in marker co-localization confirm that PD and SD patients have fundamentally different spatial immune architectures before any treatment is administered.
Treatment further separates the groups. On-treatment biopsies showed that the treatment shifted the ecology of each group, as visualized by centroid displacement in multi-dimensional scaling plots. However, the direction of change differed between groups, suggesting that the immunological response to vorinostat plus pembrolizumab was inherently constrained by the pre-existing immune landscape rather than creating a uniform shift toward immune activation across all patients.
SVM classification from single-cell data. A support vector machine (SVM) trained on single-cell marker expression using leave-one-out cross-validation achieved an overall classification accuracy of 87.5%, correctly identifying all five SD patients (100% recall) and two of three PD patients used in the analysis (66.7% recall). The markers most important for predicting SD were CD3, CD8, and PD-L1, while FoxP3 and PD-1 were most predictive of progressive disease - directly matching the spatial neighborhood findings.
Neural network species distribution model from quadrats. A feedforward neural network trained on quadrat counts achieved an overall accuracy of 75%, with 100% recall for PD patients and 60% for SD patients. A feature importance analysis identified PD-1, FoxP3, CD8, and PanCK as the markers with the greatest influence on the model's predictions, again consistent with the identified cellular neighborhoods. The model also generated spatial risk maps for each image, coloring quadrats by probability of disease progression and enabling visualization of where within each tumor the risk is highest.
PD-L1 alone performs poorly. When either model was trained using only PD-L1 expression, prediction accuracy dropped to 63% - barely better than random. Clinical PD-L1 status assigned to patients at enrollment also misclassified three of nine patients, meaning the theoretical maximum accuracy from PD-L1 status alone was 66.7%. The multi-marker spatial models at 75-88.5% accuracy significantly outperformed this ceiling.
Ratio of Geometric Means as a clinical biomarker. A single clinical biomarker was derived by calculating the ratio of geometric means (RoGM) of SD-associated neighborhoods (PanCK+PD-L1+ with CD3+CD8+) to PD-associated neighborhoods (PanCK+FoxP3+PD-1+) for each patient. This ratio clearly separated SD from PD patients without overlap, and Kaplan-Meier survival curve analysis confirmed that the RoGM-based separation matched the actual survival differences between groups, providing a potential single-number companion biomarker grounded in spatial immune architecture.
Inferring marker interaction networks. To quantify how marker interactions differ between SD and PD patients, and how they change from pre-treatment to on-treatment, a least-squares minimization framework was used to infer interaction weight matrices. The approach modeled the change in each marker's expression over time as a function of interactions with all other markers, constrained by prior biological knowledge of known marker relationships.
SD interaction pattern. The inferred weight matrix for SD patients showed increased CD8 interaction with tumor cells, decreased FoxP3 interaction with tumor cells, and a positive interaction between CD8, PD-L1, and PanCK. This network portrait captures an environment in which cytotoxic T cells are functionally engaging with tumor cells in the context of active checkpoint signaling - conditions associated with potential therapeutic response to PD-1 blockade.
PD interaction pattern. For progressive disease patients, the weight matrix showed the opposite: increased FoxP3 interaction with tumor cells, decreased CD8 interaction with tumor, and positive interactions between FoxP3, PD-1, and PanCK. This network portrait describes a deeply immunosuppressive niche where regulatory T cells are dominating the tumor border and exhausted T cells cannot mount an effective response.
Consistency across methods. The network model findings were consistent with those from the image t-SNE analysis, cellular neighborhood clustering, and quadrat neighborhood analysis. The agreement between these independently implemented approaches confirms that the identified immune ecology differences between SD and PD patients are real biological signals rather than computational artifacts of any single analytical method.
Spatial biomarkers as companions to PD-L1. The study demonstrates that pre-treatment tissue immune architecture carries predictive information about immunotherapy response that is largely invisible to PD-L1 testing alone. The identified biomarker combinations - specifically the ratio of CD8+PD-L1+ neighborhoods to FoxP3+PD-1+ neighborhoods around tumor cells - represent a potentially actionable and interpretable companion biomarker that could be measured from routine pre-treatment biopsies using multiplexed staining.
Importance of spatial context. A key finding was that performing the same ratio analysis without spatial information (ignoring cell positions) failed to separate SD from PD patients cleanly. Only when the spatial arrangement of markers was preserved did the RoGM metric achieve clean group separation. This demonstrates that spatial context is not merely informative but essential - the same cells arranged differently would yield opposite predictions.
Study limitations. The cohort was small (nine patients, 92 images), retrospective, and included only stable and progressive disease responses without complete or partial responders. All patients had already failed immunotherapy, so results may not generalize to treatment-naive patients. PanCK positivity was used as a proxy for tumor cells, but normal epithelial cells can also express PanCK. Different biopsy sites (primary vs. metastatic) across patients introduced potential confounding related to site-specific immune differences.
Ongoing validation and future directions. A larger randomized Phase II trial (NCT02638090) is underway with 78 patients in two arms comparing pembrolizumab monotherapy to pembrolizumab plus vorinostat. This will allow validation of the identified biomarkers across a wider range of responses and in a monotherapy context. Mouse studies are planned to mechanistically test whether the FoxP3+PD-1+ ecosystem actively prevents immune response or is simply a marker of pre-existing immune suppression.