How can spatial biology predict immunotherapy response?

By measuring the physical distance between specific immune cells and tumor cells, rather than only the presence or density of a single marker. Multiple studies have found that proximity metrics, such as the percentage of T cells within a defined distance of tumor cells, predict checkpoint inhibitor response more accurately than PD-L1 status alone in the same patients.

Why is PD-L1 alone a poor predictor?

Because PD-L1 expression says nothing about whether the cells expressing it are positioned to actually engage the T cells a checkpoint inhibitor depends on. A 2025 study found a spatial proximity score achieved an area under the curve of 0.79 for predicting response, compared with 0.58 for PD-L1 tumor proportion score alone, in the same low-PD-L1 patient subgroup.

What are spatial predictive biomarkers?

Biomarkers built from the physical arrangement of cells within tissue, such as the nearest-neighbor distance between immune and tumor cells or the density of one cell type within a fixed radius of another, rather than a single protein’s abundance alone. These have outperformed traditional density-based or single-marker biomarkers for predicting immunotherapy response in several recent studies across multiple cancer types.

Predicting immunotherapy response with spatial data

Highlights in this article

The limits of single-marker prediction
Adding spatial context to PD-L1
Proximity and interaction metrics
Multi-marker spatial models
Toward clinical use

Relying on PD-L1 testing alone to predict who will respond to checkpoint inhibition has always been an imperfect proxy, and a growing body of evidence now shows precisely how much predictive power is left on the table by measuring PD-L1 expression without regard to where, exactly, the relevant cells sit relative to one another. In more than one recent study, a spatial proximity measurement has directly outperformed PD-L1 status in the same patients, which is a considerably stronger claim than simply noting that spatial data adds context.


Key takeaways A 2025 study found a spatial proximity score outperformed PD-L1 tumor proportion score directly, with an area under the curve of 0.79 versus 0.58, in the same low-PD-L1 patient subgroup. A separate 2024 study found immune cell density did not discriminate between responders and non-responders, while a specific nearest-neighbor distance metric did. In hepatocellular carcinoma, a proximity-defined interaction between PD-L1-positive macrophages and CD8-positive T cells, measured within 25 micrometers, associated with response to an approved checkpoint combination. A breast cancer study found the proximity effect held across tumor subtypes and treatment regimens, with a proposed chemokine mechanism explaining why cells end up that close together in responders. Multi-marker spatial models increasingly combine proximity with co-expression and other spatial features rather than relying on any single metric.

Key takeaways

A 2025 study found a spatial proximity score outperformed PD-L1 tumor proportion score directly, with an area under the curve of 0.79 versus 0.58, in the same low-PD-L1 patient subgroup.
A separate 2024 study found immune cell density did not discriminate between responders and non-responders, while a specific nearest-neighbor distance metric did.
In hepatocellular carcinoma, a proximity-defined interaction between PD-L1-positive macrophages and CD8-positive T cells, measured within 25 micrometers, associated with response to an approved checkpoint combination.
A breast cancer study found the proximity effect held across tumor subtypes and treatment regimens, with a proposed chemokine mechanism explaining why cells end up that close together in responders.
Multi-marker spatial models increasingly combine proximity with co-expression and other spatial features rather than relying on any single metric.

The limits of single-marker prediction

PD-L1 immunohistochemistry remains the most widely used biomarker for selecting patients for checkpoint inhibitor therapy, and it has never been a strong one in isolation. Response rates among PD-L1-positive patients are meaningfully better than among PD-L1-negative patients on average, but the overlap between the two groups is substantial enough that a considerable share of PD-L1-negative patients still respond, and a considerable share of PD-L1-positive patients do not.

A specific mechanistic reason for that overlap is now well documented: PD-L1 expression alone says nothing about whether the cells expressing it are actually positioned to engage the T cells a checkpoint inhibitor is meant to unleash. A tumor with abundant PD-L1 expression scattered throughout, far from any T cell, presents a fundamentally different clinical picture than a tumor with the same total PD-L1 expression concentrated immediately adjacent to an exhausted T cell population, and a bulk or even a simple density-based measurement cannot distinguish the two.

Adding spatial context to PD-L1

The clearest available evidence that spatial context adds real predictive power, rather than a marginal refinement, comes from a direct head-to-head comparison in the same patient cohort.


A direct comparison, in the same patients A 2025 study analyzed pretreatment tumor samples from 46 patients with advanced non-small cell lung cancer treated with PD-1/PD-L1 inhibitors, using multiplex immunohistochemistry to assess the spatial proximity of CD8-positive, FoxP3-positive, PD-1-positive cells to tumor cells. In patients with low PD-L1 tumor proportion scores, the spatial proximity score predicted response with an area under the curve of 0.79, compared with 0.58 for PD-L1 tumor proportion score alone measured in the same patients. A low proximity score was also identified as an independent risk factor for shorter progression-free survival, with a hazard ratio of 6.16. That is a direct, quantified answer to whether spatial context outperforms the standard biomarker in the same population, not an inference drawn by comparing separate studies with different patients and different methods.

A direct comparison, in the same patients

A 2025 study analyzed pretreatment tumor samples from 46 patients with advanced non-small cell lung cancer treated with PD-1/PD-L1 inhibitors, using multiplex immunohistochemistry to assess the spatial proximity of CD8-positive, FoxP3-positive, PD-1-positive cells to tumor cells. In patients with low PD-L1 tumor proportion scores, the spatial proximity score predicted response with an area under the curve of 0.79, compared with 0.58 for PD-L1 tumor proportion score alone measured in the same patients.

A low proximity score was also identified as an independent risk factor for shorter progression-free survival, with a hazard ratio of 6.16. That is a direct, quantified answer to whether spatial context outperforms the standard biomarker in the same population, not an inference drawn by comparing separate studies with different patients and different methods.

A separate 2024 study makes the same point from a different angle, focused specifically on what density-based measurement misses. Analyzing 24 baseline urothelial and head and neck cancer samples from patients treated with combination checkpoint inhibitors, researchers found that immune cell density did not discriminate between response groups. However, modeling the nearest-neighbor distance from CD8-positive T cells or macrophages to their closest cancer cell using a Weibull distribution did positively associate with response. In the same dataset, the traditional metric failed while the spatial metric succeeded, which is a cleaner demonstration of the underlying problem than two independent studies reaching different conclusions in different populations.

Proximity and interaction metrics

Several related but distinct proximity measurements recur across this literature, and understanding what each specifically quantifies matters for interpreting or building a spatial biomarker correctly.

Metric	What it measures	Example finding
Nearest-neighbor distance	The distance from each cell of interest to its single closest neighbor of another defined type	Modeled with a Weibull distribution, associated with response in urothelial and head and neck cancer despite density showing no association
Proximity score	A composite or normalized measure of how close a defined cell population sits to tumor cells across a whole sample	Outperformed PD-L1 TPS directly, AUC 0.79 versus 0.58, in low-PD-L1 NSCLC patients
Interaction variable within a fixed radius	The count or density of one cell type within a specific, defined distance, such as 20 or 25 micrometers, of another	CD8-positive T cells within 25 micrometers of PD-L1-positive macrophages associated with response to an approved checkpoint combination in hepatocellular carcinoma

Table 1. Three distinct proximity metrics used across the current literature. Each answers a slightly different question, and comparing results across studies requires checking which specific metric was used.

The hepatocellular carcinoma example in that table is worth expanding, because it ties a proximity metric directly to an approved combination regimen rather than a research-stage question. A study of patients treated with atezolizumab plus bevacizumab defined an interaction variable using nearest-neighbor analysis, specifically counting CD8-positive T cells within 25 micrometers of PD-L1-positive tumor-associated macrophages, normalized to overall cell counts. That spatially defined interaction, rather than either cell population’s density measured separately, associated with tumor shrinkage and progression-free survival in patients receiving this specific, currently used treatment regimen.

Multi-marker spatial models

Proximity alone is not the whole story, and a separate breast cancer study supplies both cross-cancer generalizability and a proposed mechanism that helps explain why physical closeness itself matters biologically rather than simply correlating with response for unrelated reasons.

That study found a higher percentage of CD8-positive T cells within 20 micrometers of cancer cells strongly correlated with pathological complete response, disease-free survival, and overall survival, and this held regardless of tumor subtype or treatment regimen, a meaningfully broader claim than a single-cancer-type finding. The same study identified a positive correlation between CXCL9 expression and that proximity measurement, suggesting this specific chemokine may facilitate the recruitment of CD8-positive T cells toward cancer cells in the first place. That proposed mechanism matters because it moves the finding from a purely descriptive correlation, T cells that end up close to tumor cells predict better outcomes, toward an explanation of why: a chemokine gradient may be actively drawing them there, which is itself a potential target or biomarker in its own right.

The direction current spatial biomarker development is heading, evident across all four studies above, is toward combining proximity metrics with co-expression data, cell density, and increasingly transcriptomic or chemokine signaling context, rather than relying on any single spatial measurement alone. The multi-marker co-expression approach specifically, and how it combines with the proximity metrics described here, is developed in full in Spatial Biology in Immuno-Oncology: Targets, Signatures, and Combinations.

Toward clinical use

Moving any of these proximity metrics from a research finding into routine clinical use faces the same regulatory and analytical validation requirements as any other spatial biomarker, covered in full in Spatial Biomarkers and Companion Diagnostics: The Next Frontier. Two considerations specific to proximity metrics are worth flagging here.

The specific distance threshold used matters and is not yet standardized. Studies described above use 20 and 25 micrometer thresholds for defining an interaction variable, and a Weibull-distribution approach that does not use a fixed threshold at all; a clinical assay will need a validated, consistent definition rather than each laboratory choosing its own cutoff.
Multiplex tissue imaging platforms and analysis software differ in how they compute distance. Reproducing a proximity score across different imaging platforms and different analysis pipelines is its own validation question, distinct from confirming the underlying biological association is real.

The broader translational and clinical workflow questions these validation requirements feed into are addressed in [LINK: Translating Spatial Biology Into the Clinic: Digital Pathology and Beyond].

For the broader tumor immune microenvironment argument this spoke builds on, see Spatial Biology in Oncology: Decoding the Tumor Microenvironment, and for where this sits within the full spatial biology pipeline, Spatial biology in drug discovery: From target discovery to translational medicine.

This article was produced under Drug Discovery News’s AI editorial policies.