Multimodal profiling for prediction of primary resistance to anti-PD-(L)1 therapy in advanced non-small-cell lung cancer: the prospective PIONeeR biomarkers study

Beyond PD-L1: Decoding Why Some Lung Cancer Patients Resist Immunotherapy

Why do some patients respond to immunotherapy while others see no benefit at all? For patients with advanced non-small-cell lung cancer (NSCLC), the standard of care often involves immune checkpoint inhibitors (ICIs). These drugs unmask cancer cells so the immune system can attack them. However, a significant portion of patients experience primary resistance. This means the treatment fails to stop the disease from progressing from the very start.

Researchers have studied various markers to predict this failure. Existing tools like PD-L1 expression and tumor mutational burden (TMB) are often insufficient. A new prospective study, the PIONeeR biomarkers study, suggests the answer lies in a complex, multimodal signature. By combining data from blood tests, tumor genetics, and immune cell mapping, the authors report a predictive model that significantly outperforms current clinical standards.

The Complexity of Immune Resistance

The core challenge in modern oncology is moving toward precision medicine. In NSCLC, doctors need to know before starting treatment if a patient is likely to be a "non-responder." If a patient is predicted to have primary resistance (PrR), clinicians might opt for alternative therapies. They might also enroll them in next-generation clinical trials immediately. This avoids wasting precious time on ineffective drugs.

Resistance to anti–PD-(L)1 therapy likely reflects the combined influence of multiple factors. These include tumor-intrinsic traits, the microenvironment, and systemic inflammatory states. Because these factors interact, a single biomarker is rarely enough. The PIONeeR study operates on this principle. It treats resistance as a composite biological state driven by many interacting layers.

The Biological Layers of PIONeeR

To capture this complexity, the authors conducted a prospective, multicentre study involving 439 patients. They performed "deep phenotyping" across six distinct biological layers [Figure 1B-C]. These included routine clinical data and blood chemistry. They also used high-dimensional circulating immune profiling (analyzing immune cell subsets in the blood). Other layers included soluble vascular markers, tumor genomics and transcriptomics (the study of DNA and RNA), and a detailed analysis of the tumor immune contexture using digital pathology.

The study aimed to solve a specific problem. Existing models often use "overall survival" as a metric. This can be problematic because it is influenced by many factors occurring after the initial treatment. Instead, the authors focused on primary resistance (PrR). This was defined by strict guidelines as disease progression occurring within a specific early window after starting therapy.

To ensure the machine-learning results were not merely "memorizing" noise, the researchers used a rigorous mathematical safeguard. This is called the .632 optimism-correction framework. This method trains the model on various subsets of the data. It then calculates a weighted average of performance on seen versus unseen data. This provides a more honest estimate of how the model would work in a real hospital [Figure 1D].

Outperforming the Clinical Standards

The results of this multimodal approach were striking. The authors report that while individual biomarkers showed only modest associations with resistance, the integrated signature was far more powerful. Specifically, the best performing individual biomarkers reached an AUROC (a measure of how well a model distinguishes between two groups) of only 0.64 [Figure 2A]. In contrast, the integrated multimodal signature achieved an optimism-corrected AUROC of 0.73 $\pm$ 0.05.

More importantly for clinical decision-making, the model's Positive Predictive Value (PPV) was significantly higher. The PPV represents the probability that a patient predicted to be resistant actually is resistant. The multimodal signature achieved an overall PPV of 0.63 $\pm$ 0.07. In the critical first-line treatment setting, the model achieved a PPV of 0.51 $\pm$ 0.11. This significantly outperformed the current gold standards. PD-L1 expression achieved a PPV of only 0.36 $\pm$ 0.02. Similarly, TMB achieved a PPV of 0.35 $\pm$ 0.03 [Figure 3A].

The researchers benchmarked ten machine-learning models, including random forests, gradient boosting, and multi-layer perceptrons. They found that the "drivers" of resistance vary by treatment context. In patients receiving chemotherapy combined with immunotherapy, the model relied heavily on routine blood chemistry. In contrast, for patients receiving immunotherapy alone, the model leaned more on the specific makeup of the tumor's immune environment [Figure 2C].

Mapping the Routes to Resistance

By distilling the massive dataset into an 18-feature signature, the authors visualized different ways a patient might become resistant [Figure 4A]. This signature includes surprisingly common clinical variables. The researchers found that routine blood parameters accounted for nearly half of the signature's importance. These included factors like chloride, albumin, and C-reactive protein (CRP).

Using SHAP (a method to explain the output of complex machine-learning models), the authors demonstrated that resistance is highly heterogeneous [Figure 4C-F]. For one patient, resistance might be driven by high systemic inflammation. For another, it might be caused by a lack of specific regulatory T-cells in the tumor stroma. This confirms that "resistance" is not a single biological event. Instead, it is a variety of different failures in the immune-tumor interaction.

Where the Edges Are

Despite the impressive performance, the authors define the boundaries of their work. First, the study lacks an independent external validation cohort. While the internal "optimism-correction" is mathematically sound, it does not replace testing on a completely different group of patients.

Second, the model currently relies on several high-tech layers. These include multiplex immunofluorescence and whole-exome sequencing. Such technologies are not yet available in every hospital. While the inclusion of routine blood tests offers a path toward clinical utility, the full signature remains difficult to implement broadly today. Finally, the study does not compare the model against a non-immunotherapy control arm. Therefore, it is difficult to determine if these markers are purely predicting drug resistance or simply predicting disease aggressiveness.

Figures from the paper