Clear cell renal cell carcinoma (ccRCC) is the most common type of kidney cancer, and when it spreads to other organs (metastatic ccRCC), it becomes very difficult to treat. Two main types of drugs are used: tyrosine kinase inhibitors (TKIs), which block blood vessel growth that tumors need, and immune checkpoint inhibitors (ICIs), which help the immune system attack cancer cells.
While combinations of TKIs and ICIs have improved outcomes, overall response rates are only 55-70%, and most patients eventually see their disease progress. A major challenge is that doctors currently lack reliable tools to predict which patients will respond to which treatment - meaning many patients receive therapy that may not work for them while experiencing significant side effects.
Past attempts to use single biomarkers - such as PD-L1 expression, tumor mutation burden, or CD8+ T cell counts - to predict treatment response have been inconsistent across different patient groups. This highlights the urgent need for a more comprehensive, multi-factor approach to treatment selection.
This study addressed that need by building the largest known harmonized database of ccRCC gene expression data, combining information from 14 research cohorts totaling 3,621 patients, to develop an AI-powered system for predicting which therapy each patient is most likely to respond to.
The research team assembled 3,621 ccRCC tumor samples from 14 different research databases and clinical trials, making it the largest harmonized collection of its kind. Data from all cancer stages, both primary tumors and metastatic sites, and multiple measurement platforms were included to ensure broad coverage.
To make data from different labs and technologies comparable, a normalization process was applied using a technique called single-sample gene set enrichment analysis. This corrected for technical differences between batches, allowing meaningful comparisons across all samples.
For each sample, researchers calculated 33 functional gene expression signatures (FGES) - standardized scores representing different biological features of the tumor and its surrounding environment, such as immune activity, blood vessel formation, and metabolic patterns.
A separate validation cohort of 193 patients from Washington University (WU-RCC) was assembled independently to test whether models built on the larger dataset would hold up in a completely new group of patients - a critical step for confirming real-world applicability.
Using an advanced clustering algorithm on the gene expression data, researchers identified five distinct tumor microenvironment (TME) subtypes, called HiTME (Harmonized immune TME) subtypes: IE (immune-enriched), IE/M (immune-enriched with suppressive myeloid cells), F (fibrotic), V (highly vascularized), and D (immune desert).
The IE subtype showed strong immune cell infiltration with active anti-cancer immune activity. The IE/M subtype also had many immune cells but was undermined by immunosuppressive myeloid cells and scar-forming (fibrotic) features. The F subtype was dominated by fibrosis, suppressive macrophages, and cancer-promoting cytokines.
The V subtype was rich in blood vessels and expressed high levels of angiogenic factors - the targets of TKI drugs. The D subtype had very few immune cells but high metabolic activity. Patients with fibrotic subtypes (IE/M and F) had the worst overall survival, while IE subtype patients had more favorable outcomes.
These HiTME subtypes strongly overlapped with previously published classification systems, validating their biological relevance. Importantly, the subtypes were significantly associated with treatment response - IE and IE/M patients responded better to immunotherapy, while IE and V patients responded better to TKI therapy.
To physically confirm that the gene-expression-based subtypes reflected real differences in tumor tissue, the team used multiplex immunofluorescence (MxIF) - a technology that simultaneously images dozens of different proteins in tumor tissue slices using fluorescent tags.
A custom panel of 19 protein markers was applied to tumor samples from 34 patients. This allowed precise mapping of different cell types - including cancer cells, immune cells, blood vessel cells, and support cells - in their exact physical locations within the tumor.
The visual patterns seen under the microscope closely matched the gene-expression profiles. For example, IE/M tumors showed dense immune cell infiltration at the tumor border, while V subtype tumors showed elaborate branching blood vessel networks. The correlation between imaging and gene expression data was strong, particularly for CD8+ T cells (r=0.71) and endothelial cells (r=0.84).
This spatial analysis also revealed that tertiary lymphoid structures - organized immune cell clusters associated with better immunotherapy response - were frequently found in IE/M tumors, providing additional evidence that this subtype may benefit from immunotherapy despite its immunosuppressive features.
Separate AI prediction models were built for immunotherapy (ICI) and targeted therapy (TKI) response. The ICI response model was trained on 214 patients and validated on 430 additional patients from independent clinical trials. It generated a score from 0 to 1 reflecting the probability of responding to immunotherapy.
The ICI model achieved strong predictive accuracy with a ROC-AUC of 0.77-0.78 - meaning it correctly distinguished responders from non-responders about 77-78% of the time. Patients with high ICI responder scores had significantly longer progression-free survival in validation datasets. Importantly, this model outperformed all previously published single-marker approaches.
The TKI response model was trained on 240 patients and validated on 822 patients. It focused on features including angiogenic activity, macrophage presence, and tumor proliferation rate - biological characteristics that reflect whether a tumor depends on blood vessels for growth. This model also demonstrated strong performance with a ROC-AUC of 0.74.
Crucially, the ICI and TKI scores were independent of each other - a high score on one did not predict response on the other. This independence is essential for using them together as complementary tools for treatment decision-making.
Combining the ICI and TKI responder scores, researchers built an integrated decision-tree model to guide treatment selection. When applied to two large clinical trial datasets (IMmotion151 and JAVELIN), the model classified 56% of patients as likely to benefit from immunotherapy-based combinations and 41% as better suited for TKI-alone therapy.
Patients classified as ICI-preferred showed significantly longer progression-free survival when treated with immunotherapy combinations (atezolizumab+bevacizumab or avelumab+axitinib) compared to TKI alone. In contrast, patients classified as TKI-preferred had better outcomes with single-agent TKI therapy such as sunitinib.
The model also identified a small group of patients (about 2-3%) with low scores on both models - meaning they were unlikely to respond well to either standard treatment. These patients had the worst survival outcomes and showed high levels of immunosuppressive and pro-tumor features, suggesting they may need entirely different treatment approaches.
The decision-tree model also reflected known genetic associations - for example, patients with MTOR gene mutations were more likely to have low ICI scores, while BAP1 mutations correlated with low TKI scores - providing a biological rationale for the AI predictions.
The research team analyzed patterns of cytokines - signaling proteins that regulate immune activity - across all tumor subtypes. Three distinct groups of cytokines were identified, each associated with a different type of immune environment and treatment response profile.
Group 1 cytokines (including CXCL13, IFNG, CCL4, and CCL5) were linked to active anti-cancer immune responses, produced mainly by T cells and macrophages. These cytokines strongly correlated with high ICI responder scores and were concentrated in immune-enriched tumor subtypes - suggesting they mark tumors primed to respond to immunotherapy.
Group 2 cytokines (including CXCL8 and CCL18) were produced by immunosuppressive macrophages and were associated with the fibrotic tumor subtypes. These signals promote chronic inflammation that suppresses immune killing and encourages cancer spread - potentially explaining why fibrotic subtypes respond poorly to immunotherapy.
Group 3 cytokines (including VEGFA, HIF1A, and TGFB) were angiogenesis-related, linked to hypoxic and poorly immune-infiltrated tumors. These strongly correlated with high TKI responder scores, suggesting that tumors driven by blood vessel growth signals are the best candidates for anti-angiogenic TKI treatment.
A key insight of this study is that no single biomarker - not PD-L1, not TMB, not angiogenesis score alone - consistently predicts treatment response across different patient groups. The complexity of kidney cancer biology requires combining multiple layers of information to make reliable predictions.
The HiTME subtype framework accounts for multiple biological features simultaneously, including immune cell types, fibrotic activity, vascular patterns, and metabolic state. This systems-level view of tumor biology captures interactions that single-marker approaches miss.
The AI approach is described as a white-box model - meaning the features it uses are biologically interpretable, not hidden. Each component of the responder score reflects a known aspect of tumor biology, making it easier for clinicians to understand and trust the predictions.
The study also highlights a previously unrecognized patient subgroup - those unlikely to respond to either immunotherapy or targeted therapy - who may benefit from novel approaches such as anti-fibrotic agents or anti-proliferative drugs. Identifying these patients early could save them from ineffective treatments and direct them to clinical trials exploring new options.
This study represents a major step toward personalized medicine for kidney cancer. By integrating genomic, transcriptomic, and tumor microenvironment data from thousands of patients, researchers have developed validated tools that can predict which treatment - immunotherapy or targeted therapy - is most likely to benefit each individual patient.
The ICI and TKI responder scores were validated across multiple independent clinical trial datasets, demonstrating that they generalize beyond the training data. This is an important indicator of real-world usefulness - a model that only works on the data it was trained on has limited clinical value.
The researchers envision the decision-tree model becoming a practical clinical tool, helping oncologists select first-line therapy for patients with metastatic ccRCC. As new drugs continue to enter clinical practice - such as newer TKI+ICI combinations - the framework can potentially be updated to accommodate them.
Limitations include the reliance on retrospective clinical data, the lack of data for some newer treatment combinations (such as lenvatinib+pembrolizumab), and relatively small sample sizes for some subgroups. Prospective clinical validation in real-world trials will be an important next step before routine clinical adoption.