Clear cell renal cell carcinoma (ccRCC) is the most common subtype of kidney cancer, accounting for 75% to 80% of metastatic cases. It is characterized by extensive genomic intratumoral heterogeneity (ITH), meaning that different regions of the same tumor harbor distinct genetic mutations and cellular phenotypes. Prior work in the TRACERx and TCGA-KIRC cohorts showed that this molecular heterogeneity correlates with worse progression-free survival. Nuclear grade, the standard histopathologic measure of how dedifferentiated tumor nuclei appear, is a primary prognostic feature and can provide a histologic window into ITH. However, measuring ITH from histology slides at scale has been impractical because pathologists cannot manually quantify the spatial distribution of grade phenotypes across thousands of whole-slide images (WSIs).
The immunotherapy puzzle: Immune checkpoint inhibitors (ICIs) are a standard therapy in ccRCC, yet this cancer defies many of the molecular conventions that predict ICI response in other solid tumors. Tumor mutational burden, microsatellite instability, and PD-L1 expression, which are reliable biomarkers in cancers like melanoma or lung cancer, have limited predictive value in ccRCC. Identifying which patients will benefit from ICI remains a major clinical challenge, and current molecular sequencing approaches lack the spatial resolution needed to simultaneously capture tumor-intrinsic heterogeneity and its relationship to the immune microenvironment.
The study's hypothesis: Researchers from Dana-Farber Cancer Institute, Harvard Medical School, the Broad Institute, and Yale Cancer Center hypothesized that spatially aware deep-learning models applied to standard H&E-stained diagnostic slides could provide a unified understanding of tissue structures that dictate biological and clinical states in ccRCC. Rather than relying on expensive multi-region molecular sequencing, the team aimed to extract spatial heterogeneity patterns from routine pathology images that are already collected for every patient. The study encompassed WSIs from 1,102 patients across five independent cohorts.
The authors built a three-layer computational framework to extract biologically meaningful features from H&E whole-slide images. The first layer used a ResNet-50 convolutional neural network (CNN), fine-tuned with pixel-level expert pathologist annotations from a retrospective Dana-Farber (DFCI-PROFILE) cohort of 208 patients, to distinguish tumor tissue from adjacent non-tumor tissue. The second layer used a separate ResNet-50 CNN, also trained on the DFCI-PROFILE cohort, to classify putative tumor regions as low-grade (G2) or high-grade (G4). The third layer deployed a fine-tuned HoVerNet CNN to identify tumor-infiltrating lymphocytes (TILs) within the tissue.
Validation performance: The grade classification model was validated on two completely independent, unseen test cohorts: TCGA-KIRC (n = 421 patients) and CM-025 (n = 439 patients). The model achieved an area under the receiver operating characteristic curve (AUROC) of 0.88 in TCGA-KIRC and 0.944 in CM-025, demonstrating strong generalizability. Continuous grade scores were derived by averaging predictions across entire slide images and then discretized into tercile bins mirroring G2/G3/G4 categories. These computationally inferred grade categories produced significant patient stratifications for both progression-free interval (PFI) and overall survival (OS), with multivariate log-rank test p-values below 1e-5 for both endpoints.
Region adjacency graphs (RAGs): To capture spatial relationships, the authors formed region adjacency graphs that describe where distinct tumor and grade phenotypes occur in a slide and whether these regions directly or indirectly contact one another. Local predictions from each CNN were grouped via watershed segmentation and assembled into graph representations. This approach condenses the tens of thousands of tile-level predictions per WSI into a compact, information-rich latent representation that encodes both the composition and the spatial arrangement of tumor phenotypes.
Cohort overview: In total, the study included 208 DFCI-PROFILE patients (training), 421 TCGA-KIRC patients (validation), 439 CM-025 clinical trial patients (validation and therapeutic analysis), 21 multi-block nephrectomy cases (micro-to-macro heterogeneity validation), and 13 paired multiplex immunofluorescence (mIF) cases (tumor-immune interaction analysis). All images underwent quality control using HistoQC. The deep-learning pipeline was implemented in PyTorch.
When the authors inspected the model's graph representations, they observed a striking pattern: some WSIs showed co-occurrence of different grade phenotypes within the same slide, while others were markedly homogeneous. They termed this co-occurrence "microheterogeneity" and identified two primary forms. "Proximal" microheterogeneity occurs when regions of different grades directly contact one another in the tissue. "Distal" microheterogeneity occurs when differing grade regions are separated by stromal barriers or physical gaps in the slide. These two forms are not mutually exclusive and can co-exist within a single WSI.
Prevalence across cohorts: Microheterogeneity was identified in 40.6% of TCGA-KIRC cases and 34.7% of CM-025 cases. Importantly, this feature did not simply track with pathologist-assigned grade labels. In TCGA-KIRC, the frequency of microheterogeneity was 0.36 in G2 tumors, 0.494 in G3 tumors, and 0.317 in G4 tumors. In CM-025, it was 0.524 in G2, 0.333 in G3, and 0.230 in G4. This inconsistent pattern across grade categories demonstrates that microheterogeneity captures something fundamentally different from overall grade assignment.
Quantifying the degree of heterogeneity: To move beyond a binary present/absent classification, the authors calculated a continuous "f-weighted" heterogeneity score for each WSI. This score represents the weighted sum of heterogeneous contacts (edges in the RAG), with larger weights given to contacts between tumor regions of similar area. Among slides that exhibited microheterogeneity, both cohorts showed a wide distribution of heterogeneity abundance, meaning that the degree of spatial grade mixing varied substantially from patient to patient.
Why pathologists cannot detect this: Standard pathology practice assigns a single categorical nuclear grade to a tumor based on the highest-grade component observed. This approach collapses all spatial information into a single label and cannot capture the pattern and extent of grade mixing across a slide. The deep-learning framework, by making predictions at the tile level and then assembling them into spatial graphs, reveals structure that is fundamentally beyond the resolution of manual review.
A natural concern about any single-slide analysis is whether it reflects the entire tumor or just an isolated region. To address this, the authors analyzed a cohort of 21 multi-block nephrectomy cases, where multiple spatially separated tissue blocks from the same tumor were scanned and evaluated. For each patient, the maximum microheterogeneity score observed in any single WSI correlated with the presence of microheterogeneity across all WSIs from that tumor. This correlation was not driven by the number of blocks sampled per patient.
Predictive modeling: The team further demonstrated through Bayesian predictive modeling that a single WSI could predict microheterogeneity status for the remaining WSIs from the same patient, with a minimum log10(Bayes factor) of 3.04, which represents very strong evidence. This finding is clinically significant because it means that routine diagnostic slides, which are already collected for every kidney cancer patient, contain latent information about spatial structures present throughout the entire tumor.
Molecular correlates: The authors also examined whether computationally derived microheterogeneity correlated with known somatic driver mutations in ccRCC. WSIs from tumors with PBRM1 loss of function (LOF), a chromatin regulator gene previously associated with molecular ITH, showed higher frequency of microheterogeneity compared with non-LOF tumors. A similar trend was observed for SETD2 LOF mutations. In contrast, 9p21.3 chromosomal deletions, previously implicated in ccRCC oncogenesis, were enriched for microhomogeneity patterns. Notably, global molecular ITH itself (as measured by sequencing) was decoupled from histologic microheterogeneity, indicating these are related but distinct phenomena.
Prognostic value: In the TCGA-KIRC cohort (untreated primary tumors), the presence of microheterogeneity was negatively correlated with survival, with hazard ratios above 1 in both univariate and bivariate Cox proportional hazards models. In bivariate models that included either pathologist-assigned grade or computationally inferred continuous grade, microheterogeneity added independent prognostic information. The continuous grade score had a stronger concordance index (C-Index) for progression-free interval but not for overall survival. This means microheterogeneous tumors, regardless of their overall grade, carried greater metastatic potential, consistent with findings from multi-region molecular profiling studies like TRACERx.
Predictive value for immunotherapy: The CM-025 cohort provided a unique opportunity to test whether microheterogeneity predicted response to specific therapies. CM-025 was a phase 3 randomized clinical trial comparing anti-PD1 blockade (nivolumab) with mTOR inhibition (everolimus) in patients with anti-angiogenic refractory metastatic ccRCC. The presence of microheterogeneity was associated with improved overall survival (OS) and progression-free survival (PFS) in the nivolumab (ICI) arm, but not in the everolimus arm. This treatment-specific association is striking because microheterogeneity was a poor prognostic marker in the untreated setting, yet it predicted better outcomes specifically under immunotherapy.
Independence from overall grade: Within the microheterogeneous cases in the ICI arm, overall grade score did not contribute a statistically significant predictive signal for PFS or OS (though it trended toward significance for OS). This suggests that the spatial pattern of grade mixing, rather than the overall level of dedifferentiation, is what drives the association with ICI response. The finding implies that microheterogeneity encodes something about the tumor-immune interface that is distinct from bulk tumor aggressiveness.
Immune infiltration measured by CD8 immunofluorescence alone was not associated with ICI response in ccRCC in prior analyses of CM-025, despite being predictive in other immune-responsive cancers. The authors hypothesized that TIL patterns might still be relevant but that joint inference with tumor spatial heterogeneity was required. They used the HoVerNet-based TIL detection model to quantify immune infiltration across the CM-025 WSIs and combined this with microheterogeneity status, creating four patient subgroups: microheterogeneous with high TIL density, microheterogeneous with low TIL density, homogeneous with high TIL density, and homogeneous with low TIL density.
The key subgroup: Among ICI-treated patients, the microheterogeneous and highly infiltrated subgroup demonstrated significantly improved OS compared with all remaining patients (p = 0.0220, log-rank test). This subgroup also showed a consistent trend toward improved PFS, though it did not reach statistical significance (p = 0.0662). Neither immune infiltration alone nor microheterogeneity alone captured this subpopulation as effectively as the combined feature. Critically, this association was absent in the everolimus arm, confirming its specificity to ICI response.
Comparison with molecular features: For OS in the ICI arm of CM-025, predictive models using computer vision features (microheterogeneity + TIL density) performed similarly to models using previously nominated genomic features (PBRM1 LOF, 9p21.3 deletion). Combining both feature sets produced net improvements in concordance index while retaining consistent parameter associations, meaning PBRM1 LOF and microheterogeneity each maintained positive coefficient weights. Adding clinical risk covariates produced further improvements to C-Index metrics. This demonstrates that histology-derived spatial features and genomic features capture complementary, non-redundant information about ICI response.
To understand the biological mechanism linking microheterogeneity to ICI response, the authors analyzed 13 advanced ccRCC cases with paired H&E and multiplex immunofluorescence (mIF) images from the same tissue. The mIF panel included markers for tumor cells (PAX8), CD8+ T cells, PD-1, PD-L1, FOXP3 (regulatory T cells), and DAPI (nuclei). The team built nearest-neighbor graphs of CD8+ and tumor cells and classified cells as "tumor-immune interacting" if they were adjacent to a different cell type in the graph.
Greater tumor-immune interaction: In regions of high tumor-immune interaction density, microheterogeneous tumors had higher CD8+ cell density while tumor cell density was similar between heterogeneous and homogeneous cases. The frequency of tumor cells adjacent to CD8+ cells was significantly higher in heterogeneous cases (p = 0.00215, Wilcoxon rank-sum test), indicating deeper immune infiltration into tumor-dense regions rather than a uniform increase across the microenvironment. Homogeneous tumors, by contrast, had more "desert-like" regions of non-infiltrated tumor tissue.
PD-1 activation: Among CD8+ cells engaged in direct tumor interaction, microheterogeneous cases had a significantly higher frequency of PD-1-high cells compared with homogeneous cases (p = 0.00908, Wilcoxon rank-sum test). PD-1-high CD8+ cells are generally considered "tumor-experienced," meaning they have been activated by encountering tumor antigens. This finding provides a biological rationale for why microheterogeneity predicts ICI response: these tumors foster an immune compartment that is both more abundant and more actively engaged with the tumor, making it more amenable to checkpoint blockade.
PDL1 context: In general, PD-1-high CD8+ cells were common across the cohort (median frequency 0.480), while PDL1-high tumor cells were relatively sparser (median frequency 0.150). This pattern aligns with broader observations in ccRCC that PD-L1 expression on tumor cells is not a reliable standalone biomarker, while the activation state of the infiltrating immune cells provides more actionable information about the tumor-immune dynamic.
Temporal disconnect: The histological data analyzed from CM-025 consisted of pretreatment primary tumor samples, which may differ from the tumor state at the time of trial enrollment due to ongoing tumor evolution. Since patients in CM-025 had already received and failed anti-angiogenic therapy, the specimens analyzed may be partially uncoupled from the eventual metastatic progression that determined trial outcomes. This is a common limitation in retrospective analysis of clinical trial tissues and underscores the need for prospective validation.
Sample size and generalizability: While the study included over 1,100 patients across five cohorts, the authors acknowledge that larger sample sizes in additional clinical cohorts are necessary to generalize findings about both the prognostic value of continuous grading and the predictive value of microheterogeneity for ICI response. The paired mIF cohort was particularly small (n = 13) and composed of varying biopsy sites, limiting the statistical power of the tumor-immune interaction analysis. Larger paired cohorts using emerging highly multiplex spatial imaging technologies will be needed to extend these findings.
Model scope: The current framework focuses on nuclear grade and TILs but does not incorporate other histological features that could further describe ccRCC biology, including necrosis, refined TIL subtypes (e.g., CD4+ vs. CD8+, B cells vs. T cells), proximal vs. distal microheterogeneity patterns as separate features, and stromal heterogeneity. The TIL detection model, based on morphology alone, cannot distinguish between CD8+ and CD4+ T cells or between B and T cells. Adding these features to the RAG framework could provide a more complete picture of the tumor microenvironment.
Annotation and validation gaps: The approach to validating spatial features is limited by a lack of exhaustive pixel-level annotation of all images. Nuclear grade itself is a composite of several imprecise tumor microenvironment phenotypes and does not have a tractable definition at the resolution required for consistent pathologist annotation. Future work could address this through larger-scale pathologist evaluation to reach consensus annotations. Additionally, the relationship between molecular features and microheterogeneity patterns needs further investigation using model systems and additional patient cohorts.
Broader potential: The authors note that this spatially aware approach could be extended to other cancers exhibiting phenotypic heterogeneity, such as Gleason grade patterns in prostate cancer. As multiple-instance learning and saliency mapping methods continue to evolve in digital pathology, integrating these techniques with the human-interpretable RAG approach could further enhance the ability to understand spatial structures that determine tumor-immune interactions across cancer types and therapeutic modalities.