Spatially aware deep learning reveals tumor heterogeneity patterns that encode distinct kidney cancer states

Cell Rep Med 2023 6 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Page [1, 2]
Why Tumor Heterogeneity in Kidney Cancer Matters

Clear cell renal cell carcinoma (ccRCC) is the most common type of kidney cancer. Even among tumors that look similar under a microscope, patients have very different outcomes - some do well for many years while others see their cancer spread quickly. Part of the explanation lies in 'tumor heterogeneity': within a single tumor, different regions can have distinct biological properties.

Standard pathology assigns a single tumor grade (a measure of how abnormal cancer cells look), but this single number may miss important variation within the tumor. Deep learning models applied to whole-slide pathology images can potentially capture this spatial variation - mapping which regions of a tumor are more or less aggressive - in ways that a single grade cannot.

TL;DR: This study used spatially aware deep learning to map tumor variation within kidney cancer slides, uncovering patterns linked to patient outcomes and treatment response.
Pages 4-4
Deep Learning Approach and Patient Cohorts

The study analyzed whole-slide pathology images from 1,102 ccRCC patients across three independent cohorts: DFCI-PROFILE (from Dana-Farber Cancer Institute), TCGA-KIRC (The Cancer Genome Atlas kidney data), and CM-025 (a clinical trial cohort). This large, multi-cohort design is important because findings that replicate across different patient groups are more likely to reflect real biology rather than data collection artifacts.

A key innovation was the use of 'spatially aware' deep learning. The model did not just classify the whole tumor - it analyzed individual small regions (tiles) of the slide and then built a spatial map of which regions had high-grade versus low-grade features. Region adjacency graphs (RAGs) were used to represent the spatial relationships between neighboring tissue patches, capturing how different tumor areas are arranged relative to each other.

TL;DR: Deep learning analyzed 1,102 kidney cancer patients across three cohorts, mapping spatial grade patterns across tumor regions using region adjacency graphs.
Pages 7-7
Grade Prediction Performance

The deep learning grade classifier performed well: AUROC of 0.88 on the TCGA cohort and 0.944 on the CM-025 trial cohort. This means the model correctly identified higher versus lower grade tumors about 88 to 94% of the time, matching or approaching the performance of expert pathologists on the same task.

More interestingly, the model could detect fine-grained spatial differences within tumors that a single assigned grade score cannot capture. By mapping grade patterns across the whole slide, the model created a spatial 'heterogeneity landscape' for each tumor - a picture of how mixed or uniform the tumor is at a cellular level.

TL;DR: The grade classifier achieved AUROC of 0.88 to 0.944 across cohorts and created spatial maps of tumor heterogeneity that go beyond what a single pathology grade can convey.
Pages 10-10
Microheterogeneity: A New Tumor Concept

The study identified a phenomenon called 'microheterogeneity' - tumors containing a mixture of low-grade and high-grade regions at small spatial scales. Microheterogeneity was found in 40.6% of TCGA-KIRC tumors and 34.7% of CM-025 tumors. This means that in over a third of patients, the tumor is not uniformly one grade but is a mosaic of different behaviors.

Microheterogeneity was strongly associated with loss of function of the PBRM1 gene - a tumor suppressor gene frequently mutated in ccRCC. PBRM1 loss destabilizes the way cells regulate their growth, potentially causing different tumor regions to evolve along different paths. This genetic connection gives biological grounding to the imaging observation and suggests microheterogeneity is not a random artifact but a real biological property.

TL;DR: Microheterogeneity - tumors mixing high and low-grade regions - was found in 35-41% of patients and strongly linked to PBRM1 gene loss, providing a biological explanation for imaging variation.
Pages 13-13
Predicting Response to Immunotherapy

One of the most clinically important findings was that microheterogeneity predicted response to immune checkpoint inhibitor (ICI) therapy - a class of drugs that harness the immune system to fight cancer. Patients whose tumors were both highly infiltrated by immune cells AND microheterogeneous had the best outcomes when treated with ICIs (p=0.0220 for overall survival). This combination identified a specific subgroup of patients most likely to benefit.

This is particularly valuable because ICI therapy is expensive and can have serious side effects. If doctors can identify who will benefit before starting treatment - using pathology images rather than additional expensive molecular tests - they can target this therapy to those most likely to respond, avoiding unnecessary treatment for patients unlikely to benefit.

TL;DR: Tumors that are both immune-infiltrated and microheterogeneous respond best to immunotherapy - a finding that could guide which kidney cancer patients receive this expensive treatment.
Pages 16-18
Clinical Promise and Limitations

This study demonstrates that spatial deep learning analysis of pathology slides can reveal meaningful tumor biology in kidney cancer that standard pathology grading misses. The discovery of microheterogeneity as a distinct, biologically grounded tumor state - linked to PBRM1 mutations and ICI response - represents a genuinely new way to think about and classify kidney cancer.

However, the study is retrospective, meaning it looked back at existing data rather than testing the model in prospective clinical practice. Validation in prospective trials where the model's output actually influences treatment decisions is needed before this approach changes clinical care. Additionally, the models require specialized computational infrastructure and pathology slide digitization that is not yet universal in all hospitals.

TL;DR: Spatial deep learning reveals tumor microheterogeneity linked to PBRM1 loss and immunotherapy response in kidney cancer, but prospective trials are needed before this changes clinical practice.
Citation: Open Access, . Available at: .