AI Predicts Kidney Cancer Drug Response Using Routine Pathology Slides

An innovative AI model provides a powerful and cost-effective alternative to genetic testing by analyzing routine pathology slides to predict a patient's response to kidney cancer therapy, thereby bringing personalized treatment within reach.

A Schematic outlining the H&E DL Model development, validation and practical application of the model. The model predicts Angioscores directly from an H&E-stained slide and is validated against the RNA-based Angioscore using multiple independent datasets. The validated model is applied to independent, previously unseen, clinical datasets where its predicted Angioscore is correlated with response to antiangiogenic (AA) therapy. B H&E DL Model is an interpretable machine learning model to predict Angioscore from H&E images. Given an input H&E image, the model predicts a vascular mask (green), and the proportion of positive pixels is the output H&E-based Angioscore. Training data consists of two datasets with H&E images matched with RNA-based Angioscores and CD31 IHC (basis of the vascular mask), providing the target ground truth. The model is trained to predict the vascular mask matching the CD31 and the RNA-based Angioscore (see Supplementary Fig 1 and methods for details).  

An artificial intelligence (AI)-based model developed by researchers at UT Southwestern Medical Center can accurately predict which kidney cancer patients will benefit from anti-angiogenic therapy, a class of treatments that's only effective in some cases. Their findings, published in the journal Nature Communications, could lead to viable methods for using AI to inform treatment decisions for this and other types of cancer.

"There's a real unmet need in the clinic to predict who will respond to certain therapies. Our work demonstrates that histopathological slides, a readily available resource, can be mined to produce state-of-the-art biomarkers that provide insight on which treatments might benefit which patients," said Satwik Rajaram, Ph.D., Assistant Professor in the Lyda Hill Department of Bioinformatics and member of the Harold C. Simmons Comprehensive Cancer Center at UT Southwestern. Dr. Rajaram co-led the study with Payal Kapur, M.D., Professor of Pathology and Urology and a co-leader of the Kidney Cancer Program (KCP) at the Simmons Cancer Center.

Every year, nearly 435,000 people are diagnosed with clear cell renal cell carcinoma (ccRCC), making it the most common subtype of kidney cancer. When the disease metastasizes, anti-angiogenic therapies are often used for treatment. These drugs inhibit the formation of new blood vessels in tumors, thereby limiting access to molecules that fuel tumor growth. Although anti-angiogenic medications are widely prescribed, fewer than 50% of patients benefit from them, Dr. Kapur explained, exposing many to unnecessary toxicity and financial burden.

No biomarkers are clinically available to accurately assess which patients are most likely to respond to anti-angiogenic drugs, she added. However, a clinical trial conducted by Genentech suggested that the Angioscore (a test that assesses the expression of six blood vessel-associated genes) may have promise. However, this genetic test is expensive, hard to standardize among clinics, and introduces delays in treatment. It also tests a limited part of the tumor, and ccRCC is quite heterogeneous, with variable gene expression in different regions of the cancer.

 Illustration of the model output on a ccRCC case with intra-slide heterogeneity and available multiregional RNA sequencing data. The central plot shows the model H&E Angioscore (blue-white-red colormap) applied to all tumor areas on the slide. To contrast against the output from patch-level models (e.g., based on MIL), we calculated a local average of the percentage positive vascular mask prediction within a 416 x 416px grid, which shows a far lower level of explainability than the pixel-level vascular masks shown alongside. Three areas (circles) are marked where we obtained the ground truth RNA-based Angioscores (circle colors in yellow, using a red colormap). There is a broad qualitative agreement in the amount of vasculature and the H&E-based and RNA-based Angioscores, and both capture ITH, which would have been missed by the standard slide-level bulk sequencing approaches.

To overcome these challenges, Drs. Kapur, Rajaram, and their colleagues at the KCP developed a predictive method using AI to assess histopathological slides – thinly cut tumor tissue sections stained to highlight cellular features. These slides are nearly always part of a patient's standard workup at diagnosis, and their images are increasingly available in electronic health records, said Dr. Rajaram, also Assistant Professor in the Center for Alzheimer's and Neurodegenerative Diseases and the Department of Pathology.

Using a type of AI based on deep learning, the researchers trained an algorithm using two sets of data: one that matched ccRCC histopathological slides with their corresponding Angioscore, and another that matched slides with a test they developed to assess blood vessels in the tumor sections.

Importantly, unlike many deep learning algorithms that don't offer insight into their results, this approach is designed to be visually interpretable. Rather than producing a single number and directly predicting response, it generates a visualization of the predicted blood vessels that correlates tightly with the RNA-based Angioscore. Patients with a greater number of blood vessels are more likely to respond to therapy; this approach enables users to understand how the model arrived at its conclusions.

When the researchers evaluated this approach using slides from more than 200 patients who weren't part of the training data, including those collected during the clinical trial that showed the potential value of Angioscore, it predicted which patients were most likely to respond to anti-angiogenic therapies nearly as well as Angioscore. The algorithm showed a responder will have a higher score than a non-responder 73% of the time, compared to 75% with Angioscore. 

The study authors suggest that AI analysis of histopathological slides could eventually be used to help guide diagnostic, prognostic, and therapeutic decisions for various conditions. They plan to develop a similar algorithm to predict which patients with ccRCC will respond to immunotherapy, another class of treatments that only some patients respond to.

Other UTSW researchers who contributed to this study include first author Jay Jasti, Ph.D., former Data Scientist in the Rajaram Lab; James Brugarolas, M.D., Ph.D., Professor of Internal Medicine, Director of the Kidney Cancer Program, and a member of the Simmons Cancer Center; Dinesh Rakheja, M.D., Professor of Pathology and Pediatrics; Hua Zhong, Ph.D., Computational Biologist; Vandana Panwar, M.D., Medical Resident; Vipul Jarmale, M.S., Data Scientist; Jeffrey Miyata, B.S., Histology Technician; and Alana Christie, M.S., Biostatistical Consultant.

Dr. Kapur holds the Jan and Bob Pickens Distinguished Professorship in Medical Science, in Memory of Jerry Knight Rymer and Annette Brannon Rymer and Mr. and Mrs. W.L. Pickens.

The study was funded by the Department of Defense (KC200285), the Cancer Prevention and Research Institute of Texas (RP220294), the Lyda Hill Department of Bioinformatics, a National Institutes of Health-sponsored Kidney Cancer SPORE grant (P50CA196516), and a National Cancer Institute (NCI) Cancer Center Support Grant (P30CA142543).