(UroToday.com) The 2023 American Society of Clinical Oncology Genitourinary (ASCO GU) cancers symposium held in San Francisco, CA between February 16th and 18th was host to a Renal Cell Cancer; Adrenal, Penile, Urethral, and Testicular Cancers poster session. Sandra Alonso Paz presented the results of her group’s study evaluating deep learning techniques in RNA sequencing data for the prediction of response to immune checkpoint inhibitors in patients with metastatic RCC.
Immune checkpoint inhibitors have demonstrated PFS and OS benefits compared to sunitinib in patients with metastatic clear cell RCC.1,2 Current research lines are using genomic and transcriptomic data to identify predictive biomarkers for response to treatment. Since healthcare domain datasets are heterogenous in nature with diverse patient populations, deep learning approaches are needed to help better select patients that would benefit from personalized approaches.
In this study, the investigators retrospectively collected clinical data and RNA sequencing data (43,893 transcripts) from 181 patients treated with nivolumab (anti-PD-1) in Checkmate 009, 010, and 025. The investigators subsequently created 8 datasets using combinations of clinical and RNA sequencing data. Patients were classified as having clinical benefit versus no clinical benefit using a PFS cut-off of 3 months.
The primary objective was to assess the performance of different deep learning models (Deep Autoencoder [DA] and Convolutional Neural Network [CNN]) for predicting PFS in this cohort. Within that scope, they followed several research lines including the creation of combined datasets with clinical and RNA sequencing data, and compared the results of the deep learning models to those of traditional machine learning (ML) models. They then came up with an interpretability analysis of those black-box models using LIME and SHAP values. 
Clinical and transcriptomic data were available for all 181 nivolumab-treated patients. Outcomes achieved confirmed that the response to Nivolumab could be modeled using RNA sequencing data.
However, deep learning models were not significantly better than traditional machine learning methods for predicting response (p= 0.068). Deep autoencoder provided 68.9% accuracy, whereas the most accurate model was logistic regression classifier which achieved 86.4% accuracy.
Interpretability results revealed that the most relevant genes for decision making were related to the immune response and the regulation of kinases. Although the best results were achieved by integrating transcriptomic and clinical data, interpretability results revealed that clinical features seemed to have more relevance when making the decision.
The authors concluded that the integration of clinical and molecular data could lead to more accurate predictions of outcome than any dataset by its own. However, further research is intended in the field of the deep learning analysis, as data codification and data structure could bias the results. The ongoing study ART (Artificial Intelligence in Renal Tumors) will address this issue prospectively.
Presented by: Sandra Alonso Paz, Master in Computational Biology, Universidad Complutense de Madrid, Madrid, Spain
Written by: Rashid Sayyid, MD, MSc – Society of Urologic Oncology (SUO) Clinical Fellow at The University of Toronto, @rksayyid on Twitter during the 2023 Genitourinary (GU) American Society of Clinical Oncology (ASCO) Annual Meeting, San Francisco, Thurs, Feb 16 – Sat, Feb 18, 2023.
References:
- Motzer RJ, et al. Nivolumab plus Ipilimumab versus Sunitinib in Advanced Renal-Cell Carcinoma. N Engl J Med 2018;378(14):1277-1290.
- Rini BI, et al. Pembrolizumab plus Axitinib versus Sunitinib for Advanced Renal-Cell Carcinoma 2019;380(12):1116-1127.