Abstract 1103: Multi-modal machine learning identifies predictive biomarkers of response to neoadjuvant axitinib in renal cell carcinoma with venous tumor thrombus
- Abstract:
- Venous tumor thrombus (VTT) is present in 10-15% patients with renal cell carcinoma (RCC), where the primary tumor invades the renal vein and inferior vena cava, increasing the morbidity and mortality of curative surgery. Neoadjuvant anti-angiogenic therapies have shown potential to reduce VTT length and improve nephrectomy success rates. The NAXIVA trial, a phase II study of neoadjuvant axitinib in 20 RCC patients with VTT, demonstrated that 35% of patients experienced significant downstaging of VTT. However, the biological mechanisms driving response or resistance in the remaining 65% are unclear. Machine learning has the potential to uncover mechanistic relationships between features in these patients and guide biomarker discovery for early response.In this study, we present a predictive machine learning framework based on multi-modal integration of digital pathology, flow cytometry and plasma cytokine profiling, using tissue and blood samples from the NAXIVA trial. A machine learning model incorporating recursive feature elimination and logistic regression with stochastic gradient descent was developed to predict response. To avoid overfitting, a leave-one-out nested cross validation approach was used for 20 iterations, leaving one patient out at a time and training the model on the remaining 19 patients. Two separate model ensembles were produced, using baseline features (N=62) and features from both baseline and two weeks post-treatment initiation (N=69) respectively. We used a statistical consensus approach to identify key predictive factors.The baseline model achieved an AUC of 0.868, identifying tissue microvessel density, plasma IL-12p70 and plasma CCL17 as key predictive factors. Responding patients had lower plasma IL-12p70, lower CCL17 and higher microvessel density than non-responding patients. Bulk RNA-seq data from pre-treatment tumor biopsies revealed elevated IL-12R expression in non-responders. Incorporating features post-treatment initiation improved model performance (AUC = 0.945), with fold changes in PlGF and sTie-2, alongside baseline IL-12p70 and CCL17, emerging as key predictors. Responders showed greater PlGF and sTie-2 induction following treatment onset.In conclusion, our machine learning framework accurately predicts response to neoadjuvant axitinib in RCC patients with VTT. The high microvessel density and growth factor induction in responders suggests they exhibit an “angiogenic” phenotype (clusters 1/2 from IMmotion151), while higher cytokine levels in non-responders suggest an “immunogenic” phenotype (cluster 4). Early changes in PlGF and sTie-2 may serve as actionable biomarkers for patient stratification. Further validation in independent datasets is essential to facilitate clinical integration and improve personalized treatment strategies for RCC patients with VTT. Citation Format: Rebecca Wray, Hania Paverd, Ines Machado, Johanna Barbieri, Farhana Easita, Abigail R. Edwards, Ferdia A. Gallagher, Iosif A. Mendichovszky, Thomas J. Mitchell, Maike De La Roche, Jacqueline D. Shields, Stephan Ursprung, Lauren Wallis, Anne Y. Warren, Sarah J. Welsh, Mireia Crispin-Ortuzar, Grant D. Stewart, James O. Jones. Multi-modal machine learning identifies predictive biomarkers of response to neoadjuvant axitinib in renal cell carcinoma with venous tumor thrombus [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 1103.
- Authors:
- R Wray, H Paverd, I Machado, J Barbieri, F Easita, AR Edwards, FA Gallagher, IA Mendichovszky, TJ Mitchell, M De La Roche, JD Shields, S Ursprung, L Wallis, AY Warren, SJ Welsh, M Crispin-Ortuzar, GD Stewart, JO Jones
- Journal:
- Cancer Research
- Citation info:
- 85(8_Supplement_1):1103-1103
- Publication date:
- 21st Apr 2025
- Full text
- DOI