Artificial intelligence (AI)–augmented analysis of quantitative circulating tumor DNA (ctDNA) burden and longitudinal kinetics to identify prognostic risk stratification in metastatic clear cell renal cell carcinoma (mccRCC).

Abstract

4522 Background: The relevance of ctDNA burden and longitudinal kinetics in RCC remains poorly defined. We applied AI–based analysis of quantitative ctDNA trajectories to identify prognostically informative patterns beyond baseline detection. Methods: We analyzed patients with mccRCC undergoing longitudinal tumor-informed ctDNA testing (Signatera) during first-line immune checkpoint inhibitor–based therapy. Quantitative ctDNA burden in mean tumor molecules per milliliter (MTM/mL) was measured serially and aligned to treatment initiation, with radiographic response classified as complete response (CR), partial response (PR), stable disease (SD), or progressive disease (PD). ctDNA trajectory similarity was independently assessed using dynamic time warping (DTW), trajectory-based K-means clustering, and a neural network sequence autoencoder (TensorFlow) with clustering to learn latent representations of ctDNA kinetics. Cross-method concordance was used to define robust ctDNA kinetic phenotypes. Results: Survival and response analyses included 93 patients with mccRCC with longitudinal ctDNA profiling. Baseline ctDNA detectability was not significantly associated with OS. Conversion from ctDNA-positive to ctDNA-negative status during therapy was associated with improved OS (p<0.05). Patients with radiographic disease control (CR/PR/SD) had significantly lower peak ctDNA levels than those with progressive disease (median 0.18 vs 8.86 MTM/mL; p=0.007). Increasing peak ctDNA burden was strongly associated with inferior OS (HR 2.21, 95% CI 1.63–3.01; p<0.001). A machine learning–based spline-regularized Cox survival model identified a high-risk ctDNA threshold >20 MTM/mL associated with worse OS (p=0.034). Trajectory-based K-means identified an optimal solution at k=4 clusters, independently supported by DTW and autoencoder-based analyses. Four ctDNA kinetic phenotypes repeatedly emerged: early ctDNA clearance, early ctDNA rise, delayed ctDNA clearance, and persistently low-level ctDNA. Autoencoder-derived phenotypes demonstrated strong internal robustness (Adjusted Rand Index, ARI ≈ 0.80) and high concordance with DTW-based clustering. Phenotypes were associated with distinct radiographic response and survival patterns, with early ctDNA rise conferring the poorest OS (p=0.022). Conclusions: AI-driven modeling of quantitative ctDNA burden and longitudinal kinetics identifies biologically interpretable and reproducible prognostic phenotypes in RCC, capturing higher-order temporal features of ctDNA dynamics and providing prognostic stratification beyond baseline detectability and burden alone, thereby informing clinical decision-making regarding earlier therapy switching.