Machine learning–derived B-cell epitopes classifiers for early detection of renal cell carcinoma.

Abstract

4554 Background: Renal cell carcinoma (RCC) remains a significant cause of cancer mortality in the United States, with poor outcomes for advanced-stage disease and limited tools for early detection. Tumor-specific antibodies, known to develop early in other solid tumors, offer biomarker opportunities for early RCC detection. This study aimed to leverage Serum Epitope Repertoire Analysis (SERA), an advanced platform for profiling B cell epitopes, to develop a machine learning-based classifier capable of distinguishing patients with RCC from those with benign renal masses and from individuals without known renal neoplasms. Methods: We obtained 564 serum or plasma samples from 1) 260 patients with pathologically confirmed RCC, spanning all stages; 2) 21 patients with benign renal masses (predominantly oncocytoma and angiomyolipoma); and 3) 283 age-matched non-RCC controls (self-reported healthy donors). The SERA platform uses a library of 8 billion unique 12-mer peptides, each expressed on a DNA-barcoded E. coli strain. Number and type of peptides bound by an antibody are identified by next-generation sequencing, enabling comprehensive profiling of B cell epitopes. Using machine learning, a classifier was trained on a subset of 178 samples (88 RCC and 90 healthy controls) to predict the presence of RCC in the validation cohort of 386 samples (172 RCC, 21 benign renal masses, and 193 healthy controls). The area under the receiver operating characteristic curve (AUC) served to evaluate the classifier's performance, overall and stratified by RCC stage. Results: Using the SERA platform, 26.4 million potential amino acid motifs were scored based on enrichment in RCC versus controls, yielding 7,244 motifs that met the predefined thresholds for inclusion in the classifier. These features were used to train a 10,000-tree random classification forest. In validation, the model achieved an AUC of 0.76 (95% confidence interval [CI]: 0.72 - 0.81), and scores were not significantly different (Mann-Whitney U test, alpha = 0.05) in the benign renal lesion control samples vs. healthy controls. Performance was consistent across both early- and late-stage RCC, with an AUC of 0.78 (95% CI: 0.70–0.85) for stage 1, 0.72 (95% CI: 0.49–0.95) for stage 2, 0.81 (95% CI: 0.70–0.92) for stage 3, and 0.75 (95% CI: 0.68–0.81) for stage 4 RCC, each compared to controls, demonstrating robust detection across all disease stages. Conclusions: Our findings suggest that a non-invasive SERA-based classifier can distinguish RCC from benign renal masses and healthy controls, with consistent performance across all stages of RCC. The robust detection of early-stage RCC underscores the potential of this approach to enhance early diagnosis of RCC and to guide clinical management while obviating the need for renal mass biopsy. Future studies will focus on refining the classifier and validating its performance in larger, multi-institutional cohorts.