Detection of early-stage cancers using circulating orphan non-coding RNAs in blood

Background

• Orphan non-coding RNAs (oncRNAs) are a novel category of small RNAs (smRNAs) that are frequently detected in cancer and largely absent in noncancerous tissues.
• First identified in breast cancer samples from The Cancer Genome Atlas (TCGA)¹, novel oncRNAs have, since then, been discovered in additional cancer tissues from TCGA and validated in an independent cohort of tumor and adjacent normal tissues².
• We recently assessed the oncRNA content of serum and demonstrated their potential to detect colorectal³, breast⁴, and lung⁵ cancers in a liquid biopsy strategy. These investigations, however, have been limited to single cancer cohorts, and the broader applicability of oncRNAs as biomarkers in a single multi-cancer blood test has yet to be determined.
• In this study, we investigate the utility of oncRNAs as serum biomarkers for early cancer detection across eight cancer types.

Goals

• Develop and validate an artificial-intelligence (AI)-driven blood test for cancer detection in multiple cancers across a range of cancer stages.
• Build and evaluate a cancer tissue-of-origin classification model using each patient’s serum-oncRNA profile.

Methodology

• We collected 3,317 serum samples from individuals with known cancers of the bladder (n = 164), breast (n = 220), colon and rectum (n = 143), kidney (n = 293), lung (n = 295), prostate (n = 96), pancreas (n = 346), and stomach (n = 286), as well as donors with no history of cancer at time of collection (n = 1,474).
• Patients had provided informed consent and contributing centers had obtained IRB approval.
• We used 0.5 mL serum aliquots to generate and sequence smRNA libraries at an average depth of 20 million 50-bp single-end reads. Individuals were split into age-, sex-, and smoking status-matched training (1,377 cancer; 1112 control) and test (466 cancer; 362 control) sets. We then profiled all the serum samples with our catalog of TCGA-derived oncRNAs².
• We trained generative AI models with batch effect removal, library-size estimation, and expression normalization modules to predict cancer presence and tissue-of-origin (TOO) through five-fold cross-validation within the training set. For individuals with cancer and high-confidence AI prediction, we also reported TOO.
• We evaluated the generalizability of our model by predicting cancer status and TOO in the held-out test set. Predictions were averaged across the five models optimized on the training set folds.

Overview of oncRNA Profiling and AI-Driven Model for Cancer Prediction

Figure 1. Schematic of oncRNA Profiling and Modeling Pipeline

• Our generative AI model utilizes tumor-derived oncRNAs discovered in TCGA tissue samples for downstream applications in serum.

Study Demographics

Result 1: Ability of oncRNA-Based Model for Prediction of Overall Cancer Status in Serum

Figure 2. Overall Model Performance by ROC and Sensitivity at 95% Specificity

• (A) The ROC curve demonstrated an AUC of 0.96 (95% CI: 0.96-0.97) in the training set, and (B) an AUC of 0.97 (95% CI: 0.96-0.98) in the test set. AUC 95% confidence intervals were calculated by bootstrapping.
• (C) Sensitivities at 95% specificity for the training set (purple) and test set (orange) stratified by tumor stage. Confidence intervals report the 95% confidence intervals calculated using the Clopper-Pearson method within each group. Bar plots in the top panel show the number of samples corresponding to each tumor stage and training/test set.

Result 2: Discrimination Between Cancer Patients and Cancer-Free Controls in Serum Across Cancer Diagnoses

Figure 3. Model Performance By Cancer

• Our model had high accuracy (AUC ≥ 0.94), demonstrating robust prediction across eight cancer types, within both the training (A) and held-out test (B) sets.
• AUCs ranged from 0.95 (95% CI: 0.92–0.98) in urothelial cancer to 0.99 (95% CI: 0.98–1.00) in lung cancer within the test set (B).
• Sensitivities at 95% specificity were lowest for urothelial cancer (0.71, 95% CI: 0.54–0.84) and highest for lung cancer (0.99, 95% CI: 0.93–1.00) in the test set (B).

Result 3: Model Accurately Predicts Tumor Tissue-of-Origin for Cancer Patients within Serum

Figure 4. Performance and Score Distribution of the Tissue-of-Origin Model

• (A) For samples with cancer tissue-of-origin prediction, our held-out testing cohort had an accuracy of 0.88 (95% CI: 0.84-0.92) using the top predicted cancer and 0.95 (95% CI: 0.92-0.97) for the top two predictions. 95% CIs were computed through bootstrapping. The baselines shown in gray represents the expected performance of a random guess given cancer prevalence within our study. (B) The heatmap shows the predicted scores of the model for each sample and cancer type.

Conclusions

• Our results show that circulating serum oncRNAs captured through a liquid biopsy assay can be used to accurately detect a shared cancer signal in a single, multi-cancer early detection test.
• We also demonstrate that a multi-modal generative AI model trained on oncRNA profiles can robustly and accurately predict cancer tissue-of-origin within serum for samples with detectable cancer signals.
• Given the limitations of retrospective studies and the use of frozen, archival samples, we plan to further validate our results in the future through a larger, prospective study.