Serum-based colorectal cancer detection using orphan noncoding RNAs
Hani Goodarzi1, Jeffrey Wang2, Oluwadamilare I. Afolabi2, Lisa Fish2 , Helen Li2 , Kimberly H. Chau2, Patrick Arensdorf2, Fereydoun Hormozdiari2, Babak Alipanahi2
1UCSF School of Medicine, University of California, San Francisco, CA, 2Exai Bio Inc., Palo Alto, CA
Small non-coding RNAs (sncRNAs) have established roles as posttranscriptional regulators of cancer pathogenesis.
We previously reported a novel and previously unannotated class of sncRNAs that were found in breast cancer tissue but not in normal tissue adjacent to the tumor, which we termed orphan non-coding RNAs(oncRNAs).1 Since then, we have identified and validated novel oncRNAs in multiple cancer tissues, using data from The Cancer Genome Atlas (TCGA) and other independent cohorts.2
We recently showed that these oncRNAs can also be detected in sera and demonstrated prognostic value for treatment response among breast cancer patients.3
Early detection of colorectal cancer (CRC) can drastically improve survival odds, reduce treatment complexity and side effects, and improve patient quality of life.4
We hypothesize that oncRNAs can be used as biomarkers in a liquid biopsy strategy to detect CRC across a range of cancer stages and tumor sizes
Develop and validate a methodology that uses machine learning (ML) to accurately predict CRC status based on oncRNA profiles detected in patient sera.
Our study cohort consists of 191 frozen serum samples from clinically diagnosed colorectal cancer patients (n=96) and age- and sex-matched individuals from the general population with no known diagnosis of cancer (n=95). Samples were acquired from three commercial biobanks and processed for small RNA (smRNA) sequencing. Dates of blood draw for serum collection range from 2009 to 2022.
Subjects were treatment-naive at sample collection and were selected to represent all stages of CRC (I–IV) as well as a broad range of ages of onset, including patients <45 years old.
Patients had provided informed consent and contributing centers had obtained IRB approval.
RNA was extracted from frozen serum samples of ≤1.0ml volume and prepared for sequencing. Sample libraries were sequenced to an average depth of 18.8 million 50 bp single-end reads per sample.
oncRNAs were previously identified in multiple cancer tissues, using data from TCGA as a discovery cohort. Of this multicancer library of oncRNAs, 57,663 were significantly present in TCGA CRC samples. To refine our TCGA library of CRC-associated oncRNAs for applications in serum, we filtered out smRNA sequences found in sera of an independent non-cancer control cohort (N=31). OncRNAs that were detected in more than one control serum sample were filtered out, yielding a final set of 53,814 CRC-significant oncRNAs.
This filtered library of oncRNAs was used as a reference to generate oncRNA expression profiles by cataloguing and quantifying oncRNAs for each individual serum sample (N=191).
• These oncRNA expression profiles were used to build an ensemble of logistic regression models to make predictions of CRC vs. control. The ensemble model was trained and evaluated using a 5-fold cross-validation setup. Within each training fold only oncRNAs observed in >4% of samples and yielding an odds ratio for CRC >1 were used to train and validate the model.
oncRNA Library Creation and Profiling
Result 1: oncRNA Content Differentiates Cancer Status
Figure 1. oncRNA Content in Control and Breast Cancer Serum Samples
Of the 53,814 TCGA CRCspecific oncRNA species, 36,282 (67.4%) were observed in the study cohort (N=191).
Total sequencing depthnormalized oncRNA content, the aggregate count of all detected oncRNAs within each sample, was significantly higher in cancer samples (one-sided Mann-Whitney U test, P=3.5e-14).
Result 2: Prediction of Colorectal Cancer Status
Figure 2. ROC Curve of an Ensemble Model
A five-fold cross validation of an ensemble of logistic regression model’s CRC prediction performance on our study cohort (N=191).
On average, 3,285 oncRNAs were used as features within each fold.
Overall area under the ROC curve (AUC) across folds is 0.964 (95% CI: 0.938–0.99).
The model achieved an overall sensitivity (true positive rate) of 90.6% (95% CI: 82.9%–95.6%) with specificity set at 90% across folds.
Figures 3 & 4. Sensitivities for CRC Detection by Cancer Stage (I–IV) and Tumor T Category (T1–T4)
For each subgroup, using the model, sensitivity was calculated with specificity set at 90% (based on recent CMS reimbursement publication). 95% confidence intervals were calculated using the Clopper-Pearson method.
Sensitivities for CRC detection were similar and high across all stages and tumor T categories
Analyzing oncRNA data with machine learning models accurately predicted colorectal cancer (CRC) across all cancer stages (I–IV) and tumor categories (T1–T4).
This oncRNA-based liquid biopsy technology is compatible with standard sample requirements enabling integration into conventional clinical workflows.
The results will be validated prospectively in further population studies.
JW, OA, LF, HL, KC, FH are full-time employees of Exai Bio. BA and PA are cofounders, stockholders, and full-time employees of Exai Bio. HG is co-founder, stockholder, and advisor of Exai Bio.