Beyond detection: AI-based classification of breast cancer invasiveness using cell-free orphan non-coding RNAs

Background

• Approximately 1 in 8 women will be impacted by breast cancer in their lifetimes¹. Earlier detection of breast cancer through screening has improved survival.
• Over-diagnosis of ductal carcinoma in-situ (DCIS), however, has brought significant costs and complications associated with over-treatment².
• Liquid biopsies have the potential to complement existing screening methods by enabling earlier detection and differentiating invasive breast cancer (BC) from DCIS.
• We previously demonstrated high sensitivity and specificity for early detection of invasive BC by using a blood-based liquid biopsy platform to analyze a novel category of cancer-associated small RNAs, termed orphan non-coding RNAs (oncRNAs)³.
• We have shown that our oncRNA-based liquid biopsy assay can not only detect cancer, but also risk stratify BC.

Methods

• We utilized The Cancer Genome Atlas (TCGA) small RNA profiles to discover a library of 20,538 oncRNAs that were significantly enriched among 1,103 breast tumors compared to 349 controls from normal tissues spanning multiple tissue sites and limited to samples from females.
• The diagnostic performance of these oncRNAs was assessed in an independent cohort of serum samples from 708 women, including 380 breast cancer patients (221 invasive BC and 159 DCIS; mean age: 58.0 ± 13.4 years) and 328 age-matched controls (mean age: 58.4 ± 13.7 years).
• Samples were sourced from Indivumed (149 invasive BC, 101 DCIS, and 121 controls), Proteogenex (72 invasive BC, 58 DCIS, and 9 controls), and MT Group (198 controls).
• We sequenced the small RNA content from 1 mL of serum from these patients at an average depth of 21.6 million 50-bp single-end reads. We detected 19,736 (96%) of the breast cancer-specific oncRNA library within at least one sample in this cohort.
• We then trained a multi-class generative AI model using 5-fold cross-validation to predict breast cancer and then distinguish invasive BC from DCIS.

Results

• Our oncRNA-based generative AI model achieved an overall AUC of 0.95 (95% CI: 0.94–0.97) for prediction of breast cancer versus cancer-free controls. At 90% specificity, overall model sensitivity is 89.7% (86.5%–92.8%). For DCIS and stage I invasive BC, the model has a sensitivity of 88.1 (82.0%–92.7%) and 90.4% (81.9%–95.8%), respectively, both at 90% specificity.
• In the second step, restricting to samples flagged as cancer, we observed an overall AUC of 0.9 (0.87–0.93). We had a sensitivity of 70.5% (63.1%–77.2%) at 90% specificity for distinguishing DCIS from invasive BC.
• For distinguishing DCIS from invasive BC, we observed that the sensitivity of the model in detecting DCIS decreases from grade 1 DCIS samples (0.87, 95% CI: 0.47–1.0) to grade 2 (0.63 95% CI: 0.35–0.85) and grade 3 (0.44, 95% CI: 0.28–0.62).
• We observed an increase in oncRNA risk score as a function of DCIS grade and tumor stage, with control samples showing the lowest scores and stage IV invasive BC samples showing the highest score.

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

Figure 1. Schematic of oncRNA Profiling and Modeling Pipeline

Our generative AI model utilizes tumor-derived oncRNAs discovered in TCGA breast cancer samples for downstream applications in serum. Our AI uses a two-arm semi- supervised variational auto-encoder. Triplet margin loss allows our AI model to minimize the impact of technical variations without explicit encoding of known confounders. The model predicts tumor absence, presence of DCIS, or presence of invasive BC.

Table 1: Study Demographics

Ability of oncRNA-Based Model to Predict Breast Cancer Risk

Figure 2. Performance for detection of invasive BC and DCIS.

(a) oncRNA-based model achieved sensitivity of 89.7% at 90% specificity for detection of breast cancer. (b) Model sensitivity increases for invasive BC compared to DCIS and is higher for later-stage cancers. (c) oncRNA-based model can distinguish DCIS from invasive BC. (d) Sensitivity of the model decreases for high-grade DCIS samples, which resemble invasive BCs.

Figure 3. A two-stage test to detect breast cancer tumor invasiveness.

(a) Sankey plot shows an initial test with 88% sensitivity for DCIS patients and 91.4% for invasive BC patients. The second rule-out test from the same oncRNA profile has sensitivity of 62.4% at 90% specificity. (b) The oncRNA risk score is lowest when tumor is not present and increases for DCIS and later-stage cancer samples.

Conclusions