Harnessing cell-free orphan non-coding RNAs as a predictive measure of long-term survival in neoadjuvant breast cancer therapy

Background

• Pathologic complete response (pCR) is a valuable metric for predicting survival following neoadjuvant therapy (NAT) in breast cancer. Blood-based biomarkers may further refine the prediction of treatment outcomes to inform follow-up treatment decisions.
• Orphan non-coding RNAs (oncRNAs) are a novel category of small RNAs (smRNAs) that are frequently detected in cancer and largely absent in non-cancerous tissues.
• While first identified in breast cancer samples from The Cancer Genome Atlas (TCGA)¹, novel oncRNAs have now been discovered in additional cancer tissues from TCGA and validated in an independent cohort of tumor and adjacent normal tissues².
• In this study, we explore the feasibility of a tumor-naive oncRNA-based liquid biopsy assay for predicting distant recurrence-free survival (DRFS) in breast cancer patients using a standard blood draw.

Goals

• Investigate the utility of an independently trained, oncRNA-derived AI model for predicting prognosis in the I-SPY2 trial through a tumor-naive liquid biopsy assay measured following NAT.

Methodology

• Eligibility for the I-SPY2 (NCT01042379) trial includes women diagnosed with stage II/III breast cancer with MammaPrint ”High Risk” of recurrence.
• Our cohort consisted of 538 patients in the I-SPY2 trial with available serum following NAT and prior to surgery (timepoint T3).
• We isolated small RNA from 0.8 mL of patient serum, generated sequencing libraries, and sequenced them at an average depth of 50 million 100-bp single-end reads.
• We developed an AI model for predicting an oncRNA risk score using our catalog of tumor-derived oncRNAs discovered in The Cancer Genome Atlas. Our AI model was trained on oncRNA profiles from an independent cohort of 719 serum samples collected at diagnosis from individuals with cancer using tumor size as the outcome of interest.
• We assessed risk of recurrence by oncRNA in the I-SPY2 cohort post-NAT(T3) and associated this measure with DRFS using cutpoints established for each of the two pCR groups: high- (n=40) and low-risk (n=498). Cutpoints were verified through cross-validation and almost all folds converged on asingle value for each pCR group, which was used in our analysis.

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

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.

I-SPY2 Trial Design

Figure 2. Study schema for the I-SPY2 trial

I-SPY2 uses adaptive randomization to assess efficacy of novel drugs in sequence with standard chemotherapy, with the goal of identifying treatments for patients based on molecular characteristics to achieve a high patient pCR rate.

Study Demographics

Table 1. Study demographics for our I-SPY2 cohort

HR status was missing for 3 patients, RCB was missing for 2 patients, and T stage was missing for 9 patients.

Figure 3. Breakdown of patient receptor subtypes by treatment arm

Our I-SPY2 study cohort consisted of 1 control arm and 9 experimental (3 Completed [gray], 6 Graduated [black]) treatment arms. Patients were randomized based on ER and HER2 status and MammaPrint score.

Results: Prognostic Value of oncRNA Risk Score in Combination with Existing Clinicopathologic Variables

Table 2. Continuous oncRNA risk score is predictive of distant recurrence in univariate and multivariate model

Hazards ratios (HR) and 95% Confidence Intervals (CI) from Cox models were reported with p-values from the Wald test.

^# Reported at timepoint T3 (post-NAT).

^† Continuous variables are normalized sothat the unit of measure for the HR is per standard deviation.

^‡ All significant variables from univariate analysis remain significant when combined in multivariate analysis.

Figure 4. oncRNA risk score separates risk of distant recurrence beyond pCR alone.

Kaplan-Meier survival curves for predicting DRFS when stratifying by (A) pCR alone, (B) oncRNA risk alone, or pCR further stratified by oncRNA risk score group within patients that (C) achieved pCR and those that (D) did not achieve pCR.

Through a Cox multivariate analysis, we identified oncRNA predicted risk (high vs. low;HR=4.1 [95% CI: 2.4–6.8]) as an independent prognostic factor when combined with pCR. The continuous oncRNA predicted risk measure was also significant in a multivariate analysis with pCR (p = 2.38x10^-3).

Conclusions

• We demonstrate the utility of a tumor-naive oncRNA-based liquid biopsy assay and AI model as a predictive measure of treatment outcome tool and as a prognostic biomarker that can be used in addition to pCR or RCB to further stratify patients as low- or high-risk of recurrence.
• The integration of our oncRNA assay with existing biomarkers and clinicopathologic variables has the potential to add clinical value in predicting outcomes and prioritizing patients who would benefit from additional care.