Breast cancer recurrence risk stratification using rapid, cost-effective TempO-Seq profiling and an XGBoost classifier outperforms Oncotype DX

Introduction: The market standard Oncotype DX 21 gene Recurrence Score predicts risk of distant recurrence in HR+/HER2- breast cancer, has a 7-14 day turnaround (TAT), requires 50-300 ng total RNA from FFPE sections with ≥5 mm² tumour, and costs ~$4,000. The low-cost extraction free targeted sequencing TempO-Seq® assay uses lysates of FFPE without RNA extraction or reverse transcription, has a 2 day TAT, and measures the whole transcriptome or any actionable subset of genes. We evaluated if TempO-Seq profiling of lysates from a single 1 mm 2 tissue microarray (TMA) core could match or surpass the clinical performance of Oncotype DX. Methods: A TempO-Seq panel of the 21 Oncotype DX signature genes was implemented which incorporated attenuation of highly expressed gene signals. Sample processing and reaction conditions were optimized to deliver robust data from a single 5 µm thick, 1 mm diameter TMA core. Results: Biological reproducibility of the TempO-Seq workflow was robust, with a median intercore correlation r>0.80. We developed a Support Vector Machine (SVM) learning model using 70% of the samples from a cohort of 86 patients, for which we had 10-year recurrence data and Oncotype DX scores from matched sections. Testing the other 30% of patients, this model accurately identified 100% of the patients with recurrence and 71% of the patients without recurrence. In comparison, Oncotype DX identified 65% of the patients with recurrence as high/intermediate risk and 52% of the patients without recurrence as low risk. Recognizing that TempO-Seq's performance could have been biased by training and testing on the same dataset, we trained an XGBoost algorithm, which is more reliable for handling independent cohorts with potential batch effects, using 70% of an independent cohort of 245 samples for which we only had 10-year recurrence data. Testing with the remaining samples, the XGBoost model correctly identified 80% of the patients with recurrence and 48.7% of the patients with no recurrence. Additionally, the model maintained its consistency when classifying all the samples in the 86 sample cohort, correctly identifying 76.8% of the patients with recurrence and 48.9% of the patients with no recurrence. We also trained the model to classify 90% of patients with recurrence as high risk, and 31% of patients with no recurrence as low risk. Conclusions: TempO-Seq outperformed Oncotype DX results from the 86 patient cohort in classifying patients with recurrence, and models classifying patients from multiple cohorts could be built that were equivalent to or outperformed published Oncotype DX data. Thus, TempO-Seq enables accurate and reproducible prediction of recurrence from a minimal FFPE input, suggesting its potential as a fast TAT, lower cost alternative for risk stratification in HR⁺/HER2⁻ breast cancer.