28MO Real-world Whole Sequencing Data of Ovarian Cancer Patients

BackgroundPoly(ADP-ribose)-polymerase (PARP) inhibitor therapy has dramatically changed outcomes for women with high-grade serous ovarian carcinoma. Germline or somatic pathogenic variants in BRCA1 or BRCA2 are strong predictive biomarkers for PARP therapy, but patients with other causes of homologous recombination deficiency (HRD) also strongly benefit. NHS patients have commercial testing for HRD using the Myriad MyChoice assay, but using WGS through the GLH could provide a more cost-effective and universally accessible assay for NHS patients.MethodsWe compared the performance of 3 bioinformatics algorithms for HRD detection (CHORD, HRDetect, Dragen) for ovarian cancer patients included in the 100K dataset to perform real-world response and survival analysis.ResultsHRD was associated with better overall survival outcome across all 3 algorithms in patients treated with platinum-based therapy (CHORD: n=259, HR=0.32 [95% CI 0.19–0.52], Wald test=20.94, p«0.001; Dragen: n=268, HR=0.38 [95% CI 0.23–0.62], Wald test=14.9, p<0.001; HRDetect: n=239, HR=0.28 [95% CI 0.16–0.48], Wald test=21.24, p«0.001). The results of the analysis showed substantial concordance, both in between each bioinformatics algorithm and in correlation with the BRCA1/2 status. However, for accurate correlation of the WGS-based HRD scores with the measures of response, there is a prerequisite of high-quality clinical data.ConclusionsClinically relevant HRD prediction is feasible using WGS data and supports further development of these algorithms for NHS use in the Genomics Medicine Service. Future improvements will use a multimodal machine-learning (ML) model to define separate weights for classes of clinical data or genomic data (HRD) and subclasses of data points. We will also include multimodal data as additional covariates in the models, including radiomic and pathology image analysis data. Incorporating multi-dimensional data, in particular pathology and radiological imaging parameters, into an ML model is fundamental for improved integrative biomarkers of HRD.Legal entity responsible for the studyThe authors.FundingHas not received any funding.DisclosureJ.D. Brenton: Financial Interests, Personal, Invited Speaker: GSK; Financial Interests, Personal, Invited Speaker: AstraZeneca; Financial Interests, Personal, Advisory Board: AstraZeneca; Financial Interests, Personal, Invited Speaker: Tailor Bio; Financial Interests, Personal, Stocks/Shares: Tailor Bio; Financial Interests, Institutional, Invited Speaker: Clovis Oncology; Financial Interests, Institutional, Invited Speaker: Aprea AB; Non-Financial Interests, Personal, Member: Association of Cancer Physicians. All other authors have declared no conflicts of interest. BackgroundPoly(ADP-ribose)-polymerase (PARP) inhibitor therapy has dramatically changed outcomes for women with high-grade serous ovarian carcinoma. Germline or somatic pathogenic variants in BRCA1 or BRCA2 are strong predictive biomarkers for PARP therapy, but patients with other causes of homologous recombination deficiency (HRD) also strongly benefit. NHS patients have commercial testing for HRD using the Myriad MyChoice assay, but using WGS through the GLH could provide a more cost-effective and universally accessible assay for NHS patients. Poly(ADP-ribose)-polymerase (PARP) inhibitor therapy has dramatically changed outcomes for women with high-grade serous ovarian carcinoma. Germline or somatic pathogenic variants in BRCA1 or BRCA2 are strong predictive biomarkers for PARP therapy, but patients with other causes of homologous recombination deficiency (HRD) also strongly benefit. NHS patients have commercial testing for HRD using the Myriad MyChoice assay, but using WGS through the GLH could provide a more cost-effective and universally accessible assay for NHS patients. MethodsWe compared the performance of 3 bioinformatics algorithms for HRD detection (CHORD, HRDetect, Dragen) for ovarian cancer patients included in the 100K dataset to perform real-world response and survival analysis. We compared the performance of 3 bioinformatics algorithms for HRD detection (CHORD, HRDetect, Dragen) for ovarian cancer patients included in the 100K dataset to perform real-world response and survival analysis. ResultsHRD was associated with better overall survival outcome across all 3 algorithms in patients treated with platinum-based therapy (CHORD: n=259, HR=0.32 [95% CI 0.19–0.52], Wald test=20.94, p«0.001; Dragen: n=268, HR=0.38 [95% CI 0.23–0.62], Wald test=14.9, p<0.001; HRDetect: n=239, HR=0.28 [95% CI 0.16–0.48], Wald test=21.24, p«0.001). The results of the analysis showed substantial concordance, both in between each bioinformatics algorithm and in correlation with the BRCA1/2 status. However, for accurate correlation of the WGS-based HRD scores with the measures of response, there is a prerequisite of high-quality clinical data. HRD was associated with better overall survival outcome across all 3 algorithms in patients treated with platinum-based therapy (CHORD: n=259, HR=0.32 [95% CI 0.19–0.52], Wald test=20.94, p«0.001; Dragen: n=268, HR=0.38 [95% CI 0.23–0.62], Wald test=14.9, p<0.001; HRDetect: n=239, HR=0.28 [95% CI 0.16–0.48], Wald test=21.24, p«0.001). The results of the analysis showed substantial concordance, both in between each bioinformatics algorithm and in correlation with the BRCA1/2 status. However, for accurate correlation of the WGS-based HRD scores with the measures of response, there is a prerequisite of high-quality clinical data. ConclusionsClinically relevant HRD prediction is feasible using WGS data and supports further development of these algorithms for NHS use in the Genomics Medicine Service. Future improvements will use a multimodal machine-learning (ML) model to define separate weights for classes of clinical data or genomic data (HRD) and subclasses of data points. We will also include multimodal data as additional covariates in the models, including radiomic and pathology image analysis data. Incorporating multi-dimensional data, in particular pathology and radiological imaging parameters, into an ML model is fundamental for improved integrative biomarkers of HRD. Clinically relevant HRD prediction is feasible using WGS data and supports further development of these algorithms for NHS use in the Genomics Medicine Service. Future improvements will use a multimodal machine-learning (ML) model to define separate weights for classes of clinical data or genomic data (HRD) and subclasses of data points. We will also include multimodal data as additional covariates in the models, including radiomic and pathology image analysis data. Incorporating multi-dimensional data, in particular pathology and radiological imaging parameters, into an ML model is fundamental for improved integrative biomarkers of HRD.