A1 Refereed original research article in a scientific journal
HR-SC—an academic-developed machine learning framework to classify HRD-positive ovarian cancer patients and predict sensitivity to olaparib.
Authors: Beltrame, L.; Mannarino, L.; Sergi, A.; Velle, A.; Treilleux, I.; Pignata, S.; Paracchini, L.; Harter, P.; Scambia, G.; Perrone, F.; González-Martin, A.; Berger, R.; Arenare, L.; Hietanen, S.; Califano, D.; Derio, S.; Van Gorp, T.; Dalessandro, M.L.; Fujiwara, K.; Provansal, M.; Lorusso, D.; Buderath, P.; Masseroli, M.; Ray-Coquard, I.; Pujade-Lauraine, E.; Romualdi, C.; D’Incalci, M.; Marchini S.
Publisher: Elsevier BV
Publication year: 2025
Journal: ESMO Open
Journal name in source: ESMO Open
Article number: 105060
Volume: 10
eISSN: 2059-7029
DOI: https://doi.org/10.1016/j.esmoop.2025.105060
Web address : https://doi.org/10.1016/j.esmoop.2025.105060
Self-archived copy’s web address: https://research.utu.fi/converis/portal/detail/Publication/492099043
Background: High-grade serous ovarian cancer (OC) patients with defects in the homologous recombination repair (HRR) pathway benefit from poly (ADP-ribose) polymerase inhibitor (PARPi) maintenance therapy. Clinically approved methods for identifying HRR status suffer from limitations, such as high failure rates and costs, leading to the clinical need for innovative approaches. To this aim, we developed Homologous Recombination Signature Classifier (HR-SC), a machine learning (ML) algorithm that integrates BRCA1/BRCA2 status and copy number signatures, leveraging the availability of OC samples recruited from two international clinical trials, namely PAOLA-1 (dataset A) and MITO16A/MaNGO-OV2 (dataset B).
Patients and methods: 569 DNA samples from datasets A and B were sequenced using a custom library design covering a backbone of structural regions and the full-length sequence of 375 genes. Data were used to train, validate (dataset A), and test (dataset B) HR-SC, using BRCA1/BRCA2 status and a compendium of previously annotated copy number signatures. Lastly, HR-SC was compared with already established approaches to evaluate its predictive and prognostic role.
Results: In dataset A, where the failure rate was 6.4%, HR-SC showed a sensitivity of 92%, a specificity of 94.73%, an accuracy of 93.18%, a positive predictive value (PPV) of 95.83%, and a negative predictive value (NPV) of 90%. In dataset B, where the failure rate was 4%, HR-SC showed a sensitivity of 90.16%, a specificity of 82.86%, an accuracy of 87.5%, a PPV of 90.16%, and an NPV of 82.86%. Univariate and multivariate survival analyses demonstrated its predictive role [progression-free survival (PFS): hazard ratio (HR) = 0.42, P < 0.0001; overall survival (OS): HR = 0.63, P = 0.036] and its prognostic role (PFS: HR = 0.56, P = 0.0095).
Conclusions: The study demonstrates that HR-SC is a novel, clinically feasible solution with a low failure rate for predicting HRR status in OC patients and underscores the importance of leveraging ML approaches for advancing precision oncology in the era of personalized medicine.
Keywords: copy number signatures; homologous recombination deficiency; machine learning; ovarian cancer.
Downloadable publication This is an electronic reprint of the original article. |  |
Funding information in the publication:
This work was supported by the Associazione Italiana per la Ricerca sul Cancro [grant numbers IG 2021-ID 25932 projects to SP (principal investigator), CO-2018-12367051 (Ministero della Salute) to SP (principal investigator), IG 29071 to CR, IG 19997 and IG 30381 to SM], and from Ricerca Corrente [grant L3/13 from Ministero della Salute to SP and PNRR-MAD-2022-12375663 to SP]; the MITO16A/MaNGO-OV2 trial was partially supported by Roche; the Fund for Scientific Research—Flanders [grant number FWO Vlaanderen 18B2921N to TVG (senior clinical investigator)].
