Naslov
Detecting homologous recombination deficiency in tumors using machine-learning algorithms

Autori
Štancl, Paula ; Foulkes, William D. ; Polak, Paz ; Karlić, Rosa

Vrsta, podvrsta i kategorija rada
Sažeci sa skupova, sažetak, znanstveni

Izvornik
Systems approaches in cancer: System Book of abstracts / Štagljar, Igor ; Polychronidou, Maria ; Klingmüller, Ursula - Split - Croatia : MedILS - Mediterranean institute for life sciences, 2021, 26-26

ISBN
978-953-55188-3-9

Skup
EMBO Workshop Systems approaches in cancer

Mjesto i datum
Split, Hrvatska, 21.09.2021. - 26.09.2021

Vrsta sudjelovanja
Predavanje

Vrsta recenzije
Međunarodna recenzija

Ključne riječi
defekti homologne rekombinacije, strojno učenje, genomika
(homologous-recombination deficiency, machine learning, genomics)

Sažetak
Cancer whole-genome sequencing provided new insights into the mutational landscape of various tumors, processes that drive their development, and how to exploit these genetic instabilities for a personalized therapeutic approach. The most common and widely accepted way for designing a specific therapy for each patient involves identifying certain somatic or germline mutations in genes like DNA-repair genes. Such treatments include the best-known example of poly ( ADP- ribose) polymerase (PARP) inhibitors in patients with deficient homologous recombination (HR) pathways. In the clinics, several PARP inhibitors have demonstrated improvement in progression-free survival in patients harboring mutations in certain HRrelated genes. Nowadays, there are four FDA-approved PARP inhibitors for treatments of patients with pathogenic BRCA1/2 somatic or germline mutation. Despite the success of selecting and treating HRdeficient patients based on testing for pathogenic BRCA1/2 mutations, there are patients sensitive to PARP inhibitors with a similar phenotype to HR-deficient tumors but are lacking mutations in BRCA genes. An increasing number of patients that can benefit from PARP inhibitor therapy can now be discovered by the latest methods based on the WGS analysis. Different genomic scars and damages, from mutational signatures to large-scale transition, have been associated in cancers with HR-deficient tumors harboring BRCA1/2 mutations . Two tools, CHORD and HRDetect, based on machine-learning approaches utilize these different genomic features of HR- deficient cancers to estimate the probability of a certain tumor to be HR-deficient. Both of these tools have a high accuracy of above 90% for detecting HR-deficient tumors, however, their application in clinical diagnostics needs to be validated further across a variety of tumors. Here we have benchmarked CHORD and HRDetect prediction for breast, pancreatic, ovarian, or prostate cancer patients from previous studies. The goal was to assess the performance of these classifiers and to investigate new alterations in genes related to HR deficiency. For each patient, we looked into the biallelic inactivation of HR- related genes alongside the promoter hypermethylation of those genes to evaluate the performance of the two tools. The average sensitivity across these four tumor types was above 98% (area under the curve (AUC) = 0.98) with both tools. The majority of HR-deficient samples contained BRCA1/2 germline mutations followed by BRCA1/2 somatic mutations. The promoter methylation status of BRCA1 genes in breast and ovarian cancer patients with HR deficiency was shown as a valuable resource in determining a tumor's HR status. These samples do not contain clear pathogenic variants in HR-related genes and therefore would not receive the appropriate therapy using standard targeted panel sequencing in clinical diagnostics. Using 10-fold nested cross-validation we estimated the optimal threshold values for each cancer type in both tools since the threshold for determining HR status has not been optimized for specific cancer types. Although some optimal thresholds may significantly vary from the default ones, the accuracy in majority tumors was the same or higher for about 2-3% except for the optimal HRDetect threshold for pancreatic cancer where an increase was 9%. Alongside using these WGS approaches in improving clinical diagnostics, they can expand existing knowledge of gene repertoire associated with HR defects tumors in those samples declared to be HR deficient by these tools.

Izvorni jezik
Engleski

Znanstvena područja
Biologija

POVEZANOST RADA