Naslov
Prediction of Cell-of-Origin of Cancers Using Gene Mutation Profiles

Autori
Štancl, Paula ; Karlić, Rosa

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

Izvornik
“HDIR-6: Targeting Cancer” The 6th Meeting of the Croatian Association for Cancer Research with International Participation – Book of Abstracts / Ozretić, Petar - Zagreb : Hrvatsko društvo za istraživanje raka (HDIR), 2022, 49-49

ISBN
978-953-48672-1-1

Skup
6th Meeting of the Croatian Association for Cancer Research with International Participation: Targeting Cancer (HDIR-6)

Mjesto i datum
Zagreb, Hrvatska, 10.11.2022. - 12.11.2022

Vrsta sudjelovanja
Poster

Vrsta recenzije
Nije recenziran

Ključne riječi
ishodišna stanica tumora ; strojno učenje ; epigenetika
(cell-of-origin ; machine-learning ; epigenetics)

Sažetak
Cancers of unknown primary origin (CUP) account for 3-5% of all cancers and still present a challenge for treatment in clinical practice. We previously developed a method that successfully identifies the cellof-origin (COO) of a cancer by modeling the relationship between chromatin features of the COO fand the cancer’s mutational landscape, obtained from whole-genome sequencing (WGS) data. Although several high-performing WGS- based methods to predict the COO were developed recently, methods of similar accuracy based on whole-exome sequencing (WXS) data are still missing. Since the cost and computational time needed for analysis is much lower for WXS compared to WGS data, the development of such a method might be very useful for diagnosis and treatment of cancer patients. Our study aimed to develop a statistical model that utilizes mutational profiles in genes obtained by either WXS or WGS, and chromatin state across those genes, to predict the COO. We analyzed a publicly available melanoma cohort from the International Cancer Genome Consortium, consisting of 183 patients with mean 224 single-base substitutions per gene. ChIP-seq data for six histone modifications (H3K27ac, H3K27me3, H3K36me3, H3K4me1, H3K4me3, H3K9me3) and input were downloaded from the ENCODE project. Read counts were normalized using FPKM over all genes on the hg19 human genome. We used wavelet transformation to determine the optimal scale for analysis of diverse data types and trained a multiple linear regression model with 10-fold cross validation to compute the amount of variance of aggregated mutations across genes explained by the epigenome of each COO. The model with the highest variance explained indicates the COO for a specific cancer type. Our WXS-based model correctly identified melanocytes as the COO of melanoma. The variance explained for the bestperforming model was ~33% and ~52% for all and only protein-coding genes, respectively, when using non-normalized epigenetic features. Wavelet transformation caused a significant increase of prediction accuracy, with explained variance of around ~60% for both all and only protein-coding genes. Residual and over-representation analyses detected a specific group of protein-coding genes involved in melanin metabolic processes and pigmentation as informative for predicting COO. The results show that we are able to use the cell’s epigenome and cancer’s gene mutation profile to predict the cell-of-origin.

Izvorni jezik
Engleski

Znanstvena područja
Biologija

POVEZANOST RADA