Deconvolution of Transcriptomic Data Shows Biologically and...

WHY

?

Aliaksandra Kakoichankava1,3,4, Andrea S. Bauer2, Aliaksandr Siatkouski3, Yury Arlouski1, Ivan Zhyltsou1, Konstantin Dedyulya4,

Gunnar Dittmar5, Jörg D. Hoheisel2 and Petr V. Nazarov5,*

Deconvolution of Transcriptomic Data Shows Biologically

and Clinically Relevant Signals in Pancreatic Tumors

(1) Vitebsk State Medical University, Belarus; (2) German Cancer Research Center, Germany; (3) Vitebsk Regional Hospital, Belarus; (4) NatiVita, Belarus; (5) Luxembourg Institute of Health, Luxembourg; (*) Corresponding author: petr.nazarov@lih.lu

Pancreatic cancer is a significant challenge to oncology. The earlystages of the disease are asymptomatic, which limits diagnosis andtreatment of the neoplastic process and thus lead to bad survivalprognosis [1,2].

High mortality

Large social impact

High progression rate

Asymptomatic?

Tran

scri

pt-

om

ics

Bailey:[2]

96 samples

TCGA:183 samples

DKFZ:[1]

457 samples

ICA

Stat

isti

cally

ind

epen

den

tsi

gnal

s (S

): m

etag

en

es

×

Independent component analysis transforms the data into amatrix product of statistically independent transcriptional signalsand their weight.

gen

es, n

samples, m

Biological functions enriched by

contributing genes of each component

components, k

Clinical information: gender, tumor subtype, survival, sample purity.

Acknowledgements: AK was supported by the travel grant ofUniversity Grenoble Alpes. This work was supported by theLuxembourg National Research Fund C17/BM/11664971 "DEMICS" .

References1. Bauer et al., International Journal of Cancer, 2018, 142(5):1010-212. Bailey et al., Nature, 2016, doi:10.1038 3. Nazarov et al., BMC Genomics, 2019, 12(1):132

Fraction of immune cells observed by a pathologist

Co

mp

on

ent

linke

d t

o

lym

ph

ocy

te p

rese

nce

Weights (M): metasamples

Research goals:1. Identify pathophysiological processes affecting survival ofpatients with pancreatic cancer.2. Characterize tumor purity in unsupervised manner.3. Predict survival of new patients.

1

2

3

Datasets (D):

Dnm = Snk × Mkm

Components identified by ICA were annotated by biologicalfunctions (GO) and linked to survival using Cox regression asis described in [3].

com

po

nen

ts, k

Increased risk:• keratinization• cell cycle• response to hypoxia• neoangiogenesis• cornification• activation of ERK-

signaling

Reduced risk:• hormone secretion activity

(normal function)• digestion• antigen binding

No effect:• immune

response• gender• axon

development

Unlike in melanoma [3], no direct link wasfound between immune response and survival:perhaps due to a dual / antinomic effect.

Risk score (RS) is calculated as theweighted sum on scaled rows of M andCox log hazard ratio (H). Stability of thecomponents (R2) is also considered.

Here we combined DKFZ (training) and Bailey (testing)datasets. ICA was performed on the joint data. Risk scores(RS) were calculated as in [3] and visualized:

― high RS (above median), ― low RS (below median)

Conclusions:

• Pathophysiological processes that affect survival in patientswith pancreatic cancer are cell proliferation andkeratinization. No strong effect of immune components onsurvival was detected.

• Tumor purity was characterized. It strongly correlated withindependent observations of immune cells in DKFZ cohort andin silico estimation in Bailey dataset.

• Risk score calculated for the testing dataset (Bailey) stronglycorrelated with the survival (p-value < 0.001).

ICA was able to detecttranscriptional signals specific to

stroma and tumors. Abundance ofnormal pancreas tissue as well asimmune cells was detected in anunsupervised manner and correlatedto an independent observation of animmuno-histopahthologist. For Baileydataset such correlation wasobserved with in silico predictions.

https://gitlab.com/biomodlih/consica