PSB 2018 presentation

Samir Rachid Zaim, Qike Li, A. Grant Schissler, and Yves A. Lussier
Emergence of pathway-level composite biomarkers
from converging gene set signals
of heterogeneous transcriptomic responses
http://lussiergroup.org/publications/PathwayMarker
@lussiergroup
@UA_CB2
#PrecisionMedicine

Problem: unproductive assumptions
for discovery of transcriptome biomarkers in common diseases
• 30,000 NIH “biomarker” grants in 25 yrs (> $2.5 billion/year) [1]
o unproductive: only 12 FDA-approved cancer biomarkers
(2012-2017)
o limited success in clinical practice
• Conventional transcriptome biomarker discovery designed for
the average patient:
o single biomolecule assumed concordantly altered across
patients
o patient-specific biomarker signal remains undetected
[1] Ptolemy, A.S. and N. Rifai, Scand J Clin Lab Invest Suppl, 2010. 242: p. 6-14.

Single-subject (SS) pathway-level studies  emerging cross-subject pathway signal
• Hypothesis: pathway-level signal emerges from heterogeneous dysregulated genes in each
patient (responsive genes), as they coordinate to alter a multi-gene function (e.g., pathway)
• Pathway Biomarker Framework:
o Identify responsive genes (red & blue below) and altered pathways in each subject
(single-subject studies)
o Followed by cross-subject pathway-level statistics
Figure: Three SS studies. Same altered pathway in each patient,
discoverable in each single-subject study
Patient 1 = SS study 1 Patient 2 = SS study 2 Patient 3 = SS study 3
Simulation parameters:
20% responsive genes
50% up regulated genes

Enabling precision medicine using transcriptomes:
pathway-level interpretation in a single subject (one study per subject)
.

.
Background: single-subject transcriptome analyses of altered pathways
Vitali F., Li Q., et. al., Briefings in bioinformatics, 2017.
Pathifier (PNAS 2013;110:6388);
IndividPath (Brief Bioinform
2016;17:78);
iPAS (Bioinform2014;30:I422)
N-of-1-pathways methods:
(Lussier Group).
Wilcoxon: JAMIA 2014;21:1015;
Mahalanobis Distance:
Bioinformatics 2015;31:i293;
ClusterT: Statistical Methods in
Medical Research 2017;
MIxEnrich: BMC Medical Genomics
2017;10:27
kMEn: J Biomed Inform 2017;66:32
Methods

Simulation Study
Simulation Overview
1. Simulation parameters
2. Conventional cohort-based analysis from the transcriptome
3. Single-subject analysis followed by pathway-level analysis across subjects
4. Evaluation: contrasting approaches

1. Simulation Parameters
Model a variety of biological and clinical conditions
Noise distributed according to negative binomial distribution
(parameters estimated from real TCGA data)
Understand the impact of bidirectional expression (p)
Assess the power of conventional methods in small samples vis-à-vis single subjects
Table 3: Simulation Parameters, 54 combinations x 1000 simulations = 54,000 datasets

Normal Tumor
Gene 1
Patie
n
t 1
Patie
n
t 2
Patie
n
t N
Normal Tumor
Gene M
Normal Tumor
Gene 2

Normal Tumor
Gene 1
Patie
n
t 1
Patie
n
t 2
Patie
n
t N
Normal Tumor
Gene M
Normal Tumor
Gene 2
Paired t-test

kMEn: Li, Q., et. al., J Biomed Inform 2017;66:32

4. Evaluation
Gene set size = 40
N = 10 patients
Legend:
Black = SS-anchored discovery
Red = Conventional discovery
Recall
Precision

Take Home Message
Individual response of transcriptomes can contain
valuable pathway-level biomarkers
• Undetectable by conventional cohort-based
studies except with a large fraction of transcripts
concordantly responsive
• Unveiled by paired transcriptome analysis in each
single subject, then detectable across subjects

Acknowledgement
Yves A. Lussier, MD
Francesca Vitali, PhDColleen Kenost, EdD
@lussiergroup
@UA_CB2
#PrecisionMedicine
Helen Zhang, PhD
Acknowledgement: This study was supported in part by The University of Arizona Cancer Center, The
University of Arizona BIO5 Institute, The University of Arizona Center for Biomedical Informatics and
Biostatistics, and the University of Arizona Health Sciences Center.
Samir Rachid Zaim A. Grant Schissler, PhD
Haiquan Li, PhDJoanne Berghout, PhD

5.1 Evaluation
Gene set size = 200
N = 30 patients
Legend:
Black = SS-anchored discovery
Red = Conventional discovery
Precision
Recall

2. Generation of 54,000 simulated datasets
1. Estimate patient-specific distributions using brain tissue data from GTeX (method of
moments estimation)
2. Select a combination of simulation parameter combination
3. Select a random ‘patient-specific’ distribution (i.e., select a (μi,δi) pair of parameter
estimates at random)
4. Simulate normal and tumor transcriptomes
5. Repeat 1000 simulations x 54 combinations of parameters
Table 4: Simulated Paired Transcriptomes
Gene subject1 subject1
A1BG 214 298
A1CF 0 0
A2M 2827 5372
A2ML1 625 474
A3GALT2 4 1

Problem Statement
Gene mu delta
1 CTAGE15 0.4 1.9
2 GATA5 0.7 3.1
3 KLK11 3.8 0.2
4 TSSC4 1253.8 0.1
5 EIF1B 4486.4 0.2
6 MOK 1041.2 0.3
7 RAX2 20.7 1.5
8 KRT15 18.2 0.0
9 PAMR1 932.0 0.3
10 ANKRD24 1570.1 1.5
• Table 1: Method of Moments
Negative Binomial parameter
estimates for brain-tissue RNA-Seq
data1
• Assuming isogenic conditions, all
patients’ gene expression
distributions are identical, differing
only by random chance
• Differing baseline risks or levels of
variability masked when everyone
gets clumped together
1. https://gtexportal.org/home/datasets
Table 1: Isogenic MME Parameter Estimates
s 2
= m +dm2

Problem Statement
Gene μ1 δ1 μ2 δ2 μ3 δ3
CTAGE15 1.333 0.938 0.358 1.934 0.358 1.934
GATA5 0.667 0.750 1.333 0.375 0.364 2.200
KLK11 2.444 0.181 6.222 0.232 4.000 0.388
TSSC4 963.111 0.060 1132.111 0.190 1017.545 0.134
EIF1B 3132.556 0.044 4658.111 0.137 2755.182 0.426
MOK 1301.333 0.124 1177.000 0.164 1006.091 0.425
RAX2 2.444 0.056 5.111 0.919 2.455 0.966
KRT15 25.000 0.529 45.111 0.357 21.818 0.274
PAMR1 2406.444 0.447 1237.556 0.935 2130.273 0.371
ANKRD24 1496.667 0.753 2283.778 1.296 1640.182 0.693
Table 2: Heterogenic MME Parameter Estimates
• Table 2: Now – assuming enough biological replicates exist – we estimate each
patient’s distributions. The above parameters relate the mean and variance as
follows:
sij
2
= mij +dijmij
2

Masking Individual Patient Behavior
Gene PAMR1: Pat1 v. Pat2 Distributions
Gene Expression
Frequency
0 2000 4000 6000 8000 10000
0100200300400500600700
Patient1
Patient2
Gene RAX2: Pat1 v. Pat2 Distributions
Gene Expression
Frequency 0 5 10 15 20 25 30
050010001500200025003000
Patient1
Patient2
Gene μ1 δ1 μ2 δ2
RAX2 2.444 0.056 5.111 0.919
PAMR1 2406.444 0.447 1237.556 0.935

Each SS pathway-level signals  emerging cross-subject pathway signal
• SS Hypothesis: pathway-level signal
emerges from heterogeneous
dysregulated genes in each patient
(responsive genes), as they
coordinate to alter a multi-gene
function (e.g., pathway)
• Pathway biomarker framework:
o Identify responsive genes and
altered pathways in each
subject (single-subject studies)
o Followed by cross-subject
pathway-level statistics

PSB 2018 presentation

More Related Content

What's hot (20)

Similar to PSB 2018 presentation (20)

Recently uploaded (20)

PSB 2018 presentation

Editor's Notes