Research has shown that circulating tumor DNA (ctDNA) is informative of tumor load and tumor evolution in both solid and hematological cancer. The ability to detect mutations in ctDNA holds the promise for an accurate and non-invasive approach to assess minimum residual disease as well as treatment response in the future. However, as ctDNA often makes up only a small fraction of cell-free DNA recovered from the plasma, traditional methods of targeted sequencing research often face a poor signal-to-noise ratio that cannot be overcome with deep coverage. Here we present a novel research method that is capable of detecting ultra-rare mutations at allelic frequency below 0.5%. This approach leverages target multiplexing capabilities of the Ion AmpliSeq™ technology with some important modifications to the sample preparation procedures. The new protocol requires as little as 20 ng of input DNA and offers a sample-toanswer turn-around time in under 24 hours. To support the analysis of this new approach, we have further developed a novel Bayesian statistics that models the propagation of potential artifacts introduced during amplification and sampling effects during sequencing to differentiate false positives (variants observed in sequencing data that were not present in input DNA) from true mutations that were present at very low levels in the original research sample. We successfully applied this new method to detect spike-in mutant DNA in both cell line (Coriell GM24385) and cfDNA samples. Specifically, we demonstrated the detection of 140 COSMIC genomic aberrations found in 23 frequently mutated genes. In preliminary study, the method achieved greater 90% sensitivity and specificity.

1. Ann Mongan, Richard Chien, Dumitru Brinza, Kelli Bramlett, and Fiona Hyland
Thermo Fisher Scientific  180 Oyster Point Blvd South San Francisco CA 94080
II.B Bayesian modelI. ABSTRACT
Research has shown that circulating tumor DNA (ctDNA) is
informative of tumor load and tumor evolution in both solid and
hematological cancer. The ability to detect mutations in ctDNA
holds the promise for an accurate and non-invasive approach to
assess minimum residual disease as well as treatment
response in the future. However, as ctDNA often makes up only
a small fraction of cell-free DNA recovered from the plasma,
traditional methods of targeted sequencing research often face
a poor signal-to-noise ratio that cannot be overcome with deep
coverage.
Here we present a novel research method that is capable of
detecting ultra-rare mutations at allelic frequency below 0.5%.
This approach leverages target multiplexing capabilities of the
Ion AmpliSeq™ technology with some important modifications
to the sample preparation procedures. The new protocol
requires as little as 20 ng of input DNA and offers a sample-to-
answer turn-around time in under 24 hours. To support the
analysis of this new approach, we have further developed a
novel Bayesian statistics that models the propagation of
potential artifacts introduced during amplification and sampling
effects during sequencing to differentiate false positives
(variants observed in sequencing data that were not present in
input DNA) from true mutations that were present at very low
levels in the original research sample.
We successfully applied this new method to detect spike-in
mutant DNA in both cell line (Coriell GM24385) and cfDNA
samples. Specifically, we demonstrated the detection of 140
COSMIC genomic aberrations found in 23 frequently mutated
genes. In preliminary study, the method achieved greater 90%
sensitivity and specificity.
II. MATERIALS AND METHODS
II.A Sample preparation
cfDNA isolation: Plasma from blood sample was obtained by
centrifugation at 1600 x g for 10 min at 4C. cfDNA was
extracted from the plasma fraction using MagMAX™ cfDNA
isolation protocol. 20 ng of cfDNA was used for library prep.
Library prep: Libraries were generated with a modified Ion
AmpliSeq™ protocol targeting a subset of the Cancer Hotspot
Panel v2. The protocol provides several improvements for a
ligation-free and automation-friendly workflow for variant
detection at low allelic frequencies. DNA was amplified using a
novel high fidelity polymerase master mix with gene specific
primers for a limited number of cycles. The resulting products
were further amplified with primers containing Ion adapter
sequences. The libraries were purified and size-selected with
bead-based reagents. Libraries were quantified by qPCR.
Sequencing: Sequencing template preparation and 318™ Chip
loading were performed on the Ion Chef™ Instrument using the
Ion PGM™ Hi-Q™ Chef Kit. Sequencing was performed on the
Ion PGM™ System with the Ion PGM™ Hi-Q™ Sequencing Kit.
Successful detection of 40 COSMIC hotspot
mutations at allelic frequency below 0.5%
IV. SUMMARY
• We have developed an assay and an analysis solution
that is capable of detecting variants at allelic frequency
below 0.5%.
• The assay has been tested on GM23485 spiked with
AcroMetrix sample as well as mixed cfDNA.
For Research Use Only. Not for use in diagnostic procedures.
© 2015 Thermo Fisher Scientific Inc. All rights reserved. All trademarks are
the property of Thermo Fisher Scientific and its subsidiaries unless
otherwise specified.
CHR POS NAME REF OBS GENE CDS Mutation AA Mutation
chr9 133750319 COSM49071 C A ABL1 c.1150C>A p.L384M
chr9 133750356 COSM12604 A G ABL1 c.1187A>G p.H396R
chr2 29432664 COSM28056 C T ALK c.3824G>A p.R1275Q
chr5 112173917 COSM18852 C T APC c.2626C>T p.R876*
chr5 112173930 COSM19230 T C APC c.2639T>C p.I880T
chr5 112173947 COSM19330 C T APC c.2656C>T p.Q886*
chr11 108205769 COSM21636 G C ATM c.8084G>C p.G2695A
chr7 55211080 COSM21683 G A EGFR c.323G>A p.R108K
chr7 55211097 COSM174732 G A EGFR c.340G>A p.E114K
chr2 212530084 COSM232263 C T ERBB4 c.1835G>A p.R612Q
chr2 212530091 COSM573362 G T ERBB4 c.1828C>A p.P610T
chr2 212530135 COSM1405173 T C ERBB4 c.1784A>G p.D595G
chr7 148508727 COSM37028 T A EZH2 c.1937A>T p.Y646F
chr4 153258983 COSM22971 G A FBXW7 c.832C>T p.R278*
chr8 38282147 COSM1292693 G A FGFR1 c.816C>T p.N272N
chr10 123274774 COSM36906 A G FGFR2 c.1144T>C p.C382R
chr10 123274794 COSM36904 T C FGFR2 c.1124A>G p.Y375C
chr10 123274810 COSM1346272 T C FGFR2 c.1108A>G p.T370A
chr4 1808331 COSM24802 G T FGFR3 c.2089G>T p.G697C
chr20 57484420 COSM27887 C T GNAS c.601C>T p.R201C
chr15 90631838 COSM33733 C T IDH2 c.515G>A p.R172K
chr15 90631879 COSM1375400 T C IDH2 c.474A>G p.P158P
chr15 90631934 COSM41590 C T IDH2 c.419G>A p.R140Q
chr4 55961023 COSM48464 C A KDR c.2917G>T p.A973S
chr4 55561702 COSM77973 C T KIT c.92C>T p.P31L
chr4 55561764 COSM1146 G A KIT c.154G>A p.D52N
chr4 55144148 COSM22415 C A PDGFRA c.1977C>A p.N659K
chr4 55144172 COSM1430086 A G PDGFRA c.2001A>G p.S667S
chr10 89720874 COSM1349625 A G PTEN c.1025A>G p.K342R
chr10 43617416 COSM965 T C RET c.2753T>C p.M918T
chr18 48586262 COSM14163 C T SMAD4 c.931C>T p.Q311*
chr18 48586291 COSM14167 G C SMAD4 c.955+5G>C p.?
chr22 24143240 COSM992 C T SMARCB1 c.472C>T p.R158*
chr7 128845101 COSM13145 C T SMO c.595C>T p.R199W
chr19 1221293 COSM29005 C T STK11 c.816C>T p.Y272Y
chr19 1221319 COSM21355 C T STK11 c.842C>T p.P281L
chr17 7574003 COSM11073 G A TP53 c.1024C>T p.R342*
chr17 7574012 COSM11286 C A TP53 c.1015G>T p.E339*
chr17 7574018 COSM11071 G A TP53 c.1009C>T p.R337C
chr17 7574026 COSM11514 C A TP53 c.1001G>T p.G334V
H0: the observed variant is a result of a PCR error
H1: the observed variant is a true mutation
The likelihood of encountering a PCR error is dependent on which
round of the N PCR cycle the error is introduced, we can model a
latent variable fj which represents the fraction of reads containing
an error if the error is introduced at round j of N PCR cycles.
P(H0|k, mi) can be expressed in terms of P(f |H0), P(f |H1), P(k|mi,f ),
and prior probabilities P(H0), P(H1) using Bayesian inference (Eq
1). Eq 2 and 3 describe the probability of observing k reads with
the variant, taking in account all the possibilities where errors can
be introduced.
 
 
 
 
   
   
       1100
00
10
0
0
0
,,
,
,,
,
,
,
HPmHkPHPmHkP
HPMHkP
mkHPmkHP
mkHP
mkP
mkHP
mkHP
ii
ii
i
i
i
i






   
   



j
jji
j
iji
HfPfmkP
HmfkPmHkP
0
00
,
,,,
   
   



j
jji
j
iji
HfPfmkP
HmfkPmHkP
1
11
,
,,,
(Equation 1)
(Equation 2)
(Equation 3)
P(fj |H0) and P(fj|H1) are discrete probabilities describing the
fraction of reads carrying an error introduced at PCR cycle j. For
simplicity, the summation of these probabilities over all possible j is
approximated by summing the error probabilities from the first 3
rounds with an average error term representing errors from the
remaining rounds (Table 1).
III. RESULTS
0 10 20 30 40 50
01020304050
Family size
Supportingreadsrequired
0 5 10 15 20 25 30 35
707580859095100
Supporting reads
%familysize
0 5 10 15 20 25 30 35
0.000.010.020.030.040.050.06
Supporting reads
Error
Experimental design and results
• Pooled cfDNA from 5 donors
• Pooled cfDNA is diluted into a known cfDNA background
sample to a target allelic frequency below 0.5%
• 10 true positive variants targeted
• 47 random positions chosen to evaluate specificity
• Average results in 4 runs
Sensitivity: 92.5%
Specificity: 93.6%
III.A Detection of COSMIC variants in AcroMetrixTM
control in GM24385 background
III.B Analysis of ultra-rare variants in cfDNA samples
Table 1: Model: Fraction of molecules carrying an error is dependent on
which PCR cycle the error is introduced
Cycle 1 2 3 1 and 2 1 and 3 Others
P(fj |H0) e 22e 23e e 2 e 3 0.99
fj 0.5 0.25 0.125 0.75 0.875 0.001
Table 2: Table of selected 40 COSMIC variants in frequently mutated
cancer genes
Coverage Coverage
cfDNA
isolation
Library
Prep
Templating
Sequencing
Data
Analysis
MagMAX™ cfDNA isolation
20 ng cfDNA
Ion Chef™ Instrument
Ion PGM™ Hi-Q™ Chef Kit
Ion AmpliSeq™ Panels
Novel Ion AmpliSeq™
Library Kit
Ion PGM™ 318™ Chip
Ion PGM™ Hi-Q™
Sequencing Kit
Torrent Suite™
Variant Caller
5 hrs
4.5 hrs
4 hrs
1 hr
0.5 hr
Figure 1: Sample preparation and sequencing workflow
Figure 5: Schematic representation of cfDNA sample mix
Figure 2: (A) The fraction of molecules required to support a variant
decreases with increasing number of observations derived from a single
template. (B) Max errors associated with each required faction in (A).
%ofmoleculesderivedfromthesametemplate
Figure 3: Sequencing metrics
Figure 4: (A) Allelic frequencies of detected variant are in the range of
0.05-0.4%. (B) The number of molecules captured increases with
sequencing depth.
A. Evidence required to
support variant call
A. Allelic frequency of 40 selected variant B. Number of molecules captured
B. Probability of error at different
threshold requirements
True variant
Error
Figure 1: Propagation of true variant and
hypothetical errors during amplification.
All daughters of the original templates are
expected to carry the true variant, but PCR
errors are expected to be uncorrelated.
Template DNA
1.01 G
Total Alignment Bases
0.3 X
Average Coverage
Depth of Reference
92% Aligned Bases
8% Unaligned
6,303,959
Total Reads
60% Usable Reads
ISP Summary
Position in Read
177 bp
Mean
198 bp
Median
201 bp
Mode
Read Length
15000
20000
25000
30000
35000
40000
45000
0.0
0.1
0.2
0.3
0.4
Allelicfrequency(%)
15000
20000
25000
30000
35000
40000
45000
1500
2000
2500
3000
3500
Numberofmoleculescaptured