Abstract
Spectral Counts approaches (SpCs) are largely employed for the
comparison of protein expression profiles in label-free (LF)
differential proteomics applications. Similarly, to other comparative
methods, also SpCs based approaches require a normalization procedure
before Fold Changes (FC) calculation. Here, we propose new Complexity
Based Normalization (CBN) methods that introduced a variable adjustment
factor (f), related to the complexity of the sample, both in terms of
total number of identified proteins (CBN(P)) and as total number of
spectral counts (CBN(S)). Both these new methods were compared with the
Normalized Spectral Abundance Factor (NSAF) and the Spectral Counts log
Ratio (Rsc), by using standard protein mixtures. Finally, to test the
robustness and the effectiveness of the CBNs methods, they were
employed for the comparative analysis of cortical protein extract from
zQ175 mouse brains, model of Huntington Disease (HD), and control
animals (raw data available via ProteomeXchange with identifier
PXD017471). LF data were also validated by western blot and MRM based
experiments. On standard mixtures, both CBN methods showed an excellent
behavior in terms of reproducibility and coefficients of variation
(CVs) in comparison to the other SpCs approaches. Overall, the CBN(P)
method was demonstrated to be the most reliable and sensitive in
detecting small differences in protein amounts when applied to
biological samples.
Introduction
In recent years, Proteomics has gained centrality in Omics studies for
basic and translational applications, especially for diagnostic
purposes and for targeted and/or personalized medicine [[40]1–[41]3].
Differential proteomics approaches are attracting particular attention
due to the possibility to compare protein expression profiles from
multiple biological conditions, e. g. wild type vs mutant or vs
pharmacologically treated, etc [[42]4,[43]5].
Different methodologies have been developed to carry out the
qualitative-quantitative analysis of the protein content in samples
using both labeled and label-free approaches. The labeling based
methods employed fluorescent (e. g. Difference in Gel Electrophoresis
(DIGE)) [[44]6,[45]7] or isotopic reagents (as for Isobaric Tag for
Relative and Absolute Quantitation (iTRAQ), Stable Isotope Labeling
with Amino acids in Cell culture (SILAC) and Tandem Mass Tagging (TMT)
[[46]8–[47]12]) to differently label the two or more proteomes under
investigation. These strategies provide higher levels of
reproducibility because are based on the contemporaneous
electrophoretic separation and/or tandem mass spectrometry (MS/MS)
analysis of the samples. Despite these advantages, labeling procedures
are time-consuming and very expensive for the high cost of the labeling
reagents [[48]13].
More recently, Label-Free (LF) approaches have been introduced for
quantification of proteomic profiles by exploiting liquid
chromatography coupled with tandem mass spectrometry (LC-MS/MS)
analyses [[49]14–[50]16]. These procedures [[51]17–[52]19] represented
an effective solution to overcome these drawbacks due to their reduced
costs and simplified sample preparation. Nevertheless, longer analysis
time and high-performance MS/MS instruments are needed.
Label-Free quantification has emerged as a consequence of the great
technical advances in the development and design of high-resolution
LC-MS/MS instruments [[53]20], such as the Orbitrap mass analyzer
[[54]21]. This analytical approach relies on the measurement of two
main parameters, Extracted Ion Chromatogram (XIC) and Spectral Counts
(SpCs) [[55]19,[56]22]. Measurement of XIC is a MS1-based strategy in
which the ion current associated with each peptide ion is individually
considered. The relative quantification of proteins can then be
obtained by evaluating the total current associated with all the
peptides belonging to a specific protein in the two conditions
[[57]23]. SpCs quantification is an MS/MS-based procedure. The number
of fragmentation events, (i.e. the spectral counts), measured for all
the peptides belonging to the same protein in each sample are summed up
resulting in the relative quantification of the protein in the
different conditions [[58]24,[59]25].
Both XIC and SpCs methods need massive use of tandem mass spectrometry,
producing a large amount of data that needs to be normalized, i.e.
corrected for instrumental or uncontrollable variations before becoming
suitable for quantitative purposes. In particular, specifically,
designed bioinformatic tools are required to manage and process data,
while appropriate normalization methods have to be used
[[60]26,[61]27].
Normalization procedure has been assessed in multiple ways by several
authors [[62]28,[63]29]. In general, XIC methods require steps of
higher complexity such as isotopic patterns resolution, feature
detection, retention time alignment, etc. Therefore, well defined and
automated processing and normalization algorithms have been integrated
into specifically designed software packages, as for MaxQuant
[[64]30–[65]32].
In SpCs methods, instead, fragmentation events are crucial for both
protein identification and quantification. Commonly, software workflows
for SpCs quantification result in discrete, unnormalized data with
normalization left to the user [[66]33,[67]34]. Although XIC-based
methods show a greater accuracy and linear range, the access to the
last generation high-resolution instruments and/or computational
processing for XIC data are often prohibitive, with the consequence
that SpC methods are still a valuable alternative for a large number of
applications, as demonstrated by recent literature [[68]35–[69]38].
Here we propose a new Complexity Based Normalization (CBN) method
introducing new adjustment factors in the SpC normalization formula
related to the complexity of the analyzed samples in terms of total
identified proteins (CBN(P)) or spectral counts (CBN(S)). Both CBN
methods are compared with the multiple label-free quantification modes
based on the SpC approaches, namely Ratio of Spectral Count (R[SC])
[[70]33] and Normalized Spectra Abundance Factor (NSAF) [[71]39].
The new proposed normalization approach was optimized on standard
protein mixtures with different complexity, and then applied to the
differential analysis of the proteome profiles extracted from the brain
cortex of zQ175 knock-in mice [[72]40], an animal model of Huntington
Disease (HD), in comparison with the wild-type counterpart. The CBN(P)
method was demonstrated to be the most reliable and sensitive in
detecting small differences in protein amounts when applied to
biological samples.
Materials and methods
Preparation of standard mixtures
Six standard tryptic peptide mixtures (Waters, Milford, Massachusetts,
US) including Bovine Hemoglobin Chain A (HBA, Uniprot ID [73]P01966, B.
Taurus), Bovine Hemoglobin Chain B (HBB, [74]P02070, B. Taurus), Bovine
Serum Albumin (BSA, [75]P02769, B. Taurus), yeast Enolase (ENO,
[76]P00924, S. Cerevisiae), yeast Alcohol Dehydrogenase (ADH,
[77]P00330, S. Cerevisiae) and rabbit Glycogen Phosphorylase (PYG,
[78]P00489, O. Cuniculus) were dissolved in 0.2% formic acid (FA) and
spiked into an E. Coli total tryptic digest (Waters, Milford,
Massachusetts, US) solution, so that final concentration of the digest
reaches 67 ng/μL. In all mixtures obtained and indicated with A-E
letters, HBA, HBB, BSA and PYG were always added in a fixed amount,
while only ENO and ADH were added in variable amounts ranging from 1/25
to 5 times the fixed standards (i.e. 0.94–23.5ng/μl for ENO and 0.76–19
ng/μl for ADH). The obtained standard mixtures have been employed for
the optimization of mass spectrometry analyses and for setting
quantification methods.
The zQ175 mouse model
All protocols involving animals were carried out in accordance with
institutional guidelines in compliance with Italian law (D. Lgs no.
2014/26, implementation of the 2010/63/UE) and authorization
n.324/2015-PR issued May 6, 2015, by Ministry of Health. The Ethics
Committee of the University of Milano approved studies in mice (Ethical
Approval 74/13; Ethical Approval 74/14). JAX stock number was
B6J.129S1-Htt tm1Mfc /190ChdiJ. For biochemical analyses, animals were
euthanized by dislocation and all efforts were made to minimize
suffering. Genotyping of the zQ175 (C57BL/6J) mouse colony was
performed by PCR of genomic DNA obtained from tail samples (Nucleo Spin
Tissue, Macherey-Nagel, catalog 740952.250) at weaning. CAG repeats of
zQ175 mice were sized by using the following PCR primers: forward:
CATTCATTGCCTTGCTGCTAAG; reverse: CTGAAACGACTTGAGCGACTC. Cycling
conditions were 94°C for 10 minutes, 30 cycles × (96°C for 30 seconds,
57°C for 30 seconds, 72°C for 30 seconds), 72°C for 7 minutes.
Preparation of total protein lysates from mouse cortical tissues
For proteomics and biochemical analyses, animals were euthanized by
dislocation and the brains were rapidly removed and the cerebral
cortices were immediately excised from the brain, frozen in liquid
nitrogen and smashed. Specifically, cortical tissues from three
symptomatic homozygous zQ175 mutant and three wild type (WT) mice at 50
weeks of age were used. The samples were then lysed by combining
chemical and mechanical methods. Cortical tissues were crushed and
repeatedly pipetted in the lysis buffer (50 mM Tris-HCl pH 7.4, 150 mM
NaCl, 0.1% v/v Sodium Dodecyl Sulphate (SDS), 1 mM Phenylmethylsulfonyl
fluoride (PMSF), 1% v/v Nonidet P-40 (NP40), protease inhibitors
EDTA-free), then passed ten times each through a syringe. Following
these mechanical processes, they were put on a stirring wheel, (20
minutes at 4°C), and then centrifuged, (16,100 x g, 30 minutes, 4°C).
After the lysis procedure, supernatants from all samples were collected
and quantified according to Bradford protein assay and 50 μg from each
were used for the Differential Proteomics experiment and Western Blot
assays.
SDS-PAGE and in situ hydrolysis
50μg from each sample were suspended in Laemmli buffer containing 100
mM dithiothreitol (DTT), boiled for 10 minutes and loaded on a 20x20 cm
4% - 15% bis-acrylamide gradient gel. After SDS-PAGE separation, the
gel was stained with Coomassie Blue Brilliant and each lane was excised
in 29 slices by using the same cutting scheme. All gel slices were in
situ hydrolyzed as previously reported [[79]41]. Finally, the peptide
mixtures were vacuum-dried and resuspended in 0.2% formic acid (FA)
solution to be analyzed by LC-MS/MS.
LC-MS/MS analyses
Quantitative data were collected on an LTQ Orbitrap XL
(ThermoScientific, Waltham, MA) coupled to a nanoLC system (nanoEasy
II). All peptide mixtures were analyzed using the same chromatographic
conditions, i.e. 3 microliters of each sample were fractionated onto a
C18 capillary reverse-phase column (100 mm, 75 μm, 5 μm) working at 250
nl/min flow rate, using a linear gradient of eluent B (0.2% formic acid
in 95% acetonitrile) in A (0.2% formic acid and 2%acetonitrile in
MilliQ water) from 5% to 40% in 80 minutes was run. MS/MS analyses were
performed using Data-Dependent Acquisition (DDA) mode: one MS scan
(mass range from 400 to 1800 m/z) was followed by MS/MS scans of the
five most abundant ions in each MS scan, applying a dynamic exclusion
window of 40 seconds. All samples were run in duplicates.
Protein identification and quantification
Raw data obtained from nanoLC-MS/MS were analyzed with MaxQuant 1.5.2
integrated with the Andromeda search engine [[80]42]. To that end, an
appropriate Fasta file was generated by downloading from UniProt and
subsequently merging amino acid sequences of both standard proteins and
whole E. Coli proteome.
The selected parameters for protein identification were the following:
minimum 2 peptides, at least 1 unique; variable modifications allowed
were methionine oxidation and pyroglutamate formation on N-terminal
glutamine; accuracy for the first search was set to 10 ppm, then
lowered to 5 ppm in the main search; 0.01 FDR was used, with a reverse
database for decoy; retention time alignment and second peptides search
functions were allowed. Protein quantification has been performed only
using unique unmodified peptides.
As output, MaxQuant resulted in discrete spectral counts, which were
then normalized by using NSAF, R[SC], or CBNs approaches.
As concerns data collected for the Differential Proteomics analysis on
zQ175 mice, following MaxQuant analysis data, each couple of technical
duplicate was mediated for every protein, obtaining a set of 6 data (3
for WT and 3 for zQ175). These values were employed for manual
normalization according to the above mentioned SpCs methods.
Western blot and densitometric analyses
Mice brains total protein extracts were separated by SDS-PAGE and then
transferred onto nitrocellulose membranes (Bio-rad, Hercules,
California, US). Membranes were blocked with 5% non-fat milk and then
incubated with the following antibodies: rabbit polyclonal anti-hnRNPH
(Abcam, ab154894), rabbit polyclonal anti-SERPIN B6 (Abcam, ab233229;
Proteintech, 14962-1-AP), rabbit polyclonal anti-IRGM (Abcam,
ab118569), rabbit polyclonal anti-UQCRQ (Proteintech, 14975-1-AP),
rabbit monoclonal anti-HOMER1 (Abcam, ab184955), rabbit monoclonal
anti-HTT (Cell Signalling Technology, mAb#5656), rabbit polyclonal
anti-OSBPL2 (Proteintech, 14751-1-AP), mouse monoclonal anti-β-Actin
(Origene, TA811000) and mouse monoclonal anti-Vinculin (Sigma-Aldrich,
V9131). Membranes were incubated with horseradish peroxidase-conjugated
secondary antibody (1:5000 for mouse host antibodies and 1:10000 for
rabbit host antibodies) for 45 minutes at room temperature and the
signals were detected by enhanced chemiluminescence (ECL) detection
system (Thermo Fisher Scientific, Inc., Waltham, MA).
For densitometric analyses, the software Quantity One 4.6.8 (Bio-Rad,
Hercules, California, US) has been employed; all protein band
intensities were normalized with the corresponding β–actin signal,
except HTT, whose intensities were normalized on vinculin signal.
Multiple-Reaction Monitoring (MRM) analyses
As additional validation method multiple-reaction monitoring (MRM)
approach was employed. 50 μg of cell lysates obtained as described
above were digested by trypsin onto S-Trap filters, according to the
manufacturer protocol (Protifi, Huntington, NY). The three biological
replicates of zQ175 and wild-type mice were pooled respectively, and
HTT, SERPIN B6, HOMER1, OSPBL2, SAMM50, APO-A4, CAMK2A, STIP1, RIC8A,
ATP2A2 peptide transitions were monitored with a Xevo-TQS mass
spectrometer coupled to a nanoAcquity UHPLC (Waters, Milford, MA, US)
equipped with IonKey CHIP interface. Peptide mixtures were separated on
peptide BEH C18 130Ǻ, 1.7μm, 150 μm x 50mm, iKey by using a linear
gradient of eluent B (95% acetonitrile LC-MS grade (Sigma Aldrich, St.
Louis, Missouri, US), 0.2% formic acid (Sigma Aldrich, St. Louis,
Missouri, US)) from 7% to 95% over 115 minutes working at a flow rate
of 3μl/min. Raw data were processed with Skyline v20.1.0.155 (MacCoss
Lab Software, Dept of Genome Sciences, UW) and the total area of each
peptide transition was used for the relative quantification of the
specific proteins. For each protein, at least two prototypic peptides
were selected and at least two transitions for each parent ion were
monitored, as in silico predicted by using Skyline. Proteins were not
monitored all together but in three different runs, both for wild type
and zQ175: run A, included HTT, HOMER1, RIC8A; run B, ATP2A2, APO-A4,
OSBPL2; run C, SAMM50, SERPIN B6, CAMK2A, STIP1 and in each run ACTIN
was monitored as the internal standard. Each run was analyzed in
duplicate and the total area of each peptide transitions was used for
the relative quantification of the specific proteins, normalized for
actin FC.
Statistical analyses
Multiexperiment Viewer v4.9.0 [[81]43], (MeV) was employed to perform
statistical analysis of MaxQuant output. Western blots results were
evaluated by univariate statistical analysis using GraphPad Prism 8.0,
and the results are presented as the mean ± standard deviation (SD) by
three biological replicates. The statistical significance of the
observed difference in western blot analysis was determined by
parametric (Welch’s t) or non-parametric (Mann–Whitney test) tests when
data failed the Shapiro–Wilk normality test. Mass spectrometry data
(LF) statistical significance was determined by the unpaired Student’s
t-test and differences were considered statistically significant at
Benjamini-Hochberg corrected p-value (FDR)<0.05 [[82]44].
Functional over-representation analysis
STRING database ([83]https://string-db.org/) [[84]45] (STRING v11:
Protein-Protein Association Networks With Increased Coverage,
Supporting Functional Discovery in Genome-Wide Experimental Datasets)
was used to perform a pathways enrichment analysis on differentially
expressed proteins in mouse brains identified in each method (NSAF,
Rsc, CBN(P), CBN(S)).
Results
Development of a new method for Spectral Counts normalization
A new mathematic method, called Complexity Based Normalization (CBN),
was developed for the normalization of Spectral Counts data in
differential proteomics experiments to calculate the relative Fold
Change for each protein between two experimental conditions. The new
formula ([85]1) employed for SpCs normalization is reported as follows
in comparison with the Normalized Spectra Abundance Factor (NSAF)
([86]2) and the Spectral Counts log Ratio (R[SC]) ([87]3) methods.
[MATH:
CBNx,<
/mo>M=Sx,MtM+f :MATH]
(1)
[MATH:
NSAFx<
/mi>,M=Sx,MLx∑i=1PSi,MLi
:MATH]
(2)
[MATH:
RSCx,M
=log2
(Sx,M+f
tM−Sx,M
+f)
:MATH]
(3)
In CBN methods ([88]1), S[x,M] are the Spectral Counts of protein x in
mixture M, t[M] are the total spectral counts of mixture M, f is the
complexity-based adjustment factor. In addition to common variables, in
NSAF method ([89]2) L[i] represents the amino acid length of "i"
protein, while P is the total number of proteins identified in the
analysis. Finally, in the R[SC] method ([90]3) f is a fixed adjustment
factor = 0.5.
The output of these formulas consists of a numeric value representing
the quantity of a specific protein in each mixture that is used to
calculate the relative Fold Change (FC). Both R[SC] and CBN require the
use of an adjustment factor "f" to avoid the presence of missing
values. In the R[SC] method, Beissbarth et al. [[91]46] have fixed the
f value constant to 0.5. For both CBN methods we propose the employment
of a variable adjustment factor according to sample complexity.
Specifically, the CBN formula has been developed with f = 1/P (CNB(P)),
where P is the number of proteins occurring in the MaxQuant output data
or with f = 1/t (CBN(S)), where t is the sum of all spectral counts
associated to all proteins in all mixtures. As a consequence of these
definitions, the adjustment factor becomes complexity-based, since the
P and t values are both related to the total number of proteins in the
sample.
Setting up the normalization methods on E. Coli proteome standard sample
The newly developed CBN formulas were evaluated in comparison with
existing normalization methods in differential proteomics experiments
using the label-free procedure applied to the E. Coli proteome.
Commercial tryptic digests from six different proteins (HBA, HBB, and
BSA from B. Taurus, ENO and ADH from S. Cerevisiae and PYG, from O.
Cuniculus) were spiked into a fixed matrix consisting of a constant
amount of E. Coli total proteins tryptic digest, to mimic a real sample
environment. Five mixtures (A, B, C, D, E) were prepared as described
above, in which only ADH and ENO tryptic peptides were added in
different amounts. Each mixture was analyzed by nanoLC-MS/MS in
duplicates and the raw data processed by using MaxQuant. Following
MaxQuant analysis, a set of unnormalized SpCs data was obtained and
used to test the different normalization methodologies, NSAF, Rsc, and
the newly developed CBNs, either in terms of total identified proteins
(CBN(P)) or spectral counts (CBN(S)).
A total of 409 proteins (P), including the six spiked standards were
identified in all samples by MaxQuant analysis, together with 17837
total spectral counts (t) ([92]S1 Table). The P and t experimental
values were then introduced in the CBN formulas for the computation of
the f factor.
As a first step, the reproducibility of the different approaches was
tested making use of the two technical replicates prepared and analyzed
for each mixture. As an example, [93]Fig 1A–1D shows the corresponding
scatter plots concerning the quantitative measurements of all proteins
identified in the two replicates of mixture A by the four different
normalization methods used, i.e. NSAF, R[SC], CBN(P) with f = 1/P and
CBN(S) with f = 1/t.
Fig 1. Scatter plots of the quantitative measurements of all the E. Coli
proteins.
[94]Fig 1
[95]Open in a new tab
Scatter plots of the quantitative measurements of all the E. Coli
proteins identified in the two replicates of mixture A by the four
different normalization methods used, i.e. NSAF (panel A), R[SC] (panel
B), CBN(P) with f = 1/P (panel C) and CBN(S) with f = 1/t (panel D). On
the X-axis is reported the normalized Spectral Counts of replicate 1,
and on the Y-axis the normalized Spectral Counts of replicate 2.
When the data produced by the applied method were perfectly
reproducible between the two technical replicates, the scatter plot
should result in a perfectly linear behavior (R^2 = 1) and, most
importantly, should have a unitary slope.
As showed in [96]Fig 1A–1D, a very poor correlation occurred when data
were elaborated by NSAF and Rsc in comparison to CBN methods, with a
high number of points largely scattered, therefore suggesting a lower
level of reproducibility. Linear regression best-fit values are
summarized in [97]S2 Table. It should be underlined that the scatter
plots associated with the two CBN methods showed quite similar shape
although the value of the adjustment factor was greatly different.
Furthermore, as shown in panel 1B, a large scattering of data that
could not be accommodated in a linear behavior was observed in the
R[SC]normalization method.
[98]Fig 2A reports the slopes calculated for the best fitting lines in
each pair of technical replicates analyzed with the four methods for
all samples except mixture B, since it showed a very low
reproducibility in all methods ([99]S1 Fig), suggesting the occurrence
of technical problems in the LC-MS/MS analysis. Findings reported in
[100]Fig 2A showed that all the SpCs methods displayed an acceptable
behavior except for R[SC,] whose median value was much lower than the
expected value (i.e. 1) suggesting that a fixed adjustment factor "f"
strongly affects the median value. The reproducibility of each method
was also evaluated by calculating the coefficient of variation (CV) in
the quantification of E. Coli proteins throughout the ten LC-MS/MS runs
in which the amount of E. Coli tryptic peptides was constant ([101]Fig
2B). The R[SC] method showed a very good behavior with a low
coefficient of variation whereas the NSAF displayed a greater
dispersion of data and a very high CV value. This is very likely due to
the absence of any adjustment factor in this method making NSAF very
susceptible to quantitative biases caused by the absence of data. As
showed, CBN(S) performed well, showing a CV dispersion intermediate
between NSAF and Rsc. The best result occurred in CBN(P), which
presented the lowest CV values contained in the narrowest dispersion
window, suggesting it as the most precise among the investigated
methods.
Fig 2. Comparative analysis of all spectral counts normalization methods
applied to E. Coli proteome.
[102]Fig 2
[103]Open in a new tab
(A) Median values for the best fitting slopes calculated for each pair
of technical replicates for all samples analyzed (except mixture B)
with each normalization method. (B) Coefficient of variation (CV) for
the evaluation of data dispersion for all the normalization methods
used in the analysis of E. Coli proteome. (C) Representation of the
logarithmic Fold Change distribution around the theoretical value
indicated by the point line for all normalization methods. (D) Box
plots of the CV for the fold change in the mix E / mix D pair,
calculated on all the four possible pairs of the technical replicates.
In panels B and D the box and whisker extremes represent 25–75%.
These preliminary data strengthened our hypothesis that differences in
Fold Change values are due to the different sizes of the adjustment
factor, since the shape of the data is identical.
The normalization methods were then tested for their ability to
correctly assess protein quantification when two different sets of data
were compared. Therefore, the five mixtures and the corresponding
replicates were compared by two according to the scheme: A1/B1, A2/B1,
A1/B2, A2/B2, etc., and the Fold Change (FC) values of all proteins in
each pair were calculated. As the amount of E. Coli proteome was
constant in all samples, the expecting Folding Change value for each
protein should have been equal to 1 (Log[2] (FC) = 0, [104]Fig 2C).
For each pair, the mean value of the Fold Change and its associated CV
was calculated. As an example, in [105]Fig 2D the data related to the
four possible combinations of mixture E/mixture D pair were reported.
Inspection of the results showed a very regular distribution of data
points for the CBN(P) method, with more than 75% of values having an
associated CV of 25% or less. NSAF and Rsc displayed a slightly more
dispersed distribution of Fold Change values, while CBN(S) showed the
highest dispersion of data.
Then, we moved to evaluate the statistical relevance of proposed
methods, through the calculation of False Discovery Rate (FDR). An
unpaired Student's t-test was performed to extrapolate the E. Coli
statistically significant Differentially Expressed Proteins (DEPs): the
latter represented the pool of false positives, since no variations
were expected in E. Coli proteins. The FDRs, calculated as the
percentage of false positives of total identified proteins, resulted in
5.1% for Rsc, 6.1% for both CBNs, and 7.6% for NSAF. These values are
very close to the reference one (5%), indicating a high reliability of
all methods, excepted NSAF, which showed the highest value of false
positives.
Furthermore, for each method the values of Fold Change cutoffs were
estimated, treating the standard mixtures A—C and D—E as a couple of
samples to be compared (A and C vs D and E).
The FC cutoffs were calculated for each method by evaluating the
dispersion of FCs of the unchangeable proteins of E. Coli proteome,
considered not statistically significant according to unpaired
Student's t-test (p<0.05). The FC thresholds were defined by the lower
Q1-(IQR*1.5) and upper Q3+(IQR*1.5) extremes, where Q1, Q3 represented
25 and 75 percentiles respectively, and IQR is defined as the
inter-quartile distance. The obtained FC cutoffs were summarized in
[106]Table 1:
Table 1. Methods fold changes cutoffs.
METHOD Lower FC value Upper FC value
CBN(P) 0.79 1.20
CBN(S) 0.49 1.53
NSAF 0.45 1.60
RSC 0.51 1.51
[107]Open in a new tab
Fold Changes cutoffs calculated for each method are reported,
indicating upper and lower significant values.
These findings showed very similar results for CBN(S), Rsc, and NSAF,
while CBN(P) performed quite differently. In particular, its narrowest
cutoffs again confirmed that CBN(P) is the most reliable method in the
detection of slight differences in protein expression levels (low FCs),
which in biological samples might be relevant.
Comparison of normalization methods for the quantification of the six
standard proteins spiked within the E. Coli proteome
The different normalization methods were then evaluated in the
quantification of the Fold Changes of the six standard proteins spiked
within the E. Coli proteome.
The Fold Change values were determined by all methods for the six
standard proteins spiked within the E. Coli proteome in all mixtures
([108]S3 Table).
The performances of the different normalization methods were evaluated
for the quantification of BSA, HBA, HBB, and PYG proteins, whose
concentration was constant in all mixtures, and for quantification of
ENO and ADH, whose amount changed in the samples. [109]Fig 3A reports
plots showing the distribution of the logarithmic Fold Change for
unchangeable proteins obtained with the four normalization methods. All
methods displayed similar results indicating a general high accuracy in
the FC measurements, as mainly demonstrated in BSA and HBB, where the
mean values agree to the true value (FC = 0). In HBA the best fit
occurred in CBN(P) method, confirming very good performances in terms
of accuracy and precision. However, almost all measured values
scattered within 0.3 logarithmic units from the theoretical values,
corresponding to an absolute Fold Change of 1.2, which is reasonably
considered as an unchanged amount of protein.
Fig 3. Fold change analysis comparing all spectral counts normalization
methods.
[110]Fig 3
[111]Open in a new tab
(A) The mean Fold Change values for unchangeable proteins (BSA, HBA,
HBB, and PYG) within the ten different pairs of mixtures calculated by
the four normalization methods. The theoretical values are 0 and it is
indicated by a point line. (B) Comparison of the experimental and
theoretical Fold Change values for the relative quantification of ENO
and ADH obtained by using the SpC-based methods and reported in the
logarithmic scale. (C) Table reporting linear regression best-fit
values for all ENO and ADH linear regressions.
The linear correlations between experimental and theoretical values
obtained for all different normalization methods were investigated for
ENO and ADH, whose concentrations were changeable ([112]Fig 3B). Very
similar correlation values were obtained for both proteins in the
calculation of relative abundance determined by NSAF, Rsc and CBNs,
suggesting that all methods are equally able to correctly evaluate
differences in the amount of proteins. For ENO, linear correlations
were confirmed also decreasing 10 times the concentration of the lowest
point reported in [113]Fig 3B (data not shown), confirming a good
behavior for all methods.
In conclusion, the quantitative measurement using standard mixtures
demonstrated that both the newly proposed CBN methods resulted reliable
in the estimation of all expected FCs, either when the protein amount
changed or when it was constant, for which they showed a very good
adherence to theoretical data even in a complex mixture.
Application of CBN methods to the differential proteomics analysis of samples
from cerebral cortex from Huntington’s Disease mice
Once the effectiveness of the newly developed SpC-based normalization
methods was defined, the two CBN approaches were challenged with the
analysis of more complex samples, in which the expression level of a
large number of proteins is expected to significantly change. A
label-free differential proteomics experiment was performed on the
brain cortex proteome from mice affected by Huntington’s Disease (HD)
(homozygous mutant HTT knock in line zQ175) [[114]40] and corresponding
control samples (WT).
Equal amounts of total protein extracts from three wild type and three
mutant mice cortices were separated by SDS-PAGE ([115]S2 Fig) and each
of the six lanes was cut in 29 slices that were in situ hydrolyzed with
trypsin. The corresponding 174 peptide mixtures were then analyzed by
nanoLC-MS/MS in duplicates.
The resulting set of 348 LC-MS/MS raw data (available via
ProteomeXchange with identifier PXD017471) was processed by MaxQuant
for protein identification and quantification ([116]S4 Table). The set
of unnormalized SpCs data was then elaborated by the newly developed
CBN methods (CBN(P) and CBN(S)). The other methods (NSAF and Rsc) were
used for comparison. The lists of proteins were then filtered according
to their differential expression and statistical significance (FDR<5%)
by using the Multi experiment Viewer, as described above. Statistically
significant up- and down-regulated proteins, identified by each method
according to FC cut off values previously calculated, were reported in
the histogram in [117]Fig 4A. It is surprising to note that the two CBN
methods led to the identification of a quite different number of
statistically significant proteins. In CBN(S) as well as Rsc, it was
nearly 140, whereas the CBN(P) as well as the NSAF approach led to the
identification of only about a hundred of significant proteins.
Fig 4. Visualization of identified proteins with all methods in the WT and HD
mouse model.
[118]Fig 4
[119]Open in a new tab
(A) The histogram shows for each method the number of proteins that
appeared to be both differentially expressed and statistically
significant. (B) Venn diagram referred to all proteins. The central
area represents the proteins common to all methods.
In particular, the specific and shared proteins were highlighted in a
Venn diagram ([120]Fig 4B). Details of statistically significant
identified proteins were summarized in [121]S5 Table.
The reliability of the proposed normalization methods was further
confirmed by quantification of some selected proteins by two
independent methodologies, i.e. western blot analyses and Multiple
Reaction Monitoring (MRM) mass spectrometry analysis.
In western blot assays ([122]Fig 5A), the expression levels of IRGM1,
OSBPL2 and SERPIN B6, which were up-regulated in HD mice and hnRNP H,
HTT, SAMM50, UQCRQ and HOMER1, whose expression was instead decreased
in HD mice were evaluated on total protein extracts of three WT and
three HD mice (the same samples employed for the proteomic experiment),
in duplicate.
Fig 5. Validation of a selected group of proteins differentially expressed in
WT and HD mouse model.
[123]Fig 5
[124]Open in a new tab
(A) Western blot assays performed on total protein extracts from three
mutant (zQ175(1), zQ175(2), zQ175(3)), and three wild-type (wt(1),
wt(2), wt(3)) mice with antibodies against the selected proteins.
β-actin was used for normalization. (B) MRM superimposed traces of
transitions of one of CAMK2A proteotypic peptide reported for zQ175
(left panel) and WT (right panel) together with one ACTIN peptide
transition. (C) Densitometric analysis of data from the western blot of
panel A. The indicated values in the graph represent the percentage of
arbitrary units compared to WT to which 100% was assigned. Results are
represented as the as mean ± SD (standard deviation). The statistical
significance was evaluated by parametric (Welch’s) or non-parametric
(Mann-Whitney) tests when data failed the Shapiro–Wilk normality test.
* p < 0.05, ** p < 0.01 *** p < 0.001, **** p < 0.0001. (D) Fold Change
measured by MRM analysis of pooled HD and WT samples, respectively.
Upon the densitometric analysis, each sample was normalized with the
intensity of β-actin developed on the same membrane, except HTT
normalized with the intensity of vinculin. The normalized intensities
were then evaluated by univariate statistical analysis and the results
summarized in [125]Table 2 and graphically reported in [126]Fig 5C.
Table 2. Summary of proteins FCs validated by western blot and/or MRM.
Protein FC MRM FC WB FC CBN(S) FC CBN(P) FC RSC FC NSAF
HTT 0.59 0.42 0.08 0.59 0.14 /
CAMK2A 0.44 / / 0.67 / /
SAMM50 0.80 0.66 0.44 0.69 0.47 /
ATP2A2 0.69 / / 0.69 / /
HOMER1 0.60 0.72 0.45 0.71 0.45 /
IRGM1 / 1.93 19.73 1.22 4.09 UP
APO-A4 1.27 / / 1.29 / /
RIC8A 1.30 / 4.46 1.30 / /
STIP1 1.35 / 2.38 1.32 2.11 2.50
OSBPL2 1.32 1.18 6.37 / 2.92 /
SERPIN B6 1.33 1.99 2.55 / 2.02 2.75
HNRNP H / 0.46 0.30 / / /
UQCRQ / 0.71 0.36 / / /
[127]Open in a new tab
In the table proteins selected for validation and their FC values
calculated by LF, MRM, and WB methods are reported.
Since in several cases the differences in expression levels were about
20–30%, the western blot technique might not be enough reliable in
detections on a restricted number of replicates.
Therefore, we integrated validation experiments by quantifying some
proteins by using a tandem mass spectrometry methodology: the Multiple
Reaction Monitoring (MRM). HTT, SERPIN B6, HOMER1, OSBPL2, and SAMM50
were also monitored by MRM, together with other proteins, APO-A4,
CAMK2A, STIP1, RIC8A and ATP2A2, whose antibodies were not available or
not efficient in western blot analyses. ACTIN was monitored too and
employed as an internal standard for normalization procedure.
In detail, three zQ175 and WT cortex brain samples (the same samples
employed for the proteomic experiment) were pooled respectively, and
few micrograms were employed for a shotgun proteomics experiment by
using the MRM targeted approach. Peptide mixtures were separated by
nanoLC-MS/MS in duplicate and two or more transitions of at least two
proteotypic peptides were monitored for each of the above proteins and
employed for relative protein quantification. Areas of monitored
transitions ([128]Fig 5B, [129]S6 Table) were measured, mediated among
transitions, and all peptides belonging to the same protein, employed
to calculate FC, which were normalized with FC measured for ACTIN in
each couple of runs. FC results are summarized in [130]Table 2 and
graphically reported in [131]Fig 5D.
A strong agreement among all measured FCs emerged by comparing the
results obtained by western blot and MRM, confirming the reliability in
the validation of both methods. Moreover, for all proteins, the
variation trends detected by LF approaches were confirmed by both
methods ([132]Table 2) also in the detection of slight FCs.
In detail, levels of HTT protein are significantly reduced in HD mice
compared to controls, according to previously observed [[133]47]. This
protein was identified by Rsc and CBN approaches as statistically
significant, although the three methods measured very different FC
values: CBN(S) FC = 0.08, CBN(P) FC = 0.59, Rsc FC = 0.14. By comparing
these values with those obtained by western blot (0.42) and MRM (0.59),
the FC measured by CBN(P) was the only perfectly in agreement with both
values. This finding was not a random or a sporadic event, but it was
recurrent in all proteins, from those quantified also in other methods
(HTT, SAMM50, HOMER1, IRGM1, RIC8A, STIP1, OSBPL2, UQCRQ) to those
statistically significant only for CBN(P) (CAMK2A, APO-A4, ATP2A2)
([134]Table 2).
hnRNP H and UQCRQ are proteins identified as statistically significant
only by CBN(S), while SERPIN B6 varied significantly for all methods
excepted CBN(P). hnRNP H and SERPIN B6 were endorsed by western blot
assays, founding their densitometric mean values statistically
significant. Surprisingly, FC values calculated by CBN(S) for these
proteins strongly agreed with those obtained from the other methods.
Altogether these data confirm that the quantification of results
obtained by normalization of the differential proteomic data performed
by the CBN methods is reliable, although CBN(P) was confirmed to be the
best in terms of precision and reliability in comparison with CBN(S)
and the other investigated methods.
The results of the differential proteomics experiments were then
evaluated on a functional basis by gathering the differentially
expressed proteins within biological functional networks. A
bioinformatic analysis involving all the differentially expressed
proteins identified by all methods were carried out using KEGG pathways
and the STRING databases. The results are summarized in [135]Table 3
where the main pathways showing FDR<0.05 are reported. Functional
classes showing the lower FDR values include metabolic pathways,
oxidative phosphorylation, and neurodegenerative diseases (Parkinson's
Disease, Huntington's Disease, Alzheimer's Disease), confirming the
effectiveness of the newly developed methods not only in terms of
sensitivity in the detection of single up- or down-regulated proteins,
but also looking at the biological meaning of aggregated results.
Table 3. List of functional pathways identified by STRING analysis.
Pathways Description Observed Gene Count (NSAF) FDR Observed Gene Count
(RSC) FDR Observed Gene Count (CBN(P)) FDR Observed Gene Count (CBN(S))
FDR
Metabolic pathways 19 2.62E-09 42 1.10E-09 22 2.30E-06 42 1.51E-10
Oxidative phosphorylation 7 0.0037 14 9.68E-09 7 0.000105 12 7.09E-07
Parkinson's disease 9 0.00049 14 1.77E-08 8 1.53E-05 12 1.11E-06
Huntington's disease 9 0.00156 14 2.78E-07 9 9.82E-06 13 1.24E-06
Alzheimer's disease 9 0.00111 13 8.40E-07 7 0.000397 12 3.94E-06
Calcium signaling pathway 8 0.00412 7 0.0328 7 0.000566 8 0.0109
Val, Leu and Ile degradation 6 0.00049 5 0.00648 3 0.0269 5 0.00578
cGMP-PKG signaling pathway 8 0.0037 8 0.0094 5 0.02 7 0.0247
Fatty acid degradation 5 0.00194 4 0.0254 3 0.0213 4 0.0223
[136]Open in a new tab
Statistically significant proteins identified by each method were
analyzed by STRING, a bioinformatic tool for the functional
clusterization in protein networks; all pathways are reported in
descending order of FDR.
Discussion
Label-Free methods for differential proteomics investigations generate
a large amount of data that needs to be normalized, i.e. corrected for
instrumental or uncontrollable variations before becoming employable
for quantitative purposes. The choice of suitable normalization formula
is a not trivial aspect in these approaches, because it might strongly
affect the experimental results.
For Spectral Counts based Label-Free approaches, normalization
procedure has been assessed in different ways, resorting in some cases
to the use of fixed and arbitrary corrective factors, such as in Rsc
formula [[137]33], to overcome biases associated with the presence of
missing values in the analysis, as occurs in NSAF method [[138]39]
lacking of adjustment elements.
In light of these considerations, we decide to test new formulas for
SpCs data normalization, in which the corrective factor was not an
arbitrary choice of the operator, but strictly and directly defined by
an intrinsic property of the system under investigation, such as its
complexity.
A new approach defined Complexity Based Normalization (CBN) was
introduced for the normalization of Spectral Counts data in
differential proteomics experiments to calculate the relative protein
Fold Changes. The new formula employed a complexity-based adjustment
factor generating two different methods, CBN(P) where f = 1/P and
CBN(S) in which f = 1/t with P and t referring to the total number of
identified proteins (P) or the total spectral counts in the LC-MS/MS
analysis (t), respectively.
The performances in terms of reliability, precision, accuracy,
linearity, and sensitivity of the two CBN methods in the analysis of
raw data from differential proteomics experiments carried out according
to the label-free strategies were evaluated in comparison with existing
SpC-based normalization approaches, NSAF and Rsc.
Methods evaluation was carried out by analyzing a set of samples
consisting of six different standard proteins spiked in an E. Coli
total tryptic digest matrix and the capability of the methods to
quantify both the total matrix and the spiked proteins were monitored.
Both CBN(P) and CBN(S) showed very good reproducibility in the analysis
of the technical replicates and were able to correctly quantify the
unchanged amount of E. Coli proteins. When the accuracy of all methods
was compared, CBN(P) showed the best behavior, while the CBN(S)
performance was comparable with the one showed by NSAF and RSC methods;
a similar trend has been detected also for unchangeable spiked
proteins.
In addition, other differences could be observed between all methods
when the coefficient of variation in the quantification of E. Coli
proteins was examined. Although both CBN methods displayed a low
dispersion of data and low CV, CBN(P) displayed a better performance
than CBN(S) suggesting that the precision is affected by the size of
the adjustment factor.
The evidence that the choice of the adjustment factor value might
affect performances of SpC-based normalization approaches is confirmed
also for NSAF and R[SC] methods. In fact, NSAF showed low
reproducibility greater dispersion of data and very high CV value
especially in the quantification of the lowest abundant proteins. This
is likely due to the absence of any adjustment factor in this method
making NSAF very susceptible to quantitative biases caused by the
absence of data. The Rsc method showed low reproducibility in the
quantification of lowest abundance proteins in the quantitative
analysis of E. Coli proteome but low coefficient of variation in the
quantitation of spiked proteins suggesting again that a fixed
adjustment factor "f" might affect the analysis of a large set of data.
Otherwise, all quantitative methods were characterized by a wide
linearity concentration-range when changeable spiked proteins were
quantified, proving a response directly proportional to the protein
amount in a wide concentration range.
When CBNs methods were tested on biological complex samples (protein
extracts from the brain of HD mice vs WT), both methods generated
reproducible, accurate, sensible, and reliable quantitative data. Their
FC trends were always confirmed by western blot and/or MRM validations,
suggesting that a complexity-based adjustment factor properly works in
the correction of output data, leading to suitable final quantitative
measurements of protein expression variations.
However, the two CBN methods performed differently in the Fold Change
calculation. Although CBN(P) recognized a lower number of changeable
significant proteins, it has proven to be the most reliable method to
appreciate minimal statistically significant changes in protein
expression levels, since its FC cutoff was the narrowest among those
calculated for other methods.
Moreover, when the FC values calculated by all methods for selected
proteins were compared with those calculated by western blot and/or
MRM, a surprisingly perfect accordance was found solely with FCs
derived from CBN(P), suggesting that this method performs as the
highest reliable and accurate among all. CBN(S), in most cases,
performed more like Rsc and NSAF than CBN(P).
Nevertheless, CBN(S) method showed a higher sensitivity in the
detection of statistically relevant FC of low abundant proteins that
give rise to a low number of SpCs, such as UQCRQ and hnRNP H. These
proteins were statistically significant solely for CBN(S), but their
FCs were confirmed by western blot assays with a good accordance.
Differences in performances between CBN(P) and CBN(S) might be due to
the dimension of the adjustment factor, which is not negligible,
considering that the difference between total SpCs and total proteins
is about two orders of magnitude (in mice analysis, total SpCs =
2.45x10^5 and P = 2860).
Moreover, over-representation analysis of CBNs data carried out by
STRING reveals that the most of proteins flows in molecular pathways
affected in HD brain tissue and other neurodegenerative model systems,
thus confirming the robustness and the biological coherence of data
provided by these new methods. Indeed, processes such as energy
metabolism [[139]5,[140]48–[141]51], oxidative phosphorylation and
mitochondria functionality [[142]52–[143]54], calcium homeostasis
[[144]55,[145]56], already known to be deregulated in cortical samples
from HD mice, were identified, thus indicating the highest correlation
level of functional data.
In conclusion, in this study, we demonstrated how all SpCs
normalization methods are strongly affected by the presence/absence or
by the value of adjustment factors “f”. The presence of a correction
factor allows overcoming the effect of the absence of data in FC
calculation and leads to methods with the lowest coefficient of
variation. Moreover, the association of the "f" with sample complexity
makes the operator free from the choice of the best value for the
correction factor.
Our data have shown that CBN(S) and CBN(P) are both a viable
alternative to other existing SpC-based quantification methods.
Furthermore, both CBN(S) and CBN(P) are two sensitive methods, although
each shows a different reactivity: CBN(P) is capable of appreciating
small but statistically and biologically significant FC variations in
proteins well represented in the proteome; on the contrary, the CBN(S)
enhances the differences in levels of protein expression in lower
abundant proteins.
Finally, if someone would ask which methods we prefer, our choice falls
on CBN(P), because it performs as the most reliable and sensitive
“sensor” for FCs. In fact, from our point of view, it is preferable to
work on a lower number of data which, however, are the result of more
stringent and reliable selection standards, rather than getting lost in
a sea of data often difficult to interpret from a biological point of
view.
Supporting information
S1 Fig. Median values for the best fitting slopes calculated for each
pair of technical replicates including mixture B for all samples
analyzed with each normalization method.
(PDF)
[146]Click here for additional data file.^ (134.1KB, pdf)
S2 Fig. Images of the SDS-PAGE of mice samples.
Three biological replicates cortices of zQ175 and WT mice were loaded
and gel slices were cut following the scheme reported.
(PDF)
[147]Click here for additional data file.^ (181.3KB, pdf)
S3 Fig. Images of the entire membranes whose inserts are reported in
[148]Fig 5A.
Full-length western blots. Technical replicates of HTT, IRGM1, Homer1,
OSBPL2, hnRNP H, UQCRQ and Samm50 in zQ175 and WT mice were developed
with same antibodies. Technical replicates for SerpinB6 were developed
with two different antibodies, as reported below the images.
(PDF)
[149]Click here for additional data file.^ (697.3KB, pdf)
S1 Table. Details of the qualitative and quantitative mass
spectrometric analysis of standard mixtures spiked in standard E. Coli
proteome.
In sheet 1, raw data generated by MaxQuant following the bioinformatic
analysis for protein identification and quantification are reported.
The processed protein groups applying each SpCs method normalization
formulas are disclosed in sheets 2, 3, 4, and 5.
(XLS)
[150]Click here for additional data file.^ (1.2MB, xls)
S2 Table. Linear regression best-fit value referred to MixA.
Included is Slope, Y-intercept, and X-intercept calculated by NSAF,
RSC, CBN(P), and CBN(S) methods.
(XLS)
[151]Click here for additional data file.^ (25KB, xls)
S3 Table. Fold Change value (FC) and expected Fold Change (exp FC)
value by NSAF, RSC, CBN(P) and CBN(S) methods.
Included is Fold Change value (FC) and expected Fold Change (exp FC)
value of standard proteins (HBA, HBB, BSA, ADH, PYG, and ENO) measured
and/or calculated in B/A, C/A, D/A, E/A, C/B, D/B, E/B, D/C, E/C and
E/C mixture by using NSAF, RSC, CBN(P), CBN(S).
(XLS)
[152]Click here for additional data file.^ (44.5KB, xls)
S4 Table. Details of the qualitative and quantitative mass
spectrometric analysis of zQ175 and Wild-Type mouse HD model.
In sheet 1, raw data generated by MaxQuant following the bioinformatic
analysis for protein identification and quantification are reported.
The processed protein groups applying each SpCs method normalization
formulas are disclosed in sheets 2, 3, 4, and 5.
(XLS)
[153]Click here for additional data file.^ (22.6MB, xls)
S5 Table. Statistically significant identified proteins.
In the table are reported the FCs and p-values of statistically
significant identified proteins (p value<0.05) and FDR (<5%). Protein
names, gene names, Uniprot codes and peptides are also indicated as
headers.
(XLSX)
[154]Click here for additional data file.^ (58.6KB, xlsx)
S6 Table. Summary of MRM results.
In the table are reported the FCs values of proteins validated by
Multiple Reaction Monitoring (MRM). Run, protein names, peptide
sequences, Collision Energy, m/z of precursor ions, m/z of fragment
ions, total transition area, peptide FC, and averaged and normalized FC
for each protein.
(XLSX)
[155]Click here for additional data file.^ (37.5KB, xlsx)
Acknowledgments