Transcript Improving the Reliability of Peptide Identification by

Improving the Reliability of
Peptide Identification by
Tandem Mass Spectrometry
Nathan Edwards
Department of Biochemistry and Molecular & Cellular Biology
Georgetown University Medical Center
Mass Spectrometry for
Proteomics
• Measure mass of many (bio)molecules
simultaneously
• High bandwidth
• Mass is an intrinsic property of all
(bio)molecules
• No prior knowledge required
2
Mass Spectrometer
Sample
+
_
Ionizer
• MALDI
• Electro-Spray
Ionization (ESI)
Mass Analyzer
• Time-Of-Flight (TOF)
• Quadrapole
• Ion-Trap
3
Detector
• Electron
Multiplier
(EM)
High Bandwidth
% Intensity
100
0
250
500
750
4
1000
m/z
Mass is fundamental!
5
Mass Spectrometry for
Proteomics
• Measure mass of many molecules
simultaneously
• ...but not too many, abundance bias
• Mass is an intrinsic property of all
(bio)molecules
• ...but need a reference to compare to
6
Mass Spectrometry for
Proteomics
• Mass spectrometry has been around
since the turn of the century...
• ...why is MS based Proteomics so new?
• Ionization methods
• MALDI, Electrospray
• Protein chemistry & automation
• Chromatography, Gels, Computers
• Protein sequence databases
• A reference for comparison
7
Sample Preparation for
Peptide Identification
Enzymatic Digest
and
Fractionation
8
Single Stage MS
MS
m/z
9
Tandem Mass Spectrometry
(MS/MS)
m/z
Precursor selection
m/z
10
Tandem Mass Spectrometry
(MS/MS)
Precursor selection +
collision induced dissociation
(CID)
m/z
MS/MS
m/z
11
The big picture...
• MS/MS spectra provide evidence for the
amino-acid sequence of functional
proteins.
• Key concepts:
•
•
•
•
Spectrum acquisition is unbiased
Direct observation of amino-acid sequence
Sensitive to minor sequence variation
Observed peptides represent folded proteins
12
Peptide Identification
• For each (likely) peptide sequence
1. Compute fragment masses
2. Compare with spectrum
3. Retain those that match well
• Peptide sequences from protein sequence
databases
• Swiss-Prot, IPI, NCBI’s nr, ...
• Automated, high-throughput peptide identification
in complex mixtures
13
Peptide Identification, but...
• What about novel peptides?
• Search compressed ESTs (C3, PepSeqDB)
• What about peak intensity?
• Spectral matching using HMMs (HMMatch)
• Which identifications are correct?
• Unsupervised, model-free, result combiner
with false discovery rate estimation
14
Why don’t we see more
novel peptides?
• Tandem mass spectrometry doesn’t
discriminate against novel peptides...
...but protein sequence databases do!
• Searching traditional protein sequence
databases biases the results towards
well-understood protein isoforms!
15
What goes missing?
• Known coding SNPs
• Novel coding mutations
• Alternative splicing isoforms
• Alternative translation start-sites
• Microexons
• Alternative translation frames
16
Why should we care?
• Alternative splicing is the norm!
• Only 20-25K human genes
• Each gene makes many proteins
• Proteins have clinical implications
• Biomarker discovery
• Evidence for SNPs and alternative splicing
stops with transcription
• Genomic assays, ESTs, mRNA sequence.
• Little hard evidence for translation start site
17
Novel Splice Isoform
• Human Jurkat leukemia cell-line
• Lipid-raft extraction protocol, targeting T cells
• von Haller, et al. MCP 2003.
• LIME1 gene:
• LCK interacting transmembrane adaptor 1
• LCK gene:
• Leukocyte-specific protein tyrosine kinase
• Proto-oncogene
• Chromosomal aberration involving LCK in leukemias.
• Multiple significant peptide identifications
18
Novel Splice Isoform
19
Novel Splice Isoform
20
Novel Mutation
• HUPO Plasma Proteome Project
• Pooled samples from 10 male & 10 female
healthy Chinese subjects
• Plasma/EDTA sample protocol
• Li, et al. Proteomics 2005. (Lab 29)
• TTR gene
• Transthyretin (pre-albumin)
• Defects in TTR are a cause of amyloidosis.
• Familial amyloidotic polyneuropathy
• late-onset, dominant inheritance
21
Novel Mutation
Ala2→Pro associated with familial amyloid polyneuropathy
22
Novel Mutation
23
Searching ESTs
• Proposed long ago:
• Yates, Eng, and McCormack; Anal Chem, ’95.
• Now:
• Protein sequences are sufficient for protein identification
• Computationally expensive/infeasible
• Difficult to interpret
• Make EST searching feasible for routine searching
to discover novel peptides.
24
Searching Expressed
Sequence Tags (ESTs)
Pros
• No introns!
• Primary splicing
evidence for
annotation pipelines
• Evidence for dbSNP
• Often derived from
clinical cancer
samples
Cons
• No frame
• Large (8Gb)
• “Untrusted” by
annotation pipelines
• Highly redundant
• Nucleotide error
rate ~ 1%
25
Compressed EST Peptide
Sequence Database
• For all ESTs mapped to a UniGene gene:
•
•
•
•
Six-frame translation
Eliminate ORFs < 30 amino-acids
Eliminate amino-acid 30-mers observed once
Compress to C2 FASTA database
• Complete, Correct for amino-acid 30-mers
• Gene-centric peptide sequence database:
• Size: < 3% of naïve enumeration, 20774 FASTA entries
• Running time: ~ 1% of naïve enumeration search
• E-values: ~ 2% of naïve enumeration search results
26
PepSeq FASTA Databases
• Organisms:
• HUMAN, MOUSE, RAT, ZEBRA FISH
• Peptide Evidence:
• Genbank mRNA, EST, HTC
• RefSeq mRNA, Proteins
• Swiss-Prot/TrEMBL, EMBL, VEGA, H-Inv, IPI
Proteins
• Swiss-Prot variants
• Swiss-Prot signal peptide & init. Met removal
• Singe FASTA entry per Gene
27
Spectral Matching for
Peptide Identification
• Detection vs. identification
• Increased sensitivity & specificity
• No novel peptides!
• NIST GC/MS Spectral Library
•
•
•
•
•
Identifies small molecules,
100,000’s of (consensus) spectra
Bundled/Sold with many instruments
“Dot-product” spectral comparison
Current project: Peptide MS/MS
28
NIST MS Search: Peptides
29
Peptide DLATVYVDVLK
30
Protein Families
31
Protein Families
32
Peptide DLATVYVDVLK
33
Hidden Markov Models for
Spectral Matching
• Capture statistical variation and consensus in
peak intensity
• Only need 10 spectra to build a model
• Capture semantics of peaks
• Extrapolate model to other peptides
• Good specificity with superior sensitivity for
peptide detection
• Assign 1000’s of additional spectra (p-value < 10-5)
34
Hidden Markov Model
Delete
Insert
Ion
(m/z,int) pair emitted by ion & insert states
35
The devil in the details
• Intensity normalization
• Discretize (m/z,int) pairs
• Viterbi distance as score
• Compute p-value using “random”
spectra
36
Random Spectra
# of spectra
• Uniform sample of (m/z,int)
• Permutation (m/z) of true spectra peaks
• M/z distribution between true spectra and
uniform sample (parameter)
True
False
Random
Viterbi Score
37
HMM Peptide Identification
Results – DLATV
DLAT (viterbi)
True_test(0.0001)
True_test(other)
False_test(0.0001)
False_test(other)
240
220
200
160
140
120
100
80
DLAT (-logP)
60
True_test(0.0001)
40
20
True_test(other)
False_test(0.0001)
False_test(other)
100
0
0
0
28
26
029
0
027
0
24
025
0
22
80
70
# of spectra
Viterbi Distance
023
0
20
021
0
18
019
0
16
017
0
14
015
0
013
12
011
-9
0
10
80
-5
0
-7
0
60
40
-3
0
90
20
010
# of spectra
180
60
50
40
30
20
10
0
0-1 1-2 2-3 3-4 4-5 5-6 6-7 7-8 8-9
9- 10- 11- 12- 13- 14- 15- 16- 17- 18- inf
10 11 12 13 14 15 16 17 18 19
-log(p-value)
38
Spectral Matching of
Peptide Variants
DFLAGGIAAAISK
DFLAGGVAAAISK
39
HMM model extrapolation
40
Mascot Search Results
41
Peptide Identification Results
• Search engines always provide an
answer
• Current search engines:
• Hard to determine “good” scores
• Significance estimates are unreliable
• Need better methods!
42
Common Algorithmic
Framework
• Pre-process experimental spectra
• Filter peptide candidates
• Score match between peptides and
spectra
• Rank peptides and assign
43
Comparison of search engines
• No single score is
comprehensive
Mascot
OMSSA
10%
• Search engines
disagree
4%
2%
69%
• Many spectra lack
confident peptide
assignment
9%
5%
2%
44
X!Tandem
Lots of published solutions!
• Treat search engines as black-boxes
• Apply supervised machine learning to
results
• Use multiple match metrics
• Combine/refine using multiple search
engines
• Agreement suggests correctness
• Use empirical significance estimates
• “Decoy” databases (FDR)
45
PepArML
• Peptide identification arbiter by
machine learning
• Unifies these ideas within a modelfree, combining machine learning
framework
• Unsupervised training procedure
46
PepArML Overview
• Unify Tandem, Mascot, and OMSSA results
X!Tandem
Mascot
PepArML
OMSSA
Other
47
Identified
Unidentified
Voting Heuristic Combiner
• Choose peptide ID with the most votes
• Use best FDR as confidence
• Break ties (single votes) using FDR
• Strawman for comparison
48
Dataset construction
Extract Features
Search Tools
Spectra
compare
Matched Ions
X!Tandem
Peak_intensity
Mass delta
Mascot
# of missed
cleavages
x
Peptide length
OMSSA
Tandem Score
Mascot Score
Other
OMSSA Score
49
Machine
Learning
Dataset construction
• Build feature vectors
Tandem
Mascot
OMSSA
( S1 , P1 )
T
(S1, P2 )
F
(S2 , P1 )
T
……
( Sn , Pm )
T
50
Dataset construction
• Synthetic protein mixtures provide ground truth
• C8
• 8 standard proteins (Calibrant Biosystems)
• 4594 MS/MS spectra (LTQ)
• 618 (11.2%) true positives
• S17
• 17 standard proteins (Sashimi Repository)
• 1389 MS/MS spectra (Q-TOF)
• 354 (25.4%) true positives
• AURUM
• 364 standard proteins (AURUM 1.0)
• 7508 MS/MS spectra (MALDI-TOF-TOF)
• 3775 (50.3%) true positives
51
Machine learning improves
single search engines (S17)
52
Multiple search engines are better
than single search engines (S17)
53
Feature Evaluation
54
Application to Real Data
• How well do these models generalize?
• Different instruments
• Spectral characteristics change scores
• Search parameters
• Different parameters change score values
• Supervised learning requires
• (Synthetic) experimental data from every instrument
• Search results from available search engines
• Training/models for all
parameters x search engine sets x instruments
55
Model Generalization
56
Rescuing Machine Learning
• Train a new machine-learning model for
every dataset!
• Generalization not required
• No predetermined search engines, parameters,
instruments, features
• Perhaps we can “guess” the true proteins
• Most proteins not in doubt
• Machine learning can tolerate imperfect labels
57
Unsupervised Learning
• Heuristic selection of “true” proteins
• Train classifier, predict true peptide IDs
• Update “true” proteins
• Heuristic selection of “true” proteins from
classifier predictions
• Iterate until convergence
58
Unsupervised Learning
Performance
59
Unsupervised Learning
Convergence
60
Conclusions
• Proteomics can inform genome annotation
• Eukaryotic and prokaryotic
• Functional vs silencing variants
• Peptides identify more than just proteins
• Untapped source of disease biomarkers
• Computational inference can make a
substantial impact in proteomics
61
Conclusions
• Compressed peptide sequence
databases make routine EST
searching feasible
• HMMatch spectral matching improves
identification performance for familiar
peptides
• Unsupervised, model-free, combining
PepArML framework solves peptide
identification interpretation problem
62
Acknowledgements
• Chau-Wen Tseng, Xue Wu
• UMCP Computer Science
• Catherine Fenselau
• UMCP Biochemistry
• Cheng Lee
• Calibrant Biosystems
• PeptideAtlas, HUPO PPP, X!Tandem
• Funding: NIH/NCI, USDA/ARS
63