The first thing I should say is that none of the material presented is original research done at Proteome Software Interpreting MS/MS Proteomics Results but we do strive to make the tools presented here available in our software product Scaffold. With that caveat aside… Brian C. Searle Proteome Software Inc. Portland, Oregon USA Brian.Searle@ProteomeSoftware.com NPC Progress Meeting (February 2nd, 2006) Illustrated by Toni Boudreault This is an foremost an introduction so we’re first going to talk about Organization SEQUEST Identify how you go about identifying proteins with tandem mass spectrometry in the first place Then we’re going to talk about the motivations behind the development of the first really useful bioinformatics technique in our field, SEQUEST. This technique has been extended by two other tools called X! Tandem and Mascot. X! Tandem/Mascot We’re also going to talk about how these programs differ Differ Combine and how we can use that to our advantage by considering them simultaneously using probabilities. A Start with a protein A I K H Q G K A L T N V I T I D V P So, this is proteomics, so we’re going to use tandem mass spectrometry to identify proteins-- hopefully many of them, and hopefully very quickly. L K E D C G R T A I R A Cut with an enzyme A I K H Q G K E A T And to use this technique you generally have to lyse the protein into peptides about 8 to 20 amino acids in length and… L L K P N V I T I D V D C G R T A I R A Select a peptide A I K H Q G K E A P T I L Look at each peptide individually. L K R N V I T I D V D C G R T A We select the peptide by mass using the first half of the tandem mass spectrometer Impart energy in collision cell A E P T I The mass spectrometer imparts energy into the peptide causing it to fragment at the peptide bonds between amino acids. R H2O Intensity Measure mass of daughter ions A E P A E P A E T 399.2 A 298.1 201.1 72.0 M/z The masses of these fragment ions is recorded using the second mass spectrometer. These ions are commonly called B ions, based on nomenclature you don’t really want to know about… Intensity B-type Ions A E P T I 72.0 129.0 97.0 101.0 113.1 M/z But the mass difference between the peaks corresponds directly to the amino acid sequence. R 174.1 H2O Intensity B-type Ions A E P T I 72.0 129.0 97.0 101.0 113.1 174.1 A-0 AE-A AEP -AE AEPT -AEP AEPTI -AEPT AEPTIR -AEPTI For example, the A-E peak minus the A peak should produce the mass of E. R H2O You can build these mass differences up and derive a sequence for the original peptide This is pretty neat and it makes tandem mass spectrometry one of the best tools out there for sequencing novel peptides. M/z But there are a couple confounding factors. So, it seems pretty easy, doesn’t it? For example… B ions have a tendency to degrade and lose carbon monoxide producing… B-type Ions Intensity A E CO P CO T CO M/z I CO R CO H2O CO A ions. A-type Ions A E P T I R H2O Furthermore… CO CO CO M/z CO CO CO … The second half are represented as Y ions that sequence backwards. And, unfortunately, this is the real world, so… R I T Intensity H2O Y-type Ions M/z P E A … All the peaks have Y-type Ions R I T Intensity H2O M/z P different measured heights and many peaks can often be missing. E A All these peaks are seen together simultaneously and we don’t … even know B-type, A-type, Y-type Ions R I T Intensity H2O M/z P E A What type of ion they are, making the mass differences approach even more difficult. Intensity Finally, as with all analytical techniques, M/z There’s noise, Intensity producing a final spectrum that looks like… M/z Intensity ….This, on a good day. M/z And so it’s actually fairly difficult to… … compute the mass differences to Intensity sequence the peptide, certainly in a computer automated way. A E P T I 72.0 129.0 97.0 101.0 113.1 M/z R 174.1 H2O So the community needed a new technique. Now, it wasn’t all without hope… Known Ion Types We knew a couple of things about peptide fragmentation. Not only do we know to expect B, A, and Y ions, but… B-type ions A-type ions Y-type ions Known Ion Types … We also B-type ions A-type ions Y-type ions know a couple of other variations on those ions that come up. We even know something about the… B- or Y-type +2H ions B- or Y-type -NH3 ions B- or Y-type -H2O ions … likelihood of seeing each type of ion, Known Ion Types B-type ions • 100% A-type ions • 20% Y-type ions • 100% where generally B and Y ions are most prominent. B- or Y-type +2H ions • 50% B- or Y-type -NH3 ions • 20% B- or Y-type -H2O ions • 20% So it’s actually pretty easy to guess what a spectrum should look like If we know the amino acid sequence of a peptide, if we know what the peptide sequence is. we can guess what the spectra should look like! Model Spectrum So as an example, consider the peptide ELVIS LIVES K that was synthesized by Rich Johnson in Seattle ELVISLIVESK *Courtesy of Dr. Richard Johnson http://www.hairyfatguy.com/ Model Spectrum We can create a hypothetical spectrum based on our rules B/Y type ions (100%) Where B and Y ions are estimated at 100%, plus 2 ions are estimated at 50%, B/Y +2H type ions (50%) and other stragglers are at 20%. A type ions B/Y -NH3/-H2O (20%) Model Spectrum So if we consider the spectrum that was derived from the ELVIS LIVES K peptide… Model Spectrum We can find where the overlap is between the hypothetical and the actual spectra… Model Spectrum And say conclusively based on the evidence that the spectrum does belong to the ELVIS LIVES K peptide. But who cares? The more important question is “what about situations where we don’t know the sequence?” We guess! And so this was an approach followed by a program called PepSeq PepSeq … AAAAAAAAAA AAAAAAAAAC AAAAAAAACC AAAAAAACCC which would guess every combination of amino acids possible build a hypothetical spectrum, and find the best matching hypothetical. … ELVISLIVESK WYYYYYYYYY YYYYYYYYYY J. Rozenski et al., Org. Mass Spectrom., 29 (1994) 654-658. PepSeq This was a start, but it’s clearly impossibly hard with larger peptides and there’s a lot of room to overfit the data. • Impossibly hard after 7 or 8 amino acids! • High false positive rate because you consider so many options PepSeq So obviously this isn’t going to work in the long run. Another strategy is needed! • Impossibly hard after 7 or 8 amino acids! • High false positive rate because you consider so many options Sequencing Explosion We needed a new invention to come around and that was shotgun Sanger-sequencing … • 1977 Shotgun sequencing invented, bacteriophage fX174 sequenced. • • • • • • In 89 and 90 the Yeast and Human Genome projects were announced 1989 Yeast Genome project announced 1990 Human Genome project announced 1992 First chromosome (Yeast) sequenced 1995 H. influenza sequenced 1996 Yeast Genome sequenced 2000 Human Genome draft followed by the first chromosome in 92 et cetra, et cetra Sequencing Explosion • 1977 Shotgun sequencing invented, bacteriophage fX174 sequenced. … Eng, J. K.; McCormack, A. L.; Yates, J. R. III J. Am. Soc. Mass Spectrom. 1994, 5, 976-989. • • • • • • 1989 Yeast Genome project announced 1990 Human Genome project announced 1992 First chromosome (Yeast) sequenced exploit genome sequencing 1995 H. influenza sequenced 1996 Yeast Genome sequenced 2000 Human Genome draft In 1994 Jimmy Eng and John Yates published a technique to for use in tandem mass spectrometry. And the idea was … SEQUEST .…instead of searching all possible peptide sequences, search only those in genome databases. Now, in the postgenomic world this seems like a pretty trivial idea, but back then there was a lot of assumption placed on the idea that we’d actually have a complete Human genome in a reasonable amount of time. SEQUEST 2*1014 2*1010 1*108 4*106 -- All possible 11mers (ELVISLIVESK) -- All possible peptides in NR -- All tryptic peptides in NR -- All Human tryptic peptides in NR So, In terms of 11amino acid peptides we’re talking about a 10 thousand fold difference between searching every possible 11mer those in the current non-redundant protein database from the NCBI And a 100 million fold difference for searching human trypic peptides So that was huge, it made hypothetical spectrum matching feasible. Instead of trying to make a better model, SEQUEST made a couple of other interesting improvements as well they decided just to make the actual spectrum look like the model with normalization… Jimmy and John noted that there was a discontinuity between the intensities of the hypothetical spectrum and the actual spectrum. SEQUEST Model Spectrum For a scoring function they decided to use Cross-Correlation, Like so. which basically sums the peaks that overlap between hypothetical and the actual spectra SEQUEST Model Spectrum And then they shifted the spectra back and …. SEQUEST Model Spectrum … Forth so that the peaks shouldn’t align. They used this number, also called the Auto-Correlation, as their background. SEQUEST Model Spectrum SEQUEST XCorr This is another representation of the Cross Correlation and the Auto Correlation. Correlation Score Cross Correlation (direct comparison) Auto Correlation (background) Offset (AMU) Gentzel M. et al Proteomics 3 (2003) 1597-1610 The XCorr score is the Cross Correlation divided by the average of the auto correlation over a 150 AMU range. SEQUEST XCorr Correlation Score The XCorr is high if the direct comparison is significantly greater than the background, which is obviously good for peptide identification. Cross Correlation (direct comparison) Auto Correlation (background) Offset (AMU) CrossCorr XCorr = avg AutoCorr offset=-75 to 75 Gentzel M. et al Proteomics 3 (2003) 1597-1610 And this XCorr is actually a pretty robust method for estimating how accurate the match is, and so far, there really haven’t been any significant improvements on it. SEQUEST DeltaCn XCorr 1 XCorr 2 XCorr 1 The DeltaCn is another score that scientists often use. It measures how good the XCorr is relative to the next best match. As you can see, this is actually a pretty crude calculation. Here’s another representation of that sentiment. The XCorr is a strong measure of accuracy, whereas the DeltaCn is a weak measure of relative goodness. . SEQUEST Accuracy Score Strong (XCorr) Relative Score Weak (DeltaCn) Obviously, there could be an alternative method that focuses more on the success of the relative score. Mascot and X! Tandem fit that bill. Alternate SEQUEST Method Accuracy Score Relative Score Strong (XCorr) Weak (DeltaCn) Weak Strong X! Tandem Scoring by-Score= Sum of intensities of peaks matching B-type or Y-type ions HyperScore= Now the X! Tandem accuracy score is rather crude. by-Score N ! N ! y b It only considers B and Y ions and and attaches these factorial terms with an admittedly hand waving argument. Fenyo, D.; Beavis, R. C. Anal. Chem., 75 (2003) 768-774 Distribution of “Incorrect” Hits But instead of just considering the best match to the second best, it looks at the distribution of lower scoring hits, assuming that they are all wrong. # of Matches This is somewhat based on ideas pioneered with the BLAST algorithm. Here, every bar represents the number of matches at a given score. The X! Tandem creators found that the distribution decays (or slopes down) exponentially… Second Best Hyper Score Best Hit Estimate Likelihood (E-Value) …and the log of the distribution is relatively Log(# of Matches) linear because of the exponential decay. Best Hit Hyper Score Estimate Likelihood (E-Value) Log(# of Matches) Hyper Score Expected Number Of Random Matches Best Hit If the distribution represents the number of random matches at any given score, the linear fit should correspond to the expected number of random matches. Estimate Likelihood (E-Value) Log(# of Matches) Score of 60 has 1/10 chance of occurring at random Best Hit And from this, you can calculate the likelihood that the best match is random. In this case, a score of 60 corresponds with a log number of matches being -1 which means the estimated number of random matches for that score is 0.1 This is called an E-Value, or Expected-Value. X! Tandem and Mascot Now, X! Tandem calculates this E-Value empirically. Likelihood that match E-Value= is incorrect relative to N guesses Empirical (X! Tandem) Likelihood that match P-Value= is incorrect (E~P·N) Theoretical (Mascot) Another search engine, Mascot, tries to get at the same kind of number using theoretical calculations, most likely based on the number of identified peaks and the likelihood of finding certain amino acids in the genome database. They’ve never explicitly published their algorithm, so we’ll never really know, but I suspect it’s something smart. I just want to bring up a point that we’ll touch on a little later… …the E-Value that X! Tandem calculates and the P-Value that Mascot calculates are probabilistically based, but they can only estimate the likelihood that the match is wrong. X! Tandem and Mascot Likelihood that match E-Value= is incorrect relative to N guesses Empirical (X! Tandem) Likelihood that match is incorrect (E~P·N) Theoretical (Mascot) P-Value= Probability= Likelihood that Note match is correct (Probability≠1-P)! This is realistically not nearly as useful as knowing the probability that a peptide identification is right, which is NOT 1 minus the P-Value. Now, let’s go back and fill in the X! Tandem part of our accuracy/relativity scoring grid. SEQUEST Relative Score XCorr DeltaCn X! Tandem Accuracy Score HyperScore E-Value To reiterate, the XCorr is an excellent measure of accuracy… SEQUEST Relative Score XCorr DeltaCn X! Tandem Accuracy Score HyperScore E-Value …whereas the E-Value is an excellent measure of how good the best score is relative to the rest. If we assume that accuracy and relativity scores are independent measures of goodness, could we use both the SEQUEST’s XCorr and X! Tandem’s E-Value together? SEQUEST Relative Score XCorr DeltaCn X! Tandem Accuracy Score HyperScore E-Value And the answer is a resounding yes. Each point on this graph is a spectrum, where correct identifications are marked in red, while incorrect identifications are marked in blue. We know what’s correct and incorrect because this is a control sample. X! Tandem: -log(E-Value) 10 Protein Control Sample SEQUEST: Discriminant Score Although in general the spectra SEQUEST scores well are spectra X!Tandem also scores well, there is considerable scatter between the search engines. One might wonder if X! Tandem and Mascot use similar scoring approaches, would they benefit as much, but the answer is surprisingly still yes! X! Tandem: -log(E-Value) 10 Protein Control Sample Mascot: Ion-Identity Score Now, why are the scores so different? Why So Different? • Sequest – Considers relative intensities • X! Tandem – Considers semi-tryptic peptides – Considers only B/Y-type Ions Well, here are a couple of possible reasons. SEQUEST is the only method to consider relative intensities. • Mascot – Considers theoretical P-Value relative to search space Why So Different? • Sequest – Considers relative intensities • X! Tandem – Considers semi-tryptic peptides – Considers only B/Y-type Ions X! Tandem is the only method to consider peptides outside the standard search space by default, such as semi-tryptic peptides. However, it’s the only score that considers only B and Y ions, as opposed to a complete model. • Mascot – Considers theoretical P-Value relative to search space Why So Different? • Sequest – Considers relative intensities • X! Tandem – Considers semi-tryptic peptides – Considers only B/Y-type Ions And Mascot is the only search engine to compute a completely theoretical P-Value • Mascot – Considers theoretical P-Value relative to search space X! Tandem: -log(E-Value) So we clearly want to consider multiple search engines simultaneously, Consider Multiple Algorithms? but how? Mascot: Ion-Identity Score How To Compare Search Engines? – SEQUEST: XCorr>2.5, DeltaCn>0.1 – Mascot: Ion Score-Identity Score>0 – X! Tandem: E-Value<0.01 You can’t use a thresholding system because it’s impossible to find corresponding thresholds. For example, a SEQUEST match with an XCorr of 2.5 doesn’t mean the same thing as an X! Tandem match with an E-Value of 0.01. How To Compare Search Engines? – SEQUEST: XCorr>2.5, DeltaCn>0.1 – Mascot: Ion Score-Identity Score>0 – X! Tandem: E-Value<0.01 The simplest way would be to convert the scores into probabilities and compare those. Need to convert scores to probabilities! We advocate for Andrew Keller and Alexy Nesviskii’s Peptide Prophet approach because it actually calculates a true probability, not just a p-value. 10 Protein Control Sample (Q-ToF) X! Tandem approach # of Matches Other Incorrect IDs for Spectrum Possibly Correct? So if you remember, X! Tandem considers the best peptide match for a spectrum against a distribution of incorrect matches Mascot: Ion-Identity Score 10 Protein Control Sample (Q-ToF) Peptide Prophet approach # of Matches ALL Other “Best” Matches Well, Peptide Prophet looks across the entire sample, and not at just one spectrum at a time. It compares the best match against all of the other best matches in the sample, which is clearly bimodal. Possibly Correct? Mascot: Ion-Identity Score Keller, A. et al Anal. Chem. 74, 5383-5392 10 Protein Control Sample (Q-ToF) Peptide Prophet approach # of Matches ALL Other “Best” Matches The low mode represents matches that are most likely wrong while the high mode represents matches that are probably right. Possibly Correct? Mascot: Ion-Identity Score Keller, A. et al Anal. Chem. 74, 5383-5392 10 Protein Control Sample (Q-ToF) Peptide Prophet approach Peptide Prophet curve fits two distributions to the modes, # of Matches “Incorrect” Possibly Correct? following the assumption that the low scoring distribution is “Incorrect” and that the higher scoring distribution is “correct”. “Correct” Mascot: Ion-Identity Score 10 Protein Control Sample (Q-ToF) p(D | ) p() p( | D) p(D | ) p() p(D | ) p() # of Matches “Incorrect” These two distributions can be analyzed using Bayesian statistics with this formula. Now that formula looks pretty complex, Possibly Correct? but… “Correct” Mascot: Ion-Identity Score 10 Protein Control Sample (Q-ToF) p(D | ) p() p( | D) p(D | ) p() p(D | ) p() # of Matches “Incorrect” It just calculates the height of the correct distribution at a particular score, divided by the height of both distributions. “Correct” Mascot: Ion-Identity Score 10 Protein Control Sample (Q-ToF) This is essentially the probability of having that score and being correct divided by the probability of just having that score p(D | ) p() p( | D) p(D | ) p() p(D | ) p() “Incorrect” prob of having score and being correct prob of having score “Correct” Mascot: Ion-Identity Score # of Matches “Incorrect” Possibly Correct? “Correct” Mascot: Ion-Identity Score This is a neat method because it actually considers the likelihood of being correct, rather than X! Tandem and Mascot, which only calculate the probability of being incorrect. It’s because of this that Peptide Prophet can get produce a true probability, which is important when the sample characteristics change. # of Matches “Incorrect” Q-ToF: Possibly Correct? “Correct” Mascot: Ion-Identity Score For example, the control sample we’ve been looking at was derived from Q-ToF data which produces pretty high quality results # of Matches If you compare that to the same sample on run on an Ion Trap, the probability of being correct is greatly diminished. If you’ll note, the Incorrect distribution doesn’t change very much between the two analyses, however, the likelihood that the identification is right changes dramatically! “Incorrect” Q-ToF: Possibly Correct? “Correct” # of Matches Mascot: Ion-Identity Score “Incorrect” Ion Trap: Possibly Correct? “Correct” As Peptide Prophet considers the correct distribution, it is immune to fluctuations between samples. P-Values and E-Values don’t consider this information, so they can’t be compared across multiple samples, or different examinations of the same sample hence the reason why we need to use Peptide Prophet for comparing two different search engines # of Matches Mascot: Ion-Identity Score “Incorrect” Ion Trap: Possibly Correct? “Correct” So going back to the scatter plot between X! Tandem and Mascot, X! Tandem: -log(E-Value) Consider Multiple Algorithms? Mascot: Ion-Identity Score we can use Peptide Prophet to compute the score threshold that represents a 95% cut-off… Like so. X! Tandem: -log(E-Value) Consider Multiple Algorithms? Mascot: -2.5=95% X! Tandem: 2.6=95% Mascot: Ion-Identity Score This allows you to fairly consider the answers from both search engines simultaneously. The important thing to note, is that if you looked at a different sample, these thresholds should change depending on the height of the correct distributions Conclusion So in conclusion, • All search engines use different criteria, producing different scores • Using multiple search engines simultaneously yields better results • Peptide Prophet can normalize search engine results all of the search engines look at different criteria Conclusion • All search engines use different criteria, producing different scores • Using multiple search engines simultaneously yields better results • Peptide Prophet can normalize search engine results And we can leverage this to identify more peptides Conclusion • All search engines use different criteria, producing different scores • Using multiple search engines simultaneously yields better results • Peptide Prophet can normalize search engine results And that Peptide Prophet is a great mechanism for doing that because it calculates true probabilities, instead of p-values The End