Expert System for Computer-assisted Annotation of MS/MS Spectra

An important step in mass spectrometry (MS)-based proteomics is the identification of peptides by their fragment spectra. Regardless of the identification score achieved, almost all tandem-MS (MS/MS) spectra contain remaining peaks that are not assigned by the search engine. These peaks may be explainable by human experts but the scale of modern proteomics experiments makes this impractical. In computer science, Expert Systems are a mature technology to implement a list of rules generated by interviews with practitioners. We here develop such an Expert System, making use of literature knowledge as well as a large body of high mass accuracy and pure fragmentation spectra. Interestingly, we find that even with high mass accuracy data, rule sets can quickly become too complex, leading to over-annotation. Therefore we establish a rigorous false discovery rate, calculated by random insertion of peaks from a large collection of other MS/MS spectra, and use it to develop an optimized knowledge base. This rule set correctly annotates almost all peaks of medium or high abundance. For high resolution HCD data, median intensity coverage of fragment peaks in MS/MS spectra increases from 58% by search engine annotation alone to 86%. The resulting annotation performance surpasses a human expert, especially on complex spectra such as those of larger phosphorylated peptides. Our system is also applicable to high resolution collision-induced dissociation data. It is available both as a part of MaxQuant and via a webserver that only requires an MS/MS spectrum and the corresponding peptides sequence, and which outputs publication quality, annotated MS/MS spectra (www.biochem.mpg.de/mann/tools/). It provides expert knowledge to beginners in the field of MS-based proteomics and helps advanced users to focus on unusual and possibly novel types of fragment ions.In MS-based proteomics, peptides are matched to peptide sequences in databases using search engines (1–3). Statistical criteria are established for accepted versus rejected peptide spectra matches based on the search engine score, and usually a 99% certainty is required for reported peptides. The search engines typically only take sequence specific backbone fragmentation into account (i.e. a, b, and y ions) and some of their neutral losses. However, tandem mass spectra—especially of larger peptides—can be quite complex and contain a number of medium or even high abundance peptide fragments that are not annotated by the search engine result. This can result in uncertainty for the user—especially if only relatively few peaks are annotated—because it may reflect an incorrect identification. However, the most common cause of unlabeled peaks is that another peptide was present in the precursor selection window and was cofragmented. This has variously been termed “chimeric spectra” (4–6), or the problem of low precursor ion fraction (PIF)¹ (7). Such spectra may still be identifiable with high confidence. The Andromeda search engine in MaxQuant, for instance, attempts to identify a second peptide in such cases (8, 9). However, even “pure” spectra (those with a high PIF) often still contain many unassigned peaks. These can be caused by different fragment types, such as internal ions, single or combined neutral losses as well as immonium and other ion types in the low mass region. A mass spectrometric expert can assign many or all of these peaks, based on expert knowledge of fragmentation and manual calculation of fragment masses, resulting in a higher degree of confidence for the identification. However, there are more and more practitioners of proteomics without in depth training or experience in annotating MS/MS spectra and such annotation would in any case be prohibitive for hundreds of thousands of spectra. Furthermore, even human experts may wrongly annotate a given peak—especially with low mass accuracy tandem mass spectra—or fail to consider every possibility that could have resulted in this fragment mass.Given the desirability of annotating fragment peaks to the highest degree possible, we turned to “Expert Systems,” a well-established technology in computer science. Expert Systems achieved prominence in the 1970s and 1980s and were meant to solve complex problems by reasoning about knowledge (10, 11). Interestingly, one of the first examples was developed by Nobel Prize winner Joshua Lederberg more than 40 years ago, and dealt with the interpretation of mass spectrometric data. The program''s name was Heuristic DENTRAL (12), and it was capable of interpreting the mass spectra of aliphatic ethers and their fragments. The hypotheses produced by the program described molecular structures that are plausible explanations of the data. To infer these explanations from the data, the program incorporated a theory of chemical stability that provided limiting constraints as well as heuristic rules.In general, the aim of an Expert System is to encode knowledge extracted from professionals in the field in question. This then powers a rule-based system that can be applied broadly and in an automated manner. A rule-based Expert System represents the information obtained from human specialists in the form of IF-THEN rules. These are used to perform operations on input data to reach appropriate conclusion. A generic Expert System is essentially a computer program that provides a framework for performing a large number of inferences in a predictable way, using forward or backward chains, backtracking, and other mechanisms (13). Therefore, in contrast to statistics based learning, the “expert program” does not know what it knows through the raw volume of facts in the computer''s memory. Instead, like a human expert, it relies on a reasoning-like process of applying an empirically derived set of rules to the data.Here we implemented an Expert System for the interpretation for high mass accuracy tandem mass spectrometry data of peptides. It was developed in an iterative manner together with human experts on peptide fragmentation, using the published literature on fragmentation pathways as well as large data sets of higher-energy collisional dissociation (HCD) (14) and collision-induced dissociation (CID) based peptide identifications. Our goal was to achieve an annotation performance similar or better than experienced mass spectrometrists (15), thus making comprehensively annotated peptide spectra available in large scale proteomics.