Jennifer Colby (Presenter)
Bio: Jennifer Colby is an assistant professor in the department of Pathology, Microbiology, and Immunology and an associate director of clinical chemistry at Vanderbilt University Medical Center. Jennifer earned her BS in Molecular and Cellular Biology from Vanderbilt University in 2007 and her PhD in Molecular Toxicology from the University of California Berkeley in 2012. She completed her postdoctoral training in clinical chemistry and toxicology at the University of California San Francisco in 2015. She is a member of SOFT, CAT, AACC, ACLPS, AACT, and IATDMCT. Her research interests include developing new applications for mass spectrometry in toxicology and laboratory medicine. When not in the lab Jennifer runs, competes in dressage, and enjoys exploring Tennessee with her Bouvier des Flandres.
Authorship: Jennifer Colby (1), Jeffery Rivera (2), Lyle Burton (2), Dave Cox (2), and Kara Lynch (3)
(1) Vanderbilt University, Nashville, TN (2) SCIEX, Concord, ON Canada (3) University of California San Francisco, San Francisco, CA
Library searching encompasses the computerized identification of compounds based on the similarity between an acquired mass spectrum and a spectrum found in a database. Library searching using commercially available algorithms and spectral libraries is a common practice in laboratories that use GC-MS, but it is also applicable to systems that collect product ion spectra, like liquid chromatography quadrupole time of flight mass analyzers (LC-QqTOF). We compared the ability of three library searching algorithms to match collected spectra against a spectral database, using 41 patient urine samples which were analyzed by LC-QqTOF. Our results indicate that an optimized library search algorithm can improve the efficiency of compound detection and reduce the manual review of putative matches.
Background: Mass spectrometry based techniques have long been used to identify small molecules in biological samples. A variety of liquid chromatography mass spectrometry (LC-MS) platforms, including ion trap and quadrupole time of flight (QqTOF) instruments, are capable of producing robust product ion spectra though collision induced dissociation. These product ion spectra contain structural information about the original molecule. With help from computer algorithms, the collected spectrum can be matched against a database or library containing spectra of known compounds, and an identification can be made. Several commercially available algorithms exist, and previous publications have assessed performance in a variety of circumstances. Our goal in this work was to compare the performance of 3 different library search algorithms, NIST, LibraryView/MasterView (LV/MV) and ProLS, using one set of data files and one in-house built spectral database. Both the data files and the library were collected on a high resolution LC-QqTOF system. We evaluated the performance of each algorithm based on scores assigned to compounds known to be present in or absent from the sample, and overall efficiency of compound detection.
Methods: Our study included 41 patient urine samples that were submitted to our laboratory for routine drug screening. Samples were centrifuged to remove particulates and diluted 1:5 into starting chromatographic conditions. Separations were performed by a Shimadzu HPLC using a Phenomenex Kinetex C18 column (50x3 mm, 2.6 µm) held at 30°C. Mobile phase A consisted of 0.05% formic acid in 5 mM ammonium formate. Mobile phase B consisted of 0.05% formic acid in 50% methanol 50% acetonitrile. The injection volume was 10 μL. The elution gradient was ramped linearly from 2% to 100% B over 10 minutes. QTOF datasets were acquired for 41 patient urine samples using an ABSCIEX TripleTOF® 5600 system operating in positive ion mode, collecting full scan data with IDA triggered acquisition of product ion spectra. TOF and MS/MS mass accuracy was calibrated every 5 injections using the DuoSpray source and automatic calibrant delivery system.
Data were processed using PeakView® (version 2.2, ABSCIEX) and MasterView® (version 1.1, ABSCIEX) software. Research versions of the software containing alternative algorithms (NIST, ProLS) were provided by ABSCIEX. Each dataset was analyzed three times, once with each search algorithm. Common parameters included a mass extraction window of ±30 ppm, a retention time window of ±0.5 minutes, and an extracted ion chromatogram (XIC) list containing 206 compounds. The product ion spectral library was home-built, and contained compounds from a variety of drug classes. Compounds were considered positive when their combined score was ≥70, using a weighting combination of 10% mass error, 10% retention time error, 10% isotope pattern error and 70% library score.
An extensive analysis was conducted on the 41 patient urine samples used in this study. Results from immunoassay screens, mass spectrometry confirmations, broad spectrum mass spectrometry based drug screens and patient prescription records were used to establish the drugs present in each sample.
Results and Discussion: Each of the 3 test algorithms was used to search the same set of 41 LC-QqTOF data files, using the same product ion spectral library and the same XIC database. Each algorithm assigned a score to putative matches between spectra in the test datasets and spectra in the library. Scores were correlated with the known status (present or absent) of each compound in each sample. The NIST algorithm assigned scores from 0-100 to compounds present in the sample, and scores from 0-100 to compounds absent from the sample, leading to distributions that overlap significantly. In contrast, the PV/MV and ProLS algorithms assigned few scores <60 to compounds known to be present. Relative to PV/MV, the ProLS algorithm did assign more scores <60 to compounds present in the sample, which is likely to result in a slightly higher false positive rate. However, ProLS assigned fewer scores between 20 and 60 than did PV/MV. The range of 20-60 is an equivocal scoring range that often requires the user to visually examine the match and decide whether it meets acceptance criteria. Reducing the number of scores that fall in this range could provide some time and therefore cost savings to laboratories.
Sensitivity, specificity, and efficiency were calculated based on previously defined weighting and threshold criteria and were used to assess the diagnostic performance of each of the three algorithms. Though sensitivity and specificity were important performance indicators, we chose to use efficiency as the primary measure. Efficiency captures both the true positive rate and the false positive rate, and we reasoned that a library searching algorithm should maximize true positives while minimizing false positive results. The NIST algorithm showed the lowest sensitivity (89.3%) and consequently the lowest efficiency (83.6%) of the 3 algorithms we tested. The highest sensitivity (98.8%) was observed using the PV/MV algorithm, but this algorithm also had the lowest specificity (70.1%). The ProLS algorithm had a minor 0.6% reduction in sensitivity compared to PV/MV, however ProLS had a 7% higher specificity than PV/MV. The increased specificity coupled with only a slight decrease in sensitivity meant that ProLS had the highest overall efficiency (90.7%), by 2%.
Conclusions: Previous publications have demonstrated that optimization of library searching algorithms can improve matching ability, and our analysis supports this idea. Our results indicate that the ProLS algorithm offers increased diagnostic efficiency relative to the predicate algorithm PV/MV and the freely available NIST algorithm. In addition, the ProLS algorithm assigns fewer intermediate scores, which may limit the need for manual review of matches. This may offer a substantial time savings for laboratories. Extending this result to other types of mass spectrometers, including nominal mass ion traps, will be necessary for widespread adoption of the ProLS algorithm.
References & Acknowledgements:
IP Royalty: no
|Planning to mention or discuss specific products or technology of the company(ies) listed above:||