Almost all sectors of analytical chemistry are finding applications for ambient ionization mass spectrometry (AI–MS) because of its ease of use, speed of analysis, and sensitivity. Although emphasis has been placed on developing new hardware that can help analyze unique samples across various applications, there has not been much innovation in the functionality of software tools and mass spectral libraries to support applications like non-targeted searching. In this article, we discuss new algorithms and libraries that have enabled non-targeted analysis of small molecules using AI–MS, as well as some of the key considerations and outstanding questions in the field.

From forensic science (1) to healthcare (2) to food analysis (3), ambient ionization mass spectrometry (AI–MS) continues to find new application areas (4). The appeal of AI–MS—which is listed in almost every article on the topic—includes minimal to no sample preparation, rapid analysis times, the ability to alter ionization chemistry on the fly, and desirable sensitivity. However, there is a one (major) drawback to AI–MS, which is that the resulting mass spectra can be difficult to interpret. Given the lack of chromatography, an AI–MS spectrum consists of ions generated from all compounds in the sample overlaid on top of one another.

For some applications, like olive oil authentication (5) or diabetes detection (6), it may be logical to consider an AI–MS mass spectrum as a fingerprint of the sample. In this case, a barrage of statistical tools (for example, principal component analysis [PCA], linear discriminant analysis [LDA], and machine learning) can be applied to the collections of spectra to group which samples are similar, identify the potential source of a sample, or answer another classification-related question. Other applications, like forensic seized drug analysis, require the analyst to understand the precise chemical makeup of a sample. In these applications, treating an AI–MS mass spectrum as a fingerprint and using classification tools is inadequate.

So how does one accomplish non-targeted analysis with AI–MS? Given that most mass spectral search tools are designed for chromatography approaches where the spectra of pure compounds are compared, we need to use new approaches. Recognizing that most AI–MS sources are soft ionization techniques, and therefore typically produce intact molecular ions (that is, protonated or deprotonated molecules), one approach is to identify individual peaks in the spectrum using simple lookup tables of compound 

 values. Although this approach can be fruitful, lookup tables can provide inconclusive results when several compounds share the same 

 (for example, isomeric or isobaric compounds). This limitation is exacerbated when using unit resolution instrumentation.

For example, take the AI–MS spectrum of a seized drug sample shown in Figure 1. The main peak at nominal 

 238 could be attributed to ketamine, a common drug of abuse, 4’-chloro-α-pyrrolidinopropiophenone, an isomeric synthetic cathinone, or both. As a result, we need a different approach if we want to determine whether one or both compounds are truly in the sample.

There are currently two approaches that improve the specificity of the lookup table approach: tandem mass spectrometry (MS/MS) with collision induced dissociation (CID), or in-source CID (is-CID). In MS/ MS, we can specify that the 

 238 ion should be isolated, fragmented, and a product ion scan collected. These product ions can then be used to help elucidate the structure of the ion. Of course, specifying an ion to be isolated is not ideal when performing a true non-targeted analysis, especially in AI–MS when your sample is only ionized for a few seconds. With is-CID, one or more “fragment spectra” are generated at multiple is-CID energies without any precursor ion isolation. The issue here is that product ions in fragmentation spectra cannot be directly associated with their precursors. If we rethink the data analysis approach, these high energy is-CID fragmentation spectra can actually be used as a rich source of information in non-targeted analysis.

The inverted library search algorithm (ILSA) is a new data analysis approach developed at the National Institute of Standards and Technology (NIST) to use multiple is-CID mass spectra and unlock non-targeted screening with AI–MS (7–9). As the name implies, the ILSA inverts the traditional mass spectral library search process. Instead of asking the traditional question of “Does my questioned spectrum match this known spectrum?”, we instead ask “Does this known spectrum match my questioned spectrum?” Given that AI–MS spectra are mixture spectra, there is an important additional piece to the ILSA question. Specifically, we ask, “Does the known spectra match a sub-pattern within my questioned spectrum?” The sub-pattern piece is critical because we must assume our questioned spectrum is not pure. Although we can do this on just one is-CID spectrum (a low energy is-CID spectrum that produces intact protonated ions), the value is increased when we combine multiple is-CID spectra, enabling us to not only match intact ions but also match fragment ions that are produced at higher energies. By implementing the ILSA alongside a similarity metric of choice, we can obtain scores that can be used for deciding the identity of a compound in the mixture. Consider our example in Figure 1. When utilizing the ILSA along with the reverse match factor and three is-CID spectra, we can see the scores from ketamine are much higher than those of 4’-chloro-α-pyrrolidinopropiophenone, allowing us to have more confidence that the sample contains ketamine.

To truly make use of any library search algorithm, one needs a high-quality mass spectral library. The ILSA currently uses the NIST Direct Analysis In Real Time MS (DART–MS) Foren- sics Database (10,11), a regularly updated resource that contains multiple is-CID mass spectra of over 1100 compounds of interest to the forensic community. The library was built and evaluated with the assistance of a combination of automated and interactive software tools that could be used to produce libraries of measurements collected with other AI–MS platforms as well.

There is still one major question that needs to be addressed before we can be fully certain about best practices for non-targeted analysis using AI–MS and the ILSA, and that has to do with the effectiveness of centrally collected spectral libraries. Given the rapidly changing landscape of AI sources and MS platforms, are the data collected using a particular source on a specific MS platform useful to a chemist using a different configuration? The corollary to that question is equally important—even using the same configuration of instruments, is there other laboratory-to-laboratory variability (12,13) that impacts the effectiveness of a centrally collected AI–MS spectral library? In traditional chromatographic techniques where ionization occurs under vacuum, spectral reproducibility across instruments and laboratories is quite high and so libraries (for example, the NIST series of electron ionization [EI]–MS libraries) are reliable resources. Some preliminary studies in our laboratory focused on the platform-to-platform utility of a central library are promising, but much more work needs to be done. Unfortunately, there is currently limited literature on the laboratory effects of AI–MS measurements (14–17). Answering these questions will help us determine the costs and benefits of centrally created libraries like the NIST DART- MS Forensics Database compared to using custom libraries created under the specific conditions of the laboratory.

Although our recent work has made non-targeted analysis with AI–MS practically possible, the final piece of the puzzle is the identification of novel compounds. The ILSA, like most mass spectral library search algorithms, is limited to identifying known compounds that are contained within the search library. For chromatographic techniques producing mass spectra of pure compounds, NIST has a few approaches (for example, MS Interpreter, hybrid search, fentanyl classifier) to assist in novel compound classification. These methods are not naturally amenable to is-CID mass spectra of mixtures but variations of them might be applicable with further work. Developing additional tools and resources for novel compound identification, using AI–MS or other mass spectral platforms, is a worthwhile pursuit.

Non-targeted screening using AI–MS and is-CID mass spectra could potentially provide critical, and previously unthinkable, capabilities to nearly all sectors of analytical chemistry in the future. Accomplishing this objective would require the development of algorithms and data analysis procedures that improve capabilities even more. Near-instantaneous non-targeted compound identification may seem far-fetched, but the research is moving in that direction and the road to success seems clear.

Certain commercial products are identified to adequately specify the procedure. This does neither imply endorsement or recommendation by NIST nor does it imply that such products are necessarily the best available for the purpose.

These opinions, recommendations, findings, and conclusions do not necessarily reflect the views or policies of NIST or the U.S. Government.

https://doi.org/10.1016/j.forc.2020.100294

(2) H. Su, M.-Z. Huang, J. Shiea, and C.-W. Lee, 

(3) A. Arrizabalaga-Larrañaga, J.F. Ayala-Cabrera, R. Seró, J.F. Santos, and E. Moyano, in 

, C.M. Galanakis, Ed. (Academic Press, Cambridge, 2021), pp. 271–312.

https://doi.org/10.1016/j.trac.2020.116046

https://doi.org/10.1007/s11306-020-1637-8

(8) A.S. Moorthy, S.S. Tennyson, and E. Sisco, 

(9) E. Sisco, A.S. Moorthy, S.S. Tennyson, and R. Corzo, 

(10) E. Sisco and A.S. Moorthy, NIST DART-MS Forensics Database (is-CID), (2020). 

(11) E. Sisco, A.S. Moorthy, and L.M. Watt, 

(19), 9603–9611 (2014). https://doi. org/10.1021/ac502075t

(14) G.A. Newsome, L.K. Ackerman, and K.J. Johnson, 

(15) S.R. Kumbhani, L.M. Wingen, V. Perraud, and B.J. Finlayson-Pitts, 

https://doi.org/10.1007/s13361-018-2105-9

(17) G.A. Newsome, L.K. Ackerman, and K.J. Johnson, 

https://doi.org/10.1007/s13361-015-1259-y

 is with the National Institute of Standards and Technology, in Gaithersburg, MD. Direct correspondence to: 

Articles

Non-targeted analysis of small molecules with ambient ionization mass spectrometry (AI-MS) is enabled by new mass spectral libraries and algorithms, such as the inverted library search algorithm (ILSA) developed at NIST.

Algorithms and Databases: Unlocking Non-Targeted Screening of Small Molecules with Ambient Ionization Mass Spectrometry

In recent years, there has been an effort to categorize how emerging contaminants are identified in water samples by liquid chromatography–mass spectrometry (LC–MS). This process is often referred to as a workflow, and it has been the focus of many talks and conference sessions around the world. Many scientists have looked to unify and generalize the workflow, but do we really need a generic workflow to identify a non-target in an environmental sample? Would the scientific method approach (applied to each individual problem to solve) be enough and reliable?

One would not want to write about a topic without defining the key terms that are used frequently in the environmental community. When focusing on non-targeted screening for environmental contaminants in mass spectrometry (MS), we often use terms such as 

There is no doubt about what a target analyte constitutes. It is one that the user is looking for, for which a standard is available (usually in the laboratory), and the properties of which are already known, such as its chromatographic retention time, tandem mass spectrometry (MS/MS) spectra, and chemical structure.

The ambiguous terminology comes when one refers to 

. These compounds might be analytes that one is not looking for, but for which standards are available somewhere (they might even be in the laboratory freezer, but one has not made the standards or analyzed them!). They could also be compounds that one is not aware of (such as new pesticides coming onto the market or new pharmaceuticals being prescribed). Once a non-targeted compound is discovered, it becomes a target. Each individual user will have a different approach for this particular process. Mine is usually to analyze the new identified standard (if available), obtain the accurate mass, record the retention time, perform an MS/MS analysis, and understand the fragmentation of the analyte. It is then that it becomes a target.

However, my problem starts when someone talks about 

. What are those? Some may claim those are analytes that are in a database but for which the user does not have the standard. As a result, the suspect compound has never been analyzed through the methodology; therefore, nothing is known about the compound. But wait, aren’t we getting ahead of ourselves? Databases are part of a workflow scheme, and we have not gotten to it yet. Suspect is the worst ambiguous term to use. Do you really have to start the process of suspecting which compounds will be in your sample? How often do you do that? Again, it depends on your approach, what type of sample you have, and what the end goal is for it. For now, I will leave this term up in the air. We have yet to talk about 

 is next. This term gives me the most trouble. What really is an unknown? How do we definite it? It might be unknown to you, but not to me, or vice versa. Wouldn’t that be the same as non-target, a compound someone knows is out there in the environment, but for which the particular user does not own the standard and has never been analyzed using the method? Once the standard for that compound is analyzed, it becomes a 

 would be more fitting for a new natural product, such as one discovered in plants, which has never been reported before. It might also be some analyte for which the standard does not exist commercially yet, some new drug that has just been developed, or some pesticide that has just been manufactured and is ready to be released. The environmental community has used this term 

 to describe compounds that are not in databases, and compounds that the user had no idea were out there. To me, that is more of an educated guess. Once you have gathered enough information and done your research, you should know what that particular unknown (to you) is!

 is relative to the observer. One might be an experienced analyst and realize some of those non-targets, suspects, or unknowns are in fact targets and very well-known compounds that just happen to be in one’s memory. I myself have memorized hundreds of 

 masses for pesticides and pharmaceuticals and can quickly make an identification. That comes with experience, and everybody’s learning curve is dif- ferent. So, when defining terms, I think one should be aware of experience, knowledge, and the methodology used by the analyst. A beginner analyst might need another set of terms, and maybe a different approach.

This might be a simplistic view, but it has proven to work time after time. To me, there are no suspect compounds or even unknowns; a compound is either a target (you have the standard in your laboratory) or a non-target (you might not even know that the compound exists). Here, I give some examples and refer to some of the work our group has published successfully when referring to non-target analyses.

Instrumentation for Non-Targeted Analysis: High-Resolution
Mass Spectrometry (HRMS)

There is no doubt that the most commonly used and accepted methodology for non-targeted analysis is high-resolution accurate mass spectrometry. It has been approximately 25 years since the development of the first benchtop liquid chromatography (LC)–accurate mass MS instruments using time-of-flight (TOF) or high-resolution MS (HRMS). The instrumentation has evolved tremendously, making these instruments very sensitive, with high resolving power capable of distinguishing between very close isobaric compounds (which have the same nominal mass but different accurate masses). The environmental community has finally accepted these methodologies as unique and capable of identifying an analyte in a sample. Back in the early 2000s, there was still a debate in food and veterinary chemistry of what constituted a positive finding, depending on a point system if using low-resolution mass spectrometry or HRMS (1).

Moreover, there have been attempts to use tandem MS instruments for non-targeted analyses, and those are valid as long as one follows the golden rules of MS and identification. In fact, ion traps were used at the beginning to aid in the identification of non-targeted analytes because of their fragmentation powers and structural elucidation capabilities. Those instruments were often underestimated as to what could be identified in a sample following a few simple approaches.

This overview focuses on using accurate mass MS for the identification of non-targeted compounds.

 is often used to refer to the process that leads to an identification of an analyte in a sample. To me, the identification process is simple, maybe because I learned to do it before any workflow software was available. But new generations of analysts have lost sight of the “golden rules” of MS, and other processes that are intuitive and follow pattern recognition. But I would still ask, do we really need a workflow?

As an author, reviewer, and editor, I have become aware of the increase in papers relating to workflows for non-targeted analyses. There has been an effort to unify this process to make it easier for the analyst to identify as many compounds as possible in a given sample. One important aspect of a workflow is the software. Nowadays, every accurate mass MS instrument manufacturer has its own software package that aids the user in the identification process. However, some of these software programs are not easy to use, have tedious learning curves, and are quite expensive. On top of that, software does not guarantee that you will correctly identify everything that is in a sample. False positives and negatives are the most common drawbacks encountered. In the next subsections, I discuss what I think needs to be considered when identifying non-targets in samples.

Sometimes, the most important process in environmental analysis is forgotten: sample preparation. Every sample, every project, and every problem is different. Sample preparation is an important step in environmental analysis, and it will determine which compounds you identify or don’t identify at the end. For example, fire-related compounds that are organic acids need a pH adjustment before the sample is processed through solid-phase extraction (SPE). If that step is not done, then no identification of polycarboxylic acids will be possible (2). Some workflows described in the literature do not take into account the important step of sample preparation.

There are myriad instruments and software packages, so where does one start? How about not having a generic workflow? How about developing a unique workflow each time a sample arrives in your laboratory? After giving careful consideration to sample preparation, one has to decide the acquisition method one wants to use. That alone will also determine what we will be able to identify. There are quite a few different acquisition methods depending on instrumentation and manufacturer type. Mainly, there are two methods that I can generalize when using quadrupole HRMS: data-independent acquisition, and data-dependent acquisition. In my view, data-independent acquisition (using a high in-source voltage) has been proven very effective in identifying compounds and their metabolites as a first step (3). Of course, data-dependent acquisition (MS/MS triggered) is also a highly effective approach for unequivocal identification using MS/MS spectral mass, libraries, and chemical structural elucidation.

Depending on which workflow one follows, at some point there is a step that involves formula generation from a measured accurate mass. Again, it is the user (not the software!) that has to set the elements that are included in the formula generation. A visual inspection of the mass spectra data will reveal if a sulfur atom or halogen has to be included. For pesticide analysis, phosphorus and fluorine are also elements to consider. These are all user-dependent parameters; the analyst has to make the choice and guide the software, not the other way around. Thus, the analyst should be well trained in accurate mass MS analysis. In one study, for example, halogen signals proved to be a key to the identification of a new anti-depressant in water samples (4).

Then, there are the databases. There are hundreds of them, and accurate mass MS data for each individual compound and their fragments can be acquired easily nowadays. There are homemade and commercial databases. Again, the database that is used determines what you find at the end. There have been attempts to have accurate mass spectral libraries as well, but each instrument is different and will give different ions and abundances for each fragment ion, depending on collision energy. Therefore, one has to decide if that is the best approach for a given sample. The advantage of using generic accurate mass databases is that this parameter of accurate mass for the protonated or depronated molecule is universal; it will not depend on the instrument or acquisition method used (5).

Therefore, the approach one follows is highly dependent on the sample type, sample preparation, acquisition method used, software (or lack of) used to extract data, and finally, user interpretation. Can that be unified into one general workflow? I don’t think so. The analyst still has to do the final interpretation. The instrument will give you accurate data (assuming it is well calibrated and performing correctly), and the software will guide you and try to interpret, but only to a point. The final decision is based on the knowledge, data, and experience of the analyst. That is what will prevail at the end.

I mentioned above some of the golden rules in MS that have been forgotten or seldom considered. I will comment on a few of them here.

It might seem obvious, but accurate mass has to be calculated correctly. Not too long ago, there were still some software programs that did not take into account the mass of an electron when calculating accurate mass (6).

A second rule is the nitrogen rule, which is an easy one to take into account if we only consider even-electron ions for a formula generation. However, there are software programs that will not account for this (7).

Mass defect has proven to be very useful for the identification of homologue series (such as surfactants) by measuring differences in alkyl chain lengths (8).

Obvious halogen and sulfur isotopic signals (abundance and intensity) are also highly useful in the structural elucidation and identification of compounds.

There is the potential for adduct formation with sodium, ammonium, and potassium. The identification of sucralose is an example of considering the formation of a strong sodium adduct in positive ion mode (9).

Labile analytes fragment easily and do not show a protonated or deprotonated ion; instead, one of the fragments is observed in electrospray. Examples of this are diphenhydramine and bupropion.

Finally, we cannot forget fragmentation patterns and fragmentation trees, some of which have also been included in advanced software programs. Again, the user has to guide the software to make an unequivocal and ultimate identification.

HRMS instruments are capable of unequivocally identifying many targeted and non-targeted analytes. The approach to follow will always depend on the type of sample, the analyst’s knowledge, the end goal to achieve, and the tools available (software packages). Generic workflows are not necessarily needed because rapidly changing variables and parameters can be used, depending on the need of each individual problem and user. Trying to unify all these complex pathways in a single workflow or identification system will never work. In the end, it is the analyst who is in charge of the data obtained from the instrument. The instrument produces the data, and the software makes an interpretation, but the analyst has to guide the whole process for successful analysis.

(1) A.A.M. Stolker, E. Dijkman, W. Niesing, and E.A. Hogendoorn, 

(2) I. Ferrer, E.M. Thurman, J.A. Zweigenbaum, S.F. Murphy, J.P. Webster, and F.L. Rosario-Ortiz, 

(5) E.M. Thurman, I. Ferrer, O. Malato, and A.R. Fernández-Alba, 

(7) E.M. Thurman, I. Ferrer, O.J. Pozo, J.V. Sancho, and F. Hernandez, 

(9) I. Ferrer, J.A. Zweigenbaum, and E.M. Thurman, 

 is with the University of Colorado, in Boulder, Colorado. Direct correspondence to: 

Are generic workflows really needed in environmental analysis? In the end, it is the analyst— not instrumentation or software—that is in charge of what data are obtained, after thoughtfully considering what sample preparation and acquisition method to use.

What’s Your Workflow? Non-Targeted Analysis of Water Samples

Conventional mass spectrometry is limited by heterogeneity to relatively small molecular weights from a biologics perspective. Charge detection mass spectrometry—where masses of individual ions are directly measured—overcomes this limitation and allows measurements into the gigadalton (GDa) range. An important application of this emerging technology is the analysis of gene therapy vectors. Their large size and complexity present analytical challenges that must be addressed for this promising therapeutic approach to reach its full potential.

During the last few years, gene therapy has re-emerged as a promising approach (1,2). There are now two FDA-approved therapies, with hundreds more in clinical trials. Both approved therapies use adeno-associated virus (AAV) as a vector. There are also clinical trials using adenovirus and lentivirus vectors. Adenovirus is also the basis of several novel SARS-CoV-2 vaccines. AAV, with a molecular weight (MW) of approximately 5 megadaltons (MDa), is a small virus. Adenoviruses and lentiviruses are much larger, with MWs well over 100 MDa. All these species are beyond the range of conventional mass spectrometry (MS). Their large size and complexity make analytical characterization challenging (3,4), while, at the same time, the prospect of expanding patient numbers makes detailed characterization more critical. Can MS play a role?

Electrospray allows large biomolecular assemblies to be ionized and launched into the gas phase where accurate MWs can be determined by MS (5,6). Ions generated by electrospray are often multiply charged, with megadalton-sized ions typically having more than 100 charges. In conventional MS, only the 

 ratio is measured. Because the ions are produced in a distribution of charge states, a series of 

 peaks occur for each mass. The charge can be deduced from the spacing between the peaks, and then the 

 of each peak is multiplied by its charge to give the mass. Heterogeneity, because of post-translational modifications, counterions, and trapped solvents, broadens the 

 peaks, making it difficult to assign charge states. Heterogeneity generally increases with size, and, in most cases, the charge cannot be determined for ions larger than 1 MDa without prior knowledge of the mass. Charge detection mass spectrometry (CDMS) is one way to overcome this limitation. CDMS is a single particle technique where the m/z and charge are measured simultaneously for each ion. These quantities are then multiplied to give the mass for each ion. The measurement is repeated for many different ions, and then the masses are binned into a histogram to give the mass distribution.

The charge on a single ion is small, and measuring it accurately has been the main challenge in CDMS since the first measurements on electrosprayed ions in the 1990s (7,8). To assign a single ion to its correct integer charge state, the measured charge must fall in a window ±0.5 e (elementary charges) of the integer charge. If the uncertainty in the charge measurement is too large, many ions will be assigned to the wrong (neighboring) charge states. As explained in Figure 1, when the root mean square deviation (RMSD) increases beyond 0.2 e, the number of misassigned ions increases rapidly (9).

To perform CDMS measurements, ions are trapped in an electrostatic linear ion trap where they oscillate back and forth through a detection cylinder (10–12). When an ion enters the cylinder, it induces a charge that disappears when the ion exits. The periodic signal generated by the trapped ions is sensed by a charge sensitive amplifier, and the resulting time domain signal is analyzed by fast Fourier transforms. The oscillation frequency is used to determine the 

 ratio, and the magnitude of the fundamental is related to the charge. If the detection cylinder is long enough, the full charge of the ion is induced on the cylinder, and the resulting signal is independent of the ion’s trajectory. This independence is key to measuring the charge accurately. The main factor limiting the accuracy of the charge measurement is then electrical noise. For a fixed noise level, the uncertainty in the charge measurement can be reduced by extending the measurement time. An uncertainty approximately 0.2 e is achievable routinely (13,14). The measurement time can be reduced by performing measurements on several ions at a time (15), though it is still longer than conventional MS.

Figure 2a shows CDMS measurements for AAV vector serotype 8 (AAV8) derived from HEK cells. AAV is a small virus consisting of a capsid (a protein shell) that surrounds a single-stranded DNA genome. First, the capsid assembles, and then the genome is packaged, leading to particles that package the full genome and empty particles. The blue line in Figure 2a shows the mass distribution measured for empty AAV8 particles separated using a cesium chloride (CsCl) gradient. There is a single peak centered on 3.74 MDa. The AAV8 capsid is icosahedral and contains 60 capsid proteins. However, there are three different types of capsid protein: VP1, VP2, and VP3 that are present in a ratio of approximately 1:1:10. The sequence masses of VP1-3 for AAV8 are 81,667 Da, 66,692 Da, and 59,805 Da, respectively. Thus, the expected average mass of the empty capsid is 3.732 MDa. The measured mass (3.74 MDa) is slightly larger. It is common in native MS for measured masses to be slightly larger than the expected mass because of counterions, salt molecules, and trapped solvents. The peak is broad because the capsids contain different combinations of VP1–3. The width of the peak matches a multinomial distribution, indicating that the number of VP1-3 in each capsid is random (16).

The orange line in Figure 2a shows the CDMS mass distribution measured for AAV8 with a CMV-eGFP genome. The AAV genome is single-stranded DNA, approximately 4700 nt for the wt virus. The CMV-eGFP genome is 2207 nt, with an average mass of 0.680 MDa. The peak in Figure 2a at 4.49 MDa is attributed to particles that have packaged the full CMV-eGFP genome. The expected mass is the mass of the empty capsid plus the mass of the genome; in this case, 3.74 + 0.68 = 4.42 MDa. The measured mass is 70 kilodaltons (kDa) larger than expected. The difference is attributed to counterions and small DNA fragments. The CMV–eGFP genome is small enough that two genomes could fit into the capsid. There is a peak at 5.22 MDa, which could be because of two genomes, but other studies suggest that this peak is probably because of a single strand of heterogeneous DNA (fragments of plasmid and host DNA) (17).

The packaged DNA can be better characterized by extracting it and analyzing it by CDMS. The green line in Figure 2a shows the mass distribution measured after incubating AAV8-CMV-eGFP at 80 °C for 15 min. After incubating the peaks at approximately 5 MDa, the intact capsids with cargo have disappeared, and there are many low mass fragments. Although not evident from Figure 2, there are also a significant number of high mass species with masses that extend beyond 100 MDa. These species result from the aggregation of AAV disassembly products. The notable feature of the green spectrum in Figure 2a is the sharp peak at 1.37 MDa. This peak is at a mass slightly larger than expected for the double-stranded CMV-eGFP genome (1.359 MDa). Although the packaged genome is ssDNA, + and – strands are packaged with roughly equal probability. Once released, the ssDNA can base pair in solution. The small deviation between the measured and expected masses of the dsDNA (11 kDa) is attributed to counterions.

Figure 2b shows the charge versus mass scatter plots for the empty AAV8 (blue) and AAV8-CMV-eGFP (orange). The charge deposited on a large electrosprayed ion depends on its size. In this case, the empty and full AAV8, and the overpackaged AAV8-CMV-eGFP all have approximately 160 charges. The fact that the charges are similar indicates that the genome is being packaged inside the capsid. If it was on the capsid surface, the charge would be larger.

The packaging capacity of AAV is relatively small, thus limiting the size of the therapeutic genome that can be used. Adenovirus, one of the largest non-enveloped viruses, is much larger than AAV (see insets in Figure 2a and 2c), and can accommodate a large genome. A mature adenovirus contains 13 different proteins, and over 2500 protein molecules. The mass distribution measured for human adenovirus type 5 (HAdV5) is shown in Figure 2c (18). The peak at 156.1 MDa is attributed to the mature particle with a genome, and the smaller peak at 129.6 MDa is assigned to immature particles that have not packaged a genome (19). The corresponding charge versus mass scatter plot is shown in Figure 2d. The full particle has an average charge approximately 1100 

. The average charge for the empty particle is approximately 900 

. The surface of adenovirus’s icosahedral capsid is decorated with fibers that play a role in cell entry. The empty particle is missing the fibers, which is why the empty particles have a significantly lower charge than the full.

The results presented here for AAV and the adenovirus illustrate the value of extending the mass range of conventional MS into the gigadalton regime. At the time of writing, CDMS is only available in a few research laboratories (20–22). However, it is being commercialized and will soon become more widely available. Recently, individual ion measurements have been performed on an orbital ion trap (23,24). The induced charge depends on the ion trajectory in this case. Consequently, the charge uncertainty is large, approximately 4% for individual ions, and the resolving power is low.

The author is a shareholder in Megadalton Solutions, a company engaged in commercializing CDMS.

This work was supported by the National Institutes of Health under award number R01GM131100.

(4) A.L. Gimpel, G. Katsikis, S. Sha, A.J. Maloney, M.S. Hong, T.N.T Nguyen, J. Wolfrum, S.L. Springs, A.J. Sinskey, S.R. Manalis, P.W. Barone, and R.D. Braatz, 

(6) S. Mehmood, T.M. Allison, and C.V. Robinson, 

(8) S.D. Fuerstenau, W.H. Benner, J.J. Thomas, C. Brugidou, B. Bothner, and G. Siuzdak, 

10.1002/1521- 3773(20010202)40:3<541::AID-ANIE541>3.0.CO;2-K

(9) D.Z. Keifer, D.L. Shinholt, and M.F. Jarrold, 

(13) A.R. Todd, A.W. Alexander, and M.F. Jarrold, 

(15) C.C. Harper, A.G. Elliott, L.M. Oltrogge, D.F. Savage, and E.R. Williams, 

(16) E.E. Pierson, D.Z. Keifer, A. Asokan, and M.F. Jarrold, 

(17) L.F. Barnes, B.E. Draper, Y.-T. Chen, T.W. Powers, and M.F. Jarrold, 

(18) L.F. Barnes, B.E. Draper, and M.F. Jarrold, 

(19) E.K. Mullins, T.W. Powers, J. Zobel, K.M. Clawson, L.F. Barnes, B.E. Draper, Q. Zou, J.J. Binder, S. Dai, K. Zhang, O. Friese, H.A. Runnels, M.F. Jarrold, and L.C. Thompson, 

(21) C.C. Harper, D.D. Brauer, M.B. Francis, and E.R. Williams, 

(22) B.L. Barney, R.T. Daly, and D.E. Austin, 

(23) J.O. Kafader, R.D. Melani, K.R. Durbin, B. Ikwuagwu, B.P. Early, R.T. Fellers, S.C. Beu, V. Zabrouskov, A.A. Makarov, J.T. Maze, D.L. Shinholt, P.F. Yip, D. Tullman-Ercek, M.W. Senko, P.D. Compton, and N.L. Kelleher, 

(24) T. P. Wörner, J. Snijder, A. Bennett, M. Agbandje-McKenna, A.A. Makarov, A.J.R. Heck, 

 is with the Chemistry Department of Indiana University, in Bloomington, Indiana. Direct correspondence to: 

Gene therapy vectors in the gigadalton (GDa) regime are beyond the range of conventional mass spectrometry. Charge-detection MS overcomes this limitation.

Charge Detection Mass Spectrometry: Analysis of Gene Therapy Vectors

Extracting thousands of metabolic features from liquid chromatography–mass spectrometry (LC–MS)–based metabolomics data is not easy. Although many feature extraction algorithms have been developed over the past few decades, automated feature extraction is still not a “white box” process. For instance, it is challenging to quickly determine the optimal parameters for the best feature extraction outcome. It is also impossible to extract every true metabolic feature. Moreover, there is contamination from false metabolic features of different sources, such as signal noise and in-source fragmentation. Our laboratory has recently developed a suite of bioinformatics tools to address these metabolic peak-picking challenges. The goal is to improve the peak-picking outcome quality, so we can effectively obtain biological information from the metabolomics data.

Liquid chromatography–mass spectrometry (LC–MS) is a prominent analytical platform that has been widely used in untargeted metabolomics. Owing to its high sensitivity and throughput, LC–MS-based metabolomics generates a large amount of data, containing the identities of and quantitative information for thousands of metabolic features from a single biological sample.

The first step of processing the megabytes or gigabytes of metabolomics data is to automatically extract all the metabolic features. The past decade has witnessed the development of many bioinformatics programs for automated metabolic feature extraction. Some of the most commonly used academic software include XCMS, MS-DIAL, and MZmine 2 (1–3). These data processing programs have been successfully applied to thousands of metabolomics studies, revealing the metabolic signatures associated with aging, cancer, and host-microbiome interactions, among others (4). However, none of the existing metabolomics peak-picking algorithms can completely fulfill its task, the main reason being the complexity and large scale of LC–MS-based metabolomics data. In particular, given the diverse chemical structures and broad concentration ranges (sub-femtomolar to sub-millimolar) (5), metabolic features present various extracted ion chromatogram (EIC) peak shapes. As a result, it is challenging to find the best peak-picking parameters that can recognize all of the true metabolic features. Therefore, we often find that some high-confidence true metabolic features are missed by the automated peak-picking process. The same challenge also leads to various false metabolic features that are picked up and included in the feature list. These false features can be contributed by signal noise, system contamination, and in-source fragmentation. The combination of an incomplete list of true features with the existence of many false features diminishes the peak-picking quality, and makes the downstream data interpretation problematic. Unfortunately, because of the large size of metabolomics data and the many features that can be picked up, it is very difficult to manually check the results to fix the problems. All these issues show that current metabolomics peak picking is still a “black box” process.

Figure 1 illustrates the workflow of metabolomics peak picking and summarizes the abovementioned issues generated during automated peak picking via conventional data processing software. Our recent work from the past two years aims to overcome the hurdles in metabolomics data processing so that we can obtain high-quality metabolomics data for downstream metabolite annotation and biological interpretation (Figure 2). Notably, all the developed programs are freely available on GitHub (

), along with revised and regularly updated content.

Measuring Optimal Data-Processing Parameters: Paramounter

The first challenge in processing metabolomics data is choosing an appropriate set of peak-picking parameters. Essentially, the parameters in all data processing programs are a set of thresholds defining what types of chromatographic peaks should be recognized as metabolic features. The parameters heavily rely on the LC–MS analytical conditions and can dramatically influence the number and quality of the extracted features. Given the different chromatographic peak shapes of metabolic features, it is difficult to directly decide the best parameters unless users manually check all the features. Conventionally, researchers use the design of experiments (DOE) strategy to test many different parameter combinations until the best performance is reached (6). However, the DOE approach is time-consuming, and does not provide a mechanistic explanation of the chosen parameters. To address the challenge, we tried to view this question from a different angle. Because the parameters are a set of definitions based on the features’ chromatographic attributes, we believe that it is possible to directly measure these parameters if we can plot the chromatographic attribute distributions from the raw LC–MS data. The distribution plots allow us to turn optimization from a time-consuming DOE trial into a direct measurement. To achieve this, we studied the peak-picking algorithms of five commonly used open-access data processing programs, including XCMS, MS-DIAL, MZmine 2, El-MAVEN, and OpenMS (1–3,7,8). We summarized and discovered four universal parameters, namely mass tolerance, peak height, peak width, and instrumental shift. We then developed an R script, Paramounter, to automatically measure these universal parameters directly from the raw data (9). More importantly, these universal parameters can be automatically translated to the specific parameters of the data processing software. The Paramounter developed in this work is a much more advanced approach than using time-consuming DOE trials.

Rescuing True Features Missed by Peak Picking Algorithms: JPA 

Although peak-picking algorithms can automatically extract thousands of metabolic features from raw LC–MS data, many true metabolic features are missed, which is because of the diversity in mass accuracy, peak shapes, and abundances. At the current stage, no existing feature extraction algorithms can pick up 100% of the features. It is important to recognize the high cost of missing true features in metabolomics. If a true feature relevant to the given biological question is missed, it will be very hard to retrieve later, and the downstream biological interpretation will be affected. Through manually investigating missed real metabolic features, we found that many have high-quality tandem mass spectrometry (MS/MS) spectra collected during the LC–MS analysis using data-dependent acquisition. To rescue these true positive features, we proposed an integrated feature extraction method for features with good and bad chromatographic peak shapes (10). Our results show that a significant number of high-quality metabolic features can be extracted using our integrated approach. To streamline automatic peak picking, precise quantification, and metabolite annotation, we further developed joint metabolic feature extraction (JPA) and automated metabolite annotation. JPA integrates three peak picking strategies for the comprehensive extraction of true positive features (11). The first incorporated algorithm is centWave, a conventional method that extracts features with Gaussian peaks. The second algorithm rescues the metabolic features that are missed by the first algorithm but have associated MS/MS spectra. The third strategy is a targeted approach to extract metabolites with known 

 values and retention times. This module can further rescue the metabolic features that are important to the given biological question but not found by conventional peak picking. JPA can drastically reduce the number of missed true features and facilitate much higher metabolite coverage with twofold more metabolic features.

An important step in processing metabolomics data is to manually check the extracted metabolic features to ensure that they are not false. A significant portion of false metabolic features are noise signals. These false features are extracted as their peak shapes accidentally pass the peak-picking criteria. Although these features can be eas- ily recognized by a trained expert, manually checking over thousands of metabolic features in a raw metabolomics data file takes days. It is even less realistic if the study is made up of tens or hundreds of samples. Because the pattern of false features can be easily recognized by humans, our laboratory sought to replace manual labor with automated recognition through deep learning. In particular, we leveraged a state-of-the-art convolutional neural network (CNN) and trained a CNN model with over 25,000 manually recognized EIC peaks from data of various sample types, LC–MS configurations, and spectra acquisition rates (12). The trained model was then developed into a Windows application, termed EVA (short for evaluation). The diversity of the training EIC plots ensures that EVA has over 90% accuracy in true and false positive feature recognition. Furthermore, EVA has a user-friendly interface, easily accessed by users with minimum programming expertise, to filter out false positive features with a few clicks.

Recognizing Features from In-Source Fragmentation: ISFrag 

Besides background noise, false metabolic features can also be contributed by in-source fragmentation (ISF), where metabolic precursor ions are inevitably fragmented during ionization. These fragment ions are then falsely recognized as true metabolic features. In general, it is difficult to differentiate ISF features from true metabolic features in an LC–MS analysis as both can have good chromatographic peak shapes. Traditionally, ISF features can be identified by common neutral losses (loss of H

, for example) and MS/MS spectral similarity comparison against the chemical standards of the parent metabolites. However, the diverse neutral loss patterns and the limited MS/MS of metabolite standards make the confirmation of ISF features hard. To address this challenge, we developed a program that can automatically identify ISF features in a de novo manner (13). In brief, we rely on three useful patterns for automated ISF feature recognition: (a) ISF features are coeluted with their precursor ions; (b) the 

 of an ISF feature exists in the MS/MS spectra of their precursor ion; and (c) ISF features and their parent features have similar fragmentation patterns. Following these rules, we developed ISFrag, an R package that can precisely recognize ISF features. Tested using a standard mixture of 125 metabolites, ISFrag achieves 100% accuracy in identifying the ISF features.

The development of bioinformatics tools is critical to the further advancement of metabolomics in the post-genomic era of biology. While the developments in instrumentation and analytical methods are critical to high-quality metabolomics data generation, it is equally important to develop bioinformatics tools to enable efficient extraction of biologically relevant metabolic information for downstream data interpretation. As demonstrated in this short article, advanced bioinformatics programs can significantly improve the quality of metabolomics data, thus making its biological applications more convenient and confident. We hope that this article can also help researchers become more aware of the bioinformatics challenges in metabolomics, and encourage the metabolomics community to further develop bioinformatics solutions to address them.

(1) C.A. Smith, E.J. Want, G. O’Maille, R. Abagyan, and G. Siuzdak, 

(2) T. Pluskal, S. Castillo, A. Villar-Briones, and M. Orešič, 

(4) C.H. Johnson, J. Ivanisevic, and G. Siuzdak, 

(8) E. Melamud, L. Vastag, and J.D. Rabinowitz, 

(13) J. Guo, S. Shen, S. Xing, H. Yu, and T. Huan, 

 are with the Department of Chemistry at the University of British Columbia, in Vancouver, British Columbia, Canada. Direct correspondence to: 

Extracting thousands of metabolomics features from LC–MS–based metabolomics data is not easy. A new suite of free bioinformatics tools helps address these metabolic peak-picking challenges and thus obtain more accurate biological information.

Hot Topics in Mass Spectrometry | LCGC Supplements

Algorithms and Databases: Unlocking Non-Targeted Screening of Small Molecules with Ambient Ionization Mass Spectrometry

What’s Your Workflow? Non-Targeted Analysis of Water Samples

Charge Detection Mass Spectrometry: Analysis of Gene Therapy Vectors

Turning Metabolomics Data Processing from a “Black Box” to a “White Box”