MS puzzle: strategies for automated structure elucidation

Dec 1, 1987 - Solving the MS/MS puzzle: strategies for automated structure elucidation. Christie G. ... Journal of Chemical Information and Computer S...
11 downloads 10 Views 10MB Size
.

Solving the I ’ \ MS/MS

I ‘

I

Strategies for Automated Structure

Artificial intelligence software approaches are providing new insights into MS/MS spectral feature/substructure relationships Christie 0. E k e Adrian P. Wade Peter T. Palmer Kevin J. Hart

Department of Chemislv Mlchigan Stale University East Lansing, Mich. 48824 There has been considerable interest recently in advancing the state of automated structure elucidation. To some extent this interest has been a necessary reaction to growth in the so-called hyphenated techniques ( I ) , improvements in data collection speeds (2),and the ever-increasing ability of new instrumentation to generate large quantities of multidimensional data. Recent advances in mass spectrometry instrumentation include femtogram detection limits (3);the ability to collect a complete mass spectrometry/ mass spectrometry (MS/MS) fragmentation map in a few seconds (4); and multidimensional instrumentation such as Fourier transform mass spectrometry (FT-MS), which has produced five consecutive stages of MS (5),and gas chromatography/infrared/ mass spectrometry (GC/IR/MS), which produces five dimensions of information (6). The sheer volume of data produced by such techniques mandates some automated method for extracting chemically relevant information. As such multidimensional instrumentation becomes more common, one can expect that traditional structure elucidation tools (including human experts) will fail to extract all the valuable analytical information within a reasonable time interval. Thus development of

new automated structure elucidation procedures has become a priority. Van Dalen notes that “there can be little doubt that the future of analysis is inextricably linked with that of the computer” (7).Any intelligent instrument should be an adaptive system; ideally, it should be able to learn from experience and its operation should adapt to changes in external circumstances (e.g., self-optimization). Intelligent instruments perform operations normally left to a human expert. As such, they will incorporate aspects of artificial intelligence which, when applied to chemical instrumentation, is defined as “the scientific discipline which attempts to endow computercontrolled machinery with the ability for actions which, if done by a human being, would be thought to require intelligence” (8). Expert systems are a practical application of artificial intelligence that attempts to capture the interpretive skills of experts in a form that can then he consulted by less knowledgeable users. Most expert systems use knowledge formalisms in which expertise may be represented by rules (e.g., if Q is true and b is true, then conclude c). Rule-based programming offers several advantages over conventional algorithmic approaches. It is simpler to understand and modify; one explicitly stated rule may be equivalent to several instances of implicit knowledge represented by conventional code scattered throughout a large program. Problems that proved intractable by conventional programming styles have been shown to be solvable using a rulebased approach (9). The results from a rule-based system may be reviewed by a human user in terms of the rules ap-

I

plied to solve the problem. Experts in mass spectral interpretation are highly trained individuals and, like any experts, they are hard to find. Mature expertise in tandem mass spectral interpretation has not yet had a chance to develop. An instrument containing an expert system with just part of the human experts’ skills will be a significant advance over what has gone before. Truly expert spectral interpretation systems should be able to deal with facts, rules, and metarules. Facts are simple statements (e.g., the collision gas pressure is 1 millitorr) that may have some degree of uncertainty assoeiated with them. Rules are the mechanisms by which an expert establishes new facts, based on what is already known (e.g., if neutral loss of 28

amu occurs, then the carbonyl substructure is likely). Metarules indicate the pathways by which an expert formulates new rules and plans how to solve problems. They describe the formalisms and procedures by which the application rules are developed (i.e., the mechanisms for learning from experience). Metarules can also assist in conflict resolution (when available evidence suggests two or more conflicting conclusions) and temporal reasoning (when evidence that was obtained previously has been refuted or is now no longer valid for some other reason). Thus metarules are essential when the data space is very large, complex, and inherently empirical. Automated tedvllques for mass spectral Interpretation Many approaches to automating the interpretation process have been tried, each aimed at decreasing the level of expertise needed by a user or increasing the amount of useful information

ANALYTICAL CHEMISTRY, VOL. 59, NO. 23, DECEMBER 1, 1987

1363,.

derived. In a recent article Small categorized such automated spectral interpretation techniques as either direct or indirect database methods (10).Direct database methods, commonly called spectral matching methods, require a library of reference spectra and some means for comparing sample and reference spectra. The prominence of these methods can be attributed to the success and popularity of GC/MS as a mixture analysis technique. Several different spectral matching methods for mass spectrometric data are currently available (11, ZZ), the most well-known of which is the ProbabilityBased Matching system (13). Considering that Chemical Abstract Services currently recognizes more than I million different organic compounds, the most serious drawback to hal matchi !thod

that a library of mass spectra can never be complete. Furthermore, the precision of this technique decreases as the library becomes more complete, hecause the number of spectra similar to any given spectrum will increase. Experimental irreproducihility that results in differences between sample and reference spectra of the same compound leads to increasingly erroneous results. Although spectral matching methods are valuable aids for limited-domain problems involving known compounds, for true unknowns (compounds whose spectra may not exist in the library) one must often resort to indirect database or interpretive methods for structure elucidation. The SelfTraining Interpretive and Retrieval System (STIRS) is one such method. STIRS deduces si ictura ‘ma-

rigure l.Schematlc of the DENDRAL approach to structure elucMation using pian, generate, and test stages. 1364A

ANALYTICAL CHEMISTRY, VOL. 59, NO. 23, DECEMBER 1. 1987

tion about an unknown by analyzing its mass spectrum for the presence of 26 different classes of mass spectral data (13).These data classes correspond to combinations of fragment masses or inferred neutral losses that are known to have structural significance. This teehnique directly uses information from all available library spectra without resorting to predefined spectrumlsuhstructure correlations; hence the selftraining appellation. Perhaps the most well-known indirect database method is the DENDRAL project, which began a t Stanford University in 1965 (14). DENDRAL, a classical approach to the solution of a problem with a large state space, employs plan, generate, and test stages (Figure 1).The plan stage (Heuristic DENDRAL) derives constraints on the unknown structure using empirically derived fragmentation rules that are automatically inferred from mass spectral data of known compounds by Meta-DENDRAL (IS).The generate stage (GENOA) provides all possible nonredundant structures consistent with the constraints (16). The test stage ranks the resulting list of structures hy comparing their simulated mass spectra to the unknown spectrum. This simulation uses fragmentation rules derived by Meta-DENDRAL. DENDRAL has been applied to several problems, and its performance has been shown to equal or exceed the performance of a human expert in structure elucidation (16). Its power is derived not from “knowing” more than any human expert, but from a thorough application of constraints and a systematic search through the space of possible structures. However, in many cases mass spectral data alone were insufficient to determine the complete structure of an unknown. Thus DENDRAL used NMR data to provide additional substructural constraints. DENDRAL’s simulation of mass spectra can best he described as an approximation. A complete and accurate simulation of mass spectra for all molecules under various experimental conditions is currently unobtainable. MSIMS Another dimension in structure elucklatlon Mass spectra of compounds that contain common substructures often show patterns of features that are caused by those substructures. Some common fragment ions and inferred neutral losses have been recognized as fairly specific indicators for certain suhstructures. These have been tabulated and are widely used for mass spectral interpretation (17,18). Until recently, there has been no attempt to exhaustively

The most complete and versatile line of metal

-

Laboratory Furniture in lab furniture by Duralab really pays dividends .in years of service.

..

-

Constructed of superior materials, such as lead-coated sheet steel with baked-on enamel finish providing corrosion and acid resistance for the utmost in performance, durability

III Q U I CATALOG IN

and

Custom features include concealed hinges, doublepanel

insulated doom and drawers with removable panels for and decontamination. Get the full story from our engineering team.

iDURALAB 'ME N T CORPORATION

SWEETS

CIRCLE 32 ON READER SERVICE CARD \1

THE MICHELSON 100 FT-IR THE PERSONAL TEACHING ASSISTANT7 * Full FT-IR performance is available starting at S2l.995 in the US'

* 2 Year Warranty* Large and easily accessible sample compartment

-

Easy to operate just turn it on and its ready

* 48 software functions available

Sealed spectrometer housing - protects against accidents * Install it

anywhere

Menu driven software

\

rl

I

* HP Plotter (optional IBM" AT compatibli

\

~

.._-----

c -

-

Single keystroke operation

U

30 day guarantee of satisfaction 1OW to 1 SIN at 2000 cm-'

* Full mid-IR spectral coverage

(optional FAR-IR)

CIRCLE 18 ON READER SERVICE CAR0 ~.~ ~~

~

ANALYTICAL CHEMISTRY. VOL. 59, NO. 23. DECEMBER 1. 1987

1365A

“=duce and organize the correlations between tandem mass spectral features and substructures. Although MS/MS instruments have been in use for more than 10 years, there is still no general database of MS/MS spectra or any agreement on standard conditions for collecting such spectra (19). A major difficulty in the interpretation of mass spectra is that the products of all the fragmentation processes are overlapped in a mass spectrum. Electron impact ionization imparts ions with excess energy. These ions can then undergo fragmentation within the ion source, and the subsequent ionmolecule reactions and decompositions can often give a wide variety of products. Rearrangements furtber complicate interpretation. A mass spectrum indicates only the presence of ions and gives no information about their parentage, Isotopic labeling is required to determine parenedaughter relationships from mass spectra. MS/MS has several advantages over conventional MS for structure elucidation, the most obvious of which is the second dimension of information. Three types of features can be derived from an MS/MS data space: specific daughter ions, neutral losses, and parent-to-daughter transitions. The parenedaugbter relationships and neutral losses can thus be determined directly rather than inferred. Considering the mass range of 1500 amu, the

MS/MS data space yields 125,750 potential features (500 possible daughter ion masses, 500 possible neutral losses, and 124,750possible parent-to-daughter transitions), and the corresponding MS data space yields only 500. With higher resolution MS/MS instruments, the number of potential features increases even further. Not only are certain features in the MS/MS database more specific than individual mass spectral features, but many more specific combinations of features are possible from MS/MS data than from MS data alone. Thus by using combinations of MS/MS features, the template for the particular contribution that a substructure makes to the mass spectral data set can be more adequately specified. By selecting a parent ion and fragmenting it, information about an isolated portion of a molecule can he obtained. Thus conceptually it is reasonable to expect that parts of a molecular structure can he identified from characteristic features within the MS/MS data space. Correlating hlS/MS features with SUbstruetUreS

An extension of spectral matching to MS/MS has been reported hy C r w and Enke (20). In this work, individual daughter spectra were correlated with specific substructures. The presence of substructures in unknowns was determined by matching daughter spectra

I

gure 2. Components and data pathways for the Automated Chemical structure Eirc cidation System (ACES), including the Triple Quadrupole Mass Spectrometer (TQMS), the Method for Analyzing Patterns in Spectra (MAPS), the Empirical Formula Generator (EFG), the constrained structure generator (GENOA), and a routine to organize the output of GENOA into groups based on known and unknown pations of the complete structure (STRCHK). Dashed lines represent the learning mod%;soild lines represent the identificationmode. 1366A

ANALYTICAL CHEMISTRY, VOL. 59, NO. 23, DECEMBER 1, 1987

from an unknown against a database of reference daughter spectra. This method still suffers from drawbacks inherent to spectral matching and does not take full advantage of the extra dimension of information that MS/MS affords. For instance, the carbonyl substructure gives rise to a characteristic neutral l w of 28 amu, which may be seen in daughter spectra of any ionic substructures that contain this moiety. We have developed a computer method that automatically searches for and identifies the relationships between MS and MS/MS spectral features and substructures. This algorithm, the Method for Analyzing Patterns in Spectra (MAPS), assumes that much information lies within patterns of features in MSn spectra, and not just in the presence of individual masses or neutral losses. A more complete description of this software will appear elsewhere (21).MAPS expresses the relationships between MS and MS/MS spectral features and substructures in the form of production rules that may then be used to help identify the presence or absence of substructures in unknown compounds. A database of a few thousand rules could in theory be used to identify the structures of millions of compounds. ACES using MS and MSlMS data Several artificial intelligence and machine learning methodologies are being developed in this laboratory for automatic structure elucidation from MS and MS/MS data. Together they form an integrated set of software tools known as the Automated Chemical structure Elucidation System (ACES). The individual components and data pathways of this system are shown in Figure 2. A triple quadrupole mass spectrometer (TQMS) is included as the source of MS/MS data. The MAPS software operates in two different modes: the learning mode (dashed line) and the identification mode (solid line). In the learning mode, MAPS identifies the relationships hetween substructures and the characteristic features they produce in the MS and MS/MS data spaces using data from known compounds and stores these in the form of rules. In the identification mode, spectra from an unknown are searched for the diagnostic features contained in the rules. The substructures identified as present or absent by MAPS can then be used as constraints for an empirical formula generator (EFG) and for a structure generator (GENOA). The main assumption behind this system is that if enough substructures can be identified as present or absent in an unknown, the

FUCUS

1

Figure 3. Schematic of the rule generation procedure in MAPS.

complete structure can be determined. The function and state of development of each of the main components in this system are described below. MAPS. This software was developed in InterLISP-D on a Xerox 1108 AI workstation. MAPS deduces the relationships between substructures and the characteristic features they produce in the MS and MS/MS data space without prior assumptions regarding fragmentation pathways. These relationships may then he used to determine the presence and absence of substructures in unknown compounds not

requiring that these spectra be in the database. MAPS uses supervised learning to formulate the rules as shown in Figure 3. First, MS and MS/MS data are obtained for a set of known compounds; these data comprise the training set. From this, the "feature bucket" and "substructure bucket" data structures are created to facilitate the next step: , correlation of features with substructures. A minimum level of correlation is specified in this stage in the rule generation. Because each spectral feature has some level of correlation with a

UtrtNUAtjLt LUNI INUl I Y

substructure, this minimum level of correlation affects the number of features in the rules. Chemical knowledge is then used to filter out spurious features from the rules. These filters include the minimum and maximum fragment masses that can logically be attributed to each substructure and constraints to define legal fragment masses and compositions based on the elemental composition of the substructure and rules of valence. This process results in inclusion and exclusion rules that predict the presence and absence of substructures, respectively. +

1

.

SHAKERBATH

RECIPROCATING SHAKERS

ACCENT ON PERFORMANCE WARING CONTAINERS

Cal No. 8017 wilh 8020 Cmlainer

BLENDER POWER UNnu ( p PROOF ~ ~ ~

Hdlav spndle w W utilllsd in EbwbKh Slnw lins muw addibDnal gripping d aglm and lsciiilales changing cmlmrs. AdIMe horn lno IhN 1110 hp.

Cat No. 8581

pnsl

U"l

WNO. Mrn

WNO.

8sm

Eberbach ANALYTICAL CHEMISTRY, VOL. 59, NO. 23, DECEMBER I. 1987

1367~

~

Labeled compounds from Merck Sharp & Dohme/Isotopes.

, \

\

MSD Isotopes - research products that work, bringing results to researchers in biology, chemistry, physics, medicine and related fields.

that even if you require a compound which is not available from us immediately 'off-the shelf, we have the world's best facilities to custom synthesize it for you?

Did you know that we now offer thousands of compounds labeled with deuterium, carbon 13, nitrogen 15 and other stable isotopes too? And did you know

Just call or write to us for more information on what we have in stock and can make for you.

-

MSD W8D ISOTOPES

OlVlSlON OF MERCK FROSST CANAOA INC

Montreal Canada

For technical information call 1-800-361-0460.

.. l '

WFm corn

EAST C O N & CENTRAL

PO.Box 2951

CA~GADA

4545 Oleatha Avcnue St. Lnuis. MO 63116 Outside state of MO: (800) 325.9034 Slate of Missouri: (314)353-7000

PO.Box 899

Terminal Annex

La Angeler. CA 90051 Los Angcler A m : (213)723-9521 Outside State of CA:(800)4234977 Stafe ofCA: (Roo) 372.6454

ClRUE 94 ON READER SERVICE CARD

Pointe ClairciDorval. Quebec Canada H9R4P7

Telephone: (514)697-2823

An example of one such rule is the inclusion rule for the benzyl substructure shown in the box. Note that MAPS provides the level of correlation between the feature and the snbstructure (the fraction enclosed in brackets) and also formulates plausible fragment formulae for the fragment masses represented by each feature in the rule. Intensities in MS/MS scans are dependent on a large nnmher of instrumental parameters and are not as important as the presence or absence of a feature; hence they play a decreased role in identifying the presence and absence of substructures (as opposed to most spectral-matching techniques). Tberefore intensity classes rather than numerical intensities were used in the rules. Three such intensity classes are currently recognized strong, medium, and weak, which correspond to relative intensities of 10-100%,1-lo%, and 0196, respectively. Incorporation of intensities and implementation of a fuzzy logic matching algorithm were found to improve the performanee of the rules. If a compound provides greater than a certain percentage of features from an inclusion rule, that substructure is said to be present. The predictive capabilities of the rules were ascertained by applying them to the training set compounds, and thus the two categories of results are indicated correct and incorrect predictions. Rule performance can be ascertained a t several different levels of correlation or “match factors.” This is shown for the benzyl rule in Table I. Different criteria exist for predicting the presence and absence of substrnctures. One can expect that certain features will be present in the MS and MS/MS data space whenever a specific substructure is present in compounds analyzed under similar instrumental conditions. Similarly, the absence of these features suggests the absence of the substructure. Thus rules for excluding a substructure should ideally contain features that correlate strongly with the presence of each substructure. These features may not be very useful, however, in rules for predicting the

presence of that substructure if they also have moderate-to-high correlations with other Substructures as well, because they produce incorrect predictions of the presence of Substructures (false positives). Uniqueness factors were calculated for each feature in several inclusion rules (22). The uniqueness factor is defined as the ratio of the number of occurrences of afeature for a substructure to the number of occurrences of that feature in the database. It was found that MS/MS features generally had higher uniqueness factors than MS features and increased the reliability of the rules. T o minimize false positives, the inclusion rules should contain features or combinations of features that have a high uniqueness factor for each substructure. MAPS currently uses simple matching of features in the rules against the features from the MS and MS/MS data set from compounds in the training set to ascertain the predictive capabilities of the rules. Further optimization of the rules, implementation of more sophisticated matching, and continued expansion of the training set have resulted in improved rule performance. EFG. Software bas been developed to determine the possible empirical formulae for unknown compounds nsing medium resolution (0.1-1.0 amu resolution) MS and MS/MS data (23). Direct determination of empirical formulae cannot be accomplished by using such data because many formulae are often consistent with the molecular

lncludon rule ror rhe benzyl wasrmcAre If [28/32] strong intensity daughter ion at m/z 51 and [30/32] strong intensity daughter ion at m/z77 and [25/32] strong intensity neutral loss of 2 amu and [25/32] medium intensity neutral loss of 26 amu and [30/32] medium intensity neutral loss of 28 amu and [30/32] strong intensity neutral loss of 28 amu and [27/32] strong intensity daughter of m/z 51 fr- m/z 77

weight. In the empirical formula generator we have developed, MS and MS/ MS data are used to develop constraints on the elemental composition of an unknown and thus reduce the list of empirical formulae. Constraints can be developed from daughter spectra of isotopic molecular ions (241, from isotopic clusters from conventional mass soectra. and from substructures identified by MAPS. GENOA. The substructures identified by the application of MAPS rules or by other means can be used to formulate constraints for GENOA, a constrained structure generator developed during the course of the DENDRAL project. Substructural constraints take the form of a substructure definition and the number of occurrences (e.g., “constraint benzyl a t least 1”). GENOA allows for overlapping substructures; each substructure does not have to encompass a unique portion of a molecule. The presence of alternate substructures (Le., either substructure A or substructure B ) can also be specified. Negative information (substructures known to be absent) can also be used to constrain the structure generation. Results from the application of exclusion rules are used to trim the list of candidate structures. Given an empirical formula and substructural constraints, GENOA produces all possible nonredundant structures. Most importantly, the structure of the unknown will always be contained within the set of structures produced by GENOA using correctly identified substructures as constraints. Structure elucidation methods that rely on spectral matching cannot guarantee that the list of “closest hits” will contain the structure of the unknown. Nor can they guarantee that the list of “closest hits” will reliably reflect the substructures (or functional groups) present in the unknown because these substructures are not directly taken into account. An identification procedure that uses known substructure information should therefore be inher-

ANALYTICAL CHEMISTRY, VOL. 59, NO. 23, DECEMBER 1, 1987

136SA

FUCUS enuy more reiianie than spectral matching methods for true unknowns. The commercial GENOA software package includes a program called STRCHK (shown in Figure 2) that performs substructure searching. Given the structure of a compound and a library of predefined substructures, this procedure provides a list of substructures contained in the compound. These data are used by the MAPS software in developing spectral feature and substructure correlations. GENOA is being modified in this laboratory to better suit the purposes of ACES. A GENOA session originally required a single empirical formula. However, several formulae are often consistent with elemental composition data. Therefore additional software is being developed to automatically run GENOA structure generation sessions for each empirical formula. These sessions may use alternate empirical formulae and substructural constraints and will provide the user with the appropriate set of candidate structures. Modifications to STRCHK are being

made to provide automated substructure searching of a library of structures for training set compounds. The library of predefined substructures used for this purpose is also being constructed. In addition, STRCHK provides a method for organizing candidate structures into groups based on discriminating substructures and thus assists the user in determining what portions of the complete structure are unidentified.

A simple example To illustrate how these three software tools interact, di-n-octyl phthalate was treated as an unknown. This compound has a molecular weight of 390 and an empirical formula of C24H3~04. When only the empirical formula was used as a constraint, GENOA produced more than 5000 structures before the program exceeded the memory capabilities of the computer. MS and MS/ MS spectra from this compound were fed into MAPS to obtain a list of substructures likely to be present in the unknown. The following substructures

Future pmpects Complete structure elucidation of unknowns is not always possible, necessary, or desirable. The analytical requirement may not be the complete identity of a sample compound. For large-molecular-weight species such as those of biological importance, determination of key substructures may be sufficient and will be made possible hy using the kind of empirically based correlation techniques currently being developed. The ultimate goal of this work is to produce an intelligent system for structure elucidation that includes the TQMS in its feedback loop (shown in Figure 4). The TQMS will carry out diagnostic and confirmatory experiments, each time feeding its results back through “expert” interpretive tools to the user. The automated integration of these software tools is still being developed. The ACES approach is not limited to MS; it can develop rules from both known expertise and empirically derived correlations and can be extended to other multidimensional techniques to provide greatly enhanced diagnostic power for structure elucidation.

1. An e 1370A

ANALYTICAL CHEMISTRY, VOL. 59,

are quickly indicated as being present with match factors of greater than 50% benzoyl, methyl, ethyl, butyl, propyl, pentyl, hexyl, heptyl, octyl, carbonyl, carboxyl, ester, phthalate, phthalate ester, x-phenyl, and 1,2-phenyl. This information was used to formulate constraints for the EFG as well as for GENOA, because the empirical formula of substructures found will necessarily restrict the number and types of atoms present in the unknown. A molecular weight of 390 and the presence of the phthalate ester (CaH404)and the octyl (CaHi7) substructures were used as constraints for the EFG (i.e., a t least C16H2104).The resulting two empirical formulae produced were C24H3804and CzdHzzOs. The latter empirical formula was rejected by GENOA because it could not incorporate the suhstructures identified by MAPS. When the substructures identified by MAPS and the former empirical formula were used as constraints for GENOA, it took just a few minutes to indicate that only 89 possible structures exist. These represented the 89 isomeric possibilities for the second octyl substructure. If a future system could tell that there were two n-octyl groups in this molecule, then only one structure would be possible, that of di-n-oetyl phthalate. When dimethylaniline was treated in the same fashion, an unambiguous identification of the structure resulted.

NO. 23. DECEMBER 1, 1987

crneus (15) Buchanan, B. G.; Smith, D. H.; White,

Reterences (1) Hirschfeld, T. Anal. Chem. 1980, 52, 297 A. (2) Holland, J. F., Enke, C. G.; Allison, J.; Stults, J. T.; Pinkston, J. D.; Newcome, B.; Watson, J. T. Anal. Chem. 1983, 55, 997 A. (3) Johns0n.J. V.:Yost.R. A. Anal. Chem. _ ~ ~~~~~~~~~, , , 1985,57,758 A. (4) Eekenrode, B. E.; Watson, J. T.; Holland, J.; Enke, C. G. Int. J. Mass Spectrom. I o n Process., in press. (5) Laukien, F. H. Abstracts of Papers, 14th Annual FACSS Meeting, Detroit, Mich.; American Chemical Society, Washington, D.C., 1987; Abstract 524. (6) Wilkins. C. L. Anal. Chem. 1987. 59. 571 A (7) Van Dslen, P Presented at the French Srwntrfir and Technical Press at MesuCora Physque. Paris. France, December ~

-Proceedings Ot the 1987EPAIAW Symposium on

~~~.

,"*C

IIIYY.

Artificial lntelligenceMaking Machines Think; Tab Books:

(8) Graham, N.

Blue Ridge Summit, Pa., 1979. (9) McDermott, J. AI Magazine 1982, 2, 9.7

(l;?Small, G. W. Anal. Chem. 1987, 59, 535 A. (11) Rasmussen. G. T.: Isenhour. T. L. J. dhem. lnf. Cohput. Sci. 1979,19,179. (12) Martinsen, D.P.; Song, B.H. Mass Spectram. Rev. 1985,4,461. (13) . ~,McLaffertv. F. W.: Stauffer. D. B. J. Chem. inf. Comput. Sci. 1985, is,245. (14) Barr, A,; Feigenbaum, E.A. The Handbook of Artificial Intelligence; Heuristeeh Stanford, 1982; Val. 11.

W. C.: Gntter. R. J.: Feieenbaum. E. A,: Ledeibera, J.:'Dierassi. J. Am.'Chem: SOC. 1976;98wV. 6168. r16) Carhart, R. E.; Smith, 1): H.; Gray, X.A.R.: Nourse..I. G.: Dierasai. C. J. Am.

c.

Chem. SOC.mi. 46, i168.

(17) MeLafferty, F. W. Interpretation of Mass Spectra; University Science Books: Mill Valley, Calif., 1980. (18) MeLafferty, F. W.; Venkatnraghavan, R. Mass Spectral Correlations;American Chemical Society: Washington, D.C., 1982. (19) Martinez, R. I.; Cooks, R. G. Towards

Building

a

Practical MSIMS Database;

35th Annual Conference on Mass Spectrometry and Allied Topics; Denver, Colo., June 1987, pp. 117&76. (20) Cross, K.P.; Enke, C. G. Cornput. Chem. 1986,10,175. (21) Wade,A. P.; Palmer, P. T.; Hart, K.J.; Enke, C. G., submitted for publication in

Anal, Chim. Acta.

(22) Hart,K.J.;Wade,A.P.;Palmer,P.T.; Enke, C. G., submitted for publication in

A d . Chim. Acta.

(23) Palmer, P. T.; Enke, C. G., submitted for oublication in Int. J. Mass SDeetrom.

Ion'Proe.

(24) Bozor adeh, M. H.; Morgan, R. P.; Beynon, Y H . Annlyst 1978,103,613. This work was funded by grant GM-28254 from the National Institutes of Health. Thanks are due toFinnigan MAT for theuseofandsupport forthe Xerox 1108 AI Workstation. The authors also thank Molecular Design Ltd. for assisting in the acquisition and modification of GENOA.

ence sponsored by APCA and US. EPA's Environmental Monitoring Systems Laboratory. Measurement of Semi Volatiles in Ambient Air Indoor Toxic Air Contaminants Acidic Deposition Volatile Organic PoIIutants in Ambient Air * PhysicallChemlcal Propenles of Toxlcs * Dry Deposition-Methods Cornparisor Atmospheric Transformations ChernometricslData Analysts Hazardous Waste Emissions Integrated Air Cancer Project Source Monitoring Environmental Ouality Assurance Reviewed and approved for publication in accordance with U 5 EPA policy this 775-paqe landmark work I S now available t ; o i APCA APCA Publications P.O. Box 2861 Pittsburgh, PA 15230 Phone (412) 232-3444

Christie G. E n k e is a professor of chemistry at Michigan State University (MSU). H e received his Ph.D. from the University of Illinois in 1959. H e is coauthor of Electronics and Instrumentation for Scientists, codirector of the NIHI M S U Mass Spectrometry Facility, and co-inventor of the TQMS. His research interests are in the broad area of computer applications i n chemical analysis. Adrian P. Wade received his Ph.D. from the University of Wales in 1985 and is now a n assistant professor of analytical chemistry at the University of British Columbia. I n 1985 he was awarded the Harry Hallam Memorial Prize. Peter T. Pdlmer received his B.S. degree from Canisius College i n 1983.H e is a graduate student at M S U and is also a Quill Fellow. His research interests are automated structure elucidation and intelligent instrumentation. Kevin J. Hart received his B.S. f r o m the University of Notre Dame i n 1984. H e is a graduate student at MSU. His interests include mass spectrometry and computer applications in chemistry.

Order code VIP-8A Price 5 70 00. APCA Members 545 00 Please send prepayment NameCompany Address

0 Send VIP-8A (payment enclosed)

0 Send APCA CIRCLE 2

Publications Catalog

ON READER SERVICE CARD

ANALYTICAL CHEMISTRY, VOL. 59, NO. 23, DECEMBER 1, 1987

1371A