From: Subject: Textual and chemical information processing: different domains but similar algorithms Date: Fri, 15 Oct 2004 17:11:10 +0100 MIME-Version: 1.0 Content-Type: multipart/related; boundary="----=_NextPart_000_0000_01C4B2D9.F7DD9D80"; type="text/html" X-MimeOLE: Produced By Microsoft MimeOLE V6.00.2800.1441 This is a multi-part message in MIME format. ------=_NextPart_000_0000_01C4B2D9.F7DD9D80 Content-Type: text/html; charset="Windows-1252" Content-Transfer-Encoding: quoted-printable Content-Location: http://informationr.net/ir/5-2/paper69.html Textual and chemical information processing: = different domains but similar algorithms vice = versa. Applications discussed include: an algorithm for = distribution sorting that has been applied to the design of screening = systems for rapid chemical substructure searching; the use of measures = of inter-molecular structural similarity for the analysis of hypertext = graphs; a genetic algorithm for calculating term weights for relevance = feedback searching for determining whether a molecule is likely to = exhibit biological activity; and the use of data fusion to combine the = results of different chemical similarity searches."=20 name=3Ddescription> vice = versa. Applications discussed include: an algorithm for = distribution sorting that has been applied to the design of screening = systems for rapid chemical substructure searching; the use of measures = of inter-molecular structural similarity for the analysis of hypertext = graphs; a genetic algorithm for calculating term weights for relevance = feedback searching for determining whether a molecule is likely to = exhibit biological activity; and the use of data fusion to combine the = results of different chemical similarity searches."=20 name=3DDC.Description>

Information Research, Vol. 5 No. 2, January = 2000

3D"" 3D"" 3D"" 3D""

Textual and chemical information processing: different domains but = similar=20 algorithms


Peter = Willett
Department=20 of Information Studies and
Krebs Institute for Biomolecular=20 Research
University of Sheffield
Sheffield S10 2TN, = UK



Abstract
This paper discusses the extent to which algorithms = developed for=20 the processing of textual databases are also applicable to the = processing of=20 chemical structure databases, and vice versa. = Applications=20 discussed include: an algorithm for distribution sorting that has been = applied=20 to the design of screening systems for rapid chemical substructure = searching;=20 the use of measures of inter-molecular structural similarity for the = analysis=20 of hypertext graphs; a genetic algorithm for calculating term weights = for=20 relevance feedback searching for determining whether a molecule is = likely to=20 exhibit biological activity; and the use of data fusion to combine the = results=20 of different chemical similarity searches.


Introduction

Over the years, researchers in information retrieval (IR) have = developed many=20 different techniques for the processing of databases of textual = information.=20 Established examples of such techniques include document clustering, = relevance=20 feedback, stemming algorithms and text compression, and there are many = other=20 emerging applications, such as categorisation, event detection and = information=20 filtering (Spark Jones and Willett, 1997). One such application is = multimedia=20 retrieval, where there is much current interest in extending algorithms = and data=20 structures developed for processing textual databases, for the storage = and=20 retrieval of speech, image and video data (Maybury, 1997). This paper = argues=20 that at least some of the algorithms that are used in textual = information=20 retrieval can also be applied to another type of data, viz the=20 two-dimensional (2D) and three-dimensional (3D) chemical structure data = that=20 forms one of the principal components of chemical information systems = (Ash et=20 al., 1991). These systems were first developed principally for = archival=20 purposes, but now play an important role in research programmes to = discover=20 novel bioactive molecules for the pharmaceutical and agrochemical = industries=20 (Martin and Willett, 1998).

The University of Sheffield has been involved in research on both = chemical=20 and textual information retrieval for many years (Lynch and Willett, = 1986).=20 These studies have led us to believe that both areas have much to offer, = with=20 research in one providing a fertile source of ideas for research in the = other.=20 In some cases, the relationship is obvious, with algorithms and data = structures=20 being transferable with little or no change from one application to = another;=20 while in other cases, the relationship is less direct, involving more a = general=20 view of the sorts of information processing techniques that are = required, rather=20 than a direct transfer of technology. Here, we consider the former, more = direct,=20 type of relationship.

There are clear similarities in the ways that chemical and textual = database=20 records are characterised. The documents in a text database are each = typically=20 indexed by some small number of keywords, in just the same way as the 2D = or 3D=20 molecular representations in a chemical database are each characterised = by some=20 small number of substructural features chosen from a much larger number = of=20 potential attributes (as discussed further in the next section of this = paper).=20 Moreover, both types of attribute follow a well-marked Zipfian = distribution,=20 with the skewed distributions that characterise the frequencies of = occurrence of=20 characters, character substrings and words in text databases being = mirrored by=20 the comparable distributions for the frequencies of chemical moieties. = Thus, the=20 overwhelming majority of all of the many millions of molecules that have = ever=20 been made contain the element carbon but even the tenth most frequent = element,=20 iodine, occurs only about one thousandth as frequently, with the great = majority=20 of the elements having vanishingly small frequencies of occurrence; = similar=20 distributions are observed for other types of chemical substructure = (Lynch,=20 1977). These shared characteristics mean that the two types of database = are=20 amenable to efficient processing using the same type of file structure. = Thus one=20 of the first examples of what would now be referred to as text signature = searching (Barton et al., 1974) arose from previous studies of = chemical=20 bit-string processing (Adamson et al., 1973), and similar = comments apply=20 to the use of an inverted file for rapid database clustering (Willett, = 1981,=20 1982). Finally, in just the same way as a document either is, or is not, = relevant to some particular user query, so a molecule is active, or is = not=20 active, in some particular biological test, thus allowing comparable = performance=20 measures to be used to assess search effectiveness in the two types of = retrieval=20 system (Edgar et al., 1999).

These similarities mean that it is often possible to apply similar = algorithms=20 to the two different sorts of database, as we describe in some detail = below.=20 That said, there are obvious differences, most obviously in the = semantics of the=20 representations that are used. A 2D chemical structure diagram bears a = much=20 closer relationship to the molecule it describes than does the set of = words=20 comprising a textual document, and this relationship is still stronger = when, as=20 is increasingly the case, a 3D structure with XYZ atomic co-ordinate = data is=20 available for a molecule (Martin and Willett, 1998). Both the structure = diagram=20 and the co-ordinates can be regarded as direct manifestations of the = underlying=20 wave equations that describe a molecule, and it has thus proved possible = to=20 develop powerful simulation techniques to predict the activities and = properties=20 of molecules from a knowledge of their 2D or 3D structure. Many of these = molecular modelling tools have no direct textual equivalent, as the use = of=20 natural language raises a host of linguistic problems that do not arise = in the=20 chemical context (although it should be noted that linguistic parsing = and=20 recognition techniques can be used to represent and search the generic = chemical=20 structures that occur in chemical patents (Barnard et al., 1984; = Welford=20 et al., 1981)).

The remainder of this paper discusses several applications to support = the=20 belief that there may be at least some degree of overlap between the = techniques=20 used to process chemical and textual databases. The first application, = which is=20 taken from a previous paper discussing a potential relationship between = these=20 two domains (Willett, 1997), is that of chemical substructure searching; = this=20 section also introduces the basic components of chemical information = systems,=20 thus providing some of the necessary background for the more recent = applications=20 discussed in the following sections.

Design of screening systems for chemical substructure searching

Chemical information systems have historically represented molecules = by means=20 of their 2D chemical structure diagrams: these are encoded as labelled = graphs in=20 which the nodes and edges of a graph encode the atoms and bonds of a = molecule=20 (Ash et al., 1991). Searches for substances that contain a = specific=20 partial structure, such as a cephalosporin ring, can then be effected by = graph-matching techniques that permit the identification of all = occurrences of=20 the user. s query substructure in each database structure (Barnard, = 1993). The=20 time-consuming nature of these subgraph isomorphism searches = means that=20 an initial screening stage is required that can eliminate the = great bulk=20 of the database from subsequent processing (in just the same way as a = text=20 signature search is used to reduce the amount of pattern matching that = is=20 required in serial searches of text databases). The question then arises = as to=20 what sorts of substructural characteristics should be used as indexing = keys in=20 the screening search.

The Zipfian distribution of the occurrences of the elements has been=20 mentioned previously, this implying a vast divergence in the = discriminatory=20 powers of searches of chemical databases that are based on elemental = type. Work=20 in Sheffield by Lynch and his co-workers in the early Seventies (see,=20 e.g., Adamson et al., 1973) showed that improved = discriminatory=20 power could be obtained by the use of more sophisticated indexing keys = that were=20 based on variably-sized chemical fragment substructures, = i.e.,=20 groups of atoms and bonds. The fragments were chosen so as to occur with = approximately equal frequencies of occurrence in the database that was = to be=20 searched, an idea that now forms the basis for many current systems for = 2D=20 substructure searching (Barnard, 1993).

It was soon realised that the concept of equifrequency was of great=20 generality, and this led to an extended investigation of the application = of=20 equifrequency ideas to the searching and processing of text databases = (Lynch,=20 1977): here, we consider the studies that were carried out on the = approach to=20 sorting normally refererd to as distribution sorting (Cooper = et=20 al., 1980; Cooper and Lynch, 1984). The suggested approach involves = sorting=20 a very large file in two stages: first, an initial, approximate sorting = stage=20 that sub-divides the file into an ordered series of sub-files; and then = an exact=20 sort of each of the sub-files. To those of us familiar with the old days = when=20 catalogue cards had to be sorted by hand, the initial stage is analogous = to=20 dividing the cards into the sub-files A to D, E to K, L to R and S to Z. = In the=20 present context, each of these rough pigeonholes represents a separate = sub-file=20 on disk and each of the ranges is chosen to ensure that approximately = equal=20 numbers of records are allocated to each such sub-file, with the aim of=20 maximising the efficiency of the final, in-core sorts of these = sub-files.

The 3D structure of a molecule plays a vital role in determining its=20 biological activity. The "lock-and-key" theory suggests that a molecule = may be=20 able to act as a drug if it can fit into the active site of a protein in = much=20 the same way as a key fits into a lock (Martin and Willett, 1998), and = there is=20 thus great interest in being able to identify molecules that are = appropriately=20 "key-shaped". This is done by searching for pharmacophores, = i.e.,=20 the patterns of atoms in 3D space that are thought to be responsible for = a=20 molecule binding to an active site in a protein molecule. When our = studies of=20 pharmacophore searching started in the mid-Eighties, it was soon = realised that=20 the graph-theoretic methods developed for 2D substructure searching were = also=20 applicable here, with a 3D molecule being described by a graph in which = the=20 nodes were the atoms and the edges were the inter-atomic distances = (Willett,=20 1995). However, the implementation of subgraph isomorphism matching on = such 3D=20 chemical graphs is still more demanding of computational resources than = is 2D=20 substructure searching, with a consequent need for the development of = efficient=20 methods of screening.

The graphs representing 2D chemical molecules are composed of atoms = and=20 bonds, and this is reflected in the compositions of the screens that are = used=20 for 2D substructure searching. It thus seemed appropriate to investigate = screens=20 for 3D searching based on the information contained in 3D chemical = graphs,=20 i.e., atoms and inter-atomic distances, and we decided to = evaluate=20 screens consisting of a pair of atoms and an inter-atomic distance = range. Thus,=20 a screen might represent the presence within a molecule of an oxygen = atom and a=20 nitrogen atom separated by a distance range, e.g., between 5 and = 7=20 =C5ngstroms. Initial experiments demonstrated the vastly skewed = distributions of=20 distances that characterise databases of typical 3D molecules and there = was thus=20 a need to identify such inter-atomic screens so that they occurred with=20 approximately equal frequencies of occurrence (Jakes and Willett, 1986). =

The approach adopted involved varying the width of the distance range = associated with each pair of atoms, so that highly populated parts of = the=20 distance distribution were associated with very narrow ranges, while = less=20 populated parts of this distribution, or very infrequently occurring = pairs of=20 atoms, were associated with more extended ranges. The algorithm that was = finally=20 developed was a simple modification of the distribution sorting = algorithm=20 described above with, in essence, the distance ranges here corresponding = to the=20 character ranges that underlie the text-sorting application (Cringean = et=20 al., 1990). This proved to be both effective and efficient, and = screens=20 based on the distances between pairs of atoms now form the basis for = nearly all=20 existing systems for 3D substructure searching; indeed, the same basic=20 methodology can be used to characterise the valence or dihedral angular=20 relationships that exist between sets of three or four atoms, = respectively, thus=20 allowing database searches to be carried out that involve the = specification of=20 both distance and angular information. More recently, we have = demonstrated that=20 analogous procedures can be used to search databases where account is = taken of=20 the fact that most 3D molecules are not completely rigid, but are = actually in a=20 state of constant flux (so that the "keys" referred to previously might = be=20 considered as being more akin to a jelly than to a piece of rigid metal) = (Willett, 1995). This further application demonstrates clearly the = generality of=20 equifrequency-based approaches for database processing.

Manipulation of hypertext graphs using measures of inter-molecular=20 structural similarity.

Text retrieval systems were initially based upon the Boolean = retrieval model,=20 but the systems were soon extended to permit best match searching (in = which the=20 documents are ranked in decreasing order of similarity to the query) = (Spark=20 Jones and Willett, 1997). A similar progression has occurred in chemical = information systems, with the substructure searching systems described = above=20 increasingly being complemented with facilities for what is referred to = as=20 similarity searching. This generally involves the specification = of an=20 entire query molecule, the target structure, rather than the = partial=20 structure that is required for substructure searching. The target is=20 characterised by one or more structural descriptors that are compared = with the=20 corresponding sets of descriptors for each of the molecules in the = database to=20 find those nearest neighbour molecules that are most similar to = the=20 target structure. Two near-contemporaneous studies in the mid-Eighties=20 demonstrated that counts of the numbers of fragment substructures common = to a=20 pair of molecules provided a computationally efficient, and surprisingly = effective, basis for quantifying the degree of structural resemblance = between=20 the two molecules under consideration (Carhart et al, 1985; = Willett et=20 al., 1986). Specifically, the use of a simple association = coefficient=20 (usually the Tanimoto coefficient) in conjunction with the lists of = screens=20 associated with the target structure and each of the database structures = provided a simple way of investigating inter-molecular structural = similarities.=20 Such fragment-based methods for similarity searching are now widely used = as a=20 complement to the established routines for substructure searching = (Willett=20 et al., 1998) and we have since demonstrated that these = measures=20 are also applicable to the comparison of textual, rather than chemical, = graphs.=20 Specifically, as described below, we have used these measures to = determine the=20 degree of consistency with which hypertext documents are created; other=20 relationships between textual and chemical similarity measures are = discussed by=20 Willett (1997).

The creation of the intra-document links between the individual = components of=20 a hypertext document is a difficult, and time-consuming, task, but one = in which=20 human intervention has commonly been thought necessary if the semantic=20 relationships that exist between the components of the document are to = be made=20 explicit. A similar view has prevailed for many years with regard to the = indexing of documents in IR systems, where the existence of = well-established=20 systems for automatic indexing has not prevented the widespread use of = trained=20 library and information specialists for indexing and classifying = documents prior=20 to their incorporation in an online database. The importance of the = indexing=20 task in IR has led to many studies of inter-indexer consistency,=20 i.e., of the extent to which agreement exists among different = indexers on=20 the sets of index terms to be assigned to individual documents. These = studies=20 (as reviewed, e.g., by Markey (1984)) have consistently concluded = that=20 recorded levels of consistency vary markedly, and that high levels of=20 consistency are rarely achieved; similar examples of manual = inconsistency are=20 provided by related tasks, such as query formulation and the assessment = of=20 document relevance (Salton, 1989). The insertion of links in hypertext = documents=20 may be viewed as being analogous to the assignment of index terms to = such=20 documents, and we hence undertook a study to determine the extent to = which=20 different people produce similar link structures for the same hypertext=20 documents (Ellis et al., 1994, 1996).

The hypertext documents were generated from five printed full-text = documents,=20 each a thesis, journal article or book written by a member of the = Department of=20 Information Studies at the University of Sheffield. Five copies were = made of=20 each of the chosen documents, and each of the twenty-five resulting = copies was=20 allocated to a different student volunteer from the Department. The = volunteers=20 were instructed in the use of an interactive system that allowed them to = create=20 explicit representations of links between paragraphs whose contents they = decided=20 were related, and thus to create hypertext versions of the source = documents.=20 These hypertexts were stored as graphs in which the nodes represented = portions=20 of a text (specifically paragraphs in our work but section-based or=20 sentence-based portions could also have been used), with an edge linking = a pair=20 of nodes if the human linker had created a link between the = corresponding=20 paragraphs. Pairs of these graphs (describing the same textual document = but=20 processed by different volunteers) were then compared using a range of=20 similarity measures, most of which were based on those used for chemical = similarity searching. For example, a commonly used type of chemical = fragment is=20 the augmented atom, which consists of an atom and the = atoms that=20 are bonded to it, and it is possible to generate a comparable hypertext = fragment=20 consisting of a paragraph and those paragraphs that are linked to it. A = measure=20 of the similarity between two hypertext graphs can then be calculated in = terms=20 of the number of such fragments that the hypertexts have in common, and = the=20 resulting similarity is taken as a measure of the degree of inter-linker = consistency.

Five hypertext versions were produced for each source document, = giving a=20 total of ten possible pairs of sets of hypertext links for that document = (i.e., fifty possible pairs for the entire dataset). Similarity = values=20 were calculated using many different combinations of graph = representation,=20 fragment descriptor and similarity coefficient. When this was done, it = was=20 possible to draw a simple, unequivocal conclusion: that levels of = inter-linker=20 consistency are generally low, but can also be quite variable. It would = hence=20 seem that different people tend to have very different views of the = semantic=20 relationships that exist among the components of a full-text document, = with the=20 possibility that different people will tend to impose different = hypertext link=20 structures on the same source documents. This lack of agreement leads = one to=20 question the effectiveness of manually-assigned links in supporting the = browsing=20 strategies of subsequent hypertext readers, since if inter-linker = consistency is=20 found to be low, then linker/reader consistency (i.e., the = level=20 of agreement as to the semantic relationships that exist between the = components=20 of the text, and that therefore should explicitly be represented by = links that=20 the reader then has the opportunity to follow) can also be expected to = be low.=20 Indeed, our results provide some justification for the numerous and = continuing=20 efforts to develop techniques for the automatic generation of hypertext = links=20 (see, e.g., Crestani and Melucci, 1998) where the set of links = created by=20 a particular technique will be produced in a consistent manner, to which = users=20 might become accustomed as they browse and which might well be no worse = than=20 those created by human indxers.

A genetic algorithm for calculating relevance feedback and = substructural=20 analysis weights.

The calculation of weights that describe the importance of document = and query=20 terms is an important component of best-match text IR systems. The most=20 effective results are obtained from using relevance feedback information = (Robertson and Sparck Jones, 1976; Salton and Buckley, 1990) but efforts = continue to develop new weighting schemes that might further increase = search=20 effectiveness. This striving for improved performance has raised the = question of=20 what is the upperbound, i.e., the maximum possible level, = that can=20 be achieved by a particular retrieval strategy. We have hence developed = a=20 genetic algorithm (hereafter a GA) to investigate the upperbound that = can be=20 achieved by ranked-output retrieval systems that employ weighted query = terms.=20 Indeed, the GA was developed not only for this purpose but also for = calculating=20 weights reflecting the extent to which particular types of molecular = feature=20 contribute to the biological activity of molecules (using the analogies = between=20 indexing terms and structural features, and between query relevance and=20 biological activity that we have noted in the introduction to this = paper). We=20 will illustrate the operation of this GA by discussing the textual = application=20 (Robertson and Willett, 1996) before proceeding to an account of the = results=20 obtained in our chemical experiments (Gillet et al., 1998).

A GA operates on chromosomes, which are linearly-encoded=20 representations of possible solutions to the problem under = investigation, with=20 each element of a chromosome describing some particular component of the = encoded=20 solution. A fitness function is used to calculate a = numeric score=20 that measures how "good" a solution is represented by each chromosome. = The set=20 of chromosomes at any particular stage of the processing is referred to = as the=20 current population, and the chromosomes in this generation are = processed=20 by genetic operators to yield the population that = corresponds to=20 the subsequent generation. The two principal operators are = crossover,=20 which combines parts of the fittest chromosomes, and mutation, = which=20 introduces new information at random. The chromosomes in the new = generation are=20 evaluated by means of the fitness function and the genetic operators = invoked=20 again, this process continuing until convergence is achieved, = i.e, until=20 there is no increase in the average fitness of each successive = generation.

Each chromosome in our text GA contains a set of elements, with one = element=20 for each of the terms comprising the query and with each such element = containing=20 a weight for one of these terms. The weights for those query terms that = occur in=20 a particular document are summed, the documents are ranked in decreasing = order=20 of these sums of weights, and a cut-off is then applied to the ranking = to=20 retrieve some fixed number of the top-ranked documents. Our experiments = have=20 used standard document test collections for which full relevance data is = available, and it is hence possible to determine the recall of the = search by=20 noting the number of relevant documents that have been retrieved above = the=20 cut-off. These recall values are used as the fitness function for the = GA. Once=20 termination has occurred, i.e., the recall values have ceased to=20 increase, the term weights are noted for the chromosome that has the = largest=20 fitness. Thus, if the GA has successfully explored the full range of = possible=20 weights, then the final weights provide an upperbound to the retrieval=20 performance obtainable for that query using relevance weighting of = single=20 terms.

The experiments involved the following seven document test = collections: Keen=20 (800 documents and 63 queries), Cranfield (1400 documents and 225 = queries),=20 Harding (2472 documents and 65 queries), Evans (2542 documents and 39 = queries),=20 LISA (6004 documents and 35 queries), SMART (12694 documents and 77 = queries) and=20 UKCIS (27361 documents and 182 queries). The GA was run for each query = in each=20 collection, and the final set of weights used to calculate recall values = for the=20 top-10, top-20 and top-50 documents. The F4 retrospective relevance = weights of=20 Robertson and Spark Jones (1976) were used for comparison with the GA = weights=20 since the former have been found to provide a consistently high level of = performance in previous studies of relevance feedback searching.

Table 1 (below) summarises the results obtained when the top-20 = documents=20 were retrieved from a ranking; many other results are presented and = discussed by=20 Robertson and Willett (1996). An inspection of this table shows that (as = would=20 be expected) the GA weights generally perform better than the F4 = weights, in=20 terms of both average recall and numbers of queries where one was = superior to=20 the other. However, there are generally many searches where the two = approaches=20 give the same level of performance (to two decimal places in mean = recall), and=20 often at least a few queries (and a fair number in the case of the = Harding=20 collection) where the F4 search is better. Where there are differences, = in=20 either direction, between the two weights, these are overwhelmingly due = to one=20 of the weights finding a single additional relevant document above the = cut-off=20 position in the ranking: thus, the two approaches give very similar = levels of=20 performance. A study of those few queries where the F4 weights were=20 (unexpectedly) superior suggested that these queries were ones in which = at least=20 some of the F4 weights were negative (this corresponding to terms that = occur=20 frequently in a collection but only infrequently in the relevant = documents). The=20 initial version of the GA did not allow for the possibility of negative = weights:=20 when it was modified to allow for such occurrences, there was a = substantial=20 reduction in the (already small) number of queries where the F4 weights = were=20 more effective (Robertson and Willett, 1996). In general then, the F4 = weights=20 gave a level of performance that was only marginally inferior to those = provided=20 by the GA weights, which had been obtained by a detailed exploration of = the=20 term-weight space defined by the query terms. This being so, it seems = reasonable=20 to conclude that the F4 weights give a practicable upperbound to the = performance=20 that is achievable in relevance feedback searches of text databases.

Document=20 Mean Recall=20 Number Of Queries =
Collection=20 F4=20 GA=20 Same=20 F4 Better=20 GA Better
Keen 0.55 0.60 31 2 30
Cranfield 0.58 0.65 117 6 102
Harding 0.34 0.33 33 16 16
Evans 0.38 0.42 11 7 21
LISA 0.61 0.64 13 5 17
SMART 0.30 0.34 23 10 44
UKCIS 0.18 0.21 105 7 70

Table 1. Retrieval effectiveness in = relevance=20 feedback searches using F4 and GA weights.
[The Mean = Recall=20 portion of the table is the mean recall when averaged over all of the = queries=20 for a particular test collection using the two types of weight, while = the=20 right-hand portion shows the number of queries where the GA and F4 = weights gave=20 the same recall value (Same), and where either the F4 weights or the GA = weights=20 were better.]

The GA that was used for these relevance feedback experiments was = also=20 designed to calculate weights for substructural analysis, (Cramer = et=20 al., 1974). Here, weights are calculated that relate the presence of = a=20 molecular feature to the probability that that molecule is active in = some=20 biological test system (cf relating the presence of a specific = index term=20 in a document to the probability that that document is relevant to a = particular=20 query). Given some training set of compounds for which the biological = activities=20 are available, the aim of substructural analysis is to develop weights = that can=20 then be used to select new compounds for biological testing. = Specifically the=20 sum of weights is calculated for the fragment substructures present in a = molecule, and then the compounds are ranked in order of decreasing=20 sums-of-weights, so that chemical synthesis and biological assays can be = focused=20 on those compounds that have a high a priori probability of = activity.=20

The use of substructural analysis methods with fragment substructures = (such=20 as those discussed in the second section of this paper) is well = established,=20 with many different types of weighting scheme having been described in = the=20 literature (Ormerod et al., 1989). The project to be described = here used=20 a rather different level of molecular description, specifically = high-level=20 molecular characteristics suggested by medicinal chemists at = GlaxoWellcome=20 Research and Development (our collaborators in this project) as = affecting the=20 drug-like behaviour of molecules. The features are the distribution of = property=20 values for the molecular weight, the =B2k = a shape index (Kier, 1987), and the numbers of = aromatic=20 rings, rotatable bonds, hydrogen-bond donor atoms and hydrogen-bond = acceptor=20 atoms, in the molecules comprising a database. Each property was = allocated a=20 total of 20 bins; for example, the first bin for the hydrogen-bond donor = feature=20 describes those molecules in a dataset that have no donor atoms, the = second bin=20 those that have one donor atom, and so in until the 20th and last, which = represents those that have 19 or more donor atoms. The molecular weight = and=20 shape index bins contained ranges of values, rather than specific = integer=20 counts, e.g., the first two bins for molecular weight described = the=20 molecules with weights in the range 0-74.99 and 75-149.99, etc. =

The GA described previously was used to calculate weights for each = bin of=20 each feature, with each such weight reflected the extent to which = possession of=20 that feature-value combination resulted in a molecule having some = specific=20 activity. The fitness function was then based on the occurrences of = these=20 feature-value pairs in sets of molecules for which activity data are = available.=20 For this purpose, we used sets of structures extracted from the World = Drugs=20 Index (or WDI, available from Derwent Information at URL=20 http://www.derwent.co.uk) and SPRESI (available from Daylight = Chemical=20 Information Systems Inc. at http://www.daylight.com) databases: the = former=20 contains molecules for which biological activities have been established = while=20 the latter contains a large number of molecules that are assumed to be = inactive.=20 As with the text application, the chromosome in the GA encodes possible = weights=20 (in this case for each of the feature-value bins), and then the score = for a=20 particular molecule is the sum of the weights for those feature values = that are=20 associated with it. The fitness function of the GA is simply the number = of=20 top-ranked active molecules once the molecules have been ranked in = decreasing=20 order of these sums-of-weights.

Sets of 1000 molecules with some particular activity were chosen from = the WDI=20 and then combined with a set of 16,807 SPRESI structures. The GA was = used to=20 calculate weights for this combined dataset, and a note made of the = number of=20 actives in the top 1000 positions once convergence had occurred. The=20 effectiveness of the weights is measured by the degree of initial = enhancement, i.e., the ratio of the observed number of = top-1000=20 actives to the number of actives that would be obtained by selecting = 1000=20 compounds at random (Gillet et al., 1998). These results are = summarised=20 in Table 2 and it will be seen that the GA weights give consistently = better=20 results than merely selecting compounds at random for testing, although = the=20 extent of the improvement is clearly dependent upon the specific = activity under=20 investigation. Gillet et al. (1998) report comparable results = from=20 extended experiments using various forms of these weights on a range of=20 datasets; these involved not only retrospective studies (as here) but = also=20 experiments where the weights were used in a predictive manner. It was = concluded=20 that the GA provided a simple and effective way of ranking sets of = compounds in=20 order of decreasing probability of activity, thus enabling the = prioritisation of=20 compounds for synthesis and biological testing.

Drug Activity=20 Initial Enhancement
Hormones 8.3
Anti-cancers 7.6
Anti-microbials 7.6
Anaesthetics 4.1
Blood 3.6
Central nervous system 3.9
Psychotropics 2.5

Table 2. Effectiveness of ranking active compounds = using the=20 GA-based substructural analysis weights.
[The Drug = Activity=20 listed is that contained in the keyword activity field for each of the = compounds=20 in the WDI database.]

Combination of chemical similarity measures using data fusion

The many types of similarity measures that are available for the = measurement=20 of molecular similarity has led to comparative studies in which = researchers try=20 to identify a single, "best" measure, using some quantitative = performance=20 criterion. For example, Willett et al. (1986) discuss the merits = of=20 different association coefficient for chemical similarity searching and = conclude=20 by recommending the Tanimoto coefficient for this purpose. Such = comparisons, of=20 which there are many in the chemical information literature, are limited = in that=20 they assume, normally implicitly, that there is some specific type of = structural=20 feature (similarity coefficient, weighting scheme or whatever it is that = is=20 being investigated) that is uniquely well suited to describing the = type(s) of=20 biological activity that are being sought for in a similarity search. = The=20 assumption cannot be expected to be generally valid, given the = multi-faceted=20 nature of biological activities, and this has led us to consider = chemical=20 applications of data fusion (Hall, 1992).

Data fusion was developed to combine inputs from different sensors, = with the=20 expectation that using multiple information sources enables more = effective=20 decisions to be made than if just a single sensor was to be employed. = The=20 methods are used in a wide range of military, surveillance, medical and=20 production engineering applications (see, e.g., Arabnia and Zhu = (1998)):=20 our interest was aroused by a paper by Belkin et al. (1995), in = which=20 data fusion was used to combine the results of different searches of a = text=20 database, conducted in response to a single query but employing = different=20 indexing and searching strategies. A query was processed using different = strategies, each of which was used to ranking the database in order of=20 decreasing similarity with the query. The ranks for each of the = documents were=20 then combined using one of several different fusion rules, the output of = the=20 fusion rule was taken as the document. s new similarity score and the = fused=20 lists were then re-ranked in descending order of similarity.

Ginn et al. (1999) have described the application of these = ideas in=20 the context of matching a target structure against a database, using = several=20 different measures of chemical similarity as summarised in Figure 1.

  1. Execute a similarity search of a chemical database for some = particular=20 target structure using two, or more, different measures of = inter-molecular=20 similarity.=20

  2. Note the rank position, ri, of each database = structure=20 in the ranking resulting from use of the i-th similarity = measure.=20

  3. Combine the various rankings using a fusion rule to give a new = combined=20 score for each database structure=20

  4. Rank the resulting combined scores, and then use this ranking to = calculate=20 a quantitative measure of the effectiveness of the search for the = chosen=20 target structure.
Figure 1. Combination of similarity rankings using = data=20 fusion

The fusion rules used were those identified by Belkin et al. = (1995)=20 and shown in Figure 2, and the combined scores output by the fusion rule = are=20 then used to re-order the database structures to give the final ranked = output.=20 It will be seen from Figure 2 that the MIN and MAX rules represent the=20 assignment of extreme ranks to database structures and it is thus hardly = surprising that both can be highly sensitive to the presence of a single = "poor"=20 similarity measure amongst those that are being combined. The SUM rule, = where=20 each database structure is assigned the sum of all the rank positions at = which=20 it occurs in the input lists, is expected to be more stable against the = presence=20 of a single poor or noisy input ranking, and this was generally found to = be the=20 rule of choice in our experiments (Ginn et al., 1999).

Name

Fusion Rule

MIN

minimum (r1, r2 = ,&ri=20 &rn)

MAX

maximum (r1, = r2,&ri=20 &rn)

SUM

Figure 2: Fusion rules for combining n ranked = lists,=20 where ri denotes the rank position of a specific = database=20 structure in the i-th (1 =A3 i = =A3 n) ranked list.

Our experiments involved combining searches of several different = datasets=20 using several different similarity measures in each case. The dataset = considered=20 here contained 75 compounds selected by Kahn (1998) in a discussion of = various=20 types of structural descriptor, with each of the compounds belonging to = one of=20 14 well-defined biological activity classes (such as = angiotensin-converting=20 enzyme inhibitors and HIV-1 protease inhibitors). Six different types of = similarity measures were used in the experiments, as detailed by Ginn = et=20 al. (1999); for the present, we note merely that they encoded = information=20 about the steric, electrostatic and hydrophobic characteristics of = molecules=20 (similarity measures denoted by the symbols "F" and "J"), about the 3D=20 arrangement of pharmacophore points in molecules (denoted by the symbols = "3" or=20 "T") and about the occurrences of chains of up to 7 non-hydrogen atoms = (denoted=20 by the symbols "2" or "N"). The SUM rule was used to generate all = possible=20 combinations of rankings from these six similarity measures. Each of the = members=20 of the dataset was used as the target structure for a similarity search = and=20 Table 3 details the mean numbers of actives (i.e., molecules with = the=20 same activity as the target structure) found in the top-10 nearest = neighbours=20 when averaged over all 75 searches. The values of c at the top of = the=20 table denote the number of similarity measures that were fused (so that, = e.g., c=3D1 represents the original measures and = c=3D2=20 represents the fusion of a pair of the original measures) and a shaded = element=20 indicates a fused combination that is better than the best original = individual=20 measure (which was one of the 3D pharmacophore measures).

c=3D1=20 c=3D2=20 c=3D3=20 c=3D4=20 c=3D5=20 c=3D6
2> 0.80 23 1.10 23F 1.28 23FJ 1.52 23FJN 1.45 23FJNT 1.43
3 1.12 2F 1.04 23J 1.39 23FN 1.23 23FJT 1.69    
F 0.89 2J 1.01 23N 1.04 23FT 1.43 23FNT 1.36    
J 1.08 2N 0.68 23T 1.24 23JN 1.31 23JNT 1.43    
N 0.63 2T 0.95 2FJ 1.35 23JT 1.45 2FJNT 1.43    
T 0.69 3F 1.09 2FN 1.08 23NT 1.25 3FJNT 1.51    
    3J 1.25 2FT 1.28 2FJN 1.28        
    3N 1.00 2JN 1.03 2FJT 1.53        
    3T 1.32 2JT 1.10 2FNT 1.28        
    FJ 1.20 2NT 0.95 2JNT 1.17        
    FN 0.91 3FJ 1.40 3FJN 1.35        
    FT 1.11 3FN 1.19 3FJT 1.55        
    JN 0.89 3FT 1.33 3FNT 1.41        
    JT 0.93 3JN 1.25 3JNT 1.36        
    NT 0.85 3JT 1.45 FJNT 1.32        
        3NT 1.20            
        FJN 1.11            
        FJT 1.21            
        FNT 1.11            
        JNT 1.12            

Table 3. Mean number of actives found in the = top-10=20 positions for chemical similarity searches when combining various = numbers,=20 c, of different similarity measures.
[The = pale-green=20 shading indicates a fused result at least as good as the best original=20 similarity measure.]

It will be seen that very many of the fused combinations in Figure 3 = are=20 shaded, thus supporting the idea that improvements in effectiveness can = be=20 achieved by using more than just a single similarity measure. The table = also=20 shows that the fraction of the combinations that are shaded increases in = line=20 with c, so that all combinations with c=B3=20 4 perform at least as well as the best of the individual similarity=20 measures. That said, the best result overall was obtained with 23FJT = (rather=20 than with 23FJNT, the combination involving all of the individual = measures):=20 thus, while simply fusing as many individual measures as are available = appears=20 to work very well, superior results may be obtained (for this dataset at = least)=20 from fusing a subset of the individual measures.

Other datasets studied by Ginn et al. (1999) were: 8178 = molecules from=20 the Starlist file (available from BioByte Corp. at URL = http://clogp.pomona.edu)=20 for which experimental values of the octanol/water partition = coefficient, an=20 important parameter in statistical methods for the prediction of = biological=20 activity, were available; several sets of 3,500 molecules from the WDI = database=20 for which biological activity data were available; and 136 biological = dyes used=20 to stain cells so as to visualise various organelles. In all cases, it = was found=20 that use of a fusion rule such as SUM generally resulted in a level of=20 performance (however quantified) that was at least as good as (and often = noticeably better than) the best individual measure. The best individual = measure=20 often varies from one target structure to another in an unpredictable = manner,=20 and the use of a fusion rule will thus generally provide a more = consistent level=20 of searching performance than will a single measure of chemical = similarity. With=20 the increasing number of such measures available that can be implemented = with a=20 reasonable degree of efficiency (Willett et al., 1998), it would = seem=20 appropriate to consider the use of more than one of them for similarity=20 searching of chemical structure databases.

Conclusions

This paper has presented several examples of algorithms and data = structures=20 that are applicable to the processing of both textual and chemical = databases.=20 These examples suggest that each area has much to offer to the other, = with the=20 transfer of methodology occurring from chemical to textual applications, = and=20 vice versa.

The substructure searching methods discussed in the second section = provide an=20 excellent example of this transfer of ideas. Here, an approach that was = first=20 developed for chemical retrieval (specifically in the context of 2D = substructure=20 searching) was soon shown to be applicable to many applications in the = general=20 area of textual databases. One such application, distribution sorting, = resulted=20 in an algorithm that was then adapted for use in a rather different area = of=20 chemical retrieval (3D substructure searching). In fact, although not = discussed=20 here, the idea of using distance and angular information for database = searching=20 has now been extended to the computationally demanding task of searching = for=20 patterns in the 3D structures of proteins, where our search methods have = led to=20 the discovery of many previously unknown structural relationships = between=20 proteins (see, e.g., Artymiuk et al., 1996, 1997). In the = case of=20 the hypertext-comparison project, the similarity measures discussed = above were=20 originally developed for calculating the similarities between pairs of = 2D=20 molecules, while the GA for the calculation of weights was designed from = the=20 start for use in both relevance feedback and substructural analysis. = Finally,=20 our use of data fusion for combining chemical similarity measures was a = simple=20 application of work done by others in the textual domain. Given the = range of=20 applications we have already been able to identify, we trust that other=20 researchers will be encouraged to investigate the many similarities that = exist=20 between chemical and textual database processing.

Acknowledgements. I thank all of my colleagues, past and present, = for=20 their contributions to the research that has been summarised here, and = the many=20 organisations that have provided financial support for my research over = the=20 years. The Krebs Institute for Biomolecular Research is a designated=20 Biomolecular Sciences Centre of the Biotechnology and Biological = Sciences=20 Research Council.

References

  • Adamson, G.W., Cowell, J., Lynch, M.F., McLure, A.H.W., Town, W.G. = and=20 Yapp, A.M. (1973) "Strategic considerations in the design of screening = systems=20 for substructure searches of chemical structure files." Journal of = Chemical=20 Documentation, 13, 153-157.=20
  • Arabnia, H.R. and Zhu, D., editors (1998). Proceedings of the=20 International Conference on Multisource-Multisensor Information = Fusion,=20 Fusion. 98. CSREA Press.=20
  • Artymiuk, P.J., Poirrette, A.R., Rice, D.W. and Willett, P. (1996) = "Biotin=20 carboxylase comes into the fold." Nature Structural Biology, = 3,=20 128-132.=20
  • Artymiuk, P.J., Poirrette, A.R., Rice, D.W. and Willett, P. (1997) = "A=20 polymerase I palm in adenylyl cyclase?" Nature, 388, = 1997,=20 33-34.=20
  • Ash, J.E., Warr, W.A. and Willett, P., editors (1991) Chemical=20 Structure Systems. Chichester: Ellis Horwood.=20
  • Barnard, J.M. (1993) "Substructure searching methods: old and = new."=20 Journal of Chemical Information and Computer Sciences, = 33, 532 -=20 538.=20
  • Barnard, J.M., Lynch, M.F. and Welford, S.M. (1984) "Computer = storage and=20 retrieval of generic chemical structures in patents. Part 6. An = interpreter=20 program for the generic structure language GENSAL." Journal of = Chemical=20 Information and Computer Sciences, 24, 66-70.=20
  • Barton, I.J., Creasey, S.E., Lynch, M.F. and Snell, M.J. (1974) = "An=20 information-theoretic approach to text-searching in direct-access = systems."=20 Communications of the Association for Computing Machinery, = 17,=20 345-350.=20
  • Belkin, N.J., Kantor, P., Fox, E.A. and Shaw, J.B. (1995) = "Combining the=20 evidence of multiple query representations for information = retrieval."=20 Information Processing and Management, 31, 431-448.=20
  • Carhart, R.E., Smith, D.H., Venkataraghavan, R. (1985) "Atom pairs = as=20 molecular features in structure-activity studies: definition and=20 applications." Journal of Chemical Information and Computer = Sciences,=20 25, 64-73.=20
  • Cooper, D., Dicker, M.E. and Lynch, M.F. (1980) "Sorting of = textual=20 databases: a variety generation approach to distribution sorting."=20 Information Processing and Management, 16, 49-56.=20
  • Cooper, D. and Lynch, M.F. (1984) "The use of binary search trees = in=20 external distribution sorting." Information Processing and = Management,=20 20, 547-557.=20
  • Cramer, R.D., Redl, G. and Berkoff, C.E. (1974) "Substructural = analysis. A=20 novel approach to the problem of drug design." Journal of Medicinal = Chemistry, 17, 533-535.=20
  • Crestani, F. and Melucci, M. (1998) "A case study of automatic = authoring:=20 From a textbook to a hyper-textbook." Data and Knowledge = Engineering,=20 27, 1-30.=20
  • Cringean, J.K., Pepperrell, C.A., Poirrette, A.R. and Willett, P. = (1990)=20 "Selection of screens for three-dimensional substructure searching."=20 Tetrahedron Computer Methodology, 3, 37-46.=20
  • Edgar, S.J., Holliday, J.D. and Willett, P. (1999) "Effectiveness = of=20 retrieval in similarity searches of chemical databases: a review of=20 performance measures." In preparation.=20
  • Ellis, D., Furner-Hines, J. and Willett, P. (1994) "On the = creation of=20 hypertext links in full-text documents: Measurement of inter-linker=20 consistency." Journal of Documentation, 50, 67-98.=20
  • Ellis, D., Furner-Hines, J. and Willett, P. (1996) "On the = creation of=20 hypertext links in full-text documents: Measurement of retrieval=20 effectiveness." Journal of the American Society for Information=20 Science, 47, 287-300.=20
  • Gillet, V.J., Willett, P. & Bradshaw, J. (1998) = "Identification of=20 biological activity profiles using substructural analysis and genetic=20 algorithms." Journal of Chemical Information and Computer = Sciences,=20 38, 165-179.=20
  • Ginn, C.M.R., Willett, P. and Bradshaw, J. (1999) "Combination of=20 molecular similarity measures using data fusion." Perspectives in = Drug=20 Discovery and Design, submitted for publication.=20
  • Hall, D.L. (1992) Mathematical Techniques in Multisensor Data=20 Fusion. Northwood, MA: Artech House.=20
  • Jakes, S.E. and Willett, P. (1986) "Pharmacophoric pattern = matching in=20 files of 3-D chemical structures: selection of inter-atomic distance = screens."=20 Journal of Molecular Graphics, 4, 12-20.=20
  • Kahn, S.D. (1998) "Combinatorial libraries: structure activity = analysis,=20 in: Schleyer, P.v.R., Allinger, N.L., Clark, T., Gasteiger, J., = Kollman, P.A.,=20 Schaefer III, H.F. and Schreiner, P.R., editors, Encyclopedia of=20 Computational Chemistry. Chichester: John Wiley. Vol 1, pp. = 417-425.=20
  • Kier, L.B. (1987) "Indexes of molecular shape from chemical = graphs."=20 Medicinal Research Reviews, 7, 417-440.=20
  • Lynch, M.F. (1977) "Variety generation - a re-interpretation of = Shannon. s=20 mathematical theory of communication and its implications for = information=20 science." Journal of the American Society for Information=20 Science, 28, 19-25.=20
  • Lynch, M.F. and Willett, P. (1987) "Information retrieval research = in the=20 Department of Information Studies, University of Sheffield." = Journal of=20 Information Science, 13, 221-234.=20
  • Markey, K. (1984) "Inter-indexer consistency tests: A literature = review=20 and report of a test of consistency in indexing visual materials." = Library=20 and Information Science Research, 6, 155-177.=20
  • Martin. Y.C. and Willett, P., editors (1998) Designing = Bioactive=20 Molecules: Three-Dimensional Techniques and Applications. = Washington:=20 American Chemical Society.=20
  • Maybury, M.T., editor (1997) Intelligent Multimedia Information = Retrieval. Cambridge MA: MIT Press.=20
  • Ormerod, A., Willett, P. and Bawden, D. (1989) "Comparison of = fragment=20 weighting schemes for substructural analysis." Quantitative=20 Structure-Activity Relationships, 8, 115-129.=20
  • Robertson, A.M. and Willett, P. (1996) "An upperbound to the = performance=20 of ranked-output searching: optimal weighting of query terms using a = genetic=20 algorithm." Journal of Documentation, 52, 1996, 405-420. =
  • Robertson, S.E. and Sparck Jones, K. (1976) "Relevance weighting = of search=20 terms." Journal of the American Society for Information = Science,=20 27, 129-146.=20
  • Salton, G. (1989) Automatic Text Processing. Reading, MA:=20 Addison-Wesley.=20
  • Salton, G. and Buckley, C. (1990) "Improving retrieval performance = by=20 relevance feedback." Journal of the American Society for = Information=20 Science, 41, 288-297.=20
  • Sparck Jones, K. and Willett, P., editors (1997) Readings in=20 Information Retrieval. San Francisco: Morgan Kaufmann.=20
  • Welford, S.M., Lynch, M.F. and Barnard, J.M. (1981) "Computer = storage and=20 retrieval of generic chemical structures in patents. Part 3. Chemical=20 structure grammars and their role in the manipulation of chemical = structures."=20 Journal of Chemical Information and Computer Sciences, = 21,=20 161-168.=20
  • Willett, P. (1981) "A fast procedure for the calculation of = similarity=20 coefficients in automatic classification." Information Processing = and=20 Management, 17, 53-60.=20
  • Willett, P. (1982) "The calculation of inter-molecular similarity=20 coefficients using an inverted file algorithm." Analytica Chimica = Acta,=20 138, 339-342.=20
  • Willett, P. (1995) "Searching for pharmacophoric patterns in = databases of=20 three-dimensional chemical structures." Journal of Molecular=20 Recognition, 8, 290-303.=20
  • Willett, P. (1997) "Information retrieval research in the = University of=20 Sheffield", ACM SIGIR Forum, 31(2), 7-13.=20
  • Willett, P., Barnard, J.M. and Downs, G.M. (1998) "Chemical = similarity=20 searching." Journal of Chemical Information and Computer = Sciences,=20 38, 983-996.=20
  • Willett, P., Winterman, V., Bawden, D. (1986) "Implementation of = nearest=20 neighbour searching in an online chemical structure search system = 26,=20 36-41.

How to = cite this=20 paper:

Willett, Peter (2000)  = "Textual=20 and chemical information processing: different domains but similar=20 algorithms"  Information Research, = 5(2)=20 Available at: http://informationr.net/ir/5-2/paper69.html

=A9 the author, 2000.   Last = updated: 5th=20 January 2000

Contents

Web=20 Counter

Home

------=_NextPart_000_0000_01C4B2D9.F7DD9D80 Content-Type: image/gif Content-Transfer-Encoding: base64 Content-Location: http://informationr.net/ir/figs/contents.gif R0lGODlheAAUAPcAAAAAAAAAMwAAZgAAmQAAzAAA/zMAADMAMzMAZjMAmTMAzDMA/2YAAGYAM2YA ZmYAmWYAzGYA/5kAAJkAM5kAZpkAmZkAzJkA/8wAAMwAM8wAZswAmcwAzMwA//8AAP8AM/8AZv8A mf8AzP8A/wAzAAAzMwAzZgAzmQAzzAAz/zMzADMzMzMzZjMzmTMzzDMz/2YzAGYzM2YzZmYzmWYz zGYz/5kzAJkzM5kzZpkzmZkzzJkz/8wzAMwzM8wzZswzmcwzzMwz//8zAP8zM/8zZv8zmf8zzP8z /wBmAABmMwBmZgBmmQBmzABm/zNmADNmMzNmZjNmmTNmzDNm/2ZmAGZmM2ZmZmZmmWZmzGZm/5lm AJlmM5lmZplmmZlmzJlm/8xmAMxmM8xmZsxmmcxmzMxm//9mAP9mM/9mZv9mmf9mzP9m/wCZAACZ MwCZZgCZmQCZzACZ/zOZADOZMzOZZjOZmTOZzDOZ/2aZAGaZM2aZZmaZmWaZzGaZ/5mZAJmZM5mZ ZpmZmZmZzJmZ/8yZAMyZM8yZZsyZmcyZzMyZ//+ZAP+ZM/+ZZv+Zmf+ZzP+Z/wDMAADMMwDMZgDM mQDMzADM/zPMADPMMzPMZjPMmTPMzDPM/2bMAGbMM2bMZmbMmWbMzGbM/5nMAJnMM5nMZpnMmZnM zJnM/8zMAMzMM8zMZszMmczMzMzM///MAP/MM//MZv/Mmf/MzP/M/wD/AAD/MwD/ZgD/mQD/zAD/ /zP/ADP/MzP/ZjP/mTP/zDP//2b/AGb/M2b/Zmb/mWb/zGb//5n/AJn/M5n/Zpn/mZn/zJn//8z/ AMz/M8z/Zsz/mcz/zMz/////AP//M///Zv//mf//zP///wAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAACwAAAAAeAAUAAAI/gB//YpD sKDBgwgTKlzIsKHDhxAfCoxIsaLFixgjotjIsaPHjyBDihxJsqTJkyYJolzJsqXLlynjwJxJs6ZN jipvWtu586bPnx5z0uTZ0drMnkONAg0qsyZSnUqPRl26USjMpyF5Tu2pdaPWrUQ5cg37VWzYmlZf Yv2I9OnZtlNRwDUbda5ZoGldri0K1mtfuX/9Cr4LmC/eplJFYrU7ePFfx4P9xqWZt+Vewo0fg/2q FHJhvpddVmYZ+nNmw6ZToz7NtvTK0aQnR/7s2XNr1KFdo4S98qxcwm41e7xceyvmmbxj+wZs27Rv zrOD110OMznV69hJWs/OvTuK7d7DAy8NCAA7 ------=_NextPart_000_0000_01C4B2D9.F7DD9D80 Content-Type: image/gif Content-Transfer-Encoding: base64 Content-Location: http://informationr.net/ir/figs/iauthori.gif R0lGODlheAAUAPcAAAAAAAAAMwAAZgAAmQAAzAAA/zMAADMAMzMAZjMAmTMAzDMA/2YAAGYAM2YA ZmYAmWYAzGYA/5kAAJkAM5kAZpkAmZkAzJkA/8wAAMwAM8wAZswAmcwAzMwA//8AAP8AM/8AZv8A mf8AzP8A/wAzAAAzMwAzZgAzmQAzzAAz/zMzADMzMzMzZjMzmTMzzDMz/2YzAGYzM2YzZmYzmWYz zGYz/5kzAJkzM5kzZpkzmZkzzJkz/8wzAMwzM8wzZswzmcwzzMwz//8zAP8zM/8zZv8zmf8zzP8z /wBmAABmMwBmZgBmmQBmzABm/zNmADNmMzNmZjNmmTNmzDNm/2ZmAGZmM2ZmZmZmmWZmzGZm/5lm AJlmM5lmZplmmZlmzJlm/8xmAMxmM8xmZsxmmcxmzMxm//9mAP9mM/9mZv9mmf9mzP9m/wCZAACZ MwCZZgCZmQCZzACZ/zOZADOZMzOZZjOZmTOZzDOZ/2aZAGaZM2aZZmaZmWaZzGaZ/5mZAJmZM5mZ ZpmZmZmZzJmZ/8yZAMyZM8yZZsyZmcyZzMyZ//+ZAP+ZM/+ZZv+Zmf+ZzP+Z/wDMAADMMwDMZgDM mQDMzADM/zPMADPMMzPMZjPMmTPMzDPM/2bMAGbMM2bMZmbMmWbMzGbM/5nMAJnMM5nMZpnMmZnM zJnM/8zMAMzMM8zMZszMmczMzMzM///MAP/MM//MZv/Mmf/MzP/M/wD/AAD/MwD/ZgD/mQD/zAD/ /zP/ADP/MzP/ZjP/mTP/zDP//2b/AGb/M2b/Zmb/mWb/zGb//5n/AJn/M5n/Zpn/mZn/zJn//8z/ AMz/M8z/Zsz/mcz/zMz/////AP//M///Zv//mf//zP///wAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAACwAAAAAeAAUAAAI/gB//YpD sKDBgwgTKlzIsKHDhxAfCoxIsaLFixgjotjIsaPHjyBDihxJsqTJkyYJolzJsqXLlynjwERpzdrM mxtr4gSp0qROlz93ogh6kqhQFD1LGhVJdOlMpyOh3kwadahNpVdzZt0pNWRXmFSZWu0YVGfNs1qt blW71ezasWPRknU7t2zTkmG9wk3Lt+lPu3zVer0qF7DgvlnLxlTK0TBcv40Jt5XMlDLiuYEhvw2Z 92Piz5ED7/07WbRH0qH3Pj4rV+tmzjKjsi4MWrVhyKpPW16NObfu1x87k9Wdefdo46gHFxedHKTN rxyFNya+Gnlt3qZ7OyZ9fXh2nrH1KnqW/Pzzbr+4nZ+fXL7u37p4wx+dT78+Uvn28+tvKX2///8d 9QfggPoFBAA7 ------=_NextPart_000_0000_01C4B2D9.F7DD9D80 Content-Type: image/gif Content-Transfer-Encoding: base64 Content-Location: http://informationr.net/ir/figs/isubji.gif R0lGODlheAAUAPcAAAAAAAAAMwAAZgAAmQAAzAAA/zMAADMAMzMAZjMAmTMAzDMA/2YAAGYAM2YA ZmYAmWYAzGYA/5kAAJkAM5kAZpkAmZkAzJkA/8wAAMwAM8wAZswAmcwAzMwA//8AAP8AM/8AZv8A mf8AzP8A/wAzAAAzMwAzZgAzmQAzzAAz/zMzADMzMzMzZjMzmTMzzDMz/2YzAGYzM2YzZmYzmWYz zGYz/5kzAJkzM5kzZpkzmZkzzJkz/8wzAMwzM8wzZswzmcwzzMwz//8zAP8zM/8zZv8zmf8zzP8z /wBmAABmMwBmZgBmmQBmzABm/zNmADNmMzNmZjNmmTNmzDNm/2ZmAGZmM2ZmZmZmmWZmzGZm/5lm AJlmM5lmZplmmZlmzJlm/8xmAMxmM8xmZsxmmcxmzMxm//9mAP9mM/9mZv9mmf9mzP9m/wCZAACZ MwCZZgCZmQCZzACZ/zOZADOZMzOZZjOZmTOZzDOZ/2aZAGaZM2aZZmaZmWaZzGaZ/5mZAJmZM5mZ ZpmZmZmZzJmZ/8yZAMyZM8yZZsyZmcyZzMyZ//+ZAP+ZM/+ZZv+Zmf+ZzP+Z/wDMAADMMwDMZgDM mQDMzADM/zPMADPMMzPMZjPMmTPMzDPM/2bMAGbMM2bMZmbMmWbMzGbM/5nMAJnMM5nMZpnMmZnM zJnM/8zMAMzMM8zMZszMmczMzMzM///MAP/MM//MZv/Mmf/MzP/M/wD/AAD/MwD/ZgD/mQD/zAD/ /zP/ADP/MzP/ZjP/mTP/zDP//2b/AGb/M2b/Zmb/mWb/zGb//5n/AJn/M5n/Zpn/mZn/zJn//8z/ AMz/M8z/Zsz/mcz/zMz/////AP//M///Zv//mf//zP///wAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAACwAAAAAeAAUAAAI/gB//YpD sKDBgwgTKlzIsKHDhxAfCoxIsaLFixgjotjIsaPHjyBDihxJsqTJkyYJolzJsqXLlynjhLRG05pJ miJxwhypc2XPnShUfvxZkihQjkY7JiW5tKVQpTZ9Rj3qsSkKqzOnAn2KVGtXqBtx1qyqtabXq2bT Zg1r0yzZtmXHdi37kSvbsz91ys07VW9csjyjykX7FW1fwX/DgrR7NzHbx4UJS4Zs1KpfsJMl8wV7 diPjuZE1O75cuHLn0ok3Z+a79DPo1Ygxi04bG3DO2pRT0/ba1DVqzKpFr7U9XLXx041D+s4NfDRu 4qGHPifNPGfmjowPR95LffJSopanOWs3HB1p+aAypYMfe3kw5Mas/Z7uzhq226tQzy6nep2//537 UYXVfwSWFCBM9xWooFPpLejggywFBAA7 ------=_NextPart_000_0000_01C4B2D9.F7DD9D80 Content-Type: image/gif Content-Transfer-Encoding: base64 Content-Location: http://informationr.net/ir/figs/isearch.gif R0lGODlheAAUAPcAAAAAAAAAMwAAZgAAmQAAzAAA/zMAADMAMzMAZjMAmTMAzDMA/2YAAGYAM2YA ZmYAmWYAzGYA/5kAAJkAM5kAZpkAmZkAzJkA/8wAAMwAM8wAZswAmcwAzMwA//8AAP8AM/8AZv8A mf8AzP8A/wAzAAAzMwAzZgAzmQAzzAAz/zMzADMzMzMzZjMzmTMzzDMz/2YzAGYzM2YzZmYzmWYz zGYz/5kzAJkzM5kzZpkzmZkzzJkz/8wzAMwzM8wzZswzmcwzzMwz//8zAP8zM/8zZv8zmf8zzP8z /wBmAABmMwBmZgBmmQBmzABm/zNmADNmMzNmZjNmmTNmzDNm/2ZmAGZmM2ZmZmZmmWZmzGZm/5lm AJlmM5lmZplmmZlmzJlm/8xmAMxmM8xmZsxmmcxmzMxm//9mAP9mM/9mZv9mmf9mzP9m/wCZAACZ MwCZZgCZmQCZzACZ/zOZADOZMzOZZjOZmTOZzDOZ/2aZAGaZM2aZZmaZmWaZzGaZ/5mZAJmZM5mZ ZpmZmZmZzJmZ/8yZAMyZM8yZZsyZmcyZzMyZ//+ZAP+ZM/+ZZv+Zmf+ZzP+Z/wDMAADMMwDMZgDM mQDMzADM/zPMADPMMzPMZjPMmTPMzDPM/2bMAGbMM2bMZmbMmWbMzGbM/5nMAJnMM5nMZpnMmZnM zJnM/8zMAMzMM8zMZszMmczMzMzM///MAP/MM//MZv/Mmf/MzP/M/wD/AAD/MwD/ZgD/mQD/zAD/ /zP/ADP/MzP/ZjP/mTP/zDP//2b/AGb/M2b/Zmb/mWb/zGb//5n/AJn/M5n/Zpn/mZn/zJn//8z/ AMz/M8z/Zsz/mcz/zMz/////AP//M///Zv//mf//zP///wAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAACwAAAAAeAAUAAAI6AB//YpD sKDBgwgTKlzIsKHDhxAfCoxIsaLFixgjotjIsaPHjyBDihxJsqTJkyYJolzJsqXLlynjwCxpraa1 mStr4mypcidInT5RAg0ak2jHoUZJIk0asmfSpUdtfrR5k6POoVSrXoXKFIVTo1Knaq26kSxWpFDD ck36FWzYsmRRcD0b9afZuF29yswrl65VvHD/1hUrmO/GtkzVZkVLtXBgwo8NI048ViRQv5EH9zV8 eK/hy4AjY94MmbRkz2Adm9Y8erXo0GxRu02b9e/WzHAVq847mbPv3xx7Ax/OVzjx40aNI1+OMyAA Ow== ------=_NextPart_000_0000_01C4B2D9.F7DD9D80 Content-Type: image/gif Content-Transfer-Encoding: base64 Content-Location: http://informationr.net/ir/figs/ihome.gif R0lGODlheAAUAPcAAAAAAAAAMwAAZgAAmQAAzAAA/zMAADMAMzMAZjMAmTMAzDMA/2YAAGYAM2YA ZmYAmWYAzGYA/5kAAJkAM5kAZpkAmZkAzJkA/8wAAMwAM8wAZswAmcwAzMwA//8AAP8AM/8AZv8A mf8AzP8A/wAzAAAzMwAzZgAzmQAzzAAz/zMzADMzMzMzZjMzmTMzzDMz/2YzAGYzM2YzZmYzmWYz zGYz/5kzAJkzM5kzZpkzmZkzzJkz/8wzAMwzM8wzZswzmcwzzMwz//8zAP8zM/8zZv8zmf8zzP8z /wBmAABmMwBmZgBmmQBmzABm/zNmADNmMzNmZjNmmTNmzDNm/2ZmAGZmM2ZmZmZmmWZmzGZm/5lm AJlmM5lmZplmmZlmzJlm/8xmAMxmM8xmZsxmmcxmzMxm//9mAP9mM/9mZv9mmf9mzP9m/wCZAACZ MwCZZgCZmQCZzACZ/zOZADOZMzOZZjOZmTOZzDOZ/2aZAGaZM2aZZmaZmWaZzGaZ/5mZAJmZM5mZ ZpmZmZmZzJmZ/8yZAMyZM8yZZsyZmcyZzMyZ//+ZAP+ZM/+ZZv+Zmf+ZzP+Z/wDMAADMMwDMZgDM mQDMzADM/zPMADPMMzPMZjPMmTPMzDPM/2bMAGbMM2bMZmbMmWbMzGbM/5nMAJnMM5nMZpnMmZnM zJnM/8zMAMzMM8zMZszMmczMzMzM///MAP/MM//MZv/Mmf/MzP/M/wD/AAD/MwD/ZgD/mQD/zAD/ /zP/ADP/MzP/ZjP/mTP/zDP//2b/AGb/M2b/Zmb/mWb/zGb//5n/AJn/M5n/Zpn/mZn/zJn//8z/ AMz/M8z/Zsz/mcz/zMz/////AP//M///Zv//mf//zP///wAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAACwAAAAAeAAUAAAI0wB//YpD sKDBgwgTKlzIsKHDhxAfCoxIsaLFixgjotjIsaPHjyBDihxJsqTJkyYJolzJsqXLlynjwJxJs6ZN jipvlrRmzSNPnUBR5gwa8mdHo0STghyq9GhPp02j4pQplSPSjVd5Zu2pFetVFFqfVmUq9StSo2fR qrX69KtSslHDyvUKFazYtHTzSoXb1GxbsXb14rU7dyzVqoHrbhV89y9ij3yV+tWbeHBlwI8jJ52c 17JnzIYfJ2brdDHj0m6TahbNmujq1rBtvo5N++Xs2rhRBgQAOw== ------=_NextPart_000_0000_01C4B2D9.F7DD9D80 Content-Type: image/gif Content-Transfer-Encoding: base64 Content-Location: http://informationr.net/ir/5-2/p69fig1.gif R0lGODdhLQAeAPcAAAAAAP//zv////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////// /////////////////////////////////////////////////////ywAAAAALQAeAAAIlgADCBxI sKDBgwgTKlzIsKHDhxAjSpxIMSKAixgzaswYAEBFghgXXvwX8uPAkgkBkCRpEuRFkR1bFkR50ONL mSdv4nxIcyfDnj4VAuVp0yZRjxRfIoW4VKJOnT8/No3ZcKrTmVVNTrVaE6bBpzO5gqwoNqtIsWW9 nkWYNmVVpRypMt1It6RGgW2/1t3LdWjQv4ADCw4cEAA7 ------=_NextPart_000_0000_01C4B2D9.F7DD9D80 Content-Type: application/octet-stream Content-Transfer-Encoding: base64 Content-Location: http://counter.digits.com/wc/-d/-z/6/-b/FF0033/paper69 R0lGODlhWgAUAIQAAL6+vv8AMwL///+Qpv94k//4+f/AzP9AZv/g5v+In/9gf/+ouf+Amf/w8//Q 2f+4xv+YrP+gs//o7P/Y3//I0/+wv0BAQDtP/zYgIzJiIzVvIzJlIyNzI0dwI2VjI2FlIyH5BAEA AAAALAAAAABaABQAQAX+YCCOZGmeaKqubCs2RTIaRWXCskjbZlRQgUKBsUr4JC1f4TRoOBrQxYrY orqu2OyJIBQ2UtzudwWLpbrdVNl7RhcMJlpkNBnGC3NR3TqiuIVSN2MiEkgnDggIWQUTWo6PkFo+ iiJCJ5MjligKeEI5JpoBCaEkQgSZSyacgW2fNHAlnjNvJn6opyxPDioRUEAouAFHJ0IPDHWNxAXG yCykJhPR0ijS1ckmCGjXkdzd3t/g4SwMA1XlLQQUFMEnD0K7KhDq5ygD9vfsqBQD2ZSgP/wK+Dux oMunEu4eiMgGr4QDgSIeDhxxQOCABBhP+GGV7aCIjSM6ojDQwI8ZFj5wGpKgAKtSqhKj+ChzRSvW yQCvejByKfPfzhVh8iCEYoBkAwg3bu5ISpPHygILonbpeQJKCgOFhDrbJmLUtWddfwZ52YoagkAV EKgcESEaCwTXBhRYi0qIIRRo7qoAW6LtBK0lrE0zW1Kc4cOIExsOAQA7 ------=_NextPart_000_0000_01C4B2D9.F7DD9D80 Content-Type: text/css; charset="iso-8859-1" Content-Transfer-Encoding: quoted-printable Content-Location: http://informationr.net/ir/IRstyle.css BODY { MARGIN-LEFT: 15%; MARGIN-RIGHT: 10% } DIV { FONT-WEIGHT: bold; FONT-SIZE: small; FONT-FAMILY: Arial; TEXT-ALIGN: = center } H1 { FONT-WEIGHT: bold; FONT-SIZE: x-large; COLOR: maroon; FONT-STYLE: = normal; FONT-FAMILY: serif; TEXT-ALIGN: center } BLOCKQUOTE { FONT-WEIGHT: 500; FONT-SIZE: medium; MARGIN-LEFT: 50px; COLOR: black; = MARGIN-RIGHT: 50px; FONT-STYLE: italic; FONT-FAMILY: serif; = BACKGROUND-COLOR: #ffffce; TEXT-ALIGN: justify } H2 { FONT-WEIGHT: bold; FONT-SIZE: medium; MARGIN-LEFT: -8%; COLOR: navy; = FONT-STYLE: normal; FONT-FAMILY: Verdana, Geneva, Arial, Helvetica, = sans-serif; TEXT-ALIGN: left } H3 { FONT-WEIGHT: bold; FONT-SIZE: medium; MARGIN-LEFT: -5%; COLOR: teal; = FONT-STYLE: normal; FONT-FAMILY: Verdana, Geneva, Arial, Helvetica, = sans-serif; TEXT-ALIGN: left } H4 { FONT-WEIGHT: bold; FONT-SIZE: large; COLOR: maroon; FONT-STYLE: italic; = FONT-FAMILY: Times, Times Roman, serif; TEXT-ALIGN: center } P { FONT-WEIGHT: normal; FONT-SIZE: large; COLOR: black; FONT-STYLE: = normal; FONT-FAMILY: serif; TEXT-ALIGN: justify } LI { FONT-WEIGHT: normal; FONT-SIZE: medium; COLOR: black; FONT-STYLE: = normal; FONT-FAMILY: serif } OL { FONT-WEIGHT: normal; FONT-SIZE: medium; COLOR: black; FONT-STYLE: = normal; FONT-FAMILY: serif } UL { FONT-WEIGHT: normal; FONT-SIZE: medium; COLOR: black; FONT-STYLE: = normal; FONT-FAMILY: serif } ------=_NextPart_000_0000_01C4B2D9.F7DD9D80--