Everything you wanted to know about the University of Minnesota's analysis of Arabidopsis ESTs but were afraid to ask

Kristi L. Swope(1,2), Thomas C. Newman(3), Elizabeth Shoop(2), Paul Bieganski(2), Ed Chi(2), Olaf Holt(2), John Carlis(2), John Riedl(2) and Ernest F. Retzel(1)

(1)Medical School and (2)Department of Computer Science, University of Minnesota, Minneapolis, MN
(3)DOE Plant Research Laboratory, Michigan State University, East Lansing, MI

email: comments@lenti.med.umn.edu

Table of Contents

1.0 Overview

At the Plant Molecular Informatics Center at the University of Minnesota, Arabidopsis EST sequences from the Michigan State University and the French consortium are being processed for quality and analyzed for similarity against public database sequences (GenBank, GenInfo, PIR), and the resulting data is made available on the World Wide Web (Shoop et al., 1995). In short, we supply information pertaining to:

In addition to processing data for Arabidopsis, we are processing sequences for other plant genomes, including rice, maize, and loblolly pine, with Brassica soon to come. All of the results are made available via our Web server [http://lenti.med.umn.edu] under a queriable WAIS index.

The purpose of this electronic tutorial is to explain the work that we have done to date and to demonstrate methods for accessing the mountains of data. In addition, we will take you step by step through example reports, providing tips regarding our presentation of the data.

2.0 Making mountains of data (or, What exactly do we analyze?)

2.1 Preparation of the raw sequence

We obtain raw sequence information directly from Michigan State University (Newman, et al., 1994), and the first step in processing the sequences is to filter artifacts from machine sequencing out of the data set. Each EST is checked for quality by checking for the ratio of unknown bases to the total number of bases in that sequence. A sequence with less than 4 or 5% of unknown bases is considered to be of reasonably good quality, and sequences with a higher percent of unknown bases are trimmed from the 3' end until sequence quality is satisfactory. In addition, 5' and 3' vector are removed from each sequence to prevent the potential similarities to sequences of interest being overshadowed by matches to vector sequences in the public databanks (Gish and States, 1993). While it would be ideal to remove all vector before checking for quality, the strict rules that are necessary for a computer to remove vector automatically sometimes requires that the quality of the sequence be improved before vector removal.

2.2 Similarity analysis of the sequence

The next step in the sequence processing involves conducting similarity searches on the filtered sequences. The current suite of similarity algorithms executed includes BLASTN, BLASTX, and, if low compositional complexity regions are found, BLASTP (Altschul et al., 1990). By and large, the statistical array of amino acids in proteins are very close to random, and for some proteins, there is a bias toward a certain class of amino acids, which has biological function, such as acidic or basic proteins. However, frequently, there are regions of low compositional complexity (LCC); such regions contain a non-random distribution of amino acids. These LCC regions often lack real biological function, resulting in spurious similarities. We use the XNU program of Claverie and States (Claverie and States, 1993) to identify potential LCC regions and to mask out those regions in peptide sequences translated from the ESTs. However, because of the nature of our anonymous ESTs, we do not make value judgements: we present both the unmasked BLASTX results and the LCC-masked BLASTP results. For example, sometimes the low complexity regions seem to be the only part of a relatively poor hit:

 Score: 58  Frame = 2   
   F+  +  + T+   +  ++   +  +     ++  LL   +  LL+  L +++L+++ +S
and sometimes the hit contains LCC regions, but still appears important:

 Score: 249  Frame = 1   

In both of the above examples, lysine occurs much more frequently than would be expected for a random distribution, where there would be approximately one lysine in every 20 residues.

A few other items are taken into consideration in processing the sequences. When DNA libraries are made using directional cloning, only the positive strand is analyzed. If the clones are not directionally cloned, as is the case for some of our other projects, then all six frames are processed. The blasts are performed against current versions of GenBank, PIR and GenPept databases using a PAM score (Dayhoff et al., 1978) of 250. The high PAM score has been chosen because it allows for similarity matches, or hits, between sequences of large evolutionary distance, and plant DNA makes up a relatively small proportion of the public databases. This high PAM value does not affect matches to very similar sequences (i.e., other plants). Finally, the EST sequences are re-blasted on a reasonably regular basis to keep up with additions to the public databases.

2.3 Making the data easy to access, read and retrieve

The results of the automated analysis performed on a set of ESTs consists of several files that contain information about each individual EST. Some of the files, like the similarity search outputs, are hundreds of lines long and contain information that is interrelated. It is difficult for anyone to peruse each file for an EST manually and get the whole picture of the analysis for that EST. It is equally difficult to obtain a summarized view of the results of the similarity searches for a set of ESTs. Therefore, we package the results from the various analyses performed and produce a single document for each EST with hypertext links connecting related material. You can browse this single document using a hypertext browser like Mosaic or Netscape and move through it using the links (Section 3.3). In order to view blast results for a single sequence at a glance, we have included in this document images from Alignment Viewer, a visualization tool developed as part of this work (Chi et al., 1995). The Alignment Viewer (AV) tool provides a graphical representation of all the hits for an input EST sequence, color-coding the hits to different frames, and organizing the hits by score and the locations of the amino acid residues where the hit occurred along the probe sequence. The resulting display provides for rapid interpretation of the BLASTX or BLASTP hits most likely to provide clues about the function of the unknown sequence. Samples of these documents and their interpretation is provided below (Section 3.2).

As another step in making data accessible, we submit all sequences that are at least 200 base pairs in length with no more than 5% unknown bases (these criteria were defined experimentally, Shoop, et al., 1994) to dbEST (Boguski et al., 1993) at NCBI. The individual clones, which contain the EST sequences, are also made available at the ABRC Stock Center at Ohio State University.

3.0 Searching through mountains of data to find your buried treasure

3.1 Conducting a WAIS search

You can pull up a form for conducting a search on this data by linking to our server [http://lenti.med.umn.edu] and clicking first on the link "Arabidopsis cDNA Sequence Analysis Project", and next on the "WAIS index search" link. One way to think of WAIS searching is to liken it to conducting a literature search, except instead of searching through many journals for articles of interest, you are searching through quality and blast reports for sequences of interest. The form used is shown in Figure 1. You need to make a choice as to which groups of sequences you wish to search. Any combination of one or all of the groups is acceptable. The yellow boxes indicate that the group of sequences will be included in the search. Next, type in the key word or words into the query search string box. There are two important points to keep in mind when choosing your key word. First, WAIS does not treat a phrase as a phrase per se, but as individual words. For example, if you get a hit on the query "early light inducible" it could be because the phrase "early light inducible" was found in the document, or it could be because the word "early" was found in the document, but not "light" or "inducible". The second key to remember is that, while in most literature searches you can restrict hits to the title or the abstract or look for hits anywhere in the report, in WAIS searches, you do not have a choice of restricting searches to particular fields. You can find more tips on conducting a WAIS search under the "tips and pitfalls" link which is just above the form shown in Figure 1. Our server also points to a very detailed explanation on how to conduct searches [http://ls6-www.informatik.uni-dortmund.de/freeWAIS-sf/fwsf_5.html#SEC45]. After typing in your query, choose the number of hits you want returned, and then click the start search button.

After searching, the server will provide you with a list of documents appearing as links (Figure 2). Each of these documents contain all the information described above (Section 2.0) for a single cDNA, and only the documents containing the key word(s) for which you searched are listed. If any the fields that you selected have no hits, this is reported at the top of this list. For example, when searching the MSU Arabidopsis sequences and the USDA loblolly pine sequences for "catalase", we find that there are no hits for the pine sequences; therefore, the hits listed are all from the Arabidopsis sequences (Figure 2). The default value for the number of documents returned is 200. If you get 200, it is probably because there are a lot more than 200 documents that contain your search words. This list of documents is ordered based on the score that the freeWAIS program, which is used to make the WAIS index, assigns to each document. This score is based on the number of occurrences of the search words in a document, the location of the words in a document, the frequency of those words within the collection, and the size of the document. It is important to keep this in mind as you survey your results, because, for example, a high score may point to a document that contains very few or no data.

3.2 Looking at the results of your WAIS search

After the WAIS search comes back with the list described above (Section 3.1), you may click on one of the links to open up one of the documents containing the sequence, quality information and several forms of blast data. For example, by clicking on the link


that is seen in Figure 2, the file containing the sequence analysis information for clone 154N18T7 is opened (Figure 3). This document can be very large, and it is generally advantageous to review its contents by scrolling through various parts of the document and clicking on links to hop quickly to other parts. Another way to move to items of interest is to use the "Find In Current" (or "Find") feature under the "File" (or "Edit") menu, which is part of Mosaic (or Netscape) itself. Here you can type in a key word, such as what you used for the WAIS search, and find the location of the words in the current document. For now, we will use the scroll bar on the right side of the file to take you through the document. First, we move directly to information about the EST studied in this file (Figure 4). The DNA sequence is presented at the top of this file so that you may copy it for your own records.

Following the background sequence information, a visualization of BLASTX similarity results is presented with an image from the Alignment Viewer tool (Figure 5), followed by a visualization of BLASTP hits. In these images, all the alignments to the sequence in question are plotted according to score on the y axis and according to the location and length of the alignment on the x axis. Alignments with scores of 150 or greater are considered strong hits, and a putative function can be assigned ESTs which have hits with this high a score (Shoop et al., 1994). Alignments with scores of less than 80 should be viewed critically, as such a low score suggests that the alignment was largely due to chance. Also, short regions of high similarity may have an artificially high score.

Sometimes the BLASTP image appears empty (Figure 6), and this can occur for various reasons. For instance, if there are no LCC regions found in the sequence, the BLASTP program is not run. As another example, there may be LCC regions found in frame 1, but alignments only in frame 3. In this case, a BLASTP was run on frame 1, but since there are no hits to frame 1, there are no data to graph. If there are LCC regions found in a frame that shows sequence similarity to a public sequence, the location of the regions is presented, as shown in Figure 7.

Underneath the description of LCC regions, a link allows you to go directly to a summary list of protein alignments that were obtained with BLASTP (Figure 8). This summary enables you to determine at a glance what frame the alignments occurred in, whether or not there was a frame shift, and a brief description of the sequence to which the alignment occurred. Additionally, the higher scores and lower p-values indicate which hits are the strongest. By clicking on the "goto" link, you can see the actual alignment between the protein sequences, as well as a longer description of the sequence from the databank (Figure 9). You can also get to these alignments by continuing to scroll down the document from the summary list. Note that the top sequence is the test (or probe) sequence that blast was run against and the bottom sequence is the similar sequence from the target public databank. If you are interested in researching the latter sequence, there are links, which connect directly to the databank, in both the summary report and the individual alignment reports.

If, instead of using the link to go to the BLASTP data, you continue to scroll down the report, there is a summary list of protein alignments that were obtained with BLASTX. These data are collected without masking out any low complexity regions, but the presentation of the data is identical to the BLASTP data. In cases where LCC regions are found, it is important to compare the BLASTX and BLASTP results to determine which data are appropriate for your research purposes.

The summary list of BLASTX alignments is followed by the summary list of BLASTN alignments, which provides similar information on the DNA alignments obtained with BLASTN. After the summary lists are the detailed lists of BLASTX, BLASTN, and BLASTP hits (Figure 9), respectively. These lists are ordered according to increasing p-value, and hits with p-values greater than 0.1 are omitted from the list. For some ESTs, when you look for the BLASTP hits, you may find the heading of this section, but no entries (Figure 10). This may seem puzzling at first. However, if there are no low complexity regions, or hits occurred with a p-value greater than 0.1, there are no reportable BLASTP results, and hence, this section should be empty. If you go back to the table of contents at the top of the document, and click on the "Low Complexity Regions" title, you can double check whether or not LCC regions were found.

3.3 A word about jumping around web pages via links

As you explore these web pages and others, it is useful to keep in mind a few features about links. First, sometimes when you click on a link, you may get a response indicating that a server is not available. This occurs when the local machines that store the information you are requesting is busy or the network itself is busy, in which case the query times out before the server responds. In these cases, it is best to try the link at another time, and you are more likely to get through. Second, one of the attractive features about links, is that server A can point to the information in server B without needing to worry about keeping the information up to date. Server B controls this information and is responsible for updates. However, server B controls the link name as well. If server B has a need to change the name of the link, it is not a trivial task to inform server A of this change. Therefore, the link to server B may be outdated without server A being aware that a change has taken place. Webmasters, or the people who make the web pages available, do a great job in keeping these links current, but occasionally, you may find an outdated link before they do. In these cases, a brief email sent to a contact provided on the web pages, describing the link and the problem, would be greatly appreciated.

4.0 Future directions

We are generally quite proud of the server as it is, with nearly 30,000 sequences and their analyses across the species. However, we are far from done with it! There are many things in queue and planned that you should be aware of, if for no other reason than to encourage you to register with the User Group on the system so you will be notified of new developments and additions.

4.1 Software

There are several pieces of attached (within the server) and associated (clients of the server) software coming on-line soon. By "soon," we mean they already exist, and are either in the final stages of testing and documentation, or are getting their respective databases updated.

First, we have a local BLASTN search tool ready, and it will likely be on our server before you read this paper. This will allow you to search our data sets with your DNA sequence(s). It is interactive, and web-based, and all you need do is select the target genomes you are interested in, paste your sequence into the window, and click on the "submit your query" button. An obvious follow-up for this program is a version for BLASTX to run your data against the translations of coding sequences in our data. All we are waiting for to bring this up is the new version of Xgrail (Xgrail 1.3; unfortunately, the current version isn't suitable for our needs) from Oak Ridge National Labs so we can have a smarter version of the translations of the sequences.

Second, Alignment Viewer (AV) (Chi et al., 1995) is the tool we use to produce the static images in the EST analysis reports. In its "real" incarnation, the program is 3-dimensional and interactive, and additional information is calculated and displayed "on the fly" (specifically, it shows you a curve of what the substitution matrices will do if you were to run them all; both PAM (Dayhoff et al., 1978) and BLOSUM (Henikoff and Henikoff, 1993) matrices are included. If you have a Sun (SunOS 4.1.3, Solaris 2.4), SGI or Linux computer running the "motif" window system, you will be able to run this client. It functions much like other external viewers for the Web (xv being one of them for Unix-based machines), in that it resides on your machine, but reads data from ours.

Third, the MotifExplorer (Bieganski, et al., 1996) is based on a something called a "suffix tree" (Bieganski, et al., 1994; Bieganski, 1995). The underlying algorithm and data structure is described in the papers above; what it *is*, however, is a very fast pattern matching tool. This is the only large scale implemetation that we know of, and presently allows you to explore PIR using a variety of patterns, including ProSite patterns or those of your own design. Our own data will be available under this tool, shortly after we get Xgrail 1.3 (see note above).

One of our longest term and most complex projects has been the development of a relational database management system (DBMS) for sequences, lab, analytical and derived information. This system will allow for very complex searches, or queries, against the entire data set. The data structure has been defined, and, indeed, data has been loaded into the database. It is still very much under development, and is queriable only by Structured Query Language (SQL) at the moment; however, we are designing web-based "query-building" screens so that one will not have to understand SQL (or even know that it exists) to use the DBMS. One of the unique additions to this system has been the development of database operators (termed "datablades") for the DBMS that understands protein motifs (Lundberg, 1995). This allows the searching of sequences in the database with any pattern and particularly the ProSite patterns.

4.2 Connecting the Dots (or, Related projects)

We have several projects which are primarily collaborative in nature. Among these is the addition of mapping information when ESTs have been intentionally (Bertrand Lemieux, York University, Ontario, CA), or accidentally (Amir Sherman, MIT, Boston, USA) mapped onto YACs, BACs or cosmids. We are also adding information supplied by other groups, when families of proteins have been identified with ESTs and other, sometimes characterized clones are available (for example, Estelle Hrabak, at the University of Wisconsin, Madison, USA, has identified ESTs that are isoforms of CDPK). Finally, we are trying to figure out a means to include identified phenotypic mutants from laboratory experiments, particularly when that information maps onto our EST data. Write to us at comments@lenti.med.umn.edu if you have similar information you would like us to make available to the community. You will be the key contact regarding any information you provide.

As happy as we are with the directions we have taken for this project, EST data, even extended by a variety of analyses, needs a context to exist in. To that end, we have been volunteering our time to work on other genomes, feeling that some of the real treasures may be in the comparative genome analysis. While this part of the project is presently unfunded, we are gathering ESTs from rice, maize and loblolly pine presently, and we will shortly add some of the Brassica sequences. In addition, some preliminary discussions are in progress with other projects. We feel very strongly that this database of multi-genus plant ESTs has the ability to begin to provide direction both for molecular and for biological experimentation.

5.0 Acknowledgments

The University of Minnesota Plant Data Acquisition, Analysis and Distribution Project is funded under NSF Grant BIR 940-2380, and the Michigan State University DOE Plant Research Laboratory Arabidopsis cDNA Sequencing Project is funded under NSF Grant BIR 931-3751. In addition, we are supported by major resources from the following: The University of Minnesota Medical School, with special thanks to Dean Frank Cerra; Computing and Informations Services, with special thanks to Professor and Vice President Don Riley; University Networking Services, with special thanks to Director Larry Dunn; Sun Microsystems, with special thanks to Sandra Swenson; IBM with special thanks to Norm Troullie and Pat Carey; and Cray Research, Inc., with special thanks to John Carpenter and Bill King.

We would also like to give our thanks to Mary Anderson of the Nottingham Arabidopsis Stock Centre, and Carolyn Tolstoshev, Mark Boguski and Jane Weisemann of NCBI's dbEST; their encouragement and assistance in this work has been very important to us.

6.0 References

S.F. Altschul, W. Gish, W. Miller, E.W. Myers, and D. J. Lipman. 1990. "Basic Alignment Search Tool." Journal of Molecular Biology, 215:403-410.

P. Bieganski. 1995. "Genetic Sequence Data Retrieval and Manipulation based on Generalized Suffix Trees." Ph.D. Thesis, University of Minnesota, Minneapolis, MN.

P. Bieganski, J. Riedl, J.V. Carlis and E.F. Retzel. 1994. "Generalized Suffix Trees for Biological Sequence Data: Applications and Implementation." In: Proceedings of the IEEE 27th Hawaii International Conference on System Sciences. Oahu, Hawaii. L. Shriver and L. Hunter, (Eds.). IEEE Computer Society Press. V:35-44.

P. Bieganski, J. Riedl, J.V. Carlis and E.F. Retzel. 1996. "Motif Explorer--A Tool for Interactive Exploration of Amino Acid Sequence Motifs." Pacific Symposium on Biocomputing, Hawaii. Submitted.

M.S. Boguski, T.M.J. Lowe, and C.M. Tolstoshev. 1993. "dbest - database for expressed sequence tags." Nature Genetics, 4:332-333.

Ed Huai-hsin Chi, Phillip Barry, Elizabeth Shoop, John V. Carlis, Ernest Retzel, John Riedl. 1995. "Visualization of Biological Sequence Similarity Search Results" Accepted for "IEEE Visualization '95" October Conference. Atlanta.

Jean-Michel Claverie and David States. 1993. "Information enhancement methods for large scale sequence analysis." Computers and Chemistry, 17(2):191-201.

M. O. Dayhoff, R. M. Schwartz, and B. C. Orcutt. 1978. "A model of evolutionary change in proteins." In: Atlas of Protein Sequence and Structure, M. O. Dayhoff, (Ed.). National Biomedical Research Foundation, Vol. 5, Suppl. 3, chapter 22, 345-352.

Warren Gish and David States. 1993. "Identification of protein coding regions by database similarity search." Nature Genetics, 3:266-272.

Steven Henikoff and Jorga Henikoff. 1993. "Performance evaluation of amino acid substitution matrices." Proteins: Structure, Function, and Genetics, 17:49-61.

Ann M. Lundberg. 1995. "Extension of a DBMS with Protein Motif Search Capabilities." M.S. Thesis, University of Minnesota, Minneapolis, MN.

T. Newman, F. de Bruijn, P. Green, K. Keegstra, H. Kende, L. McIntosh, J. Ohlrogge, N. Raikhel, S. Somerville, M. Thomashow, E.F. Retzel and C. Somerville. 1994. "Genes Galore: A Summary of Methods for Accessing Results from Large-Scale P artial Sequencing of Anonymous Arabidopsis cDNA Clones." Plant Physiology. 106:1241-1255.

E. Shoop, E. Chi, J.V. Carlis, P. Bieganski, J. Riedl, N. Dalton, T. Newman and E.F. Retzel. 1995. "Implementation and Testing of an Automated EST Processing and Similarity Ana lysis System." In: Proceedings of the IEEE 28th Annual International Conference on System Sciences. Maui, Hawaii. L. Shriver and L. Hunter, (Eds.). IEEE Computer Society Press. 5:52-61.

E. Shoop, J.V. Carlis and E.F. Retzel. 1994. "Automating and Streamlining Inference of Function of ESTs within a Data Analysis System" In: Proceedings of the IEEE 27th Hawaii International Conference on System Sciences. Oahu, Hawaii. L. Shriver and L. Hunter, (Eds.). IEEE Computer Society Press. V:45-46.