Appendix E. Design issues

This appendix addresses a number of key design issues on the implementation of MAExplorer and the implications they have on its efficiency. The ordinary user of MAExplorer need not be concerned with any of these issues. A PowerPoint presentation describing the class structure of the "Software design of the MAExplorer data mining tool" is available as either an Adobe Acrobat file
(PDF) or a PowerPoint file (PPT).

E.1 Internal data structures design to facilitate direct manipulation

MAExplorer was constructed using a number of fundamental data objects including clones (genes), hybridized samples (membranes or glass arrays), tables, etc. organized using an object-oriented methodology enforced by Java. Sets of genes are implemented as bit sets for efficiency in both storage and set-theoretic operations. With a set being implemented as 64-bits/word, a set intersection, union or difference can be performed on 64 genes in parallel in one logical (i.e. AND, OR, XOR) computer instruction. This makes the data filter quite efficient when computing the intersection of many gene sets. When ordered gene lists are required, memory and compute intensive lists are used - but only when needed. Tab-delimited ASCII is used as the basic I/O file type for all types of data. This simplifies I/O and allows data to be prepared with a variety of systems including Excel, array quantification programs, relational database systems, etc.

Another major decision was to use multiple pop-up windows for 2D plots, histograms, expression profiles, clustergrams, reports, dialog boxes, etc. rather than sharing a single window. These windows are maintained by a special pop-up registry that handles many of the bookkeeping chores involved with tracking and updating multiple windows viewing the same underlying data. Whenever an event occurs which may change the set of data filtered genes, the current gene or the current cluster set of genes, the registry is notified. Some of the events are the current clone changed, the Filter parameters changed, the sample labels changed, the normalization method changed, etc. It in turn notifies all relevant active plots, tables and reports - requesting them to update themselves if necessary. This object-oriented design greatly simplifies the process of synchronizing the various data presentations with changes in the database.

E.2 Approaches to data mining: client-centric and server-centric models

There is a range of approaches for performing data mining of microarray data over the Internet. However, all assume rapid access to underlying databases and the ability to transform data from one presentation mode to another where differences might be easily observed. One extreme is the server-centric model using CGI or Applets in Web browser. This assumes that all data search and analysis is performed on a back-end server and graphic or tabular results from the server are sent back to the researcher over the Internet. The server-centric model has the advantage of keeping all user data up-to-date, but the disadvantage of performing all computations and graphics generation on the back-end server. Relying so much on the server for major computations and graphics generation can result in significant delays if the networks or servers are heavily loaded. The other extreme is the client-centric model. Here all of the data being analyzed is copied to a user's computer and computationally expensive analyses are done there. This has the disadvantage for the user of possibly not having the most up-to-date data to analyze as well as setup time overhead. However, it does distribute the computational load, allowing more effective data mining with many alternate views and avoiding excessive delays during a data mining session. In both the Web browser applet and the stand-alone application, data is downloaded to MAExplorer. The difference being access to the local file system with some additional capabilities in the case of the latter.

A good intersection of the server-centric and client-centric methods is to distribute the computation and data to the systems where they can be handled most effectively. Because Java enables computation in a Web browser, PCs currently available have enormous power and memory, and high-speed Internet connections are readily available, it is now possible to distribute some of the data and computations to the desktop. If high-speed direct manipulation methodology is to be made available on the Internet for microarray data mining, then it must be brought to the user's desktop browser or local computer rather than residing solely on the back-end server. This is the approach taken in designing the MAExplorer.


Table E.2 Comparison of client-centric vs. server-centric data mining. The table shows a comparison of some of the features of client-centric and server-centric (using CGI and/or Applet) data mining analysis methods. The client-centric approach presented here primarily uses Java with data downloaded to the client's computer. A server-centric approach might use a mix of HTML, CGI, servlet and Java. However, even a client-centric approach may take advantage of server support for additional functionality (e.g. accessing genomic servers to gain additional information about specific genes or sets of genes).

Approach Advantage (+)
disadvantage (-)
Feature
Client-centric  a) + Java programs run (pretty much) on all operating system platforms as either stand-alone or applets (in browsers)
Client-centric  b) + handles rapid response required for direct manipulation on the new generation of very fast desktop computers
Client-centric  c) + stand-alone version may be restarted quickly from local data or data cached from the Web server
Client-centric  d + size limitations are not a problem with stand-alone Java applications
Client-centric  e) + Java plug-ins allows prototyping new local and Web DB analysis method functionality by any group of users
Client-centric  f) - for the applet version, there is slow startup because the program and all data has to be downloaded each time it is run
Client-centric  g) - difficult to build large stable Web-applets handling very large data sets. However, stand-alone applications don't have this problem
Client-centric  h) - for the stand-alone application version, it must be installed on client's computer where there nmight be some level of incompatibility

Approach Advantage (+)
disadvantage (-)
Feature
Server-centric  a) + may have better resources for very large data sets but with dependence on server
Server-centric  b) + faster startup than downloaded applet since minimal GUI is required and data does not have to be loaded before computation requests may be made to the server
Server-centric  c) + may be easier to prototype and distribute new functionality using third party software such as RDBMS, S-plus, etc. using centralized CGI or servlets where only one copy is required on the server
Server-centric  d) - susceptible to Internet traffic bandwidth problems for large numbers of users
Server-centric  e) - susceptible to server-load dependencies for large numbers of users
Server-centric  f) - difficult to get very rapid response for direct manipulation for data mining


E.3 Conversion of microarray data files to MAExplorer format using Cvt2Mae

A tool is being developed that converts microarray data files, both commercial and one-of-a-kind research data to a complete MAExplorer data format. Input data will be tab-delimited, although it may be possible to use XML data at some point. When the tool becomes available, it will be announced on the MAExplorer home page and in this manual.

Cvt2Mae data converter

Because it is difficult to manually edit user's microarray quantified data files, we constructed the
Cvt2Mae data converter program (also see Appendix C.6). The idea is to create array layouts for known array chips and to let the user define their own for specialized arrays. These user-defined layouts may then be saved and used in subsequent data conversions. The basic problem of data conversion is that of "field picking" to map user data fields to those required by MAExplorer, and of setting the appropriate options in the MAExplorer configuration files. User-interactive wizards query the user and then does this information to perform the conversion generating the output data files that are ready to use with MAExplorer. Cvt2Mae then generates the directory tree of required data files described in Appendix C.

The MAEPlugins Web site is under construction E.4 Extending MAExplorer functionality using Java Plugins

We are adding the ability for users to add their own Java Plug-in Extensions to MAExplorer. These will extend capabilities of the core MAExplorer program to other analysis methods by users. The
MAEPlugins Web site will be an Open Java API, open-source Java code examples, our plugins and donated plugins, links to plugins at other Web sites. Typical plug-ins include: normalization, Filters, PCA, clustering, client-server, Web-server functional analysis of cluster results, etc. We group these into three types of new analytic functionality:
  1. Using Java code to implement the complete plugin. This means it will be portable across all systems.
  2. Accessing local programs written in any language ( eg. the R statistics package). This method may not be portable across all systems.
  3. Web-CGI or client-server to specialized genomic DBs Plug-ins. This method should be portable across all systems.
The MAExplorer Open Java API (Applications Programming Interface) will allow users to get at all data structures without understanding the details of the system. The specialized application classes are derived from the GatherScatterAPI class which can access all of the internal MAExplorer data structures. This allows us to improve and change the internal data structures without causing problems with plugins using those data structures.

The following figures show the top level plugin design.

Overall MAEPlugin design for MAExplorer

Figure E.4.1 Overall MAEPlugin design for MAExplorer. Plugins are dynamically loaded into MAExplorer where they may be invoked from a menu entry or by various other means such as startup, normalization, etc.


Open Java API for MAEPlugins showing the specialized Java classes

Figure E.4.2 Open Java API for MAEPlugins. Each type of application could be derived from specialized Java classes that contain most of the access methods required for that type of analysis.


E.5 Web database server design

Although MAExplorer can be run stand-alone on a user's computer, additional capabilities may be made available with support from the back-end Web database server. This server design, used with the MGAP database, includes several distinct functions (Figure 1). The primary one is the hosting of login-protected microarray quantitative data and auxiliary flat files required to support basic MAExplorer operations. These "flat files" could be synthesized on the fly from searches on a relational database server that is part of the microarray database Web server. The public database does not require a login while the collaborator subset of the database does.

In support of the MGAP server, additional software was written to automate the pre-processing of the microarray quantitative data from Research Genetics' Pathways array quantification analysis program and perform compression and Web server updates for this data. The Web server also hosts several common gateway interface (CGI) programs. These include user login support, a Web proxy server (to access other genomic Web sites from the Java applet), support of login-protected user state file access, custom database creation, user state files, and "groupware" user-access support.