Incorporating recent advances in statistical computing, computer graphics, machine learning, and user interface design, Panmo has the following features:

• Tools to generate information-rich visualizations:Scatterplot, 2D histogram, boxplot, table plot, histogram, barplot, variable resolution bivariate plot, Tukey sum-difference plot, profile plot (or parallel coordinate plot), X-rayed profile plot, 3D point cloud rotation, various types of trellis graphs, and browse.
• Powerful numerical tools:SOM (Self-Organizing Map), K-means, hierarchical clustering, principal component analysis, classification and regression trees, sign test, minimal spanning tree planing, multivariate Smirnov test, various profile searching tools, and lowess smoothing. All these numerical tools present their result graphically and are seamless integrated with the intuitive user environment via dynamic graphics.
• Reporting tools:PostScript/PDF files of publishing/production quality and PNG images can be generated at the click of one finger tip. The whole flow of data analysis and user's train of thought can be recorded and published on the internet easily.
• Multi-window interactive dynamic graphics
• Focusing and linking
• Graphical query formulation
• The most advanced graphics to cope for overstriking
• Logical zooming
• On-line context help
• Object-oriented data representation
• Three layers of user interface
• Smart menu system
• Inspection
• HCS image module
• HCS impressionist density plot module

The combined power of Panmo's tools and operations working together is multiplicative, instead of being additive in traditional informatics systems.

A bit more details about the above features follow:

Multi-window interactive dynamic graphics

Visualizing big and complex data sets using a single display window is difficult at best. Multiple windows are used in Panmo to provide users with simultaneous multiple views into data space. Each display window serves as a 2-way communication link between the system and users: the system shows data in display windows and users can look at a display window and, by interacting with graphical symbols in the display window, issue commands to the system.
Top

To display complicated information, like that contained in a big and complex data set, a common instinct is to draw a plot that is equally complicated, such as presenting the data as a tableau of Chernoff faces. Attempts at such dense encoding are seldom successful. It is usually more effective to construct a number of simple, easy to understand plots, each focused clearly on a particular aspect of the underlying data. Each plot conveys partial information about the data. Panmo integrates the information in multiple plots into a coherent image of the data as a whole by linking the contents of individual plots. Painting is one of many interactive techniques available in Panmo to link contents in plots.

The following 2 plots based on the yeast microarray data illustrate the basic idea.

	Panel A: SOM output
	Panel B: 3D rotation of the yeast data plotted on the first 3 principal components.

A set of genes with gradual increase in expression level during diauxic shift and relatively stable expression level during sporulation in the SOM output was painted red. You can now see where they fall in the 3D space spanned by their first 3 principal components. This linkage allows us to relate the cluster assignment of a gene with its location in the data space. (and vice versa).

Acknowledgement: The data used for the above 2 plots are from Spellman et al. (1998) and are available at the Yeast Cell Cycle Analysis Project website at Stanford University.


Top

Plots in traditional data analysis systems only serve as passive, one-way communication links from the system to the user. There are few ways for users to interact with these systems through plots. Panmo considers components of plots as visual representations of underlying entities (e.g., cells, credit card customers, data sets, etc.), which opens entirely new possibilities for interaction. With such an arrangement, users can look at a plot and, by interacting with graphical symbols in the plot, initiate the retrieval of data from the underlying database. Analysis routines to be applied to the data can be selected graphically. The results of the analysis or data retrieval can again be in the form of plots and ready for further graphical manipulation.

With Panmo, if you see any pattern of interest in a plot, you can retrieve the data generating the pattern immediately. Panmo's way of graphical query formulation is especially useful when there are patterns that are apparent in plots but are tedious or difficult to describe with a textual query language.
Top

Scatterplots are the method of choice for displaying the distribution of points in two dimensions. They are used to discover patterns such as holes, outliers, modes, and association between the two variables. A common problem is overstriking, the overlap on the plotting surface of glyphs representing individual observations. Overstriking can create a misleading impression of the data distribution.

As a simple measure to cope for overstriking, the default plotting glyphs in Panmo's scatterplots are unfilled circles. If there is only partial overlap and no exact overlap, using an unfilledcircle as the plotting glyph can improve the distinguishability of individual points. Filled circles do not share this property and can be very misleading, as illustrated by the following plots.

Panel A	Panel B	Panel C

Panel A and Panel B display 40 points each. Panel A only reveals points not buried by other points and does not give any sense of density. A glance at Panel B will alarm viewers that there are many points closely together. Panel C is a composite of Panel A and Panel B for better comparison. You'll have to make your web browser repeat animation continuously to properly see Panel C.

As a more sophisticated measure to cope for overstriking in scatterplots, the novel variable resolution bivariate plots (or Varebi plots for short) in Panmo deals with the problem of overstriking by mixing display of a density estimate and display of individual observations. The idea is to determine the display format by analyzing the actual amount of overstriking on the screen. Thus, the display format will depend on the sample size, the distribution of the observations, the size and shape of individual icons, and the size of the window. It may change automatically when the window is resized. Varebi plots reveal detail wherever possible, and show the overall trend when displaying detail is not feasible.

Here is an example comparing a scatterplot with a varebi plot (Data Credit: Professor P. Dee Boersma):



Here is an example of a varebi plot with different amount of drawing space (Data Credit: Professor P. Dee Boersma):


You'll have to make your web browser repeat animation continuously to properly see these 2 plots. If there is no serious overstriking, the appearance of a varebi plot will approach that of a scatterplot as the drawing area is increased.

Other areas where Panmo takes special care to deal with overstriking are:
• Boxplots use jittering to alleviate the problem of overstriking.
  
  

  Traditional way to draw a boxplot.	Panmo's way to draw a boxplot.

• X-ray like images of parallel coordinate plots reveal the internal structure obscured by overstriking.

Top

Zooming is one way to get a better look at a portion of the data in a display window. There are two types of zooming: geometric zooming and logical zooming. Geometric zooming produces a blown up version of the region in which each pixel in the source image is represented by a small square of the same color. Logical zooming produces a plot based on the actual observations in a source region. As a result, more details can be revealed.

   We'd like to zoom into the area bounded by the blue rectangle.

   This is the result of a geometric zooming.

   This is the result of a logical zooming.

Logical zooming in Panmo is recursive, which means users can zoom into a plot resulting from a previous logical zooming again. The power of magnification and the area zoomed into can be changed on the fly.
Top

Users can press F1 or Shift-F1 to conveniently get context-sensitive help messages. All help messages are displayed in a Netscape browser window.
Top

Object-oriented data representation allows easy mapping of conceptual entities in the problem domain into computational entities in the system domains.
Top

User interface of Panmo consists of graphical direct manipulation, menus, and textual commands. A textual user interface can be concise and accurate. If a task needs to be carried out repeatedly, this can be simplified by writing a script. A menu based user interface is easy to master. Graphic direct manipulation is good at taking advantage of patterns that are apparent on the screen but are tedious or difficult to describe in a textual command language. Any one of the three alone is not enough to gracefully handle all users' requests. There is usually more than one type of user interface suitable for a task; Panmo lets users determine for themselves the most convenient way to accomplish a task.
Top

Panmo knows what type of menu to use for a tool and only includes relevant items in the menu. This greatly speeds up work flow and reduces finger fatigue.
Top

Users can click any graphical icons and get detailed information on data they represent.
Top

With this module, all plots are fully linked to the original scan images. You can get the original scan images of the cells from any graphical symbols in any plots. Data of any cells in a scan image can be instantly retrieved to pass to any analytic function or to make any plot. For example, the following DNA profile plot was based on the data grabbed out of Panel A; the green cells in the DNA profile plot were grabbed out again to make Panel B. All these were done with only a few mouse clicks.



Panel A	DNA Profile	Panel B



For a more extensive demonstration of this HCS image module, please take a look at here.
Top

Impressionist density plots provide a compact, eye-pleasing way to compare the data from different treatments in HCS experiments. For examples:

For a more extensive demonstration of this HCS impressionist density plot module, please take a look at here.
Top

You'll have to make your web browser repeat animation continuously to properly see this plot.
Top

Minimal spanning tree (MST) planing is also known as multivariate planing. MST planing is similar in intent to multidimensional scaling (MDS) onto a 2-dimensional plane, which tries to find N points in the 2-dimensional plane such that the inter-point distances in the plane match the inter-point distances in high-dimensional space. MST planing is much faster to compute than MDS because it does not require nonlinear optimization. Planing 40,000 points is a perfectly reasonable task. Principal component analysis is just "poor man's MDS."

The following 2 plots are the result of MST planing of 5431 points from a 14-dimensional space:

Acknowledgement: The data used for the above 2 plots are from Spellman et al. (1998) and are available at the Yeast Cell Cycle Analysis Project website at Stanford University.
Top

Parallel coordinate plots take a unique approach to draw high-dimensional data. Since plotting more than 3 mutually perpendicular axes is impossible, parallel coordinate plots draw all the axes parallel to each other and equally spaced in a 2D plane. For example, the following 2 parallel coordinate plots are based on the same data set:



	This parallel coordinate plot uses a common vertical range that encompasses all the values of all the variables being visualized.
	Each axis in this parallel coordinate plot uses a different scale so that the minimum and the maximum values of the variable at an axis are mapped to the bottom and the top of the available drawing room.

Overstriking can frequently be a problem for parallel coordinate plots. Panmo provides several novel ways to cope for it. The following plot demonstrates one of them:



Notice the internal structure and the x-ray like appearance of this plot.

Panmo uses parallel coordinate plots as the launch point for several utilities that help users find similar or dissimilar observations.

Acknowledgement: The data used for the above 2 plots are from Spellman et al. (1998) and are available at the Yeast Cell Cycle Analysis Project website at Stanford University.
Top

This is a smart tool. It knows what plots to draw no matter how many and what types of variables are passed to it. It cuts down dramatically the cognitive effort required during graphical query formulation because users don't have to worry about what types of plots to look at. All they have to do is know what variables to look at.
Top

This plot allows users to study more effectively the deviations of the points from the 45-degree line y = x.



This is a tradition way to compare 2 variables: using a scatterplot with the line y = x drawn inside. It is the vertical deviation from a point to the line y = x that matters here, not the shortest distance from a point to the line y = x. It is hard to compares vertical deviations of the points from the line because the non-zero slope of the line y = x affect our visual perception.

	This is a 45-degree clockwise rotation of the above plot followed by an expansion of the vertical scale and allows users to study more effectively the deviations of the points from the line y = x. See in which plot it is easier to compare those 2 red points.


Top