Data Analysis Pipeline for High Content Screening in Drug Discovery - High-Throughput Image Reconstruction and Analysis - page 215

parameters for each object, such as dimension, intensity, and texture features. Cel-
lomics and Metamorph software, for example, provide more than 30 parameters
for each channel. A more comprehensive list was described by Haralick [24].
In addition to vendor software there are several attempts being made to
develop generic image and data analysis software that can be used with all
acquisition platforms. These include the free and open source image and data
analysis software Cell Profiler, from the Broad/MIT and Whitehead Institutes
(http://www.cellprofiler.org/index.htm) and Zebrafish image analysis software
(ZFIQ) from the Methodist Hospital in Houston, Texas (http://www.cbi-
platform.net/download.htm). More advanced image analysis software can be
found in the Definiens software package (http://www.definiens.com/), which
uses a novel over-segmentation and reconstruction approach to identify cellular
regions of interest. Finally, another example of a free and robust (but rudimentary)
software package is Image J from the NIH (http://rsb.info.nih.gov/ij). This package
is extremely useful for viewing and quickly manipulating images for display and
analysis. This software has a recordable macro language for analysis of images in
batch and has a large user group base to freely exchange macros.

9.7 Data Analysis

In contrast to all the assays, hardware, and image analysis software developed
for HCS in the past 10 years, little has been done on the data analysis aspects,
one reason being that most application developers have neither experience nor the
desire to handle a large amount of data, and without proper integration of data
analysis pipeline, HCS has little use for high throughput screening and remains
on par with a manually operated fluorescence microscope. In an effort to make
HCS practical in a drug screen scenario, we have developed a series of tools for
data extraction, normalization, dose response estimation, and multiparameter cell
classification. We will use most of the rest of the chapter to discuss these tools,
and introduce examples when necessary.

9.7.1 Data Process Pipeline
HCS generates a large amount of primary image data. For example, a typical
medium scale screen of 25,000 compounds in duplicate can easily be achieved
in one week with one HCS instrument and limited liquid handling automation
access. Assuming four individual channels and four image fields were collected in
each treatment condition, the experiment generates

600 GB of raw image data,
and a full HTS screen at 1M compounds will generate over 25 TB of raw image
data (Figure 9.2). For small scale exploration such as HCS assay development, a
local hard drive is sufficient. A corporate storage and archive system is necessary
if screening effort runs at full scale. Some image analysis routines also generate a
large amount of data, especially when every single cell analysis data is stored. We
estimated that 2.5 billion parameters will be generated in the typical medium-sized
experiment, as mentioned above, and a full million compound screen will produce
over 100 billion data points. The dataset is big enough that it cannot be handled by

∼