Data Analysis Pipeline for High Content Screening in Drug Discovery - High-Throughput Image Reconstruction and Analysis - page 216

Figure 9.2 HCS data size. (a) A typical medium sized HCS screen with 25,000 compound treat-
ments in duplicates took one week to complete, assuming imaging takes one hour per 384 well
plates. (b) 600 GB of raw image data was generated, assuming four image fields for each treatment
and four marker channels were taken with a 2 × 2 binned 1M pixel CCD camera and 12-bit digitiz-
ing format. (c) 2.5 billion measurements were generated, assuming 500 cells were quantified with
100 descriptors each in every treatment condition.
an Excel spreadsheet used by most bench biologists unless dramatic data reduction
was achieved.
In our lab, a relational database was provided by Cellomics for storage of
all raw image and image analysis data. Cellomics, like many other vendors, also
provides simple data visualization software at the plate, well, and cell levels, but not
at the screen level. Data treatment is limited to simple statistics such as averaging
and Z function. None of the current commercial vendors provides tools for large
scale data normalization and data mining, which are absolutely necessary for a
successful high throughput screening campaign.
Here we developed/adopted a series of data analysis tools for a large numer-
ical dataset resulting from image analysis algorithms. These data analysis tools
are grouped into three modules, preprocessing, data mining, and data integration
modules (Figure 9.3). The preprocessing module handles retrieval of cell level data
from the image analysis software, data normalization, and generation of quality
control plots. Most of the data reduction was achieved by a variety of data min-
ing modules generating treatment level data. A series of data mining modules has
been built in our lab. We will describe the dose response curve/confidence estima-
tion module and the automated cytometry classification module in detail. A data
driven factor model was also recently published, and we will briefly introduce the
concept.
9.7.2 Preprocessing Normalization Module
Systematic errors, such as plate-to-plate and well-to-well variations, often exist in
a high content screen. Data normalization is a necessary procedure to minimize the