Human Cytome Project - A framework for cytome exploration - Update 19 April 2005

Hi,

As the on-line version of my article on the Human Cytome Project and the
application of cytomics in medicine and drug discovery (pharmaceutical
research) evolves, I put the updated version in this newsgroup for
reference. The original "question" on a Human Cytome Project was posted in
this newsgroup on Monday 1 December 2003.

On-line version (split version):
A Human Cytome Project - an idea
[Only registered users see links. ]

Human Cytome Project and Drug Discovery
[Only registered users see links. ]

Human Cytome Project - How to Explore
[Only registered users see links. ]

A framework for cytome exploration
[Only registered users see links. ]

================================================== ==========================

A framework for cytome exploration

By Peter Van Osta

Goal

To create an analog to digital workflow concept which can be applied to
ultra large scale research of human cellular diversity to improve our
understanding of cellular disease processes and to develop better drugs
(less attrition due to better functional predictions). Allow for managing
a highly diverse quantitative processing of cellular structure and
function. Create in-silico representations of cellular (and maybe beyond)
structure and function to make them accessible to quantitative content and
feature extraction.

Introduction

An entire organism is an anisotropic, densely packed, 4D grid (or matrix)
of a high order of “recursive” information levels. We can study its
structure and function at multiple levels, where the structure and
function at each level is intertwined with over- and underlying structures
and their function. The genotype and the phenotype both exist in a
continuum of (bidirectional) interacting organizational levels.

Here I want to present and discuss some ideas on the exploration of the
cytome and the conversion of the spatial, spectral and temporal properties
of the cytome and its cells into their in-silico digital representation.
It is a set of ideas about a concept which is still changing and growing,
so do not expect anything final or polished yet.

A modular and distributed framework should provide a unified approach to
the management of the quantitative analysis of space (X, Y and Z),
spectrum (wavelength) and time (t) related phenomena. We want to go from
physics to quantitative features and finally come to a classification and
understanding of the underlying biological process. We want to extract
attributes from the physical process which are giving us information about
the status and development of the process and its underlying structures.

First we have to create an in-silico digital representation starting from
the analogue reality captured by an instrument. The second stage (after
creation of an in-silico representation) is to extract meaningful parts
(objects) related to biologically relevant structures and processes.
Thirdly we apply features to the extracted objects, such as area and
(spectral) intensity, which represent (relevant) attributes of the
observed structure and process. Finally we have to separate and cluster
objects based on their feature properties into biologically relevant
subgroups, such as healthy versus disease.

In order to quantify the physical properties of space and time of a
biological sample we must be able to create an appropriate digital
representation of these physical properties in-silico. This digital
representation is then accessible to algorithms for content extraction.
The content or objects of interest are then to be presented to a
quantification engine which associates physical meaningful properties or
features to the extracted objects. These object features build a
multidimensional feature space which can be inserted into feature
analyzers to find object/feature clusters, trends, associations and
correlations.

Managing the flow

My personal interest is to build a framework in which acquisition,
detection and quantification are designed as modules each using plug-ins
to do the actual work and which operate on objects being transferred
through the framework. Data representing space, time and spectral sampling
are distributed throughout a data management system to be processed. The
data flow through the framework and are subjected to plug-in modules which
operate on the data and transform the content into another content
representing space, such as physics to features. The focus is not on the
individual device to create the data or on individual algorithms, but on
the management of the dataflow through a distributed system to convert
spatial, spectral and temporal data into a feature (hyper-) space for
quantitative analysis.

The software framework manages the entire flow and transformation of data
from physics to features, like a ball which is thrown from player to
player. As long as digital information is transferred from module to
module, it is nothing more than a chunk of data whose actual data layout
is only important for those modules which act upon its data content. The
dimensionality of its content (XYZ, spectral, time) only matters for those
modules which have be aware of it for extracting content in the process
from converting physics into features and finally attributing a meaning to
the events being observed.

Up- and downscaling of cell-based research is dynamically managed by the
system as the scale of processing does not require a change in basic
design. Expanding and collapsing data and feature dimensionality is a
dynamic process in itself and leading to a continuously variable
exploration system. Methods and algorithms for content extraction and
feature attribution are overloaded for a multiplicity of data types and
dimensionality.

I will mostly focus on imaging technology, but the basic principles should
be applicable on any digitized content extraction process. Images are
digital information matrixes of a higher order; they only become images as
such when we want to look at them and have to transform them into
something which is meaningful for our visual system. Visualization
provides us with a window on the data content, but not necessarily on the
data as such. Probing the sample

We want to extract from the sample its structure and its dynamics or the
flow of its structural changes through time. When applying digital imaging
technology to a biological sample, a clear understanding of the physical
characteristics of the sample and its interaction with the “sampling”
device is a prerequisite for a successful application of technology.

The basic principle of a digital imaging system is to create a digital
in-silico representation of the spatial, temporal and spectral physical
process which is being studied. In order to achieve this we try to let
down an equidistant sampling grid on the biological specimen. The physical
layout of this sampling grid in reality is never a precise isomorphic
cubical sampling pattern. The temporal and spectral sampling inner and
outer resolution is determined by the physical characteristics of the
sample (electromagnetic spectral range and spectral sampling layout) and
the interaction with the detection technology being used.

The instrument which converts the spatial (scale, dimensions), spectral
(electromagnetic energy, wavelength) and temporal continuum of the sample
into its digital representation allows us to take a view on biology beyond
the capacity of our own perceptive system. It rescales space, spectrum and
time into a digital representation accessible to human perception
(contrast-range, color) and ideally also to quantification. Instruments
rescale spatial dimensions, spectral ranges and time into a scale which is
accessible to the human mind. The digital image acts as a see-through
window on a part of the physical properties of the biological sample, not
on the instrument as such.

We want to insert a probe system into the sample which changes its state
according to the physical characteristics of the sample. A probe is in
general a dual system, a structure/function reporter on one side and an
appropriate detector on the other side. The changes in the probe system
are ideally perfectly aligned in a spatial-spectral and temporal space
with the physical properties of the sample itself in space and time. Each
probe system senses the state of the specimen with a finite aperture and
so provides us with a view on the biological structure. All sensing is
done in a 5 dimensional environment, in 3D space, spectrum (wavelength)
and time. It is the inner an outer resolution of our sampling which
changes. When we do 2D imaging, this is the same as 3D with the 3rd
dimension collapsed to one layer, but due to the Depth of Focus (D.O.F.)
of the optical system we use, this represents a physical Z-slice.

In the spectral domain we also probe electromagnetic energy along the
spectral axis with a certain inner and outer resolution. We slide up and
down the spectral axis within the spectral limits of the probing system,
which transforms analogue electromagnetic energy into its digital
representation. A single CCD camera probes the visible spectrum (and
beyond) in one sweep, with a rather bad inner resolution. A 3CCD camera
uses 3 probes to do its spectral sampling and gives us a threefold
increase in inner resolution. Increasing or decreasing the density of the
spectral sampling is only a matter of spectral dynamics. By using n
cameras (or PMTs, etc.), each individually controlled (spectral) we can
expand or collapse our spectral inner and outer resolution. We tend to use
“spectral imaging” for anything which samples the visible spectrum
with more than the spectral resolution of a 3CCD camera. Up-and
downscaling our spectral sampling from broad to narrow, parallel or
sequential, continuous or discontinuous is a matter of applying an
appropriate detector array. A system can manage 1 to n spectral probing
devices such as cameras or PMTs (or a spectral filter in front of a single
detector), each sampling a part of the spectrum and spatially aligned
allows probing the spectrum in a dynamic way.

The time axis is also probed with a varying temporal inner and outer
resolution and depending on the characteristics of the detection device;
the time-slicing can be collapsed or expanded. Time can be sampled
continuously or discontinuously (time-lapse). We can expand or collapse
the temporal resolution of the detector in order to capture (temporal
integration) weak signals or shorten the time-slicing down to the minimum
achievable with a given detector.

In order to compensate for sensitivity deficits of a detector, three
strategies for improvement can be followed, but all three decrease the
sampling resolution. Spatial, spectral and temporal signal integration can
be used by expanding the physical scale of capturing along the spatial,
temporal or spectral axis or in combinations. Using a B/W camera instead
of a 3CCD camera is a way of spectral integration, but gives a threefold
reduction in spectral sampling.

The result of the detection is a 5-dimensional system expanding or
collapsing each dimension (XYZ, lambda, time) according to the
requirements of exploration. The device and its components attached to the
exploration core, imposes the inner and outer resolution limits upon the
system. In-silico these are only high-order matrix arrays representing a
5D space. We could call this a continuously variable in-silico
representation.

The inner an outer resolution of the probing system is determined by the
physical XYZ sampling characteristics of the sampling device, such as its
point spread function (PSF). For a digital microscope the resolving power
of the objective (XYZ) and its depth of view/focus are important issues in
experimental design and determining the application range of a device. The
interaction of the detection device with the image created by the optics
of the system such as Nyquist sampling demands, distribution of spectral
sensitivity, dynamic range, also plays an important role.

Multiplexing

In order to increase and improve the extraction of content from our
experiments, we try to increase their information density by multiplexing.
To increase the throughput of exploration we try to do multiple
experiments simultaneous to obtain multiple readouts at once. We
miniaturize the experiments (multi-well plates, arrays) and we use
biological entities which can be multiplexed in relatively small volumes
(cells, tissue samples). We place multiple molecular structural and/or
functional markers or labels into each biological unit (labeled molecules,
structural contrast), so we can make functional and structural
cross-correlations between biological events. The more events and
structures we can explore in parallel, the more chance we have to detect
potential meaningful events (shotgun, grid, and mesh or spider web type
exploration). From each structural or functional label we extract multiple
attributes as quantitative features. It is the choice of the appropriate
markers and their features which are co-changing with functional
attributes (cell division, apoptosis, cell death …) which is open for
exploratory research.

Arrays are actually a type of miniaturized assays; they allow us to do
more experiments on a smaller footprint. The exploration of samples is
organised in an array-pattern (in general 2D due to technical
limitations), ranging form a single tissue slice on a glass slide up to a
large scale grid of for instance a cell or tissue expression arrays.
Biological samples, up to tissue samples are small enough to allow for
multiplexing experiments and they do not require large amounts of reagents
in huge containers. Multiplexing experiments with entire elephants would
be somewhat cumbersome, but DNA, protein, cells and parts of tissue nicely
fit into our instruments. Scaffold cultures would allow us to use the 3rd
dimension if we can properly capture its content. Dynamic scaffold
culturing, would allow us to disassemble the culture for manipulation or
content exploration and reassemble them for continuation of the experiment
(the ultimate scaffold culture is the organism itself).

DNA and protein arrays are arrays of the first degree, as each sample in
an array in itself provides us with a scalar readout; there is no further
spatial differentiation. Cell arrays are of the second or third degree,
depending on the content (how many cells per array coordinate) and the
resolution of the readout. In an array of the second degree each array
coordinate is in itself an array as it is not a homogeneous sample
(multiple cells), but readout resolution is limited to the sub elements.
In an array of the third degree each of the sub elements is also
compartmentalized (e.g. tissue arrays, sub-cellular organelles, nuclear
organization) and each array coordinate is explored at sufficient
resolution. By using arrays with multiple cells at each coordinate, we can
create readout cascades at multiple readout resolutions. This way we can
combine speed and simplicity for a quick overview and switch to more
detail, to find out about cellular heterogeneity and/or sub-cellular
compartmentalization.

At each array position we can add additional spatial, spectral and
temporal multiplexing strategies. Spatial multiplexing in arrays is done
in cell based assays or bead assays. Spectral multiplexing is done by
using multiple spectral labels, either static or by using spectral shift
signalling (dynamic spectral multiplexing). Temporal multiplexing is done
by sequential readouts at each array position to study dynamics or
kinetics. By combining arrays with multiplexing we can increase the
content readout of experiments. By combining DNA-, RNA-, protein-, cell-
and tissue arrays with each other we can also multiplex information from
different biological processes, e.g. massive parallel RNAi transfection of
stem cells.

When we construct arrays with compartmentalized elements, we can up- and
downscale our exploration without the need to redo an entire experiment
and so extract more content from the experiment when wanted. The
experiment is arranged and its content is extracted in a way like Russian
dolls fit into each other. When the array consists of living cells or
tissue, we can add the time dimension to our experiment and create a 4D
array for experimental multiplexing.

The granularity or density of the array pattern is determined by the
experimental demands and upstream and downstream processing capacity. Of
course the optical characteristics of the sample carrier (glass, plastic)
will determine the spatial sampling limits in its inner and outer
resolution. The optical and mechanical characteristics of the device used
to explore the (sub) cellular physical domain will also lead to a spatial,
spectral and temporal application domain. The coarse grid-like pattern of
samples on a sample carrier is being explored at each array position at
the appropriate inner and outer resolution, within the optical physical
boundaries of the device used to capture the data. The outer resolution
barrier of the individual detector in space and time is extended by both
spatial and temporal tiling at a range of intervals. Spectral multiplexing
is being done by using spectral selection devices with the appropriate
spectral characteristics for the spectral profile of the sample.

Feedback loops on the content-flow

The detection cascade is not a one way passive flow of events, but we can
place content-driven feedback systems into the dataflow. Adaptive content
generation manages a source content driven digitalization process. Active
feedback and control depends on the degree of automation and flexibility
of the detection system. The spatial content capturing can be driven by a
plug-in which controls the spatial sampling in order to sample within the
physical boundaries of a sample (e.g. adaptive tissue scanning in 2D or 3D
and beyond). A plug-in is docked into the system to modify its behavior
and make it respond to content changes. The decision process can be
implemented, based on a set of rules implemented as a neural network,
fuzzy logic or whatever is appropriate. Spatial, spectral and temporal
events can drive the process to create a content-driven acquisition
process. Feedback loops cross the dimension and scale boundaries, a
spectral change can drive a change in spatial layout, etc. A content
driven time-lapse will change its temporal pacing whenever a meaningful
event is detected and allow for aniso-temporal sampling. An acquisition
system can be equipped with an active search plug-in making it search for
interesting regions at low resolution and switching to high resolution for
spectral and/or time-slicing. Liquid dispensers, incubators, robot arms
and other automated components can be controlled by a content driven
control system. Object extraction

Robust operating algorithms for object extraction are a prerequisite for a
large scale endeavor. A semi-interactive approach is not acceptable for
large volume processing. The challenges are enormous as robust unattended
large scale object extraction is still not achieved in many cases. The
failure rate of the applied object extraction procedures must be less than
1 to 0.1 percent if we are to rely on large scale automated exploration of
the human cytome.

The detection of appropriate objects for further quantification is done
either in-line within the acquisition process or distributed to another
process dealing with the object extraction. Objects should be aligned with
biological structures and processes. The pixel or voxel representation
in-silico however is basically “unaware” of this meta-information
about how the digital density pattern was created. The physical meaning of
one data point will change depending on the spatial, temporal and spectral
sampling and its inner and out resolution. The digital data build a
(dis-)continuous representation of a spatial, spectral and temporal
continuum which expands or collapses in an anisotropic way.

The content of the data is of no meaning for a data-transfer system as
such, it only transfers the content throughout its dependencies.
Analytical tools operating on the data content need to be informed about
the layout of the data. Detection and quantification algorithms act on the
digital information as such and only the back-translation into physical
meaningful data requires a back-propagation into the real-world layout and
dimensions. The resulting discrete representation of the sampled spatial,
spectral and temporal grid at each array position is being sent to a
storage medium (file system, database…) to provide an audit trail for
quality assessment and data validation.

Content extraction

The selected objects are sent to a quantification module which attaches an
array of quantitative descriptors (shape, density …) to each object. We
expand or collapse the content extraction according to their meaning for
describing the biological phenomenon. Content extraction is being
multiplexed, just as the experiment itself.

Objects belonging to the same biological entity are tagged to allow for a
linked exploration of the feature space created for each individual
object. The resulting data arrays can be fed into analytical tools
appropriate for analysing a high dimensional linked feature space or
feature hyperspace. The dynamics of the attributes of the biological
system need not be aligned with the features we extract to create a
quantitative representation. An attribute change and a feature of which we
expect to represent this change may not be perfectly aligned, so we may
only capture a fraction of the actual change itself. Changes may occur in
a combined spatial-spectral and temporal space of which we can only
capture certain features, such as length, intensity, volume, etc.

The feature sets can be fed into analytical systems for statistical data
analysis, exploratory statistics, classification and clustering.
Classification performance can be improved by combining several
independent classifiers on the feature sets. The resultant vector of a
multiparametric quantification may point in the most meaningful direction
to capture a change. Both parametric and nonparametric approaches to
classification can be used.

We often try to do our experiments on a non-changing background (genetic
homogeneity) or average the background noise by randomisation. What we
call noise is in many cases not well understood but maybe meaningful
dynamic behaviour of a system? Trying to describe changes relative to
underlying oscillations, e.g. cell cycle, by using dynamic background
reporters could help to find dynamic correlations between events.

Copyright notice and disclaimer

My web pages represent my interests, my opinions and my ideas, not those
of my employer or anyone else. I have created these web pages without any
commercial goal, but solely out of personal and scientific interest. You
may download, display, print and copy, any material at this website, in
unaltered form only, for your personal use or for non-commercial use
within your organization. Should my web pages or portions of my web pages
be used on any Internet or World Wide Web page or informational
presentation, that a link back to my website (and where appropriate back
to the source document) be established. I expect at least a short notice
by email when you copy my web pages, or part of it for your own use. Any
information here is provided in good faith but no warranty can be made for
its accuracy. As this is a work in progress, it is still incomplete and
even inaccurate. Although care has been taken in preparing the information
contained in my web pages, I do not and cannot guarantee the accuracy
thereof. Anyone using the information does so at their own risk and shall
be deemed to indemnify me from any and all injury or damage arising from
such use. To the best of my knowledge, all graphics, text and other
presentations not created by me on my web pages are in the public domain
and freely available from various sources on the Internet or elsewhere
and/or kindly provided by the owner.

If you notice something incorrect or have any questions, send me an email.

Email: pvosta at cs dot com

First on-line version published on 9 Jan. 2005, last update on 16 April
2005

The author of this webpage is Peter Van Osta, MD.