IEEE Technology and Society Magazine - June 2019 - 60

data as well as DNA mutations (see Figure 2(a)), it is possible to generate heterogeneous populations of agents
(see Figure 2(c)) that resemble the variability known to
be present in tumors [17].
Simulation of multi-cellular systems is computationally demanding, but because of the simulation structure, it
is very suitable for running in HPC environments. Since
at each time step the internal state of each agent can be
computed independently of the other agents, these operations can be run in parallel and thus performance may
scale linearly with the number of agents. There are
already different packages for simulating multi-scale
models, and some of them have been developed to run
in HPC environments (see Figure 2(d)). For example,
PhysiCell has already been designed and implemented
for the HPC environment and its deployment is straightforward [19]. However, simulating the evolution of multicellular systems of realistic sizes, involving billions of
cells, as those found in in-vivo tumors, remains a challenging task even for HPC infrastructures. Moreover, to
study the effects of drug synergies on such systems
requires a very large number of such simulations.
To support the simulation of multicellular systems
with the actual number of cells found in in-vivo tumors,
the framework outlined in the prior section ("Method,
Architecture, and Apparatus") would host, in the respective components, i) machine and deep learning techniques for obtaining dynamic cell-cycle models; ii)
complex event forecasting techniques for the early
detection of undesired simulation outcomes, e.g., when
performing large-scale parameter screening, as well as
various events of biological interest, such as forecasting
the emergence of resistant cells via the detection of
small changes in physiological or molecular markers;
and iii) graphical workflow designs and interactive learning techniques enhancing the efficiency of model calibration and parameter selection during the simulation
process. All the above would be efficiently orchestrated
by an optimization and runtime adaptation component.
Such a virtual laboratory will lead to the exploitation
of simulation outcomes for the timely indication of personalized therapies. Socio-economic implications of
improved therapies, in turn, will aid patients in maintaining their working status and quality of life, as well as
in receiving equal treatment from insurance, banking
(such as loans), or other services in their everyday life
[20]. Moreover, shorter treatment durations will cut
down on public healthcare costs, with the possibility of
savings being invested in other social services.

Broader Application Domains
In principle, the generic architecture of Figure 1 suits several other application domains dealing with massive
data flows. Taking advantage of such data requires

60

sophisticated analytics tools, capable of extracting insights
on-the-fly, from a multitude of voluminous, correlated,
high-velocity data streams, but also, harvesting ever-growing historical data repositories. This imposes similar challenges, compared to life sciences, to computer scientists,
system engineers, and industrial stakeholders.

Maritime Monitoring
Maritime surveillance applications need to combine highvelocity data streams, including global maritime surveillance systems, such as the Automatic Identification System
(AIS), with acoustic signals of autonomous, unmanned
vehicles, such as wave gliders [21]. Employing an architecture that incorporates the discussed algorithmic apparatus
helps improve maritime situational awareness by enabling
the accurate identification (via learning new patterns) and
forecasting the activities of "dark targets" that are (intentionally) hidden from traditional monitoring systems.

Finance
In the financial domain, stock price forecasting and
portfolio management rely on stock ticker data combined with rich, real-time information sources on various pricing indicators. Financial data involve a variety of
market data, including stock market and crypto-currencies market data, arriving in tens of thousands of correlated, high-velocity streams. The distributed complex
event forecasting component of Figure 1 can be used to
forecast price swings of stocks, currencies, and commodities. The distributed learning and data mining component can serve systemic risk prediction purposes and
offer decision support for investment opportunities.

Battling Cancer with Extreme Scale Analytics
We presented the method, architecture, and algorithmic
apparatus for materializing interactive extreme-scale
analytics in the battle against cancer. Our discussion
expands on the whole processing pipeline, from when
time-distributed data streams resulting from simulations
of multi-scale biological processes are digested into a
Big Data/HPC platform, to the extraction of real-time
knowledge and event forecasting. The apparatus
includes architectural components equipped with novel
algorithms necessary for constructing concise data
summaries, machine learning models for knowledge
extraction, and event forecasting facilities. These components are assisted by integrated workflow design
tools, minimizing programming effort, and by an optimizer transparently devising the execution of the whole
data processing pipeline. In this way, the time required
to set up on-line processing pipelines is diminished, and
real-time discovery and monitoring of events is enabled,
paving the way for searching, discovering, and employing new personalized cancer therapies.

IEEE TECHNOLOGY AND SOCIETY MAGAZINE

∕

JUNE 2019
IEEE Technology and Society Magazine - June 2019

Table of Contents for the Digital Edition of IEEE Technology and Society Magazine - June 2019
