FiatFlux-P - User-Level Workflow Design: A Bioinformatics Perspective - page 106

GC-MS data
input

file format
conversion

RATIO
analysis

flux ratio
output

NETTO
analysis

flux distribution
output

specification of experimental data,
diagnosis of faulty measurements,
exclusion of redundant and faulty data

configuration of reaction network,
specification of experimental parameters,
(repeated) invocation of the flux estimation

Fig. 5.5 Abstract FiatFlux workflow

As detailed in Section 5.1.3, FiatFlux consists of the modules ratio and
netto , where the former performs the flux ratio computation and the latter
calculates the net flux distributions. Figure 5.5 on the next page summarizes
the general FiatFlux workflow: The input data (GC-MS data in netCDF
format, the output format of many MS devices) is first converted into the
internal FiatFlux (.ff) data format, then the ratio and netto analyses are
performed and their respective results are output. In the original FiatFlux
software, both ratio and netto require several user interactions via the
graphical user interface (GUI) FiatFlux. These are indicated by the two note
boxes at the bottom of the figure.
Preparing GeneFisher for its realization as GeneFisher-P was compara-
tively simple: its building blocks had to be integrated as services in Bio-jETI
and combined to (new) workflows. Essentially, only suitable “glue code” was
needed. With FiatFlux, however, the case is not as easy. The user interactions
of FiatFlux require specific expert knowledge, as GC-MS data quality and
relevance have to be assessed and the resulting data has to be compared with
biochemical knowledge in the framework of metabolic flux analysis. Conse-
quently, this expert knowledge had to be translated into quantifiable criteria,
which could be used for the automated determination of intracellular flux dis-
tributions. Thus, for turning FiatFlux into FiatFlux-P it was not sucient
to turn its components into services, but it was also necessary to implement
functionality that emulates the interaction with an expert user as close as
possible.
For analysis procedures that do not involve human interaction, it is easy to
see that the automation of the in silico experiment using workflow technology
increases the speed of the analyses without influencing the results at all.
However, also workflow realizations of usually interactive analysis processes
do not necessarily impact the quality of the results: it is often possible to
identify quantifiable criteria in the human expert's analysis behavior, and
apply these for at least heuristic user interaction emulation. In the case of
FiatFlux-P, most automatically obtained results were as good as the manually
acquired ones (cf. [89, Chapter 6], [90]). This means in particular that in such
cases automated experiments can be used for a (fast) pre-screening or initial
analysis of large amounts of data, and only the remaining “dicult” data
sets have to undergo (time-consuming) manual analysis.