Biology Reference
In-Depth Information
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
-
5
-
4
-
3
-
2
-
1
0
1
2
3
4
5
Input
FIGURE 2.13
The sigmoid function.
data were labelled with, for example, 0 for proteins with a cytoplasmic location and 1
for secreted proteins, and set of variables which led to an output less than 0.5 would
be classified as having a cytoplasmic location, while any with an output of more than
0.5 would be classified as secreted. It can be demonstrated that a neural network can
learn any function with arbitrary precision but, as so often the case with machine
learning algorithms, the selection of requisite architecture and training regime is a
“black art”, usually addressed using trial and error.
Neural networks can be a very powerful approach to classification and model-
ling of complex systems. They do require a lot of data; one rule of thumb is that
there should be at least five cases in the training set for each weight in the network.
Dataset size is, however, often a moot point with microbiological data. Input data
can be real valued, categorical or binary, making ANNs a flexible approach to
many problems. One criticism often directed at ANNs is that they are “black
boxes”. Unlike classifiers such as decision trees, it is not easy (although not impos-
sible) to untangle the relative importance of the variables that contribute to the
classification.
Neural networks have been used in microbiology to address a very wide range of
problems, such as the estimation of hydrogen production by genetically modified
E. coli
(
Rosales-Colunga
et al.
, 2010
); optimization of biomolecule production
(
Singh
et al.
, 2009; Nelofer
et al.
, 2012
); predicting bacterial community assem-
blages (
Larsen
et al.
, 2012
); elucidating the relationship between growth and envi-
ronmental factors in
Staphylococcus aureus
(
Fern
´
ndez-Navarro
et al.
, 2010
);
optimising fermentation conditions for
E. coli
(
Silva
et al.
, 2012
) and predicting
essential genes in microbial genomes (
Palaniappan and Mukherjee, 2011
) (this
research also incorporated SVMs and decision trees). There have even been attempts
to build ANNs using microbes (
Ozasa
et al.
, 2009
).