Predicting gene essentiality from expression patterns in Escherichia coli

Essential genes are genes whose loss of function causes lethal-
ity. In the case of pathogen organisms, the identification of these genes
is of considerable interest, as they provide targets for the development
of novel antibiotics. Computational analyses have revealed that the posi-
tions of the encoded proteins in the protein-protein interaction network
can help predict essentiality, but this type of data is not always avail-
able. In this work, we investigate prediction of gene essentiality from
expression data only, using a genome-wide compendium of expression
patterns in the bacterium Escherichia coli, by using single decision trees
and random forests. We first show that, based on the original expression
measurements, it is possible to identify essential genes with good accu-
racy. Next, we derive, for each gene, higher level features such as average,
standard deviation and entropy of its expression pattern, as well as fea-
tures related to the correlation of expression patterns between genes. We
find that essentiality may actually be predicted based only on the two
most relevant ones among these latter.We discuss the biological meaning
of these observations.