Global Gene Expression Assays: Quantitative Noise Analysis - Systems Biology: Networks, Models, and Applications - page 117

Figure 4.9 Standard deviation of measurement noise s as a function of signal
level q (see text) for MPSS measurements on LPS-activated macrophages at
8 h after activation. Four replicate MPSS runs were taken and the noise level
calculated separately for signatures with no zero measurements ( ° ), one ( ),
and two ( ) zero measurements (see text). Note that those signatures with no

zero measurements exhibit significantly lower noise at higher expression levels.

♦

replicates yielded a zero count. The results are shown in figure 4.9.
It can be seen that the abscissa for the nonzero statistics reaches values
more than one order of magnitude larger than those for the one- or
two-zero statistics. Further, the observed noise strength is considerably
smaller for the nonzero statistics than for the other two, which are of
similar magnitude.
If a measurement of zero tpm for a given signature represented
simply the absence of that signature in the sample being studied, then
the three curves in figure 4.9 would be identical. However, it is clear
from figure 4.9 that data on signatures for which some of the sequenc-
ing replicates are zero exhibits significantly higher variability. This
suggests that while some of the observed zeros truly represent the
absence of a given signature in the data, other zeros represent the
nonmeasurement of a finite signal, or alternatively, some nonzero
measurements result from spurious signatures not actually present in
the sample. While the physical mechanism of these spurious meas-
urements is not clearly understood, it is noteworthy that signatures
that yield zero counts in some replicates and nonzero counts in others
are most often (but not always) signatures with less reliable association
to a gene in the genome [13]. In any case, it is clear from figure 4.9 that
the absence of a signature in one of the sequencing replicates indicates
the need for statistical modeling different from that used when the
signature is present in all replicate measurements.