ABSTRACT
The codon bias in
Escherichia coli
for all two-fold degenerate amino acids was studied as dependent on the context from
the six bases in the nearest surrounding codons. By comparing the results in
genes at different expression levels, effects that are due to differences in
mutation rates can be distinguished from those that are due to selection.
Selective effects on the codon bias is found mostly from the first neighbouring
base in the 3
'
direction, while neighbouring bases further away influence mostly the
mutational bias. In some cases it is also possible to identify specific
molecular processes, repair or avoidance of frame shift, that lead to the
context dependence of the bias.
The preferential use in
Escherichia coli
of some synonymous codons over others, the codon bias, increases in genes with
higher expression levels, at least for those genes whose expression level can
be ranked (
1
,
2
). Conversely, the level of codon bias has been used as a surrogate measure of
relative gene expression level. Because of the dependence on gene expression,
codon bias is thought to be based on translational efficiency, possibly speed
or accuracy.
The choice of base pair at a site is influenced also by the surrounding
sequence, the context in which it appears (
3
-
7
). By studying the context effects of synonymous codon choice only, all effects
that are due to selection at the protein level are excluded. Synonymous codon
usage can be influenced by context by a number of mechanisms: intrinsic
mutability through DNA damage, replication error, or efficiency of various
repair processes (
8
); selection on DNA or RNA structure favouring or avoiding certain sequence
combinations; or selection of translational efficiency where translation of
certain codons may be more efficient in certain contexts (
3
,
4
). Bulmer (
6
) studied genes with low bias and found that the context effects on the third-position base choice are largely the same when the complementary sequences
are considered. This suggests that the codon bias in the low-expression genes depends largely on mutational effects acting equally on
both strands, and that these mutational effects are dependent on the base
context in the immediate surroundings.
In the present study we focus explicitly on the context effects of the codon
bias of all two-fold degenerate amino acids individually. As a consequence, the context
from the bases in the same codon is already accounted for and we consider the
additional influence due to the context from the bases in the two surrounding
codons. The two-fold degenerate amino acids were chosen since they satisfy a particularly
simple theoretical relationship between selection and bias. Rather than asking
what codon is preferred in which context, or vice versa, we are asking how much
a difference in context influences the bias. By studying the effects separately
in genes of different overall bias, it becomes possible to distinguish between
effects that are due to mutation and those that are due to selection. Those
that are reasonably ascribed to mechanisms that involve translation depend
strongly on overall codon bias, while those that involve mutation do not. In
some cases it is also possible to identify a molecular mechanism that could be
responsible for the effects.
Our starting point was the ECDC (
9
), a non-redundant compilation of
E.coli
K-12 genes, release 25 (January 1996). ECDC classifies its sequences into
several divisions (genes, ORFs, tRNAs, etc.). We based our dataset on the ECDC
genes only; in particular, we did not use any unidentified putative open
reading frames.
To reduce statistical noise due to small gene length, we deleted all genes
smaller than 150 codons from our data set. Furthermore, we carefully screened
the remaining genes for possible errors, making a number of corrections (which
have been communicated to the maintainers of ECDC) and further deletions. All
corrections were based on the EMBL (
10
) and SWISS-PROT (
11
) databases.
The ECDC entry for the
rhl
F gene is obviously wrong, as it lists the (correct) size of the gene as 4617
base pairs (bp), but gives only the first 391 bp; the full sequence was
extracted from EMBL and used. Only 354 bp are given for the
suc
D gene in ECDC, but the full sequence (870 bp) is available, so we used it. ECDC
gives only the
tor
A sequence corresponding to the mature peptide, but provides the full sequence
for many other similarly processed proteins; we used the complete sequence for
this gene also. The first six bases of the
uxu
B gene are missing in ECDC; we recovered them.
The ECDC entries for all the following genes start three bases before they
should:
nuo
E,
nuo
F (starts with a stop codon!),
nuo
H,
nuo
I,
nuo
J,
nuo
K,
nuo
L,
nuo
M and
nuo
N; we used the correct starting points. The ECDC entries for the following genes
end a number of bases after they should, resulting in apparent internal stop
codons:
asm
A,
glt
L, HRA-1,
rlp
B and
sbc
B; we used the correct gene sequences. The ribosomal protein gene
rpm
I appears to have many internal stop codons, because of the insertion of an
extra base early in the sequence; this was corrected by deleting the spurious
base.
prf
B has an internal programmed frame shift; ECDC starts the sequence after the
frame shift; we did not use this gene. The sequence of the following genes are
not available from the start:
cys
Z,
dgo
D,
emr
X,
fru
B,
his
M,
inf
C,
irp
2,
men
D,
nac
,
nan
T,
neu
E,
rhs
E,
usp
T and
xas
A. Since some of our analyses require information about distance from the start
of the gene, we deleted them from our database.
Our main dataset then comprised 1649 non-redundant sequences of
E.coli
K-12 protein coding genes, at least 150 codons long. This was divided into
groups according to their CAI (codon adaptation index) values. The CAI is a
measure that accounts for the unequal usage of synonymous codons (
12
). The CAI value of a gene correlates with its expression level (
13
). However, there may also be other effects that influence the codon bias. Six
gene groups of approximately equal size were constructed, corresponding roughly
to genes of very low (VL), low (L), medium-low, (ML), medium-high (MH), high (H), and very high (VH) expression level (Table
1
).
Table 1
.
The codon adaptation index represents an average codon bias for all degenerate
amino acids in a gene. On the other hand, for any particular 2-fold degenerate amino acid, the codon bias can be defined as the ratio of
the frequencies,
f
M
and
f
m
, of major and minor codons (
14
,
15
).
B = {{f sub M} over {f sub m}} = {{u sub 1} over {u sub 2}} {e sup {{{2 N} sub
e} s}}
1a
or
ln ( B ) = ln {left ( {{u sub 1} over {u sub 2}} right )} + 2 {N sub e} s
1b
Here
u
1
and
u
2
are the rate constants for the mutation of minor to major codons and vice
versa.
s
is the selection coefficient for the major codon over the minor and
N
e
is the effective population size.
The codon bias in Eqn 1 has two components, the mutational bias,
u
1
/
u
2
, and the selective bias, exp(2
N
e
s
). The mutational part is expected to be independent of the varying selection
pressure on the genes (unless the ratio of the mutation rates varies between
the genes in the different CAI groups). On the other hand, ln(
B
) depends linearly on the selection coefficient for the major codon relative to
the minor. Thus, effects that are the same in all CAI groups will be considered
mutational and those that vary will be considered selective.
The basic assumption behind the use of Eqn 1 is that synonymous changes are
easier to achieve than non-synonymous ones and therefore the synonymous bias is `equilibrated' under
fairly fixed context conditions in the immediate surroundings. Thus, the
difference in ln(
B
) for two different contexts is a direct measure of the contextual influence on
the mutational bias or selective preference for one base-pair sequence over the other.
The small-sample uncertainty in ln(
B
) for an amino acid in a gene or group of genes is approximately
sigma = {{1 + B} over sqrt {N cdot B}}
2
based on a binomial sampling.
N
denotes the number of occurrences of the amino acid considered. In the data set
used this corresponds to ~0.05-0.1 for the more common amino acids and possibly as large as 0.15
for a rare one like Cys.
The relative selective advantage,
s
, of one synonymous codon over another is determined from the ratio of the
respective substitution rates. Only for the 2-fold degenerate amino acids is this ratio the same as the ratio of the
codon frequencies as in Eqn 1. The ratio of the frequencies of two synonymous
codons for an amino acid with higher degeneracy will in general involve not
only their relative selective advantage, but also that of the other synonymous
codons in the family. In fact, for amino acids of degeneracy higher than two,
it is not possible to resolve the pairwise selective advantages from the codon
usage data alone. (e.g. for a 4-fold degenerate amino acid there are six independent pairwise codon
comparisons but only three independent numbers in the codon usage).
It has been found experimentally (
16
) that mutations in ribosomes or in elongation factor Tu that slow down the
elongation rate lead to a proportional slow down in growth rate. Thus if the
average elongation time per codon,
t
el
,
is increased by [delta]
t
el
, the relative change in the growth rate is
s = {{delta {k sub 0}} over {k sub 0}} = - {{delta {t sub {e l}}} over {t sub {e l}}}
3
For organisms that compete through growth, this relative change corresponds to
the selection coefficient
s
for the variant that carries the mutation. The growth rate
k
0
is related to the average translation time
t
el
through
k
0
tel
= [Rib]/[rho]
0
, where [Rib] is the concentration of ribosomes in the cell and [rho]
0
is the total concentration of amino acids in protein (
17
). If the translation time from a certain gene,
j
, increases by [Delta]
t
, e.g. through a synonymous substitution to a codon that is slower to translate,
the average translation time per codon overall increases by [delta]
t
el
= [Delta]
t
[P
j
]/[rho]
0
, where [P
j
] is the concentration of gene product
j
in the cell. Thus, from Eqn 3
s
= -[Phi]
j/Rib
k
0
[Delta]
t
4
[Phi]
j/Rib
= [P
j
]/[Rib] is the mole ratio of the gene product
j
and ribosomes in the cell. This can provide a very substantial selection. It is
similar in magnitude to that calculated by Bulmer (
14
) under somewhat different assumptions.
The codon bias was calculated for all 2-fold degenerate amino acids with all four possible base choices at the
three positions before (i.e. positions 1, 2, and 3 of the previous codon,
labelled P1, P2 and P3) and the three following (i.e. positions 1, 2, and 3 of
the next codon, labelled N1, N2 and N3) the codon considered. Except for some
special cases discussed further below, only one position at a time was
considered. The resulting ln(
B
) has been plotted for the six CAI groups in Figure
1
. For all amino acids in almost all contexts, the curves have a general upwards
tendency. This shows that the bias for the individual amino acids in the
different contexts almost always follows the general one given by the CAI
groupings. The influence of the context shows up as a difference in the curves.
Although strong context effects can also be found from the positions farther
away, the most dramatic effects come, not surprisingly, from the position
immediately following the codon considered (i.e. N1), in agreement with
previous results (
3
-
6
).
There are strong context effects, both mutational and selective, in the
synonymous codon usage bias of the 2-fold degenerate amino acids. While the selective preference for the major
codon in most contexts increases with increasing expression level-corresponding to the upwards slopes in most curves in Figure
1
-we find that the degree of selection, i.e. the slope of the curves, is
influenced by context mostly from position N1. The context at other positions
seem to influence primarily the mutational bias giving parallel curves.
The strongest of the context effects are consistent with some simple models. In
one model that depends on translational efficiency, in this case avoidance of
potential frame shifts, the context effects increase with increasing overall
bias (CAI), as expected. This supports the notion (
13
) that the overall bias is also determined by translational efficiency. The
context dependence of the Phe bias is also strongly position dependent (Fig.
2
). This suggests that the position dependence may be used as a general marker
for effects that are due an avoidance of potential frame shifts or other
processivity errors.
The context effects from the VSP repair, on the other hand, are largely
independent of CAI level, giving mostly parallel lines in Figure
1
, indicating that the efficiency of VSP is not influenced by the transcription
level, in contrast to some other repair processes (
29
). This resolves a recent controversy in the literature where it has been argued
(
30
,
31
) that the expression level dependent distribution of sequences that are targets
or products for VSP indicate that VSP is dependent on expression level and
counter argued (
32
) that this changing distribution is simply a consequence of the changing codon
preferences. By looking at the codon bias as a function of VSP context, we have
automatically accounted for the change in codon preferences and study directly
the extra effect due to VSP: there is no influence on VSP efficiency from the
expression level, at least not for the sequences involved in the bias CAG/CAA,
in agreement with the suggestion by Eyre-Walker (
32
).
Some context effects that show up in Figure
1
appear to be outside of the three types discussed; there are clearly other
mechanisms that can contribute context effects to the selection and/or mutation
between synonymous codons. These could come, for instance, from intrinsic
mutability perhaps based on DNA polymerase interactions, from repair mechanisms
other than VSP, or from selection effects based on tRNA-tRNA interactions on the ribosome.
By looking at individual amino acids, the first and second position context is
already accounted for, and the curves in Figure
1
show the additional influence from the simultaneous context at one position in
the surrounding codons. Thus, the context studied involves the simultaneous
presence of three bases. Some effects are expected to depend on the
simultaneous context at a number of positions in the nearby codon(s). We have
not studied such higher-order effects systematically, because the number of possible context
combinations is huge, resulting in very large small-number uncertainties as the sample is subdivided further.
Previously, the synonymous divergence between
E.coli
and
Salmonella typhimurium
has been interpreted as showing that mutation rates decrease with increasing
expression level (
15
), possibly due to a transcription-repair coupling. If such a coupling influences all mutation rates to the
same relative extent, the mutational bias will still be independent of the
overall bias; if not, some mutational bias effects could show up as selective,
giving diverging lines in Figure
1
. Conversely, any selective effects that are independent of gene expression (or
CAI value), e.g. selection at the level of DNA structure, would be identified
as mutational in the present discussion.
The sequence CCAGG is methylated on the second C from the 5' end of each strand. These methylated `C's can mutate to T through
deamination and this process has been identified as providing mutational
hotspots (
23
). A function of the VSP repair process could be to counteract such mutations,
some of which would lead to potentially lethal amber mutations. However, VSP
repair interferes with other mismatch repair processes and will very strongly
drive other mutations which are not influenced by the hotspots (
18
,
19
,
21
,
22
), as is evidenced here by the very strong context effects for the Gln bias.
Another curious effect of VSP is that it will not only preserve but also create
the very sequence CCAGG that lead to the mutational hotspots.
The translation rates of the Glu codons have been measured
in vivo
(
33
) and GAA is translated 3.4 times faster (or about 0.11 s faster) than GAG,
independently of the following nucleotide being a G or a C. Since GAA is the
major codon, this suggests that codon bias may be determined by a selection for
speed of translation. However, when codon bias is considered in context (Fig.
1
), the curve before C is flat showing that there is no selective preference for
GAA in this case. There is a mutational bias for GAA in general and a selective
preference only before G. Thus, the faster translation, at least before C,
leads to no discernible selective advantage. With a growth rate of
k
0
= 10
-5
s
-1
(or about one doubling per 24 hours which may be reasonable under natural
conditions), Eqn 4 predicts that the selection coefficient would be about
s
= 10
-6
in high expression level genes. It does not appear likely that there is some
conflicting selection that exactly compensates for the slow translation. Since
there is no selective difference, suggesting that
N
e
s
<< 1 in Eqn 1, we would expect that the effective population size is
N
e
<10
5
. A recent analysis of the silent site diversity among
E.coli
strains (
34
) suggests that
N
e
= 2*10
8
. However, it is not certain that this number is also applicable in combination
with the selection coefficient as in Eqn 1 (
35
). Alternatively, codon evolution has taken place mostly at very slow growth
where the linear dependence between
s
and translation time, Eqn 3, may not hold. Whatever the reason, if this fairly
large difference in translation time does not lead to selection, it seems
likely that most of the codon bias is based on translational efficiency in some
other way than the speed, possibly accuracy (
36
).
The Glu and Lys bias before G is curious. The presumed Shine-Dalgarno matching does not seem to be inducive to frame shifts as
indicated by the lack of a strong position dependence beyond the first 100
codons. Furthermore, the lack of strong selective effects from other S-D matching bases simultaneous with the G at N1 may be an indication that
the resemblance to the S-D sequence is merely a coincidence. Based on the observed translation
rates discussed above, the bias is not caused by differences in the speed of
translation either. However, the position dependence in the first 100 codons,
which is similar to that of the overall bias, indicates that the origin of the
Glu and Lys bias before G is the same as that for the bulk of the codon bias.
Interestingly, there is no position dependence in the Glu and Lys bias before
bases other than G (data not shown); thus it is only the selective component of
the bias that is relaxed in the beginning of genes. This speaks against the
suggestion (
27
), at least for these codons, that there is a conflicting selection operating
only in the beginning of genes that is responsible for the reduction of bias in
the first 100 codons.
The positions P3 and N3 are special since they do not really conform to the
basic assumption of this study; when the context is considered at either of
these positions, a certain sequence can be avoided by a synonymous change
either in the codon considered or in the context. Thus, in this case it is not
truly a 2-fold degenerate choice. For instance, if the sequence G[brvbar]GA
The codon bias of a gene is very heterogeneous. It has been shown (
26
; see also Fig.
2
) that codon bias is substantially smaller in the first 50-100 codons of a gene and possibly also in the last 20 codons (
37
). In addition, there is a strong context effect that should be considered. The
average Lys bias, for instance, varies hardly at all across the groups of
different overall bias, suggesting that there is little or no selection on
synonymous codon choice in this case. However, the context dependent curves
(Fig.
1
) clearly show that there is large and even conflicting selection on these
codons, as discussed above. Similarly, the average bias of the Phe codons,
calculated without regard to context, shows a small increase with increasing
overall bias (cf. Fig.
2
), while the selective differences with regard to N1 context is very large (Fig.
1
).
Berg and Martelius (
15
) studied the relationship between selection, as given by the codon bias using
Eqn 1 without regard to context, and the synonymous substitution rates, as
calculated from the
E.coli
-
S.typhimurium
divergence. The heterogeneous selection from the context effects reported here
will change the predicted relationship between the apparent bias and apparent
substitution rate, calculated without regard to context. However, this change
will be small (O.G. Berg, unpublished), except possibly for Phe in the very
high CAI group, and will not influence the general conclusions drawn.
The results discussed above suggest that the translation time is of little
importance for the evolution of the codon bias. However, there is also strong
evidence that translation time is important: the tRNA levels seem to be such
that overall translation time is minimised (
38
; Berg and Kurland, in preparation). The estimated selection coefficient for
this optimisation is only marginally larger than that estimated above for the
selection of GAA over GAG. This could suggest that the effective population
size is very narrowly constrained around
N
e
= 10
5
to make one selection effective and not the other. Alternatively, a major part
of the codon bias may have evolved under conditions where
E.coli
is not under growth competition (
39
). The tRNA levels, on the other hand, are growth-rate dependent (
38
,
40
,
41
), and could be optimised primarily under fast growth where, presumably,
translation efficiency is most important.
We thank Mattias Martelius who carried out some preliminary studies on the
context dependence of the bias. This work was supported by grants from the
Swedish Natural Sciences Research Council and from the Human Capital and
Mobility Program of the European Union.
Expression level
CAI values
Number of genes
VL
CAI <= 0.270
258
L
0.270 < CAI <= 0.315
283
ML
0.315 < CAI <= 0.355
285
MH
0.355 < CAI <= 0.400
288
H
0.400 < CAI <= 0.475
256
VH
0.475 < CAI
279
Total
1649
REFERENCES
Return