A connection between bacterial chemotactic network and optimal filtering

Kento Nakamura and Tetsuya J. Kobayashi∗

Department of Mathematical Informatics, Graduate School of Information Science and Technology, the University of Tokyo
(Dated: May 28, 2020)
The chemotactic network of Escherichia coli has been studied extensively both biophysically and
information-theoretically. Nevertheless, the connection between these two aspects is still elusive.
In this work, we report such a connection by showing that a standard biochemical model of the
chemotactic network is mathematically equivalent to an information-theoretically optimal filtering
dynamics. Moreover, we demonstrate that an experimentally observed nonlinear response relation
can be reproduced from the optimal dynamics. These results suggest that the biochemical network
arXiv:2005.13208v1 [q-bio.CB] 27 May 2020

of E. coli chemotaxis is designed to optimally extract gradient information in a noisy condition.

Living things have developed sensory systems to be- difference ft between active and inactive states:
have appropriately in changing environments. One of the
most-analyzed such systems is the sensory system of Es- 1
at = . (1)
cherichia coli for chemotaxis. In E. coli chemotaxis, a cell 1 + exp(ft )
obtains information of a spatial gradient of a ligand from
The free energy difference ft comprises the additive effect
the temporal change in the ligand concentration that it
of the methylation level mt and of the ligand concentra-
experiences by swimming in the gradient. An E. coli cell
tion [L]t as
can sense a positive change in the ligand concentration
when it swims along the direction of the gradient and ft = N (−αmt + log[L]t ), (2)
vice versa. The swimming trajectory of E. coli consists
of a series of ballistic swimming called run interrupted where we omit a constant term and N, α > 0 are bio-
with random reorientations of direction called tumbling. chemical constants. Equations (1) and (2) take the form
By inhibiting the frequency of tumbling when it senses of the Monod-Wyman-Changeux (MWC) model describ-
a positive change in an attractant concentration, the E. ing allostery [15] where N specifies the receptor cooper-
coli cell can elongate the run length toward the direction ativity producing high sensitivity [4, 6, 7]. The methy-
of the higher concentration. lation level mt is modulated by the receptor activity at
The mechanism of the sensory system has been inten- as
sively studied both experimentally and theoretically. Ex- dmt
perimental studies have revealed the response of E. coli to = F (at ), (3)
various temporal profiles of concentration by measuring dt
behaviors of motor rotation [1, 2] and signaling molecules where F is assumed to be a monotonically decreasing
[3, 4]. Theoretical studies have proposed and analysed function. Since dat /dmt > 0 and F ′ (at ) < 0, the dy-
biochemical models that can reproduce properties of the namics of the methylation level mt with the function F
experimentally observed responses such as high sensitiv- constitutes a negative feedback regulation over the re-
ity to weak changes in concentration [4–7] and sensory ceptor activity at . Due to the negative feedback, this
adaptation [8]. Based on these works, Tu et al proposed biochemical network displays the sensory adaptation [8],
a simplified biochemical model [9], which can explain var- that is, when the concentration [L]t is stationary, the re-
ious aspects of the responses simultaneously [10]. This ceptor activity converges to a single value ā such that
biochemical model has been widely employed for various F (ā) = 0 which is independent of background concentra-
purposes such as analysis of sensory-motor coordination tion.
[11], fold-change detection [12, 13], and thermodynamics Although the biochemical model captures the integral
of sensory adaptation [14]. parts of the sensory system and its behaviors, there is
In Tu’s model [9], the sensory system consists of re- room for discussion from the view point of noise toler-
ceptor complexes, each of which takes either active or ance. Because the sensory system relies on stochastic
inactive state. Active receptors send a signal via media- ligand-receptor interactions and receptor modifications,
tor proteins and control the rotation of flagellar motors. sensing signal inevitably contains noise. This noise would
The ratio of active receptors, termed receptor activity cause a fatal influence on the chemotactic performance
at , is subjected to a feedback regulation through recep- because it can bury the actual temporal changes in con-
tor modification characterised by methylation level mt . centration and could end up with misdirections of the
The receptor activity at is determined by the free energy motor control. Therefore, the sensory system of E. coli
is expected to have a certain noise filtering property, and
several works have investigated impacts of noise in in-
formation transmission and favorable traits for noise fil-
∗ Also at Institute of Industrial Science, the University of Tokyo; tering [16]. However, these works focused on linear re-
[email protected] sponse by ignoring the underlying biochemical network
 2

and resultant nonlinear properties of the E. coli sensory describes the posterior probability Zt = P(Xt = −1 |
system. Even though some others considered a possi- Y0:t ) of the descending direction given the time series of
ble biochemical implementation of an ideal noise-immune the noisy sensing Y0:t := {Yt′ |t′ ∈ [0, t]} when its pa-
system based on nonlinear filtering theory [17], the corre- rameter values matches those of tumbling, run, gradi-
spondence with actual biological systems, especially that ent, and noise as R = ROPT := r+ + r− , p̄ = p̄OPT :=
of the gradient sensing in chemotaxis, is still elusive. r− /(r+ + r− ), K = KOPT := 2vc/σ.
In this paper, we utilize nonlinear filtering theory to de- Under this set of the optimal parameter values, the first
rive noise tolerant gradient sensing dynamics and demon- term represents the prediction based on a prior knowl-
strate its biochemical implementation in E. coli ’s cell. In edge about switching dynamics of direction Xt (Eq. (4)).
particular, we find that the derived ideal noise-filtering Thereby, without the second term (sensing signal), Zt
system excellently coincides with Tu’s biochemical model converges to the stationary probability of the direction
for the E. coli sensory system [9] and reproduces a non- p̄ for t → ∞. The second term corresponds to the up-
linear response relation measured experimentally. date of the posterior by new observation (Eq. (5)). The
As a minimal model of the temporal gradient sens- optimal gain of this term, KOPT , describes the signal-to-
ing, we consider a run-tumble motion of E. coli on one noise ratio because σ and 2vc specifies the noise intensity
dimensional axis along with monotonically increasing lig- and the steepness of the temporal change in the ligand
and concentration. This assumption is mainly due to the concentration during swimming, respectively. We call
limited capacity of the cell that may not be able to rec- the dynamics of Zt described by Eq. (6) the filtering
ognize the three dimensional physical space. Let ξt ∈ R dynamics hereafter.
and Xt ∈ {−1, +1} be the location and the direction of Next, we reveal the relation between the filtering dy-
swimming at time t ∈ [0, ∞). We assume that an E. namics and the biochemical network of E. coli chemotaxis
coli cell runs ballistically with a constant speed v > 0 by demonstrating that Eq. (6) can be equivalent to Eqs.
as dξt /dt = vXt and that each run and its direction (1),(2), and (3) if noise is neglected.
is interrupted by a stochastic tumbling motion. By ap- To this end, we introduce a coordinate transform from
proximating the tumbling motion by an instantaneous the posterior probability Zt to the log-likelihood ratio
event[18], we model the random changes in direction Xt θt := log Zt /(1 − Zt ). From the chain rule for deriva-
due to tumbling with a continuous-time Markov chain: tives, dθt /dt = (dθt /dZt )(dZt /dt), we obtain the follow-
ing equivalent representation of the filtering dynamics:
−r− r+
 
dpt
= pt , (4)
dt r− −r+ dθt Zt − p̄ dYt
=R −K ◦ . (7)
dt Zt (1 − Zt ) dt
T
where pt = (P(Xt = +1), P(Xt = −1)) , and r+ and
r− are the time-independent transition rates from −1 By further defining a new variable µt for the prediction
to +1 and from +1 to −1, respectively. Note that the dynamics as
transition rate of direction Xt would be smaller than the
rate of tumbling event because each tumbling does not dµt R Zt − p̄
:= − , (8)
always lead to the flipping of the direction. dt κ Zt (1 − Zt )
Next, we assume that the ligand concentration depends
exponentially on the location as [L]t ∝ exp(cξt ) where then we can formally integrate Eq. (7) as
c > 0 is a constant. This assumption is natural because  √ 
θt = −κµt + K log[L]t + σWt . (9)
the spatial distribution of a ligand typically obeys diffu-
sion. Then, we define a noisy sensing of the ligand by where we use Eq. (5) and κ > 0 is an arbitrary constant.
adding a noise term to the ligand-dependent term in Eq. Finally, Zt in Eq. (8) can be obtained by the inverse
(2) as transformation from θt to Zt :
√
Yt = − log[L]t − σWt (5) 1
Zt = . (10)
1 + exp(θt )
where Wt is the standard Wiener process and σ is the
intensity of noise. It should be noted that Wt can also These transformations unveil that Eqs. (10),(9), and
be interpreted as the noise from methylation [19] because (8) for the filtering dynamics are equivalent to Eqs.
the methylation level mt additively appears in Eq. (2). (1),(2), and (3) for the biochemical model of E. coli
By applying the nonlinear filtering theory under the chemotaxis, respectively (see also table S1 in SM for com-
above settings and assumptions [20], we can derive the parison).
following stochastic differential equation as The posterior probability Zt corresponds to the re-
dZt dYt ceptor activity at and they are both described by the
= −R(Zt − p̄) + KZt (1 − Zt ) ◦ , (6) sigmoidal function of θt and ft , respectively. The log-
dt dt
likelihood ratio θt is determined by the logarithm of the
where ◦ is the Stratonovich integral (See supplementary ligand concentration [L]t and the prediction term µt ,
material (SM) for details of derivation). This equation which corresponds to the dependence of the free energy
 3

difference ft on the ligand concentration [L]t and the 0.02

methylation level mt in Eq. (2). Finally, the dynamics filtering
of prediction term µt corresponds to that of the methy- 0.01 experiment
lation level mt .
0.00
Because the right-hand-side of Eq. (8) is a decreasing
function of Zt in the same way as the feedback func- −0.01

F(a)
tion F (at ) of mt , µt works as a negative feedback com-
ponent to Zt . Even though F (at ) in Tu’s model can- −0.02
not be determined biochemically but inferred only ex- −0.03
perimentally, the filtering dynamics provide a concrete
functional form for the feedback function, FOPT (Z) := −0.04
−(R/κ)·(Z−p̄)/(Z(1−Z)). Thus, if E. coli has developed
−0.05
the sensory system being tolerant to sensing noise near 0.0 0.2 0.4 0.6 0.8 1.0
optimally, the feedback function F describing the methy- a
lation dynamics can have a similar form as FOPT . To
test this expectation, we compare the feedback function
FIG. 1. Theoretically derived FOPT (red curve) fitted to the
FEXP inferred experimentally by a FRET measurement
experimentally obtained FEXP (black points) [21]. FOPT in
[21] with the theoretically predicted FOPT by adjusting the figure is obtained by modulating two parameters, R and
two free parameters R/κ and p̄. Figure 1 shows a notable p̄, as R/κ ≈ 2.8 × 10−3 and p̄ ≈ 0.28 (see also SM for the
agreement between the experimental data and theoretical fitting procedure).
prediction. Both FEXP and FOPT share a characteristic
nonlinearity; a gentle slope around a = 0.5 and a sharp
decline near a = 1. This result implies that the E. coli at the level of parameter values, though it is so at the
chemotactic network is designed structurally to be robust level of network structure. By considering the correspon-
to the sensory noise. In addition, because p̄ in FOPT cor- dence of N with the gain KOPT , which is determined by
responds to the stationary probability that the direction the speed of swimming, steepness of the gradient, and
of swimming is down the gradient, the parameter values intensity of sensing noise, N should not be fixed at cer-
p̄ ≈ 0.28 obtained by fitting implies that E. coli has a tain value but be variable depending on environmental
prior expectation that it likely swims up the gradient. situations. Several studies suggested that N as well as
We further investigate whether the biochemical pa- other parameters are diversified in a population of cells
rameters observed experimentally in laboratory environ- for hedging environmental uncertainties [25].
ments can satisfy the optimality in terms of filtering. To perform chemotaxis under the limitation in pa-
From the fitting of FOPT to FEXP , we have R/κ ≈ rameter adjustment, the robustness against the mis-
2.8 × 10−3 . κ can be estimated as κ = αN ≈ 12 by match of parameters could be beneficial. We inves-
comparing Eq. (2) and Eq. (9) and by employing a pre- tigate whether such robustness is endowed by exam-
vious estimate of α and N [21]. Thus, R is calculated as ining the filtering dynamics with misspecified param-
R ≈ 3.4 × 10−2 . In contrast, the optimal ROPT can be eter values of K. We measure the performance of
estimated from ROPT = r+ + r− and measurements of the dynamics using mean square error (MSE) defined
tumbling rate as 10−0.5 ≤ ROPT ≤ 100 s−1 [2, 22]. Thus, h R
T
i1/2
the obtained biochemical parameter R is much smaller as T1 t=0 {Xt − (1 − 2Zt )}2 dt in which 1 − 2Zt =
than the estimate ROPT from tumbling measurements. 1 − 2P(Xt < 0 | Y0:t ) = E[Xt | Y0:t ] holds for the opti-
This discrepancy may be attributed to three possi- mal parameter set. We define a reference value of K as
bilities: First, experimental conditions for the measure- Kref := N = 6 according to the correspondence between
ments of tumbling rate might not capture a wild condi- K and N . We set swimming speed to a physiologically
tion where E. coli cells are supposed to perform chemo- relevant value: v = 20µm · s−1 . The rates of directional
taxis. Recent studies suggest that swimming behavior in changes are determined as r+ = R(1−p̄), r− = Rp̄ so that
polymeric solutions or soft agar is different from that un- the values of R and p̄ obtained by fitting in Fig. 1 be-
der a liquid condition used in most experiments [23]. In come optimal. We define the reference of the steepness of
particular, the tumbling frequency is shown to decrease gradient as cref := 10−3 µm−1 by taking into account con-
with addition of polymeric molecules due to remodeling ditions in previous simulation studies [11]. We also de-
of signaling pathway downstream of sensory system or fine the reference of noise intensity as σref := 2cref v/Kref
possibly due to motor load. In such a case, ROPT may such that the reference parameter Kref is optimal un-
take smaller value. Second, the values of R might be un- der c = cref and σ = σref . Note that Kref is also opti-
derestimated because of the difficulty in estimating the mal on the half-line, (σ, c) = η(σref , cref ), η > 0, because
biochemical parameter N . Although we used an estimate 2vc/σ = 2vcref /σref = Kref holds on it.
N = 6 in previous studies [7, 9, 21], other estimates of Figure 2 shows MSEs of Eq. (6) for different K as
N are larger, N = 15 ∼ 20 [7, 24]. The last possibility functions of σ with fixed c = cref (A) and as functions
is that the system is not or cannot be always optimized of c with fixed σ = σref (B). The error with fixed K is
 4

(A) (B) pair the performance of the dynamics for any K. We can
see a similar trend in Fig. 2 (C) and (D). These results
indicate that even under the misspecification of K asso-
ciated with parameters σ and c, the filtering dynamics
still reliably and robustly estimate temporal gradient if
the change in σ and c is one that increases SNR.
Small value of gain K is optimized to a low SNR situ-
ation, and variation of MSE between low and high SNRs
(C) (D) is small (Fig. 2). In contrast, large K adjusted to a
high SNR one shows a significant variation in MSE be-
tween low and high SNR cases. This means that low K
can work moderately well for most of conditions whereas
large K can work much better if the environmental SNR
is large enough at the cost of lower performance under
low SNR situations. Thus, K modulates the balance of
risk-averting and -taking strategies of sensing.
FIG. 2. MSE of the filtering dynamics as a function of σ The growth-dependent variability of K can coordi-
with fixed c = cref (A), as a function of c with fixed σ = σref nate such risks at the level of population [26]. More-
(B), as a function of K and c with fixed σ = σref (C), and over, N , which biochemically corresponds to K, is sug-
as a function of σ and c with fixed K = Kref (D). Curves in gested to vary temporally at the single-cell-level [24, 27]
(A) and (B) represent MSEs with fixed parameter K = Kref via a receptor cluster rearrangement. The integration of
(blue) and with the optimal parameter K = KOPT = 2vc/σ biochemical modeling and optimal filtering theory could
(red). White lines in (C) and (D) represent the parameter work for further analysis of such a gain adaptation of
region on which the parameter K is set optimal i.e. 2vc/σref = cells. This approach may also apply to other sensory
K (C) and 2vc/σ = Kref (D). systems with allosteric receptors and a negative feedback,
e.g., G protein-coupled receptors for vision and EGF re-
ceptor in animal cells, whose models can be reduced to
always greater than or equal to that with K adjusted similar biochemical models to Tu’s [9]. Furthermore, we
to KOPT . For each fixed gain K, MSE monotonously might extend it to directly include the closed cycle be-
increases as the signal-to-noise ratio (SNR) decreases ei- tween sensing of environment and the resultant actions
ther by the increase in the noise intensity σ (Fig. 2(A)) of cells.
or by the decrease in gradient steepness c (Fig. 2(B)), This research is supported by JSPS 19H05799 and
indicating that greater SNR than optimal one never im- JSPS 20J21362.

[1] S. M. Block et al. Cell 31, 215 (1982); J. E. Segall, S. M. Wolde, and F. Tostevin, Phys. Rev. Lett. 109, 148101
Block, and H. C. Berg, Proc. Natl. Acad. Sci. U.S.A. 83, (2012); Y. S. Dufour, X. Fu, L. Hernandez-Nunez, and
8987 (1986). T. Emonet, PLoS Comput. Biol. 10, e1003694 (2014);
[2] S. M. Block, J. E. Segall, and H. C. Berg, J. Bacteriol. R. He, R. Zhang, and J. Yuan, Biophys. J. 111, 430
154, 312 (1983). (2016); J. Long, S. W. Zucker, and T. Emonet, PLoS
[3] V. Sourjik and H. C. Berg, Proc. Natl. Acad. Sci. U.S.A. Comput. Biol. 13, e1005429 (2017).
99, 123 (2002). [12] O. Shoval, L. Goentoro, Y. Hart, A. Mayo, E. Sontag,
[4] V. Sourjik and H. C. Berg, Nature 428, 437 (2004). and U. Alon, Proc. Natl. Acad. Sci. U.S.A. 107, 15995
[5] D. Bray, M. D. Levin, and C. J. Morton-Firth, Nature (2010); M. D. Lazova, T. Ahmed, D. Bellomo, R. Stocker,
393, 85 (1998); T. Duke and D. Bray, Proc. Natl. Acad. and T. S. Shimizu, ibid. 108, 13870 (2011).
Sci. U.S.A. 96, 10104 (1999). [13] N. Olsman and L. Goentoro, Proc. Natl. Acad. Sci.
[6] J. E. Keymer, R. G. Endres, M. Skoge, Y. Meir, and U.S.A. 113, E4423 (2016).
N. S. Wingreen, Proc. Natl. Acad. Sci. U.S.A. 103, 1786 [14] G. Lan, P. Sartori, S. Neumann, V. Sourjik, and Y. Tu,
(2006). Nat. Phys. 8, 422 (2012); A. C. Barato, D. Hartich, and
[7] B. A. Mello and Y. Tu, Biophys. J. 92, 2329 (2007). U. Seifert, New J. Phys. 16, 103024 (2014); S. Ito and
[8] N. Barkai and S. Leibler, Nature 387, 913 (1997); T.-M. T. Sagawa, Nat. Commun. 6, 1 (2015).
Yi, Y. Huang, M. I. Simon, and J. Doyle, Proc. Natl. [15] J. Monod, J. Wyman, and J.-P. Changeux, J. Mol. Biol.
Acad. Sci. U.S.A. 97, 4649 (2000). 12, 88 (1965).
[9] Y. Tu, T. S. Shimizu, and H. C. Berg, Proc. Natl. Acad. [16] B. W. Andrews, T.-M. Yi, and P. A. Iglesias, PLoS
Sci. U.S.A. 105, 14855 (2008). Comput. Biol. 2, e154 (2006); F. Tostevin and P. R.
[10] Y. Tu, Annu. Rev. Biophys. 42, 337 (2013). Ten Wolde, Phys. Rev. Lett. 102, 218101 (2009); T. Mora
[11] L. Jiang, Q. Ouyang, and Y. Tu, PLoS Comput. and N. S. Wingreen, ibid. 104, 248101 (2010); P. Sartori
Biol. 6 (2010); M. Flores, T. S. Shimizu, P. R. ten and Y. Tu, J. Stat. Phys. 142, 1206 (2011); G. Aquino,
 5

L. Tweedy, D. Heinrich, and R. G. Endres, Sci. Rep. 4, (2010).

5688 (2014); C. C. Govern and P. R. ten Wolde, Proc. [22] O. Pohl, M. Hintsche, Z. Alirezaeizanjani, M. Seyrich,
Natl. Acad. Sci. U.S.A. 111, 17486 (2014); N. B. Becker, C. Beta, and H. Stark, PLoS Comput. Biol. 13, e1005329
A. Mugler, and P. R. ten Wolde, Phys. Rev. Lett. 115, (2017).
258103 (2015); G. Micali and R. G. Endres, Sci. Rep. 9, [23] A. Patteson, A. Gopinath, M. Goulian, and P. Arra-
1 (2019). tia, Sci. Rep. 5, 15761 (2015); J. D. Partridge, N. T.
[17] T. J. Kobayashi, Phys. Rev. Lett. 104, 228104 (2010); Nhu, Y. S. Dufour, and R. M. Harshey, MBio 10, e00316
C. Zechner, G. Seelig, M. Rullan, and M. Khammash, (2019).
Proc. Natl. Acad. Sci. U.S.A. 113, 4729 (2016); T. Mora [24] R. G. Endres, O. Oleksiuk, C. H. Hansen, Y. Meir,
and I. Nemenman, Phys. Rev. Lett. 123, 198101 (2019). V. Sourjik, and N. S. Wingreen, Mol. Syst. Biol. 4 (2008).
[18] We assume that the change of direction occurs only at [25] V. Sourjik and N. S. Wingreen, Curr. Opin. Cell Biol. 24,
the tumbling event, in other words, we neglect directional 262 (2012); A. J. Waite, N. W. Frankel, and T. Emonet,
change in run phase caused by viscosity. Annu. Rev. Biophys. (2018).
[19] E. Korobkova, T. Emonet, J. M. Vilar, T. S. Shimizu, [26] C. M. Khursigara, G. Lan, S. Neumann, X. Wu, S. Ravin-
and P. Cluzel, Nature 428, 574 (2004); T. Emonet and dran, M. J. Borgnia, V. Sourjik, J. Milne, Y. Tu, and
P. Cluzel, Proc. Natl. Acad. Sci. U.S.A. 105, 3304 (2008). S. Subramaniam, EMBO J. 30, 1719 (2011).
[20] A. H. Jazwinski, Stochastic processes and filtering theory [27] V. Frank and A. Vaknin, Mol. Microbiol. 88, 634 (2013).
(Courier Corporation, 2007).
[21] T. S. Shimizu, Y. Tu, and H. C. Berg, Mol. Syst. Biol. 6