of 12
On noise processes and limits of performance in biosensors
Arjang Hassibi
a

Electrical and Computer Engineering Department, University of Texas at Austin, Austin, Texas 78712
Haris Vikalo and Ali Hajimiri
Electrical Engineering Department, California Institute of Technology, Pasadena, California 91125

Received 11 January 2007; accepted 4 May 2007; published online 11 July 2007

In this paper, we present a comprehensive stochastic model describing the measurement uncertainty,
output signal, and limits of detection of affinity-based biosensors. The biochemical events within the
biosensor platform are modeled by a Markov stochastic process, describing both the probabilistic
mass transfer and the interactions of analytes with the capturing probes. To generalize this model
and incorporate the detection process, we add noisy signal transduction and amplification stages to
the Markov model. Using this approach, we are able to evaluate not only the output signal and the
statistics of its fluctuation but also the noise contributions of each stage within the biosensor
platform. Furthermore, we apply our formulations to define the signal-to-noise ratio, noise figure,
and detection dynamic range of affinity-based biosensors. Motivated by the platforms encountered
in practice, we construct the noise model of a number of widely used systems. The results of this
study show that our formulations predict the behavioral characteristics of affinity-based biosensors
which indicate the validity of the model. ©
2007 American Institute of Physics
.

DOI:
10.1063/1.2748624

I. INTRODUCTION
Affinity-based biosensors use selective binding and in-
teraction between certain biomolecules

recognition probes

and specific target analytes to determine the presence of the
latter in the biological samples. The essential role of the
biosensor platform

e.g., microarrays,
1
,
2
and immunoassays
3

is to facilitate the binding of the probe-target complexes to
produce a
detectable
signal, which correlates with the pres-
ence of the target and conceivably its abundance. The mini-
mal components required for affinity-based detection include
a molecular recognition layer

capturing probe

integrated
within or intimately associated with a signal-generating
physiochemical transducer
4
and a readout device.
To generate target-specific signal, the target analytes in
the sample volume first need to collide with the recognition
layer, interact with the probes, and ultimately take part in a
transduction process. The analyte motion in typical biosensor
settings

e.g., aqueous biological mediums

is dominated by
diffusion spreading, which from a microscopic point of view
is a probabilistic mass-transfer process

i.e., random-walk
events for a single analyte molecule
5

. Accordingly, the ana-
lyte collisions with the probes become stochastic processes.
Moreover, because of the quantum-mechanical nature of
chemical bond forming,
6
9
interactions between the probes
and the analyte molecules are also probabilistic, adding more
uncertainty
to the biosensing procedure. On top of these two
processes, we also have the transducer and readout circuitry,
which likely adds more noise to the already noisy process.
Beside the inevitable uncertainty associated with the
binding, transduction, and readout, in all practical biosen-
sors, binding of other species to the probes

nonspecific
binding

is also possible. Nonspecific binding

e.g., cross-
hybridization in DNA microarrays
1

is generally less prob-
able than the specific binding when target analyte and the
interfering species have the same abundance. Nonetheless, if
the concentration of the nonspecific analyte becomes much
higher than the target analyte, nonspecific bindings may
dominate the measured signal and hence limit the minimum-
detectable level

MDL

of the biosensor platform.
10
In this
paper, we examine all of the aforementioned processes and
assess their contributions to the biosensor output signal. Our
goal is to introduce a general methodology to model the
noise in biosensors, and based on that, evaluate the MDL,
highest detection level

HDL

, and detection dynamic range

DR

.
Initially, in Sec. II, we review the existing stochastic
model of the biosensor platforms,
11
which is essentially a
Markov model for the analyte capturing. We proceed by add-
ing to it a probabilistic model for the transduction, probe
saturation, and the readout process. In Sec. III, we implement
this model to evaluate the uncertainty of the detected signal
and the relative noise contributions of all underlying pro-
cesses as a function of the analyte concentrations, capturing
probe density and other characteristics of the biosensor sys-
tem. Finally, in Sec. IV, demonstrating the practical use of
our model, we implement the model to describe the stochas-
tic behavior of impedimetric
12
and fluorescent-based
13
bio-
sensors.
II. BIOSENSOR STATISTICAL MODEL
A. Analyte motion
Molecules, cells, and many other analytes immersed in
the aqueous mediums of biosensor platforms are subject to
thermal fluctuations and, in particular, sensory platforms,
a

Previously at Electrical Engineering Department, California Institute of
Technology, Pasadena, CA 91125; electronic mail: arjang@mail.utexas.edu
JOURNAL OF APPLIED PHYSICS
102
, 014909

2007

0021-8979/2007/102

1

/014909/12/$23.00
© 2007 American Institute of Physics
102
, 014909-1
Downloaded 25 Jul 2007 to 131.215.225.175. Redistribution subject to AIP license or copyright, see http://jap.aip.org/jap/copyright.jsp
subject to electromagnetic or mechanical forces. While me-
chanical movement

e.g., convection

and electromagnetic
forces

e.g., electrophoresis

are in some sense deterministic,
thermal fluctuations are random in nature. Thermal fluctua-
tions of a particle from a microscopic point of view follow
the characteristics of a typical random-walk process
5

i.e.,
Brownian motion

, which results in a diffusive spreading
phenomenon in macroscopic systems.
It is challenging to statistically describe the motion of
individual analyte molecules in the general case using the
continuity equation formulation. However, we can utilize a
discrete-time Markov process
14
to model the particle sto-
chastic behavior within the reaction chamber.
11
In this
model, we discretize the coordinates of the reaction chamber,
where the states of the Markov model correspond to a fixed
set of coordinates within the chamber

Fig.
1

. The transition
probabilities of the Markov process are defined as the prob-
abilities of an analyte moving from one coordinate

state

to
another in the given time epoch

t
. Such a simple model
captures the stochastic nature of the process yet is still trac-
table enough to allow for computation of the probability den-
sity function

pdf

of the analyte particles. Additionally, we
can easily derive the expected behavior

ensemble average

of the system, a quantity that is generally assessed using a
continuity-based equation formulation.
For two arbitrary coordinates
v
i
and
v
j
in the system,
associated with states
i
and
j
, the state transition probability
from state
i
to
j
, for an analyte particle in the time interval

t
, is denoted by
m
i
,
j
,
m
i
,
j
=Pr

v


t

=
v
i

v

0

=
v
j

,

1

where
v

0

and
v


t

specify the coordinate of an analyte
particle at times 0 and
t
, respectively. The
m
i
,
j
defined in

1

represents the

i
,
j

element of the analyte transition matrix
M


t

M


t


R
N

N
and
N
is the total number of states

.If
an analyte particle has the probability distribution
x

0

at
time zero across the state space, where for all
tx

t


R
N
,
then at

t
,
x


t

=
M


t

x

0

.

2

Typical biosensor structures have a large number of analyte
particles within the reaction chamber. Provided that their mo-
tion is statistically independent,
M


t

and
x

t

become in-
dependent, resulting in a homogeneous Markov process.
Note that this is, in fact, a realistic assumption for most bio-
sensors, because statistical motion in such systems is gov-
erned by analyte interaction with the medium molecules and
temperature, and not by analyte-analyte interactions. Hence,
analytes seldom affect the probabilistic motion of one an-
other. Because of the homogeneity of the Markov process,
we are able to calculate the spatial probability distribution of
the analyte particles at all time instants that are multiples of

t
, given the initial distribution
x

0

and
M


t

,by
x

k

t

=

M


t

k
x

0

,

3

where
k
is an integer. If we have
n
analyte particles in the
system we can also define a spatial concentration distribution
of the analyte particles
X

t

, where
X

t

=
nx

t

and hence
X

t

=
nx

k

t

=
n

M


t

k
x

0

=

M


t

k
X

0

,

4

where
X

0

is the initial concentration of the analyte par-
ticles. It is important to realize that the matrix
M


t

can
theoretically be estimated for any small time increment

t
,
given the exact statistics of the mass transport process de-
scribed by the continuity equation.
11
B. Analyte capture
Now that we have a statistical model of analyte motion
within the solution, we need to incorporate the boundary
conditions within the biosensor chamber. The boundary con-
ditions for conventional biosensors are either purely reflec-
tive

e.g., inert chamber walls and solution surfaces

or se-
lectively absorbing

e.g., surfaces where the capturing probes
exist

. Incorporation of reflective boundaries into the Markov
model is carried out by assigning zero probability for par-
ticles to move beyond the boundaries. However, accommo-
dation of selective absorption requires a probabilistic model
for the collision and interaction processes.
Analyte collision with the binding sites may be reactive,
elastic, or inelastic.
15
We have previously introduced a
probabilistic model to describe the specific binding of an
analyte particle
X
to a single probe
Y
in an affinity-based
biosensor.
11
To begin with, we assume that the probe
Y
is
confined

immobilized

in a certain coordinate. Now we can
argue that any meaningful interaction between
X
and
Y
at
time
t
only occurs if
v

t

=
v
1

i.e., molecule
X
is in intimate
proximity of
Y
, as shown in Fig.
2

. If the bulk-phase reac-
tion between the analytes
X
, and the capturing probes
Y
has
an association rate
k
1
, and disassociation rate
k
−1
, then we
can write
X
+
Y
k
1
,
k
−1
XY
d

XY

dt
=
k
1

X

Y

k
−1

XY

,

5

where the symbol

indicates concentration of the species.
Now based on

5

we can apply the following approximation
to find the transition probabilities between the captured state
c
of the analyte and the collided state 1:
11
FIG. 1. Markov model for probabilistic motion in the reaction chamber.
Each coordinate corresponds to a state of the process into which target
analytes can move or from which they can leave.
v

t

represents the location
of the analyte particle at time
t
, and
v
i
defines the coordinates of the state
i
in the system.
014909-2 Hassibi, Vikalo, and Hajimiri
J. Appl. Phys.
102
, 014909

2007

Downloaded 25 Jul 2007 to 131.215.225.175. Redistribution subject to AIP license or copyright, see http://jap.aip.org/jap/copyright.jsp
m
1,
c
=

k
1

t
,
m
c
,1
=
k
−1

t
.

6

The parameters
m
1,
c
and
m
c
,1
are also defined as the associa-
tion and disassociation transition probabilities, respectively.
The quantity

basically describes the effects of immobiliza-
tion of probes and the finite reactive distance between the
probe and target. Using

6

for absorbing boundaries and

4

for the mass transfer and reflection cases, we can analyze the
binding and motion of the analytes in the biosensor systems
in its entirety.
The formulation in

6

assumes infinite capturing capac-
ity for probes, which is not always realistic. If there are
n
1,
Y
immobilized capturing sites in state 1, then the association
and disassociation probabilities for the immobilized
Y
spe-
cies become functions of
n
c

t

, the number of captured ana-
lytes at state
c
, such that
m
c
,1

t

=

k
1

n
1,
Y
n
c

t

n
1,
Y


t
,
m
1,
c
=
k
−1

t
,

7

where the term

n
1,
Y
n
c

t

/
n
1,
Y
is basically the probability
of finding a free probe at the capturing site
c

see Fig.
3

. The
incorporation of

7

into

4

clearly results in a nonhomoge-
neous Markov process, since the transition probabilities have
become a function of target analyte distribution, i.e.,
M


t

in

4

becomes a function of
n
c

t

. Finding a closed form for
the dynamics of such system is challenging; however, we
will demonstrate in Sec. III how in chemical equilibrium the
fluctuations of the system can be quantified.
C. Transduction
We define the transduction process and the output signal
of the biosensor as a function of
X

t

, the target concentra-
tion distribution within the system. We can represent trans-
duction process by matrix
T
, where
T

R
1


N
+
P

and
P
being
the number of detectable capturing states. Accordingly, the
output signal
s

t

is defined by
s

t

=
TX

t

+
u
T

t

,

8

where
u
T

t

is a random process describing the additive noise
of the transducer. The matrix
T
in

8

signifies the process
wherein binding results in a detectable signal or parameter.
For example, the unit of
T
in a chemiluminescence-based
biosensor
16
is photons/s target, while in impedimetric biosen-
sors are

−1
/target

see Sec. IV

. As implied by

8

, the
transduction process itself may introduce some uncertainty in
the form of an additive noise component which is repre-
sented by
u
T

t

. We know from

4

that
X

t

is a random
process, but if we assume that its expected behavior

en-
semble average

is denoted by
X

t

, we can rewrite

8

as
s

t

=
T

X

t

+
u
X

t


+
u
T

t

=
T
X

t

+
Tu
X

t

+
u
T

t

.

9

In

9

, the random process
u
X

t

essentially describes the
deviation of the actual analyte concentration

perturbation

from the expected distribution
X

t

, where
u
X

t

=0 for all
t
.
The function
Tu
X

t

therefore describes the observed
bio-
chemical noise
of the system.
Since the transduction mechanism varies tremendously
among biosensor platforms, we need to reexamine

9

and
derive a formulation which makes their performances com-
parable. One approach that we propose in this paper is to
refer the biosensor output signal to the input of the biosensor
system which is essentially the number of analytes in the
sample. In this approach the biosensor equivalent signal
s
b

t

represents the number of
perceived
target analytes in the
sample plus its uncertainty. To carry out the procedure, we
divide

9

by the effective gain of the biosensor system
which is
T

=

s

t

/

n
=
T
X

t

/
n
. The parameter
T

is basi-
cally the expected output signal change per analyte concen-
tration change within the system. Now the biosensor master
equation expressed in terms of the number of analytes is
given by
s
b

t

=
n
+

T

−1
T

u
X

t

+
T

−1
u
T

t

.

10

If there are
K
stages after transducer

see Fig.
4

, each with
gains
G
i
and additive noise
u
i

t

i
=1,...,
K

, we can incor-
porate them in

10

such that
s
b

t

=
n
+
T

−1

Tu
X

t

+
u
T

t

+

i
=1
K
j
=1
i
G
j
−1
u
i

t


.

11

Equations

10

and

11

essentially provide the number of
observed

or existing

target analytes in the biosensor sys-
tem. In the next section we apply these stochastic formula-
FIG. 2. Markov model for probabilistic capture in the reaction chamber.
Only when the target analytes get into intimate proximity of the capturing
probes

state 1

there is a probability of capturing denote by
m
c
,1
.
FIG. 3. Modification of association transition probability
m
c
,1
, due to probe
saturation.
014909-3 Hassibi, Vikalo, and Hajimiri
J. Appl. Phys.
102
, 014909

2007

Downloaded 25 Jul 2007 to 131.215.225.175. Redistribution subject to AIP license or copyright, see http://jap.aip.org/jap/copyright.jsp
tions to calculate the noise characteristics of biosensor plat-
forms.
III. NOISE AND SIGNAL FLUCTUATION
At this point we can utilize the transition matrix derived
in

4

along with the correct boundary conditions of

7

to
simulate various characteristics of the biosensor, and use

9


11

to assess the output signal of the system. The simu-
lation of such a system although possible is generally cum-
bersome, since the number of states is extremely large in
practical biosensors. Hence, in the following subsections we
will derive the closed form expressions describing the fluc-
tuation of biosensor output. In particular, we will focus on
the noise at biochemical equilibrium which, for many prac-
tical reasons, is the state where the measurements are pref-
erably carried out.
A. Biochemical fluctuation in equilibrium
Tu
X
,
E
t
...‡
We can show that when there is no saturation in the
biosensor, the general solution for
x

t

and
X

t

which de-
fines the single analyte probability distribution at time
t
,
given an initial condition
x

0

, is described by the
following:
11
x

t

= exp

M


t

I


t
t
x

0

=
H

t

x

0

,

12

where
I
is the identity matrix of the same dimensions as
M


t

. Matrix
H

t

advances the initial distribution in time
and is calculated using the matrix exponential function. It
can be shown that
M


t

has a single eigenvalue equal to 1,
and all other eigenvalues of
M


t

have magnitudes less
than 1, ensuring a single equilibrium distribution,
x
E
which is
the eigenvector associated with an eigenvalue of 1. Further-
more, it can be shown that
M


t

and
H

t

share
x
E
, i.e., it
holds that
x
E
=
M


t

x
E
=
H

t

x
E
.

13

To quantify the
biochemical
noise of the system
Tu
X

t

,we
need to characterize the analyte fluctuation in every state,
and particularly in the states for which the transduction oc-
curs

i.e., the states corresponding to nonzero entries of
T

.
To perform this calculation for the biochemical equilibrium,
we first need to evaluate the autocorrelation function of the
stationary random processes describing the occupancy of the
states denoted by
O
i
,
j

t

, where 1

i

N
describes the state
of the system and 1

j

n
represents the target analytes. The
occupancy
O
i
,
j

t

is essentially an indicator function of the
form
O
i
,
j

t

=
1
if
v
j

t

=
v
i
0
otherwise,

14

which means that
O
i
,
j

t

=1 if the
j
th particle occupies state
i
at time
t
, and zero otherwise. It can be shown that
R
i
,
j



, the
autocorrelation of
O
i
,
j

t

, is identical for all
n
analytes and
can be found using the definition of
H

t

to be
R
i
,
j



=
h
ii



x
i
,
E
,

15

where
x
i
,
E
is the
i
th component of
x
E
and
h
ii



is the
i
th
diagonal entry of
H



.
11
Since we have
n
independent par-
ticles in the system, we define the stochastic process
n
i

t

counting the total number of particles occupying state
i
as
n
i

t

=

j
=1
n
O
i
,
j

t

,

16

and hence
X

t

=

n
1

t

n
2

t

̄
n

N
+
P


t

. Now by implement-
ing

15

, we can derive
R
n
i



, the autocorrelation function of
n
i

t

,as
R
n
i



=

i
=1
n
O
i
,
j

t
+



i
=1
n
O
i
,
j

t

=

n
2
n

O
i
,
j

t

2
+
nR
i
,
j



=

n
2
n

x
i
,
E

2
+
nh
ii



x
i
,
E
.

17

Accordingly, the unilateral

single-sided

power spectral den-
sity

PSD

of
n
i

t

defined by
S
n
i

f

which describes the fluc-
tuations as a function of frequency becomes
S
n
i

f

=2


n
2
n

x
i
,
E

2

f

+2
nS
O
i

f

,

18

where
S
O
i

f

=
x
i
,
E

+
h
ii



e
j
2

f

d

.

19

To find
n
i
2
, the variance of
n
i

t

, we have
n
i
2
=

i
=1
n
O
i
,
j

t


i
=1
n
O
i
,
j

t



i
=1
n
O
i
,
j

t


2
=
nx
i
,
E

1−
x
i
,
E

.

20

It is evident from

4

and

13

that the expected value of
n
independent particles in state
i
becomes
nx
i
,
E
. Thus,
X
E
the
expected concentration at equilibrium becomes
FIG. 4. Block diagram of different stages of a general biosensor platform.
Various noise processes originating from different sources may corrupt the
signal; however, all of these noise sources can be referred to as the number
of analytes in the sample.
014909-4 Hassibi, Vikalo, and Hajimiri
J. Appl. Phys.
102
, 014909

2007

Downloaded 25 Jul 2007 to 131.215.225.175. Redistribution subject to AIP license or copyright, see http://jap.aip.org/jap/copyright.jsp
X
E
= lim
t
X

t

= lim
t

n
1

t

n
2

t

̄
n

N
+
P


t

T
=
nx
E
.

21

Based on the definition of
u
X

t

, now for the equilibrium
fluctuation
u
X
,
E

t

=lim
t
u
X

t

we can derive the expected
value
u
X
,
E

t

, and the variance
X
2
, where
u
X
,
E

t

=0,
X
2
=
n

x
1,
E

1−
x
1,
E

x
2,
E

1−
x
2,
E

̄
x

N
+
P

,
E


1−
x

N
+
P

,
E

T
.

22

Now by using

22

, we easily compute for the expected value
of the output noise fluctuation
Tu
X
,
E

t

and its variance
S
2
at
equilibrium
Tu
X
,
E

t

=0,
S
2
=
T

2

X
2
,

23

where
T

2

=

t
1
2
t
2
2
̄
t

N
+
P

2

is the element-by-element
square of the transduction matrix
T
=

t
1
t
2
̄
t

N
+
P


.Itis
important to realize that
S
2
is the variance of the unavoidable
noise of the system, which is basically the biochemical shot
noise originating from the quantum nature of binding events.
This particular noise is inherent to affinity-based biosensors
and the uncertainty

and certainty

limitation which it estab-
lishes is often referred to as the quantum-limit

QL

detec-
tion level.
17
,
18
Note that while

13


23

are derived assuming a homo-
geneous process

i.e., no probe saturation and no analyte-
analyte molecular interactions

, the same methodology can
be adopted for nonhomogeneous processes, given that an
equilibrium distribution exists for the system. The only nec-
essary modification is that a transition matrix at the equilib-
rium for nonhomogeneous process must be derived to calcu-
late
h
ii



. Essentially in this approach we perform
linearization
, and the method is thus valid when the fluctua-
tions of the concentration are small compared to the en-
semble average concentration

i.e., small perturbation

.
B. Output signal fluctuation at equilibrium
Tu
X
,
E
t
...
+
u
T
t
...‡
Although the biochemical noise can easily be quantified
by the capturing probability and vector
x
E
, presenting a gen-
eral stochastic model for the transducer noise
u
T

t

appears
to be more difficult. The main reason is that the additive
noise of the transduction process is platform dependent and
may originate from fundamentally different physical pro-
cesses for different transducers. Nevertheless, we can still
evaluate the characteristics of the system for a number of
specific cases.
The simplest case is when
u
T

t

is an additive signal
independent of the binding process, with expected value of
zero and variance of
T
2
. In this case the observed output
noise of the biosensor becomes
Tu
X
,
E

t

+
u
T

t

with the fol-
lowing expected value
Tu
X
,
E

t

+
u
T

t

and variance
S
2
:
Tu
X
,
E

t

+
u
T

t

=0,
s
2
=
T

2

X
2
+
T
2
.

24

Now that we have the fluctuation characteristics of the bio-
chemical noise at equilibrium, we can revisit the biosensor
master equation. Based on

21

, we have
T

=
T
X

t

/
n
=
Tx
E
.
Therefore, the master equation at equilibrium, constraining
our considerations to only biochemical noise and additive
transducer noise, becomes
s
b
,
E

t

= lim
t
s
b

t

=
n
+
T

−1

Tu
E
,
X

t

+
u
T

t

=
n
+
Tu
X
,
E

t

+
u
T

t

Tx
E
.

25

C. Signal-to-noise ratio and noise figure
definitions
We define the signal-to-noise ratio

SNR

of a biosensor
platform as the ratio of the square of the detected target
analyte

i.e., signal power

over its noise variance. Using

25

, we can write the general expression for the SNR as
SNR =

s
b
,
E

t

2

s
b
,
E

t

s
b
,
E

t

2
.

26

By employing

24

and

25

, we can rewrite

26

as
SNR =
n
2
T

−2

T

2

X
2
+
T
2

=

nT


2
T

2

X
2
+
T
2
,

27

or
SNR =
n
T

2

t
1
2
x
1,
E

1−
x
1,
E

+
t
2
2
x
2,
E

1−
x
2,
E

+
̄
+
t

N
+
P

2
x

N
+
P

,
E

1−
x

N
+
P

,
E

+
T
2
/
n
.

28

The first term in the denominator of

26

and

27

describes
the biochemical shot noise

Poisson noise

of the biosensor,
while the second term is the added noise of the transducer
. It is again important to emphasize that the biochemical shot
noise is unavoidable and essentially inherent to the system.
Hence, the SNR of the biosensor excluding any transducer
noise

i.e., ideal detector

has an upper bound defined by
this shot noise which is generally referred to as the QL SNR.
014909-5 Hassibi, Vikalo, and Hajimiri
J. Appl. Phys.
102
, 014909

2007

Downloaded 25 Jul 2007 to 131.215.225.175. Redistribution subject to AIP license or copyright, see http://jap.aip.org/jap/copyright.jsp
We can further generalize

27

for the case where the
biosensor comprises of other amplification stages as in

11

.
In this case, the SNR becomes
SNR =

nT


2
T

2

X
2
+
T
2
+

i
=1
K
j
=1
i
G
j
−1
i
2
,

29

where
i
2
represents the variance of the noise added by the
i
th amplification stage.
In addition to SNR, we can define the noise figure

NF

of the biosensor as another figure of merit for the sensing
system. This figure of merit essentially describes the ratio of
the added noise to the input noise which is the unavoidable
shot noise. We define NF as
NF

total output noise power
output noise power due to input source
,

30

and by applying

23


25

and for simplicity referring every-
thing to the transducer output we have
NF=1+
T
2
T

2

X
2
+

i
=1
K
j
=1
i
G
j
−1
i
2
T

2

X
2
=1+
1
n
T
2
+

i
=1
K
j
=1
i
G
j
−1
i
2
t
1
2
x
1,
E

1−
x
1,
E

+
t
2
2
x
2,
E

1−
x
2,
E

+
̄
+
t

N
+
P

2
x

N
+
P

,
E

1−
x

N
+
P

,
E

.

31

Equations

28

and

31

both indicate that by increasing
the concentration

or quantity

of the analytes, the system
generates more noise although SNR increases and NF de-
creases. This phenomenon, which one expects from the in-
herent shot noise of binding events, sets no upper bound for
the concentration levels that the biosensor is able to detect.
However, from

7

we know that this is not practically jus-
tifiable since the probe saturation will eventually occur as we
keep increasing the amount of analytes. In the next subsec-
tion, we will examine saturation to assess the upper bound of
the detection and evaluate the dynamic range of the biosen-
sor.
D. Saturation and dynamic range
To find the distribution of analyte particles in the case
where the number of capturing probes is finite, we need to
incorporate

7

into

12

and compute the equilibrium con-
centration from

13

. This basically results in finding the
equilibrium state of a nonhomogeneous Markov process,
which is numerically possible but difficult. One alternative
approach is to use the simplifying assumption that the ana-
lyte molecules in the solution are not depleted by the binding
events

i.e., only a very small fraction of analytes is captured
by the probes

. Although this limits the generality of the
solution, it is still applicable to most biosensor platforms. In
this case, for each capturing site in state 1 in our model
described in Sec. II A, we write the following balance equa-
tion at the equilibrium:
n
c
,
S
m
1,
c
=
nm
c
,1
n
1,
Y
n
c
,
S
n
1,
Y
,

32

where
m
c
,1
,
m
1,
c
are the transition probabilities from

6

and
n
c
,
S
is the number of the captured particles when the system
is in saturation. Solving

32

to find
n
c
,
S
, we have
n
c
,
S
=
n

m
c
,1
/
m
1,
c

1+

n
/
n
1,
Y

m
c
,1
/
m
12,
c

.

33

Note that the parameter
m
c
,1
/
m
1,
c
is basically equivalent to
n
c

t

/
n
=
x
c
,
E
, which is the analyte capturing probability with-
out saturation. Hence, we use

33

to find
x
c
,
S
, the capturing
probability with saturation, as a function of
x
c
,
E
and
n
1,
Y
,
x
c
,
S
=
n
c
,
S
n
=

m
c
,1
/
m
1,
c

1+

n
/
n
1,
Y

m
c
,1
/
m
1,
c

=
x
c
,
E
1+

n
/
n
1,
Y

x
c
,
E
.

34

The value of
x
c
,
S
in

34

is basically an entry of
x
S
, the
equilibrium distribution vector of the system in the presence
of saturation. The limits of
x
S
and
x
c
,
S
when
n
0 are
x
E
and
x
c
,
E
, respectively, which are the nonsaturated equilibrium dis-
tribution and capturing probability. The other limit of
x
S
is
zero, obtained by letting
n
. It is important to realize that
the value of
n
1,
Y
in the general case is different for different
states of the system and can be represented by vector
n
Y

R

N
+
P

which has finite entries corresponding to the states
with capturing probes, and infinite otherwise.
The most important impact of saturation on our formu-
lations is the breakdown of the linearity assumption between
the input

analyte concentration

and the biosensor output
signal which enabled the calculation of the SNR and NF in
the previous section. Equation

33

basically describes a non-
linear and monotonic relationship between the biosensor in-
put and output. Because of the monotonicity, we can argue
that the input can always be evaluated by observing the out-
put of the system, but the concern for the validity of SNR
expression is caused by observing the effect of saturation
014909-6 Hassibi, Vikalo, and Hajimiri
J. Appl. Phys.
102
, 014909

2007

Downloaded 25 Jul 2007 to 131.215.225.175. Redistribution subject to AIP license or copyright, see http://jap.aip.org/jap/copyright.jsp

nonlinearity

on the detection uncertainty. In order to evalu-
ate this effect, we use the same approach that we had in
evaluating SNR of a linear system, which is referring the
output signal plus its noise to the input. The difference here
is their nonlinear relationship, which alters the function
T

.
First we derive the output fluctuation of the biosensor
system in the presence of saturation. Equation

22

indicates
that the fluctuation of the number of captured analytes is a
function of the capturing probability and
n
. This is indepen-
dent of the relationship between
n
and
n
c
,
S
. Thus, we can
derive the expected value and variance of the fluctuation of
the captured analytes in saturation as
u
X
,
E

t

=0,
X
2
=
n

x
1,
E
1+

nx
1,
E
/
n
1,
Y


1−
x
1,
E
1+

nx
1,
E
/
n
2,
Y

̄
x

N
+
P

,
E
1+

nx

N
+
P

,
E
/
n

N
+
P

,
Y


1−
x

N
+
P

,
E
1+

nx

N
+
P

,
E
/
n

N
+
P

,
Y


T
,

35

and we can use

35

for the biosensor output fluctuation in

24

.
In saturation, the parameter
T

which we defined as

s

t

/

n
is derived by calculating the derivative of

33

with
respect to
n
for all states where the transduction occurs. Us-
ing the derivative of

34

we have

n
c
,
S
=
x
c
,
E

1+

nx
c
,
E
/
n
1,
Y

2

n
,

36

and therefore, by considering the transduction process,

s

t

=
T

x
1,
E

1+

nx
1,
E
/
n
1,
Y

2
x
2,
E

1+

nx
2,
E
/
n
2,
Y

2
̄
x

N
+
P

,
E

1+

nx

N
+
P

,
E
/
n

N
+
P

,
Y

2

T

n
=
T


n
,

37

which essentially denotes the value of
T

in saturation.
Now that we calculated both
T

and
X
2
in

35

and

37

,
we can utilize them in

27

and

31

to find the SNR and NF
of the biosensor in saturation. Based on

27

, the SNR of a
biosensor system without saturation should increase when
n
increases. However, in saturation, according to the results of

35

and

37

, we observe that
T

becomes proportional to
1/
n
2
when
n
is large, while
X
2
becomes a constant. Hence,
the SNR of the system becomes proportional to 1/
n
2
as
n
becomes large, which suggest an increase in the uncertainty

noise

of the input referred signal when the biosensor satu-
rates. This is an extremely important outcome of our formu-
lations, since it establishes the fact that the biosensor SNR is
not a monotonic function but fundamentally a concave func-
tion where SNR
0 in both limits of the input

i.e.,
n
0
and
n

.
To evaluate the range of detection of biosensors, similar
to all sensing systems, we are required to find the MDL as
well as the HDL. On the lower side of sensing, we define the
MDL or
n
MDL
as the analyte level where the biosensor SNR
is equal to some minimum acceptable value SNR
min
. How-
ever, based on the results of

35


37

, for the same SNR
min
,
we will have another analyte level which is larger than
n
MDL
.
This value is basically HDL and the analyte concentration
level associated with it is
n
HDL
. Consequently, the dynamic
range of the biosensor is defined as
DR =
n
HDL
n
MDL
.

38

The closed form for DR is generally complicated and is de-
rived by solving

27

for SNR
min
. As we will show in Sec.
IV, the value of DR for biosensors can easily be evaluated
numerically for any given value of SNR
min
.
E. Effects of interference, non-specific binding, and
background
The expressions for noise, SNR, and DR in the previous
subsections all assumed that there is only one analyte species

the target

that can bind to the capturing probes. However,
in all practical biosensors, other species coexist with the tar-
get in the sample, and they may or may not bind to the
probes. We call such species interferers, and their capturing
incidents nonspecific bindings. Accordingly, the signal gen-
erated by the transducer which originates from nonspecific
bindings is called nonspecific signal, or in certain biological
detection platforms it is referred to as the background signal.
A nonspecific binding event

e.g., cross-hybridization in
DNA microarrays
1

is generally less probable than the spe-
cific binding when target analyte and the interfering species
have the same abundance in the sample. Nevertheless, when
the ratio of the target analyte to the nonspecific analyte is
small, contributions of nonspecific binding may dominate the
measured signal. To incorporate the effects of nonspecific
binding in the previously discussed formulations, we first
need to assess the uncertainty associated with the concentra-
tion of interferers. This basically defines whether we should
consider nonspecific binding incidents coming from a deter-
ministic or a random concentration.
To examine the general case of detection in the presence
of interference, we assume that
m
interferers coexist with the
target in the sample, and denote their amounts by
n

1

,
n

2

,..., and
n

m

, where for the
i
th specie the expected
number of analytes is
n

i

, with variance of

i

2
. Obviously,
when the interferer concentration is known

deterministic in-
terference

, we will have
n

i

=
n

i

and

i

2
=0. Now in the
general case we can rewrite

25

, the equilibrium master
equation with interference, as
s
b
,
E

t

=
n
+
T

−1
T

i
=1
m
n

i

x
E

i

+
T

u
X
,
E

t

+

i
=1
m
u
X
,
E

i


t


+
u
T

t

,

39

where
x
E

i

and
Tu
X
,
E

i


t

are the equilibrium distribution and
biochemical noise of the
i
th interferer specie, respectively. It
014909-7 Hassibi, Vikalo, and Hajimiri
J. Appl. Phys.
102
, 014909

2007

Downloaded 25 Jul 2007 to 131.215.225.175. Redistribution subject to AIP license or copyright, see http://jap.aip.org/jap/copyright.jsp
is straightforward to show that
X

i

2
, the variance of
u
X
,
E

i


t

,
can be calculated as
X

i

2
=
n

i



i

2
A

i

+
B

i


T
,

40

where
A

i

=

x
1,
E

i


2

x
2,
E

i


2
̄

x

N
+
P

,
E

i


2

,

41

and
B

i

=

x
1,
E

i


1−
x
1,
E

i


x
2,
E

i


1−
x
2,
E

i


̄
x

N
+
P

,
E

i


1−
x

N
+
P

,
E

i


.

42

By employing

40

, we can repeat the same procedure as
described in Sec. III C to find the SNR and NF of the system,
given that the values of all
n

i

and

i

2
are known. In Table
I
,
we have listed the formulation for SNR and NF with inter-
ference described by

38

.
To find the DR of the system with interference, we need
to find
n
min
and
n
max
in the presence of interference. Finding
n
min
requires solving the modified version of

29

, as shown
in Table
I
. To derive
n
max
, we should do the same thing but
solve the balance equation

32

in the presence of interferers,
which can be very complicated. It is important to realize that
in most biosensors the concentration levels of analytes ex-
posed to capturing probes are kept low enough to ensure that
the background and nonspecific binding do not drive the sys-
tem into saturation. As a result, the upper limit of detection
n
max
can still be dominated by the captured target analyte
rather than the interferers. Thus, to find
n
max
in biosensors we
generally do not need to take into account the interferer ana-
lytes, which makes the calculation less complicated.
IV. MODEL IMPLEMENTATION
To implement our model, in this section, we look at two
different biosensor platforms. The systems under analysis
have different transducers

impedimetric and fluorescence
based

; however, we will demonstrate that for both we can
use the described modeling method to evaluate and compare
their limits of detection.
A. Impedimetric biosensor
In impedimetric biosensors the detection relies on the
impedance of an electrode-electrolyte interface which can
monotonically increase

or decrease

by analyte bindings. In
some arrangement, the intrinsic characteristics of the ana-
lytes

e.g., charge

are used to the capacitance

and accord-
ingly the impedance

of the electrode-electrolyte
interface,
13
,
19
22
while in other arrangements, electroactive
labels are used
23
25
to enhance the impedance changes.
In Fig.
5
, we have illustrated an example of a microfab-
ricated impedimetric biosensor where an interdigitated
electrode-electrolyte structure is used as the transducer. The
capturing probes are immobilized between the fingers of the
interdigitated structure and binding of analytes changes the
conductivity of the solution near the surface by changing the
mobility of the free ions. This subsequently modifies the im-
pedance of the electrode-electrolyte transducer system

i.e.,
ports 1 and 2 in Fig.
5

, making the impedance an indication
of the target analyte concentration.
In Fig.
6
, we have shown the detection circuitry of this
impedimetric biosensor which consist of an excitation volt-
FIG. 5. Impedimetric biosensor where binding of analytes alters the imped-
ance
Z

or admittance
Y

between the fingers of an interdigitated electrode
structure. The change in overall impedance corresponds to the amount of
captured analytes.
TABLE I. Signal-to-noise ratio

SNR

and noise figure

NF

of biosensor systems when

a

only the target
analyte is in the system,

b

the target analyte and the
m
interfering species coexist in the system, and

c

when
the system in addition has
K
noisy amplification stages.
Signal-to-noise ratio

SNR

Noise figure

NF

a
n
2
T

2
T

2

X
2
+
T
2
1+
T
2
T

2

X
2
b
n
2
T

2
T

2

X
2
+
T

2


i
=1
m
X

i

2
+
T
2
1+
T
2
T

2

X
2
+

i
=1
m
X

i

2
X
2
c
n
2
T

2
T

2

X
2
+
T

2


i
=1
m
X

i

2
+
T
2
+

i
=1
K

j
=1
i
G
j
−1

i
2
1+
T
2
T

2

X
2
+

i
=1
m
X

i

2
X
2
+

i
=1
K

j
=1
i
G
j
−1

i
2
T

2

X
2
FIG. 6. Detection circuitry for the impedimetric biosensor.
014909-8 Hassibi, Vikalo, and Hajimiri
J. Appl. Phys.
102
, 014909

2007

Downloaded 25 Jul 2007 to 131.215.225.175. Redistribution subject to AIP license or copyright, see http://jap.aip.org/jap/copyright.jsp
age source at 1 MHz, transimpedance amplifier

TIA

, band-
pass filter, and an envelope detector. Any change in imped-
ance of the transducer changes the current going through
structure and subsequently the output of the amplifier which
is finally detected by the ideal envelope detector. In Table
II
,
we list the characteristics of this biosensor which is required
to calculate the limits of performance of this system. The
capturing probabilities are a function of target-probe interac-
tion and also the capturing area. These values might seem
quite small; however, they are a good approximate for typi-
cal biosensor systems where the surface to volume ratio is
small. In Table
II
, we have also listed the transduction ma-
trix, probe density, and the transducer

TIA

additive noise. It
is important to realize that the entries of the transduction
matrix basically correspond to the different capturing areas
of the transducer structure which have different numbers of
capturing probes and capturing probability.
In Fig.
7
, based on our modeling method we have illus-
trated the performance of this impedimetric biosensor. In
Fig.
7

a

, we show the output of the biosensor as a function
target analyte concentration. As expected the output signal of
this system initially increases linearly with the analyte con-
centration level but it plateaus eventually due to the satura-
tion. The probabilistic model not only predicts this behavior
but also quantifies the background signal of the system
which originates from the nonspecific bindings. In Fig.
7

b

,
we show the SNR of this biosensor based on

26


29

.
Without saturation we expect the SNR to increase monotoni-
cally, our formulations quantify how SNR decreases as the
analyte concentration level increases. One important obser-
vation here is the fact that SNR is more sensitive to interfer-
ence at low concentration levels, an important empirical ob-
servation which has been previously reported in different
biosensor systems.
26
The HDL of this system is limited by
saturation and finite number of capturing probe density.
However, our results show that the MDL is interference de-
pendent. The overall DR of this system for SNR
min
=30 dB
varies between 84 and 104 dB depending on the interference
concentration. The noise of the transducer and the detection
circuitry has little effect on the overall performance of the
system which indicates that the biochemical noise and probe
saturation are performance-limiting factors in this impedi-
metric biosensor.
B. Fluorescent-based biosensor
In this example we implement our model for a
fluorescent-based biosensor system. We have chosen this par-
ticular biosensor system for two main reasons. One is the
generality of this system which is currently a widely used
detection method in different assays. The second reason is to
illustrate the versatility of our modeling method which can
unify the performance metrics of different biosensors
The example biosensor platform is illustrated in Fig.
8
.
In this system, the capturing probes are immobilized at the
bottom of the microtiter plate and the analytes are all fluo-
rescent labeled. After an incubation phase, the solution con-
taining analytes is removed from the well and the captured
analytes

specific and nonspecific

are quantified using a
fluorescence intensity detector. The detector in this example
consists of an excitation light source which uniformly excites
the captured fluorescent labels, an emission filter which
blocks the excitation source while passing the emitted light
of the labels, and a photodetector. The photodetector consists
of a photodiode and a charge-integrating TIA

CTIA

.In
Table
III
we have listed the specification of this system in-
cluding the transduction gain and signal-dependent noise
FIG. 7. Performance result of the impedimetric biosensor. In

a

the biosen-
sor output signal is plotted vs the target analyte concentration. In

b

the
SNR of the system is plotted and the DR is calculated for the SNR
min
of
30 dB.
TABLE II. Characteristics of the impedimetric biosensor.
Capturing probability
x
E
=

0.3 0.45 0.6 0.6 0.45 0.3


10
−5
Probe density
n
Y
=

257752


10
5
Reaction chamber volume
V
=10
l
Admittance transfer matrix

Y
=

2 1.5 1.2 1.2 1.5 2


10
−7

−1
/
target
Transduction matrix
T
=

2 1.5 1.2 1.2 1.5 2


V/target
Transducer noise
T
=

kT

2

10
4

BW

180 nV
Interferer capturing probability
x
E

1

=

0.3 0.45 0.6 0.6 0.45 0.3


10
−8
Interferer variation

1

2
=0
014909-9 Hassibi, Vikalo, and Hajimiri
J. Appl. Phys.
102
, 014909

2007

Downloaded 25 Jul 2007 to 131.215.225.175. Redistribution subject to AIP license or copyright, see http://jap.aip.org/jap/copyright.jsp
variance of the transducer. It is important to realize that in
this system we have intentionally selected the assay specifi-
cations similar to the impedimetric biosensor example, al-
though the transduction mechanism and detector are totally
different.
In Fig.
9
, we have the performance results of this bio-
sensor. The output of the system as a function of analyte
concentration

see Fig.
9

a


is similar to the impedimetric
biosensor; however, the voltage level is lower because of the
lower transduction gain. Nevertheless, the different output
voltage level does not specify the performance of this sys-
tem. In Fig.
9

b

, we plot the SNR of this system as a func-
tion of analyte concentration. The graph shows that the HDL
of the system, similar to the previous example, is limited by
the saturation. The MDL, on the other hand, is almost con-
stant and it seems to be independent of interference. By look-
ing carefully at Fig.
9

b

, we can see that the SNR at low
analyte concentration deviates very much from the QL-SNR,
which is an indication that the additive noise of the trans-
ducer is dominating the SNR at low concentration levels. To
explore this hypothesis, in Fig.
9

c

, we have plotted the
simulation results of the same assay but with a high-
performance photodetector transducer system which has an
integration time of 100 s

ten times more transduction gain

and 100 times less dark current. As illustrated, on the HDL
side of SNR for the modified system is not improved, yet the
MDL is lowered drastically, making the low concentration
SNR interference limited and closer to QL-SNR. The overall
improvement in DR with the modified transducer is as high
as 40 dB.
By examining Figs.
7
and
9
we can compare the perfor-
mance of both. The fluorescent-based biosensor with the ini-
FIG. 8. Fluorescence-based biosensor where binding of labeled analytes is
detected after the incubation phase. The emitted light from the fluorescent
labels is detected by the photodiode connected to a charge integrating tran-
simpedance amplifier

CTIA

.
TABLE III. Characteristics of the fluorescence-based biosensor.
Capturing probability
x
E
=2.7

10
−5
Probe density
n
Y
=3.2

10
6
Reaction chamber volume
V
=10
l
Photon flux transfer function

I
flux
=100 photons/target s
Transduction matrix
T
=0.02

V/target
Transducer noise
T
=10
V+2

10
−8

targets
Interferer capturing probability
x
E

1

=2.7

10
−8
Interferer variation

1

=0
FIG. 9. Simulation result of fluorescence-based biosensor. In

a

the biosen-
sor output signal is plotted vs the target analyte concentration. In

b

the
SNR of the system is plotted and the DR is shown for the SNR
min
of 30 dB.
In

c

the SNR of the system is shown where the integration time

i.e.,
transduction gain

is increased by one order of magnitude while the photo-
detector dark-current noise power is decreased ten times.
014909-10 Hassibi, Vikalo, and Hajimiri
J. Appl. Phys.
102
, 014909

2007

Downloaded 25 Jul 2007 to 131.215.225.175. Redistribution subject to AIP license or copyright, see http://jap.aip.org/jap/copyright.jsp