A Comparative Study of the Time-Series Data for Inference of

Gene Regulatory Networks Using B-Spline

Haixin Wang, James E. Glover and Lijun Qian

Abstract In this paper, the quantitative analysis of timeseries gene expression data on inference of gene regulatory
networks is performed. Time-series gene data are modeled by
the B-Spline algorithm to improve the overall smooth expression
curves which can further reduce over-fitting. The effect of the
different sizes of observed time-series data on gene regulatory
networks inference is analyzed. The stochastic errors introduced
by the B-Spline algorithm to the system are evaluated. The
precision of different sizes of time-series data on parameter
estimations is compared. With application of the B-Spline to
generate continuous curves, simulation results can be much
more accurate and inference results are significantly improved.
Both synthetic data and experimental data from microarray
measurements are used to demonstrate the effectiveness of the
proposed method.

I. I NTRODUCTION
ENETIC regulatory networks (GRNs) are collections
of DNA segments in a cell which interact with each
other and with other substances in the cell, thereby governing
gene transcriptions. In light of the recent development of
high-throughput DNA microarray technology, it becomes
possible to discover GRNs, which are complex and nonlinear
in nature. Specifically, the increasing existence of microarray
time-series data makes possible the characterization of dynamic nonlinear regulatory interactions among genes. The
modeling, analysis and control of GRNs are critical for
finding medicine for gene-related diseases.
One issue during the inference of GRNs is that not enough
time-series experiment data are available [1]. This makes
many inference processes not realizable. Many models are
restricted by limited data. B-Spline can reconstruct unobserved gene expression data or recover the missing data.
Compared to the B-Spline algorithm, simple algorithms, for
example linear interpolation, can lead to poor estimation for
the missing time-series data, especially when the sampled
data is not uniform [2].
In this paper, the ordinary differential equation (ODE)
model using S-system is adopted to infer GRNs [3]. Different sizes of time-series data are analyzed. The B-Spline
algorithm is introduced to analyze the raw data. The effects
of the time-series data to infer gene regulatory networks are
analyzed in synthetic model and microarray experimental
expression data.

Haixin Wang and James E. Glover are with the Department of Mathematics and Computer Science, Fort Valley State University, Fort Valley,
Georgia, USA (email: [email protected]).
Lijun Qian is with the Department of Electrical and Computer Engineering, Prairie View A&M University, Prairie View, Texas, USA (email:
[email protected]).

II. I NTRODUCTION

TO

B-S PLINE

ALGORITHM

In numerical analysis, B-Spine has been widely used

in image processing, computer-aided design and statistical
analysis [4], [5]. It is one of the most flexible formulations for
many engineering applications. A spline curve is a sequence
of curve segments that are connected together to form a
single continuous curve. In this study, it is a time-related
sequence [6]. The basis B-Splines of degree m can be
expressed as
S(t) =

N m1
X

Pi bi,k (t), t [dm , dkm ]

i=0

where k is the order of the polynomial basis, Pi are called

control points, and the normalized B-Spline basis bi,k (t) can
be expressed using the Cox-deBoor recursion formula [6]:

1, di t di+1
bi,1 (t) =
0, otherwise
bi,k (t) =

(t di )bi,k1 (t) (di+k t)bi+1,k1 (t)

+
.
di+k1 di
di+k di+1

Given N + 1 time series datasets with knots (di ) that

satisfying
d0 d1 dN ,
the corresponding control points (Pi ) need to be calculated.
Suppose t [dm , dkm ] and P0i = Pi for i = m, , k m
h1
Phi = (1 h,i )Ph1
i1 + h,i Pi

where h = 1, , k; i = 0, , N m 1;
h,i =

t Pi
.
pi+k+1h Pi

A polygon can be constructed by connecting the control

points with line segments. From those control points, the
corresponding data points can be calculated.
In this study, we select the uniform cubic B-Spline to insert
time-series data, and k is chosen to be 4. Then the basis can
be expressed as

1 3 3 1
pi1

 1 3 6 3 0 pi

Si (t) = u3 u2 u 1
3 0 pi+1
6 3 0
1
4
1 0
pi+2
(1)
where u [0, 1].
The general process of inserting time series data via Bspline algorithm is shown in Algorithm 1.

Algorithm 1 General time-series data processing by B-Spline

Require:
The raw data: d0 dN
The insertion data length between two insertion data
points: (di , di+1 )
Spline order: k = 4
Ensure: The data set Si
1) Calculate the control points Pi from the original data
di 2) Determine the insertion data Si according to equation (1) with initial conditions: spline order and insertion
data length

Microarray
Experiments

Apply B-Spline
Algorithm to
Gene Expression
Data

Genetic
Programming

III. T HE ADOPTED ODE S

Kalman Filter

MODEL TO INFER GENE

Yes

REGULATORY NETWORKS

In general, modeling gene regulatory networks is a nonlinear identification problem. Assuming there are N genes
of interest and xi denotes the state (such as the microarray
reading) of the ith gene, then the dynamics of the GRN may
be modeled as
dxi
= fi (x1 , x2 , , xN )
i = 1, 2, , N
(2)
dt
where the nonlinear functions fi need to be determined
from time-series microarray measurements. In this study, we
assume that the functions (fi , i) are in the form [7]:
fi =

Li
X

[wij ij (x1 , x2 , , xN )]

i = 1, 2, , N (3)

j=1

where Li is the number of terms in fi , wij are the parameters to be estimated and ij (x1 , x2 , , xN ) is the j th
component of the nonlinear function fi . A two-step nested
optimization procedure is proposed to identify the nonlinear differential equation for each individual gene. Genetic
programming (GP) is applied to determine the nonlinear
parameters (global optimization) and then the corresponding parameters associated with each term are estimated by
Kalman filtering (local optimization) in each iteration. Such
a decomposition of the problem into a structural part solved
by GP and a parameter optimization part solved by Kalman
filtering reduces the complexity significantly and speeds
up convergence. In this paper the S-system is adopted for
function fi . The S-system model is given by [3]:
N
N
Y
Y
dxi
gi,j
h
= i
xj i
xj i,j , (i = 1, ..., N )
dt
j=1
j=1

Data Sturcture

Model Parameters

Entire Model

No
Evaluation by Fitness Function?

Yes

Find the model

Fig. 1. The flowchart for GRN identification using GP, Kalman Filtering
and B-Spline

In addition, the S-system model has been proven to be

successful in modeling GRNs [9], [10], [11], [12], [13], [14].
Hence, the S-system model is adopted for modeling GRNs
in this paper. More importantly, the B-Spline algorithm is
introduced into the process to improve the quality of the
time-series data sets. The general process of inferring the
gene regulatory networks is shown in Fig. 1.
IV. T HE

ANALYSIS OF TIME - SERIES DATA USING

B-S PLINE

In this section, the effects of gene expression data using

the B-Spline algorithm are discussed in the optimization
estimation process.

(4)

where xi is the state variable, i and i are the positive

rate constants, gi,j and hi,j are the exponential parameters
called kinetic orders. If gi,j > 0, gene j will induce the
expression of gene i. On the contrary, gene j will inhibit
the expression of gene i if gi,j < 0. hi,j will have the
opposite effects on controlling gene expressions compared to
gi,j . S-system is a quantitative model which is characterized
by power-law functions. It has the rich structure capability of
capturing various dynamics in many biochemical systems [8].

A. The effect of the sampling data size to infer GRNs

For microarray experiments, the time-series data are important to infer gene regulatory networks. The sampling data
sizes from the experiment have a significant impact on the
resulted GRNs. For insufficient data sizes, the inferred GRNs
can have errors. Hence, the effects of the various data sizes
on the inference of GRNs are analyzed in this section. In
order to show the relation, the coefficient of determination
(CoD) is applied [15]. It provides a measure of how well
future outcomes are likely to be predicted by the model. The

0.9

0.9

BSplline
BSpline

BSpline

0.8

0.8

Linear

BSpline

Linear

Linear

0.7

0.7
MSE Error

Average CoD

data4
0.6

0.5

0.6
0.5
0.4

0.4

0.3
0.3

0.2
0.2

0.1
0.1

0
0

0
0

10

12

14

16

18

10

15

20

25

Sampling Size of the BSpline Algorithm

20

The Number of Sampling Data

Fig. 2.

The effect of the sampling data size on the estimator.

CoD is given by
SSerr

N
X

(yi fi )2

N
X

(yi y)2

i=0

SStot

Fig. 3.

The data error from the B-Spline algorithm.

that the number of samples in the time-series data from

microarray experiment is usually very small. Therefore, it
is necessary to introduce the B-spline into the process of
inference of GRNs. It is also noticed that B-Spline can reach
a better result than that of a linear interpolation algorithm
before approaching the upper bound.

i=0

= 1

SSerr
SStot

where SSerr is the sum of the squared errors between

observed data and data from the inferred model. SStot is
the sum of the squared errors between observed data and its
mean. R2 is the CoD that indicates how well the regressing
line fits the data. An R2 = 1.0 indicates that the regression
line perfectly fits the data. yi is the observed data, fi is the
modelled data and y is the mean of the observed data. The
process is shown by Algorithm 2.
Algorithm 2 The general process of analyzing the effect of
data size to infer GRNs
Require:
The observed data xi with size n
Ensure: CoD of each GRNs
1) Apply two-step method to infer GRNs
2) Analyze CoD for each GRNs
As an example, the result is shown in Fig. 2 for the
following synthetic model:
x1
x2

1.2
0.2
= x1.5
1 x2 x2
0.5
= x0.1
1 x1 x2

From Fig. 2, the averaged CoD from the estimator becomes larger with the increased amount of sampling data.
The CoD does not change much after certain number of
samples, which implies that the CoD has an upper bound.
It is observed that increasing the number of sampling points
yields more exact estimates, within the upper bound. Note

B. The data error introduced by the B-Spline

B-Spline algorithm is introduced to interpolate more data
from the original sampling data with some missing data. It is
necessary to analyze whether or not the B-Spline algorithm
introducesP
cumulative data errors. The mean square errors
N
(M SE = i=0 (yi fi )2 ) for different sizes of the data in BSpline algorithm are stochastically calculated and compared
to the observed data. The result is shown in Fig.3. The
synthetic data range is [0.1, 1]. The MSE range by BSpline algorithm is [0.02, 0.25]. The MSE range by linear
interpolation is [0.029, 0.81]. From Fig. 3, it is observed that
the MSE remains very small and it does not increase with
interpolated data by B-Spline algorithm. Therefore, B-Spline
algorithm introduce very limited data errors, which have little
effect on the inference result of GRNs.
C. The effect of data size on parameter optimization estimation
After applying B-Spline to interpolate data, it is necessary
to analyze the parameter estimation under different interpolation sizes. The general process is shown by Algorithm 3.
The effect of the total data size is analyzed using the
same synthetic model. The data size is the total data size
after applying the B-Spline algorithm with different sampling
sizes. The stochastic CoD is analyzed with different intervals
from two adjacent control points. The result is shown by
Fig.4. It clearly shows that the CoD is larger with larger
sampling size. However, both interpolation algorithms are
upper bounded. Fig. 4 shows that B-Spline can get much
better results in comparison to linear interpolation.

2500

HAP1
CYB2
CYC7
CYT1
COX5A
x1 Bspline
x2 Bspline
x3 Bspline
x4 Bspline
x5 Bspline

2000

Concentration Level

Algorithm 3 The general process to analyze the effect of

data size on parameter optimization
Require:
The observed data xi
Ensure: Stochastical errors of the estimated parameters
1) Choose the sampling size N for B-Spline algorithm
2) Apply B-Spline algorithm with the size N sampling
points
3) Apply Kalman Filter to calculate the parameters
4) Calculate the stochastical errors of the estimated parameters

1500

1000

500

50

100

150

200

250

300

350

Time
BSpline

0.95
CoD of Estimated Parameters

BSpline

Fig. 5.

Linear

The dynamics of expression level of the 5 yeast genes .

Linear

0.9

0.85

0.8

0.75

0.7
0

20

40

60

80

100

120

140

160

180

200

Data Size

Fig. 4.

The effect of interpolation data size on the parameter estimator.

V. Y EAST DATA S IMULATION R ESULTS

During this part of the simulation, we consider timeseries gene-expression data corresponding to yeast protein
synthesis. Five genes (HAP1(x1 ), CYB2(x2 ), CYC7(x3 ),
CYT1(x4 ), COX5A(x5 )) are selected because the relations
among them have been revealed by biological experiments [16]. After the B-Spline algorithm is applied to the
experimental data, the trace of the time-series microarray
measurement data from [17] is shown by Fig. 5. Because the
processes using B-splines algorithm generates enough data
from the original experiment data, inference results for each
iteration are stable. Therefore, unlike [3], the complicated
processes using Z-scores are not necessary to process the
simulation results, and B-splines is useful for the experiment
data processing which improves the inference with less
complicated process and more stable results.
The relationships among the 5 genes are shown by the
branch pathway model given in Fig. 6. We observe that
HAP1 represses CYC7, and CYB2 activates CYC7. It is
also observed that HAP1 activates COX5A and CYT1. These
observations are in agreement with the biological experiment
findings in [16], [18].
VI. C ONCLUSIONS
In this paper, the effects of the B-Spline algorithm on the
gene time-series data and inference process are analyzed.

Fig. 6.

The branch pathway model of the 5 genes in yeast.

Both synthetic and yeast microarray data and models are

analyzed. The introduction of B-Spline algorithm yields
better results in both cases. In further studies, B-Spline
algorithm will be applied to analyze the effects of noisy data
on genetic programming during the inference process.
R EFERENCES
[1] T. Chen, V. Filkov, and S. S. Skiena, Identifying gene regulatory
networks from experimental data, pp. 94103, 1999.
[2] J. Aach and G. M. Church, Aligning gene expression time series with
time warping algorithms, BIOINF: Bioinformatics, vol. 17, pp. 495
508, 2001.
[3] H. Wang, L. Qian, and E. Dougherty, Inference of gene regulatory
networks using s-system: A unified approach, IEEE 2007 Symposium
on Computational Intelligence in Bioinformatics and Computational
Biology, 2007.
[4] S. Samadi, M. Ahmad, and M. Swamy, Characterization of b-spline
digital filters, IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS,
vol. 51, no. 4, pp. 808816, 2004.
[5] G. S. Knmar, P. K. Kalra, and S. G. Dhande, Parameter optimization
for b-spline curve fitting using genetic algorithms, IEEE Evolutionary
Computation, vol. 3, pp. 18711878, 2003.
[6] H. Prautzsch, W. Boehm, and M. Paluszny, Bezier and B-Spline
Techniques. Springer, 2002.

[7] L. Qian, H. Wang, and E. Dougherty, Inference of noisy nonlinear

differential equation models for gene regulatory networks using genetic
programming and kalman filtering, IEEE Transactions on Signal
Processing, vol. 56, no. 8, 2008.
[8] E. Voit and T. Radivoyevitch, Bioinformatical systems analysis of
genome-wide expression data, Bioinformatics, vol. 16, pp. 1023
1037, 2000.
[9] C. Spieth, F. Streichertr, N. Spee, and Zell., A memetic inference
method for gene regulatory networks based on s-systems, Proceedings
of the IEEE Congress on Evolutionary Computation, pp. 152157,
2004.
[10] N. Norman and H. Iba, Inference of gene regulatory networks using
s-system and differential evolution, Proceedings of the Genetic and
Evolutionary Computation Conference, pp. 439446, 2005.
[11] S. Kimura, K. Ide, A. Kashihara, M. Kano, M. Hatakeyama, R. Masu,
N. Nakagawa, S. Yokoyama, S. Kuramitsu, and A. Konagaya, Inference of s-system models of genetic networks using a cooperative
coevolutionary algorithm, Bioinformatics, vol. 21(7), pp. 11541163,
2005.

[12] S. Kikuchi, D. Tominaga, M. Arita, K. Takahashi, and M. Tomita,

Dynamic modeling of genetic networks using genetic algorithm and
s-system, Bioinformatics, vol. 19(5), pp. 643650, 2003.
[13] S. Kimura, K. Ide, A. Kashihara, M. Kano, M. Hatakeyama, R. Masui,
N. Nakagawa, S. Yokoyama, S. Kuramitsu, and A. Konagaya, Inference of s-system models of genetic networks from noisy time-series
data, Chem-Bio Informatics Journal, vol. 4(1), pp. 114, 2004.
[14] J. Almeida and E. Voit, Neural-network-based parameter estimation in
s-system models of biological networks, Genome Informatics, vol. 14,
pp. 114123, 2003.
[15] W. Mendenhall and T. Sincich, A Second Course in Statistics: Regression Analysis. Prentice Hall, 2003.
[16] P. Woolf and Y. Wang, A fuzzy logic approach to analyzing gene
expression data, Physiol. Genomics, vol. 3, pp. 915, 2000.
[17] http://www.wanghaixin.com/ss/fulldata.htm,
[18] J. Schneider and L. Guarente, Regulation of the yeast cyti gene encoding cytochrome cl by hap1 and hap2/3/4, Molecular and Cellular
Biology, vol. 11(10), pp. 49344942, 1991.