Software for Automated Analysis
of DNA Fingerprinting Gels
Daniel R. Fuhrmann,1Martin I. Krzywinski,2Readman Chiu,2Parvaneh Saeedi,2
Jacqueline E. Schein,2Ian E. Bosdet,2Asif Chinwalla,3LaDeana W. Hillier,3
Robert H. Waterston,3John D. McPherson,3Steven J.M. Jones,2and
Marco A. Marra2,4
1Department of Electrical Engineering, Washington University, St. Louis, Missouri 63130, USA;2Genome Sciences Centre,
British Columbia Cancer Agency, Vancouver, British Columbia V5Z 4E6, Canada;3Genome Sequencing Center, Washington
University School of Medicine, St. Louis, Missouri 63108, USA
Here we describe software tools for the automated detection of DNA restriction fragments resolved on agarose
fingerprinting gels. We present a mathematical model for the location and shape of the restriction fragments as
a function of fragment size, with model parameters determined empirically from “marker” lanes containing
molecular size standards. Automated identification of restriction fragments involves several steps, including:
image preprocessing, to put the data in a form consistent with a linear model; marker lane analysis, for
determination of the model parameters; and data lane analysis, a procedure for detecting restriction fragment
multiplets while simultaneously determining the amplitude curve that describes restriction fragment amplitude
as a function of mobility. In validation experiments conducted on fingerprinted and sequenced Bacterial
Artificial Chromosome (BAC) clones, sensitivity and specificity of restriction fragment identification exceeded
96% on restriction fragments ranging in size from 600 base pairs (bp) to 30,000 bp. The integrated suite of
software tools, written in MATLAB and collectively called BandLeader, is in use at the BC Cancer Agency
Genome Sciences Centre (GSC) and the Washington University Genome Sequencing Center, and has been
provided to the Wellcome Trust Sanger Institute and the Whitehead Institute. Employed in a production mode
at the GSC, BandLeader has been used to perform automated restriction fragment identification for more than
850,000 BAC clones for mouse, rat, bovine, and poplar fingerprint mapping projects.
Maps constructed from fingerprinted large-insert bacterial
clones (Marra et al. 1997) have been constructed to support
whole-genome and localized DNA sequencing activities, as
well as gene cloning studies, in plants (Marra et al. 1999;
Mozo et al. 1999; Tao et al. 2001; Chen et al. 2002), animals
(McPherson et al. 2001; Gregory et al. 2002), the nematodes
Caenorhabditis elegans (Coulson et al. 1995; The C. elegans
Genome Sequencing Consortium 1998) and Caenorhabditis
briggsae (J. Schein and M. Marra, unpubl.), insects (Hoskins et
al. 2000), fungi (Olson et al. 1986; Schein et al. 2002), and
bacteria (Wechter et al. 2002; J. Schein, I. Bosdet, and M.
Marra, unpubl.). Starting with a sufficiently redundant large-
insert library of genomic DNA, the fingerprinting process
(Schein et al. 2003) involves purification of DNA from clones,
treatment of the DNA with restriction enzymes to produce
restriction fragments, resolution of the restriction fragments
on agarose gels, identification of the restriction fragments to
produce a fingerprint, comparison of the fingerprints to each
other to generate contigs (clusters of overlapping clones rep-
resenting the genomic regions from which the clones were
derived), and finally verification of clone ordering within ev-
ery contig. Even using clones containing very large inserts
(i.e., bacterial artificial chromosome [BAC] clones; Shizuya et
al. 1992), hundreds of thousands of fingerprints may be re-
quired to accurately represent a large (i.e., mammalian-sized)
genome in large contigs. The scale of such efforts and the
need to produce data in a rapid, efficient, and cost-effective
fashion have provided impetus for the automation of various
steps in the fingerprinting procedure. Here we describe auto-
mation of the step involving identification of restriction frag-
ments (“bandcalling”), which is performed on digital images
of agarose fingerprinting gels.
Sulston et al. (1988, 1989) were among the first to de-
velop methods for automating bandcalling. They developed
software for lane tracking and band detection that was the
precursor to the IMAGE package, which is available from the
Sanger Institute (http://www.sanger.ac/Software/Image). Al-
though IMAGE is user-friendly and has impressive function-
ality in terms of image manipulation and display, in our ex-
perience the bandcalls it produces require significant manual
verification. Presumably one reason for this is that IMAGE
was designed for analysis of fingerprints generated on acryl-
amide gels from end-labeled DNA fragments (Coulson et al.
1986; Gregory et al. 1997). These end-labeled fragments rep-
resent typically only a portion of the DNA contained within
the clone. This is in contrast to the agarose method employed
currently by ourselves (Marra et al. 1997; Schein et al. 2003)
and others, in which all of the restriction fragments derived
from a clone are visualized by postelectrophoretic staining of
agarose gels with SYBR green (Molecular Probes). Indeed, this
methodological difference has made possible our bandcalling
E-MAIL firstname.lastname@example.org; FAX (604) 877-6085.
Article and publication are at http://www.genome.org/cgi/doi/10.1101/
13:940–953 ©2003 by Cold Spring Harbor Laboratory Press ISSN 1088-9051/03 $5.00; www.genome.org
approach, which aims to identify all of the restriction frag-
ments and, in the case of comigrating restriction fragments,
their copy number (or “multiplicity”).
Our software tools, collectively called BandLeader, con-
sist of a set of MATLAB routines that are capable of automati-
cally identifying and locating marker lanes and data lanes and
the restriction fragments contained therein. Gel images, col-
lected during the fingerprinting procedure, are subjected first
to “lane-tracking”, a semi-automated procedure that identi-
fies the location of the lanes on the digital gel image. This step
is performed using the excellent image manipulation tools
that are part of the IMAGE package. IMAGE-extracted gel
lanes are then passed automatically to BandLeader for band-
calling. Currently the BandLeader software is completely de-
pendent upon the gel and data format presently in use at the
British Columbia Cancer Agency (BCCA) Genome Sciences
Centre. Detailed protocols for the production of fingerprints
suitable for analysis by BandLeader have been described
(Schein et al. 2003). Briefly, each gel consists of 25 marker
lanes and 96 “data lanes” containing large-insert clone fin-
gerprints. DNA quantities and electrophoresis conditions are
strictly controlled to ensure gel-to-gel uniformity of the data.
Each marker lane contains 37 fragments, with data lanes
(BACs digested with HindIII or another suitable restriction
enzyme) containing 50 or more bands. Hence, each gel con-
tains more than 6000 fragments that must be identified. The
BandLeader suite accomplishes this task for each gel in ap-
proximately 10 min on a computer with 1 gigabyte of RAM
and an Intel-based processor running at 1 GHz. Using heuris-
tic data checking, the BandLeader routines flag potential ar-
tifact lanes and exclude them from the data set. These can be
viewed subsequently if desired.
BandLeader has been under development at both Wash-
ington University (St. Louis) and the Genome Sciences Centre
(Vancouver) for more than four years, and during this time
several versions were produced and used. The earliest ver-
sions, developed during the height of activity on the Human
Genome Project, were based on full two-dimensional image
processing techniques and were too slow to be of practical
use. The basic methodology as presented here was in place by
June 2000 (Version 2.0). At that time, the software could be
used as a first step, but manual postprocessing was required to
correct for certain artifacts, most notably overcalls, which re-
sulted from a mismatch between a narrowly defined data
model and the actual image data. Most of the development
effort of the past two years was carried out while the mouse
mapping effort at the Genome Sciences Centre was underway,
and concentrated on making the software more reliable and
the results more robust with respect to variations in the data
away from the nominal model. A full set of exception-
handling routines, to flag data that the software tools were
unable to interpret, were also implemented. The most recent
version (Version 2.3.3) is described here.
An electronic image of a typical agarose fingerprinting gel is
shown in Figure 1A. Figure 1B shows a marker lane, with each
enumerated DNA fragment annotated with the correspond-
ing size of the fragment. Gel (TIFF) images are collected on a
Molecular Dynamics Fluorimager. Although the Fluorimager
is capable of different settings, BandLeader has been tuned to
use 200-micron-square pixels. Each gel image is 1000 ? 1200
pixels. The gel image is partitioned into 121 single-lane im-
ages, each 1000 pixels long ? 9 pixels wide, as a result of the
lane tracking process in IMAGE, but is otherwise unprocessed
prior to our analysis. The lanes on the gel image in Figure 1A
are typical of the input provided to BandLeader, and illustrate
the problem BandLeader has been designed to address;
namely, the automated identification of all DNA fragments in
both the marker lanes and the data lanes. Below we describe
the considerations and approaches that we devised to auto-
matically identify fragments on gels of this type, and quantify
BandLeader’s performance on fingerprints corresponding to a
test set of fully sequenced and finished BAC clones.
Forward Synthesis Model
The methodology we adopt for the analysis of fingerprinting
gels is consistent in many respects with the general model-
based image analysis paradigm of O’Sullivan, Blahut, and
Snyder (1998). It is based on a forward synthesis model that
captures as much relevant information about the desired
quantities (the molecular fragment sizes) as possible, while
maintaining a level of simplicity that allows for computa-
tional efficiency. While the model is based in part on the
underlying physics of the fingerprinting process, the quanti-
tative aspects are derived or refined from the data themselves.
In brief, our model for the production of electrophoretic
gel data is as follows. Under the influence of an electric field,
molecular fragments of a certain size fkmigrate to a particular
location on a fingerprinting gel and form a diffuse band. The
relationship between the distance traveled, or mobility, and
the fragment size is given by a curve which is known approxi-
mately but which must be refined empirically. A typical size/
mobility curve is shown in Figure 2.
In our model, the shape of the band is a rectangle sub-
jected to Gaussian diffusion, and the diffusion parameters are
size-dependent. The brightness, or amplitude of each band is
also size-dependent. Qualitatively speaking, bands due to
smaller fragments travel further, are dimmer, and are more
diffuse, as is evident in Figure 1. In an idealized linear model,
one data lane contains the superposition of bands consistent
with this model at locations determined by the fragment
sizes. This is described by the model equation
AkB?x,y − mk,fk?
where x and y are the horizontal and vertical spatial coordi-
nates, respectively, I is the acquired image intensity, Akis the
amplitude of the kthband, mkis the mobility of the kthband,
and B is a band shape function for the kth band. This band
shape function is separable, that is,
B?x,y,fk? = Bh?x,fk? Bv?y,fk?
where Bhand Bvare the horizontal and vertical factors.
Our model also accounts for the various deleterious ef-
fects that cause the image data to depart from the idealized
linear model. These include: (1) a pointwise nonlinearity of
fluorescent signal intensity, deliberately introduced by the
Molecular Dynamics Fluorimager to compress the visual dy-
namic range, (2) an additive background function which ap-
pears data-dependent in an unknown way but which is
smoothly-varying, (3) impulsive noise due to dust specks and
other gel impurities, and (4) saturation in regions of high
signal intensity. Background instrumentation noise is not in-
cluded in the model, as the signal-to-noise ratio (SNR) is high
Automated Agarose Gel Analysis
(Continued on facing page)
Fuhrmann et al.
and we see little to be gained with a Poisson or Gaussian
“statistical inverse problem” approach.
Based on the model described, we have developed a pro-
cessing strategy which comprises the following elements: (1)
an image preprocessing step for compensating for the delete-
rious effects in the image and putting it in a form consistent
with a one-dimensional linear model, (2) marker lane analy-
sis, for determining the quantitative aspects of the model,
particularly the size/mobility curve and band shape param-
eters, and (3) data lane analysis, for the detection and sizing of
bands in data lanes.
The purpose of preprocessing is to mitigate the deleterious
effects that are present in the data and are not directly related
to the linear model and unknown fragments; in effect, it is a
“data cleaning” step. The preprocessing also includes the in-
tegration of the image in the across-lane direction to obtain a
one-dimensional trace T(y) which is used in the subsequent
model-fitting using analysis-by-synthesis.
Correction for Optical Nonlinearity
The first step in the preprocessing is to correct for the optical
nonlinearity deliberately introduced to compress the visual
dynamic range. This is accomplished by the trivial operation
of squaring every pixel value. As a result of this step, the data
are represented using real or floating-point values rather than
Background subtraction has as its goal the removal of a
slowly-varying positive function, which may depend on the
distribution of fragment sizes in some unknown way, but
which is uninformative and does not enter into the linear
model. Various algorithms for doing this can be found in
several application areas in image processing. We adopt a pro-
cedure known as the MinMax Filter (J. Mullikin, pers. comm.),
modified slightly for application to 2-D image data. In the
modified MinMax Filter, a background image which is slowly
varying in the vertical direction and constant in the horizon-
tal direction is first determined and then subtracted from the
original image. A second 1-D MinMax filter is also applied to
the trace T(y) which results from the preprocessing, as de-
Impulsive Noise Filter
Dust specks and other particulate contaminating material
that are found frequently in the gels appear as isolated bright
pixels, or spots about 2–3 pixels in size. They are usually easy
to recognize because, according to our linear model, the im-
age is smooth in the across-lane direction. Our approach to
dealing with impulsive noise is to identify outlier pixels, then
delete and spline-fit through them.
Outlier pixels are identified on a row-by-row basis. In
each row, every pixel is tested to see whether or not it exceeds
twice the median value for that row, and if so it is considered
an outlier. Pixels so identified are deleted and replaced by
values obtained by cubic spline interpolation from the re-
maining pixels in the row. In the event that the outlier pixel
lies on the edge of the lane, it can happen that the result of
cubic spline extrapolation can be negative. Any negative val-
ues obtained are set to 0.
ber rn0211c, produced at the BCCA GSC in July 2001. Resolvable
fragments range in size from about 30,000 bp to 600 bp. The gel
contains 96 data lanes and 25 marker lanes, with marker lanes oc-
curring every fifth lane. The data lanes are never more than one lane
removed from a known DNA standard. (B) Close-up of a marker lane,
indicating the size in bp of the 37 marker fragments.
A typical agarose fingerprinting gel. (A) Shown is gel num-
Automated Agarose Gel Analysis
Symmetry and Monotonicity Constraints
The band shape model we use is a rectangle diffused via con-
volution with a Gaussian kernel. In the horizontal, or across-
lane, direction, this shape is symmetric about the center pixel
and monotonically decreasing from the center pixel to the
edges. To force our image data to conform to this model,
several simple steps are taken. The “center of mass” is com-
puted for each row, and these values are fit to a straight line
running in the vertical direction. Each row is then interpo-
lated onto a grid, which moves this line to the center of the
lane. Following this, the lane is averaged with its mirror image
about the center to enforce the symmetry constraint, and fi-
nally, to enforce monotonicity, each pixel is thresholded so
that its value does not exceed the value of its neighbor toward
Conversion to One-Dimensional Trace
Because of the separability of our band shape model, it is
possible to convert the 2-D image to a 1-D trace prior to analy-
sis via model fitting. All of the desired fragment-size informa-
tion remains present in the 1-D data; this greatly simplifies
the analysis in terms of computational and memory require-
ments. At every vertical position y, we eliminate the across-
row factor by matched filtering:
T?y? =?I ?x,y? Bh?x,y?dx
In this expression, Bh(x,y) is the across-row band shape
identified with vertical position y. If the 2-D image which
results from the first preprocessing steps conforms to the
noiseless linear model, then
This matched filtering procedure would be optimal un-
der an additive Gaussian noise model, even though we have
not postulated such a model. There are other ways to obtain
T(y) in the no-noise case, such as integration across the lane or
even simply sampling the center pixel. However, matched
filtering does provide some robustness to noise and is more
amenable to modifications to handle model inaccuracies,
such as saturation and adjacent-lane effects.
All of the computations implied in Equation 3 are carried
out on a discrete grid, with the integrations being replaced by
a summation over nine pixels, corresponding to the width of
the lane extracted by IMAGE from the fingerprinting gel. All
of the shape functions Bh(x,y), or templates are stored in a
look-up table which is computed prior to analysis.
In the linear model there is a large dynamic range in the
amplitudes of the bands, due to the compounding of the ef-
fects of fragment size and band diffusion. To counteract this
effect, both for purposes of visualization and also to ensure
that all sections of the lane are treated as equally important in
the model fitting, we have introduced a normalization which
applies increasing gain to pixels at increasing mobility. The
gain applied to the row at vertical position y is the inverse of
the integrated intensity of an unnormalized band at position
y. The nominal result of this normalization is that all bands
have equal area.
Examples of the results of the preprocessing steps de-
scribed above are illustrated graphically in Figure 3. In this
figure, the various stages of the preprocessing are shown as a
sequence of panels. The top panel shows
the raw image data displayed in false color
using the MATLAB “jet” colormap. The
second panel shows this image after back-
ground subtraction and correction for op-
tical nonlinearity. In the third panel, we
see the image after the output of the dust-
speck filter. This image also has a line run-
ning down the middle, indicating the es-
timated center of the lane. The fourth
panel shows the result of additional pro-
cessing to center the bands in the lane and
enforce constraints of monotonicity away
from the center. The result of gain correc-
tion to normalize the bands is shown in
the fifth panel. Also in the fifth panel the
image data have been shifted to the left to
account for a bulk mobility shift (relative
to a fixed standard) that is determined
during marker lane analysis. Finally, in
the sixth panel is shown the 1-D trace ob-
tained after the application of the across-
lane matched filter.
Marker Lane Analysis
The purpose of the marker lane analysis is
to determine the exact mobilities of 37
different fragments (Fig. 1B) that have ex-
actly known sizes. The nominal mobilities
of these fragments are known fairly accu-
rately, for a given experimental protocol,
restriction fragment mobility. Size vs. mobility data are for the 37 marker fragments, taken from
one lane of one of our standard gels. Individual data points are indicated by asterisks.
Curve illustrating the typical relationship between restriction fragment size and
Fuhrmann et al.
Data lane preprocessing steps. Panel 1: Raw image data (after lane tracking) using MATLAB “jet” colormap. Panel 2: Result of background subtraction and correction for
pointwise nonlinearity. Panel 3: Result of impulsive noise filtering. Panel 4: Result of enforced symmetry and monotonicity constraints. Panel 5: Result of gain correction and mobility
shift to standard location. Panel 6: Extracted one-dimensional trace.
Automated Agarose Gel Analysis
but vary slightly even within one gel due to subtle variation in
the gel conditions and nonuniformities in the electric field.
Once the marker band locations have been determined, the
fragment size/mobility relation can be determined for every
data lane by interpolation across the gel and down each lane.
After the locations of the bands have been identified, the
shapes of the bands are also analyzed to develop the templates
needed for a complete linear model for the data lanes in a gel.
Marker Band Detection
The first step in the marker lane analysis is the image prepro-
cessing described previously. The templates used for the
across-lane matched filtering are taken from a “standard” gen-
erated by the analysis of a typical gel produced under a given
In a marker lane there are 37 bands (Fig. 1B), most of
which are distinct and easily identified, except for the pair
numbered 18–19, and the group of seven at high mobility,
numbers 31–37. All of the marker lanes appear similar, differ-
ing only in some translation and distortion of the mobility
axis, and in the overall lane amplitude. Thus, the primary task
in marker lane analysis is to fit a distorted version of a stan-
dard template to the marker trace. In this respect, the analysis
has much in common with algorithms in pattern matching or
pattern recognition using deformable templates (Grenander
and Miller 1994; Jain et al. 1996; Zalubas et al. 1997).
In the marker template, the first 17 bands form a distinc-
tive and easily recognized pattern. This pattern is approxi-
mated by a translated and dilated version of a standard tem-
plate, with all the band peak amplitudes equal. The top sec-
tion of the trace is matched to a set of 4000 versions of the
template (100 translations times 40 dilations) until a best fit is
As the distortion of the mobility axis may be something
other than a simple translation and dilation, each marker
band must be individually isolated. This is accomplished by
sequentially finding each band using a prediction based on
previously identified bands. This sequential procedure is car-
ried out beginning at marker band 9 and operates in both
directions, up and down the trace, from this point. Quadratic
peak-finding is used to identify peak locations to subpixel
accuracy for known singlets, whereas a slightly different ver-
sion of the previously described pattern-matching procedure
is used for bands that do not have clearly identifiable peaks.
Marker Band Verification
The accuracy of the bandcalling depends critically on the cor-
rectness of the marker lane analysis; in short, there is little
room for error at this step. Accordingly, measures must be
taken to ensure that the marker lane analysis was successful.
For verification, we generate a synthetic marker trace using
the called band locations and the nominal band shapes. The
correlation between the synthetic and the true (preprocessed)
trace is then computed. This correlation must exceed 0.95;
otherwise, the marker lane is discarded.
Additional steps are taken to verify the marker lane
analysis, once all the individual lanes have been called. The
25 marker lanes across the gel are examined for any discrep-
ancies. For each marker band k, the 25 called mobilities across
the gel, mk(I), I = 1···25, should form a smooth curve. Each
function mk(I), I = 1···25, is fit to a low-order polynomial. Any
called locations that deviate significantly from this curve are
replaced by an estimated mobility found by polynomial in-
terpolation. The same interpolation procedure can be used to
replace data from “bad” marker lanes discarded by the corre-
Figure 4 shows an image depicting the raw data from all
25 marker lanes in a typical gel, with results of the marker
lane analysis superimposed. In this figure we have shown
only the high-molecular-weight bands at low mobility
(roughly the top half of the gel) to better illustrate the perfor-
mance. Note the subtle but significant variation in the marker
band location from lane to lane, the very reason that accurate
marker lane analysis is critical.
Band Shape Analysis
The marker bands, once identified, can be used to develop a
complete band shape model for a fingerprinting gel. This is
done by an empirical analysis of the second moments of the
bands, and fitting these to a sequence of second moments
consistent with the model. Because the model band shape is
separable, we can analyze the horizontal and vertical mo-
ments separately. A horizontal band is found by summing
pixels in the vertical direction, and vice versa. Furthermore,
for shape analysis, the horizontal and vertical bands are easily
normalized to unit area.
The analysis of the band shapes proceeds by computing
the horizontal and vertical second moments of the 28 well-
resolved singlets in the marker lanes. According to our model,
the second moment of each band can be attributed to three
sources: (1) a fixed rectangular pulse, (2) a fixed Gaussian
pulse with different horizontal and vertical widths, and (3) a
variable-width Gaussian pulse with circular symmetry and
width increasing with mobility. We have found that a useful
model describing the diffusion is that the standard deviation
(square root of the second moment) of the variable Gaussian
pulse grows quadratically with mobility. A complete descrip-
tion for the band shapes is found by fitting the two sequences
all the features described above.
Because of the computational impracticality of building
a separate model for each lane in the gel, the results of the
analysis for all 25 marker lanes are combined to give an “av-
erage” band shape model for the gel. From the band shapes
and the known fragment sizes, a nominal model for the am-
plitude curve can be generated as well. All of this information
is combined to generate a set of templates and other data
structures used in the data lane analysis, which we call the
complete gel model.
h(28) and ?2
v(28) to a model that incorporates
Data Lane Analysis
After the marker lanes have been analyzed, and a full para-
metric model has been developed for the gel, the analysis of
the data lanes with the unknown fragments can be carried
out. The approach used is one of analysis-by-synthesis, wherein
synthetic data are generated and matched to the true data.
The basic data model, after the preprocessing described
previously, is given by
A˜kBv?y − mk, Fk?
The band shapes are assumed to be exactly known, and
the amplitudes Ãkare nominally all equal to a constant. The
amplitudes will be subject to slight corrections as the analysis
progresses. The objective of the data lane analysis is to deter-
mine a set of fragment sizes fkwhich when used to generate a
Fuhrmann et al.
Results of marker lane analysis, showing low-mobility, high-molecular-weight marker bands 1–16. Shown in false color are 25 marker lanes, isolated from the full
121-lane gel and juxtaposed. Superimposed are red horizontal bars indicating the marker band locations as determined by BandLeader’s marker lane analysis.
Automated Agarose Gel Analysis
synthetic trace according to the model of Equation 5, provide
the best least-squares fit to the preprocessed data.
We adopt a discrete implementation of the model, in
which the possible mobilities mkare quantized onto a grid of
1500 possible values, logarithmically spaced between a mini-
mum and maximum mobility determined by the modeling
step. Typically this leads to step sizes on the “mobility grid”,
as it is called, of approximately 0.2 pixels at low mobilities
and 1 pixel at high mobility. This corresponds roughly to the
resolution available from the band shapes, which decreases
with increasing mobility. The typical quantization error in
mobility leads to errors on the order of 0.25% in fragment
size, ignoring other bandcalling errors.
The reason for the discretization of the mobilities onto
the mobility grid is that it simplifies the search procedure. We
use a search algorithm that shares properties of both a gradi-
ent algorithm and exhaustive search. As the band shape
model is stored in a look-up table, it is not possible to com-
pute gradients analytically; a numerical approach is required.
One of the characteristics of the trace T(y) is that the bands
tend to occur in isolated groups containing typically any-
where from 1–10 or 12 bands. We call these groups clusters. In
the space between the clusters, the signal value is near 0, and
this fact can be used to isolate clusters. In effect, by searching
for signal-absent regions the trace is broken down into a se-
quence of contiguous signal-present and signal-absent re-
gions. In this way the global model-fitting problem is reduced
to a number of much smaller local model-fitting problems.
Following the partitioning of the trace and the mobility grid
into isolated clusters, each cluster is analyzed for the best
model fit. Suppose that a cluster occupies pixels N1···N2and
that these same pixels correspond to mobilities M1··· M2on
the mobility grid. Define N = N2?N1+ 1 (number of pixels)
and M = M2? M1+ 1 (number of mobilities to test). Define
the test vector as s = T [N1: N2] in MATLAB notation. We seek
a model of the form
s ? Ax
where A is an N ? M matrix whose columns contain the in-
dividual band model. x is a vector of M integers, describing
the finite combination of bands to include in the model fit.
Most of the entries of x will be either 0 or 1, but our model
does allow for multiple copies of bands at the same mobility.
The knowledge of the amplitudes of the bands, or
equivalently the fact that the entries of the solution vector x
are integers, eliminates the model-order problem which often
plagues model-fitting procedures. There is no risk of overfit-
ting the data with too many bands. Increasing the number of
bands over that which gives the optimal fit will simply in-
crease the error between the data and linear combination;
thus the fitting procedure is in a sense self-limiting.
We have crafted a hybrid numerical gradient search to
solve the model-fitting problem for one cluster. We adopt a
cost function h(s,Ax), and seek the value of the vector s
which minimizes this cost function. For simplicity, the details
of the search algorithm are omitted here. The cost function is
a modified least-squares function, where the modifications
address the uncertainty in the amplitude curve. The modified
cost function places more emphasis on the shape of the target
function, and less on its amplitude.
Amplitude Curve Estimation
The determination of the amplitude curve ak, k = 1···1500 is
critical to the success of the algorithm described above. Nomi-
nally, the amplitude curve is known to within a single scale
factor prior to the data analysis. However, the amplitude
curve varies from lane to lane, and the model based on inte-
grated intensities is not sufficiently predictive to be used with-
out modification. Accordingly, the full data lane analysis re-
quires three passes through the data, with refinements of the
amplitude curve at each pass.
Pass 1. The amplitude curve is found by scaling the nor-
malized amplitude curve by a factor ?, where ? is chosen so
that 15% of the values in the trace vector y are above the
curve, and the remaining 85% below. We have found that this
normally causes the adjusted curve to “hug” most of the
single peaks, and that it allows the multiplet peaks to exceed
the curve. Using this scaled nominal amplitude curve, the
algorithm described above is run; however, only clusters with
singlets and resolved doublets are retained.
Pass 2. The normalized amplitude curve is again scaled
by a factor ?, this time chosen to achieve a least-squares fit
between the trace vector t and the retained clusters. This new
amplitude curve is again used in the gradient search proce-
dure, and this time all the clusters are retained.
Pass 3. The amplitude curve is multiplied pointwise by a
cubic polynomial. The coefficients of this polynomial are
chosen to minimize the squared error between a synthetic
trace and the data.
The results of the analysis of a typical data lane are sum-
marized graphically in Figure 5. The top panel shows the im-
age of the data lane in false color, after preprocessing. The
second panel shows the one-dimensional trace and the nomi-
nal amplitude curve based on the 15% rule plus the Pass 1
bandcalls indicated as small black circles. The fourth panel
shows the same trace with the Pass 3 amplitude curve and the
final bandcalls, indicated with small red circles. The bottom
panel contains a synthetic trace, generated according to our
forward synthesis model using all the results of the data lane
analysis. The agreement between the model and the prepro-
cessed data is evident here; the correlation between the actual
trace and the synthetic trace is 0.98 in this example.
Several heuristic safeguards have been built into BandLeader
to detect data lanes that are defective in some sense, and also
to recognize when there has been an error in processing and
thus the results cannot be used with confidence. Specifically,
there are eight conditions that will generate errors and four
conditions that will generate warnings.
An error will cause the lane data and any bandcalling
results to be discarded, while a warning is recorded in a log file
for further manual inspection if desired. Most conditions are
tested on the preprocessed one-dimensional trace signal. Dif-
ferent tests occur at different points in the processing.
The conditions that generate errors are as follows:
1. Empty lane. The total signal level is below a threshold de-
termined from the signal level in the marker lanes.
2. Nonrecombinant lane. Thirty percent of the total signal is
found within a single 10-pixel window. Nonrecombinant
clones are those which contain the vector DNA without
any insert DNA, causing there to be just one or two me-
dium-sized bands, depending on how many enzyme mo-
tifs are contained in the vector.
Fuhrmann et al.
Data lane bandcalling steps. Panel 1: Image data after preprocessing. Panel 2: Result of bandcalling, first pass. Panel 3: Result of bandcalling, second pass.
Panel 4: Result of bandcalling, third pass, with individual bands superimposed in red. Panel 5: Synthetic trace based on called bands and estimated model parameters.
Automated Agarose Gel Analysis
3. Low-mobility concentration. The total signal in the first 100
pixels is greater than 80% of the total signal in the lane.
4. Overcount. The sum of all called fragment sizes exceeds a
user-specified limit (e.g., 350 kbp).
5. Undercount. The sum of all called fragment sizes is below a
user-specified limit (e.g., 50 kbp)
6. Poor quality measure. The correlation between the prepro-
cessed trace and a synthetic signal generated using the
bandcalls as input to our forward model is less than 0.9.
7. No singlets found. No singlets were identified in the first
bandcalling pass, thus making amplitude curve estimation
8. Unknown error. A run-time software error such as divide-by-
zero or subscript out-of-bounds is trapped by the MATLAB
error handling routines. This prevents any remaining
“bugs” in the software from halting production bandcall-
ing, although naturally it is a cause for concern and usually
leads to investigation and correction of the problem.
The conditions that generate warnings are as follows
1. Low-mobility contamination. Some lanes contain high-
molecular-weight genomic DNA which does not belong to
either the vector or the insert. There are two tests for this
condition: (a) The number of pixels that are assigned to
band clusters in the first 180 pixels exceeds 60, and (b) the
number of bands called in the first 20 pixels exceeds three.
There is some modification to the processing under these
conditions; under condition (b), all bandcalls in the first 10
pixels are disregarded.
2. Possible overcall. There are 10 or more bandcalls in any
region of four or fewer pixels.
3. Saturation. The number of pixels in the high-mobility re-
gion of the lane which are set equal to the largest output
value of the imaging A/D converter exceeds a given thresh-
4. High-mobility concentration. The total signal level in the last
200 pixels exceeds the total signal in the first 400 pixels.
All of the errors and warnings are recorded in a separate
log file for each gel analyzed. In our experience it is rare that
all 96 data lanes are analyzed without error, with an average
of nine lanes per gel generating a warning or error record in
the log files.
Assessment of BandLeader Performance
To evaluate BandLeader’s performance on agarose fingerprint-
ing gels, we identified a “test set” of 140 human BAC clones
and 185 mouse BAC clones. These were selected from among
a set of BACs, available in GenBank (http://www.ncbi.
nlm.nih.gov), that had been sequenced completely and accu-
rately (i.e., were “finished”) at Washington University Ge-
nome Sequencing Center, the Whitehead Institute for Bio-
medical Research Sequencing Center, or at the Sanger Insti-
tute. We generated HindIII fingerprints of the BACs using our
standard laboratory conditions (Schein et al. 2003). The re-
sulting gel images were analyzed with BandLeader, and the
bandcall data compared to the “in silico” fingerprints identi-
fied by computer analysis of the sequenced BACs (see Meth-
ods). In our analyses, the sizes of “in silico” and actual restric-
tion fragments that were within an arbitrarily chosen 2% win-
dow were classified as identical. This criterion was applied to
all restriction fragments except those less than 600 base pairs
(bp). These were excluded because under our standard finger-
printing gel conditions they tend to be diffuse with low levels
of signal, obviating their accurate identification by any ap-
proach. The adoption of a test set of fingerprinted BACs was
crucial, as it provided us with objective “ground truth” test
data that made critical evaluation of BandLeader performance
The results of our analyses are shown in Figure 6A. For
comparison, we have included the results of a similar analysis
performed using automatically generated IMAGE bandcalls
(Fig. 6B). Of the 140 fingerprinted human BACs and 185
fingerprinted mouse BACs considered for this analysis,
BandLeader accepted 139 and 183, respectively. Analysis of
the sequences corresponding to these clones revealed that
they contain 16,782 HindIII restriction fragments. Of these,
BandLeader correctly identified 16,134, corresponding to a
sensitivity measure of 96.13%. Of the 16,736 fragments iden-
tified by BandLeader, 16,138 correctly identified a sequence-
predicted fragment, for a specificity measure of 96.42%. For
comparison, automated IMAGE bandcalls are only 60.99%
sensitive and 88.65% specific. Hence, although BandLeader is
not perfect, it offers a remarkable improvement over IMAGE.
Further, BandLeader outperforms the manual efforts of even
our most experienced technical staff, at throughputs far ex-
ceeding those possible by manual analysis of the fingerprint-
A major goal of the BandLeader project was to produce
software that would reliably detect multiplets, which we de-
fined as restriction fragments that comigrated within the
same data lane on a fingerprinting gel. We assessed the per-
formance of BandLeader in multiplet identification as fol-
lows. First, we identified as multiplets all fragments in BAC
sequence data falling within a restriction fragment size
window of +/? 2%. A similar grouping was done for the
BandLeader bandcalls corresponding to these clones. A total
of 322 human and mouse BAC sequences were analyzed and
3431 multiplets found in the sequence, for an average of ap-
proximately 10 multiplets per sequenced BAC. BandLeader
correctly predicted band multiplicity in 96.0 % of these
cases (see Fig. 7 and Methods). BandLeader’s performance,
even on the larger fragment clusters, is striking. For example,
BandLeader correctly identifies a cluster of 11 bands. Again,
although BandLeader is not perfect, it is distinctly superior to
any other bandcalling option we are aware of, manual or au-
We have described a set of software tools for the automated
analysis of agarose gel images acquired during the DNA fin-
gerprinting process. The various steps involved are common
to many image analysis and related engineering problems.
First, a model was established for the process of electropho-
resis and the digital image acquisition. The model included
sufficient detail to allow for variations in the model param-
eters, but was not so complex as to lead to unwieldy image
analysis tasks. Based on this model, procedures were derived
for determining the model parameters (through the use of
marker lanes) and for solving the inverse problem on the data
lanes. The integrated suite of tools, written in MATLAB and
collectively called BandLeader, has been used for the mouse
fingerprint mapping project at the BCCA (Gregory et al. 2002)
and is currently being used in fingerprint mapping projects
targeting the rat and bovine genomes (J. Schein, I. Bosdet,
C. Mathewson, N. Wye, R. Chiu, C. Fjell, H. Shin, S. Jones, and
M. Marra, unpubl.).
The use of automated image analysis for high-
Fuhrmann et al.
throughput fingerprint mapping
projects has important advantages.
Chief among these are the increases
in both the rate and accuracy of
data analysis and the opportunity
to reanalyze the very large finger-
print data sets if more suitable pa-
rameters are found. Further, the op-
portunity exists to repeatedly ana-
lyze the gel images to collect
statistics. For example, our entire
set of mouse (C57BL/6) fingerprints
(3500 gels containing more than
330,000 fingerprints) can be reanal-
yzed in 600 CPU hours. Since each
gel analysis is independent, the pro-
cess is amenable to parallelization,
such that only about 24 processors
would be needed to reanalyze the
3500-gel mouse set in one day.
The BandLeader software has
already proven of enormous value
in completing mapping projects
that would otherwise be unfeasible
given time and budgetary con-
straints. We have used versions of
BandLeader to analyze 13,629 fin-
gerprinting gels, generated in fin-
gerprint mapping efforts aimed at
bacterial, fungal, plant, and animal
genomes. Although BandLeader’s
performance is excellent, there is
room for incremental improve-
ment. With continued careful mod-
eling and algorithm improvement,
we see the potential for increased
bandcalling performance, with
gains in sensitivity, specificity, and
sizing accuracy. One of the more
challenging aspects of the auto-
mated trace analysis has been the
estimation of the amplitude curve,
which facilitates detection of mul-
tiple bands. The human eye adapts
quite easily to model variations and
aberrations, and in all but the most
pathological cases it is a simple
matter for our technical staff to
identify singlets visually and hy-
pothesize a smooth curve connect-
ing the peaks. In our automated
analysis, it has proven difficult to
sort out the singlets, multiplets,
and clusters, and derive a reliable
amplitude curve. The current ver-
sion of the software is doing a sat-
isfactory job with this particular
task, but on rare occasions, errors
may cause a lane to be failed.
There are several safeguards
built into the software to recognize
bad data and to abort the process-
ing for a given lane. Causes of un-
usable data include: (1) an empty
size”, plotted on the x-axes, refers to restriction fragment size as determined by computer analysis of
“finished” human and mouse BAC sequence data. Fragment sizes determined by BandLeader or
IMAGE are plotted on the y-axes. Fragments less than 600 bp are not considered. Each data point
represents the comparison of a restriction fragment predicted by sequence analysis to a restriction
fragment identified by either IMAGE or BandLeader. Red lines on the plot indicate a 2% size window.
Points falling within the 2% window are considered to represent identical restriction fragments. “Total
called” refers to the number of restriction fragments identified by IMAGE or BandLeader. “Total
correct” refers to the number of these fragments that match the corresponding sequence-predicted frag-
ments. This is a measure of specificity. “Total real” refers to the number of sequence-predicted fragments,
and “total found” refers to the number of these detected by the software. This is a measure of sensitivity.
Comparison of BandLeader (A) and IMAGE (B) performance on test gels. “Real fragment
Automated Agarose Gel Analysis
lane, (2) contamination, due to traces of a second clone or
other genomic material in the sample, and (3) nonrecombi-
nant BACs, which yield only one or two large bands in the
lane. On very rare occasions, there are good lanes for which
BandLeader analysis fails, presumably because the fit of the
data to the model is poor. One way to recognize this is to
compute the estimated clone size by summing all of the de-
tected fragment sizes; if this sum is outside of acceptable lim-
its, the lane can be rejected. In addition, any unexpected soft-
ware errors (such as a divide by zero) are “trapped” and allow
for the analysis of subsequent lanes to continue without the
entire process halting.
Currently, BandLeader relies on the data collection for-
mat used in our laboratories, and there is no flexibility in gel
format or choice of marker DNA. As the fingerprinting data
generation protocols are published and the marker DNA is
commercially available, this inflexibility is not a major ob-
stacle in the use of BandLeader to support fingerprinting ac-
tivities in other laboratories. However, we recognize that
there are several applications for a more flexible version of
BandLeader, including restriction analysis of plasmid and
other clones, and also genotyping. Hence, near-term future
research and development will focus on methods for the gen-
eralization of our techniques to other protocols. For example,
we intend to work towards the substitution of restriction di-
gested, sequenced BAC clones in place of the commercially
prepared markers. This will permit BandLeader to generate
data models from marker lanes that are equivalent to the data
lanes, and this in turn is anticipated to positively impact
bandcalling accuracy, especially for comigrating restriction
fragments. The extent to which accuracy can be improved is
limited however, as BandLeader already is capable of 96.42%
specificity and 96.13% sensitivity when used in a fully auto-
mated mode. This performance, and the robustness and reli-
ability of the code, have made BandLeader the only restriction
fragment identification system used in all of the large-scale,
high-throughput fingerprinting activities at the GSC.
To evaluate the performance of BandLeader, we compared the
restriction fragment sizes determined by BandLeader analysis
of fingerprinted clones to those generated by in-silico digests
of the sequences of the same clones. Three hundred and
twenty-five fully sequenced and finished BACs residing in
GenBank were identified and the sequences downloaded from
the National Center for Biotechnology Information (NCBI;
http://www.ncbi.nlm.nih.gov). All clone sequences were ana-
lyzed computationally to generate “in silico” fingerprints.
The actual clones were recovered from our local copy of the
RPCI-11 library (Osoegawa et al. 2001) and fingerprinted us-
ing our standard protocols (Schein et al. 2003). The resulting
fingerprinting gels were analyzed by BandLeader version
2.3.3, and the BandLeader-identified restriction fragments
were compared to the corresponding in silico restriction frag-
ments. The comparison was performed using a modified ver-
sion of the Needleman-Wunsch algorithm (Needleman and
Wunsch 1970). The modification involved setting a cutoff of
2% of the restriction fragment size in bp, such that only those
BandLeader and in silico fragments that were within a 2% size
sequences. Plotted on the x-axis is the number of fragments per multiplet detected by sequence analysis. Plotted on the y-axis is the number of
correct bands per multiplet identified by BandLeader. Numbers under the individual data points indicate the number of multiplets detected in the
sequences for a given number of bands in the multiplet. For example, there was a single 11-plet identified in the sequences, and BandLeader
identified correctly all 11 bands in the 11-plet. In another example, there were two 8-plets identified in the sequences, and all of the bands in all
these multiplets were correctly identified by BandLeader. The inset summarizes BandLeader performance for all multiplet sizes.
Performance of BandLeader in multiplet identification: 3431 multiplets were extracted from 322 “finished” human and mouse BAC
Fuhrmann et al.
952 Genome Research
window were classified as identical. In addition, a cutoff of
600 bp was introduced, as our standard laboratory protocols
do not yield reliable data for fragments that are smaller than
this size. The in silico and BandLeader-generated datasets
were each used in turn as the reference fingerprint set, and the
percentage of matching bands for all of the test clones was
taken and designated as the “sensitivity” and “specificity”
This work was supported in part by NIH grants 1-U01-
HG02042, Sequencing the Human Genome, and 1-U01-
HG02155, Sequencing the Mouse Genome. We gratefully ac-
knowledge the support of the British Columbia Cancer Foun-
dation, the British Columbia Cancer Agency (BCCA), and all
members of the Mapping Group at the BCCA Genome Sci-
ences Centre. M.M. is a Michael Smith Foundation for Health
The publication costs of this article were defrayed in part
by payment of page charges. This article must therefore be
hereby marked “advertisement” in accordance with 18 USC
section 1734 solely to indicate this fact.
The C. elegans Genome Sequencing Consortuim 1998. Genome
sequence of the nematode C. elegans: A platform for
investigating biology. Science 282: 2012–2018.
Chen, M., Presting, G., Barbazuk, W.B., Goicoechea, J.L., Blackmon,
B., Fang, G., Kim, H., Frisch, D., Yu, Y., Sun, S., et al. 2002. An
integrated physical and genetic map of the rice genome. Plant
Cell 14: 537–545.
Coulson, A., Huynh, C., Kozono, Y., and Shownkeen, R. 1995. The
physical map of the Caenorhabditis elegans genome. Methods Cell
Biol. 48: 533–550.
Coulson, A.R., Sulston, J., Brenner, S., and Karn, J. 1986. Towards a
physical map of the genome of the nematode Caenorhabditis
elegans. Proc. Natl. Acad. Sci. 83: 7821–7825.
Gregory, S.G., Howell, G.R., and Bentley, D.R. 1997. Genome
mapping by fluorescent fingerprinting. Genome Res.
Gregory, S.G., Sekhon, M., Schein, J., Zhao, S., Osoegawa, K., Scott,
C.E., Evans, R.S., Burridge, P.W., Cox, T.V., Fox, C.A., et al. 2002.
A physical map of the mouse genome. Nature 418: 743–750.
Grenander, U. and Miller, M.I. 1994. Representations of knowledge
in complex systems. J. Royal Stat. Soc. B 56: 549–603.
Hoskins, R.A., Nelson, C.R., Berman, B.P., Laverty, T.R., George, R.A.,
Ciesiolka, L., Naeemuddin, M., Arenson, A.D., Durbin, J., David,
R.G., et al. 2000. A BAC-based physical map of the major
autosomes of Drosophila melanogaster. Science 287: 2271–2274.
Jain, A., Zhong, Y., and Lakshmanan, S. 1996. Object matching via
deformable templates. IEEE Trans. Pattern Analysis and Machine
Intelligence 18: 267–278.
Marra, M., Kucaba, T., Sekhon, M., Hillier, L., Martienssen, R.,
Chinwalla, A., Crockett, J., Fedele, J., Grover, H., Gund, C., et al.
1999. A map for sequence analysis of the Arabidopsis thaliana
genome. Nat. Genet. 22: 265–270.
Marra, M.A., Kucaba, T.A., Dietrich, N.L., Green, E.D., Brownstein,
B., Wilson, R.K., McDonald, K.M., Hillier, L.W., McPherson, J.D.,
and Waterston, R.H. 1997. High throughput fingerprint analysis
of large-insert clones. Genome Res. 7: 1072–1084.
McPherson, J.D., Marra, M., Hillier, L., Waterston, R.H., Chinwalla,
A., Wallis, J., Sekhon, M., Wylie, K., Mardis, E.R., Wilson, R.K., et
al. 2001. A physical map of the human genome. Nature
Mozo, T., Dewar, K., Dunn, P., Ecker, J.R., Fischer, S., Kloska, S.,
Lehrach, H., Marra, M., Martienssen, R., Meier-Ewert, S., et al.
1999. A complete BAC-based physical map of the Arabidopsis
thaliana genome. Nat. Genet. 22: 271–275.
Needleman, S.B., and Wunsch, C.D. 1970. A general method
applicable to the search for similarities in the amino acid
sequence of two proteins. J. Mol. Biol. 48: 443–453.
Olson, M.V., Dutchik, J.E., Graham, M.Y., Brodeur, G.M., Helms, C.,
Frank, M., MacCollin, M., Scheinman, R., and Frank, T. 1986.
Random-clone strategy for genomic restriction mapping in yeast.
Proc. Natl. Acad. Sci. 83: 7826–7830.
Osoegawa, K., Mammoser, A.G., Wu, C., Frengen, E., Zeng, C.,
Catanese, J.J., and de Jong, P.J. 2001. A bacterial artificial
chromosome library for sequencing the complete human
genome. Genome Res. 11: 483–496.
O’Sullivan, J.A., Blahut, R.E., and Snyder, D.L. 1998.
Information-theoretic image formation. IEEE Trans. Info. Theory
Schein, J., Tangen, K., Chiu, R., Shin, H., Lengeler, K.B., MacDonald,
K., Bosdet, I., Heitman, J., Jones, S.J.M., Marra, M., et al. 2002.
Physical maps for genome analysis of serotype A and D strains of
the fungal pathogen Cryptococcus neoformans. Genome Res.
Schein, J., Kucaba, T., Sekhon, M., Smailus, D., Waterston, R., and
Marra, M. 2003. High-throughput BAC fingerprinting. In
Bacterial artificial chromosomes: methods and protocols (eds. S. Zhao
and M. Stodolsky), Humana Press Inc., Totowa, NJ (in press).
Shizuya, H., Birren, B., Kim, U.J., Mancino, V., Slepak, T., Tachiiri,
Y., and Simon, M. 1992. Cloning and stable maintenance of
300-kilobase-pair fragments of human DNA in Escherichia coli
using an F-factor-based vector. Proc. Natl. Acad. Sci. 89:
Sulston, J., Mallett, F., Staden, R., Durbin, R., Horsnell, T., and
Coulson, A. 1988. Software for genome mapping by
fingerprinting techniques. Comput. Appl. Biosci. 4: 125–132.
Sulston, J., Mallett, F., Durbin, R., and Horsnell, T. 1989. Image
analysis of restriction enzyme fingerprint autoradiograms.
Comput. Appl. Biosci. 5: 101–106.
Tao, Q., Chang, Y.L., Wang, J., Chen, H., Islam-Faridi, M.N.,
Scheuring, C., Wang, B., Stelly, D.M., and Zhang, H.B. 2001.
Bacterial artificial chromosome-based physical map of the rice
genome constructed by restriction fingerprint analysis. Genetics
Wechter, W.P., Begum, D., Presting, G., Kim, J.J., Wing, R.A., and
Kluepfel, D.A. 2002. Physical mapping, BAC-end sequence
analysis, and marker tagging of the soilborne nematicidal
bacterium, Pseudomonas synxantha BG33R. OMICS 6: 11–21.
Zalubas, E.J., O’Niell, J.C., Williams, W.J., and Hero, A.O. 1997. Shift
and scale invariant detection. Proc. ICASSP (Munich, Germany)
WEB SITE REFERENCES
http://www.sanger.ac/Software/Image; IMAGE software is available at
this Sanger Institute site.
http://www.ncbi.nlm.nih.gov; NCBI home page. Access to GenBank
Received October 11, 2002; accepted in revised form February 26, 2003.
Automated Agarose Gel Analysis