ABSTRACT
Data driven computational biology relies on the large quantities of genomic data
stored in international sequence data banks. However, the possibilities are
drastically impaired if the stored data is unreliable. During a project aiming
to predict splice sites in the dicot
Arabidopsis thaliana
, we extracted a data set from the
A.thaliana
entries in GenBank. A number of simple `sanity' checks, based on the nature of
the data, revealed an alarmingly high error rate. More than 15% of the most
important entries extracted did contain erroneous information. In addition, a
number of entries had directly conflicting assignments of exons and introns,
not stemming from alternative splicing. In a few cases the errors are due to
mere typographical misprints, which may be corrected by comparison to the
original papers, but errors caused by wrong assignments of splice sites from
experimental data are the most common. It is proposed that the level of error
correction should be increased and that gene structure sanity checks should be
incorporated-also at the submitter level-to avoid or reduce the problem in the future. A non-redundant and error corrected subset of the data for
A.thaliana
is made available through anonymous FTP.
Biological sequence databases offer scientific researchers unique opportunities
for working with large quantities of genomic data. The GenBank, EMBL and DDBJ
databases aim to contain all published DNA sequences. In its feature table each
entry holds information about transcription, splicing and translation associated signals. This information may be used to create large data subsets, where the
sequences are related to their functionality.
Unfortunately, GenBank is currently not optimally suited for such extraction. Too often errors occur in the entries. Almost all major
publicly available genomic databases suffer from this condition. The problem of
corrupt databanks has been pointed out earlier (
1
,
2
). However, the situation remains more or less the same.
For the construction of a data set relating the DNA (or pre-mRNA) sequence to its alternating exon-intron structure, we extracted relevant information from the
Arabidopsis thaliana
entries in GenBank ver. 87 (
3
). The data set has been used in work aiming to predict splice sites in the
dicot by means of neural network algorithms (Hebsgaard, S.M., Korning, P.G.,
Tolstrup, N., Engelbrecht, J., Rouze, P. and Brunak, S., submitted) and (
4
). A common feature of neural network algorithms is their ability to cope with
non-linearities in the association between objects and categories. During
training on the initially extracted data training difficulties were encountered
for many of the splice sites (
1
,
2
). This led us to perform a range of `sanity' checks that revealed quite a large
percentage of errors in the
A.thaliana
entries. The most effective of these checks being examination of particular splice sites, which diverged strongly from the `extended consensus sequence' of the organism. The latter was created from a measurement of the Shannon information around the splice sites and will be
described below. Of the 999
A.thaliana
entries in GenBank rel. 87, 167 contained enough splicing information to be of
interest to our work.
A list of corrupt
A.thaliana
GenBank entries has been compiled and in most cases we explain how to correct
the errors. When possible, the original contributers of the entries have been
contacted and the errors should be corrected in future versions of GenBank. The
errors detected are often shift errors, which are not caused by misprints only.
Instead, many errors are created when splice sites are assigned by homology
with wrong cDNA assignments. We suggest that the consistency of the annotation
is assessed prior to the submission in order to reduce the error rate. Such
proof-reading may be assisted by using internet available servers, which are
able to evaluate the correlation between the nucleotides in a single splice
site from a consensus view point, rather than relying only on a more individual
assignment by hand or by cDNA homology.
An initial data set was extracted from GenBank 87.0 by use of a software
scanner. All
A.thaliana
entries were examined. The criteria for inclusion of an entry in the data set
were the following: (i) it must contain two or more introns; (b) it must not in
any way be indicated as being partial; and (iii) there must be no logical
conflicts between the entry's description of its components, e.g. a substring
which is defined as being both an intron and an exon. (i) Means that all genes
in the data set contain internal exons, while (ii) means that the CDS in a
given entry in the data set is neither mentioned as being partial in the
DEFINITION, nor as being partial or as having uncertain borders in the FEATURES section. Due to the
third criterion 20 entries were excluded, having direct annotation conflicts
not caused by alternative splicing.
Table 1
The
A.thaliana
entries found in GenBank rel.87 likely to be erroneous
After this reduction 167 entries remained in the data set and training of the
neural network algorithms following (
5
) was performed. The output neuron of the networks contained one unit, trained to
classify the central nucleotide in the DNA segment as being either a splice
site or a non-splice site. It soon became clear that unless very large networks were
used, certain sites in the data set could just not be learned. This phenomenon
is symptomatic of impurities in the input data. As a result of this, a number
of sanity checks was performed on the data set. These included: checks for
reading frame inconsistencies, checks for presence/lack of start and stop
codons, checks for introns below the minimal functional intron length, and
finally investigation of splice sites deviating significantly from their
extended consensus sequence.
The general sequence patterns were found by plotting the Shannon information
content in the context of aligned splice sites in the data set (
6
). The formula used was
H ( i ) = - {sum from {alpha = 1} to 4} {P sub i sup alpha} {log sub 2} {P sub i sup
alpha}
1
where
H
(
i
) is the Shannon information at position
i
, and
P
a
i
is the probability of finding nucleotide [alpha] ([alpha] {
A
,
C
,
G
,
T
}) at position
i
. The probabilities were computed from the frequencies of the nucleotides in the
data set. Plotted together with the nucleotide frequencies at each locus, such curves can be said to constitute an extended consensus sequence of the splice sites. The most conserved nucleotides
in the
A.thaliana
consensus sequences found were AG'GTAAGT and TGYAG'GT for donor and acceptor sites respectively.
First, in each gene it was examined whether the sum of the number of nucleotides
in the translated parts of the exons did obey the modulo three rule and thus
did not contain incomplete codons. There turned out to be one entry that did
not obey the modulo three rule, namely ATDNABFS1.
Secondly, a scan for the lack of start codons (ATG), for the lack of stop codons
[TAA TAG and TGA,
A.thaliana
uses the standard genetic code (ftp://weeds.mgh.harvard.edu/pub/codon/ath.cod)],
and for the presence of stop codons inside the reading frame was performed. The
entry ATDNABFS1 turned out to lack a start codon and contained no less than 52
stop codons in the reading frame. Correspondence with the author revealed that
a fair CDS is found when the first nucleotide of the start codon is taken as
position 393 instead of position 374, and the first nucleotide of the stop
codon is taken as position 3134 instead of position 3305.
Thirdly, the data set was scanned for entries containing very short introns. As they have to form a lariat, splicing becomes functionally impossible if the intron is very short. According to a recent review (
7
,
8
) a minimal functional length of 64 nt is reported for dicots. While it seems
that this number should be lowered somewhat (our scan revealed seemingly
correct introns of lengths between 58 and 64 nt), an intron of length 18 nt as
intron four in ATU11033 was an error (has later been corrected by the authors).
This entry was removed from the data set.
The largest number of errors was found by examining apparent deviations from the
donor and acceptor splice site consensus sequences. Of the 878 donor sites and
880 acceptor sites in the extracted data set, 45 donor sites and 44 acceptor
sites turned out to lack the normal dinucleotides, GT and AG. The non-consensus splice sites were found in 26 entries only, which were examined
manually by comparing to the original and related publications, and by
searching for related entries in the databases. From this careful examination
we strongly suggest that
all of the acceptor sites and all the donor sites but three are errors
!
The three donor sites represent true exceptions. They appear in the entries ATHACOACAR (donor
site 15), ATHFUS6A (donor site 3) and ATHHANKA (donor site 6). The use of GC
instead of GT at the donor site is observed in the order of ~1% of the
A.thaliana
introns. Their occurrence as true donor sites has recently been reported in the
myrosinase genes from
Brassicaceae
, the family including the species
A.thaliana
(
9
).
A list of the remaining entries, with comments on the errors, can be found in
Table
1
. The errors may be divided into a small number of categories. A large part of
them (21/24) are caused by bad localization of a feature in the sequence, most
often the splice sites. If the assignment is performed by comparing the gene sequence and
homologous cDNA without caring for a proper consensus, this kind of mistake may
occur (see Fig.
1
). Since the sequences around the 5' and 3' splice sites often are similar, alignment ambiguities will result
and one among several good solutions may be discarded. For this kind of errors,
anyone has access to the same basic information as the authors were having
themselves and sometimes even more due to new data and publications that have
appeared later. In a few cases other kinds of errors are indicated. Their
occurence may be related to sequence errors coming from editing or from the
sequencing itself. Here the authors only may be able to track down and confirm
the suggested errors, albeit the indication of errors is always very strong.
Examples of such indications are given by redundant entries for the same gene
(ATHRD29AB), or by entries for very close homologs (ATACP, ATU18969), or by
clear biological inconsistencies (ATU11033). This latter case is among those where entries have been modified in a later version of GenBank,
confirming our suggestion of error. In several cases we have been able to
contact the authors, ending up with consistent modifications of the entry for
most of the genes.
It is well known that the data in GenBank have a certain degree of redundancy.
Therefore the length of the longest common substring which would occur in a collection of random strings corresponding in size to the
data set was calculated. Then a scan for common substrings above this length
was performed and entries where such substrings did occur were investigated
manually. There are basically two classes of redundancies: First, there are
cases where the same gene has been submitted twice; either by the same author or by
different authors. These sequences tend to vary a bit in length, but otherwise
be more or less identical. Secondly, there are cases where two different genes
are members of the same gene family and so closely related that only one of
them should be included in the data set (see Table
2
for the redundant sequences and comments).
The redundancy reduction has been made in order not to overestimate the
performance of intron splice site finding computer programs. If the aim was,
for example, to study alternative splicing this would be another matter. For
quite some time a rigorous procedure for this removal has been common practice
in the field of protein structure prediction (
10
-
12
). We are in the process of determining a proper alignment threshold for
nucleotide data making it possible to distinguish between cases where the
splice sites may be found safely by alignment, and those where genuine
prediction is needed (Tolstrup, N., Dalsgaard, K., Engelbrecht, J. and Brunak,
S., in preparation.) Today, algorithms are tested on an ill-defined mixture of the two.
A search for `virtual contradictions' in the data set was also performed. A
virtual contradiction for a given substring-size is defined as two identical sequences, where the central nucleotides
have different functionalities, e.g. splice site and non-splice site, or coding nucleotide and non-coding nucleotide. Four such pairs of virtual contradictions were
found (see Table
2
). The contradiction between the pair ATRPBl and ATRPII seems to be caused by
alternative splicing. More interestingly, however, are the contradictions
between the three pairs ATU18968/ATHRPSlSA, ATU18970/ATHRPS15A and
ATU18970/ATU18968 respectively. Although further analysis is needed and actually undertaken (Rouzé, P., Breyne, P. and Van Gysel, A., in preparation), these virtual
contradictions may stem from a new kind of recombination and/or transposition
event. Interestingly enough, this is not mentioned by the contributors in the
corresponding papers (
13
,
14
).
It is clear that the errors found may be divided into a few categories. Many of
the errors are simple shift errors, created somewhere along the path from the
laboratory to the database, which can be easily corrected by comparison of the
sequence to its corresponding cDNA if the latter is available (taking into
account well documented consensus sequences). Other errors are caused by
sequencing or typing errors. There is however a class of errors, which are more
or less impossible to correct, unless one wants to go back to the laboratory.
These errors occur when a DNA string, having been sequenced correctly, is
assigned splice sites by homology with a more or less distant species. It would
be highly desirable if such speculative assignments were stated clearly in
GenBank, as the quality of such entries is often not good enough for creating
powerful data driven computer based methods. Splice sites shifted a few
nucleotides, constitute directly `negative information' to for example neural
network algorithms, and are extremely harmful to their learning and
generalisation capabilities if the errors are frequent.
After the above mentioned erroneous entries had been corrected or removed, very
clear consensus sequences did emerge from the Shannon plots. They correspond to
the ones mentioned above, but with much more certainty at each position. The
contrast to the Splice Site Consensus Table in AAtDB
(ftp://weeds.mgh.harvard.edu/aatdb-info/Splice_Site_Consensus) is striking: in AAtDB a more disorderly
picture emerges; >6% (maybe up to 12%, only single nucleotide frequencies are
given) of the donor sites do not contain the GT dinucleotide. The situation is
the same for AG at the acceptor sites. The
A.thaliana
sequences from AAtDB (rel. 3-5) may have been selected by other criteria than the ones adopted in this
work. Still, observing differences as large as these, we do believe that AAtDB
might benefit from a thorough sanity scan as well.
One reason for the success of finding the above mentioned errors in GenBank, is
the fact that complete coding sequences were used only. These have to obey a
large number of rules in order to be processed correctly by the cell's
machinery. However, the consensus sequence check can be used anywhere splicing
occurs.
Biological research does now, and will even more in the future, depend on
information retrieved from large genomic databases. As they control the
information flow between scientists, it would be an obvious place for proof-reading of the data, at least by computers and preferably checked by human
experts. The examples from this paper demonstrate the necessity of such proof-reading. From similar work on human genes, and on genes from
C.elegans
, we estimate that the GenBank error rate is somewhat smaller for boths
organisms, in the order of 5-10%. The error rate is likely to depend of the average amount of work
done on the entries, but also the diversity of the splice sites, from a given
organism.
The cleaned
A.thaliana
data set is made available to other researchers by anonymous FTP from
ftp.cbs.dtu.dk:/pub/arabidopsis/NetPlantGene.seq. The current version contains
147 genes from 145 GenBank entries all in all. The intron splice site prediction method
has also been made available by electronic mail and through the WWW (Hebsgaard,
S.M., Korning, P.G., Tolstrup, N., Engelbrecht, J., Rouzé, P. and Brunak, S., submitted). Send a file containing the word `help' to the internet address
NetPlantGene@cbs.dtu.dk to obtain information on sequence formats and other
details.
The authors wish to thank Anders Gorm Pedersen for helpful discussion, and
Krzysztof Rapacki for providing the primary GenBank scanner. This work was
supported by the Danish National Research Foundation.
REFERENCES
Return