mtDNA Data Mining in GenBank Needs Surveying

Main Text

To the Editor: Since the first sequencing of the complete human mtDNA genome,¹ both the sequencing techniques and the quality of commercial kits have improved greatly. This has led to a growing number of reports for complete mtDNA sequences from the fields of molecular anthropology, medical genetics, and forensic science; and there are now over 6700 complete or near-complete mtDNA sequences available for study.² However, in comparison to the pioneer manual-sequencing efforts in the early nineties, the overall mtDNA data quality, especially in the medical field, is still far from satisfactory.³ Sequencing errors and inadvertent mistakes in the reported mtDNA data are not infrequent.^4–10 Deficient mtDNA data sets of complete genomes can have important consequences for the conclusions achieved in many studies and may also pose problems for any subsequent reanalyses.

Most recently, Pereira and colleagues¹¹ discussed the overall picture of the mtDNA genome diversity in worldwide human populations with a comprehensive reanalysis of 5140 published complete or near-complete (lacking some control region information) mtDNA sequences. Their study represents an important advance in defining the effects of gene structures on limiting mtDNA diversity and may have valuable implications for mtDNA studies in the medical field.¹¹ However, all of the data used in the study by Pereira et al.¹¹ were directly retrieved from GenBank without any scrutiny for problematic or flawed data that should have been excluded. Many of the mtDNA sequences analyzed in their study¹¹ have in fact already been questioned in the literature or even corrected by their authors, but unfortunately, in several instances the new corrected versions of the sequences have not been made generally available or updated in GenBank.

In Table 1, we list some of the problematic data sets and single sequences used by Pereira et al. in their study.¹¹ Among them is the original data set of Herrnstadt et al.,¹² which was announced by the authors¹³ as having been corrected, although the new sequences have never been entered into GenBank. Portions of those coding-region data (in either corrected or uncorrected form) were augmented by the associated control-region data and published in several papers; thus, none of these expanded data can be downloaded from GenBank but have to be retrieved from the figures in the corresponding articles. To cite a more recent example, the African mtDNA data set published by Gonder et al.¹⁴ is of particularly poor quality. These sequences are incompletely recorded (as already mentioned by Behar et al.¹⁵); the most extreme instance of this is the haplogroup L0k1 sequence EF184609 that lacks as many as 25 expected variants scattered along the whole pathway from the haplogroup root to the revised Cambridge reference sequence (rCRS).¹⁶ Also, several different phantom mutations appear throughout the data set; in particular, five sequences have been affected by phantom base changes to G within the short 9949–9978 stretch. We have annotated problems in 14 sequences by way of example, but nearly all sequences of Gonder et al.¹⁴ may suffer from overlooked variants, except for the three sequences from the well-described West Eurasian haplogroups J1 and N1. Additional details are given in the Supplemental Data, available online.

Table 1.

List of Some Flawed Data and Uncorrected Sequences Employed in the Study by Pereira Et Al.¹¹

GenBank Data	Cause of Error	Reference	Errors Detected or Corrected
DQ156212, DQ156214	NUMT contamination	Montiel-Sosa et al.27	Yao et al.28
DQ112878	NUMT contamination	Kivisild et al.29	Yao et al.28
DQ112952	Missed mutation	Kivisild et al.29	this study
DQ341068.1	Artefactual recombination	Torroni et al.30	Behar et al.;15 DQ341068.2 (updated May 5, 2009)
AP008259, AP008269, AP008278, AP008306, AP008552, AP008776, AP008777, AP008798, AP008799, AP008801, AP008803	Artefactual recombination	Tanaka et al.23	Kong et al.21
Various	Missed mutations	Maca-Meyer et al.31	Palanichamy et al.32
Various	Phantom mutations and documentation errors	Herrnstadt et al.12	Herrnstadt et al.;13 Bandelt et al.19
Various	Missed mutations	Rajkumar et al.33	Sun et al.34
Various	Various	Gonder et al.14	Behar et al.;15 this study
AY963586.1	Editing error in GenBank submission	Bandelt et al.4	AY963586.3 (updated June 29, 2009)
EF660912–EF661013	Phantom mutations and missed mutations	Gasparre et al.17	This study
AM260596–AM260597	Missed mutations	Annunen-Rasila et al.35	This study
AY289073	Missed mutations and recombination	Ingman and Gyllensten36	This study
AY195745, AY195756, AY195767, AY195775	Phantom mutations and missed mutations	Mishmar et al.37	Brandstätter et al.;38 this study
EU095205, EU095208, EU095250	Phantom mutations and missed mutations	Fagundes et al.39	Perego et al.;40 this study
AY339437, AY339463.2, AY339546, AY339549,AY339581.2, AY339582	Phantom mutations and missed mutations	Finnilä et al.41	This study
AF46968, AF346973, AF347006	Missed mutations, phantom mutations, and recombination	Ingman et al.42	Kong et al.;21 this study
Various	Phantom indels and missed mutations	Kumar et al.43	This study
EU597580	Missed mutation	Hartmann et al.44	This study
DQ826448, DQ834253-DQ834261	Various	Phan et al. (unpubl. data)a	This study
DQ418488, DQ437577, DQ462232–DQ462234, DQ519035	Various	The State Key Laboratory of Forensic Sciences (unpubl. data) a	This study
DQ358973–DQ358977	Documentation errors (position 750)	Detjen et al. (unpubl. data)a	This study
EF446784, EF488201	Poor sequencing quality (artefactual heteroplasmy)	Noer et al. (unpubl. data)a	This study

^aUnpublished data were released by GenBank, and detailed annotation of the potential errors is given in the Supplemental Data.

Again, if one examines the ten Vietnamese complete mtDNA sequences that were submitted to GenBank by Phan et al. and used in the Pereira et al. study,¹¹ it is possible to see errors of many kinds. First, all sequences miss three expected variants (A263G, 315+C [or written as 315insC], and C14766T). Second, there are many phantom mutations that are not observed elsewhere. Third, several sequences are incomplete; e.g., the haplogroup M7b1 sequence DQ826448 lacks an additional nine expected variants by oversight or artefactual recombination. This sequence also has a base-shift error and harbors six suspicious transversions. Finally, the haplogroup N9a sequence (DQ834258) has a problem with artefactual recombination. Detailed annotations for these Vietnamese mitochondrial genomes and a few more GenBank complete mtDNA sequences with similar problems are listed in the Supplemental Data.

It is likely that most conclusions in the Pereira et al. study¹¹ would essentially remain unaltered after the flawed data sets or single problematic sequences were filtered out. Nonetheless, the results reported in their tables would benefit from a reanalysis using an improved version of the complete genome database. It depends on the particular aspect under study as to whether a small residue of errors would matter or not. A good example of where it would cause problems is with the estimation of the transition:transversion ratio, because transversions are relatively rare and flawed data are often enriched in transversions (see phantom mutations in the Supplemental Data). The number of artefactual transversions from some of the data sets does appear to be raised, in particular in the sequences from Gasparre et al.¹⁷ (Table 1 and Supplemental Data). In addition, misalignment of seven sequences (DQ341085–DQ341090 and EU600343) in the Pereira et al. study¹¹ has produced at least another 21 artefactual transversions at positions 292, 296–299, 300, 302, and 303. Similarly, the insertion 5436insG in DQ246818 has been shifted by four base pairs and scored as C5437G 5440insC, so that a transversion is created artificially. Suboptimal alignment induced further artificial transversions: e.g., the two sequences AY922293 and AY922275 are identical in the 54–60 region (GTTATT versus GTATTTT in the rCRS) and yet the former was interpreted as 55insT-59delTT and the latter as 56T-57A-60delT in that region by Pereira et al.¹¹ Inconsistent alignment is also seen in the long C stretch in regions 16184–16193 and 303–315 in the Pereira et al. study.¹¹

Another instance in which a small amount of error could have a significant influence involves the estimation of the positional rate spectrum along the molecule. For instance, the change C12705T (characteristic of non-R status) is a rare mutation but was overlooked by Gonder et al.¹⁴ half a dozen of times, and the mutation T10810C (characteristic of non-L2′6 status) was overlooked an additional eight times.¹⁴ Similarly, the estimated rate of any mutation scored between the roots of frequent haplogroups in the mtDNA phylogeny gets inflated by the use of incomplete or recombinant sequences. Thus, the incorporation of flawed data considerably distorts the estimation of rates for a number of positions. The same effect may be caused by systematic documentation errors, as in the case of the 14766 transition, which has often been misrecorded because of the discrepancy at 14766 between rCRS and a partly corrected CRS (which was in use for a long time).^3,10 Moreover, for parts of the mtDNA phylogeny in which numerous mutations are missed in the data used, estimation of haplogroup coalescent times becomes distorted. The consequences of using wrong data can be dramatic under particular circumstances, as we have discussed before.^{3–10,18–21} Fortunately, the standard and quality of sequencing from the large laboratories (with long-standing experience) has improved over the years, and the results from these laboratories are now setting the standard against which all smaller institutions should compare themselves. This does not preclude the possibility that single sequences from data sets released by large laboratories may have minor problems.

Bioinformatics-based projects are more and more popular, drawing conclusions from whatever can be retrieved from GenBank (e.g., Gonder et al.'s data¹⁴ were also employed by Atkinson et al.²²). However, the common practice of mining mtDNA data from GenBank or other genomic resources should be carried out with the necessary caution in order to avoid erroneous claims in future studies. For instance, one could foresee that the use of the original incorrect sequences by Tanaka et al.²³ would easily lead to erroneous signals of mtDNA recombination.²¹ To eliminate errors in the published mtDNA data or at least to exclude the suspicious GenBank entries from any subsequent reanalyses, we call for a stringent scrutiny of reported data and a bookkeeping annotation of errors in the public databases, such as in Phylotree.org (maintained by Mannis van Oven)² and some personally owned websites (e.g. Ian Logan's website). For the benefit of science, submissions to GenBank should be revised as promptly as possible by the authors responsible for the data in question. And, importantly, when submitting a new paper for publication, authors should provide evidence that their data has been checked for the more common errors that come from poor sequencing technique and data handling, as well as for discrepancies between the actual submissions to GenBank and what has been shown or inferred in the paper. But instances will remain in which authors either do not react or claim that they did everything right (as in the prominent case analyzed by Bandelt and Kivisild²⁴ and Parson²⁵). Therefore, when one plans to perform a cumulative reanalysis of mtDNA data, one cannot avoid making a substantiated, though partly subjective, decision as to which data are to be included and which are to be excluded, as has been exemplified in a recent paper by Soares et al.²⁶

Supplemental Data

Supplemental Data include one appendix and can be found with this article online at http://www.cell.com/AJHG.

Web Resources

The URLs for data presented herein are as follows:

• GenBank, http://www.ncbi.nlm.nih.gov/Genbank/

• Ian Logan's website, http://www.ianlogan.co.uk

• PhyloTree.org, http://www.phylotree.org/