Likebrief communications
Bioinformatics
Finding genes in Plasmodium falciparum
he completion of the sequencing of chromosome 3 of the malarial parasite Plasmodium falciparum1 is a major step forward in our understanding of the Plasmodium genome. We have analysed this chromosome using GlimmerM, a freely available gene-finder developed specifically for the P. falciparum species2. GlimmerM was highly effective in finding nearly all the genes that were reported on P. falciparum chromosome 2 (ref. 2), and a newly retrained version was even more effective on chromosome 3, confirming virtually all reported genes and finding several additional ones. Using GlimmerM, we found 25 possible new genes on chromosome 3 in regions currently annotated as non-coding1 (Table 1). Although some of these are relatively short and may not be genuine coding sequences, six of them are longer than 400 base pairs (bp) and one is 2,187 bp. This last gene represents a 729-amino-acid protein that is encoded by one very large exon and one shorter exon. Open reading frames of this length in a chromosome with 80% A+T content are virtually certain to represent real genes. Three of the table entries, G802, G803 and G740, have detectable homology to var gene fragments on chro-
T
mosome 2; G802 and G803 are close enough that they might represent two portions of the same gene. Of the 215 protein-coding regions reported by Bowman et al.1, GlimmerM automatically finds 214 of them. (An earlier version of GlimmerM was provided to the annotators of chromosome 3.) The only gene missed is a short hypothetical protein (PFC0360w, 114 amino acids) with no homology to any known gene1. Finally, given that chromosome 3 is 12 per cent larger than chromosome 2, an extrapolation based on gene density would predict 234 genes on chromosome 3, consistent with our finding that additional genes could be present. GlimmerM is very accurate at identifying splice sites, having been trained on a carefully curated set of experimentally confirmed introns from the P. falciparum genome2. GlimmerM’s predictions suggest different splice sites for 49 of the 215 genes annotated on chromosome 3; all of these are hypothetical proteins. As with the annotation of hypothetical proteins for chromosome 2, substantial additional laboratory studies are needed in order to determine with confidence the exon structure of these genes. For chromosome 2, we used the polymerase chain reaction with reverse transcription for 13 hypothetical genes predicted by GlimmerM: all 13 predictions were confirmed2,3. Of course, as emphasized
previously2,3, GlimmerM is only one step, albeit an important one, in a process that should involve many other computational methods as well as careful human curation of the results produced by those methods. To achieve the highest quality in analysing genome sequences, peer-reviewed bioinformatics methods should be used when they are available. We consider it an oversight that the chromosome 3 annotation effort neglected to consider the predictions of GlimmerM, especially as a substantial part of the chromosome 3 analysis involves comparing the two chromosomes (see, for example, Table 2 of Bowman et al.1).
Mihaela Pertea, Steven L. Salzberg, Malcolm J. Gardner
The Institute for Genomic Research, 9712 Medical Center Drive, Rockville, Maryland 20850, USA e-mail: salzberg@tigr.org
1. Bowman, S. et al. Nature 400, 532–538 (1999). 2. Salzberg, S. L. et al. Genomics 59, 24–31 (1999). 3. Gardner, M. J. et al. Science 282, 1126–1132 (1998).
Table 1 Possible new genes on chromosome 3
ORF G802 G803 G815 G036 G070 G121 G124 G141 G144 G233 G256 G273 G282 G290 G294 G313 G356 G408 G410 G529 G612 G614 G702 G734 G740 No. of exons BP 1 1 1 2 1 1 1 1 4 1 1 1 2 1 1 2 2 2 1 5 3 2 2 1 1 450 468 210 192 220 531 411 255 348 240 459 282 249 243 285 285 293 213 231 339 249 279 2187 237 312 Length AA 150 156 70 64 73 177 137 85 116 80 153 94 83 81 95 95 131 71 77 113 83 93 729 79 104 PFC0010c PFC0010c PFC0030c PFC0125w PFC0175w PFC0260w PFC0260w PFC0280c PFC0280c PFC0415c PFC0440c PFC0465c PFC0480c PFC0485w PFC0490w PFC0505c PFC0555c PFC0580c PFC0580c PFC0810c PFC0910w PFC0910w PFC1010w PFC1060c PFC1065w PFC0025c PFC0025c PFC0035w PFC0130c PFC0180c PFC0265c PFC0265c PFC0285c PFC0285c PFC0420w PFC0445w PFC0470w PFC0485w PFC0490w PFC0495w PFC0510w PFC0560c PFC0590c PFC0590c PFC0815c PFC0915w PFC0915w PFC1015c PFC1065w PFC1070c + + + + + + + + + + Flanking genes Strand
Analysis of chromosome 3 of P. falciparum by GlimmerM1 generated 25 new gene models not included in current annotation. Columns show the number of exons in the predicted gene model, the length in base pairs (BP) and amino acids (AA), the nearest flanking genes on either side of the predicted new gene and the strand (+, forward strand; , reverse complement). A version of this table, linked to the DNA and protein sequences for each gene, is available at http://www.tigr.org/tdb/edb/pfdb/pf3table.html. ORF, open reading frame.
Lawson et al. reply — Pertea et al. report their interpretation of the published Plasmodium falciparum chromosome 3 sequence1 using The Institute for Genomic Research’s gene-prediction algorithm GlimmerM2. We believe, however, that their analysis is flawed, and highlights the problem that overreliance on a single algorithm can lead to mistakes in annotation. The accurate ab initio prediction of a eukaryotic coding sequence is extremely difficult because of the presence of intronic DNA. Gene predictions are distillates of the results from several algorithms that differentiate coding from non-coding segments. The role of the annotator is to identify putative genes and achieve the best prediction using the tools available. Ultimately, predictions can only be validated by laboratory studies. GlimmerM finds 25 new gene models, which Pertea et al. suggest could be an oversight in the original analysis. Predictions G802, G803 and G740 map to regions of chromosome 3 already assigned as varC pseudogenes, as the region of similarity is heavily frameshifted1. We do not believe that the prediction of exons from within these regions correctly interprets the data. Until laboratory investigation can show that these regions are coding, the parsimonious (and correct) annotation is that these regions are pseudogenes. Many gene-prediction algorithms are designed to have a minimum gene size (quite often 100 amino acids). These limits are implemented not because of any biological rationale, but because the inclusion of such open reading frames presents major difficulties in the interpretation of the data as most are likely to be non-coding. There is
NATURE | VOL 404 | 2 MARCH 2000 | www.nature.com
34
© 2000 Macmillan Magazines Ltd
brief communications
no indication that the 15 novel small predictions are coding, and the hexamer algorithm suggests that 12 of the 15 are non-coding. The remaining seven GlimmerM predictions form part of gene models that we consider to be borderline and are validating experimentally before inclusion in the chromosome annotation. Pertea et al. suggest that peer-reviewed bioinformatics methods are essential. Certainly, if the diverse outputs of multiple gene-prediction algorithms can be included, then the annotation should be good, but that does not make the gene prediction correct — it remains a prediction and not a confirmed gene model. Overreliance on the output of a predictive tool can lead to erroneous prediction and bad annotation. The well-established tools used in the analysis of chromosome 3, Genefinder, hexamer and ACEDB (P. Green and L. Hillier, unpublished software; R. Durbin, unpublished software) have a proven track record in eukaryotic gene prediction3. Our analysis of chromosome 2 using Genefinder/hexamer indicates that 40 modifications need to be made to the original gene predictions (20 per cent of gene models). Although these new predictions require confirmation, they highlight the inaccuracy of first-pass annotation from uncharacterized genomic DNA. We encourage re-analysis of genomic data to increase sequence accuracy: Genefinder, hexamer and GlimmerM all have roles to play in the (re)annotation of P. falciparum chromosomes and in improving the interpretation of sequence data.
Dan Lawson, Sharen Bowman, Bart Barrell
Pathogen Sequencing Unit, The Sanger Centre, Wellcome Trust Genome Campus, Hinxton CB10 1SA, UK e-mail: dl1@sanger.ac.uk
1. Bowman, S. et al. Nature 400, 532–538 (1999). 2. Salzberg, S. L. et al. Genomics 59, 24–31 (1999). 3. The C. elegans Sequencing Consortium Science 282, 2012–2018 (1998).
Figure 1 Comparison of acoustic data collected by an autonomous underwater vehicle (AUV) and a research vessel. a, Fish biomass at integrated intervals of 186 m along three transects on 21 July 1999; fish were identified as herring (Clupea harengus) by trawling. b, Strata-averaged biomass estimates from the AUV and research vessel were significantly correlated (r 0.935, P<0.001), lying around the (dotted) one-to-one line (Wilcoxon signed-rank test, P>0.05). c, Composite echogram collected by the AUV at 105 m (equipped with upward- and downward-looking transducers) as it passed less than 7 m beneath a large midwater herring school (left) and only 10 m above another close to the sea bed (right).
Oceanography
Fish do not avoid survey vessels
he precarious condition of the world’s fisheries is making ever-greater demands of the scientific assessment of fish stocks. Traditional assessments that rely on commercial catch statistics can have major shortcomings1 (as shown, for example, by the collapse of Canada’s northern cod stock2), increasing the need for more fishery-independent data. Acoustic surveys can provide such information3, but oceangoing research vessels have high operating
T
costs, and there is also widespread concern that fish avoid these vessels because of the noise they make, thereby biasing abundance estimates4. Here we present new data gathered by an autonomous underwater vehicle (AUV) showing that vessel avoidance is not a significant source of bias. Our investigation also heralds the arrival of AUVs as effective survey platforms. During acoustic surveys, sound pulses are transmitted vertically downwards into the water at regular intervals (typically 1 s) from a survey vessel travelling along defined transects. Fish density is calculated by integrating the intensities of the returning echo5, and is then interpolated to give an estimate of the abundance in the survey area3. We deployed the AUV Autosub-1 (ref. 6) 200–800 m ahead of the research vessel Scotia on eight transects in water 60–180 m deep during an acoustic survey of herring in the North Sea. Herring have the most sensitive hearing of the commercially exploited species examined so far7 and are more likely than any to react to vessel noise. Autosub-1 is unmanned and follows pre-programmed
© 2000 Macmillan Magazines Ltd
mission trajectories. In comparison to the 68-m Scotia, it is small (torpedo shaped, 7 1 m) and extremely quiet (being propelled by an electric motor). Avoidance of Autosub-1 by herring is minimal: passing unprecedently close to a school (Fig. 1c), the vehicle caused only the localized school compression that typically occurs on close approach of predators8. Autosub-1 was equipped with the same type of 38-kHz scientific echosounder as Scotia, and gathered equivalent acoustic data before the research vessel arrived. If fish avoided the Scotia, then it should have detected fewer fish than Autosub-1. At the integrated resolution, there are small-scale temporal and spatial differences between the AUV and research-vessel data (Fig. 1a), but the underlying similarity in magnitude and trend is clear. For statistical comparison, data from each source were aggregated into independent strata of equivalent geographical area. The amount of fish detected by the research vessel was not significantly different from that detected by the AUV (Fig. 1b). Scotia is very quiet, having been built
35
NATURE | VOL 404 | 2 MARCH 2000 | www.nature.com
Readers
Add Comment