Limits...
Accurate splice site prediction using support vector machines.

Sonnenburg S, Schweikert G, Philips P, Behr J, Rätsch G - BMC Bioinformatics (2007)

Bottom Line: We apply our method to the genome-wide recognition of splice sites in Caenorhabditis elegans, Drosophila melanogaster, Arabidopsis thaliana, Danio rerio, and Homo sapiens.Our performance estimates indicate that splice sites can be recognized very accurately in these genomes and that our method outperforms many other methods including Markov Chains, GeneSplicer and SpliceMachine.We provide genome-wide predictions of splice sites and a stand-alone prediction tool ready to be used for incorporation in a gene finder.

View Article: PubMed Central - HTML - PubMed

Affiliation: Fraunhofer Institute FIRST, Kekuléstr, 7, 12489 Berlin, Germany. Soeren.Sonnenburg@first.fraunhofer.de

ABSTRACT

Background: For splice site recognition, one has to solve two classification problems: discriminating true from decoy splice sites for both acceptor and donor sites. Gene finding systems typically rely on Markov Chains to solve these tasks.

Results: In this work we consider Support Vector Machines for splice site recognition. We employ the so-called weighted degree kernel which turns out well suited for this task, as we will illustrate in several experiments where we compare its prediction accuracy with that of recently proposed systems. We apply our method to the genome-wide recognition of splice sites in Caenorhabditis elegans, Drosophila melanogaster, Arabidopsis thaliana, Danio rerio, and Homo sapiens. Our performance estimates indicate that splice sites can be recognized very accurately in these genomes and that our method outperforms many other methods including Markov Chains, GeneSplicer and SpliceMachine. We provide genome-wide predictions of splice sites and a stand-alone prediction tool ready to be used for incorporation in a gene finder.

Availability: Data, splits, additional information on the model selection, the whole genome predictions, as well as the stand-alone prediction tool are available for download at http://www.fml.mpg.de/raetsch/projects/splice.

Show MeSH

Related in: MedlinePlus

Precision Recall Curve for the three methods MC, WD, WDS estimated on the genome-wide data sets for worm, fly, cress, fish, and human in a nested cross-validation scheme. In contrast to the ROC the random guess in this plot corresponds to a horizontal line, that depends on the fraction of positive examples in the test set (e.g. 2% and 3% in the case of the worm acceptor and donor data sets, respectively).
© Copyright Policy - open-access
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC2230508&req=5

Figure 3: Precision Recall Curve for the three methods MC, WD, WDS estimated on the genome-wide data sets for worm, fly, cress, fish, and human in a nested cross-validation scheme. In contrast to the ROC the random guess in this plot corresponds to a horizontal line, that depends on the fraction of positive examples in the test set (e.g. 2% and 3% in the case of the worm acceptor and donor data sets, respectively).

Mentions: Confirming our evaluations in the pilot studies, kernel methods outperform the MC methods in all eight classification tasks. Figure 3 displays the precision recall curves for all five organisms comparatively, Table 4 the corresponding auPRC scores. For worm, fly and cress the improvement in the performance accuracy for the SVM in comparison to MC lies in a similar range of 4–10% (absolute), both for donor and for acceptor tasks. However, for fish and especially for human the performance gain is considerable higher. For human, MCs only achieve 16% and 25% auPRC scores, whereas WDS reaches 54% and 57% for acceptor and donor recognition, respectively. The severe decrease in performance from worm to human for all classification methods in the auPRC score can partially be explained by the different fractions of positive examples observed in the test set. However, a weaker decline can also be observed in the auROC scores (also Table 4) which are independent of the class skew (e.g. for acceptor sites from 99.6% on worm to 96.0% on human for MC, and from 99.8% to 97.9% for WDS). The classification task on the human genome seems to be a considerably more difficult problem than the same one on the worm genome. We may speculate that this can be partially explained by a higher incidence of alternative splicing in the human genome. These sites usually exhibit weaker consensus sequences and are therefore more difficult to detect. Additionally, they often lead to mislabeled examples in the training and testing sets. Finally, it might also be due to the used protocol for aligning the sequences which may generate more false splice sites in human than in other organisms. This hypothesis is supported by the fact that the performance significantly increases, if one only considers cDNA confirmed genes (data not shown).


Accurate splice site prediction using support vector machines.

Sonnenburg S, Schweikert G, Philips P, Behr J, Rätsch G - BMC Bioinformatics (2007)

Precision Recall Curve for the three methods MC, WD, WDS estimated on the genome-wide data sets for worm, fly, cress, fish, and human in a nested cross-validation scheme. In contrast to the ROC the random guess in this plot corresponds to a horizontal line, that depends on the fraction of positive examples in the test set (e.g. 2% and 3% in the case of the worm acceptor and donor data sets, respectively).
© Copyright Policy - open-access
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC2230508&req=5

Figure 3: Precision Recall Curve for the three methods MC, WD, WDS estimated on the genome-wide data sets for worm, fly, cress, fish, and human in a nested cross-validation scheme. In contrast to the ROC the random guess in this plot corresponds to a horizontal line, that depends on the fraction of positive examples in the test set (e.g. 2% and 3% in the case of the worm acceptor and donor data sets, respectively).
Mentions: Confirming our evaluations in the pilot studies, kernel methods outperform the MC methods in all eight classification tasks. Figure 3 displays the precision recall curves for all five organisms comparatively, Table 4 the corresponding auPRC scores. For worm, fly and cress the improvement in the performance accuracy for the SVM in comparison to MC lies in a similar range of 4–10% (absolute), both for donor and for acceptor tasks. However, for fish and especially for human the performance gain is considerable higher. For human, MCs only achieve 16% and 25% auPRC scores, whereas WDS reaches 54% and 57% for acceptor and donor recognition, respectively. The severe decrease in performance from worm to human for all classification methods in the auPRC score can partially be explained by the different fractions of positive examples observed in the test set. However, a weaker decline can also be observed in the auROC scores (also Table 4) which are independent of the class skew (e.g. for acceptor sites from 99.6% on worm to 96.0% on human for MC, and from 99.8% to 97.9% for WDS). The classification task on the human genome seems to be a considerably more difficult problem than the same one on the worm genome. We may speculate that this can be partially explained by a higher incidence of alternative splicing in the human genome. These sites usually exhibit weaker consensus sequences and are therefore more difficult to detect. Additionally, they often lead to mislabeled examples in the training and testing sets. Finally, it might also be due to the used protocol for aligning the sequences which may generate more false splice sites in human than in other organisms. This hypothesis is supported by the fact that the performance significantly increases, if one only considers cDNA confirmed genes (data not shown).

Bottom Line: We apply our method to the genome-wide recognition of splice sites in Caenorhabditis elegans, Drosophila melanogaster, Arabidopsis thaliana, Danio rerio, and Homo sapiens.Our performance estimates indicate that splice sites can be recognized very accurately in these genomes and that our method outperforms many other methods including Markov Chains, GeneSplicer and SpliceMachine.We provide genome-wide predictions of splice sites and a stand-alone prediction tool ready to be used for incorporation in a gene finder.

View Article: PubMed Central - HTML - PubMed

Affiliation: Fraunhofer Institute FIRST, Kekuléstr, 7, 12489 Berlin, Germany. Soeren.Sonnenburg@first.fraunhofer.de

ABSTRACT

Background: For splice site recognition, one has to solve two classification problems: discriminating true from decoy splice sites for both acceptor and donor sites. Gene finding systems typically rely on Markov Chains to solve these tasks.

Results: In this work we consider Support Vector Machines for splice site recognition. We employ the so-called weighted degree kernel which turns out well suited for this task, as we will illustrate in several experiments where we compare its prediction accuracy with that of recently proposed systems. We apply our method to the genome-wide recognition of splice sites in Caenorhabditis elegans, Drosophila melanogaster, Arabidopsis thaliana, Danio rerio, and Homo sapiens. Our performance estimates indicate that splice sites can be recognized very accurately in these genomes and that our method outperforms many other methods including Markov Chains, GeneSplicer and SpliceMachine. We provide genome-wide predictions of splice sites and a stand-alone prediction tool ready to be used for incorporation in a gene finder.

Availability: Data, splits, additional information on the model selection, the whole genome predictions, as well as the stand-alone prediction tool are available for download at http://www.fml.mpg.de/raetsch/projects/splice.

Show MeSH
Related in: MedlinePlus