Limits...
Regression based predictor for p53 transactivation.

Gowrisankar S, Jegga AG - BMC Bioinformatics (2009)

Bottom Line: The extent of this regulation depends in part on the binding affinity of p53 to its response elements (REs).Experimentally validated known p53-REs along with their transactivation capabilities are used for training.Taking into account both nucleotide interactions and the spacer length of p53-RE, we have created a novel in-silico regression-based transactivation capability predictor for p53-REs and used it to analyze validated and novel p53-REs and to predict the impact of SNPs overlapping these elements.

View Article: PubMed Central - HTML - PubMed

Affiliation: Division of Biomedical Informatics, Cincinnati Children's Hospital Medical Center, Cincinnati, OH, USA. sgowrisankar@partners.org

ABSTRACT

Background: The p53 protein is a master regulator that controls the transcription of many genes in various pathways in response to a variety of stress signals. The extent of this regulation depends in part on the binding affinity of p53 to its response elements (REs). Traditional profile scores for p53 based on position weight matrices (PWM) are only a weak indicator of binding affinity because the level of binding also depends on various other factors such as interaction between the nucleotides and, in case of p53-REs, the extent of the spacer between the dimers.

Results: In the current study we introduce a novel in-silico predictor for p53-RE transactivation capability based on a combination of multidimensional scaling and multinomial logistic regression. Experimentally validated known p53-REs along with their transactivation capabilities are used for training. Through cross-validation studies we show that our method outperforms other existing methods. To demonstrate the utility of this method we (a) rank putative p53-REs of target genes and target microRNAs based on the predicted transactivation capability and (b) study the implication of polymorphisms overlapping p53-RE on its transactivation capability.

Conclusion: Taking into account both nucleotide interactions and the spacer length of p53-RE, we have created a novel in-silico regression-based transactivation capability predictor for p53-REs and used it to analyze validated and novel p53-REs and to predict the impact of SNPs overlapping these elements.

Show MeSH

Related in: MedlinePlus

Leave-one-out cross-validation results showing a straight line between the actual and observed transactivation capabilities. The average predicted values for lower levels of transactivation do not exactly follow the observed levels.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC2719629&req=5

Figure 2: Leave-one-out cross-validation results showing a straight line between the actual and observed transactivation capabilities. The average predicted values for lower levels of transactivation do not exactly follow the observed levels.

Mentions: We used ten-fold and leave-one-out cross validations to test the performance and usability of our model. Pearson correlation coefficients were calculated between observed and predicted transactivation capabilities. For ten-fold cross-validation we obtained correlations of 0.71 and 0.73 (0.71 ± 0.06 and 0.73 ± 0.05 respectively if correlation is calculated for each fold separately) for models without and with spacers, respectively. In the case of leave-one-out cross-validation, we obtained correlations of 0.71 and 0.70 for models without and with spacers, respectively. We were unable to find correlation for each fold separately as each has only one test case in leave-one-out cross-validation. Surprisingly, we did not observe a significant difference between training with and without spacers. This could probably be because the training data spacer distribution is highly skewed toward the lower values. In other words, only 12 of the 263 p53-REs had a spacer of length 8 bp or higher. Nevertheless, we noted some improvement in the performance (ten-fold cross-validation) when spacers was used as a feature, although it is not statistically significant. To test whether the correlation results are skewed toward a specific transactivation capability value, we obtained the average predicted capability for each level of true capability. Figure 2 shows the "predicted" and "observed" transactivation capabilities for leave-one-out cross-validation. Both the models – without and with spacers – performed similarly. However, toward the lower levels of observed capability, we noticed a slight increase in the average predicted capability levels, though not statistically significant. This was especially apparent for levels 0 and 1, which correspond to "Non-responsive" and "Poor" transactivation capabilities, respectively. Both models performed well in predicting the higher capability values.


Regression based predictor for p53 transactivation.

Gowrisankar S, Jegga AG - BMC Bioinformatics (2009)

Leave-one-out cross-validation results showing a straight line between the actual and observed transactivation capabilities. The average predicted values for lower levels of transactivation do not exactly follow the observed levels.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC2719629&req=5

Figure 2: Leave-one-out cross-validation results showing a straight line between the actual and observed transactivation capabilities. The average predicted values for lower levels of transactivation do not exactly follow the observed levels.
Mentions: We used ten-fold and leave-one-out cross validations to test the performance and usability of our model. Pearson correlation coefficients were calculated between observed and predicted transactivation capabilities. For ten-fold cross-validation we obtained correlations of 0.71 and 0.73 (0.71 ± 0.06 and 0.73 ± 0.05 respectively if correlation is calculated for each fold separately) for models without and with spacers, respectively. In the case of leave-one-out cross-validation, we obtained correlations of 0.71 and 0.70 for models without and with spacers, respectively. We were unable to find correlation for each fold separately as each has only one test case in leave-one-out cross-validation. Surprisingly, we did not observe a significant difference between training with and without spacers. This could probably be because the training data spacer distribution is highly skewed toward the lower values. In other words, only 12 of the 263 p53-REs had a spacer of length 8 bp or higher. Nevertheless, we noted some improvement in the performance (ten-fold cross-validation) when spacers was used as a feature, although it is not statistically significant. To test whether the correlation results are skewed toward a specific transactivation capability value, we obtained the average predicted capability for each level of true capability. Figure 2 shows the "predicted" and "observed" transactivation capabilities for leave-one-out cross-validation. Both the models – without and with spacers – performed similarly. However, toward the lower levels of observed capability, we noticed a slight increase in the average predicted capability levels, though not statistically significant. This was especially apparent for levels 0 and 1, which correspond to "Non-responsive" and "Poor" transactivation capabilities, respectively. Both models performed well in predicting the higher capability values.

Bottom Line: The extent of this regulation depends in part on the binding affinity of p53 to its response elements (REs).Experimentally validated known p53-REs along with their transactivation capabilities are used for training.Taking into account both nucleotide interactions and the spacer length of p53-RE, we have created a novel in-silico regression-based transactivation capability predictor for p53-REs and used it to analyze validated and novel p53-REs and to predict the impact of SNPs overlapping these elements.

View Article: PubMed Central - HTML - PubMed

Affiliation: Division of Biomedical Informatics, Cincinnati Children's Hospital Medical Center, Cincinnati, OH, USA. sgowrisankar@partners.org

ABSTRACT

Background: The p53 protein is a master regulator that controls the transcription of many genes in various pathways in response to a variety of stress signals. The extent of this regulation depends in part on the binding affinity of p53 to its response elements (REs). Traditional profile scores for p53 based on position weight matrices (PWM) are only a weak indicator of binding affinity because the level of binding also depends on various other factors such as interaction between the nucleotides and, in case of p53-REs, the extent of the spacer between the dimers.

Results: In the current study we introduce a novel in-silico predictor for p53-RE transactivation capability based on a combination of multidimensional scaling and multinomial logistic regression. Experimentally validated known p53-REs along with their transactivation capabilities are used for training. Through cross-validation studies we show that our method outperforms other existing methods. To demonstrate the utility of this method we (a) rank putative p53-REs of target genes and target microRNAs based on the predicted transactivation capability and (b) study the implication of polymorphisms overlapping p53-RE on its transactivation capability.

Conclusion: Taking into account both nucleotide interactions and the spacer length of p53-RE, we have created a novel in-silico regression-based transactivation capability predictor for p53-REs and used it to analyze validated and novel p53-REs and to predict the impact of SNPs overlapping these elements.

Show MeSH
Related in: MedlinePlus