Skip to main page content
U.S. flag

An official website of the United States government

Dot gov

The .gov means it’s official.
Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

Https

The site is secure.
The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
. 2013;9(11):e1003314.
doi: 10.1371/journal.pcbi.1003314. Epub 2013 Nov 7.

Systematically differentiating functions for alternatively spliced isoforms through integrating RNA-seq data

Ridvan Eksi  1 , Hong-Dong Li, Rajasree Menon, Yuchen Wen, Gilbert S Omenn, Matthias Kretzler, Yuanfang Guan
Affiliations

Systematically differentiating functions for alternatively spliced isoforms through integrating RNA-seq data

Ridvan Eksi et al. PLoS Comput Biol. 2013.

Abstract

Integrating large-scale functional genomic data has significantly accelerated our understanding of gene functions. However, no algorithm has been developed to differentiate functions for isoforms of the same gene using high-throughput genomic data. This is because standard supervised learning requires 'ground-truth' functional annotations, which are lacking at the isoform level. To address this challenge, we developed a generic framework that interrogates public RNA-seq data at the transcript level to differentiate functions for alternatively spliced isoforms. For a specific function, our algorithm identifies the 'responsible' isoform(s) of a gene and generates classifying models at the isoform level instead of at the gene level. Through cross-validation, we demonstrated that our algorithm is effective in assigning functions to genes, especially the ones with multiple isoforms, and robust to gene expression levels and removal of homologous gene pairs. We identified genes in the mouse whose isoforms are predicted to have disparate functionalities and experimentally validated the 'responsible' isoforms using data from mammary tissue. With protein structure modeling and experimental evidence, we further validated the predicted isoform functional differences for the genes Cdkn2a and Anxa6. Our generic framework is the first to predict and differentiate functions for alternatively spliced isoforms, instead of genes, using genomic data. It is extendable to any base machine learner and other species with alternatively spliced isoforms, and shifts the current gene-centered function prediction to isoform-level predictions.

PubMed Disclaimer

Conflict of interest statement

The authors have declared that no competing interests exist.

Figures

Figure 1
Figure 1. Overview of the computational approach for predicting functions for alternatively spliced isoforms.
We collected RNA-seq data from the sequence read archive (SRA) database and estimated isoform-level expression values using state-of-the-art software , . We then generated a gene-level gold standard using Gene Ontology (GO) annotations. For each biological function, this gold standard contains positive genes (annotated to the function under investigation) and negative genes (other genes). Our study contains two major parts: cross-validation for performance estimation and bootstrap bagging for generating final predictions as well as performance evaluation. For cross-validation, we partitioned the examples into a training set for model development and a test set for model validation. For generating final predictions for all isoforms, we sampled with replacement to construct a training set, and then used this training set to construct models to assign prediction probabilities to the out-of-bag set. The final predictions for all isoforms were made by calculating the median prediction values of all out-of-bag sets. For each training set, a model was derived from the RNA-seq data to delineate the positives and the negatives. This model was used to classify the training set and update the labels of the isoforms of the positive genes, under the criterion that at least one isoform of a positive gene must remain positive. This new assignment is then used to construct the model in the next iteration. This process is iterated until the assignment of positive isoforms no longer changes, and then the final model was used to assign a prediction value to the test or the out-of-the-bag set. Bootstrap was done for 30 iterations and the median value for each out-of-the-bag isoform was taken as the final prediction value. The predictive performance of our model was assessed through three approaches: (1) cross-validation of gene-level predictive performances, focusing on comparison between single-isoform genes and multiple-isoform genes, (2) literature validation and (3) experimental validation of top predictions using proteomic data.
Figure 2
Figure 2. Performance comparison of different formulations of the SVM-MIL algorithm in predicting isoform functions.
A. The histogram shows the score distribution of the instances in the positive bags and the negative bags in the training set. Different threshold choices in mi-SVM are based on the distribution of scores of negative genes. The first threshold is equal to the mode of distribution of scores from negative instances in the training set. The second threshold is equal to the 75% percentile of scores of the negative instances in the training set. The third threshold is equal to the maximum score of negative instances in the training set. B. This panel illustrates how different thresholds and formulations can divide the isoforms in a positive bag into positive, negative and neutral classes. Three thresholds in mi-SVM represent different degrees of strictness for assigning labels. The first threshold is the least strict, which assigns most of the isoforms from positive genes as positive, whereas the third threshold is the strictest, which in general leaves only one positive instance in every positive bag. For the MI-SVM formulation, only one isoform per positive gene is assigned as positive, and other isoforms are dropped (i.e. neutral class). C. Performance comparison of three different threshold choices for the mi-SVM formulation, the MI-SVM formulation and the MI-SVM formulation with random witness selection. This plot shows that the mi-SVM formulation with threshold-2 performs best in terms of AUC.
Figure 3
Figure 3. Robust performance of our algorithm to predicting functions using RNA-seq data.
We carried out five-fold cross validation to test the performance of our algorithm. For each function, the prediction value for each gene is assigned the maximum prediction value of all of its isoforms, under the assumption that at least one of its isoforms should carry out the function. Because the number of known genes of each GO term systematically affects the prediction performance, we group these terms into 5 groups according to their GO term sizes. (A)–(D) shows the distribution (10, 25, 50, 75, 90%) of the AUCs, the AUPRCs, the precisions at 1% recall and the precisions at 10% recall, respectively.
Figure 4
Figure 4. Prediction performance comparison of single-isoform genes (green) and multi-isoform gene (blue) based on AUC (upper panel) and AUPRC (lower panel).
We separately evaluated its prediction performance for single-isoform genes and multiple-isoform genes. Two-fold cross-validation was carried out to ensure enough examples in both groups. To ensure comparability, the negatives were randomly selected to ensure that the ratios of positive to negative genes for the multi-isoform group and the single isoform group are the same for each GO term. GO terms were grouped according to the number of genes in the test set. Shown in the box-plot are the AUC (A) and AUPRC (B) at 10, 25, 50, 75 and 90 percentile, respectively.
Figure 5
Figure 5. Prediction precision between single-isoform genes (green) with multi-isoform gene (blue).
Two-fold cross-validation was carried out to ensure that enough examples are included in both the single-isoform group and the multi-isoform group. The negatives were randomly selected to ensure that the ratios of positive to negative genes for the multi-isoform group and the single isoform group are the same for each GO term, so that the baseline precision for each GO term is equal for the two groups. GO terms were grouped according to the number of genes in the test set. Each dot represents the precision value of an individual GO term. A. Precision at one percent recall. B. Precision at ten percent recall.
Figure 6
Figure 6. Robust performance of our algorithm in predicting isoform functions.
A. Genes are grouped according to their expression levels averaged across all samples in our RNA-seq data collection. The distribution of the performance in AUC across all GO terms is plotted using box-plot. B. The performance in AUC across all GO terms by partitioning the genes according to homologous groups between the training and the test set is compared against the performance of partitioning the genes randomly.
Figure 7
Figure 7. Predicted functions for isoforms of CDKN2a and their predicted protein structures. A
. Gene model for NM_001040654.1 and NM_009877.2. B. Predicted functions for NM_001040654.1 and NM_009877.2. C. The computationally modeled structure of NM_001040654.1 is characterized by five ankryin repeats. D. The modeled structure of NM_009877.2 has a CDKN2a N-terminus domain.
See this image and copyright information in PMC

References

    1. Schmitz R, Young RM, Ceribelli M, Jhavar S, Xiao WM, et al. (2012) Burkitt lymphoma pathogenesis and therapeutic targets from structural and functional genomics. Nature 490: 116–120. - PMC - PubMed
    1. Guan Y, Ackert-Bicknell CL, Kell B, Troyanskaya OG, Hibbs MA (2010) Functional genomics complements quantitative genetics in identifying disease-gene associations. PLoS Comput Biol 6: e1000991. - PMC - PubMed
    1. Chen KF, Crowther DC (2012) Functional genomics in Drosophila models of human disease. Briefings in Functional Genomics 11: 405–415. - PubMed
    1. Liang H, Cheung LWT, Li J, Ju ZL, Yu SX, et al. (2012) Whole-exome sequencing combined with functional genomics reveals novel candidate driver cancer genes in endometrial cancer. Genome Research 22: 2120–2129. - PMC - PubMed
    1. Nelson AC, Pillay N, Henderson S, Presneau N, Tirabosco R, et al. (2012) An integrated functional genomics approach identifies the regulatory network directed by brachyury (T) in chordoma. Journal of Pathology 228: 274–285. - PMC - PubMed

Publication types

LinkOut - more resources