A novel diagnostic tool reveals mitochondrial pathology in human diseases and aging

www.mitodb.com

Using various sources such as the United Mitochondrial Disease Foundation webpage (http://www.umdf.org), Pubmed (http://www.ncbi.nlm.nih.gov/pubmed) and the Online Mendelian Inheritance in Man database (http://www.omim.org) we identified 31 monogenic diseases that all have been characterized as mitochondrial describing a total of 1,265 patients (Supplementary table 1). To get as broad a spectrum of mitochondrial phenotypes as possible we added respiratory chain defects, fatty acid oxidation defects and other primary mitochondrial diseases. 25 non-mitochondrial monogenic diseases, used as controls, describing a total of 21,032 patients, were selected based on well established pathogenesis with no or very minimal involvement of mitochondria. A database, www.mitodb.com, was constructed by writing ~8,000 lines of code in the languages html and php and using the aforementioned data sources, referenced signs, symptoms, laboratory and paraclinical findings (collectively referred to as clinical parameters or parameters) were recorded for each disease by clinically experienced physicians. The prevalence of each parameter was listed based on what was reported in the most authoritative literature we could identify. The publication date was as current as possible (Supplementary figure 1). For any particular disease we used the largest cohorts we could identify. Case reports were excluded. If we used data from two or more studies the prevalence was the average of the combined number of patients. All the references including their abstracts, publication date, periodical and links to Pubmed can be found online at www.mitodb.com. Each disease was thus characterized by a number of clinical parameters with a prevalence from 0-100 percent. See supplementary figure 2 or www.mitodb.com for visual examples of how the diseases are represented in the database. We will present a number of figures that can be generated online at www.mitodb.com. We also recommend reading our tutorial on the following URL: www.mitodb.com/tutorial.html.

Mitochondrial diseases form a distinct clinical group

Based on the data we could now analyze the clinical parameters in the mitochondrial group by averaging the occurrence of a parameter across all the diseases in each group. This allowed us to investigate whether any particular parameter would be more prevalent among the mitochondrial diseases than among the non-mitochondrial diseases. Figure 1A shows the top 20 most prevalent parameters and supplementary figure 3 and 4 shows all the parameters in the mitochondrial group and the non-mitochondrial group, respectively. Our analysis shows that, out of 117 clinical parameters seen in mitochondrial diseases, lactate accumulation, hypotonia, muscle weakness, developmental delay and seizures are among the most common. Since it has been suggested that the clinical parameters of mitochondrial diseases depend upon the age of onset of the disease we investigated whether this was the case for our dataset. We identified the mean age of onset and standard deviation for each disease in the database (See supplementary table 1). Using this data we could then model the age of onset of each clinical parameter letting us identify the most common symptoms for diseases with onset before age 20 and most common symptoms for diseases with onset after age 20 (Figure 1B and C) and their evolution over time (Supplementary figure 5). Lactate accumulation, hypotonia, developmental delay, muscle weakness and seizures were the most prevalent parameters in the early onset diseases while hearing loss, muscle weakness, ophthalmoplegia, seizures and lactate accumulation were the most prevalent in the late onset diseases. Some differences in clinical parameters were thus found when comparing early onset to late onset disease. However, quantifying these differences as a Pearson coefficient resulted in a value of 0.5 indicating an overall similar phenotype in early and late onset diseases. In addition, there was no published data for the individual parameter onsets for every disease preventing us from using age of onset for further analysis. We therefore decided to continue our analysis without including age of onset.

By overlaying the prevalences of all the mitochondrial and non-mitochondrial parameters it became evident that some mitochondrial clinical parameters occurred at a high frequency indicating a high degree of similarity between the diseases in the mitochondrial group while there was much more heterogeneity amongst the clinical features in the non-mitochondrial disease group (Supplementary figure 6A). Based on this resemblance among the mitochondrial diseases, we investigated how these diseases would associate with each other when clustering them using the software cluster 3.0 [18]. Supplementary figure 6B shows an example of how some mitochondrial and some non-mitochondrial diseases cluster with each other based on clinical parameters. The clustering of the diseases is shown in the top, while the clinical parameters are shown on the right with the clustering of the parameters on the left. The tint of each square in the center of the cluster map reflects the prevalence of a parameter in the disease based on published literature. Interestingly, there seemed to be excellent separation between the mitochondrial and non-mitochondrial diseases. Supplementary figure 6C shows an overview of the cluster while a detailed look can be seen online at www.mitodb.com. In summary, we were able to use cluster analysis to compare the diseases in the database.

The qualitative determination of a mitochondrial phenotype using mitodb.com

With this established database we proceeded to develop tools to test diseases for potential mitochondrial involvement. We started by implementing hierarchical clustering algorithms using the coding languages html, javascript and php. The phenotype of a disease is entered under the “Test-disease” tab on mitodb.com by adding the prevalence of all the observed parameters. After the analysis has been completed a circular graphic illustrating (dendrogram) how the diseases cluster with each other will be generated. The connecting lines (branches of the tree) illustrate how similar two diseases are. As can be seen in figure 2A there is substantial separation of mitochondrial (red) and non-mitochondrial (green) diseases using this approach. To further visualize how the clinical parameters are interconnected we generated a disease network tool (Figure 2B) showing the association between diseases. Each dot represents a disease and each line represents a shared parameter, with the distance between the dots and the thickness of the line representing the degree of similarity (See supplementary equations). Interestingly, a central tightly associated hub of mitochondrial diseases formed with the non-mitochondrial diseases loosely associated in the periphery (Figure 2B). This is caused by the lack of overlapping parameters seen in the non-mitochondrial group compared with the highly overlapping parameters seen in the mitochondrial. Due to the substantial complexity we recommend investigat-ing this network online on www.mitodb.com, where different thresholds can be applied to selectively visualize the most common connections. To better depict the correlation between diseases we developed a mitochondrial barcode-algorithm that shows which parameters are found in other diseases in the database (Figure 2C). Each vertical colored bar represents a disease that shares a clinical parameter with the disease being tested. The tint is given by the prevalence of the parameter in the disease tested, multiplied by the prevalence in the disease in the database that shares the parameter. Thus, the barcode appears more red with increasing mitochondrial features and more green or blue with fewer or no mitochondrial features. The list of shared clinical parameters for a tested disease can be found online underneath the barcode.

The quantitative determination of a mitochondrial phenotype using mitodb.com

To get more quantitative results we developed a tool assigning a mito-score to a disease from 0-100 with a score above 50 indicating a larger overlap with mitochondrial diseases than non-mitochondrial diseases (See supplementary equations). This method gave consistent results with mitochondrial diseases scoring high and non-mitochondrial scoring low (Figure 2D).

We next developed a machine learning algorithm to separate diseases based on clinical parameters and to get a quantitative score. Briefly, we chose a model of supervised learning, a support vector machine (SVM), that is ideal for investigating complicated datasets. This type of machine learning has been used to analyze various datasets including domain mapping of amino acid sequences in proteins and facial recognition in pictures on the web. In this method, the clinical pattern of the mitochondrial diseases and the pattern of the non-mitochondrial diseases are fed to a computer that will learn from these training examples and subsequently predict whether a tested disease may show mitochondrial involvement. A negative score reflects a likely non-mitochondrial disease while a positive score indicates a mitochondrial disease (See methods). We implemented several variations of the SVM, one unfiltered taking all parameters in the database into account and various filtered ones using only parameters displaying a cumulative prevalence of 25%, 50%, 75%, 100% or 200%. A cumulative prevalence of 25% is defined as one disease with a 25% prevalence of a parameter or five diseases with a 5% prevalence each. Thus, with increasing percentages progressively more common clinical parameters will be the only ones considered by the SVM. Overall, the performance of the various classifiers was excellent with 100% separation and a correlation coefficient (measured by cross-validation) around 0.85, with the 75% filtered version yielding a marginally better correlation coefficient of 0.87 (See supplementary equations). The great performance of all the SVM's show that the method is robust, and that the prevalence of the clinical parameters can be used to indicate whether a disease show mitochondrial involvement. To keep the data as unbiased as possible we chose the unfiltered version as our standard SVM algorithm (Figure 2E, and supplementary equations).

The bioinformatics tools reveal mitochondrial dysfunction in ataxia-telangiectasia, Cockayne syndrome and others

Recent studies that have shown mitochondrial dysfunction in ataxia-telangiectasia [17] , AOA1 [16], SCAN1 [15] and Cockayne syndrome [14, 20]. Interestingly, these diseases clustered tightly with the mitochondrial diseases, grouped with mitochondrial disease in the disease network and displayed a relatively high SVM and mito-score (Figure 2A-E). This indicated that our tools could be used to investigate mitochondrial dysfunction in diseases of unknown pathogenesis simply by comparing the clinical parameters with the parameters in our database.

Aging and some progerias display mitochondrial dysfunction

Many diseases with unknown pathogenesis were strongly associated with mitochondrial pathology (Figure 3A and supplementary table 2). Because a close correlation between aging and mitochondrial dysfunction has been proposed, we next tested whether diseases characterized by accelerated aging shared clinical parameters with the mitochondrial diseases in the database. Of all the segmental progerias tested, Cockayne syndrome and ataxia-telangiectasia correlated tightly with mitochondrial pathology in the dendrogram, while Werner syndrome, Bloom syndrome, Rothmund-Thomson syndrome, trichothiodystrophy, dyskeratosis congenita, Fanconi anemia, Nijmegen breakage syndrome and Hutchinson-Gilford progeria formed an independent cluster (Figure 3A). This was corroborated by the finding in the disease network that Cockayne syndrome and ataxia-telangiectasia appeared to lie deeper in the mitochondrial hub compared to the other progerias (Figure 3B). When looking at the barcode for Cockayne syndrome and ataxia-telangiectasia, these diseases also showed a predominantly red impression while the other progerias where less red (Figure 2C and Figure 3C). The SVM and mito-score supported all these results (Figure 2D and E and figure 3D and E). The association between the mitochondrially associated progerias, Cockayne syndrome and ataxia-telangiectasia, was related to the high degree of neurological involvement in these diseases, since removing the neuronal parameters led these diseases to cluster with the other progerias (Supplementary figure 7).

We finally investigated if normal human aging was associated with mitochondrial pathology. A number of important papers were found to describe the phenotype of normal human aging as accurately as possible (Figure 4) [21-30]. By entering the parameters of aging an interesting picture emerged. Normal aging seemed to cluster with the mitochondrial diseases where the mitochondrial segmental progerias also clustered indicating a potential mitochondrial phenotype (Figure 3A-C). Reflecting this, the barcode showed a mitochondrial imprint that was corroborated by a high mito-score of 70. However, the SVM was inconclusive reflected by a score of around 0. Interestingly, using the 200% filtered SVM, aging received a score of 1.46, indicating a probable mitochondrial impact when considering only very common clinical parameters. It therefore seems likely that some features of aging, eg. cancers, may be independent of mitochondria while others, eg. neurological features, are closely correlated with mitochondrial health.

Hierarchical clustering

Hierarchical clustering was implemented using common similarity measures and linkage methods (See supplementary equations) [18, 39]. For each clustering the cophenetic correlation is calculated as a measure of how well the clustering was achieved revealing uncentered similarity and average linkage producing the best result (see supplementary figure 9 and supplementary equations). Other similarity metrics and linkage methods can be chosen online by clicking “advanced” under the “test-disease” tab on www.mitodb.com.

Support vector machine (SVM)

The support vector classifier was created by exporting the symptom-vectors of the known mitochondrial and non-mitochondrial diseases. The symptom-vectors were processed with the statistical program “R” using the “e1071” package (http://cran.r-project.org/web/packages/e1071/index.html), which is based on the library “libSVM” to train the SVMs. The SVMs were trained using a “linear kernel” and “eps-regression”, and displays (using 11-fold cross-correlation) a total mean squared error of 0.32, a 0.84 correlation coefficient, and 100% separation. The parameters of the SVMs were then exported from R and the classifier functions for the web-page was created (with php) using these parameters.

Further methods can be found in the supporting materials.