Defining the progeria phenome

Identification of progeroid syndromes

The syndrome database on https://www.mitodb.com was used in this study as a source of data concerning known progeroid syndromes and other groups of syndromes. To first identify premature aging diseases previously unrecognized by the mitodb database, putative progeroid diseases in the Online Mendelian Inheritance in Man (OMIM) were identified using the search strategy ‘Progeroid’, ‘Progeria’ and ‘Premature aging’. Hits were sorted based on a predefined inclusion and exclusion criteria allowing the identification of 32 premature aging syndromes (Figure 1 and Table 1). For each disease, published papers on PubMed were identified and prevalences of phenotypes in each disease were added to the database (see Supplementary Figures 1, 2 and Methods “Mean Prevalence of Phenotypes”). This quantified description of each disease allowed to compare and cluster diseases and phenotypes from all known progerias similar to what we have previously done with mitochondrial disorders [7].

Figure 1. Identification of progeroid syndromes for database. Flow diagram illustrating the process of identifying new progeroid syndromes for mitodb.com including search, inclusion and exclusion criteria.

Hierarchical clustering identifies five phenotypical groups of progerias

To understand the overlap and differences in clinical phenotypes between these rare diseases, principal component analysis, hierarchical clustering and networking tools were used on all 32 progeroid syndromes in the database. We also included normal human aging that consists of the human phenome as previously defined by us [8]. The hierarchical clustering (Figure 2A) showed progeroid syndromes clustered in groups indicating phenotypic diversity: one group contained syndromes such as Cockayne syndrome, Xeroderma Pigmentosum A (XPA) and Ataxia-Telangiectasia sharing phenotypes like cerebellar atrophy and short stature, a second group contained syndromes such as Werner syndrome, Rujis-Aalfs syndrome and Hutchinson Gilford sharing phenotypes like micrognathia and short stature, a third group contained Seckel syndrome and Nijmegen breakage sharing phenotypes like microcephaly, developmental delay and prominent nose, a fourth group contained Bloom syndrome and Xeroderma Pigmentosum (B, C, F, G, V and E) sharing phenotypes like cancer and sun sensitivity and a fifth group contained syndromes such as KPLB syndrome, GAPO syndrome and SHORT syndrome sharing phenotypes like micrognathia and skin wrinkles. Brachiooculofacial syndrome (BOFS), Geroderma osteodysplasticum (GO) and Acromicric dysplasia deviated from the other syndromes indicating a substantial difference in phenotypic traits and perhaps a weaker progeria phenotype.

Figure 2. Progeroid syndrome overview. (A) Agglomerative hierarchical cluster based on phenotype prevalences using uncentered similarity and average linkage. (dark green are new to the database, other colors represent different clusters). Each syndrome group is color coded in the inner circle. (B) Principal component analysis of diseases based on the prevalence of phenotypes. (C) Hierarchical clustering of publicly available dataset for some premature aging diseases and the shared pathways between closely associated diseases. (D) Support vector machine scores for premature aging diseases (available at https://www.mitodb.com).

To identify potential outliers in our dataset we used Principal Component Analysis (PCA) of the quantified traits (Figure 2B). Surprisingly, the PCA did not result in a strong separation between diseases, indicating that there are no clear phenotypic differences in the dataset. While the analysis indicated a considerable variation in the clinical presentation, it did not provide evidence to exclude any syndrome from our dataset.

Transcriptomics reveals altered pathways in premature aging diseases

To understand if the clinical phenotypes could be reflected in molecular changes, we explored gene expression data from premature aging diseases. We were able to identify published gene expression data from 13 progeroid syndromes allowing us to investigate the similarity between diseases (Figure 2C). Interestingly, we observed that Cockayne syndrome, Xeroderma Pigmentosum group A (XPA), and Ataxia-Telangiectasia, three diseases associated with premature neurological aging, formed a close cluster. Pathway analysis revealed that DNA damage and multiple inflammatory processes were common between these diseases (Supplementary Table 1). For Werner syndrome and Hutchinson Gilford Progeria Syndrome pathway analysis showed perturbation of growth signals and the DNA damage response (Supplementary Table 2). Further, Bloom syndrome was found to cluster with Nijmegen breakage syndrome supporting their biochemical role in double stranded DNA break repair and shared pathways were growth signaling, DNA damage response pathways, transcription and inflammation (Supplementary Table 3). Lastly, Seckel syndrome clustered together with Nestor-Guillermo progeria and dyskeratosis congenita perhaps suggesting issue with DNA replication and cell division leads to similar transcriptional outcomes. However, almost no pathways were found to overlap between Nestor-Guillermo progeria and Seckel syndrome (Supplementary Table 4).

A support vector machine accurately classifies progeroid diseases

A major hurdle for rare diseases is the possibility of fast diagnosis. We therefore generated a support vector machine (SVM) classifier that can identify premature aging diseases based on their phenotype alone. The SVM was trained to recognize progeroid syndromes by training it on the prevalence of phenotypes associated with the 32 progerias and to a control group of 29 non-progeroid syndromes from the database, which showed the least correlation with progeroid syndromes when using hierarchical clustering. The control syndromes are all non-mitochondrial syndromes with known pathogenesis. This yielded good results and the SVM accurately separates both progeroid and non-progeroid diseases in the database (Figure 2D). Notably, the SVM is available online at https://www.mitodb.com where inputted phenotypes will automatically receive an SVM score.

Mitochondrial diseases show minor overlap to progeroid syndromes

Mitochondrial alterations are known to occur with age and have been proposed to be a hallmark feature of aging. A connection has previously been established between mitochondrial syndromes and certain premature aging diseases such as Cockayne syndrome, xeroderma-pigmentosum and Ataxia-Telangiectasia [7, 9]. Therefore, we performed hierarchical clustering with the progeroid diseases and mitochondrial disorders (Figure 3A). Cockayne syndrome, xeroderma pigmentosum and Ataxia-Telangiectasia clustered with mitochondrial syndromes as expected. Notably, Rahman syndrome also clustered with mitochondrial syndromes suggesting phenotypic overlap. However, all other progeroid syndromes appeared to cluster exclusively with other progeroid syndromes.

Figure 3. Hierarchical clusters and syndrome networks. Agglomerative hierarchical clustering and network algorithms of mitochondrial syndromes (A, B), DNA-repair syndromes (C, D) and syndromes with abnormal autophagy (E, F).

To further investigate the link between mitochondrial and progeroid syndromes, we created a network to explore the connection between diseases. Each dot is a disease and a connecting line between dots represents a shared phenotype. The shorter the distance the more phenotypes are shared. This showed progeroid syndromes group in one end of the network indicating that they share significantly more traits among themselves even though they connect with mitochondrial syndromes (Figure 3B). Cockayne syndrome, Ataxia-Telangiectasia, Rahman syndrome and XPA were found scattered between the mitochondrial syndromes confirming their correlation as seen in the hierarchical cluster.

DNA-repair syndromes show strong overlap with progerias

The progeroid syndromes in the data set are caused by mutations in single genes often associated with defect genome maintenance resulting in acceleration of some features in aging. Hutchinson-Gilford progeria, Cockayne syndrome and Werner syndrome are progeroid syndromes known to be caused by altered DNA-repair [10]. Creating a hierarchical cluster and a network for progeroid syndromes and DNA-repair syndromes allowed further analysis of their correlation. To investigate this, we clustered the premature aging diseases with diseases known for defects in DNA repair. As opposed to syndromes with mitochondrial pathogenesis, the DNA-repair syndromes were seen scattered between progeroid syndromes within the hierarchical cluster (Figure 3C). For instance, epileptic encephalopathy, childhood-onset (EEOC) and ataxia-telangiectasia like disorder 2 (ATLD2) clustered with Cockayne syndrome and XPA, Warsaw breakage syndrome (WABS) clustered with Seckel and Nijmegen breakage syndrome and Meier-Gorlin syndrome (MGORS) clustered with acromicric dysplasia.

Additionally, we created a network to compare with the hierarchical cluster. The network (Figure 3D) illustrated how certain progeroid syndromes such as SHORT syndrome were grouped in one end, suggesting that they share more traits among themselves than they do with other syndromes in the network. However, multiple DNA-repair syndromes such as ATLD2, neurodegeneration, childhood-onset, stress-induced, with variable ataxia and seizures (CONDSIAS) and MGORS were placed close to this group of progeroid syndromes indicating a large amount of shared traits and confirming our results from the hierarchical clustering.

Syndromes characterized by abnormal autophagy show minor overlap with progerias

In healthy humans, the basal activity of autophagy in living cells decreases with age giving rise to accumulation of damaged cells. Notably, defective autophagy has recently been added as a hallmark of human aging [11]. To investigate the possible correlation between progeroid syndromes and autophagy we build a hierarchical cluster and a network comparing progeroid syndromes and syndromes displaying abnormal autophagy (Figure 3E, 3F). Surprisingly, the hierarchical clustering showed little clustering between progeroid syndromes and syndromes characterized by abnormal autophagy. Similarly, the hierarchical cluster with mitochondrial syndromes, Cockayne syndrome, XPA, Rahman syndrome and Ataxia Telangiectasia appeared to correlate phenotypically with syndromes presenting abnormal autophagy while the rest of the progeroid syndromes stayed clustered to each other. Notably defective mitophagy, a mitochondrial specific macro-autophagy pathway, has been shown for some premature aging disease [12]. The network complemented this result, showing progeroid syndromes distanced from the syndromes with abnormal autophagy. However, the syndromes spastic paraplegia 49 (SPG49), Vici syndrome and Zellweger syndrome were located close to the group of progeroid syndromes in the network making them potential candidates for further investigation.

The premature aging phenome

Finally, we calculated the mean prevalence of phenotypes in each group of syndromes. Short stature, micrognathia, sun sensitivity, alopecia, skin pigmentation changes and microcephaly were discovered as being the most common phenotypes in progeroid syndromes (Figure 4). This showed considerable overlap with diseases with abnormal DNA-repair where short stature, microcephaly, ataxia, cerebellar atrophy and developmental delay were the most common phenotypes. In mitochondrial syndromes lactate accumulation, hypotonia, developmental delay, muscle weakness, seizures and ataxia were the most common phenotypes [7]. Lastly, developmental delay, cerebellar and cerebral atrophy, ataxia and dysarthria were the most common phenotypes in syndromes with abnormal autophagy perhaps illustrating that defective authophagy can lead to mitochondrial dysfunction. In conclusion, progeroid syndromes therefore seem to share most phenotypes with DNA-repair syndromes as shown previously in our hierarchical cluster and networks.

Figure 4. Mean prevalence of phenotypes. (A) Mitochondrial syndromes, (B) DNA-repair syndromes, (C) syndromes with abnormal autophagy and (D) progeroid syndromes.

Identification of possible progeroid syndromes

In order to find potential progeroid syndromes amongst the mitochondrial syndromes, DNA-repair syndromes and syndromes with abnormal autophagy, we applied a high threshold when working with networks (see Methods). The majority of these syndromes were not associated with progeroid syndromes when the threshold was raised and were mainly connected to others of the same classification. However, three syndromes were more strongly linked to progeroid syndromes in their respective networks, suggesting a correlation of phenotypic traits. They included ATLD2 (Ataxia Telangiectasia like Disorder 2) and MGORS (Meier-Gorlin syndrome) that are both DNA-repair syndromes and SPG49 (Spastic Paraplegia 49), a syndrome displaying abnormal autophagy.

To test their potential to be progeroid, we incorporated them in a network with all progeroid syndromes (Figure 5A–5C) with the aim of investigating how the addition of the syndromes would alter the network. Firstly, we tested our network exclusively containing progeroid syndromes by raising the threshold, which led to the syndromes GO, XPC, XPF and XPV to disconnect from the rest. After adding the three syndromes ATLD2, SPG49 and MGORS the same four progeroid syndromes disconnected from the network, indicating a strong connection between the three tested syndromes and the main cluster of progeroid syndromes. Furthermore, when looking at the position of syndromes within the networks there seemed to be few changes. All tested syndromes were located near the center group of progeroid syndromes. This indicated that the incorporated syndromes have similar phenotypic profiles and do not result in change in the correlation between known progeroid syndromes.

Figure 5. Possible progeroid syndromes. (A) A network containing progeroid syndromes and MGORS, (B) network containing progeroid syndromes and SPG49, (C) network containing progeroid syndromes and ATDL2 and (D) a hierarchical cluster with progeroid syndromes, MGORS, ATLD2 and SPG49.

To further investigate this hypothesis, we created a hierarchical cluster containing the three tested syndromes and all progeroid syndromes known to the database (Figure 5D). The known progeroid syndromes clustered in their usual groups as previously visualized in Figure 2A, showing little interference when adding the three new syndromes. ATLD2 clustered with the group of progeroid syndromes containing Xeroderma Pigmentosum A and Cockayne syndrome, SPG49 clustered with Ataxia Telangiectasia, and MGORS clustered with Seckel syndrome and Nijmegen breakage syndrome. Lastly, we used the SVM to predict the syndromes’ likelihood of being progeroid. ATLD2 scored 3.49, SPG49 scored 1.75 and MGORS scored 1.56 supporting the hypothesis of these syndromes being progeroid.

To further illustrate the predictive power of our algorithms we further decided to test if Charcot-Marie-Tooth disease (CMT2Z), a disease recently associated with Cockayne Syndrome [13], could be a possible new progeria. Hierarchical clustering indeed showed close phenotypical correlation with Cockayne syndrome (Supplementary Figure 3A) and a relatively high SVM score of 1.24 suggesting that the disorder could be classified as progeroid.

OMIM.org

This study is a literature review based on the selection of relevant articles focusing on reviews, meta-analysis and clinical studies. The primary source of articles was PubMed Central (PMC) and the process of selecting articles occurred October 1st - November 5th. When designing our search, we used the website https://www.OMIM.org as a base. OMIM.org is an online compendium of the human genes and genetic phenotypes. The site contains PMC referenced full-text overviews on all known Mendelian disorders and over 16.000 genes. OMIM additionally has a unique search feature facilitating the option of focusing the search on diseases with phenotype description and known molecular basis.

The literature search was based on an “Advanced Search” using relevant medical subject heading and keywords. The relevant terms and keywords were then combined by exploiting the Boolean Operators ‘OR’, ‘AND’, ‘NOT’. The process of specifying the search originated from the terms ‘Progeriod’ and ‘Progeria’ with opt in ‘# phenotype description, molecular basis known’ due to the purpose being phenotypic analysis. This search was revealed to be insufficient, seeing as the result excluded several confirmed progeroid syndromes. Therefore, we added ‘premature aging’ in addition to the original terms as keywords with a proximity search to limit the distance in words between two keywords to 1. The second search resulted in 66 hits including all previously reported as well as unreported progeroid syndromes on https://www.mitodb.com, making this the base of further selection.

To reinforce the quality of the study we supplemented the OMIM search with an additional PMC search focused on articles published later than those referenced on OMIM. This thoroughness ensured that using OMIM as the original source of literature did not lead to insufficient or inadequate literature for the study. As an example, the Brachiooculofacial syndrome had 24 articles referenced on OMIM, the latest dating back to 2015. To account for research made after 2015, we supplemented with an additional PMC search using the keyword “branchiooculofacial” and focusing the search on articles published from Jan 1st 2015 and forward. This resulted in 7 hits, giving us a total of 31 articles on the syndrome. After screening all articles, 24 were excluded on the basis of the exclusion criteria, thereby limiting the result to accessible English articles describing the syndromes phenotype in 2 or more unrelated individuals. The articles excluded were case reports, articles without phenotypic description, non-English articles and articles without full text link. As illustrated in Figure 1A, an identical processing was performed for all 9 syndromes leading to a total of 26 articles [14, 17, 23–46].

Inclusion and exclusion

Originally, we included 19 syndromes (see Figure 1) since they all presented with phenotypic traits of progeroid syndromes and had monogenic originance therefore fulfilling our inclusion criteria. However, to ensure the quality of the study we excluded 10 disorders presenting with only 1 studied patient or based purely on biologically related patients. Other syndromes already existed in the database and therefore, naturally, were excluded. The remaining 9 mendelian disorders met the inclusion criteria by respectively being new to the database, showing phenotypic signs of progeria, having a monogenic origin and having been studied in two or more unrelated patients.

Hierarchical clustering

Hierarchical clustering illustrates how diseases cluster with each other based on correlation of their traits by creating a dendrogram [47]. The dendrogram connects diseases sharing similar traits by measuring similarity and linkage methods. In the dendrogram two closely correlated diseases will be portrayed with a short distance and short leg. Furthermore, the tool calculates a cophenetic correlation when clustering diseases to ensure resemblance of the original sample distance and the cophenetic distance in the dendrogram. The value measures how well the clustering result matches the original resemblance by identifying what uncentered similarity and average linkage will produce the best correlation [48].

The disease network

The Disease Network is a network connecting syndromes by shared phenotypic traits. Each dot represents a syndrome and each line between dots represents a shared trait. If multiple traits are shared by two syndromes, the line will appear darker, thicker and shorter. It is possible to apply different thresholds in order to limit the network so that it only displays diseases with a significant number of shared traits. Some features occur in almost all patients and are therefore perhaps more biologically relevant for the particular disease. However, patients may also suffer from features that are not seen in all patients and therefore are not as biologically relevant. By applying a threshold where only high prevalence features are considered, less common features are excluded. Thresholding is performed by summing the product of percentage prevalence for each shared phenotype. For instance, if two diseases share one phenotype (e.g. short stature) with 10% each, the product score is 100. The same product score would be calculated in the case for 1% in one disease and 100% percent in another disease. However, if one disease has 100% and another 100% then the product score is 10.000. The product score is thus exponentially higher for diseases with higher percentage phenotypic overlap. If there are more than one shared phenotype then the product score for each phenotype is summed. The lowest threshold is a product score of 1000 which indicates at least an overlap of at least 34% in at least one phenotype while the highest threshold currently available is 20.000 which indicates at least two shared phenotypes with 100% prevalence in both diseases.

Mean prevalence of phenotypes

The mean prevalence of phenotypes was calculated from manually curated literature where each phenotype was mentioned. We calculated the mean prevalence of phenotypes for a disease by calculating the sum of all individuals with the phenotype and divide it with all individuals examined for that phenotype. The prevalence of phenotypes for disease categories is the mean prevalence for each phenotype in each disease in that category.

Principal component analysis

Principal component analysis was performed on the disease and phenotype matrix using cluster 3.0 [49].

Support vector machine

The support vector machine (SVM) was created by exporting the symptom-vectors of the known progeria and non-progeria diseases. The symptom-vectors were processed in python using the ‘SVC’ class from scikit-learn to generate an SVM. The SVM was trained using a “linear kernel”. The parameters of the SVM were then exported and the classifier functions for the web-page was created (with php) using these parameters [50, 51].

Omics data collection and hierarchical clustering

Publicly available gene expression data for 13 progeroid syndromes were collected using PandaOmics [52], an AI-driven target discovery platform with its proprietary pathway analysis approach, iPANDA [53]. The syndromes included Branchio-Oculo-Facial Syndrome (GSE108521 (RNA-seq, Neural crest cell line)), Ataxia-telangiectasia (GSE75852 (RNA-seq, NPCs), E-MTAB-1217 (Microarray, NPCs), GSE142842 (RNA-seq, blood), GSE61019 (Microarray, cortex), GSE152287 (Microarray, fibroblasts), GSE35347 (Microarray, fibroblasts), GSE35347 (Microarray, IPSCs), GSE75852 (RNA-seq, IPSCs)), Bloom Syndrome (GSE54502 (Microarray, fibroblasts), GSE123447 (Microarray, fibroblasts)), Cockayne syndrome (GSE144557 (Microarray, Cerebellum), GSE36648 (Microarray, IPSCs)), Dyskeratosis congenita (GSE64023 (Microarray, MSCs), GSE77525 (Microarray, T cells), GSE83501 (RNA-seq, lung)), Hutchinson-Gilford progeria syndrome (E-MEXP-2597 (Microarray, fibroblasts), GSE28863 (Microarray, fibroblasts), GSE113957 (RNA-seq, fibroblasts), GSE137083 (RNA-seq, fibroblasts), GSE3860 (Microarray, fibroblasts)), Nestor-Guillermo progeria syndrome (GSE65170 (Microarray, fibroblast), GSE65172 (Microarray, Mesenchymal stem cell)), Nijmegen breakage Syndrome (GSE83686 (Microarray, NPCs), GSE94707 (Microarray, fibroblasts), GSE13909 (Microarray, lymphoblasts)), Seckel syndrome (GSE121384 (RNA-seq, IPSCs)), Werner Syndrome (GSE48761 (Microarray, fibroblasts)), Xeroderma Pigmentosum group A (GSE55484 (Microarray, fibroblasts)), Xeroderma Pigmentosum group C (GSE119501 (RNA-seq, fibroblasts), GSE133084 (Microarray, fibroblasts)), and Xeroderma Pigmentosum group V (GSE70818 (Microarray, fibroblasts)).

All gene expression datasets were processed according to platform-specific protocols. Differential expression analysis was performed using the limma package. For each disease, the results of gene expression comparisons were combined into a meta-analysis, which was utilized to identify disease expression signatures. Using PandaOmics, disease-expression signatures were extracted from the meta-analysis section, which allowed users to calculate logarithmic fold-changes (LFCs) and Q-values across all gene expression datasets. This was achieved using min-max normalization for LFC values and Stouffer’s method for combining p-values. Disease-expression signature vectors with combined LFC values were aggregated into a matrix, which was then used for hierarchical clustering analysis. Genes not expressed in all 13 progeroid syndromes were excluded from the expression table. Hierarchical clustering was performed using the scipy python package and the cluster.hierarchy.linkage function with metric=‘euclidean’ and method=‘average’. The hierarchical clustering results were visualized using the ete3 python package Tree, TreeStyle, NodeStyle, and TextFace functions.

Signaling pathway analysis

Pathway analysis for 13 progeroid syndromes was performed using iPANDA algorithm [53]. Reactome signaling pathway graph was used as a database for the iPANDA algorithm [54]. Using a linear combination of logarithmic fold-changes, statistical weights, and topological weights applied to each pathway member gene, iPANDA algorithm estimates the direction and intensity of pathway activation. Accordingly, the output of iPANDA represents the difference in gene expression between disease and control groups, and the iPANDA score is used to calculate the score for each signaling pathway. A high iPANDA score indicates an upregulation of a pathway, whereas a low score indicates a downregulation. Combined iPANDA scores for each disease were calculated using min-max normalization for iPANDA values and Stouffer’s method for combining p-values. Lists with significantly (Combined iPANDA p-value < 0.05) positively upregulated and downregulated pathways were collected and overlaps between groups of progeroid syndromes (Group 1 includes Xeroderma Pigmentosum group A, Cockayne syndrome, and Ataxia-telangiectasia, Group 2 includes Werner syndrome and Hutchinson-Gilford progeria, Group 3 includes Bloom syndrome and Nijmegen breakage, Group 4 includes Nestor-Guillermo and Seckel; Group 5 includes Hutchinson-Gilford progeria, Werner syndrome and Nestor-Guillermo; Group 6 includes Bloom syndrome, Xeroderma Pigmentosum group C and Xeroderma Pigmentosum group V) were performed and visualised using upsetplot python package.