Anaïs Baudot, Cécile Beust, Galadriel Brière, Morgane Térézol
Aix*Marseille Université, INSERM, Marseille Medical Genetics, France
Aix*Marseille Université, CNRS, Marseille Mathematics Institute, France
Barcelona Supercomputing Center
jupyter-notebookpython 3.10.9- Python packages:
multixrank==0.1(https://multixrank-doc.readthedocs.io/en/latest/)obonet(https://pypi.org/project/obonet/)pyhpo==3.2.2(https://pypi.org/project/pyhpo/)pandasxml.etree.ElementTreecsvnetworkxglobos
Go to the jupyter-notebooks folder, and run the following steps:
step1_Build_RareX_Layer.ipynbstep2_Build_Orphanet_Layer.ipynbstep3_Build_Bipartite_Network.ipynbstep4_Run_MultiXrank.ipynbstep5_Parse_MultiXrank_Results.ipynbstep6_Analyze_MultiXrank_Results.ipynb
Rare diseases often pose significant diagnostic challenges due to their low prevalence and the diversity of their associated symptoms. Gaining a more comprehensive perspective on the symptoms and phenotypes associated with rare diseases is crucial, as it could facilitate the early detection and more efficient management of these uncommon conditions.
In recent years, efforts have been made to organize the available knowledge about rare diseases. We can, for example, mention the Orphanet database [1], which provides a classification of these rare diseases along with various information such as the mutated genes causing the disease and the associated phenotypes.
When analyzing a dataset of symptoms associated with a cohort of patients with various rare diseases, the incorporation of knowledge from public databases is essential to uncover new unrecognized symptoms that may potentially be linked to rare diseases. However, the diversity and heterogeneity of available information can make this task challenging.
Here, we propose to use multilayer networks to represent this complex and heterogeneous information, as is has the potential to not only enhances the interpretability of the data but also provides a powerful framework for uncovering hidden patterns and relationships within the complex landscape of rare diseases, leveraging both the information contained in the input dataset and the knowledge contained in public databases.
In particular, Random Walk with Restart (RWR) approaches have demonstrated their utility in the analysis of complex biological networks [2]. In these approaches, a random walker explores the network, starting from a specified seed node. As the random walker navigates the network, it computes scores for each node in the network, reflecting their proximity to the initial seed. These scores offer a valuable metric for assessing the influence and relevance of nodes within the network.
Recently, MultiXrank, a novel RWR algorithm for multilayer networks, has been introduced [3].
MultiXrank is able to navigate very complex and heterogeneous networks, composed of various node types and interaction types. Here, we propose to use this algorithm on a two-layer network derived from the symptom dataset provided in the Rare-X challenge, supplemented by information extracted from the rare disease database Orphanet and the Human Phenotype Ontology (HPO) database [4]. We expect that the iterative application of MultiXrank, starting from each Rare-X disease seed node on the multilayer network, will yield RWR scores that effectively highlight the significance of symptom and phenotype nodes with respect to the seed disease.
The first three steps of the pipeline consist in building the multilayer network represented in the figure below.
- Building the Rare-X layer
The first layer, hereafter referred to as the Rare-X layer, contains information extracted from the Rare-X dataset of symptoms associated with a cohort of rare disease patients. This layer contains three node types:
- Rare-X disease nodes
- Patient nodes
- Symptom nodes
Every Patient node is linked to the corresponding Rare-X disease as reported in the dataset. We establish a weight of 1 for each Patient-Rare-X disease connection. Additionally, Patient nodes connect to Symptom nodes based on the symptoms documented in the dataset. In the case of boolean symptoms, where the symptom is either present or absent, we only report the presence of the symptom with an edge weighted 1. For symptoms associated with scores, such as CSHQ symptoms, we report the normalized score as the edge weight.
The code used to generate the Rare-X layer is available in the step1_Build_RareX_Layer.ipynb jupyter-notebook.
- Building the Orphanet layer
The second layer, hereafter referred to as the Orphanet layer, contains information extracted from the Orphanet and HPO databases. It contains two types of nodes:
- Orphanet disease nodes
- Phenotype nodes
This layer establishes connections between Orphanet diseases and their associated phenotypes, extracted from the Orphanet database. Additionally, we augment this layer by incorporating phenotype-phenotype associations from the HPO database. The edges connecting Orphanet diseases and associated phenotypes are weighted based on the reported prevalence of the phenotype in Orphanet. For the HPO phenotype-phenotype associations, we apply a fixed weight of 0.2.
Furthermore, we leverage the genetic information from Orphanet, specifically the mutated genes associated with diseases, to establish connections between Orphanet diseases within the layer. Two Orphanet diseases are linked with an edge weighted 1 if they share at least one mutated gene as documented in the Orphanet database. Including those associations facilitates MultiXrank's navigation between Orphanet diseases, with the rationale that diseases sharing the same causative gene(s) may present common (and potentially yet unreported) phenotypes.
The code used to generate the Orphanet layer is available in the step2_Build_Orphanet_Layer.ipynb jupyter-notebook.
- Connecting the Rare-X layer and the Orphanet layer with bipartite associations
To enable MultiXrank to navigate between the Rare-X layer and the Orphanet layer, we introduce bipartite edges connecting Rare-X diseases and Orphanet diseases. However, it's worth noting that out of the 27 Rare-X diseases, only 15 could be mapped to corresponding Orphanet diseases, resulting in just 15 bipartite edges linking the two layers. As a result, Rare-X diseases lacking correspondence with Orphanet were left without these bipartite edges.
The code used to generate the bipartite associations is available in step3_Build_Bipartite_Network.ipynb jupyter-notebook.
Note: The two layers and the bipartite interactions are stored in tsv files for MultiXrank to access:
network/multilayer/RareX_layer.tsvfor the Rare-X layernetwork/multilayer/Orphanet_layer.tsvfor the Orphanet layernetwork/bipartite/bipartite_RareX_Orpha_diseases.tsvfor the bipartite associations
After building the multilayer network, we can run MultiXrank, taking the Rare-X disease nodes iteratively as seeds, as illustrated by the blue node in the above figure. Starting from a Rare-X disease node, the random walker can navigate within and across the two layers and score each node according to its relevance with respect to the seed node.
By comparing top scored Symptom nodes from the Rare-X layer with the top scored Phenotype nodes from the Orphanet layer we expect to reveal novel symptoms associated with rare diseases that are not reported in Orphanet.
The code used to run MultiXrank is provided in the step4_Run_MultiXrank.ipynb jupyter-notebook.
- Parsing MultiXrank results
MultiXrank returns one file for each layer of the multilayer network. Those files contain the scores associated with each node composing the layer considered. To enhance result readability, we extract the nodes with the highest scores from each layer and consolidate them in a single file.
The code used to parse MultiXrank results is provided in the step5_Parse_MultiXrank_Results.ipynb jupyter-notebook.
Note: The raw MultiXrank output files are stored in the folder results_MultiXrank/output/.
Processed output files are stored in the folder results_MultiXrank/results/.
- Interpreting MultiXrank output scores
Finally, we compare the top scored Symptom nodes and the top score Phenotype nodes to find Symptoms that are highly scored with no equivalent in top scored Phenotypes for each Rare-X disease.
The interpretation of the results for each Rare-X disease is available in the step6_Analyze_MultiXrank_Results.ipynb jupyter-notebook.
Understanding rare diseases requires comprehensive insights into their diverse symptoms and phenotypes. Leveraging innovative approaches like MultiXrank on intricate biological networks, complemented by data from sources such as Orphanet and HPO, offers promising opportunities for the detection of novel symptoms associated with rare diseases. Moreover, the multilayer network can easily be completed with other sources of knowledge from public databases or include new cohorts and symptoms, enabling a more comprehensive exploration of the rare disease landscape.
[1] Orphanet: an online database of rare diseases and orphan drugs. Copyright, INSERM 1997. Available at http://www.orpha.net Accessed 08/10/2023.
[2] Valdeolivas, Alberto, Laurent Tichit, Claire Navarro, Sophie Perrin, Gaëlle Odelin, Nicolas Levy, Pierre Cau, Elisabeth Remy, and Anaïs Baudot. “Random Walk with Restart on Multiplex and Heterogeneous Biological Networks.” Bioinformatics 35, no. 3 (February 1, 2019): 497–505. https://doi.org/10.1093/bioinformatics/bty637.
[3] Baptista, Anthony, Aitor Gonzalez, and Anaïs Baudot. “Universal Multilayer Network Exploration by Random Walk with Restart.” Communications Physics 5, no. 1 (July 1, 2022): 1–9. https://doi.org/10.1038/s42005-022-00937-9.
[4] Köhler, Sebastian, Michael Gargano, Nicolas Matentzoglu, Leigh C Carmody, David Lewis-Smith, Nicole A Vasilevsky, Daniel Danis, et al. “The Human Phenotype Ontology in 2021.” Nucleic Acids Research 49, no. D1 (January 8, 2021): D1207–17. https://doi.org/10.1093/nar/gkaa1043.