The multilayer community structure of medulloblastoma

•The molecular interpretation of rare diseases is a challenging task•Multilayer networks allow patient stratification and explainability•We identify subgroup-specific genes multilayer associations in medulloblastoma•Multilayer community analysis enables the Multilayer interpreting basis diseases, which particularly where number cases small compared with size associated multi-omics datasets. In this work, we develop dimensionality reduction methodology to minimal set that characterize disease subgroups based on their persistent association network communities. We use approach study medulloblastoma, childhood brain tumor, using proteogenomic data. Our able recapitulate known medulloblastoma (accuracy >94%) provide clear characterization gene associations, downstream implications for diagnosis therapeutic interventions. verified general applicability our method an independent dataset >98%). This opens door new generation network-based methods overcome specific limitations To improve understanding complex systems, it crucial take into account multiple types relationships inherently define natural systems. The so-called (alternatively multiplex networks) has recently become one most important directions science (Kivela et al., 2014Kivela M. Arenas A. Barthelemy Gleeson J.P. Moreno Y. Porter M.A. networks.J. Complex Netw. 2014; 2: 203-271Crossref Scopus (1728) Google Scholar; Aleta Moreno, 2019Aleta nutshell.Annu. Rev. Condens. Matter Phys. 2019; https://doi.org/10.1146/annurev-conmatphys-031218-013259Crossref PubMed (57) Scholar). A organized layers representing different nodes edges (Figure S1). Despite offering means achieve comprehensive view human by accounting complexity accumulated biomedical data, biological exhibit range research challenges still require substantial investigation (Kristensen 2014Kristensen V.N. Lingjærde O.C. Russnes H.G. Vollan H.K.M. Frigessi Børresen-Dale A.L. Principles integrative genomic analyses cancer.Nat. Cancer. 14: 299-313Crossref (236) Among them, detection area promising biomedicine, facilitating evaluation relevant among identification candidate targets drug development repurposing (Halu 2019Halu De Domenico Sharma diseases.NPJ Syst Biol Appl. 5: 15Crossref (30) Valdeolivas 2019Valdeolivas Tichit L. Navarro C. Perrin S. Odelin G. Levy N. Cau P. Remy E. Baudot Random walk restart heterogeneous networks.Bioinformatics. 35: 497-505Crossref (71) Popular strategies include Louvain algorithm (Blondel 2008Blondel V.D. Guillaume J.L. Lambiotte R. Lefebvre Fast unfolding communities large Stat. Mech. Theor. Exp. 2008; https://doi.org/10.1088/1742-5468/2008/10/p10008Crossref Scholar), greedy optimization technique, maximize structural metric called modularity (Newman Girvan, 2004Newman M.E.J. Girvan “Finding evaluating structure networks.Phys. 2004; 69: 026113Crossref (8826) Modularity defined as fraction within group significantly enriched when random model. It measures strength given partition (Reichardt Bornholdt, 2006Reichardt J. Bornholdt Statistical mechanics detection.Phys. 2006; 74: 016110Crossref (1107) widely used meta-heuristics networks. outperforms other algorithms accuracy, scalability, computing time (Yang 2016Yang Z. Algesheimer Claudio J.T. comparative artificial networks.Sci. Rep. 2016; 6: 1-18Crossref (244) Moreover, implemented software, been adapted (Didier 2015Didier Brun Identifying from networks.PeerJ. 2015; 3: e1525Crossref (33) Didier 2018Didier randomized modularity.F1000Res. 2018; 7: 1042Crossref (5) Nevertheless, determination remains open problem such extent preferred formulation often domain (Porter 2009Porter Onnela Mucha P.J. Communities networks.arXiv. 2009; http://arxiv.org/abs/0902.3788Google One major conundrum modularity-based approaches intrinsic limit resolution, priori impossible rule out at certain level resolution may be composed cluster smaller (Fortunato Barthélemy, 2007Fortunato Barthélemy Resolution detection.Proc. Natl. Acad. Sci. U S 2007; 104: 36-41Crossref (1910) Lancichinetti Fortunato, 2011Lancichinetti Fortunato Limits maximization 2011; 84: 066122Crossref (273) words, topological descriptions, each its own importance, coexist scales are detected alternative values (Arenas 2008Arenas Fernández Gómez Analysis levels.New https://doi.org/10.1088/1367-2630/10/5/053039Crossref (365) As consequence, meaningful communities, groups interest robustly express strong heavily depends choice value used. limitation can through stable partitions resolution. Indeed, changing indicative modular structures Here systematically found belong same across values. view, two or more consistently will deemed strongly evidence network. applied concept altered cohort patients (MB) who were previously stratified data (Forget 2018Forget Martignetti Puget Calzone Brabetz Picard D. Montagud Liva Sta Dingli F. al.Aberrant ERBB4-SRC signaling hallmark 4 revealed phosphoproteomic profiling.Cancer Cell. 34: 379-395.e7Abstract Full Text PDF (63) Scholar) 1). aim, (see methods: “multilayer analysis” Figure 2).Figure 2Identification trajectoriesShow full caption(A–D) For genes, identified they (A). then computed pairwise Hamming distances trajectories visited (B). corresponding distance matrix (C) was represented form dendrogram (D) clustering analysis. See also S3.View Large Image ViewerDownload Hi-res image Download (PPT) (A–D) S3. MB malignant fast-growing primary central nervous system originates embryonic cells spinal cord no causes preferential manifestation children (aged 1–9 years). being rare, common cancerous tumor children. Four pediatric distinct clinicopathological features have identified: WNT, SHH, Group 3 (G3), (G4) (Taylor 2012Taylor M.D. Northcott P.A. Korshunov Remke Cho Y.-J. Clifford S.C. Eberhart C.G. Parsons D.W. Rutkowski Gajjar al.Molecular medulloblastoma: current consensus.Acta Neuropathol. 2012; 123: 465-472Crossref (1108) 2011Northcott Witt H. Hielscher T. Mack Bouffet Hawkins C.E. French al.Medulloblastoma comprises four variants.J. Clin. Oncol. 29: 1408-1414Crossref (902) WNT favorable prognosis, whereas SHH G4 intermediate-level prognosis G3 worst outcome. Seven recurrent genetic alterations (SHH group, CTNNB1 MYC MYCN G4, ERBB4, SRC, CDK6 (Kool 2012Kool David Jones W. Schlanstein Y.J. Koster Schouten-van Meeteren van Vuurden international meta-analysis transcriptome, aberrations, clinical 3, medulloblastomas.Acta 473-484Crossref (659) Ramaswamy 2016Ramaswamy V. Bailey Steven C.C. Doz Kool Dufour Vassal V Milde al.Risk era: 131: 821-831Crossref (297) Taylor 2014Northcott Lee Zichner Stütz A.M. Erkek Kawauchi Shih Hovestadt Zapatka Sturm al.Enhancer hijacking activates GFI1 family oncogenes medulloblastoma.Nature. 511: 428-434Crossref (338) Robinson 2012Robinson Parker Kranenburg T.A. Lu Chen X. Ding Phoenix T.N. Hedlund Wei Zhu al.Novel mutations target 488: 43-48Crossref (565) 2017Northcott Buchhalter I. Morrissy A.S. Weischenfeldt Ehrenberger Gröbner Segura-Wang Rudneva V.A. al.The whole-genome landscape subtypes.Nature. 2017; 547: 311-317Crossref (427) 2014Kool D.T.W. Jäger Pugh T.J. Piro R.M. Esparza L.A. Markant S.L. al.Genome sequencing predicts genotype-related response smoothened inhibition.Cancer 25: 393-405Abstract (457) 2006Clifford Lusher M.E. Lindsey J.C. Langdon J.A. Gilbertson R.J. Straughton Ellison Wnt/Wingless Pathway Activation Chromosome 6 Loss Characterize Distinct Molecular Sub-Group Medulloblastomas Associated Favorable Prognosis.Cell Cycle. 2666-2670Crossref (58) Forget Each subgroup presents heterogeneity survival differences (Jones 2012Jones Hutter B. Sultan al.Dissecting underlying 100-105Crossref (598) so much than proposed, particular concerns (Schwalbe 2017Schwalbe E.C. Nakjang Crosier Smith A.J. Hicks Rafiee Hill Iliasova Stone classification outcome prediction study.Lancet 18: 958-971Abstract (239) results show 94.94%), well better them identifying sets S2). further verify dataset, achieving very high performance case 98.29%). work represents step forward not only but also, general, research, absence sample cohorts makes supporting extremely task. implement way monitor behavior containing upon changes initially sought mentions abstracts scientific publications about “data sources genes”). By interrogating PubTator Central (PTC) (Wei 2019Wei C.-H. Allot Leaman central: automated annotation text articles.Nucleic Acids Res. 47: W587-W593Crossref (77) retrieved list 1,941 multi-species consisting 1,475 (76%), 389 murine (20%), 77 species (4%). (1,387 1,475, network) S3). conceived proof shown there plain gene. Interestingly, seven whose well-known introduction), branch off well-separated exception SRC CTNNB1, physical interactors (IntAct: EBI-15951997). these explored investigate operations dynamic (Cazabet 2017Cazabet Rossetti Amblard Dynamic detection.in: Encyclopedia Social Network Mining. 2. Springer, 2017: 1-10Google birth (a appears), death vanishes), resurgence disappears appears again later on). Along 2,186 unique text-mined experience total 2,517 events 673 S4), indicating instability (all disappear least once) commutability (some reappear several times exact composition). These observations led us realize characterized journey throughout levels reason, tested hypothesis tracing disease-related could exploited purposes “identification subgroups”). stratification. reference (ground truth) consists classical (WNT, G3, G4), represent standard categorization despite possibility granular reported introduction). investigated via fusion information reanalyzed optimally subgroups, while aiming reduce critical required lists 35 display complete datasets (DNA methylation, RNA sequencing, proteomics, phosphoproteomics) Partial available three additional (MB10, MB21, MB33) retained validation results: “sensitivity analyses”). performed hierarchical optimal selection genes. Optimality selected terms representation (parameter ?) similarity ?), maximum accuracy Matthews correlation coefficient (MCC) achieved highest (94.94%) MCC (87%) five clusters G3-G4), selecting those (? = 6) always part along (? 0) (Figures 5, Tables S1–S3). Strikingly, corresponds strict portion sufficient accomplish accurate segregation. observation implies tightly never leave trajectories. An aspect result that, clusters, few escape subtler stratas exist, suggested recent studies Archer 2018Archer T.C. Mundt Gold M.P. Krug K. Mah C.K. Mahoney E.L. Daniel C.J. LeNail Ramamoorthy al.Proteomics, post-translational modifications, reveal subgroups.Cancer 396-410.e8Abstract (68) Scholar).Figure 5Clustering patientsShow captionWard's linkage obtained ? ? 0. rectangles indicate PAM (partitioning around medoids) criteria. color indicates original Scholar): (blue), (red), (G4, green), (G3, yellow). fifth depicted purple, including originally assigned (MB47) (MB09 MB54). Figures S8 S9 S4, S5, S6.View Ward's S6. consist partial genes”), excluded samples parameter procedure set. (patient MB10) (patients MB21 similar remaining Jaccard Index (J) parametrized Patient MB10 shows MB22 (J 0.263), belongs likewise eight following ranking positions (Table S4). (MB31 J 0.2653; MB34 0.2631; MB30 0.2601). Finally, MB33 (MB30 0.2168; 0.2106). Of note, MB31 fourth 0.2080), MB16 third 0.2081). parameters optimized classifying correctly classified). ?, patients, correspond average 1,812.74 per (SD 106.97) (i.e., 87.56% 0.44) patient) S5). some uniquely all (148 patients; 83 115 46 G3-G4 260 patients). evaluated robustness analyses. first analysis, shuffled 10,000 times, maintaining yielded 54.76% 0.11) 0 distribution accuracies dramatically lower procedure, stratification, non-random gene-subgroup S6). second recursively after excluding iteration. observed progressive decrease and, expected, higher iterations, less effective S7). Overall, decay iterative removal divided phases: short initial phase (accuracies between 94.94 88.57) removed iteration (1027.72 average), long intermediate 79.76 69.96) (23.31 final (between 57.06 31.43) 1.08 before drops At end cumulative 5,950.63 (average patient; 38 effectiveness nature algorithm, even pool operates largely reduced. test if good similarities sensitivity significance assessment. first, multiscale bootstrap resampling (Suzuki Shimodaira, 2006Suzuki Shimodaira Pvclust: R package assessing uncertainty clustering.Bioinformatics. 22: 1540-1542Crossref (1500) assigns confidence value, approximately unbiased probability (pvAU), cluster. High pvAU clusters. second, Monte Carlo (Kimes 2017Kimes P.K. Liu Hayes D.N. Marron J.S. clustering.Biometrics. 73: 811-821Crossref (64) empirical p Gaussian approximate difference b