Integrative Multi-Omics and Network Analyses Reveal Pathogenic and Protective Pathways in Centronuclear Myopathies

Abstract

Centronuclear and myotubular myopathies (CNMs) are rare, inherited muscle disorders characterized by muscle atrophy, weakness, and altered muscle fiber structure, primarily caused by mutations in MTM1, DNM2, or BIN1. The molecular mechanisms driving CNM are only partially understood, and no curative therapies are available. To elucidate molecular pathways involved in CNMs, we present an integrative multi-omics analysis across several CNM mouse models untreated or treated with pre-clinical strategies, combining transcriptomic, proteomic, and metabolomic datasets with curated interaction, metabolic, tissue, and phenotype knowledge using network-based approaches. Weighted Gene Co-expression Network Analysis (WGCNA) identified gene modules commonly altered in three CNM genetic forms. Modules correlated with improved muscle function were enriched for processes such as muscle contraction, RNA metabolism, and oxidative phosphorylation, whereas modules linked to disease severity were enriched for immune response, innervation, vascularization, and fatty acid oxidation. We further integrated transcriptomic, proteomic, and metabolomic data from the Mtm1^−/y mouse model with public knowledge bases into a multilayer network, and explored it using a random walk with restart approach. These analyses highlighted metabolites closely connected to CNM phenotypes, some of which may represent candidates for nutritional or pharmacological modulation. Our findings illustrate how integrative multi-omics and network analyses reveal both pathogenic and protective pathways in CNM and provide a foundation for identifying novel therapeutic opportunities.

1. Introduction

Centronuclear and myotubular myopathies (CNMs) are a group of rare inherited muscular disorders characterized by muscle weakness, hypotonia, and abnormal positioning of nuclei within muscle fibers [1]. These diseases are primarily caused by mutations in genes such as MTM1, DNM2, and BIN1. Mutations in additional genes such as RYR1 or TTN are associated with other forms of CNM with abnormal nuclei positioning and additional histological features. Although a potential correlation between nuclei centralization and phenotype severity has been suggested for XLCNM, it is yet to be established for most forms of CNM [2]. Despite progress in identifying causative mutations, the intricate molecular pathways leading from these genetic disruptions to the clinical manifestation of CNMs are still not fully delineated, limiting the development of effective therapies. To investigate the mechanisms implicated in CNMs, several murine models have been developed and faithfully replicate the muscle phenotypes observed in patients [3]. Mtm1^−/y knock-out mice exhibit a progressive myopathy beginning at 2–3 weeks of age, characterized by marked muscle hypotrophy, reduced locomotor activity, and lethality by approximately 8 weeks [4]. Bin1^mck−/− mice (muscle-specific deletion of Bin1) maintain near-normal locomotor function but demonstrate reduced muscle strength by 8 weeks of age [5]. Knock-in mice for the most common mutation in the neonatal DNM2 form (Ser619Leu; Dnm2^SL/+) present with neonatal feeding difficulties associated with decreased body weight, pronounced muscle atrophy, and impaired locomotor activity, though they survive into adulthood [6]. These mouse models thus represent severe, mild and intermediate CNM, respectively. All three models display centronuclear myopathy (CNM)-like histopathological features by 7–8 weeks of age. Notably, these CNM genes encode for proteins functioning in the same pathway in muscle. Myotubularin (MTM1) is a lipid phosphatase, amphiphysin 2 (BIN1) binds lipids and senses and promotes membrane curvature, and dynamin 2 (DNM2) binds lipids and amphiphysin and eventually fission membrane. Moreover, we previously demonstrated that different CNM forms can be rescued by a common strategy: increasing BIN1 expression, reducing DNM2 levels, or administering tamoxifen, each of which alleviates most phenotypes in the corresponding mouse models [1,3]. These findings support shared pathogenic mechanisms and suggest that common therapeutic principles may apply across different CNM subtypes. In the present study, we aimed to identify these mechanisms and therapeutic targets through integration of multi-omics data.

Omics approaches, including transcriptomics, proteomics, and metabolomics, offer powerful tools for uncovering the broader molecular landscape involved in CNM. These approaches represent comprehensive and unbiased techniques to characterize complex biological systems, enabling the identification of key molecular mechanisms underlying various processes [7]. Single-omics analysis often involves differential expression analysis between two conditions to identify dysregulated entities. Such analyses can yield valuable insights, but they cannot take into account the intricate regulatory and interaction processes that connect different layers of omics data, and they are limited to pairwise comparisons. Therefore, integration of multi-omics data together with clinical information has shown great promise in achieving a holistic understanding of complex biological systems [8]. To our knowledge, such a multi-omics integration has never been applied in the CNM field, and provides a novel means to uncover disease mechanisms, prioritize cross-subtype targets, and identify metabolites potentially linked to pathological processes.

Integration approaches based on biological networks have gained traction in recent years. Weighted Gene Co-expression Network Analysis (WGCNA) is one such approach that can be used to construct networks based on the correlation patterns between genes across multiple samples and to identify gene modules that correlate with clinical traits [9]. Multilayer network approaches, which integrate diverse biological data types (e.g., transcriptomic, proteomic, and metabolomic data), data sources (e.g., protein–protein interactions, pathway co-membership, co-expression), and/or experimental conditions (e.g., transcriptomic data obtained from several models) offer an even more comprehensive framework [10]. By leveraging high-throughput data from multiple layers, these methods allow modeling of complex, multi-dimensional relationships that single-layer network approaches cannot capture.

Here, we used network integration techniques to uncover novel molecular interactions and pathways that contribute to disease pathomechanisms in CNMs. First, we performed WGCNA on several cohorts of CNM murine models treated with various therapeutic approaches. Then, we performed a case study focusing on the investigation of the multi-modal pathomechanisms implicated in XLCNM (X-linked CNM, caused by mutations in MTM1), for which comprehensive experimental transcriptomics, proteomics, and metabolomics data are available. Thus, we built and explored a heterogeneous multilayer network integrating the multi-omics data generated for the Mtm1^−/y mouse model with biological knowledge extracted from publicly available databases.

2. Results

2.1. Weighted Gene Correlation Network Analysis of Several Mouse Models of CNM

2.1.1. Transcriptomic and Phenotypic Data Collection

To identify genes implicated in skeletal muscle structure and function and in the pathogenesis of CNM, we performed WGCNA on several murine models of myopathies treated with various therapeutic approaches (

Table 1. Overview of the transcriptomic cohorts used for WGCNA.

Cohort	Mouse Model	Model Severity	Therapeutic Approach	Therapeutic Effect	Age (Weeks)	Genetic Background	n	Original Phenotypic Study	Original RNA-Seq Study
A	Mtm1−/y	Severe	* Dnm2+/−	Complete rescue	7	50% 129 Pas, 50% C57BL/6N	17	[12]	[3]
B			* TgBIN1	Complete rescue	7	129 Pas	15	[13]
C			Tamoxifen	Partial	7	129 Pas	14	[14]
D	Bin1mck−/−	Mild	ASO Dnm2	Partial rescue	7	C57BL/6N	16	[5]
E	Dnm2S619L/+	Intermediate	ASO Dnm2	Complete rescue	7	C57BL/6N	16	[6]
F	Bin1mck−/−	Mild	Tamoxifen	Partial rescue	8	C57BL/6N	12	[11]	This study
G	Dnm2S619L/+	Intermediate	Tamoxifen	Partial rescue	8	C57BL/6N	16		[11]

* = genetic crossing, n = total number of samples, ASO = antisense oligonucleotide.

). In particular, three models of CNM (Mtm1^−/y, Bin1^mck−/−, Dnm2^S619L/+) were considered. Therapeutic approaches included Bin1 and Dnm2 gene expression modulation by genetic crossing or antisense oligonucleotide (ASO) and tamoxifen repurposing. Genetic backgrounds differ between some groups. For each cohort, bulk RNA-sequencing data was retrieved from previous publications [3,11] or generated for this study (cohort F). For all cohorts except cohort F (Bin1^mck−/− +/− tamoxifen), 4 groups were studied: WT, disease model, WT treated, and disease model treated. In cohort F, 3 groups were considered: WT, disease, and disease treated. In total, bulk RNA-seq data of 106 mice were used. Overall, Principal Component Analysis (PCA) of normalized and variance stabilized counts revealed a separation of samples according to their status for WT, WT treated, and disease groups along PC1, which explained 22% of the variance (Figure 1). Notably, tamoxifen-treated WT samples were an exception to this pattern, clustering on the right side of PC1 alongside disease samples. For the disease groups, we observed that Mtm1^−/y samples (cohorts A–C) clustered together on the top of PC2. Disease-treated samples exhibited a variable distribution on the PCA plot, with some clustering alongside WT samples and others aligning with disease samples in a cohort-specific manner. In particular, disease models treated with tamoxifen diet supplementation (cohorts C, F, and G) tended to cluster with disease samples, correlating with a partial rescue of the phenotypes (Table 1). In contrast, disease-treated samples in other cohorts were positioned closer to WT samples, especially in cohorts B, D, and E, correlating with a rather complete correction of phenotypes (Table 1).

Phenotypic characterizations of the cohorts were retrieved from their respective publications [5,6,11,12,13,14]. In particular, we retrieved the average value of the following metrics for each group when they were available: body mass (g), tibialis anterior (TA) mass normalized by body mass (mg/g), maximal specific force of the TA relative to WT, hanging test (time to fall off an upside-down wire cage lid) (s), percentage of muscle fibers with internalized or central nuclei, percentage of muscle fibers with abnormal mitochondria localization, and percentage of small muscle fibers (fiber diameter < 40 μm). Full phenotypic data are available in Supplementary Information S2. In the absence of mouse identifiers connecting the phenotypic and transcriptomic studies, mice were assigned their average group value for each phenotypic trait.

2.1.2. Identification of Gene Co-Expression Modules Correlated to Phenotypic Traits

WGCNA was performed to build a gene expression correlation network and identify modules of highly correlated genes. Then, the correlation between module expression and phenotypic traits was computed (Figure 2). In total, 24 modules were identified, with size ranging from 47 to 1848 genes. Moderate to high correlation coefficients (absolute values ranging from 0.3 to 0.7) were observed for some modules and traits, with the strongest positive and negative correlations observed for normalized TA mass. Module–trait correlation showed low specificity, as most modules were significantly correlated with multiple phenotypic traits, suggesting potential redundancy among these traits. We defined beneficial modules as modules that were positively correlated to normalized TA mass, relative maximal force, and/or hanging time, and negatively correlated to the percentage of fibers exhibiting central nuclei, abnormal SDH staining and/or a small diameter. Modules that showed an opposite correlation pattern compared to beneficial modules were defined as pathogenic.

2.1.3. Enrichment Analysis of Modules Significantly Correlated to Phenotypic Traits

To identify the biological functions associated with each module, Gene Ontology (GO) terms over-representation analysis was conducted for the biological process (BP), cellular component (CC), and molecular function (MF) ontologies (Figure 3, Supplementary Figure S1, Supplementary Information S2, Supplementary Information S3). Among the beneficial modules, the blue module was enriched in genes associated with apoptosis regulation, calcium ion transport and muscle development and contraction (Figure 3A). The pink module was enriched in genes associated with RNA processing, transcription, and translation, and the dark red module was associated with oxidative phosphorylation (Figure 3B). Among pathogenic modules, the brown module was associated with immune system processes. The midnight blue, salmon, light green and grey60 modules were associated with vascularization and/or innervation, with enriched pathways related to angiogenesis and with the development, function, and maintenance of the neuromuscular junction (axonogenesis, synaptic transmission, cell–cell junction, semaphoring-plexin signaling). The green–yellow module was associated with beta oxidation of fatty acids (Figure 3C), and the purple module was associated with histone deacetylation and autophagy (Figure 3D). Overall, these findings underscore multiple pathways linked to muscle structure and function across three murine models of CNM, as well as their modulation through therapeutic intervention.

2.2. Multi-Omics Integration in X-Linked CNMs

To investigate the mechanisms implicated in the pathology of XLCNM (also called myotubular myopathy), we integrated experimental omics data (transcriptomics, proteomics, metabolomics) obtained from Mtm1^−/y mice and their WT littermates with public databases. To integrate these diverse and multiscale data sources, we created a multilayer network composed of a gene multiplex network, a metabolite multiplex network, and three monoplex networks containing phenotypes, metabolic reactions, and tissues (Figure 4, see the methods Section 4 for details). These networks are connected to each other by bipartite networks representing several types of relationships, such as localization, expression, pathway co-membership, or participation in a reaction described in the murine genome-wide metabolic model (

Table 2. Description of the multiplex networks present in the multilayer network.

Multiplex Network	Layer	Number of Nodes	Number of Edges	Density	Weighted	Directed
Genes	PPI (Protein–protein interaction)	5521	34,668	0.003	No	No
	Complex co-membership	4893	47,802	0.005	No	No
	Pathway co-membership	9063	674,975	0.017	No	No
	Gene co-expression	5222	879,879	0.065	Yes	No
	Protein co-expression	2046	695,171	0.358	Yes	No
Metabolites	Pathway co-membership	1604	20,590	0.017	No	No
	Metabolite co-expression	831	115,171	0.334	Yes	No
Reactions		12,987	10,557,827	0.125	No	Yes
Phenotypes		10,093	16,791	<0.001	No	Yes
Tissues		30	0

Density: number of actual edges divided by the number of potential edges of a network. Weighted: whether the edges between nodes are weighted. A higher weight indicates a stronger connection; if the network is not weighted, all edges have a weight of 1. Directed: whether the edges are directed. If the network is not directed, edges go in both directions.

and

Table 3. Description of the bipartite networks present in the multilayer network.

Bipartite Network	Number of Nodes	Number of Edges	Density	Weighted	Directed
Genes—metabolites	6062	69,470	0.004	No	No
Genes—reactions	10,115	22,788	0.004	No	No
Metabolites—reactions	15,047	18,867	<0.001	No	Yes
Reactions—metabolites	13,460	18,064	<0.001	No	Yes
Genes—tissues	16,683	533,921	0.002	Yes	No
Genes—phenotypes	15,134	262,415	0.002	No	No
Metabolites—tissues	3384	6425	0.001	No	No
Phenotypes—tissues	3926	4569	0.001	No	No

Density: number of actual edges divided by the number of potential edges of a network. Weighted: whether the edges between nodes are weighted. A higher weight indicates a stronger connection; if the network is not weighted, all edges have a weight of 1. Directed: whether the edges are directed. If the network is not directed, edges go in both directions.

).

To explore the network, random walk with restart (RWR) was performed using the Python package MultiXrank (version 0.3) [15]. This approach simulates a random particle navigating the network. At each step, the particle can either move to a neighboring node or restart its walk from a node that is randomly selected from a user-provided set of seed node(s). RWR output scores were used to rank and identify top-scoring nodes that are closely connected to the seed(s).

2.2.1. Identification and Removal of Nonspecific Hub Nodes with Random Seeding

After performing RWR on the full multilayer network with several sets of seeds, we observed that generic chemicals such as water, H⁺, or ATP were often present in the list of top-scoring metabolites. We hypothesized that these highly connected nodes, or “hubs,” could create artificially shortened paths and be inaccurately ranked among the top-scoring nodes by RWR. To assess the presence of hubs in the multilayer network, we conducted RWR using 1000 random seeds, each representing a randomly selected protein-coding gene. To assess the impact of including experimental data on hubs, RWR was performed on both the full multilayer network and a subset excluding the experimental layers based on transcriptomic, proteomic, and metabolomic data (called “subset” hereafter). For each RWR iteration, the top-scoring genes and metabolites were identified, and the number of occurrences of each node among the top 10, 50, or 100 ranked nodes was calculated (Figure 5).

When considering the top 10 genes, we observed an over-representation of mitochondrial genes using both the full multilayer network and its subset, with 250 to 750 occurrences in the top nodes out of 1000 iterations. Additionally, we observed an over-representation of water, energy-related molecules (ATP, ADP, GTP, and GDP), and several cations (H⁺, Ca²⁺, Mg²⁺) in the top 10 scoring metabolites. When considering the lists of top 50 or top 100 scoring nodes, we observed an increased number of hubs, including genes encoding ribosomal protein large and small subunits, and additional energy-related metabolites such as NADH and FAD. A higher correlation between full multilayer and subset runs can be observed for the genes compared to the metabolites. This indicates that the metabolomic co-expression layer has a higher impact on the network than the transcriptomic and proteomic layers. This may be due to the difference in composition between the gene multiplex network (three layers extracted from public databases, two experimental layers) and the metabolite multiplex network (one layer extracted from a public database, one experimental layer).

Nonetheless, when considering the top 10, 50, or 100 nodes, density plots showed that the vast majority of top-scoring nodes occurred in less than 5% of iterations. Therefore, we considered that genes and metabolites ranked in the top 50 scoring nodes in more than 5% of iterations were nonspecific hub nodes and represented false-positive hits. For the following analyses, nonspecific hub nodes were removed from the multilayer network.

2.2.2. Investigation of the Pathomechanism of XLCNM Using a Multilayer Heterogeneous Network

After removing nonspecific hub nodes from the multilayer network, we explored it using RWR in order to investigate the pathomechanism implicated in XLCNM. To do this, we performed RWR with several sets of seeds and selected the top 20 highest-scoring gene and metabolite nodes.

We sought to identify nodes that may be implicated in several phenotypes observed in XLCNM, such as skeletal muscle atrophy (HP:0003202), muscle weakness (HP:0001324), and centrally nucleated skeletal muscle fibers (HP:0003687). Therefore, we conducted RWR on the multilayer network, using each of these phenotypes and Mtm1 as joint seeds. We observed a strong overlap among the 20 top-scoring metabolites for these three sets of seeds. Shared hits included inositol (Ins), phosphatidylinositol (PtdIns), and phosphoinostide (PtdInsP) related metabolites: Ins, PtdIns pool, PtdIns(4)P, PtdIns(3,4)P₂, Ins(1,3,4)P₃, Ins(1,3,4,6)P₄, Ins(1,3,4,5)P₄, and scyllo-Ins(4)P. These metabolites are directly linked to the function of Mtm1, which encodes myotubularin, a phosphoinositides phosphatase. Other shared top metabolites were ADP, lutein, geranylgeranyl diphosphate (GGPP), arachidonate, and S-adenosylmethionine. By comparison, the overlap was smaller in the 20 top-scoring genes, with only Mb (myoglobin) and Myh7 (which are both highly expressed in slow-twitch oxidative fibers) as common hits for the three seeds. Additionally, Mtmr12 (encoding a myotubularin-related protein that binds MTM1), Ckm (encoding a cytoplasmic enzyme involved in energy homeostasis), Myl2 (encoding a sarcomeric protein expressed in cardiac muscle and slow-twitch skeletal muscle fibers), Pten (encoding a PtdIns(3,4,5)P₃ phosphatase), and mt-Te (a mitochondrially encoded tRNA) were common hits for the muscle weakness and skeletal muscle atrophy seeds. Seed-specific top-scoring nodes are detailed in

Table 4. Specific top-scoring genes and metabolites using Mtm1 and several XLCNM phenotypes as seeds.

Seeds	Genes	Metabolites
Mtm1 HP:0003202 (skeletal muscle atrophy)	Trpv4 Capn3 Lmna Neb Ryr1 Pomt2 Fkrp Crppa Tnnt1 Fktn Inpp5k Pomt1	Dexamethasone Diacetyl Epitestosterone Oxytocin
Mtm1 HP:0001324 (muscle weakness)	Fhl1 Tnnc2 Klhl41 Chrna1 Itga7 Col6a1 Col6a2 Scn4a Col12a1 Hnrnpa2b1 Col13a1 Pik3r5	9-O-acetylneuraminic acid Einecs 307-780-4 (2s)-1-(Alpha-D-Glucopyranosyloxy)-3-(Hexadecanoyloxy)propan-2-Yl (11z)-Octadec-11-Enoate 1-O-(alpha-D-galactopyranuronosyl)-N-tetradecanoyldihydrosphingosine
Mtm1 HP:0003687 (centrally nucleated skeletal muscle fibers)	Mtmr14 Myot Hnrnpa1l2-ps2 Hnrnpa1 Selenon Pyroxd1 Adgrg6 Bves Ccdc78 Syne1 Sil1 Matr3 Unc45b Fxr1 Orai1 Tpm3 Myf6	Thromboxane A2 Tetrahydrobiopterin Uric Acid 9-Anthroic acid Naadp Niflumic acid Aspirin Diacetyl

. Specific top-scoring genes for the three sets of seeds are all implicated in the neuromuscular disorders described in the Muscle Gene Table [16], except for Pik3r5, a regulatory subunit of phosphatidylinositol 3-kinases that phosphorylates PtdIns to PtdIns3P, as well as Mtmr14, which encodes a myotubularin-related protein, and Myf6, a myogenic factor. Overall, by using a multilayer heterogeneous network and RWR analysis, this study identified hypothetical metabolites and genes linked to Mtm1 function and XLCNM phenotypes, shedding light on shared and phenotype-specific molecular mechanisms.

3. Discussion

This study provides a comprehensive molecular perspective common to several forms of centronuclear myopathies by integrating transcriptomic data with phenotypic profiles obtained from several murine models of CNM treated with various therapeutic approaches. XLCNM was further investigated by integrating multi-omics data with publicly available biological knowledge. These findings contribute to a more detailed understanding of the molecular landscape of CNM and suggest new directions for targeted therapeutic interventions. The limitations of our study include the use of average group values for phenotypic traits instead of individual data points, which may mask subtle within-group variations. Additionally, while murine models provide valuable insights, they may not fully recapitulate human CNM pathology, especially concerning the distribution of muscle fiber types and genetic background diversity. Future studies should seek to verify these findings in human tissues or in vitro models to confirm the translatability of our results.

3.1. Gene Co-Expression Modules and Phenotype Correlation

Employing WGCNA facilitated the identification of gene modules linked to CNM phenotypes, highlighting both pathogenic and beneficial molecular processes relevant to muscle function and structure. Notably, beneficial modules were enriched for genes involved in apoptosis regulation, calcium ion transport and muscle contraction, oxidative phosphorylation, and RNA processing, which are essential for maintaining muscle cell integrity and function. The enrichment of genes in these modules may underpin molecular mechanisms promoting muscle resilience in response to CNM-related stressors. Conversely, pathogenic modules were linked to immune system processes and structural disruptions in the muscle tissue, such as defects in vascularization and innervation. These insights align with recent findings of immune cells infiltration in CNM-affected murine muscle [3] and reports of neuromuscular junction defects in CNM patients and animal models [17,18,19]. However, vascularization defects have yet to be reported in CNM and may represent an intriguing area for future research to uncover their potential role in disease pathology. In terms of energy metabolism, one pathogenic module was enriched in genes linked to fatty acid oxidation. Excessive fatty acid oxidation has been linked to skeletal muscle atrophy in cancer cachexia, and dietary supplementation with nicotinamide or lipids has been shown to improve muscle health in this context [20,21]. This finding suggests the possibility of using dietary interventions, such as high-fat diets, as adjunctive therapies to augment muscular resilience in CNMs.

3.2. Implications of Multi-Omics Integration and Network Exploration

Our methodological framework represents an advance for CNM research, as previous analyses relied mainly on differential expression or single-omics comparisons. By integrating transcriptomic, proteomic, and metabolomic data into a multilayer network that also incorporates curated interactions, metabolic models, HPO relationships, and tissue specificity, our study provides the first cross-scale systems view of CNM pathogenesis. This integrative strategy captures relationships that are not detectable in single-layer datasets and allows the discovery of metabolites or genes that emerge from the network topology rather than from differential abundance alone. The multilayer network, comprising genes, metabolites, reactions, phenotypes, and tissue multiplex networks, facilitated the ranking of key metabolites and genes connected to XLCNM phenotypes, including muscle atrophy, muscle weakness, and abnormal nuclear centralization in muscle fibers. The presence of Myh7 and Mb (which are highly expressed in oxidative fibers) in the top-ranking genes suggests a heightened role of oxidative metabolism in XLCNM pathology, aligning with the WGCNA results on several CNM models discussed above. The emergence of inositol and phosphoinositide metabolites as top-ranked hits across multiple XLCNM phenotypes is consistent with the function of myotubularin, the protein encoded by Mtm1, but provides limited new insights into the disease. This finding differs from mechanisms already known in other myopathies and highlight the inositol and phosphoinositide pathway as a CNM-specific finding. On the other hand, other top-scoring metabolites may provide novel therapeutic opportunities to improve muscle phenotype. Indeed, dietary lutein has been associated with a reduction in oxidative stress and inflammation in rats, increased muscle fiber diameter in chickens, and may act through downregulation of MSTN, which is a negative regulator of muscle growth that is dysregulated in several myopathies. Several CNM models exhibit downregulation of MSTN, potentially reflecting a compensatory effort to limit muscle atrophy [3,22,23]. However, this MSTN downregulation may limit the potentially therapeutic benefits of further MSTN inhibition [24,25]. GGPP depletion has also been shown to regulate MSTN and cause skeletal muscle atrophy in mice, and mutations in GGPS1, encoding a GGPP synthase, have been linked to congenital myopathies [26,27]. Additionally, GGPP depletion has been linked to apoptosis and the reduction in the expression, localization, and GTP-binding ability of the Rho-GTPase RAC1 in cell studies [28]. This metabolite may thus represent a marker directly linked to CNM as RAC1 plays a role in recruiting MTM1 to plasma membrane ruffles, which are membrane protrusions implicated in actin cytoskeleton remodeling [29]. Finally, an increase in S-adenosylmethionine through a betain-supplemented diet has been shown to delay age-related muscle loss by promoting protein synthesis [30].

3.3. Study Limitations

The functional relevance of the identified metabolites would require experimental validation before they can be used as biomarkers or therapies. Also, as the results were obtained from in-house omics data and general databases knowledge, and as public datasets of omics data related to CNM are currently lacking, we could not proceed to an independent validation.

4. Materials and Methods

4.1. Transcriptomic Data

Bulk RNA-sequencing datasets from various mouse models were retrieved from previously published studies and are available online at NCBI’s GEO through the GEO Series accession number GSE160084 [3]. Additionally, the transcriptomic characterization of tamoxifen-treated Dnm2^S619L/+ and Bin1^mck−/− mice described in [11] have been deposited on NCBI’s GEO and are accessible through GEO Series accession number GSE282489. All studies were performed in tibialis anterior (TA) muscle. Detailed information about each cohort is available in Table 1.

4.2. Proteomic Data

Proteomic profiling of cohort A was retrieved from a previous study, and the data are available in the PRIDE repository under the identifier PXD021765 [3].

4.3. Metabolomic Data

Metabolomic profiling of quadriceps muscles from 5-week-old Mtm1^−/y and WT mice was carried out by Metabolon (Morrisville, NC, USA). Briefly, 50–130 mg of muscle and 150–300 μL of serum were flash-frozen and sent to Metabolon. Samples were prepared using the automated MicroLab STAR^® system from Hamilton Company (Reno, NV, USA). Ultrahigh-performance liquid chromatography–tandem mass spectroscopy (UPLC–MS/MS) was carried out with a Waters ACQUITY (Waters, Milford, MA, USA) ultra-performance liquid chromatography (UPLC) and a Thermo Scientific Q-Exactive (Thermo Scientific, Waltham, MA, USA) high resolution/accurate mass spectrometer interfaced with a heated electrospray ionization (HESI-II) source and Orbitrap mass analyzer operated at 35,000 mass resolution. Raw data were extracted, peak-identified, and QC processed using Metabolon’s hardware and software. Peaks were quantified using the area under the curve.

4.4. Weighed Gene Correlation Network Analysis

WGCNA was carried out using the bulk RNA-seq data from the 7 cohorts (Table 1). To construct the weighed gene correlation networks, genes with more than 10 reads in at least 90% of all samples were considered. Gene expression normalization was performed with DESeq2 [31]. The gene correlation network was built with the WGCNA R package (version 1.72-5) [9]. Briefly, the adjacency matrices were computed with a soft power value of 12. Topological overlap dissimilarity matrices were computed and used to perform hierarchical clustering of the genes. Gene modules were determined with dynamic tree cut and close modules (dissimilarity < 0.25) were merged. Module–trait Pearson correlations were computed, and the associated p-values were determined with Student’s t-test and adjusted with Bonferroni correction based on the number of traits.

4.5. Multi-Omics Integration with Knowledge from Public Databases

4.5.1. Biological Network Construction

The composition of the multiplex-heterogeneous biological network used to integrate transcriptomic, proteomic, and metabolomic data with publicly available knowledge is detailed below. The full network contains five multiplex networks (one or more layers connecting one type of node) and bipartite networks encoding links between the different multiplex networks. Unless specified otherwise, the layers are unweighted and undirected.

4.5.2. Description of the Multiplex Networks (Table 2)

• The first network is a gene multiplex network composed of five layers, each encoding different gene–gene interactions. Three layers were constructed based on publicly available knowledge: protein–protein interactions (PPIs) extracted from the STRING database (experimental confidence score > 700), molecular complex co-membership extracted from the CORUM database, and pathway co-membership extracted from the Reactome database [32,33,34]. The remaining two weighted layers were constructed with WGCNA of bulk transcriptomic and proteomic datasets from WT and Mtm1^−/y mice. For the transcriptomic layer, samples of cohorts (A–C) were used. After gene expression normalization and batch effect correction with DESeq2 and limma [35], the topological overlap matrix was computed using a soft power of 10, and edges with an adjacency > 0.15 were exported. The proteomic layer was built similarly. The wrMisc (version 1.15.4) and wrProteo (version 1.13.3) R packages were used to normalize and to impute missing data with default parameters. Then, the topological overlap matrix was computed using a soft power of 12, and edges with an adjacency > 0.25 were exported.
• The second network is a metabolite multiplex network, composed of two layers, with one layer encoding pathway co-membership extracted from the Reactome database, and one weighted layer built with WGCNA from metabolomics data (WT and Mtm1^−/y samples, soft power = 5, adjacency > 0.05).
• The third network is a directed monoplex network containing phenotypes described in the Human Phenotype Ontology (HPO) [36]. Directed edges linking HPO terms to their parents were retrieved using custom SPARQL queries on the OWL file provided by the HPO and loaded in Apache Foundation’s JENA suite (v4.7.0) (Supplementary Information S1).
• The fourth network is a monoplex network containing the reactions described in the mouse genome-scale metabolic model (Mouse-GEM) [37]. Reactions belonging to the same subsystem (sharing a similar metabolic function) are linked together.
• The fifth network is a monoplex network containing tissues found in the Genotype-Tissue Expression project (GTEx) [38] and is linked to other multiplexes through bipartite networks.
• Description of the bipartite networks (Table 3):
• The first network is a weighted gene–tissue bipartite network extracted from the GTEx project. This bipartite network links genes to the tissues in which they are expressed. Median transcript per million (TPM) values were used as edge weights, only keeping edges with TPM > 5.
• The next network is a gene–metabolite bipartite network extracted from Reactome, based on pathway co-membership [34].
• The next network is a metabolite–tissue bipartite network extracted from the Human Metabolome Database (HMDB), linking metabolites to the tissues in which they have been detected [39].
• The three bipartite networks extracted from the Mouse-GEM are as follows: two directed networks linking metabolites to reactions (reactants to reaction and reaction to products), and one network linking genes to the reactions in which the proteins they encode are implicated [37].
• The two bipartite networks extracted from the HPO are as follows: one network linking phenotypes to the genes associated with diseases in which the phenotype can be observed, and one network linking phenotypes to the tissues they affect. The second network was obtained using custom SPARQL queries on the OWL file provided by the HPO and loaded in Apache Foundation’s JENA suite (v4.7.0) (Supplementary Information S1).

4.6. Heterogeneous Multiplex Exploration

The heterogeneous multiplex was explored with random walk with restart (RWR) using the Python package MultiXrank [15] with global restart probability r = 0.7 and uniform inter-layers jump, layer restart, inter-multiplex networks jump, and multiplex network restart probabilities. In practice, this means that all layers had the same weight within each multiplex network, and that all multiplex networks had the same weight within the multilayer network.

4.7. Metabolite ID Mapping

To combine experimental metabolomic data, metabolite information extracted from the Mouse-GEM, Reactome, and HMDB, metabolites were mapped to their PubChem ID if possible. Metabolite ID mapping was performed programmatically by sending custom REST queries to the PubChem Identifier Exchange Service using cross-reference IDs as input. For the metabolomics data, mapping was performed based on cross-references (PubChem, HMDB, KEGG) and INCHI keys provided by Metabolon. For the Mouse-GEM data, cross-references provided with the GEM were combined with RECON3D [40] and MetaNetX [41] information to obtain ChEMBL, BioCyc, ChemSpider, KEGG, HMDB, LipidMaps, ChEBI cross-reference IDs and map them to PubChem IDs. If ID mapping was unsuccessful, original IDs were kept. For the Reactome database, mapping was performed on ChEBI IDs, and unmapped metabolites were discarded.

4.8. Network Visualization

Cytoscape (version 3.10.2) was used to visualize networks and sub-networks [42].

4.9. Enrichment Analyses

GO term enrichment analysis was performed with the R package ClusterProfiler (version 4.2.2), and enriched GO terms were clustered based on their semantic similarity [43,44].

5. Conclusions

This study underscores the value of multi-omics integration and network analysis for uncovering novel aspects of centronuclear myopathy pathogenesis. By linking gene expression modules to phenotypic traits, we identified both protective (as muscle contraction, RNA metabolism, and oxidative phosphorylation) and pathogenic molecular processes (i.e., immune response, innervation, vascularization, and fatty acid oxidation). In the Mtm1^−/y mouse model, multilayer network exploration highlighted not only phosphoinositide metabolism but also candidate metabolites such as lutein, geranylgeranyl diphosphate, and S-adenosylmethionine, which may represent avenues for dietary or pharmacological interventions. Collectively, these findings expand the molecular landscape of CNMs and provide prioritized targets for translational studies aimed at developing innovative therapeutic strategies.