iModMix: integrative module analysis for multi-omics data

Abstract

Summary

Integrative Module Analysis for Multi-omics Data (iModMix) is a biology-agnostic framework that enables the discovery of novel associations across any type of quantitative abundance data, including but not limited to transcriptomics, proteomics, and metabolomics. Instead of relying on pathway annotations or prior biological knowledge, iModMix constructs data-driven modules using graphical lasso to estimate sparse networks from omics features. These modules are summarized into eigenfeatures and correlated across datasets for horizontal integration, while preserving the distinct feature sets and interpretability of each omics type. iModMix operates directly on matrices containing expression or abundances for a wide range of features, including but not limited to genes, proteins, and metabolites. Because it does not rely on annotations (e.g., KEGG identifiers), it can seamlessly incorporate both identified and unidentified metabolites, addressing a key limitation of many existing metabolomics tools. iModMix is available as a user-friendly R Shiny application requiring no programming expertise (https://imodmix.moffitt.org), and as a Bioconductor R package for advanced users (https://bioconductor.org/packages/release/bioc/html/iModMix.html). The tool includes several public and in-house datasets to illustrate its utility in identifying novel multi-omics relationships in diverse biological contexts.

Availability and implementation

iModMix is freely available from Bioconductor (https://bioconductor.org/packages/release/bioc/html/iModMix.html), and the example dataset package (iModMixData) is also available from Bioconductor (https://bioconductor.org/packages/release/ data/experiment/html/iModMixData.html). The R package source code and Docker are available from GitHub: https://github.com/biodatalab/iModMix. Shiny application can be accessed at: https://imodmix.moffitt.org.

1 Introduction

Integration of multiple omics modalities (i.e., “multi-omics”) is essential for a comprehensive understanding of complex biological systems and disease mechanisms. The scale and complexity of multi-omics data necessitate computational approaches that are adaptable to diverse experimental designs and data types, while remaining accessible to users without extensive programming experience. A key challenge in this space is the integration of untargeted metabolomics data, where a large fraction of detected metabolites are unidentified (Jendoubi 2021). Current pathway analysis tools generally require standardized identifiers to map features to known biological pathways, and therefore, they inherently cannot incorporate unidentified metabolites. Even for identified compounds, inconsistent naming conventions and incomplete annotation (e.g., missing KEGG identifiers) often prevent their use in pathway-based analyses. As a result, substantial portions of metabolomics data are excluded, limiting downstream analysis and reducing the potential for functional interpretation of results (Jendoubi 2021).

In general, approaches for integrating multi-omics can be categorized into vertical or horizontal integration. Vertical integration takes multiple omics datasets as input and produces a single output that no longer reflects the individual features of the original data, such as clustering assignments or a merged dataset. iClusterPlus (Qianxing et al. 2013) is an example of vertical integration: it takes multiple omics datasets from the same samples as input, extracts latent factors across omics datasets, clusters samples in latent space, and outputs the clustering assignments by sample. Another example of vertical integration is Non-negative Matrix Factorization (NMF) (Brunet et al. 2004), which is incorporated into tools like intNMF (Chalise and Fridley 2017), a tool that uses integrative NMF to cluster samples based on multiple genomic datasets. In contrast, horizontal integration analyzes multiple omics datasets in parallel to maintain the context of the original inputs. PIUmet (Pirhaji et al. 2016) is an example of horizontal integration: it takes metabolomics and proteomics data as input, and the resulting network output contains both metabolites and proteins.

iModMix is a horizontal integration framework that constructs network modules from any combination of input omics datasets (e.g., transcriptomics, proteomics, metabolomics). It first uses graphical lasso to estimate a sparse Gaussian Graphical Model (GGM) (Friedman et al. 2008) for each input dataset. GGMs capture direct associations within the input omics datasets, in contrast to Weighted Gene Correlation Network Analysis (WGCNA) (Langfelder et al. 2008a), which is based on pairwise correlations and therefore includes both direct and indirect associations. Next, a Topological Overlap Matrix (TOM) is calculated (Yip et al. 2007) to quantify the extent to which pairs of features share common neighbors. Hierarchical clustering is then performed on TOM dissimilarity to group related features into modules. iModMix then finds the first principal component (PC) of the abundances of the features in the module, called an eigenfeature. These eigenfeatures can be used as abundance measurements for each module and therefore can be used for testing for differential expression or association with experimental conditions at the module level. Finally, integration between omics types is achieved by correlating the eigenfeatures from different omics sets.

A major advantage of iModMix is its ability to generate network modules and associated eigenfeatures independent of feature annotation. In the case of metabolomics, iModMix incorporates both identified and unidentified metabolites into the analysis, overcoming a common limitation of many existing metabolomics tools. Although we emphasize metabolomics here, this same approach can be applied to other omics types where the identity and/or function of some features are unknown, such as long non-coding RNAs in transcriptomics. Because iModMix does not rely on existing pathway databases like KEGG, it also supports the discovery of novel associations between features. To highlight its functionality, we have applied iModMix to two case studies (see supplementary file). The first uses public data on clear cell renal cell carcinoma (ccRCC, RC20 dataset) to validate the results (Case Study S1), while the second demonstrates the utility of iModMix in analyzing untargeted metabolomics data that includes unidentified features (Case Study 2, available as supplementary data at Bioinformatics  online, Table 1, available as supplementary data at Bioinformatics  online, Fig. 1, available as supplementary data at Bioinformatics online).

2 Implementation

2.1 Data pre-processing

Users can integrate any number of omics datasets simultaneously with the iModMix R package, while the Shiny application (https://imodmix.moffitt.org) supports up to three omics datasets (Fig. 1A). A separate metadata file containing sample IDs and at least one sample grouping column (e.g., treatment, control) is necessary for conducting PCA, heatmaps, and boxplots. Users can provide feature-level annotation corresponding to the input omics dataset. iModMix calculates the missing values in each row and filters out rows with missing values in >10% of the columns. Based on benchmarking, it is recommended to use a matrix size of 5000 features for optimal analysis speed. iModMix then calculates the standard deviation for each column and filters out columns with a standard deviation below the 25th percentile. Missing values are imputed using the k-nearest neighbors’ algorithm, or users can also provide their own imputed data if a different method is preferred. Finally, the data is scaled by centering each feature to have a mean of zero and a standard deviation of one. In addition to the case studies, we provide ten multi-omic datasets (transcriptomics and metabolomics) in the Shiny application to further illustrate iModMix’s capabilities. These datasets (Benedetti et al. 2023) have been formatted to meet iModMix specifications, with variable columns labeled as “Feature_ID” and sample columns labeled as “Samples”. These datasets are readily accessible for download on Zenodo (https://doi.org/10.5281/zenodo.13988161). Furthermore, we provide an in-house dataset of 20 lung adenocarcinoma (LUAD) mouse samples (10 wild type, 10 knockout), with 7353 metabolomics features and 7928 protein groups, to highlight iModMix’s functionality in handling unidentified metabolites. Analysis of these example datasets are described in more detail in the supplementary file.

Figure 1.

Overview of iModMix. (A) iModMix integrates multi-omics and clinical data to understand biological systems comprehensively. (B) Co-expression network: The tool performs independent analyses of each omics dataset to identify biologically relevant modules, and it then integrates the datasets by correlating eigenfeatures. (C) iModMix evaluates associations between phenotype-related modules across different data types. (D) Module assignments and eigenfeatures for each dataset are calculated, and phenotypes are compared. A module network detailing interactions between all datasets is generated along with the correlation plot for visualization.

2.2 Module construction

iModMix first estimates partial correlations using the glassoFast R package (Friedman et al. 2008) (Fig. 1B) to build a GGM. The regularization parameter lambda was set to a default value of 0.25 after benchmarking (see benchmarking section). The function also allows users to input their desired value for lambda using the iModMix R package. The user has the option to upload precalculated partial correlations for further downstream analysis using the iModMix package. Next, the TOM is computed and TOM dissimilarity (1—TOM) is used for hierarchical clustering (Yip et al. 2007). Modules are created using the dynamic tree cut method, specifically the “hybrid” method, which refines assignments derived from a static cut height by analyzing the shape of each branch (Langfelder et al. 2008b). Finally, iModMix calculates eigenfeatures for each module (Fig. 1B) for downstream omics analyses.

2.3 Module exploration

The phenotype analysis section provides a framework for classifying phenotypes based on eigenfeatures. Users can upload metadata, select phenotypes of interest, and set significance thresholds for statistical tests. Significant eigenfeatures are visualized through boxplots. Additionally, users can explore specific modules and view associated features, PCA loading, and heatmap plots to view feature behavior across different phenotypes (Fig. 1C). iModMix supports pathway enrichment analysis of genes or proteins by utilizing all available libraries in Enrichr if gene symbols are provided (Kuleshov et al. 2016).

2.4 Multi-omic analysis

iModMix integrates modules across all available omics datasets by calculating the Spearman correlation between eigenfeatures for all possible combinations. This is followed by construction of an interactive network, where modules from Data 1 are represented as yellow diamonds, modules from Data 2 as green triangles, modules from Data 3 as blue circles, and correlation coefficients on connecting arrows. Users can select the number of top multi-omics module correlations to visualize and explore detailed correlations within these modules. Lists of metabolites and proteins/genes within highly correlated modules are provided, facilitating pathway analysis and offering insights into the relationships between different omics layers. Additionally, classification between features of each layer and phenotypes using t-tests and boxplots is provided, enabling users to visualize and identify differences (Fig. 1B).

2.5 System requirements

iModMix can run on standard computers with > 8GB of RAM and a multi-core processor. The running time depends on data size and complexity. Typical data sizes of 20 000 variables can be completed in around 10 min (Fig. 2, available as supplementary data at Bioinformatics online).

3 Benchmarking

In iModMix, the regularization parameter lambda (λ) controls the sparsity of the partial correlation matrix estimated using the glasso package (Friedman et al. 2008). This matrix determines which direct associations between features are retained and which are removed. Because the network structure derived from these partial correlations forms the basis for downstream module detection, λ is a critical parameter. If λ is too high, the resulting network is overly sparse and may not capture meaningful relationships. If λ is too low, the network retains too many weak associations, which can lead to large, uninformative modules or overly fragmented module structures, depending on the clustering method used. To identify a practical default, we benchmarked iModMix on datasets from clear cell renal cell carcinoma (ccRCC, RC20 dataset) (Benedetti et al. 2023) and an internally generated lung adenocarcinoma dataset. For each dataset, we applied the iModMix pipeline and got a λ = 0.25 for consistently balanced speed and module quality (Fig. 3, available as supplementary data at Bioinformatics online). This value is set as the default in iModMix Shiny app. Advanced users can modify this parameter in the R package to suit their data.

iModMix was also compared directly with WGCNA using the lung adenocarcinoma (LUAD) data. For metabolomics data, WGCNA generated 9 modules (Mean size = 817, SD = 2206), whereas iModMix produced 287 modules (Mean size = 23, SD = 11.2). For proteomics data, WGCNA generated 22 modules (Mean size = 360, SD = 559), while iModMix identified 412 modules (Mean size = 17, SD = 5.75) (Table 2, available as supplementary data at Bioinformatics  online, Fig. 4, available as supplementary data at Bioinformatics online). WGCNA yielded a wide range of module sizes with one or two large modules comprising most features, whereas iModMix modules were more uniform in size (Figs 3 and 5, available as supplementary data at Bioinformatics online). Smaller module sizes generated from iModMix resulted in more comprehensible grouping for pathway enrichment, identifying 181 pathways from protein modules including all 13 pathways identified by WGCNA (Fig. 6, available as supplementary data at Bioinformatics  online, Table 3, available as supplementary data at Bioinformatics online). The correlation in the integration process between metabolite and protein modules was higher with iModMix (0.93) compared to WGCNA (0.88) (Fig. 7, available as supplementary data at Bioinformatics online). This benchmarking process demonstrates that iModMix generates more evenly sized modules with less features per group by using only direct associations for network construction, thus allowing easier interpretation and downstream analysis. A comparative overview of the underlying principles and analytical frameworks of iModMix, PIUMet, iClusterPlus, and WGCNA is presented in the Table 4, available as supplementary data at Bioinformatics online.

4 Summary

The iModMix pipeline provides an empirical framework for uncovering novel associations between multi-omics data, with the unique capability to incorporate both identified and unidentified metabolites. By generating de novo network modules independent of pathway databases, iModMix enables the discovery of previously unrecognized interactions between features. Unlike methods that rely on predefined biological functions or annotations, iModMix leverages feature matrices alone, making it highly adaptable to a wide range of omics combinations, including those with limited annotation or poorly characterized interactions.

Author contributions

Isis Yanina Narvaez-Bandera (Conceptualization [equal], Data curation [lead], Formal analysis [equal], Investigation [equal], Methodology [equal], Software [lead], Validation [equal], Visualization [equal], Writing—original draft [lead], Writing—review & editing [lead]), Ashley Lui (Conceptualization [equal], Formal analysis [equal], Investigation [equal], Methodology [equal], Software [equal], Validation [equal], Visualization [equal], Writing—original draft [equal], Writing—review & editing [equal]), Yonatan Ayalew Mekonnen (Software [equal], Validation [equal], Writing—review & editing [equal]), Vanessa Rubio (Conceptualization [equal], Formal analysis [equal], Methodology [equal], Validation [equal], Writing—review & editing [equal]), Augustine Takyi (Validation [supporting]), Noah Sulman (Conceptualization [equal], Data curation [equal], Resources [equal], Supervision [equal], Validation [equal], Writing—review & editing [equal]), Christopher Wilson (Conceptualization [equal], Software [equal], Writing—review & editing [equal]), Hayley D. Ackerman (Data curation [equal], Investigation [equal], Writing—review & editing [equal]), Oscar Ospina (Data curation [supporting], Software [supporting], Visualization [supporting], Writing—review & editing [equal]), Guillermo Gonzalez-Calderon (Software [supporting], Validation [equal], Visualization [supporting]), Elsa Flores (Conceptualization [equal], Formal analysis [equal], Funding acquisition [equal], Investigation [equal], Methodology [equal], Project administration [equal], Resources [equal], Supervision [equal], Validation [equal], Writing—review & editing [equal]), Qian Li (Conceptualization [equal], Formal analysis [equal], Methodology [equal], Supervision [equal], Writing—review & editing [equal]), Ann Chen (Conceptualization [equal], Formal analysis [equal], Investigation [equal], Methodology [equal], Supervision [equal], Writing—review & editing [equal]), Brooke Fridley (Conceptualization [lead], Formal analysis [equal], Funding acquisition [lead], Investigation [equal], Methodology [equal], Project administration [lead], Resources [lead], Software [equal], Supervision [equal], Writing—review & editing [equal]), and Paul Allen Stewart (Conceptualization [lead], Formal analysis [equal], Funding acquisition [lead], Investigation [equal], Methodology [equal], Project administration [lead], Resources [lead], Software [equal], Supervision [equal], Validation [equal], Visualization [equal], Writing—review & editing [equal])

Supplementary material

Supplementary material is available at Bioinformatics online.

Conflict of interests

None declared.

Funding

This work was supported by the NCI Research Project Grant [P01CA25084], Biostatistics and Bioinformatics Shared Resource and the Proteomics and Metabolomics Shared Resource at the H. Lee Moffitt Cancer Center & Research Institute [P30CA076292], the Cancer Bioinformatics Shared Resource at Huntsman Cancer Institute at the University of Utah [P30CA042014], the Cancer Research Institute Technology Impact Award, NIH T32CA233399, NIH U54HD090258, and the Anna D. Valentine and Charles L. Oehler Award. The content is those of the authors and does not represent the NIH.

Data availability

The data underlying this article are available in the article and in its online supplementary material.