Analysis
MOFA is implemented in BiomiX following developer guidelines (biofam.github.io/MOFA2), with transformed data inputs and reduced feature size. Transformation methods depend on the omics type: transcriptomics data use variance-stabilizing transformation, metabolomics data use log transformation to approximate Gaussian distribution, and methylomics data require no transformation. Top genes and CpG islands with the highest variance are selected as integration input, except for metabolomics where fewer peaks are typically included.
BiomiX's MOFA implementation allows customization of convergence speed, Evidence Lower Bound frequency, and maximum iterations. An automated tuning mode optimizes the total factor count, stopping when three consecutive models produce a factor explaining less than 1% of variance. The top three models, prioritizing those with more statistically significant discriminant factors, are selected. Each model's factors are tested with a Mann-Whitney test, with p-values adjusted by the FDR method.
Factor interpretation
Both automatic and manual MOFA analyses include three methods to support interpretation of discriminating factors.
Correlation analysis
Users can upload a matrix of binary or numerical clinical features to integrate into the MOFA model. Numerical data are correlated with each factor using Pearson correlation; binary data are analyzed using the Wilcoxon test after dividing groups into positive and negative categories. Nominal p-values are corrected using the Benjamini-Hochberg method.
Pathway analysis
BiomiX identifies the top contributing genes, metabolites, and CpG islands for each discriminating factor. Genes are analyzed with EnrichR using Reactome, Biological Process, ENCODE, and ChEA transcription factor libraries. Metabolites are assessed via MetPath using the KEGG and HMDB databases. CpG islands linked to associated genes are also analyzed using EnrichR.
PubMed bibliography search
For each discriminating factor, top contributing features are used for PubMed searches across three priority levels: abstracts combining contributors from multiple omics are prioritized first. The output is a TSV file with PubMed articles, contributor matches, DOIs, and keywords. BiomiX also performs text-mining with litsearchr v1.0.0 to generate a vocabulary of frequent word combinations filtered against Gene Ontology and human phenotype terms.
Analysis
SNF constructs a patient similarity network for each omics layer independently, then iteratively fuses them into a single integrated network that captures complementary cross-modal information. Spectral clustering is then applied to the fused network to identify patient subgroups.
In BiomiX 3.0, SNF is implemented via the SNFtool R package. The number of clusters is selected automatically via the eigen-gap heuristic, but can also be set manually. Cluster quality is evaluated using the Average Silhouette Width (ASW), Adjusted Rand Index (ARI), Adjusted Mutual Information (AMI), and Calinski-Harabasz (CH) index.
SNF requires complete data across all omics layers. Samples with missing data in any layer are excluded before analysis. If your dataset has partial overlap between omics, consider NEMO instead.
Analysis
NEMO is specifically designed for datasets where samples are not fully overlapping across omics layers, without requiring imputation. It constructs affinity scores from nearest-neighbor relationships within each omics layer, using only the available modalities for each sample.
In BiomiX 3.0, NEMO is implemented via the NEMO R package. Like SNF, cluster quality is assessed using ASW, ARI, AMI, and CH metrics, and the number of clusters can be determined automatically or set manually.
NEMO is the recommended choice when your omics layers do not share the same set of samples — for example, when transcriptomics and methylomics were profiled on different subsets of the same cohort.
Analysis
DIABLO is a supervised integration method from the MixOmics framework that extends sparse canonical correlation analysis to multiple omics layers. It identifies correlated latent components across modalities while optimizing discrimination between predefined groups, making it well-suited for multi-omics biomarker discovery.
In BiomiX 3.0, DIABLO components are selected automatically via Mann-Whitney test with FDR correction, mirroring the MOFA factor selection strategy. Significant components are then interpreted through pathway enrichment and clinical correlation, using the same interpretation pipeline as MOFA.
DIABLO requires a known group label (condition) for each sample. Unlike the unsupervised methods, it uses this label to maximize between-group discrimination, which makes it particularly powerful for biomarker discovery but requires a clear experimental design.
Analysis
PRAMIGO is a novel supervised Heterogeneous Graph Transformer (HGT) that models cross-modal molecular interactions explicitly. It constructs a heterogeneous graph where nodes represent both samples and molecular features, and edges capture within-omic, between-omic, and sample-feature relationships.
Unlike other integration methods that summarize variation at the sample level, PRAMIGO uncovers biologically interpretable multi-omics programs and feature-level interactions — including ncRNA hubs, immunoglobulin repertoire diversity, and epigenetic co-regulation programs — that are inaccessible to conventional integration approaches.
PRAMIGO is currently available as a standalone tool at github.com/iclemente99/PRAMIGO and will be integrated into BiomiX in a future release.
Which method should I use?
Each method is designed for a different analytical objective. When in doubt, run multiple methods and compare results.
| Method | Choose it when... |
|---|---|
| MOFA | Your goal is multi-dimensional coverage of independent biological axes and you do not have a predefined group label, or you want to explore latent structure across omics without supervision. |
| SNF | Your goal is patient stratification and all samples are profiled across all omics layers (no missing samples between layers). |
| NEMO | Your goal is patient stratification but your omics layers do not share the same set of samples — NEMO handles partial overlap natively. |
| DIABLO | Your primary goal is biomarker discovery with cross-omic co-variation, and you have a known condition label for each sample. |
| PRAMIGO | Your goal is discovering cross-modal regulatory mechanisms, especially structural multi-omics features such as ncRNA hubs or epigenetic co-regulation programs. Soon |