Multi-omics tools stacking or genuine integration?

Introduction

The term "multi-omics" has become a must to have in grant applications, paper abstracts and biotech pitch decks. Mention two data layers — a transcriptome and a methylome, a proteome and a metabolome — and the study qualifies as integrative. The reality is often less straightforward.

In a significant fraction of publications claiming a multi-omics approach, integration limits to separate analyses of each data layer followed by a post hoc cross-referencing of gene or protein lists. A Venn diagram. A sentence in the discussion: "Interestingly, 47 genes were shared between the transcriptome and the proteome." The integration stops there.

This is not necessarily a lack of rigor — constraints of time, expertise, or sample size are real. But methodologically, it is closer to multi-analysis than to multi-omics in the strict sense.

The distinction is not semantic pedantry. It has direct consequences on the validity of biological conclusions. Rigorous multi-omics integration explicitly models the relationships between data layers — the co-variations, the shared and specific structures, the regulatory patterns that only emerge when multiple molecular levels are observed simultaneously. A collection of independent mono-omics analyses, however sophisticated each one may be, cannot capture these signals by design.

This article provides a framework to distinguish genuine integration from glorified list-crossing, reviews the main integration methods and their limitations, and asks a deliberately uncomfortable question: does a substantial share of what is published under the "multi-omics" label truly deserve the term?

Three levels of integration

The literature typically distinguishes three families of integration strategies. This classification, formalized in several reviews (Krassowski et al., 2020; Cai et al., 2022; Picard et al., 2021), reflects fundamental differences in how data layers interact within the analytical pipeline.

Late integration: each layer on its own

Late integration is the most common approach. Each omics layer is analyzed independently — differential expression for the transcriptome, differential methylation for the methylome, protein deregulation for the proteome — and results are cross-referenced after the fact.

Cross-referencing typically takes the form of list intersections (Venn diagrams), correlation matrices across layers, or overlay on knowledge bases (shared pathway enrichment).

It is the most accessible strategy. It requires no specialized integration method; each layer is processed with its standard toolkit. But it comes at a cost: quantitative relationships between layers are lost. Two genes may be flagged as differentially expressed in both the transcriptome and the proteome, but late integration says nothing about the consistency of direction or magnitude of these changes. Worse, it depends on arbitrary significance thresholds — a gene with an adjusted p-value of 0.049 in both layers makes the Venn; a gene at 0.051 in one layer vanishes, even if its joint signal would be biologically meaningful.

Late integration is not useless. But it cannot, by construction, reveal signals that only emerge from the simultaneous observation of multiple molecular layers.

Early integration: everything in one bag

Early integration takes the opposite approach. Data matrices from each omics layer are concatenated into a single matrix, which is then analyzed by standard methods (PCA, clustering, penalized regression).

The theoretical advantage is clear: all signals are available simultaneously, enabling the capture of cross-layer co-variations. In practice, this strategy runs into several major obstacles.

The first is the curse of dimensionality. Concatenating a transcriptome (20,000 genes), a methylome (400,000 probes) and a proteome (8,000 proteins) creates a matrix of over 428,000 variables for a few dozen to a few hundred samples. The signal-to-noise ratio collapses. Standard dimensionality reduction methods struggle to extract biologically interpretable signals from this space.

The second problem is scale heterogeneity. RNA-seq counts (integer, zero-inflated), methylation levels (beta-values bounded between 0 and 1) and proteomic intensities (continuous, log-normal) have neither the same units, nor the same distributions, nor the same noise profiles. A naive concatenation without rigorous normalization gives disproportionate weight to the most variable or highest-dimensional layer — often the methylome — drowning out signals from other layers.

The third problem is loss of group structure. By merging all variables into a single matrix, the information about each variable's origin is lost. Extracted variation axes (PCA components, clusters) indiscriminately mix within-layer technical variation and cross-layer biological variation. Nothing guarantees that the dominant signal captured reflects a biological co-regulation rather than a layer-specific artifact.

Intermediate integration: modeling the relationships

Intermediate integration (also called joint integration) represents the most accomplished level of integration. These methods explicitly model the multi-layer structure of the data: they seek to identify axes of variation shared across layers and those specific to each layer.

This is where genuine multi-omics integration happens.

The approaches fall into several families:

• Matrix factorization — MOFA (Argelaguet et al., 2018), MOFA+ (Argelaguet et al., 2020), iCluster, JIVE. These methods decompose multi-omics data into a small number of latent factors, some shared across layers, others specific. Shared factors capture biological processes that are coordinated across molecular levels.

• Network-based methods — SNF (Similarity Network Fusion, Wang et al., 2014) constructs a similarity network per omics layer and iteratively fuses them into a consensus network. The idea is elegant: each layer contributes to the global structure without any single one dominating.

• Multi-block analysis — DIABLO/mixOmics (Singh et al., 2019) seeks linear combinations of variables in each layer that maximize cross-layer covariance while predicting a phenotype of interest. It is one of the few approaches that combines integration with supervised classification.

• Deep generative models — totalVI (Gayoso et al., 2021) for CITE-seq data (RNA + surface proteins), MultiVI for multiome data (RNA + ATAC-seq). These methods jointly model the distributions of different modalities within a shared latent space.

What unites these approaches is a fundamental principle: the explicit modeling of cross-layer co-variation. They do not merely cross-reference results; they seek the biological processes that manifest simultaneously across multiple molecular levels.

Comparative overview of methods

Benchmarks (Chauvel et al., 2020; Herrmann et al., 2021) show that no method is universally superior. The choice depends on the biological question, cohort size, number of omics layers and the type of signal sought (subtypes, biomarkers, biological processes).

The Venn diagram is not integration

It is tempting to consider that a Venn diagram between differentially expressed genes in the transcriptome and deregulated proteins in the proteome constitutes multi-omics integration. It is the most commonly published approach. It is also the one whose limitations are least well understood.

Take a concrete example. A study compares tumor biopsies and healthy tissues. RNA-seq analysis identifies 1,200 differentially expressed genes (FDR < 0.05, |log2FC| > 1). Proteomic analysis identifies 350 deregulated proteins. The intersection contains 87 genes/proteins. The conclusion: "87 targets are confirmed at both transcriptomic and proteomic levels, suggesting coordinated regulation."

The problem is that this statement rests on at least three implicit assumptions, all debatable:

Assumption 1: the thresholds are biologically meaningful. A gene with an FDR of 0.04 in the transcriptome and 0.06 in the proteome is in the Venn. A gene with an FDR of 0.06 in both layers is not. Yet the second has a potentially more coherent signal than the first. The Venn transforms continuous distributions into binary categories, destroying quantitative information.

Assumption 2: directional concordance is guaranteed. A gene upregulated in the transcriptome but downregulated at the protein level appears in the intersection if both are "significant." The Venn does not check directional consistency. An active post-transcriptional regulation (accelerated protein degradation, for example) would produce exactly this pattern — but the conclusion of "coordinated regulation" would be wrong.

Assumption 3: absence of intersection means absence of relationship. Genes that do not appear in the intersection are ignored. Yet a transcriptionally stable gene whose protein is massively deregulated through post-translational modification is invisible in an expression-based Venn. Late integration cannot capture relationships that cross regulatory levels.

Intermediate integration methods solve all three limitations. MOFA, for instance, models continuous co-variations between layers without imposing thresholds. DIABLO seeks variable combinations that are simultaneously discriminant across multiple layers. None of these methods reduces data to binary lists.

When integration genuinely changes the conclusions

Multi-omics integration is not justified for its own sake. It is justified when it reveals a biological signal that is not accessible through separate analysis of the layers. Here are concrete situations where formal integration changed the conclusions.

Patient stratification: the pilocytic astrocytoma example

Picard et al. (2023) used SNF to integrate transcriptomic and proteomic data from 62 patients with pilocytic astrocytoma — the most common pediatric brain tumor. Separate analysis of each layer had failed to identify biologically relevant subgroups. Fusing the two similarity networks revealed two distinct groups, validated in three independent cohorts, with divergent immune and neuronal profiles and significant differences in progression-free survival.

This result is characteristic of what integration can deliver: a signal too weak in each individual layer, but that emerges when layers reinforce each other.

Uncovering hidden variation factors: MOFA on chronic lymphocytic leukemia

The foundational MOFA study (Argelaguet et al., 2018) applied multi-omics factorization to 200 patients with chronic lymphocytic leukemia (CLL), profiled for somatic mutations, RNA expression, DNA methylation and drug responses. MOFA identified shared variation factors, including IGHV status and trisomy 12 (known), but also an oxidative stress response axis that had not been identified by mono-omics analyses.

The value of factorization is that it distinguishes sources of variation common to all layers (coordinated biological processes) from those specific to a single layer (technical noise or local regulations). This decomposition is impossible with late integration.

Post-transcriptional regulation invisible in the transcriptome alone

One of the strongest arguments for multi-omics integration concerns the transcriptome-proteome discordance. Numerous studies have documented that the correlation between mRNA and protein levels is modest. This discordance reflects post-transcriptional regulations (mRNA stability, translational efficiency, protein degradation) that are only detectable when both layers are observed simultaneously.

A late integration approach (Venn) cannot identify a gene whose mRNA is stable but whose protein is massively degraded. Only a joint modeling of mRNA-protein co-variation can reveal these regulations.

When multi-omics adds nothing

It is equally important to recognize situations where multi-omics does not improve — or even degrades — conclusions compared to a simpler approach.

The TCGA benchmark: a result that warrants caution

Herrmann et al. (2021) conducted the most extensive benchmark to date on the predictive utility of multi-omics. Across 18 TCGA cancer datasets (35 to 1,000 patients, up to 100,000 variables), they compared 11 multi-omics integration methods for survival prediction. The result is unambiguous: only one method (block forest) outperformed a Cox model based solely on clinical variables, and only marginally.

This does not mean multi-omics is useless. It means that, in this specific context (survival prediction, TCGA data), clinical variables (age, stage, grade) already carry most of the prognostic information. Adding omics layers primarily introduced noise and overfitting.

The lesson is direct: adding data layers only adds value if they contain information complementary to what is already available. When clinical variables already explain most of the variance in the outcome of interest, multi-omics cannot improve prediction — it can only add noise.

The "more data = better" trap

The intuition that more data leads to better analyses is deeply ingrained. In multi-omics, it is often wrong. Each additional layer brings not only potential signal but also noise, missing data, layer-specific batch effects, and an exponential increase in variable space. The signal-to-noise ratio can degrade with the addition of layers if they are redundant, noisy or poorly integrated.

Han et al. (2025) showed on TCGA data that feature selection was the single most important factor for multi-omics clustering performance, improving subtype discrimination by 34%. The quality of feature selection matters more than the quantity of omics layers.

Experimental design: the mistakes that doom integration

The most costly errors in multi-omics are not computational. They happen at the study design stage.

The imperfect pairing problem

Multi-omics integration rests on a fundamental assumption: data layers come from the same biological samples, or at least from samples similar enough that cross-layer co-variations reflect biological processes rather than inter-sample differences.

In practice, this assumption is frequently violated. The transcriptome is measured on one aliquot, the proteome on another, the methylome on a third. If aliquots are not homogeneous (which is common for heterogeneous tissue biopsies), observed co-variations between layers partly reflect intra-sample heterogeneity rather than biology.

Single-cell multimodal technologies (CITE-seq, multiome RNA + ATAC) partially solve this problem by measuring multiple modalities within the same cell. But at the bulk level, the pairing question remains critical and under-discussed.

Sample sizes incompatible with dimensionality

The classic "n << p" problem (far more variables than samples) is amplified in multi-omics. If a study with 50 patients and 20,000 genes is already at the limit for RNA-seq alone, the same study with 3 omics layers and 100,000+ combined variables is in a statistically perilous situation.

Intermediate integration methods (MOFA, DIABLO) manage this dimensionality through sparsity constraints, but they do not work miracles. A small multi-omics dataset will produce unstable results, highly dependent on preprocessing choices and hyperparameters.

Cross-layer batch effects

Each omics layer is generated by a different technology, often in different laboratories, at different times. Within-layer batch effects are a well-known problem (see our previous article on batch effects in scRNA-seq). But cross-layer batch effects are less discussed and potentially more insidious.

If the transcriptome is sequenced in January and the proteome measured in March, technical variations between these two timepoints confound with cross-layer variation. An integration method may interpret this technical variation as biological signal — an apparent transcriptome-proteome discordance that is in reality a temporal artifact.

Checklist: is your study truly multi-omics?

At the design stage

• Does the biological question require observation of multiple molecular layers, or would a thorough mono-omics analysis suffice?
• Does each omics layer bring complementary (not merely redundant) information?
• Are samples paired across layers (same biological specimen)?
• Is the cohort large enough to support the combined dimensionality of the layers?
• Will layers be generated under controlled conditions to minimize cross-layer batch effects?

At the analysis stage

• Are you using a method that explicitly models relationships between layers (intermediate integration), or are you merely cross-referencing results from separate analyses (late integration)?
• If you are using a Venn diagram, have you verified directional concordance of changes?
• Have you quantified what integration adds compared to separate layer analysis (in terms of explained variance, clustering, prediction)?
• Are results robust to preprocessing choices and hyperparameters?

At the interpretation stage

• Do your conclusions depend on integration (they would not have been possible otherwise), or could they have been drawn from a single layer?
• Does the word "multi-omics" genuinely describe your methodology, or is it dressing up a collection of mono-omics analyses?
• Have you compared your integrated results to a simple reference model (clinical variables alone, most informative mono-omics analysis)?

Conclusion

Multi-omics is a powerful tool — when used for what it is: an integrative approach that models relationships between molecular levels. When it amounts to a collection of separate analyses under a common label, it does not leverage its full potential. And it can create an impression of comprehensiveness that does not reflect the reality of the analysis.

This is not an indictment of teams working under constraints. Formally integrating multiple omics layers requires time, specialized skills, and adequate sample sizes — resources that are not always available. But recognizing this distinction enables a more accurate evaluation of what each study can, and cannot, conclude.

The choice between stacking omics tools and genuine multi-omics is not a technical choice. It is a scientific one. It starts with a clear biological question, continues with an experimental design that enables integration, and materializes through methods that explicitly model cross-layer co-variation.

Sometimes, the most honest answer is that the biological question only warrants a single omics layer, analyzed rigorously. And that answer is more valuable than a superficial integration.