multiRF-Cluster — Cross-Modal Multi-Omics Integration

§01 · Summary

Separating shared and modality-specific structure in multi-omics

Abstract

Multi-omics studies are widely used across biomedical research. In many diseases, some signals are shared across data types, while others are strongest in a single omics layer. Existing integrative clustering methods either merge all data types into a single representation — which can blur biology that is strong in one layer — or rely on linear structure that misses more complex relationships. We introduce multiRF, a random-forest-based method that handles complex data types and separates shared from modality-specific structure. multiRF learns sample similarities across omics layers from multivariate random forests, combines them across data types, and uses the resulting weights to estimate the part of each omics layer predictable from the others. The remaining residual is treated as modality-specific signal, so shared and modality-specific similarities can be clustered separately. In simulations multiRF recovered shared clusters as well as or better than established integrative methods and more reliably separated modality-specific signal under nonlinear structure. In TCGA head-and-neck cancer the shared component aligned with established subtype classifications, while gene- and miRNA-specific residuals revealed additional immune and developmental biology. In ADNI, the shared cross-modal aging signal between blood DNA methylation and structural MRI tracked future conversion to MCI or Alzheimer's disease.

§02 · Framework

The multiRF clustering workflow

The method takes matched multi-omics samples and produces a pair of similarity matrices for each layer — one capturing what that layer shares with the others, and one capturing what is unique to it. These are then handed to a downstream clustering algorithm of the user's choice.

i.

Matched multi-omics input

Two or more omics layers on the same set of samples (e.g. gene expression + miRNA, or blood DNAm + structural MRI).

ii.

Cross-layer forests

For each ordered pair of layers, fit a multivariate random forest predicting one layer from another. Terminal-node co-occurrence yields a sample-by-sample similarity.

iii.

Cross-modal weights

Combine the pairwise similarities and use the resulting weights to estimate the part of each layer that is predictable from the others.

iv.

Cluster shared & residual

The predictable part defines the shared similarity; the residual defines the modality-specific similarity. Each is clustered separately.

Figure 1. Four-step multiRF pipeline. Forest-based similarities are a natural fit for cross-modal integration because they capture nonlinear structure without forcing a common low-dimensional embedding before clustering.

§03 · Decomposition

What each similarity matrix captures

For every omics layer multiRF returns a pair of similarities. Reading them side by side lets a downstream analysis distinguish a common disease axis from biology that is genuinely unique to one data type — instead of collapsing the two:

Component A

Shared

Sample similarities reproducible from the other layers via the cross-modal weights. Captures the common cross-modal structure.

Component B

Weights

Per-layer contributions estimated from the multiRF similarities. Indicate how much each layer is explained by the others.

Component C

Residual

The part of a layer's similarity that is not predictable from the others. Carries modality-specific signal for separate clustering.

Note. The decomposition is symmetric across layers — every omics block gets its own (shared, residual) pair — so a study with three or more layers does not need a privileged "anchor" modality.

§04 · Simulation

Recovery of shared clusters under linear and nonlinear structure

We simulated matched multi-omics data with a known set of shared sample clusters plus modality-specific structure that should remain in each layer's residual. Cluster recovery on the shared similarity was scored by Adjusted Rand Index (ARI) against the ground-truth labels. multiRF holds its own under linear structure and pulls clearly ahead under nonlinear structure, where established integrative methods degrade:

Figure 2

Shared-cluster recovery (ARI) across simulation regimes

Figure 2. Mean ARI for multiRF, iClusterPlus, SNF, MoCluster, and IntNMF on the shared cluster recovery task. Under linear data structure multiRF (~0.82) is competitive with the best linear methods; under nonlinear structure linear-assumption methods drop sharply while multiRF retains ARI ≈ 0.74.

Shared clusters. multiRF recovers the planted shared clusters as well as or better than established integrative methods across regimes.
Modality-specific signal. The residual similarity in each layer reliably retains structure that is unique to that layer, even when other methods absorb it into a single merged representation.
Nonlinearity. The advantage is largest under nonlinear cross-modal relationships, which is the regime in which linear-projection methods are weakest.

§05 · TCGA-HNSC

Shared subtype axis + modality-specific biology in head-and-neck cancer

We applied multiRF to TCGA head-and-neck squamous cell carcinoma (HNSCC) with matched gene expression and miRNA. The shared component aligned with the main subtype structure recognised across established HNSCC reference classifications, while the gene-specific and miRNA-specific residuals exposed additional biology that would otherwise be hidden inside a merged representation:

Component	What it captures	Biology surfaced
Shared	Cross-modal HNSCC axis	Aligns with the main subtype structure described in established HNSCC reference classifications.
Gene-specific	Expression-only residual	Surfaces additional immune-related biology not visible after a one-shot merge of expression and miRNA.
miRNA-specific	miRNA-only residual	Surfaces additional developmental biology that is strong in miRNA but weak in expression.

Table 1. Reading the three multiRF outputs side-by-side recovers a common HNSCC subtype axis while preserving biology that is unique to a single data type — addressing a known failure mode of integrate-then-cluster pipelines.

§06 · ADNI

Cross-modal aging signal between blood DNAm and brain MRI

In the ADNI cohort we paired blood DNA methylation with structural MRI on the same participants. The shared cross-modal similarity captured an aging axis that travelled between the two tissues; the DNAm-specific residual carried exploratory signal that complemented it. Risk stratification by the shared component picked up future conversion to mild cognitive impairment or Alzheimer's disease:

Shared

Cross-modal aging axis

DNAm residual

Exploratory

MRI residual

Structural-only

Figure 3

Conversion-free survival by multiRF shared-axis score

Figure 3. Kaplan–Meier-style curves (simulated for visualisation) illustrating separation of high- vs. low-risk groups defined on the multiRF shared cross-modal aging signal between blood DNAm and brain MRI.

multiRF — shared axis
MCI / AD ↑
The shared cross-modal aging signal was associated with future conversion to mild cognitive impairment or Alzheimer's disease.

DNAm residual

Exploratory

The DNAm-specific residual carried additional information beyond the shared cross-modal axis — useful for hypothesis-generating follow-up.

§07 · Citation

How to cite

Zhang et al., 2026 (preprint)

Zhang, W., Wang, L., Franzmann, E. J., & Chen, X. S. (2026). Multivariate Random Forests for Cross-Modal Multi-Omics Integration. bioRxiv. https://doi.org/10.64898/2026.06.17.732933.

Project repository at github.com/TransBioInfoLab/multiRF-cluster; the underlying R package is at github.com/novawz/multiRF with a step-by-step vignette at novawz.github.io/multiRF.