Stack expression, methylation, mutation, and copy-number panels on a shared sample axis. Cluster the columns. Read what the subtypes converge on.
A gene can be inactivated by mutation, promoter hypermethylation, deletion of its chromosome region, or overactivation of its negative regulator. A multi-omics heatmap stacks these modalities on one shared column axis (patients) and reads the patterns across layers.
Two ideas drive the figure. Layers share a column axis: every panel indexes the same patients in the same order, so each column is a multi-modal profile. Each row is normalized within its own measurement scale before it is drawn. Expression is not comparable to methylation in raw units, but their aligned patterns can still be read side by side.
The column ordering is the integrative hypothesis. In a real analysis it might come from consensus clustering or another multi-view clustering method. Here the synthetic clusters are planted so the visual grammar is easier to inspect.
Sorting twenty patients by CDKN1A expression alone produces a tidy red-to-blue continuum. Stacking methylation, mutation, and copy-number underneath splits the "low-CDKN1A" half into two mechanistically distinct subgroups: TP53-mutant patients and MDM2-amplified patients. Same phenotype, two routes, invisible in the expression panel alone.
Figure 1. Twenty synthetic patients sorted by CDKN1A expression (low, left → high, right). The thin grey lines between panels emphasise that every column is the same patient in both views. Hovering a column brings up the full multi-modal profile. The left-side bracket collapses two mechanistically distinct subgroups into one "CDKN1A-low" label — the multi-modal view separates them. This is the motivation for every figure that follows.
The full matrix is an oncoprint-like view: 80 synthetic patients across four modality row-groups, clinical ribbons above, summary bars alongside. Column ordering is a consensus-cluster stand-in based on the full multi-omics profile, not a sort by any single variable.
Consensus clustering asks what integrative subtypes the cohort falls into; single-variable sorts ask how one variable shapes the rest.
Figure 2. Row groups (top-down): expression (log₂ FC), promoter methylation (β, from 0 unmethylated to 1 methylated), somatic alterations (mutation × copy-number glyphs), and driver-locus copy number. Along the top: consensus cluster, age, sex, stage, mutation burden, vital status. Along the right: per-row alteration frequency. Hover or focus any column for details; switch the sort to see how different orderings shape the visible pattern.
The consensus-cluster ordering uncovers four subtypes. Two — TP53-mutant and MDM2-amplified — are biologically convergent: both functionally disable TP53 by different mechanisms. A sort by TP53 status alone hides MDM2 amplifications inside the wildtype column.
The cards below are subtype fingerprints: one per cluster, showing the ten most discriminating features as signed deviations (z-scores) from the cohort average. The bold bar in each row marks the most extreme subtype.
Figure 3. One card per consensus subtype. Header shows the cluster name and size; the inner strip shows median age, median mutation burden, and fraction deceased. Bars are standardised (z-scored against the whole cohort, in units of σ); axes go from −2σ to +2σ. The same ten features appear in every card, so the rows align column-wise across the grid — hover any feature to cross-highlight it in all four cards simultaneously. Colors mark modality: blue-red for expression, violet for methylation, black for mutation rate, red-blue for copy number.
Cluster A is TP53-mutant with low CDKN1A expression and promoter hypermethylation. Cluster B has wildtype TP53 but amplified MDM2, a different route to suppressing the same pathway. Cluster C is hypermutated with high TMB and a heterogeneous downstream state. Cluster D is quiet wildtype with better outcome.
Clinical ribbons are useful context, but they should not silently drive the discovery step unless that is the study design. If survival status or stage leaks into clustering, the subtype labels can become a circular explanation of the outcome they later claim to predict.
Use a multi-omics heatmap when you have multiple modalities on the same samples and the pattern of co-occurrence is the message. Diverging color scales are coarse — pair the heatmap with a forest plot for effect sizes and a waterfall for single-variable ranking.