Metagenes

A metagene is one profile that stands for a group of features that move together. PaintOmics uses them in two places where there are more values than there is room to draw: a pathway map node that holds many genes, and a pathway node in the network view, where a whole pathway has to be summarised by a single colour. This page explains what is computed, when, and how to read it.

A metagene is not a gene. It has no identifier in your data and no position of its own in a pathway; it is a summary of the features that are already there.

Where metagenes are computed

There are two levels, computed at different times by different code, and they do not cover the same omics.

	Pathway level	Node level
Summarises	Every matched feature of one omic in one pathway	The features sharing one position on a pathway map
Computed	During Step 2, once per omic and per pathway database	In the browser, when you open a crowded box in Step 4
Covers	Gene-based omics only	Every omic in the job
Appears in	The pathway network colouring and the major trends charts	The pathway diagram, behind the Genes / Metagenes toggle

The pathway-level pass runs over the job's gene-based omics, so a compound-based omic such as metabolomics never gets a pathway-level metagene and never appears as a cluster colouring on the network. Node-level metagenes in Step 4 do cover every omic in the box.

An omic that matched no feature at all is skipped rather than failing the run, and so is a pathway or database whose data is too degenerate to cluster.

How a metagene is computed

The values of the group are centred and reduced by principal component analysis. The metagene is the component's score — the profile across conditions — and the loadings say how much each feature contributed to it.

• Not every component becomes a metagene. A component is kept only while its explained variance passes a cutoff of 0.3. That is why a node or a pathway usually yields one or two trends rather than as many as it has features, and why some yield none at all.
• The sign is corrected. The sign of a principal component is arbitrary, so each metagene is flipped where necessary to follow the majority of its members' loadings. A metagene that goes up therefore means the features that built it mostly go up.

With a single condition there is no PCA

One condition gives no direction of maximum variance across conditions, so neither level runs a PCA:

• The pathway-level pass uses a median centroid of the pathway's members.
• The browser uses an outlier-trimmed mean of the box's members, dropping values more than three standard deviations from the mean.

The sign-orientation step is skipped in this regime, because there is no arbitrary component sign to resolve. Single-condition jobs are also clustered with a model-based clusterer in one dimension rather than with k-means, since with one condition the amplitude is the profile.

Pathway-level metagenes on the network

Open the pathway network on Step 3 and set Node coloring to an omic name instead of Classification. Each pathway node is then coloured by the cluster its metagene fell into, and the rail shows:

• the heading N Clusters found from M in total — the second number is every cluster the job produced, the first is how many still own a pathway that survived the filters you have applied;
• one thumbnail per cluster, plotting its member trends in grey with the cluster centroid in red, and captioned with how many metagenes it holds. Click a thumbnail to hide or show its nodes, so you can isolate the pathways that move together;
• Modify number of clusters, a slider from 1 to 20 with an Apply button.

By default PaintOmics chooses the number of clusters itself — an elbow scan for k-means, or BIC for the model-based clusterer — and clamps it to the number of distinguishable profiles the data actually holds. Asking for more than that yields fewer than you asked for rather than an error: with only two conditions, for instance, every centred profile lands on one of two points, and no amount of sliding will produce more than two meaningful groups.

Note

Apply re-runs the clustering as a queued job, and the panel says so: it is an intensive process and the results take time to come back. The slider is hidden on a job you are only viewing rather than one you own.

Major trends in a pathway

Click a pathway node in the network, or a row in the pathway table, and the details panel lists that pathway's metagenes for each omic — N major trends in this pathway — with a Heatmap / Line chart toggle. The same panel appears in the Pathway information column beside an open diagram in Step 4.

In the heatmap each row is one trend, labelled Trend n and the cluster it belongs to. Where the pathway has no data for an omic the panel says No data for this pathway.

Warning

The colours in a trend chart are scaled to the metagenes' own range, not to the omic's. A metagene is a component centred on zero, so it goes negative whatever the omic did, and its magnitude is not in the omic's units — on a real job here an omic ran 0.79 to 1.41 while its metagenes reached ±9.4. Compare the shape of two trends, never the strength of a trend colour against the colour of a gene box.

When you switch the job to the averaged view, metagene charts are collapsed the same way every other chart is, so a trend has one cell per condition and its tooltip names the condition, instead of one cell per replicate.

Node-level metagenes on a pathway diagram

Features bucket onto a map by their literal position, so co-located genes share one drawn box. When a single position holds more than five features, the box no longer paints one arbitrary member: PaintOmics computes metagenes for that group in the browser and paints those instead. A box holding exactly five features still paints an individual feature.

The feature window for such a box carries a Genes | Metagenes toggle in its title bar — the only way back from the compressed view to the individual features — and the Prev. / Next arrows step through the metagenes just as they step through genes. The N more … at this position line says how many are behind them.

Node-level metagenes are named in order: Metagene 1, Metagene 2, and so on, within that box. The name is positional and means nothing outside the box it was computed for. If the computation fails for any reason, the box falls back silently to the ordinary single-feature display.

Reading a metagene honestly

• A metagene is a component, not a measurement. Its units are not the omic's units, and two metagenes from different pathways or different boxes are not on a common scale.
• Trend n and Metagene n are positional labels. They are stable within one pathway or one box in one job; they carry no meaning between them, and re-clustering renumbers them.
• A cluster tells you that a set of pathways move alike for that omic. It says nothing about whether they are enriched — that is a separate test, described in Pathway enrichment.
• When you need the individual features, they are always still there: the Genes toggle in the feature window, and the per-feature charts described in the detail views and heatmaps.

Where to go next

• The pathway network
• Reading a pathway diagram
• A whole job, step by step