The interpretation¶
When a job finishes, PaintOmics AI interprets it by walking a graph. AI agents walk the network of every KEGG, Reactome and OmniPath interaction known for your organism, with your values laid on its nodes. They start where your relevant features concentrate, read your values at every stop against your experiment design, and hand the chain they walked to writers that may cite nothing else. A checked Results section comes out of that chain.
It is a draft for you to check, not a conclusion. It is grounded in two things you can inspect: the interactions the databases draw, and your own numbers. A sentence such as "Ccnd1 gene expression collapsed from +1.90 at 0h to −4.13 at 24h" quotes values the code has matched against your upload, on a leg the network draws. How the walk works describes the machinery.
Starting it¶
You do not press anything. If AI interpretation is enabled on the server and the job was submitted through the upload form, the walk is queued the moment Step 2 finishes. It runs while you look at your results.
The AI Interpret button in the Step 3 toolbar opens the panel. So does the circular mark in the bottom-right corner, whose badge shows the state: spinning while it works, a green tick when the result is ready, and a red exclamation if it failed. The panel is anchored to the page rather than to the results view, so it stays open over a painted pathway.
Filling in Experiment design on the upload form is the most useful thing you can do for the result. The walk reads every value against a design card built from that text and from your column headers. The card says what a value is, what the columns are (time points, doses or unordered conditions), and which omics have no column labels. Without the text, the card has your column labels and nothing else.
While it works¶
The panel lists the legs as the agent walks them, newest last, with a progress bar through its stages: reading the network, reading the design, walking, writing statements, checking them, and writing the Results section. Several agents work at once, and a walk of the whole network finishes within ten minutes, usually in about five.
What the result contains¶
- The header and the five checks: the perturbation the design named ("Ikzf1, induced (up) · anchored in the network") and five chips, one per check the walk must pass before anything below is shown -- not a graph artifact, title fits the body, direction logic, citations in context, anchored to the perturbation. Hover a chip for its numbers. When a check fails, its reason is printed, the Results and the statements are withheld, and only the walk itself is shown; see How the walk works.
- The Results section: a title, a one-sentence summary, and one paragraph per
kept statement in the order the walk found them. Each paragraph ends with chips
for the legs it rests on. Numbers in brackets such as
[1]link to the papers on PubMed, numbered in the order they are first cited. - The cited papers, numbered as the text cites them. Under each paper is the
passage the citation rests on, in the paper's own words, with where it sits:
the abstract, or the part of the main text (results, discussion, introduction).
Hover over a
[1]in the text to read its passage without scrolling. - The statements that passed the checks, each marked mechanism (it rests on signed, directed relations only) or association. Each one separates what the pathway already draws (cited as a leg) from what goes beyond it (a paper the writer read, or a hypothesis worded as one). Dropped statements are listed with the reason.
- The walk itself: every leg, the edge it followed and the database and pathway that draw it, the reading the agent gave at that stop, and each module's distance from the perturbed gene.
Every quoted value, every leg and every citation is checked before you see it. A citation is kept only when an agent reading the paper found the passage that states the claim, and code found that passage in the paper. A statement that fails the checks is rewritten once and then dropped. A Results section that fails twice is dropped too, and the statements stand alone.
There is no export
The result lives in the panel. Select the text and copy it; the line saying it was drafted by a language model comes with it.
Following a thread¶
Click a pathway named on a leg to open its diagram in Step 4. The diagram's Walk column walks that one pathway on its own map. It shows the design card, which you can correct before you start. The legs appear on the map as numbered arcs while the agent walks, and the column ends with a Results section for that pathway.
Ask a follow-up question in the box at the foot of the panel. The question goes to the model with the walk as context, plus tools that read this job's data. Those tools return a gene's values in every layer under your own column labels, and they can compare genes, list a pathway's matched genes, or walk a few steps from a gene you name. The conversation is kept with the job.
When it does not work¶
| What you see | What it means |
|---|---|
| The bar sits at 0%, "Not started", and never moves | The server has AI interpretation enabled but no API key for its provider. Nothing was spent and nothing was sent; this needs a server administrator. |
| "AI interpretation is not enabled on this server." | AI_INTERPRETATION_ENABLED is off here. |
| "The walk was interrupted (no progress for 10 min). Click Retry." | The walk stalled and was marked dead. Retry queues it again. |
| "Your session expired…" | Sign in again and reopen the job from your job list. |
| "This job is no longer stored on the server…" | The job passed its retention window: 7 days for a guest job, 14 for one belonging to a registered account. |
| A pathway's Walk column says the pathway draws no interactions | MapMan bins, and the few KEGG and Reactome pathways that draw no gene-to-gene interactions, have nothing to walk. |
What it cannot do¶
- It cannot see your uploaded files. It sees matched features, their values and their column labels.
- It cannot walk an interaction no database draws. A relation outside KEGG, Reactome and OmniPath, or between features that failed to map, is not in the network.
- It does not know your hypothesis unless you wrote it in Experiment design.
- It is a language model. It can write a fluent paragraph that is wrong about your biology. The checks verify that its values are yours, its legs are drawn and its papers contain the passage shown. Whether the argument holds is your judgement.
- It reads the main text of a paper only when PubMed Central or Europe PMC carries it. Otherwise the passage comes from the abstract.