> For the complete documentation index, see [llms.txt](https://docs.mithrl.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.mithrl.com/features/access-the-knowledge-graph.md).

# Access the Knowledge Graph

## Using the knowledge graph

The knowledge graph explains a gene signature. Give Mithrl a differential expression result and it returns the regulators most likely driving the change, the biology the change affects, and the evidence behind every link: each regulator comes with a predicted direction, a z-score and the target genes that support the call.

The examples on this page come from two studies: Lucanthone at 5 uM in U2OS cells, and a DESeq2 treatment versus control table. For how the graph is built, see [Mithrl Knowledge Graph](https://docs.mithrl.com/docs/mithrl-knowledge-graph).

### What you get

Two analyses run on the same signature:

* **Upstream regulators**: which regulators explain the signature, and in which direction.
* **Biological effects**: which outcomes the signature changes, and through which genes. Terms are ranked in five classes: pathway, process, disease, toxicity and hallmark.

For a drug, you can also ask for its **mechanism of action**: the most likely path from the drug to its outcome. Every result arrives as a network card you can view by cell compartment or as a pathway.

<figure><img src="https://626853399-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FWzmerQ3GQmDagWwm1Oeg%2Fuploads%2FkpgIFKevzq4Vbq73H1tZ%2Fanalyses.jpg?alt=media&amp;token=2275ace7-5160-4572-bd1b-7043e8746bf4" alt="Example cards for upstream regulators, biological effects and mechanism of action"><figcaption><p>The three kinds of result</p></figcaption></figure>

### Ask about a signature

Ask in your own words, naming the comparison:

> Which upstream regulators best explain the Lucanthone 5 uM vs DMSO signature in U2OS?

<figure><img src="https://626853399-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FWzmerQ3GQmDagWwm1Oeg%2Fuploads%2FG3pt8MtzfJMZcjZ1THwg%2Fask-annotated.jpg?alt=media&amp;token=3724fdf9-ba02-480c-9c0c-7ea0d33b7c38" alt="The prompt, the Process row listing each step Mithrl took, the written answer and the network card"><figcaption></figcaption></figure>

In this example the Process row shows that Mithrl loaded the table, kept 3,012 genes at an adjusted p below 0.05, ran upstream regulator analysis and queried 173 relations in the graph.

Name the comparison, the cutoffs and the organism. Prompts that work:

* Which upstream regulators best explain the Lucanthone 5 uM vs DMSO signature in U2OS?
* Using the DESeq2 table, take genes with padj < 0.01 and abs(log2FC) >= 1, top 500, and find the upstream regulators and biological effects.
* Is TP53 activation driving the Nutlin-3a response in MCF-7?
* Walk me through the likely mechanism of action of Nutlin-3a.

### Read the result

#### The network card

<figure><img src="https://626853399-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FWzmerQ3GQmDagWwm1Oeg%2Fuploads%2FkBWrkDuzVn4ijOnuLYxK%2Fnetwork-card-annotated.jpg?alt=media&amp;token=6e125cf3-2a1f-43b8-b146-7400dc134aca" alt="An upstream regulators network card with its controls, legend, ranked regulators panel and summary"><figcaption><p>Network card</p></figcaption></figure>

#### The regulator table

Behind every card is a ranked table. Four columns tell you how far to trust each regulator.

<figure><img src="https://626853399-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FWzmerQ3GQmDagWwm1Oeg%2Fuploads%2F0oU6YiwNqolJQgPQ5Y74%2Fregulator-table-annotated.jpg?alt=media&amp;token=19e0afc8-71b1-4889-832f-11a9bcb223e7" alt="Upstream regulator table with Overlap, P value, BH q, Inferred state and Z columns"><figcaption><p>Upstream regulators for a DESeq2 treatment vs control signature, 475 genes</p></figcaption></figure>

Here FOXA1 (z 6.64) and TFAP2C (z 6.17) read as activated, and palbociclib (z -5.92) as inhibited. ZNF750 and TP63 overlap well but have no inferred state, because too few of their targets carry a direction.

### Explore the evidence

Click any node or line on a card to see the evidence behind it.

<figure><img src="https://626853399-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FWzmerQ3GQmDagWwm1Oeg%2Fuploads%2FM1ZFUDXdyTWsRBN69Uwq%2Fregulator-detail-annotated.jpg?alt=media&amp;token=d7291acb-3e28-41df-b21b-e2d611a69268" alt="Details for FLI1: strongly inhibited, z-score -5.74, and the targets it acts on"><figcaption><p>Clicking a regulator</p></figcaption></figure>

<figure><img src="https://626853399-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FWzmerQ3GQmDagWwm1Oeg%2Fuploads%2FVcoBW67vGBIjUBPjASbi%2Fgene-detail-annotated.jpg?alt=media&amp;token=59057c36-44ea-4126-ba28-2ebe53bc2650" alt="Details for PLK1: up 16.7-fold in the data, and the four regulators that act on it"><figcaption><p>Clicking a gene</p></figcaption></figure>

### Switch views

The same result can be read three ways. The **Cell** view, shown in The network card, places each regulator and target in its compartment: outside the cell, membrane, cytoplasm or nucleus.

<figure><img src="https://626853399-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FWzmerQ3GQmDagWwm1Oeg%2Fuploads%2FFO8Dd5QCRlBjSk65TTaS%2Fpathway-view-annotated.jpg?alt=media&amp;token=b83ccb39-a78e-4299-89ad-e535d259a096" alt="The Pathway view, with regulators on the left and phenotypes on the right"><figcaption></figcaption></figure>

<figure><img src="https://626853399-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FWzmerQ3GQmDagWwm1Oeg%2Fuploads%2FWHSdsWneK00gg2yKXSNC%2Fall-targets-annotated.jpg?alt=media&amp;token=978d0ffe-3d85-42ee-acd8-ebf557c9d84b" alt="All 139 targets shown, with E2F4 found by search"><figcaption></figcaption></figure>

### Biological effects

Upstream regulators tell you what is driving the change. Biological effects tell you what the change does. Ask for both in one prompt:

<figure><img src="https://626853399-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FWzmerQ3GQmDagWwm1Oeg%2Fuploads%2FzeFGv7zxQzNyq5kAFFKw%2Feffects-prompt.jpg?alt=media&amp;token=536c4758-2dae-4d53-9df6-71b5a3f76681" alt="A prompt asking for upstream regulators and biological effects from a DESeq2 table"><figcaption></figcaption></figure>

<figure><img src="https://626853399-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FWzmerQ3GQmDagWwm1Oeg%2Fuploads%2FjJJa0Zja2RDRiPlLfBHb%2Feffects-table-annotated.jpg?alt=media&amp;token=5406ec75-55d1-4185-9feb-0a9298b9399c" alt="Biological effects table with Direction, Genes up, Genes down and P value columns"><figcaption></figcaption></figure>

#### The linked card

With both analyses run, Mithrl joins them in one card.

<figure><img src="https://626853399-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FWzmerQ3GQmDagWwm1Oeg%2Fuploads%2FhMrbbJHPrtHfDO4pRzdL%2Flinked-card-annotated.jpg?alt=media&amp;token=8a59e318-cb75-4ac6-9781-c174ba91af58" alt="A card linking upstream regulators FOXA1, TFAP2C, MYB and palbociclib to three outcomes, with a hypothesis at the top"><figcaption></figcaption></figure>

#### Stories

Open the full view and each outcome becomes a story.

<figure><img src="https://626853399-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FWzmerQ3GQmDagWwm1Oeg%2Fuploads%2FODxItsHGEEf4WMWoneY0%2Fstory-annotated.jpg?alt=media&amp;token=79269a29-ab7a-4e8a-a8bd-3ede1657aa2e" alt="The Proliferation regulators story with its explanation, network, known-biology note, strongest changed genes and regulators"><figcaption><p>The Proliferation regulators story</p></figcaption></figure>

### Mechanism of action

Name the drug and the experiment:

> Walk me through the likely mechanism of action of Nutlin-3a.

Mithrl scores candidate paths through the graph using curated edges only, and draws the best one.

<figure><img src="https://626853399-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FWzmerQ3GQmDagWwm1Oeg%2Fuploads%2FGaZTjDHXtSi7dMfBYpv7%2Fmoa-annotated.jpg?alt=media&amp;token=63f43261-c514-4605-a2ef-cdd05d0183cd" alt="Mechanism of action for Nutlin-3a: Nutlin-3a, MDM2, TP53, CDKN1A and its targets, then the G2/M transition"><figcaption></figcaption></figure>

For Nutlin-3a in MCF-7 cells, the best path scores 0.93 over five curated hops: Nutlin-3a blocks MDM2, stabilised TP53 induces CDKN1A, and the G2/M transition stalls. A second route through apoptosis scores 0.81.

A follow-up to try: "Which TP53 targets did not move as predicted?"

### Check before you conclude

Treat each result as a hypothesis to test. The caveats section of each answer names the checks that apply:

* **Enrichment shows overlap.** A regulator on the list has targets in your signature. That alone does not show it drives the response.
* **Direction needs signed targets.** Unknown or Direction unclear means too few targets carry a direction to call the state.
* **Only mapped genes count.** In the DESeq2 example, 214 of 475 signature genes mapped into the graph. The analysis cannot see the rest.
* **Regulators in your signature score higher.** FOXA1, TFAP2C and MYB are themselves in the signature, so part of their enrichment comes from co-regulation.

Every ranked list is saved in the Explorer as a CSV, such as `ranked_gene.csv`, `ranked_drug.csv` and `ranked_process.csv`, for your own checks. Then design the experiment: a knockdown, a ChIP, or a drug sensitivity test.

### Try it

1. Find the regulators behind your own DESeq2 signature.
2. Click the top regulator and read the targets that support it.
3. Ask for biological effects and open the strongest story.
