Services

Cortex

Contextual Omics Reasoning and Translational EXploration

An AI co-scientist for analysing the lab's genomic and multi-omics data. It has no web screen; you use it conversationally on the lab's analysis server.

What it does

Cortex is the lab's genomics and multi-omics data analysis co-scientist. It is more than a tool that writes code for you: it builds hypotheses with you and ties together the lab's own R packages, curated databases, and literature and target knowledge to design, run and interpret reproducible analyses.

Use it to start a new analysis or to pick up the analysis of a lab project already under way. Where AURORA takes raw data through alignment, quantification and variant calling, Cortex takes those results on to downstream analysis and interpretation — group comparisons, survival analysis, single-cell interpretation and so on.

What to know before you start

There is no web screen. The Cortex card appears in the company site's service list, but its button is disabled and reads "Coming soon", and going to the service address takes you back to the home page. Cortex is not a web service: it is a working environment you use by starting an AI coding agent (Claude Code or Codex) on the lab's analysis server and talking to it.

You need an account on the lab's analysis server. When a lab member logs in to the server and starts the agent in the Cortex working folder, Cortex's rules, analysis procedures and tools are loaded with it. Ask the lab administrator for the folder location and how to connect.

Originals cannot be changed. The raw sequencing data and the lab's own R packages and curated databases are read-only; writing there is blocked. All analysis output goes into a working folder created per project.

The site card also describes planned scope. The virtual-experiment and virtual-clinical-trial simulation mentioned on the card has not been confirmed as something Cortex provides today. This page lists only what works now.

What it can analyse

Cortex follows a fixed procedure (a skill) for each kind of analysis. These are the areas in place today.

Dedicated helper agents handle statistical-rigour review, reproducibility audits, literature interpretation, result interpretation, planning the next analysis, and writing paper-style reports.

How to use it

1. Create a project folder — ask the agent to create a new project, or use /new-project <name>. Lowercase letters and underscores are recommended. If the name already exists it refuses rather than overwriting
2. State the question and the controls — say what you want to find out, what you are comparing with what, and any thresholds or samples to exclude that you have already decided. This is recorded as the first instruction and becomes the reference for later work
3. Connect the data — for results handed off from AURORA, give the project name (see "Taking over from AURORA" below). For other inputs, give their location; they are read in place, not copied
4. Review the analysis plan — Cortex determines the data type and proposes the stages and the pass criteria for each. The plan is a proposal; it runs only after you approve it
5. Run — heavy preprocessing and training are submitted as jobs to the lab's compute cluster. Light work runs directly
6. Check and interpret — at each stage it shows the result as a table or figure and writes an interpretation against it. If a result looks wrong, say so and have it rerun
7. Wrap up — if needed, produce paper-style figures and a report, and record the runtime environment with /session-info <name>

Taking over from AURORA

In AURORA's results screen, open the CORTEX로 넘기기 (Send to CORTEX) tab, choose the project to hand off to and press the button. A handoff document (cortex_handoff.json) containing the analysis R objects, the sample metadata and the list of input files is written into that project. Cortex reads this document first.

• Fill in the comparison group (group) in AURORA's sample metadata table before handing off. It goes into the handoff document, so a between-group comparison can start straight away
• Large files such as BAM and VCF are not copied. They are read in place, from the locations recorded in the handoff document
• You can only hand off to a project that already exists and is writable

How to read the results

Every project folder has the same shape.

One script makes one output, and the output file has the same name as the script. From any figure you can go straight to the code that made it. When someone opens the project months later, the work log and analysis summary in reports/ are the place to start to see what was done in what order.

Must-know points

Plans and interpretations are proposals. The hypotheses, next-analysis plans and interpretations Cortex writes are material for your judgement, not conclusions. Check the figures and numbers yourself before they go into a paper.

The safeguards differ under Codex. The automatic check that blocks writing to the original data areas only runs under Claude Code. Under Codex the same rules are followed only as instructions, so if a command that touches originals is proposed, stop it yourself.

Literature and target searches call outside services. What goes to PubMed and the like is the search query — do not put unpublished information into it.

Software availability

Cortex does not compute on its own; it calls the components below. If one is missing, only the area that depends on it stops.

What to do when blocked

1. Ask the agent to tell you exactly by name what is missing or disconnected
2. If a connector does not appear, check its connection and authentication with /mcp in an interactive session
3. Package, cluster and local-model problems cannot be fixed by users — pass the name and the error text to the lab administrator
4. If it is urgent, ask whether another route does the same job — for example, if the cluster is blocked you can check on a small subset of samples first

Last updated · 2026-09-29