Data and analysis

Run Analysis

Describe what you want to know. Get the code, the figures and the write-up.

ARC reads the data, writes the analysis, runs it, looks at what came out, and iterates when something is obviously wrong. Results land in the project's Analysis folder — the outputs and the script, so the run is reproducible and reviewable rather than a black box that emitted a PNG.

It runs as a background job with a visible pill, so you can carry on working. When it finishes you get a card on the Activity board that opens the result directly in the editor.

Cloud jobs

Hand off the long-running work, shut the laptop, pick the results up later.

Run Analysis and Draft Text both have a Cloud tab. A cloud job runs an ARC instance on Heureka's infrastructure with proper compute behind it, and downloads its results into your project folder when it completes. Your machine does not need to stay awake, or even online.

Running jobs are counted on the button that launched them, so a cloud draft does not show up as a pending analysis. You can also start, watch, cancel and retry cloud work from the Activity board, or from your phone.

Raw sequencing data, pre-processed

Point Bench at the server your data landed on and get analysis-ready output back.

The Pre-processing tab takes an FTP or SFTP location holding raw sequencing files and runs an established pipeline over them in the cloud — single-cell RNA-seq, metagenomics and similar. Results stream back into the project on completion.

This is the step that usually costs a lab a week and a core facility ticket. It is one dialog.

Code, read and run

Python, R, notebooks, JSON, config — opened properly, edited in place.

Scripts in a project are first-class: syntax highlighted, editable, and available to the agent to read, modify and execute. The Code folder is where analysis scripts live, so the thing that produced a figure is one click from the figure.

Chemical structures

Drop in a .mol, .sdf or .smi and get a drawn structure instead of a wall of coordinates.

An SDF library opens as a grid you can click through: every compound drawn, the properties stored in the file, computed identifiers (canonical SMILES, InChIKey, formula, weight), a rule-of-five read, and a search box that highlights a substructure across the whole set.

Your Inventory draws itself with no setup — any column called SMILES, Structure, InChI or InChIKey renders the molecule above the text. Paste structures into a chemicals list you already keep and it lights up. Going the other way, Add to Inventory sends compounds from a structure file into your inventory, each row carrying its own InChIKey so it stays matchable later.

The agent knows its chemistry

Resolve a compound from a name, CAS number, InChIKey or structure — and get back what is actually known about it.

What it is, what it is known to hit, how far it has progressed clinically, what its hazards are, and what close analogs exist. ARC resolves structures rather than recalling them, so a compound written into your record is one that was looked up.

In Privacy Mode these lookups are blocked along with every other off-machine tool — a compound name never leaves your machine.

Protein structures

.pdb, .cif and .mmcif open in a real 3D viewer.

Cartoon, Ribbon and Trace styles, with bound ligands staying visible rather than disappearing with the backbone. mmCIF matters because it is what the PDB serves by default and the only format for very large structures — so AlphaFold models and big complexes open rather than defeating you.

The viewer follows the panel when you resize it and matches your light or dark theme.

Data Spaces

Pool your datasets into a linked graph, and let the agent reason across the lot.

Drop datasets into a space and ARC distils each into a structured card — organism, tissue, condition, the top features with their direction, enriched processes, key insights. It then links the cards by what they genuinely share into a graph you can explore, drag and read.

Link thickness and opacity track how much two datasets actually have in common, so a strong overlap is visibly different from a one-thing-in-common link. By default the graph connects datasets by their specific biology — shared genes and pathways — and a filter in the legend lets you add or remove other kinds of connection (organism, tissue, condition, comparison). Your choice is remembered.

Open ARC on a space and ask about it: what is in it, which datasets share a signal, where a gene or pathway shows up across your body of work and in which direction. That prior work then acts as context when you analyse something new or draft text about it. A whole Bench project can be added as a card, and any card can be edited by hand.

Spaces can be private to you or shared with your team, where datasets that look alike are also connected by similarity.

The paper library

Papers you rely on, kept in one place and linked to the projects that use them.

A library of references you can read inside Bench, annotate, and hand to the agent. Papers link to projects, and a project's rail shows which ones it is built on — so the literature underneath a piece of work is visible from the work rather than buried in a reference manager.