Clovertex Numen on Google Cloud

User Guide

How to sign in, choose scientific software, launch a costed environment, work in it, and shut it down - on numen-gcp.clovertex.net.

1 · What Numen is

Numen turns "I need a machine with RELION on it" into a running, costed environment in a few clicks - without a ticket, a Terraform file, or a conversation about instance types.

You pick applications. Numen works out the machine, shows you what it will cost, builds it, installs the software, and tells you when it is ready. When you are finished you stop it, and the billing stops with it.

The signed-in home page: "Research compute command center" banner with the four science pillars, the counters row (catalog tiles / apps and tools / running / my launches), and the three cards - Scientific app catalog, Infrastructure, My instances. — The home page. The counters are live: they reflect the real catalog and your own running work.

2 · How Numen thinks

Three ideas explain most of the interface.

You choose software, not hardware

Every application carries its own requirements - cores, memory, disk, how long it takes to install. Select several and Numen adds them up and proposes a machine that fits all of them. You can override it; you rarely need to.

A tile is a working environment, not a folder

Tiles group software that people genuinely use together. Choosing a tile does not commit you to everything in it - inside, you tick only the tools you want.

Nothing runs that you did not ask for

An environment exists because you launched it, and it costs money only while it is running. Stopping it keeps your disk and stops the meter.

Go to numen-gcp.clovertex.net. Pick your account from the list and enter the password your Clovertex contact gave you.

The sign-in card: six accounts in two columns - Alex Chen (Platform administrator), Riya Patel, Jide Okafor, Lena Novak, Sam Whitfield (researchers), and Demo user. Password field beneath, then the sign-in button. — Choose an account, then enter the password. The button names the account you selected.

Account type	Can launch	Can stop / delete	Admin console
Administrator	Yes - any size	Yes	Yes
Researcher	Yes - real sizing	Yes	No
Demo	Yes - capped	No	No

4 · Demo account

The demo account exists to let you prove the platform works without being able to spend meaningful money. Sign in as Demo user with the password your Clovertex contact provided.

What a demo account can do

Browse the whole catalog - every tile, every application, all the sizing and cost information
Launch a real instance. Not a mock-up: a genuine Google Cloud VM, with the software really installed on it
Connect to it and use the software
See its own instances in My instances

What is capped, and why

Guardrail	Value	Why
Machine size	`e2-standard-2` - 2 vCPU / 8 GB, about $0.067/hr	Fixed regardless of what you select, so a demo cannot become an expensive mistake
Disk	Capped at 50 GB	Same reason
Idle stop	30 minutes	The VM stops itself once nobody is using it. Nothing is left running by accident
Stop / delete	Not available	Demo accounts cannot destroy resources. The idle stop reclaims them instead

The cap is a demo guardrail, not the product. A real account gets real sizing - RELION is given its 8 vCPU, ArchR its 32 GB, and the recommendation engine picks the machine that actually fits the work. On e2-standard-2 the heavier applications will install and run, but slowly. That is the guardrail you are seeing, not the platform's ceiling.

What to demo on a capped instance

Tile	On a demo instance
Transcriptomics and quantification	Ideal - 4 tools, installs in about 2 minutes
NGS data processing	Ideal - 9 tools, about 3 minutes
Cheminformatics	Good - RDKit and docking work fine at this size
Structural biology (PyMOL)	Good - light, and scriptable headless
Variant analysis	Workable - installs fine; real callers want more cores
Single-cell, Pipelines (RELION), Bioimage	Not representative - these want 8 vCPU and 32 GB

Idle stop, precisely. The 30-minute clock only runs when the machine is genuinely unused. An open session, an active connection, or real CPU work all reset it - so reading output for 25 minutes will not cost you your machine. It also never interrupts an install.

5 · Anatomy of a tile

Open Scientific app catalog from the home page and you get the tile grid.

The "Browse by category" grid showing nine tiles, each with its icon, name, a preview list of the tools inside, and an "N tools included" link. Search box top-right. — Nine science tiles. Each preview lists the first few tools; the link opens the full set.

Inside a tile you see each application with its version and a short description of what it is for. Tick the ones you want. As you tick, the recommendation panel updates.

A tile opened - application rows with checkboxes, names, versions and descriptions, and the recommendation panel showing the proposed machine, hourly cost and install estimate. — Selecting applications drives the recommendation. Nothing is provisioned yet.

6 · Worked example - quantify RNA-seq reads

The fastest way to see the whole loop. About five minutes end to end.

1Open the catalog

From the home page, choose Scientific app catalog, then Transcriptomics and quantification.

2Select your tools

Tick Salmon and Kallisto - both quantify straight from FASTQ. Add featureCounts and HTSeq if you want the genome-alignment route too. All four together install in roughly two minutes.

3Check the recommendation

Numen proposes a machine and shows the hourly cost and install estimate. On a demo account this is fixed at e2-standard-2; on a real account it reflects what you selected.

4Launch

Give it a short name and press Launch. You are taken to the instance view.

An instance under the "Initializing" heading, captioned "Provisioning or installing - not ready to use yet". The card for numen-sn-cryoem-test-f931 shows a spinner and the live progress counter "0 of 2 · preparing", alongside its owner, machine type and hourly cost. — Progress is live and counted per application - here **0 of 2**, still preparing. As each application installs, the counter advances and names the one being worked on, so you get a real state rather than an indefinite spinner. The Running group below is unaffected.

5Wait for ready

The card turns Running when the software is installed. Two minutes for this tile; a few tiles take much longer and say so up front.

6Use it

Connect (see Connecting) and the tools are on your path:

salmon --version
kallisto version
featureCounts -v
htseq-count --version

7Stop it

On a real account, press Stop when you are done. On a demo account the machine stops itself after 30 idle minutes.

7 · What's in the catalog

Nine science tiles plus two shared sets. "Installable" is what Numen can deploy for you; the remainder are commercial or vendor-licensed products you would bring your own licence for, and they are listed so each tile states its real scope.

Tile	Installable	Full tile install
NGS data processing (QC & alignment)	9 of 9	~11 min
Variant analysis and genomics	11 of 12	~17 min
Single-cell	7 of 10	~65 min
Bioimage analysis and imaging	6 of 10	~48 min
Cheminformatics and drug discovery	5 of 10	~13 min
Transcriptomics and quantification	4 of 4	~6 min
Clinical data management (EDC/CDMS)	2 of 4	~5 min
Structural biology and molecular modelling	1 of 9	~4 min
Shared - add to any environment
Workflow orchestration	5 tools	Nextflow, Snakemake, Cromwell, nf-core
Core programming and IDEs	11 tools	R, Python, JupyterLab, VS Code, Docker

Install times are pessimistic on purpose. Some tiles pull large files from third-party mirrors whose speed we do not control - the imaging set has been seen to take 9 minutes on one day and 55 on another. We quote the slow figure, because being told 9 and waiting 55 feels like a hang.

8 · Watching the cost

Cost is visible before you commit, not after. Ticking another application updates the proposed machine and the hourly rate immediately.

Hourly rate - what the machine costs while running
Running total - accumulates for as long as it runs
Stopped - no compute charge; you keep only a small disk cost

How to read the numbers. Costs shown are Google's on-demand list price for compute. They exclude disk and network, and they do not apply sustained-use discounts, which Google grants automatically on machines left running. For a long-running instance the real invoice is typically lower than the figure shown.

9 · After you launch

The My instances view - one card per environment showing name, state, machine type, the applications installed, and the actions available for your account. — My instances is the single view of everything you have running, stopped, or still building.

State	Meaning	Costing?
Initializing	The machine is being created	Starting
Installing	Machine is up; software is being installed. Progress names the current app	Yes
Running	Ready to use	Yes
Stopped	Shut down; disk kept, so you can start it again	Disk only
Failed	Something went wrong - the card says what	No

10 · Connecting

Command line

Every instance is reachable over Google's Identity-Aware Proxy. There is no public SSH port and no bastion host to configure:

gcloud compute ssh <instance-name> --zone us-central1-a --tunnel-through-iap

Remote desktop

Some applications are graphical and have no command-line mode - Fiji, QuPath, ilastik, IGV and KNIME among them. Tick remote desktop when launching and connect with any RDP client. The tools appear in the Science menu.

Ask for the desktop at launch time. It is part of how the machine is built, so it cannot be added afterwards without relaunching.

10b · Batch jobs and GKE jobs - submit, watch, validate

Two of the four ways to run have no login: a Google Batch job and a GKE (Kubernetes) job run on workers that exist only while the job runs. You submit, watch the state, read the log in the panel, and collect the results from the outputs bucket. If you want a shell, launch a single-node machine instead.

Google Batch

Tick the application, set Infrastructure to Batch - Google Batch job, give it a short name suffix, Launch. The job (numen-bj-<suffix>-<id>) appears at once in My instances as lane Batch.
Google provisions a worker (one to three minutes), runs the script, deletes the worker: queued -> scheduled -> running -> succeeded or failed.
Click the row: state, machine, task counts, timed events, the outputs prefix, and the Cloud Logging link for the job's output.
Validate: results are under gs://numen-gcp-sandbox-outputs/<job>/ (gcloud storage ls gs://numen-gcp-sandbox-outputs/numen-bj-.../). A job that succeeded is also recorded as job evidence the assistant can quote.

GKE - a container job on the Autopilot cluster

Only applications whose maintainers publish a container image run here today: Seurat (runs the PBMC 3k tutorial workflow) and GROMACS (a version check, or the command you pass).

Tick one of those, set Infrastructure to GKE - Kubernetes job (Autopilot), add a suffix, Launch. The job (numen-kj-<suffix>-<id>) appears in My instances as lane GKE.
Autopilot provisions a node for exactly the CPU and memory requested (one to three minutes), pulls the image (a minute or two the first time), runs the container.
Click the row: state and conditions, image and resources, the pod (its phase and, if waiting, why), the events, and the container log read back through the Kubernetes API. Refresh re-reads it.
Validate: a Seurat run ends with NUMEN_GKE_SEURAT_SUCCEEDED cells=... clusters=... markers=.../12 in the log and writes summary.json, clusters.csv, markers.csv and the Seurat object under gs://numen-gcp-sandbox-outputs/<job>/.
Queued with no pod for more than five minutes means the cluster rejected the pod template; the Events in the panel say why. Delete job stops a running job or removes a finished one; outputs stay.

Nothing to log in to. The pod is gone when the job ends; the log stays readable in the panel for seven days and the outputs stay in the bucket.

Slurm clusters - the one with a login

Build a cluster from New cluster... (head node plus compute groups; a shared filesystem at /scratch if you pick one). Once it is running, log in to the head node as yourself over IAP and OS Login, submit with sbatch, watch with squeue; finished jobs stay in My instances because they are read from the cluster's accounting.

gcloud compute ssh <head-node> --zone us-central1-a --tunnel-through-iap

11 · Shutting down

Stop - shuts the machine down, keeps the disk. Compute billing ends. Start it again later.
Delete - removes the machine and its disk. Nothing is kept.

Every instance also stops itself when it goes idle. Idle means no logged-in user, no open SSH or remote-desktop session, and no CPU or disk activity. Any one of those resets the clock, so a long job, a download, or a session you are reading but not typing into all keep the machine awake - you do not lose work by walking away mid-analysis. The default is two hours; demo accounts are fixed at 30 minutes.

An idle-stopped instance is stopped, not deleted: the disk and everything installed on it are kept, and you start it again from My instances. The instance card says it stopped itself, so a machine that is off is never a mystery.

Demo accounts cannot stop or delete. This is intentional. Your instance stops itself after 30 idle minutes, and a Clovertex administrator clears it up afterwards. You cannot leave anything running by accident.

12 · FAQ

Is the demo a simulation?

No. It provisions a real Google Cloud VM and really installs the software. The only differences are the size cap and the idle stop.

Why is my demo instance slower than expected?

Because it is capped at 2 vCPU / 8 GB regardless of what you selected. A real account is sized to the work.

My instance disappeared.

If it is a demo instance, it stopped itself after 30 idle minutes. That is the guardrail working. Launch another.

Can I install something not in the catalog?

On a real account, yes - it is your machine once it is running. Anything you add by hand is not reinstalled if you rebuild, so tell your Clovertex contact if it should be in the catalog permanently.

Where does my data go when I stop an instance?

Stop keeps the disk and everything on it. Delete does not. Anything you cannot lose should be copied to Cloud Storage before deleting.

Why do some applications say "not installed"?

They are commercial or vendor-licensed - cryoSPARC, MaxQuant, Cell Ranger, ChemDraw and others. They stay listed so each tile states its real scope, and you can bring your own licence.

13 · Glossary

Tile: A group of scientific applications that are used together. You pick individual tools inside it.
Instance: One virtual machine that Numen built for you, with your chosen software on it.
Recommendation: The machine Numen proposes based on what you selected - cores, memory and disk.
e2-standard-2: A Google machine type: 2 vCPU and 8 GB of memory. The fixed size for demo accounts.
Idle stop: A demo safeguard: the machine stops itself after 30 minutes with no user, no connection and no CPU work.
IAP: Identity-Aware Proxy - how you reach an instance without any public SSH port.
Remote desktop: A graphical session for applications that have no command line. Requested at launch.
Single node: One machine. Most work fits here.
Cluster: Several machines for one job - needed only for software that genuinely splits work across nodes, such as RELION.