Meet the bioAF Assistant: run pipelines by describing what you want

A conversational, action-taking assistant that turns plain-language intent into real work in bioAF. It resolves your data, recommends and installs the right pipeline, sets up experiments and samples, launches runs, and explains results, all behind a plan you confirm and your own permissions.

Most biology teams have someone who knows the samples cold and knows exactly what question they want to ask, but has never written a line of Nextflow. Getting from “I have a 10x run sitting in Experiment 3” to a configured, launched pipeline usually means finding a bioinformatician, or spending an afternoon learning parameter flags that will be forgotten by next month.

So in this release we shipped the bioAF Assistant: a chat interface that you talk to in plain language and that takes real action on your data. Tell it what you have and what you want, and it figures out the rest: which experiment or sample you mean, which nf-core pipeline fits, how to configure it, and what the results say once it finishes.

This is the feature that lets every member of your biology team act as a computational biologist. They spend their time thinking about results and designing the next experiment, not fumbling with pipeline configuration.

What you can use it for

You tell the Assistant what you want in plain language. Some real things you can type:

  • “I want to do single-cell sequencing of the sample I just uploaded into Experiment 3.”
  • “Set up a new experiment on serotonin levels in the mouse gut and add my samples.”
  • “Import GSE123456 from GEO and run it through the standard bulk RNA-seq pipeline.”
  • “How did my last run go? What does the QC actually mean?”

You don’t need to know pipeline keys or database IDs. You just describe your biology, and the bioAF Assistant resolves it to real records in your organization.

How the bioAF Assistant actually works

The Assistant reads first, then proposes a course of action, then waits for you. Every message runs a short loop where the Assistant calls a fixed set of tools:

  1. It reads to resolve your intent. Read-only tools (list experiments, list samples, list pipelines, check run status) let it turn “the sample I just uploaded into Experiment 3” into a specific experiment and sample before it does anything. These run freely, because they change nothing.
  2. It recommends, deterministically. Pipeline recommendation is based on a rule-based bioAF service, not the model guessing. It maps a sample’s assay and organism to a pipeline and reference genome, and reports how confident it is and which signals it used. So the most common decision doesn’t ride on the model, and you can audit it.
  3. It proposes a plan. Anything consequential (install a pipeline, create an experiment or sample, launch a run) is never executed on the model’s say-so. It becomes a step in a plan the Assistant shows you, with the resolved entities spelled out in plain language, so you can catch “not that sample” before anything happens.
  4. You confirm before it acts. On confirmation, the plan executes in order. If a step spends compute, the bioAF Assistant warns you before you approve it.

If you ask for several dependent actions at once (“install scrnaseq and run it on these three samples”), the Assistant batches them into a single plan you confirm in one step, and executes them in order.

What it can do

The Assistant has a small, fixed set of tools. Grouped by what they do:

  • Discover: list your experiments, list an experiment’s samples, list installed pipelines, check the status of a run.
  • Set up: create a new experiment, add samples to it (with controlled-vocabulary assay, organism, prep method, and more).
  • Install and run: install any nf-core pipeline into your catalog, then launch a run, scoped to the specific samples you named or the whole experiment.
  • Explain: read back a run’s QC metrics and get a plain-language explanation of what the results mean. Anyone who can view a run’s results can ask.
  • Review (for permitted users): if you hold the AI review permission, you can ask the Assistant to run a full, saved agent review of a run, right from the chat, and it lands on the run’s AI Review tab like any other review.

Conversations are saved. Close the chat and reopen any past conversation from the Assistant’s History to pick up where you left off.

Why it’s safe to let it act

An assistant that can spend compute and change your records is only worth it if the safety lives in bioAF itself, not in hoping the model behaves. In bioAF, it does. The model never runs anything itself: it can only ask to, and bioAF checks every request against your permissions before anything happens.

  • It uses your account, not its own. The Assistant has no permissions of its own. Every action runs as the authenticated user and is checked against that user’s role, server-side, on every call. Launching a run requires the same pipelines:launch permission the launch button requires. A bench scientist can hold a full conversation and get a recommendation, but if their role cannot launch, the launch is declined at the tool boundary. The Assistant never grants itself rights.
  • Consequential actions always stop for confirmation. Installing a pipeline, creating an experiment or sample, and launching a run never execute until you confirm the proposed plan. The plan renders the resolved entities in plain language so someone who isn’t an expert can still spot a mistake before it costs anything.
  • Spend is called out before you approve it. When a plan includes a step that spends compute, the confirm card shows a “this will spend compute” warning. Your budget cap is still the hard limit. Confirming is just the human check on top of it.
  • Everything is audited, and marked as the Assistant. Every action is written to the audit log, attributed to you and tagged “via assistant.” A reviewer sees who took each action and that the agent was used, and the full chain (intent, plan, confirmation, tool call, result) is reconstructable. Assistant-driven entries are badged in the audit log UI.
  • What the Assistant reads is data, not instructions. Content it reads back (your sample metadata, accession metadata pulled from public databases like SRA and GEO, QC text, pipeline error logs) is treated strictly as information. Tool results are fenced as untrusted input and cannot redirect the Assistant, skip the confirmation step, or override its instructions. Your records and their provenance are never rewritten by this: the defense sits at the model boundary only.

Bring your own provider

The Assistant runs on your organization’s active LLM provider, the same single-active configuration that powers AI Review. Bring your own key for Anthropic Claude, OpenAI ChatGPT, or Google Gemini, and one provider is active for the whole organization at a time.

One requirement worth knowing: the Assistant needs a provider with native tool-calling. When the active provider and model support tools, the Assistant is available; when they don’t, it tells you exactly what to change in Settings > Integrations > LLMs. A self-hosted, non-tool-capable model can still power AI Review, but not the Assistant.

Who gets it

Access to start a conversation is gated by a new assistant:use permission, granted by default to the admin, comp_bio, and bench roles. A read-only viewer does not get it: a role that cannot act should not drive an action-taking agent.

Because the Assistant works within each person’s role, it makes bioAF easier to use without handing anyone new permissions:

  • A bench scientist can describe their data, get a pipeline recommendation, set up experiments and samples, and ask what a finished run means.
  • A comp bio user or admin can do all of that and launch real runs, install pipelines, and trigger a full saved review, all from chat.

What the Assistant is not

A few things it doesn’t do, so you know what to expect:

  • It is not a permission bypass. It can never take an action your own role could not. If you cannot launch a run in the UI, you cannot launch one through the Assistant.
  • It does not spend silently. No consequential action happens without your explicit confirmation of a plan, and spend is flagged before you approve it.
  • Its explanations are advisory. A plain-language results explanation is a research-assistant note, not part of the scientific record, exactly like AI Review. The runs, samples, and QC it describes remain the source of truth.
  • It needs a tool-capable provider. If your active provider can’t do native tool-calling, the Assistant just isn’t available. It won’t quietly run in a worse mode.
  • It does not manage your AI spend. Billing happens on your provider account, on their terms. Bring your own key.

Try it

If your organization already has a tool-capable LLM provider configured for AI Review, the Assistant is ready: click the icon in the header and describe what you have and what you want. If not, set an active provider under Settings > Integrations > LLMs, then start with something small and concrete, like asking it to characterize an experiment you already understand and recommend a pipeline for it. Confirm a plan, watch it act, and use that to build trust before you hand it a launch.

TL;DR: you describe the biology, the Assistant does the busywork, and nothing important happens until you say so.