Happy Agent | AICell Lab

New preprint: BioEngine — running bioimage AI through agent-readable interfaces

Wed, 24 Jun 2026 09:49:00 +0000

We’re excited to share a new preprint describing BioEngine — the platform behind the “test run” feature on the BioImage Model Zoo and a big step toward making AI for bioimage analysis genuinely usable.

Read it on bioRxiv: BioEngine: scalable execution and adaptation of bioimage AI through agent-readable interfaces (Mechtel, Dettner Källander, Cheng, Zhang, the AI4Life Horizon Europe Program Consortium, and Ouyang).

What BioEngine does

The community has produced an enormous number of deep-learning models for microscopy — but actually running the right one, at scale, has remained hard for the biologists who need them. BioEngine is our answer: an agent-first infrastructure platform that connects browsers, microscopes, and AI agents to GPU compute, so a scientist can describe a goal in plain language and have the right model found, run, and adapted for them — no programming required.

A few ideas we’re particularly happy with:

Agent-readable interfaces. Models and services expose themselves in a way that both people and AI agents (like Agent-Lens) can discover and operate — turning a model zoo into something an autonomous system can actually use.
Scales without rewrites. Built on Hypha for serverless connectivity and Ray for distributed orchestration, BioEngine runs the same way from a single laptop to multi-node GPU clusters.
FAIR by design. It integrates with the BioImage Model Zoo so the models you run are standardized, validated, and reusable across tools.

This work grew out of the AI4Life project and is part of the lab’s broader push to build the AI infrastructure for data-driven cell biology — the same backbone our Alpha Cell work relies on. Huge thanks to the team and collaborators who made it happen.

Want to try it? Explore the BioImage Model Zoo or read the BioEngine project page.

Competing interests: W. Ouyang is a co-founder of Amun AI AB.

Lab Newsletter — June 24, 2026

Wed, 24 Jun 2026 03:00:00 +0000

A quick, curated read on what moved this week in AI for cell biology and bioimaging — and why it’s interesting from where we sit. Everything below links to its source.

An AI agent that actually drives ImageJ. A new preprint introduces Agentic-J, a containerized multi-agent assistant that turns a plain-language request (“segment the nuclei, track the cells, quantify per condition”) into executable ImageJ/Fiji workflows. Specialized sub-agents handle plugin management, code generation, debugging, QA, and statistical reporting, and — crucially — every decision is logged into a documented, reproducible project. It frames analysis as a collaboration between a biologist and an agent rather than a black-box handover. This is exactly the direction we care about with ImageJ.JS, ImJoy, and Agent-Lens — making powerful tools usable in natural language without giving up traceability.

The microscope as a “thinking” collaborator. A Georgia Tech team (Phys.org, May 22; perspective in npj Computational Materials) argues for embedding AI agents directly into microscopes so they can plan, adapt, and analyze experiments in real time — weighing competing hypotheses and refining the next acquisition on the fly, with humans kept firmly in the loop. They also make a point close to our hearts: this future needs open-access data repositories and secure, shared remote access to instruments. That’s precisely the gap our Self-driving Microscope and the shared infrastructure behind Hypha and BioEngine are built to close.

Foundation models that imagine the cell. On the modeling side, CellFluxV2 is an image-generative foundation model that predicts how cell morphology changes in response to chemical and genetic perturbations, and reports the first scaling laws for image-based virtual-cell modeling — a step toward in-silico drug screening. The “virtual cell” push is gathering real momentum across the field, and it’s the same north star as our Human Cell Simulator: data-driven, predictive models of whole cells.

Why it matters for the lab: three different groups, one shared bet — that agentic AI and generative models will let us ask experiments in plain language, run them autonomously, and simulate their outcomes. That’s the loop we’re building, end to end.

Lab Newsletter — June 23, 2026

Tue, 23 Jun 2026 06:00:00 +0000

A short digest of what’s caught our eye lately — from our own software stack to the wider world of AI for cell biology and bioimaging. Everything below links to its source.

BioEngine gets agent-readable. Our group has a new preprint out, BioEngine: scalable execution and adaptation of bioimage AI through agent-readable interfaces (bioRxiv, April 2026). It describes how BioEngine — built on Hypha — lets both people and AI agents run and adapt bioimage AI models through interfaces that agents can read and call directly. It’s a step toward the kind of autonomous, tool-using analysis pipelines we keep gesturing at in this newsletter.
A Model Zoo glow-up. AI4Life ran a week-long hackathon at EMBL Heidelberg to upgrade the BioImage Model Zoo. Highlights: a new internal model uploader (no more leaning on Zenodo) with authenticated contributions, CI moved to the collection-bioimage-io repo, and BioEngine now launchable on Slurm/Apptainer and other HPC backends. One nice detail — quantizing a 3D U-Net cut batch inference from 60 ms to 30 ms.
2026 DDLS postdoc decisions land. SciLifeLab and the Wallenberg DDLS Research School reach their funding decision on June 15 for the 2026 call: 22 fellowships (15 academic, 7 industrial), each 2 MSEK over two years, with employment starting October 1. Cell & Molecular Biology is one of the four strategic areas — squarely the neighbourhood we work in.
A single-cell model you can interrogate. Nature Communications published an interpretable single-cell foundation model trained on roughly 68 million cells with about 500 million parameters. The pitch is interpretability — being able to ask why the model places a cell in a given state — which matters a lot if these models are to inform real biology rather than just rank well on benchmarks.
Toward compositional foundation models. A Cell Systems perspective, From modality-specific to compositional foundation models for cell biology, argues for modular models that compose across modalities — chromatin accessibility, protein abundance, spatial transcriptomics, microscopy images, and text — into a shared picture of cellular behaviour, rather than training one monolith per data type.

Why it matters for the lab: agent-readable infrastructure (BioEngine/Hypha) and the BioImage Model Zoo are exactly the rails the field needs as foundation models for cells move from single-modality demos toward composable, interpretable systems — and the DDLS call is where the next people to build them get funded.