active | AICell Lab

Autonomous Research Agents — Discovery from Minimal AI Loops

Mon, 01 Jun 2026 00:00:00 +0000

What is the minimum an AI agent needs to do real science? In this research direction we study whether genuine scientific discovery can emerge from radically simple agent loops: agents that wake up fresh each cycle, read a shared notes file, do a little work, write back what they learned, and repeat — with no long-term memory, no hardcoded tools, and no domain-specific programming.

It’s a deliberately minimal setup, and that’s the point. Much as intricate patterns arise from simple rules in nature, we’re interested in whether complex, useful autonomous research behaviour can emerge from a handful of basic ingredients. We run these loops across several life-science domains, using public scientific databases as the agents’ “instruments.”

This work is ongoing and a manuscript describing the underlying principle is in preparation, so we’re holding specific findings until they’ve been through validation and peer review. If you’re excited by autonomous AI agents, emergence, and machine-driven discovery in the life sciences, this is one of the frontiers we’re actively hiring for — get in touch.

BioEngine - Cloud-Powered AI for Bioimage Analysis

Mon, 01 Jun 2026 00:00:00 +0000

BioEngine is an agent-first platform: it connects browsers, microscopes, and AI agents to GPU compute over secure links, so a scientist can describe a goal in plain language and have an agent find, run, and adapt the right model for them. Built on Hypha (serverless RPC connectivity) and Ray (distributed task orchestration), it scales from a single laptop to multi-node GPU clusters with no code changes, and exposes agent-readable service interfaces so AI agents (such as Agent-Lens) can operate the whole platform. It integrates with the BioImage Model Zoo for FAIR, validated community models.

Nils Mechtel presented BioEngine at the Euro-BioImaging Data Days (2026) — watch the talk or browse the slides. The platform is described in our 2026 preprint, BioEngine: scalable execution and adaptation of bioimage AI through agent-readable interfaces.

The increasingly amount of data generated in life science poses challenges in managing and analysis. The conventional approach for storing and processing scientific data locally on workstations or laptops is failing to met modern needs in applications such as AI-powered image analysis. We would like to tackle the challenge by introducing the BioEngine platform, which is a computational platform consists of containerized services for scalable data management and AI model serving. It is a web platform built on top of the Hypha with an emphasis on serving models for bioimage analysis.

In this project, we aim to develop web services for providing flexible image data management solutions and also model serving in the cloud (private or public). The BioEngine is being used to support the test run feature in the BioImage Model Zoo website (also see the AI4Life project). We aim to provide deployment toolkit for users to setup their own server, either in an institutional Kubernetes cluster or a workstation. Under the AI4Life project, we aim to provide a standard for managing and sharing image data together with the BioImage Archive.

BioImage Model Zoo — FAIR AI Models for Microscopy

Mon, 01 Jun 2026 00:00:00 +0000

The BioImage Model Zoo is a community-driven, fully open resource where standardized, pre-trained deep-learning models can be shared, explored, tested directly in the browser, and deployed in many end-user tools — including ilastik, deepImageJ, QuPath, StarDist, ImJoy, and ZeroCostDL4Mic. A shared model standard makes these models cross-compatible, so a model contributed once can be reused everywhere.

The AICell Lab leads the user services and cloud infrastructure behind the Zoo: the model-upload and testing pipelines, and the BioEngine backend that runs the in-browser “test run” feature. Our goal is to make deep-learning methods for microscopy findable, accessible, interoperable, and reusable (FAIR) across the whole bioimaging ecosystem. This effort grew out of the now-completed AI4Life project and continues as a living community platform.

Recent additions push the Zoo further into the browser: in-browser model testing on cloud or HPC GPUs via BioEngine (no install, no local GPU); a collaborative annotation layer with AI-assisted segmentation (Cellpose and Cellpose-SAM); and an agent skill that lets any AI assistant guide a researcher through contributing a model end to end — packaging, validating, and submitting it. Much of this runs entirely client-side via Pyodide/WebAssembly, backed by Hypha Cloud.

Hypha — Distributed Computing for AI-Powered Science

Mon, 01 Jun 2026 00:00:00 +0000

Modern science is increasingly built from many moving parts — AI models, large datasets, compute clusters, and laboratory instruments — that rarely live in the same place. Hypha is our open-source framework for connecting them. It lets researchers and AI agents call remote functions, services, and models as if they were local, organising everything into unified virtual workspaces.

Hypha is the backbone of much of what the lab builds: it powers cloud model serving in BioEngine, instant model testing in the BioImage Model Zoo, autonomous microscopy in Agent-Lens, and agent-ready biological data. Its companion libraries — hypha-rpc, hypha-core, and hypha-compute — make it easy to expose any Python or browser service to the network and to AI agents over standards like the Model Context Protocol.

Learn more in the documentation or try the public server at hypha.aicell.io.

Euro-BioImaging Research Navigator — AI for Access

Wed, 06 May 2026 00:00:00 +0000

Finding the right microscope — or the right facility — shouldn’t require knowing the map of European imaging infrastructure by heart. The Euro-BioImaging Research Navigator is an AI-powered assistant for the Euro-BioImaging Access Portal that answers plain-language research questions like “Which nodes offer MINFLUX in Europe?” or “I need cryo-EM for structural biology — where can I go?” and returns grounded, ranked recommendations.

Built as part of the EU-funded “AI for Access” project, the Navigator uses a retrieval-first architecture: the language model generates focused search queries, a deterministic BM25 lookup retrieves real records from the facility database, and an agent synthesizes the answer — so recommendations stay grounded in actual data, with honest fallback when something isn’t in the database. It supports streaming responses, visible reasoning steps, multi-turn conversation and interactive clarification.

The Navigator is live at navigator.bioimage.io, deployed on Kubernetes on the KTH cluster, with frontend integration into the Euro-BioImaging Access Portal underway. It continues the lab’s work — alongside AI4Life and the BioImage Model Zoo — on making bioimaging infrastructure open, findable and AI-accessible.

Agent-Lens — AI Agents for Smart Microscopy

Fri, 20 Feb 2026 00:00:00 +0000

What if you could run a microscope just by describing what you want to see? Agent-Lens is a web-based platform where an AI agent turns natural-language intent — “scan well C3 and find the cells with strong green fluorescence” — into concrete microscope commands. It brings together multi-microscope hardware control with robotic sample handling, time-lapse and multi-channel imaging, segmentation (SAM), and similarity search (CLIP/DINOv2) behind a single conversational interface.

Built on Hypha, Agent-Lens connects instruments, AI models, and an interactive UI into one workspace, and manages large image datasets in OME-Zarr. It is our step toward self-driving microscopy: experiments that observe, decide, and adapt on their own, so biologists can focus on the questions instead of the controls.

ImageJ.JS — ImageJ in Your Browser

Thu, 01 May 2025 00:00:00 +0000

ImageJ.JS brings ImageJ — one of the most widely used scientific image-analysis tools — into the web browser. Powered by WebAssembly (via CheerpJ), it runs the full Java application with no installation: open it from any link, keep plugin support, work directly with local files, and get AI assistance for analysis. It serves 1,000–1,500 unique users every day — researchers and educators worldwide.

By removing the installation barrier, ImageJ.JS makes image analysis instantly shareable — ideal for teaching, reproducible workflows, and embedding interactive analysis in websites and notebooks. It integrates with the lab’s broader web-computing ecosystem around ImJoy and Hypha.

BioImage.IO Chatbot — Your AI Assistant for Bioimage Analysis

Tue, 20 Aug 2024 00:00:00 +0000

Bioimaging is evolving fast, with an ever-growing landscape of tools, data formats and workflows. For many researchers — especially those without extensive programming experience — navigating it is daunting. The BioImage.IO Chatbot bridges that gap: an AI-powered assistant, built on a knowledge base contributed by the global bioimaging community, that answers complex questions, points to the right tools and databases, and can autonomously carry out analysis tasks on demand.

Under the hood it pairs large language models with a retrieval-augmented generation (RAG) system over community-curated documentation, and a set of specialized AI agents tuned for different needs — general information, education, and hands-on bioimage analysis. Leveraging modern code-generation, tool-calling and vision capabilities, it solves bioimaging tasks with increasing autonomy and accuracy, and can be embedded into third-party websites or extended through ImJoy and Hypha.

The work was published in Nature Methods (2024, Focus on advanced AI in biology; free access) by Wanlu Lei, Caterina Fuster-Barceló, Gabriel Reder, Arrate Muñoz-Barrutia and Wei Ouyang — a collaboration between the AICell Lab (KTH), KTH’s Department of Intelligent Systems, Ericsson, and Universidad Carlos III de Madrid.

Try the BioImage.IO Chatbot →

Reef - Automated Imaging Farm

Sat, 01 Jun 2024 00:00:00 +0000

The aim of the project is to build a smart microscopy imaging farm for massive production of image data. It consists of multiple microscopes and fluidic systems, robotic arms, liquid handling robots and automatic incubators. The farm will be used for performing automated widefield/fluorescence imaging, long-term live cell imaging, tracking, spatial-omics and multiplexing imaging. With the AI-powered control software, data are analyzed in real-time, augmented views are added on the fly. By generating feedback control signals to control the microscope, the software will automatically change field-of-views, illumination power and other experimental conditions in order to optimize the phototoxicity and capture rare events in live cells.

Here is a sketch for our cell lab with the imaging farm:

Human Cell Simulator

Sun, 01 Jan 2023 00:00:00 +0000

Whole-cell modeling enables a holistic and quantitative view of cell biology and allows performing in-silico experimentation which has a great potential in revolutionizing systems biology, synthetic biology, medicine and other applications in life science. However, modeling the entire cell is an extremely complex task and is heavily limited by our understanding of the biological systems. We take on this grand challenge of building a human cell simulator through recent advances in multi-omics data generation and artificial intelligence. Our aim is to use modern deep learning techniques — transformers, structure models such as AlphaFold, and diffusion/generative models — to integrate existing multi-omics datasets with massive amounts of newly generated live-cell and multiplexed imaging data, and to model cellular behavior through generative and predictive models.

This is the heart of our contribution to SciLifeLab’s Alpha Cell program: turning molecular maps of the human cell — grounded in the Human Protein Atlas and high-resolution imaging — into predictive, simulatable AI models. We build the data engines (Reef imaging farm, Self-driving Microscope) and the AI infrastructure (Hypha, BioEngine) that make whole-cell modeling at this scale possible.

Self-driving Microscope

Sun, 01 Jan 2023 00:00:00 +0000

This is the lab’s overarching vision for autonomous microscopy. We are realizing it concretely through two active projects: Agent-Lens — AI agents that control microscopes — and the REEF Automated Imaging Farm — robotics and fluidics for imaging at scale.

The aim of the project is to develop an AI-powered self-driving microscopy system for studying cellular response under genetic and environmental stressors. The project will also be supported by the WASP-DDLS collaboration, and closely collaborate with the Jaldén group at KTH. The Jaldén group has rich experience in computer vision and automatic control systems. The work will also be supported by the Lundberg group under the Human Protein Atlas, which is a unique world-leading effort to map all the human proteins in cells, tissues, and organs in the human body. By joining the force with the Jaldén group and Lundberg group, we would like to develop an AI-powered imaging system to continuously monitor, actively acquire and track human cells under different cellular and environmental stressors. The data will be used to train large-scale AI models to study cellular responses and make predictions for cell fate.

AI4Life - AI Models for BioImaging (completed)

Thu, 01 Sep 2022 00:00:00 +0000

Status: completed. AI4Life ran from 2022 to 2025 as a Horizon Europe project. The infrastructure and community services the AICell Lab built within it continue today through the BioImage Model Zoo and BioEngine.

AI4LIFE is a Horizon Europe-funded project that brings together the computational and life science communities. Its goal is to empower life science researchers to harness the full potential of Artificial Intelligence (AI) and Machine Learning (ML) methods for bioimage analysis – and in particular microscopy image analysis, by providing services, and developing standards aimed at both developers and users. With a consortium of ten partners, AI4LIFE promises to create harmonized and interoperable AI tools & methods via Open calls and public challenges and bring these developments to researchers via strategic outreach and advanced training. The services provided and solutions developed within the AI4LIFE framework are crucial to solving today’s microscopy image analysis problems and will contribute to boosting the pace of biological and medical insights and discovery in the coming years.

For more information about the project, check it out at the AI4Life website.

The AICell Lab at KTH is a leading partner in the AI4Life consortium. We focus on supporting the user services and cloud computing infrastructure via BioImage Model Zoo (https://bioimage.io). The model zoo is a community-driven, fully open resource where standardized pre-trained models can be shared, explored, tested, and downloaded for further adaptation or direct deployment in multiple end user-facing tools (e.g., ilastik, deepImageJ, QuPath, StarDist, ImJoy, ZeroCostDL4Mic, CSBDeep). To enable everyone to contribute and consume the Zoo resources, we provide a model standard to enable cross-compatibility, a rich list of example models and practical use-cases, developer tools, documentation, and the accompanying infrastructure for model upload, download and testing. Our contribution aims to lay the groundwork to make deep learning methods for microscopy imaging findable, accessible, interoperable, and reusable (FAIR) across software tools and platforms.

For more details about the model zoo, see our publication here.