Lab Newsletter — June 23, 2026: When Models Learn to Read Cells

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.