Lab Newsletter — August 4, 2026: The Cell's Fingerprint

A few weeks ago we watched cell segmentation grow up — models that reliably find where the cells are, across microscopes they’ve never met. That’s the outline. The layer directly above it is quieter and, for a lab that wants to measure biology, more consequential: turning those outlined pixels into meaning — a compact, quantitative phenotypic fingerprint of what a cell is doing, that you can compare, cluster, and query. This is image-based profiling, and 2025–26 is the season it grew a proper foundation model — one trained, as it happens, on our own home turf.

🔬 A model that reads the cell, not just its shape

The capability anchor is SubCell (Gupta, Wefers, … Karaletsos, Lundberg; bioRxiv, updated Oct 2025 — preprint): a suite of self-supervised Vision Transformers trained on the Human Protein Atlas — protein expression and spatial distribution for more than 13,000 genes across 37 cell lines — with a proteome-aware learning objective. Point it at a fluorescence image and it emits an embedding that captures “cellular morphology, protein localization, cellular organization, and biological function.” The result that matters is the transfer: without any fine-tuning, SubCell “produces robust representations … across diverse independent datasets that vary greatly in image resolution, channel markers, cell types, and even species” — and it drives real downstream work: localization classification, cell-cycle modeling, drug- response prediction, and mechanism-of-action identification, on datasets from OpenCell to JUMP Cell Painting. From that single representation the authors build “the first image-based multiscale map of subcellular protein organization,” learned directly from pixels. Why it matters for the lab: this is close to home in two senses. The data is the Human Protein Atlas — the imaging substrate the Human Cell Simulator leans on — and the work comes out of Emma Lundberg’s group at KTH, next door. A model that reads phenotype straight from a picture is the measurement layer of everything we build.

🧱 The wall: when the channels don’t match

Capability is half the story; the field is refreshingly candid about the other half. Morphology models are usually “trained with a single microscopy imaging type,” which yields — in the words of CHAMMI-75 (Agrawal, … Caicedo; ICLR 2026) — “specialized models that cannot be reused across biological studies because the technical specifications do not match,” the classic offender being a “different number of channels.” Your beautiful five-channel Cell Painting model meets a three-channel assay and simply has nowhere to put the pixels. CHAMMI-75’s answer is a public dataset of heterogeneous, multi-channel images from 75 diverse biological studies, built to train models that are “channel-adaptive and can process any microscopy image type”; the authors show that its sheer diversity of modalities is what improves multi-channel performance. It’s the same lesson we keep relearning in a new dialect: generalization is a data problem before it’s a model problem, and the route through is heterogeneity, in the open.

📏 The honest map of what’s left

For the wider view, a 2025 review from the group that invented Cell Painting — “Progress and new challenges in image-based profiling” (Serrano, … Carpenter, Singh, Caicedo, Way) — is exactly the kind of stocktake we value. It credits how far deep learning has pushed the field (single-cell analysis, robust similarity metrics, expansion into “optical pooled screening, temporal imaging, and 3D organoid profiling,” and “the growth of public benchmarks and open-source software ecosystems”), then names what’s still unsolved without flinching: “developing methods for emerging temporal and 3D data modalities, establishing robust quality control standards and workflows, and interpreting the processed features.” That last one is the deepest — an embedding can separate two conditions perfectly and still not tell you which piece of biology moved. A Cell Systems piece this year frames the road ahead as the shift from modality-specific to compositional, assay-agnostic foundation models — open, benchmarked, and not locked to one fixed panel. Why it matters for the lab: publish the model and the honest test — the BioImage Model Zoo and AI4Life ethos — is precisely how a field at this stage compounds instead of fragmenting.

Read the three together and the arc is familiar. A picture of a cell is one of the cheapest, richest signals in biology, and we’re finally learning to read it at scale — a fingerprint dense enough to reveal a drug’s mechanism or a protein’s home. The frontier isn’t a bigger model; it’s a model that works on your microscope, on your channels, and can tell you not just that the cell changed but how. Segment the cell, fingerprint the cell, then — the whole point — predict it. This is the measurement layer a virtual cell has to be judged against, and it’s coming into focus one honest benchmark at a time.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (X/Twitter sweep was skipped today — our news API is out of credits.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.