Lab Newsletter — September 25, 2026: Reading a Cell's Future from Live Images

This week we’ve worked at the level of molecules and whole-cell identity. Today we come home to the lab’s own instrument — the microscope — and a quietly astonishing idea: a cell’s future is often visible in its present shape and motion, well before any stain, marker, or reporter reveals it. Today’s digest is about deep learning that forecasts fate, cycle stage, and behavior from live, often label-free, imaging — the perception layer beneath a self-driving microscope that can watch, predict, and decide in real time.

🔬 First, see every cell over time

You can’t predict a cell’s future until you can reliably find it in every frame. Van Valen et al. (PLoS Computational Biology, 2016) laid that groundwork with DeepCell, showing that “deep convolutional neural networks … can robustly segment the cytoplasms of mammalian cells … from phase contrast images without the need for a fluorescent cytoplasmic marker,” and do so “across multiple cell types across the domains of life.” Turning raw live-cell movies into per-cell measurements — no labels required — is the substrate everything else here builds on, and the same automated-analysis spirit as the lab’s imaging tools.

🩸 Predict lineage — generations early

The headline result is almost eerie. Buggenthin et al. (Nature Methods, 2017) built a network that “prospectively predicts lineage choice in differentiating primary hematopoietic progenitors using image patches from brightfield microscopy and cellular movement.” The finding that stopped people short: “lineage choice can be detected up to three generations before conventional molecular markers are observable.” The commitment a cell will make is already encoded in how it looks and moves — the model just reads it, “without molecular labeling.”

⏱️ Catch differentiation in minutes

How early can you catch a fate decision? Waisman et al. (Stem Cell Reports, 2019) trained a CNN on “transmitted light microscopy images to distinguish pluripotent stem cells from early differentiating cells,” reaching “an accuracy higher than 99%.” Remarkably, “successful prediction started just 20 min after the onset of differentiation” — long before any standard assay would register the switch, and robust across settings including “mesoderm differentiation in human induced PSCs.” Non-invasive, live, and early enough to act on.

👁️ Call organoid fate before the reporter lights up

The same trick scales to 3D tissue. Kegeles et al. (Frontiers in Cellular Neuroscience, 2020) built a CNN that predicts retinal differentiation in stem-cell organoids “based on bright-field imaging,” and — the key word — “before the onset of reporter gene expression.” It beat a human expert ("84% vs. 67 ± 6% of correct predictions"), in what they call “the first demonstration of CNN’s ability to classify stem cell-derived tissue in vitro.” A non-invasive, reporter- free readout of where an organoid is heading — exactly the kind of live decision a smart microscope wants.

🔄 Reconstruct the hidden clock

Fate is discrete; many processes are continuous. Eulenberg et al. (Nature Communications, 2017) showed that “deep convolutional neural networks combined with nonlinear dimension reduction enable reconstructing biological processes based on raw image data,” demonstrating it by “reconstructing the cell cycle of Jurkat cells and disease progression in diabetic retinopathy.” It even separated dying cells “in an unsupervised manner” — and, crucially for live use, ran “fast enough for on-the-fly analysis in an imaging flow cytometer.” Recovering a cell’s position along a hidden timeline, from a single snapshot.

🧠 …and explain what the model sees

The obvious objection to all this is the “black box.” Zaritsky et al. (Cell Systems, 2021) met it head-on, pairing “a generative neural network … with supervised machine learning” to classify melanoma xenografts as “’efficient’ or ‘inefficient’ metastatic” from label-free live images, then using the generator to synthesize “in silico cell images that amplify the critical predictive cell properties.” Those exaggerated images “unveiled pseudopodial extensions and increased light scattering as hallmark properties of metastatic cells” — features “too subtle to be identified in the raw imagery by a human expert.” Prediction and a biological explanation of what the cell is telling us.

🧫 Why it’s our kind of problem

Read across the six and it’s the lab’s thesis in miniature. First, this is the perception layer of the self-driving microscope: a model that reads fate and state from live images in real time lets the instrument decide — where to look, when to image, when to intervene — the closed loop that Agent-Lens and the REEF farm are built to run. Second, predictive live imaging is a cheap, non-invasive path toward the virtual cell: forecast phenotype from a cell’s observable morphology and motion, at scale, without perturbing it. And third, Zaritsky’s interpretable, generative approach is the lab’s answer to the “black box” worry — models that don’t just predict but show their reasoning, the same value behind open, inspectable tools like the BioImage Model Zoo. A cell, it turns out, keeps telling you where it’s going. The work here is learning to listen — live, label-free, and early.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (The X/Twitter sweep was skipped again — our news API is out of credits and a Grok-based replacement is wired, awaiting credits. Anchors were verified via NCBI E-utilities.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.