AI Analyzes 70 Million Cells to Identify Regulators of Autophagy and STING

About 70 million cells imaged under a microscope during a genetic screen became the training and analysis material for an AI system that searches for genes regulating autophagy and the STING signaling pathway. Jiqizhixin reported on the work on October 12, 2026. The idea is to read from an image what used to require a separate fluorescent reporter for each process.

Why cell shape is so hard to analyze

Almost any change inside a cell leaves a visible trace. An organelle shifts, an atypical structure appears, the number of vesicles grows, and behind each of these signs there may be a specific biological process. The problem is that several genes and several cellular programs affect cell shape at once. The phenotype is multidimensional, not simply "on" or "off."

Fluorescence is more convenient in this respect: the signal is either there or it is not, and setting a threshold is easy. Morphology cannot be labeled that way. Two knockouts can produce similar images for different reasons, and the same knockout can look different in neighboring cells. These are exactly the tasks where image neural networks have an advantage over hand-crafted features.

What is known about the work

The available excerpt of the source discloses only the basic parameters. It does not give the model's name, the authors, the architecture, or quality metrics.

• Data scale: about 70 million cells.
• Data format: microscope images of cells after genetic perturbations.
• What is analyzed: morphological features, including the position of organelles, abnormal intracellular structures, and the number of vesicles.
• Biological targets: autophagy, meaning the disposal of damaged cell components, and the STING pathway, which is linked to the innate immune response.
• Result: a list of genes that the system associates with the regulation of these processes.

Scale matters here for a practical reason. A screen includes many cells per perturbation, and the morphological noise between them is high. Without tens of millions of observations, it is hard for a model to separate a gene's real effect from random differences between cells and between sample batches.

How to evaluate the result

With so few published details, it is too early to draw conclusions about accuracy. Biotech teams and ML engineers following this kind of work have several checks to apply.

First: whether independent experiments confirm the genes found, for example a targeted knockout with a classic autophagy reporter. A screen without validation remains a list of hypotheses.

Second: how well the model transfers to other cell lines and other microscopes. Morphological models readily memorize the quirks of a specific setup, which is a common source of inflated metrics.

Third: whether the data and code are open. A set of 70 million cell images is valuable in its own right, because it allows comparing alternative architectures and trying other targets.

If the approach holds up, it will make it cheaper to find regulators of processes that have no convenient fluorescent marker. STING, for example, interests the pharmaceutical industry as a target in immunology and oncology. A team building its own screening pipeline should plan ahead to build batch controls and uniform staining protocols into its imaging. That way the morphological model will learn from biology rather than from equipment artifacts.