AI predicted from video which mouse blood stem cells would become “immortal”, and blocking one signal in culture increased their proportion
AI predicted from video which mouse blood stem cells would become “immortal”, and blocking one signal in culture increased their proportion
On October 7, a team at Fred Hutchinson Cancer Center in Seattle released a preprint describing a neural network trained on videos of individual blood stem cells. It predicts which cells will become long-lived with 96,7–98,5% accuracy, without destroying the cells for analysis. The study also identified a molecular signal that controls this outcome and experimentally confirmed that blocking it in culture increases the proportion of long-lived cells.
Blood stem cell transplantation treats leukemia and other blood disorders: radiation or chemotherapy destroys the patient's bone marrow, after which donor cells are infused. The procedure originated at Fred Hutch, where E. Donnall Thomas received half of the 1990 Nobel Prize in Physiology or Medicine for this work. Decades later, a problem remains: only a small fraction of the cells in a transplant can sustain blood production throughout life. Most are progenitor cells with limited capacity, and distinguishing the two beforehand is difficult. Sorting cells by protein markers yields a mixed population, while single-cell genetic analysis captures the difference at a single moment rather than following the process by which cells acquire their fate.
A year earlier, the same laboratory used the same culture system to show that genuinely long-lived cells divide symmetrically, producing two stem cells of the same kind rather than one stem cell and one specialized cell, and remain metabolically quiescent.
“Our work challenges the traditional view that blood stem cells are highly active during development,” laboratory head Brandon Hadland explained at the time in a Fred Hutch article.
He also outlined the next step:
“Looking ahead, we are also developing machine learning tools to analyze live-cell videos from our platform and predict these cells' fate from subtle features of their behavior. This could change how we select the most potent cells for clinical use.”
The laboratory's preprint puts that plan into practice. The researchers isolated blood stem cells and progenitor cells from fetal mouse liver, where these cells multiply before birth rather than in the bone marrow. They plated individual cells onto a layer of cells that mimicked this niche and recorded video every 2–3 hours for two to three days. They trained the neural network on 80% of the cells and tested it on the remaining 20%. Whether a cell became long-lived was established later through serial transplantation into mice. The model predicted cell fate with 96,7–98,5% accuracy. Low motility was the strongest predictor: cells that would become long-lived moved less, and their descendants retained this behavior after division.
The authors traced this low motility to signals from neighboring cells. Cells that would become long-lived received less signaling through the protein Activin and showed greater activation of cell adhesion genes, a program that keeps cells in place. To test this, they added Inhibin-A, a natural blocker of Activin, to the culture. Cell motility decreased, and the proportion of long-lived cells increased after two weeks. The authors describe the effect as modest but statistically significant across four independent experiments.
The authors identify adapting the protocol to human cells and clinical use as their next goal. Video-based prediction and a molecular signal that can be experimentally controlled already provide a framework for selecting and growing the desired blood stem cells. The authors suggest that this approach could eventually extend to other stem cell types.