AI-Driven Cell-Fate Prediction in Microscopy

Microscopy is evolving from a descriptive to a predictive tool in biology. Trained on time-lapse images, deep learning can forecast whether a cell will divide, die, or differentiate from how it looks and moves, in some systems, hours or even generations before the usual molecular markers appear, and without added fluorescent labels. Two questions frame this review and stay largely open: which image features actually carry the predictive signal, and whether a model foresees a genuinely future outcome or instead reads a state the cell has already entered. We organise the field by the visual signal that carries fate: changes in cell and nuclear shape over time, changes in brightness and texture, the dynamics of differentiation and competition, and patterns of movement. For each, we ask what must be measured, and over what area and time window, to predict fate. We then set out the main families of models, from networks that read a single image to those that read sequences of images over time and those that compress images into compact numerical summaries, outlining their strengths and limitations. Our central argument is that prediction has outrun validation: most reported performance is checked retrospectively against endpoint markers instead of on genuinely future cells. We close on the shared datasets, benchmarks, and interpretability work this gap demands, and on underexplored fates such as migration and senescence.