AI-built 'Virtual Cells' could speed up drug discovery

Researchers at UC San Diego used 4D lattice light-sheet microscopy to build MitoSpace, a deep-learning model that mimics mitochondrial behavior. Trained on 40,000 single-cell movies from cancer cells treated with 25 compounds, MitoSpace increased drug classification accuracy to 75 percent from 56 percent using traditional 2D images, and can predict cell energy states without retraining on new drugs.

For a century we have believed that mitochondrial form reflects function; this shows the relationship is strong enough that a model can learn it without ever being shown the answer.