After training on data from more than 1.3 million human cells, the system became accurate enough to predict gene expression in cell types it had never seen, yielding results that agreed closely with experimental data.