Cells respond to environmental and developmental stimuli by changing their transcriptomes through regulation of both mRNA transcription and mRNA decay. Our understanding of these processes has relied on careful perturbation experiments, and on metabolic labeling methods (e.g. 4-thiouracil) that allow transcription and decay to be separately measured. In this work, we replace complex experimental screens with anin silicomodel that makes time-dependent RNA kinetic predictions and can be internally perturbed, to simulate transcription factor changes. We acquire model training data by sequencing the transcriptomes of 175,000 individualSaccharomyces cerevisiaecells that are continuously sampled from a population without metabolic labeling. The rates of change for each transcript are calculated on a per-cell basis to give smooth biological estimates of RNA velocity. We then train a deep learning model with this transcriptomic and velocity information to make time-dependent predictions about RNA production (via transcription) and RNA decay. By separating RNA velocity into transcription and decay, we determine that rapamycin treatment causes existing ribosomal protein transcripts to be very rapidly destabilized, while transcription of new transcripts gradually slows over the course of an hour. The model framework is carefully designed so that causal regulatory relationships between transcription factors and their genes can be inferred, which is demonstrated by evaluating the model in the context of known regulatory relationships. Transcription factors are perturbedin silico, and the transcriptome-wide effect of perturbations on rapamycin response is predicted and compared to observed perturbation data collected at a specific timepoint.