This is a pretty good firing offense by itself

And cause it's interpretable, we can backpropagate the prediction error for the entire transcriptome to the transcription factor layer and figure out which TFs we're mispredicting (could continue fine-tuning the model this way, too). We explicitly measure that error in the context of TF activity.

It worked really nicely, with some biological constraints the decomposition is nice and clean, and you can find clear differences in decay vs transcription regulation

In order to make this work, we had to estimate RNA velocity and then break it out into separately modeled decay rates and transcription rates (so that changing TFs only changes transcription and not any other process).

We used that data to train a model, which takes proliferative cells and predicts the entire transcriptome as it responds to rapamycin.

We put together a single-cell training data set from yeast cells that are continually sampled for single-cell sequencing, which gives a beautiful time-dependent transcriptome curve.

Q: Can we make a general-purpose model of gene regulation? A tool where you can make changes to regulation and see what happens, or take observed effects and figure out what the cause is.

🤔