De novo distillation of thermodynamic affinity from deep learning regulatory sequence models of in vivo protein-DNA binding

We have very strong evidence of real generalization here www.biorxiv.org/content/10.1... We explicitly compare to copying & marginalization of the training signal in the genome vs. marginalization using the NN predictions. 4/4

So what we know is that TF models trained on ChIP-seq from weird cell-lines do infact encode intrinsic affinity of TFs really quite well and that we can extract this very effectively. www.biorxiv.org/content/10.1...

I do agree that intrinsic sequence preference is a major component of genomic binding as we also show here www.biorxiv.org/content/10.1.... But I think claiming striking overlap is quite confusing IMO.