With this, we want to show that capturing family-wide mutation effects is essential to allow the training of accurate quantitative models of protein-protein interactions.

With out extensive dataset, we trained a neural network that could accurately predict binding from sequence and even enable the design of highly specific synthetic bZIPs.

We found that mutation effects are mostly additive in the context of close family members but more complex in distant regions of the sequence landscape indicating that the genetic architecture of protein function is more complex than previously thought.

In total, we quantitatively measured about 2M interactions of all possible single amino acid mutants of all 54 human leucine zipper domains with their wild-type interaction partners.