Explore consequences of interaction effects between mutations

[timcyu]

I am opening this issue for myself to investigate as I think more deeply about how to handle interaction effects. It occurred to me that cases of epistasis (particularly negative) between mutations could make it difficult for Polyclonal to learn "true" mutation effects.

As a simple illustration:

• mutation A has a +5 effect on antibody escape at epitope 1.
• mutation B also has a +5 effect on antibody escape at epitope 1.
• mutation A and B together have an additional -5 effect on antibody escape due to some negative epistatic interaction.

Since Polyclonal does not fit any interaction terms, I worry that these types of interactions between mutations at an epitope can obscure effects of individual mutations, leading to potential misinterpretation of escape mutations (ex. mutation A or B being given a much lower individual effect than the ground truth). I am also interested in how these types of interactions, which we should expect in nature to some extent, will affect escape prediction of unseen variants.

For now, I am going to simulate some very basic data with and without interactions and compare how Polyclonal does at 1) inferring the true individual mutation effects and 2) predict escape of unseen variants.

[jbloom]

@timcyu, I'm just trying to clean up the issues. Can we close this one, since at this point it seems to be more of a conceptual question about epistasis outside the framework of polyclonal than a specific actionable issue?

[timcyu]

Yes, feel free to close it.