44 interaction effects

[timcyu]

@jbloom I was wondering if I could get your thoughts on this notebook interaction_effects.ipynb. I simulated some simple combinatorial DMS data, where the escape scores were calculated with either:

• no interactions (exactly under polyclonal model)
• pairwise interaction effects between mutations at the same epitope (and everything else exactly under polyclonal)

I found that having some pairwise interactions (at least on this simple example) can lead to poor out-of-sample escape prediction, and make it possible to misinterpret the wildtype activity terms (and also the beta coefficients to a lesser extent).

If this is an issue, I do worry if this will make it difficult to reliably infer the wildtype activities and beta coefficients on real data, since we won't know the exact source of a poor escape prediction (ex. could be epistasis, other things not captured by polyclonal, or experimental noise). I also don't know if interactions are even something that can be feasibly added on into polyclonal, given the number of possible pairwise interactions between mutations. Although, I guess it could be feasible to only look at pairwise interactions between specific mutations, like the "key epitope" mutations in the H3 HA library.

Lastly, I have no intention at the moment of adding this notebook to the docs! I opened this PR as I just thought it'd be easier to share this way if anyone wants to build the docs for themselves and take a look.

[jbloom]

@timcyu, thanks for posting this. I think we just have to wait and see how things look with actual data. Obviously any model is going to only work as well as the data conform to the assumptions of the model. I think your simulations are useful in that they provide some indication of how deviations could look in principle, but we will just have to see in practice.

Do you want to convert this to a draft pull request if it isn't really meant to be merged. I'm not sure how to do that, but I know @matsen has done it for other pull requests.