Active learning to discover pairwise genetic interactions via representation learning

Embeddings of microscopy images from single gene knockouts can be used to infer biological interactions, but are limited to interactions that are revealed by single perturbations. If we want to detect effects that are only present in pairwise knockouts we need to (1) address the quadratic scaling of experimental costs, and (2) develop a method for detecting pairwise interactions. We present a set of theoretical independence assumptions under which the sum of embedding of single perturbations predicts the pairwise embeddings. Prediction failures then correspond to violations of these assumptions, and can be used to detect biological interactions. We used this prediction error as a reward in an active search algorithm and found that we can efficiently identify these instances of non-independence, and many of the selected pairs correspond to known biological interactions.