Fig. 2

Example of how generative modeling can be employed to hunt for genetic interactions. Most deep generative models are made of two elements: the encoder, which reduces dimensionality, and the decoder, which can generate genetic profiles in silico (top panel). We present three problems where generative models can be employed. Interpretability: The output of the encoder, and input of the decoder, can be interpreted and related to phenotypes of interest. Perturbation: A patient’s genetic profile can be perturbed in silico and passed through the encoder. For instance, a patient with two wild-type alleles (green circles) can be modified by induction of A or B (orange circles) or both at the same time (red circles). Study of the corresponding perturbation in the latent space can help prioritize potentially interacting genetic pairs. Optimization: Finally, a deep generative model could be directly employed inside an optimization strategy geared towards finding epistatic interactions, benefiting from two advantages of deep generative models: the auto-differentiation of the decoder and the continuous character of the latent space