Fig. 1
From: BindVAE: Dirichlet variational autoencoders for de novo motif discovery from accessible chromatin
a Generative model captured by BindVAE. b The decoder weights \(\theta \in \mathbb {R}^{100 \times 112800}\) from the GM12878 model, where each row i contains the 8-mer distribution learned for the latent dimension i. Since visualizing the weights of all 112800 8-mers is difficult, the weights for a subset of the 8-mers are shown. This subset of 8-mers was picked to contain the top 20 8-mers of each latent dimension. See the text for details. c Motif analysis of each latent dimension shows that different binding specificities are captured by each latent factor in GM12878. Column 2 shows the name of the TF assigned by Algorithm 1, column 3 shows the CIS-BP motif corresponding to the TF, and the last column shows the motif that was found by HOMER using de novo motif discovery for the TF