Fig. 1: Overview TRAPT.

a The simplified flowchart of TRAPT indicates that the process of model inference comprises three components: an input gene set, an inference by the model, and TR activity as the output. TRAPT only requires a gene set as input and provides results through online or offline methods, with the online service providing enhanced features of visualization. b The TR-RP regulatory potential model was used to calculate the TR-RP matrix, and the epigenome-RP regulatory potential model was used to compute the Epi-RP matrix. c TRAPT predicts the regulatory potential associated with the genome-wide binding sites of the TRs. The inputs to TRAPT consist of preprocessed TR-RP and Epi-RP matrices, which are integrated to form a regulatory potential matrix and an adjacency graph. We first used a conditional variational autoencoder as the teacher network to learn the latent representation h. Subsequently, a graph variational autoencoder was applied as the student network to reconstruct the TR-epigenome adjacency graph, enabling it to learn its own network structure and latent feature representation from the teacher network. Finally, we performed an aggregation operation using the reconstructed TR-epigenome adjacency graph and the input Epi-RP matrix to obtain the matrix of the regulatory potential associated with the genome-wide binding sites of the TRs. d TRAPT predicts the upstream regulatory potential associated with the query gene. First, epigenomic samples are grouped based on their correlation with the query gene vector. Then, the teacher model extracts feature maps to guide the student model in selecting non-redundant samples using SGL constraints. Finally, a nonlinear neural network model is retrained to generate an upstream regulatory potential matrix. e The predicted matrices of the regulatory potential associated with the genome-wide binding sites of the TRs, the regulatory potential of the target gene cis-regulatory elements, and the TR regulatory potential are integrated through matrix operations to obtain the I-RP matrix. f The AUC score of each TR sample in the I-RP matrix was first computed with the queried gene set. The TR AUC scores from all epigenetic groups were integrated to derive the final activity score for each TR.