Framework of the proposed VADLP model. Given input data, (A) triple-layer heterogeneous graph is constructed to associate the similarities and correlations across lncRNAs, miRNAs and diseases with inter- and intra-layer weighted edges. Three representations are learned including (B) pairwise topology encoding by random walk and CAE, (C) node attributes by a proposed heterogeneous embedding mechanism and CAE and (D) feature distribution representation and encoding by VAE. The three representations are adaptively fused by (E) attentional representation-level integration for final lncRNA–disease association prediction. Details of each component are given in Fiures 2–4.