PTF-Vāc: An explainable and generative deep co-learning encoder–decoder system for ab initio discovery of plant transcription factor binding sites

Discovery of transcription factor (TF) binding sites (TFBSs) and their motifs in plants poses significant challenges due to high cross-species variability. The interactions between TFs and their binding sites are highly specific and context dependent. Most existing TFBS-finding tools are not sufficiently accurate to discover these binding sites in plants. They fail to capture cross-species variability, interdependence between TF structures and corresponding TFBSs, and the context specificity of binding. Because they are coupled to a predefined TF-specific model/matrix, they are strongly influenced by the volume and quality of data provided to build the motifs. All these software applications make the assumption that the user input is specific to a particular TF, limiting their use for practical applications such as genomic annotation of newly sequenced species. Here, we report PTF-Vāc, an explainable deep-learning encoder-decoder generative system. PTF-Vāc is based on PTFSpot, a universal model of deep co-learning on variability in binding sites and TF structure, making it completely free from the bottlenecks described above. PTF-Vāc decouples the process of TFBS discovery from the prior step of motif finding and the requirement for TF-specific motif models. Because it is guided by a universal model for TF-DNA interactions, it can discover binding motifs irrespective of data volume and species and without reference to TF-specific models. In a comprehensive benchmarking study across an extremely high volume of experimental data, PTF-Vāc outperformed most advanced motif-finding deep-learning algorithms. PTF-Vāc thus opens a completely new chapter in ab initio TFBS discovery through generative AI.