Virtual screening (VS) is an essential technique for understanding biomolecular interactions, particularly, drug design and discovery. The best-performing VS models depend vitally on three-dimensional (3D) structures, which are not available in general but can be obtained from molecular docking. However, current docking accuracy is relatively low, rendering unreliable VS models. We introduce sequence-based virtual screening (SVS) as a new generation of VS models for modeling biomolecular interactions. The SVS model utilizes advanced natural language processing (NLP) algorithms and optimizes deep K-embedding strategies to encode biomolecular interactions without invoking 3D structure-based docking. We demonstrate the state-of-art performance of SVS for four regression datasets involving protein-ligand binding, protein-protein, protein-nucleic acid binding, and ligand inhibition of protein-protein interactions and five classification datasets for the protein-protein interactions in five biological species. SVS has the potential to dramatically change the current practice in drug discovery and protein engineering.