The identification of small molecules binding to biological targets is the central dogma of drug discovery, a challenging endeavour marked by a failure rate exceeding 90%. Structure-based virtual screening is a ligand discovery tool used to triage chemical libraries against the structure of a protein target at a much larger scale than wet-lab screens. Recent advances in chemoproteomics are rapidly mapping all the potential druggable sites encoded by the human genome, collectively known as the human pocketome. Concurrently, innovative approaches like ultra-large chemical libraries and artificial intelligence (AI) methods are paving the way for a new era in drug discovery. Yet only 3% of human proteins have been targeted with drug-like molecules to date. Indeed, emerging targets lacking known ligands remain very challenging for methods like docking and AI generative design, due to the arduous nature of identifying target conformations suitable for computational design when no confirmed ligands are known. To fill this gap, we seek to leverage machine learning (ML) to accurately predict druggable structures of binding sites from experimental or computational ensembles, without requiring prior knowledge of ligands nor human intervention. To this end, we first focus on developing a structural database of druggable conformations for known human protein targets, starting by developing compelling decoy molecule sets that closely simulate prospective library screens. The conformational druggabilities of the targets are then explored by combining biomolecular simulations and large-scale docking of active-decoy libraries. As such, the resulting database is a valuable resource for training ML models for forecasting structural druggability, thus guiding target selection for optimized computational ligand screening and design. Our work will open the door to the investigation of many new potential drug targets, including those modeled with Alpha Fold (AF) that represent an unprecedented, yet untapped resource in the field.