With the goal to build a virtual screening tool as accurate as long and costly exptl. high throughput screening (HTS), we have developed the empirically-trained, ligand-based Profile-QSAR method, which combines all AC50 data across entire protein families. Thus, each prediction is informed by millions of AC50s covering hundreds of thousands of compounds This greatly improves the accuracy and range of chem. structures, enabling performance comparable to exptl. HTS. We have applied Profile-QSAR to over 50 drug design projects. Originally developed for kinases, we have recently extended it to GPCRs, proteases, CYPs and non-kinase adenosine binding proteins. We have predicted the activity for millions of Novartis compounds across thousands of assays. These predictions have been used in virtual screening, drug safety profiling and mode of action (MOA) fishing. Profile-QSAR can only be applied to drug targets from large protein families, but what constitutes a family? Profile-QSAR models rely on structure-activity relationship (SAR) rather than evolutionary (sequence) or protein function relationships, as demonstrated by kinase-based profile-QSAR models that also work well for many addnl. adenosine-binding proteins from other traditional families: chaperones, carboxylases, phosphodiesterases, ABC transporters, ion channels, and others. This suggests building Profile-QSAR models based on "SAR-families" formed from clustering proteins on exptl. HTS data may allow accessing addnl. drug targets outside the large traditional families. Using Profile-QSAR predicted activities, we also can compute an SAR-based dendrogram for the human kinome and compare it to the sequence (evolution)-based Sugen kinome dendrogram.