When you design from the ligands, not the target
Ligand-based drug design is what you do when no good target structure exists — a common situation for membrane proteins and novel targets. Instead of looking at the pocket, you study a set of molecules whose activity you already know and ask: what do the active ones share that the inactive ones lack? The whole discipline of structure–activity relationship (SAR) lives here, now formalized so a computer can reason about it.
A pharmacophore is the abstract idea at the center: the 3D arrangement of features — an H-bond donor here, an acceptor there, a hydrophobic group, an aromatic ring, a positive charge — that a molecule must present to be active. A pharmacophore model turns that idea into a concrete query: overlay several known actives, find the features they share in space, and use the resulting pattern to search libraries or to judge new designs.
Making molecules into numbers
A computer cannot reason about a drawing of a molecule; it needs numbers. A molecular descriptor is any computed property — molecular weight, logP, count of H-bond donors, polar surface area, ring count — that summarizes one facet of a structure. A molecular fingerprint is a different representation: a long bit-string that flags which substructures or atom environments a molecule contains, so two molecules can be compared for similarity by how many bits they share.
QSAR — quantitative structure–activity relationship — is the model that maps these numbers to activity. In its classic form it is a regression: activity ≈ a weighted sum of descriptors. A QSAR model lets you predict the potency of a molecule you have not yet made, and, just as usefully, it tells you which descriptors drive activity — pointing your chemistry in a direction.
Toy QSAR (illustrative only): pIC50 = 6.2 + 0.8 * (clogP) - 1.1 * (#H-bond donors) Read it as: in THIS series, more lipophilicity helps a little, but each extra H-bond donor hurts potency. Use it to rank ideas before synthesis -- never as physical truth.
Where ligand-based models break
These models only know what they were trained on. A QSAR built on one scaffold can be badly wrong for a different one, and every model has an applicability domain — the region of chemical space where its predictions are trustworthy. The sharpest failure is the activity cliff: two molecules that look almost identical to a fingerprint yet differ wildly in potency. Cliffs violate the smooth assumption behind most QSAR, and they are exactly where a single methyl or fluorine flips activity on or off.