Loading field map
How does anyone know which molecule will treat a disease? For most of history, the answer came from trial and error—chewing a bark, brewing a root, or testing a mineral against a symptom. The central pressure driving medicinal chemistry has been the need to make that search less accidental and more predictive. Each major framework in the field's history represents a different answer to the same question: what kind of knowledge turns the hunt for a drug into a disciplined science rather than a lucky find?
The earliest systematic approach to drugs was Materia Medica, a tradition stretching from antiquity through the eighteenth century. Materia Medica was essentially an empirical catalog: plants, minerals, and animal products were collected, dried, and recorded alongside their observed effects on the body. A physician consulting Dioscorides or a medieval herbal knew what had worked before, but had no way to explain why, nor any method to improve on nature's offerings. The framework was descriptive, not manipulative.
Iatrochemistry, which flourished from the early 1500s to the late 1600s, transformed that stance. Iatrochemists—most famously Paracelsus—argued that disease was a chemical imbalance and that drugs should be prepared in the laboratory, not simply gathered from the field. They introduced mineral remedies, distillation, and the idea that the chemist's craft could produce medicines more potent than any plant. Iatrochemistry did not replace Materia Medica overnight; the two coexisted for generations, with herbal remedies remaining central to pharmacy while chemical preparations slowly gained ground. But iatrochemistry planted a crucial seed: the active substance in a drug could be isolated, concentrated, and deliberately modified.
The Isolation of Active Principles framework, dominant through the 1800s, made that seed grow. Chemists like Friedrich Sertürner (who isolated morphine from opium) showed that a single purified compound—not the whole plant—carried the therapeutic effect. This was a narrowing move: instead of studying complex mixtures, medicinal chemists now focused on individual molecules. Isolation gave the field a concrete object of study and made possible the next conceptual leap.
If a pure molecule could produce a specific biological effect, then it must be interacting with something equally specific in the body. That intuition crystallized as Receptor Theory, first articulated by Paul Ehrlich around 1900 and still foundational today. Ehrlich proposed that drugs act by binding to cellular "receptors"—molecular targets that recognize particular chemical structures. Receptor Theory transformed medicinal chemistry from a descriptive craft into a mechanistic science. It also created a new question: if a drug's shape determines whether it fits a receptor, then systematically varying that shape should reveal which structural features matter.
That question drove Structure-Activity Relationships (SAR) , a framework that emerged alongside Receptor Theory and remains the empirical backbone of the field. SAR is the practice of synthesizing a series of related compounds, measuring their biological activity, and mapping how changes in chemical structure alter that activity. Early SAR work was entirely manual: chemists made analogues one by one, tested them in animal or tissue assays, and looked for patterns. The framework did not replace Receptor Theory; it operationalized it, turning the receptor concept into a practical experimental program. SAR is still indispensable today—even the most advanced computational pipeline ultimately depends on SAR data to train its models.
By the mid-twentieth century, SAR had accumulated thousands of empirical observations but lacked a way to predict activity for untested compounds. Quantitative Structure-Activity Relationships (QSAR) , pioneered by Corwin Hansch in the 1960s, addressed that gap. QSAR uses mathematical equations to correlate molecular descriptors—such as lipophilicity, electronic properties, or steric bulk—with biological activity. A QSAR model allows a chemist to estimate how active a new compound will be before synthesizing it. This was a major shift: SAR told you what had happened; QSAR promised to tell you what would happen.
QSAR did not make SAR obsolete. Instead, it absorbed SAR's logic into a more formal, predictive framework. The two coexist today, with QSAR providing the quantitative backbone for many ligand-based design strategies. But QSAR also revealed a limitation: its models were only as good as the descriptors and data used to build them, and they struggled when the receptor structure was unknown or when activity depended on complex three-dimensional interactions.
Rational Drug Design, which took shape in the 1970s, was a broader philosophical commitment rather than a single technique. Its core claim was that drug discovery should be hypothesis-driven: instead of screening random compounds and hoping for a hit, researchers should start from a biological hypothesis about the disease target, design molecules to interact with that target, and then test the design. Rational Drug Design is not an empty umbrella term; it coordinates a research program that insists on mechanistic reasoning at every step. Under this umbrella, two major branches developed, distinguished by the kind of information they use to guide design.
Ligand-Based Drug Design (LBDD) , emerging in the 1970s, works when the receptor structure is unknown but a set of active ligands is available. LBDD methods—pharmacophore modeling, QSAR, and later three-dimensional shape-matching—infer the receptor's requirements from the molecules that bind to it. The framework extends QSAR's logic into three dimensions and adds computational tools for aligning and comparing molecular shapes. LBDD is strongest when the target is a membrane protein or ion channel that resists structural determination.
Structure-Based Drug Design (SBDD) , which matured in the 1980s with the rise of protein crystallography and NMR spectroscopy, takes the opposite approach. If the three-dimensional structure of the receptor is known, chemists can design molecules that fit directly into the binding site. SBDD made possible the visual inspection of drug-target complexes, computational docking, and iterative cycles of design, synthesis, and structural feedback. The first drugs developed with significant SBDD input included HIV protease inhibitors and several kinase inhibitors. SBDD did not replace LBDD; the two frameworks coexist and are often used in tandem—LBDD to generate initial ideas when the structure is unknown, SBDD to refine those ideas once the structure becomes available.
By the 1990s, rational design had produced notable successes but also faced a bottleneck: even with a good hypothesis, synthesizing and testing enough compounds to find a lead was slow. Combinatorial Chemistry and High-Throughput Screening (HTS) , which peaked between 1990 and 2010, responded by scaling up. Combinatorial chemistry allowed the rapid synthesis of thousands or millions of related compounds in parallel, while HTS automated the biological testing. The framework was a deliberate narrowing of rational design's hypothesis-driven ideal: instead of designing each molecule carefully, you made a huge library and let the assay pick the winners.
Combinatorial chemistry and HTS did not reject rational design; they coexisted with it, often in the same company or laboratory. But the framework's limitations became clear over time. Many combinatorial libraries produced compounds that were chemically similar to each other, lacked drug-like properties, or failed in later development. The hit rates were often disappointing relative to the scale of effort. By the early 2000s, the pharmaceutical industry began to pull back from brute-force screening and look for more efficient strategies.
Fragment-Based Drug Design (FBDD) , introduced around 1996, offered a middle path between rational design and massive screening. Instead of screening large, drug-sized molecules, FBDD screens small, low-molecular-weight fragments—typically fewer than 20 heavy atoms—that bind weakly to the target. Because fragments are small, a library of a few thousand fragments can cover much more chemical space than a library of millions of larger compounds. Once a fragment hit is identified, it is grown or linked to other fragments to build a potent drug molecule. FBDD depends heavily on structural biology (a direct inheritance from SBDD) because detecting weak fragment binding usually requires X-ray crystallography or NMR. The framework has produced several approved drugs, including vemurafenib for melanoma. FBDD did not replace HTS; it added a complementary tool that works especially well for targets where HTS had failed.
Machine Learning in Drug Design, which has accelerated since 2010, represents the most recent transformation of the field. Machine learning (ML) models—neural networks, random forests, graph convolutional networks, and large language models—are trained on massive datasets of chemical structures and biological activities to predict properties such as potency, toxicity, solubility, and metabolic stability. ML extends QSAR's quantitative ambition but on a vastly larger scale: where QSAR used a handful of hand-crafted descriptors, ML can learn relevant features directly from the data, including from molecular graphs, three-dimensional conformations, and even text from the scientific literature.
ML does not replace earlier frameworks; it absorbs and accelerates them. Modern drug discovery pipelines often use ML to prioritize compounds for synthesis, to design libraries for HTS, to predict fragment binding modes, and to generate novel molecular structures through generative models. The relationship between ML and QSAR is one of transformation: ML is QSAR's quantitative logic writ large, freed from the constraint of linear equations and small datasets. At the same time, ML inherits SAR's dependence on high-quality experimental data—garbage in, garbage out remains a hard limit.
No single framework dominates medicinal chemistry today. The field is deeply pluralistic, with different frameworks leading in different contexts. Receptor Theory remains the conceptual foundation for almost all drug discovery; every framework from SAR to ML assumes that drugs act through specific molecular targets. SAR is still the daily work of most medicinal chemists, who synthesize and test analogues in iterative cycles. QSAR and LBDD are standard tools for early-stage hit-to-lead optimization, especially when target structures are unavailable. SBDD is the method of choice for targets with well-characterized structures, such as kinases and proteases. FBDD has carved out a niche for challenging targets where HTS failed. ML is rapidly becoming a universal infrastructure layer, used to prioritize, predict, and design across all the other frameworks.
What the leading frameworks agree on is that drug discovery must be data-driven and iterative: hypotheses are tested by synthesis and assay, and the results feed back into the next design cycle. Where they disagree is on the optimal source of that data. LBDD and QSAR trust ligand-derived information; SBDD trusts structural information; ML trusts large-scale statistical patterns; HTS trusts brute-force empirical coverage. The practical art of medicinal chemistry today lies in knowing which framework to apply to which problem—and in combining them when no single approach is sufficient. The field's history is not a story of one framework triumphing over another, but of an expanding toolkit, each tool shaped by the limitations of the ones that came before.