Drug-discovery presentations love a beautiful protein structure. The target is rendered as a solid ribbon, the binding pocket is a fixed cavity, and the candidate sits inside it like a key placed into a lock. The picture is useful. It is also incomplete in exactly the way that matters for pharmacology.

A protein in solution is not a statue. It is an ensemble of interconverting conformations distributed across an energy landscape. Ligands do not merely fit into a pre-existing rigid pocket; they can prefer rare states, shift state populations, alter motions, stabilize inactive conformations, promote active conformations, or change communication between distant sites.

What Anfinsen actually gave drug discovery

Christian Anfinsen's ribonuclease experiments established the thermodynamic hypothesis: under appropriate conditions, the amino-acid sequence contains the information required for a protein to adopt its biologically active structure. This was foundational because it connected sequence, physical chemistry, and function. But it is often simplified into the idea that every protein has one final shape.

The richer interpretation is an energy landscape. The native state is not necessarily a single microscopic arrangement. It can be a collection of closely related states, with populations determined by free energy and altered by temperature, solvent, binding partners, membranes, post-translational modifications, and ligands. Hans Frauenfelder, Peter Wolynes, and others helped make this landscape view central to protein biophysics.

Binding may select a state that was already there

Two classical descriptions of recognition are induced fit and conformational selection. In induced fit, ligand binding drives the protein into a new conformation. In conformational selection, the protein already samples multiple conformations and the ligand preferentially binds one, shifting the population toward it. Real systems can combine both mechanisms.

This matters because a therapeutically useful pocket may be invisible in the most populated apo structure. A cryptic pocket may open only transiently. An allosteric site may exist only in a particular signalling state. A mutation can change the population of states without dramatically changing the average structure. Static similarity can therefore hide dynamic pharmacological differences.

Activity is often a population shift

Many receptors, kinases, ion channels, transporters, and enzymes move among functional states. Agonists, antagonists, inverse agonists, biased ligands, and allosteric modulators can be understood partly by which states they stabilize. For a GPCR, two ligands with similar affinity can drive different downstream signalling because they reshape the receptor's conformational ensemble differently. For a kinase, inhibitors can prefer active or inactive conformations and thereby produce different selectivity and residence-time profiles.

A binding pose is a hypothesis about one state. A drug mechanism is a claim about how the distribution and kinetics of states change in biology.

Structure prediction solved one problem and exposed the next

AlphaFold-class models transformed access to high-quality structural hypotheses, especially when no experimental structure exists. That achievement is enormous. Yet a predicted structure should not be confused with a complete model of binding thermodynamics, dynamics, protonation, water networks, induced states, membrane context, or cellular complexes. Confidence scores describe aspects of structural prediction, not the probability that a proposed drug will achieve the desired pharmacology.

The next stage of structure-enabled discovery is therefore not “more docking to one structure.” It is integrating multiple structural states and multiple evidence types: experimental structures, predicted structures, molecular dynamics, NMR or HDX data, mutagenesis, biophysical binding, cellular target engagement, kinetics, and functional assays.

Residence time reveals why equilibrium affinity is not enough

Two molecules can have similar equilibrium affinity and very different association and dissociation rates. If one remains bound far longer, target suppression may persist after plasma concentration falls. Slow dissociation can be beneficial when sustained engagement is desired, but problematic when rapid reversibility is needed for safety. Kinetics also interact with target turnover, tissue distribution, and dosing interval.

This is another expression of the same principle: time belongs inside the mechanism. A static Kd, like a static structure, compresses dynamic behaviour into one number.

How to design with ensembles

  • Represent structural uncertainty: record whether a structure is experimental, predicted, homologous, ligand-bound, or modelled.
  • Generate mechanistically relevant states: active, inactive, open, closed, partner-bound, mutant, and post-translationally modified forms.
  • Test pocket persistence: distinguish stable pockets from transient or model-dependent cavities.
  • Measure kinetics and engagement: complement affinity with residence time and cellular target-engagement evidence.
  • Connect structure to function: ask whether stabilizing the proposed state changes the disease-relevant phenotype.

The conceptual shift is subtle but powerful. We are not designing a molecule for a structure. We are designing a perturbation of an energy landscape, under cellular conditions, that must propagate into a useful biological response. The best structural model is therefore not the prettiest picture. It is the ensemble that makes the next experiment more discriminating.