What Kd is, and why lower is tighter
Binding is an equilibrium. A ligand A and a target B associate to form a complex AB, and the complex comes apart again. Kd, the equilibrium dissociation constant, is the ratio of the free species to the complex at equilibrium — [A][B] divided by [AB] — which is also the off-rate divided by the on-rate, koff/kon.
Because Kd sits in the numerator of the free species, it has units of concentration, and that is what makes it intuitive. At a free ligand concentration exactly equal to Kd, half the target sites are occupied. A ligand with a Kd of 1 nM reaches half occupancy at a thousand-fold lower concentration than one with a Kd of 1 µM — which is the whole reason a lower number is the better one.
Kd is a property of the pair, not of the molecule alone. It is also a thermodynamic quantity: two ligands can share a Kd while differing entirely in how long they stay bound, because Kd only fixes the ratio of the two rates, not their magnitudes. For a construct that has to remain on a cell long enough to be internalised, residence time can matter as much as the equilibrium constant.
pKd, and why the field works in logarithms
pKd is the negative base-10 logarithm of Kd expressed in molar units. A Kd of 1 nM (10⁻⁹ M) is pKd 9; 100 pM is pKd 10; 1 µM is pKd 6. Each whole pKd unit is a factor of ten in affinity.
Three reasons the logarithm is the working scale. Affinity data spans six or more orders of magnitude, and a linear axis compresses everything interesting into the bottom corner. Experimental error in binding assays is roughly multiplicative rather than additive, so it is approximately constant in log space — an assay good to "a factor of two" is good to 0.3 pKd units whether the ligand is micromolar or picomolar. And free energy is linear in the logarithm: ΔG° = −RT ln(1/Kd), so pKd is proportional to binding free energy, which is what a physics-based method actually computes.
The practical consequence is that every correctness claim about predicted affinity should be stated on pKd. A correlation computed on Kd in nanomolar is dominated by whichever few ligands happen to be weakest, and is not the quantity anyone cares about. If you want to convert between the two on real numbers, the pKd calculator on this site does it both ways.
IC50 versus Kd: why they are not interchangeable
IC50 is the concentration of an inhibitor that halves a measured response — enzyme activity, signalling, cell viability — under one particular set of assay conditions. It is a potency readout, and it moves when the assay moves.
The classical demonstration is Cheng and Prusoff's 1973 analysis. For competitive inhibition of an enzyme, IC50 depends on how much substrate the inhibitor has to out-compete: Ki = IC50 / (1 + [S]/Km). Raise the substrate concentration and the same inhibitor, unchanged, reports a larger IC50. Their paper's own conclusion is blunt — Ki does not equal IC50 when competitive kinetics apply — and it warns explicitly against "the indiscriminate use of I50 values".
So the difference is not a matter of accuracy but of category. Ki and Kd are intrinsic equilibrium constants for a binding interaction; IC50 is a conditional answer that includes the assay's own competition. Two IC50 values from different laboratories with different substrate concentrations are not comparable, while two Kd values from well-run direct-binding measurements are. Cell-based IC50s add further layers — permeability, efflux, target expression — that have nothing to do with affinity.
Direct binding measurements avoid the problem by measuring association and dissociation instead of inhibition. Surface plasmon resonance and biolayer interferometry both track molecules accumulating on a sensor surface over time and derive kon, koff and therefore Kd. FEP+ and molecular dynamics approach it computationally, and note that free-energy perturbation is relative: it estimates the difference in binding free energy between two related ligands, not an absolute Kd.
nM and pM: what the numbers feel like
Some anchors, for a field where two digits change the decision:
- µM (10⁻⁶ M) — a fragment or a weak hit. Half occupancy needs a concentration no systemic drug reaches at its target for long.
- 100 nM (pKd 7) — a credible starting point from a display selection or a first generative round.
- 1 nM (pKd 9) — the neighbourhood of approved targeting ligands; tight enough that occupancy is limited by delivery rather than by affinity.
- 100 pM (pKd 10) and below — engineered or affinity-matured binders. Beyond this point tighter binding stops buying tumour uptake and starts buying slower clearance from everything else.
Why ranking by predicted affinity works — and where it stops
Predicting an absolute Kd is hard. Ranking a set of candidates against one target is substantially easier, and ranking is what a design campaign actually needs: from ten thousand generated sequences, which twenty are worth synthesising.
The reason is that a ranking cancels systematic error. A model that is uniformly optimistic by a factor of five still puts the right molecules near the top. That is why the useful evaluation metrics are rank-based or top-k — how many of the true best ligands appear in the predicted top twenty — rather than absolute error, and why a correlation quoted on pKd is more informative than a mean error in nanomolar.
Epitorix's published figures are of that shape and carry their fences: predicted pKd correlates with measured pKd at Pearson R 0.84 for peptides and 0.70 for small molecules in internal benchmarks, at about 1,000 molecules per second; and given 100 generated ligands, ranking recovers more than 80 % of the true top ligands in internal benchmarks. The ranking product itself is Xenoclea, and the beat that publishes the numbers is affinity ranking on the science page.
Four limits belong next to any such number, and they apply to every vendor in this space equally.
- Correlation is not rank order. Pearson R measures linear agreement; if a ranking claim is being made, Spearman's ρ or a top-k recovery rate is the honest statistic.
- Generalisation depends on the split. A model evaluated on ligands similar to its training data will look better than it is; the field's answer is similarity-aware partitioning, and a reported correlation without a described split is not interpretable.
- Affinity is one objective of several. Published reviews of AI-designed binders list avidity and expression artefacts, polyreactivity, aggregation and off-target binding as documented failure modes, and note that affinity optimisation can degrade stability and solubility — which is why multi-objective or Pareto filtering is preferred.
- For this modality, affinity is not the endpoint at all. A tight binder with the wrong clearance route irradiates the kidney. That is a biodistribution question, and no affinity model answers it.
