Binding affinity
Predicts the binding free energy and dissociation constant for the interface between two sets of chains in a complex, using PRODIGY. Use to compare how tightly related complexes bind. It assumes the complex is real and says nothing about whether two proteins interact.
Binding affinity
What it does
Takes a structure containing two binding partners, and the chains on each side of the interface, and predicts how tightly they bind — as a free energy (ΔG, in kcal/mol) and a dissociation constant (Kd, in molar).
More negative ΔG means tighter. Smaller Kd means tighter.
How the prediction is made
The engine is PRODIGY. It counts the contacts across the interface, split by the chemical character of the residues involved — charged against charged, polar against apolar, and so on — and puts those counts into a linear model fitted to a set of complexes whose affinities were measured experimentally.
That is the whole method. It is fast and needs nothing but the structure, and the trade-offs follow directly from its simplicity:
- It does not model solvent, entropy, or flexibility explicitly.
- It assumes the structure it is given is a genuine, correctly posed complex.
- It was trained on transient protein–protein complexes, and is least reliable outside that class.
Temperature affects the Kd but not the ΔG — it only enters when converting between them.
Reading the output
As a rough guide, ΔG around −6 kcal/mol is weak, −9 is moderate, and below −12 is tight. Kd follows: micromolar is weak, nanomolar is tight.
Comparisons are more trustworthy than absolute numbers. Running the same complex with and without a mutation, or comparing two related complexes, uses the model where it is strongest. Quoting a single predicted Kd as though it were measured does not.
The model is known to compress the extremes: very tight complexes come back looking less tight than they are. A famously tight interaction may be predicted in the nanomolar range when the measured value is orders of magnitude smaller.
Important limitations
It answers "how tightly", never "whether". Give it any two chains and it returns a number. If the two proteins do not actually interact, that number is meaningless — it is not evidence of an interaction.
It needs a real pose. A docked or modelled complex gives a prediction about the model, not about the molecules.
Both selections must genuinely touch. If the chains you name have no interface, the run fails rather than returning zero.