Skip to main content
A single crystal structure captures one conformation of a protein, but ligands bind to a moving target. Ensemble docking docks each ligand into several snapshots taken from a molecular dynamics simulation of the target and combines the scores. This page asks whether those scores rank a target’s ligands in the order that experiment does, across 18 targets and 451 ligands.
All results are ours. For each target, every ligand was docked into snapshots from a 50 or 100 ns simulation (five snapshots for a 50 ns run, ten for a 100 ns run) and its scores were averaged over the snapshots. Measured affinities are logKa values for the ligands’ PDB complexes.

​
What the comparison shows

Averaged over snapshots, the learned CNN affinity ranks ligands better than the plain docking energy: its mean rank correlation across targets is 0.29 against 0.19, it points the right way on 15 targets against 11, and it is ahead on 12 of the 18. Performance varies a lot by target. It is strong on PDE10A (ρ = 0.88), PTP1B (0.73), BRD4 (0.59, on 52 ligands) and PKA (0.52), moderate on about half the panel, and poor on AR, CHK1, EphA2, Factor Xa and GluA2, where it is near zero or negative. A mean of 0.29 is a modest result, and it should be read that way.
The pooled summary in our internal panel table reports a mean ρ of 0.45, because it takes the best of several scoring metrics for each target after looking at the results. That inflates every target, so this page fixes one metric in advance for all of them (the CNN affinity averaged over snapshots) and shows what it actually delivers.

​
By target

Spearman ρ and Pearson r between each ligand’s score and its measured logKa, for the 18 targets, grouped by simulation length.

​
Target by target, ranked

​
Where learned scoring helps

The difference in rank correlation for each target, CNN affinity minus docking energy.

​
Score against measured affinity

Pick a target and a score. Each point is one ligand.

​
How much the choice of metric matters

The last row is not a valid estimate of performance: it uses the answer to choose the method. It is shown only to size the effect.

​
Reading these results

Targets have between 6 and 75 ligands, and most have 15 to 30. A rank correlation over that few points moves a lot with one ligand, and JAK2 (6 ligands) is too small to read. Weight the pattern across targets over any single value.
The 50 and 100 ns runs were done on different targets, so the table cannot say whether the longer simulation helps. It only reports each target at the length it was run.
The scores are compared with measured affinity by rank and correlation only. They are not calibrated to kcal/mol, and the page makes no claim about absolute accuracy.
This page compares two scores from the same docking runs. It does not compare ensemble docking with docking into a single structure or with other programs.
The CNN scoring model was trained on public protein–ligand data, and these ligands come from the same public archives. We have not checked for overlap, so treat the CNN result as agreement with public data, not a strict test on unseen ligands.

​
Sources

  • Experimental affinities and ligands: PDB complexes with measured logKa for each of the 18 targets.
  • Scores: our ensemble docking runs, with molecular dynamics snapshots at 50 and 100 ns, scored with the CNN affinity and docking energy of the docking engine.