GOLD Docking Results of Non-covalent designs

We’ve docked the non-covalent ligands available as of the 27th April from the CDD Vault using GOLD. GOLD was configured to create an ensemble of 6 proteins taken from the Fragment Screening results selected by diversity based on overall protein backbone RMSD. The top ranked pose by fitness was kept from each docking.

This led to ~3000 docked molecules.

We then re-ranked the molecules based on certain binding properties, effectively using a binary decision tree. In summary:

  1. Structures with a fitness > 60.0 were ranked before structures with a fitness < 60.0

  2. In these two sets, the structures were further ranked. Structures forming an H-bond to GLU166 and HIS163 were ranked highest, followed by structures that formed at least one H-Bond to one of these residues.

  3. The resultant subsets were then further ranked based on whether the structures formed an H-bond to HIS41

  4. The larger array of subsets was then ranked based on buriedness calculated using GOLDMine - structures more than 90% buried were ranked ahead of structures that were less than 90% buried

  5. Finally, ties on the basis of the above were ordered by fitness.

The files are available at the dropbox link below:

The score based on the above has the tag . By way of example, a score might be


This seems big, but its actually an indexing: the first 2 signifies “Scores more than 60.0” the second 2 is a sum of two 1s - the first 1 indicates ‘doesnt form an H-bond to GLU166’, the second indicates ‘doesnt form an H-bond to HIS163’.

The next 1 indicates that this doesn’t form an H-Bond to HIS41. The final ‘2’ indicates this is buried

Finally the fitness score is added. So we can read this as a highish scoring docking by fitness, that is buried but doesn’t seem to form certain key interactions in the binding site in the pose the program has produced.

The original GOLDScore is also included in the SD Files with the tag so sorting the results on this alone is possible if the user would rather interpret the data on this basis alone.

I note that none of the poses seem to form an H-Bond to all 3 of GLU166, HIS163 and HIS41. This is unlike some known inhibitors (for example 6Y2G) in the active pocket.

1 Like

I mapped some of the compounds that had activity data (CSV download with matching compounds and plotted that against the Gold-Fitness-Score. Many of the compounds have a high Gold-Fitness-Score, ranging from 30 to 86, max is at 57. Some outliers with negative score exist.

I am not sure why I could only map 177 of the compounds, because activity data for the Average Inhibition at 50 uM is available from 277 unique compounds (as of May 20 2020). Matching was done via InchiKey and then both tables were cross-referenced via EXCEL vlookup.

Its just an initial plot (have to double check for matching errors). But there is no correlation between the Gold docking score and inhibition. The refined score which has all the scores binned into 6 different categories shows similar trends. Which might be expected, because according to the manual:

“The (Gold Score) fitness function has been optimised for the prediction of ligand binding positions rather than the prediction of binding affinities, although some correlation with the latter has been

“ChemScore was derived empirically from a set of 82 protein-ligand complexes for which
measured binding affinities were available. Unlike GoldScore, the ChemScore function
was trained by regression against measured affinity data, although there is no clear
indication that it is superior to GoldScore in predicting affinities.”


I am wondering if its possible to use the existing inhibition data, potentially also PDB data from all the new Mpro complexes and to develop an improved scoring function. Maybe only specific to Mpro docking experiments. I did not test any decoys yet, but a scoring function that gives at least a cut-off, even without showing clear correlations would be of course useful when screening combinatorial libraries.

One other reason comes to my mind, that the 3000 candidates are actually all good candidates. So out of the 6000 compounds in the postera/covid-19 repository they represent “good” candidates. But it is also known that small functional group changes (such as the tBOC groups from alpha-ketoamide inhibitors) can influence activity and non-activity during the inhibition assay. So basically the scoring function during the docking process is not sensitive enough to detect such minor changes and maybe additional protomers, tautomers and molecular dynamics steps have to performed to obtain better correlations.


Dear Tobias,

Thanks for looking at this. I suspect the reason you only have 177 candidates is that the data was generated using the CDD vault as of the 27th April and we have only included the non-covalent designs. The negative outliers are likely examples where the automated workflow failed to generate a good starting geometry for the ligand.

Its not at all surprising to me that the fitness score alone is not discriminatory, although I have to say I thought it would do a little better than that. My ‘rescore’ using features was far too optimistic on the scores discriminatory power. Given the above data, that could be adapted.

I think that’s why we tend to use the algorithms to generate poses and then search for critical features in the poses that are necessary for binding. I think creating a custom scoring function would certainly be a very useful addition (That’s why I’ve left all the terms in the SD files as this gives a start point.).I used GoldScore in these docking as historically it has tended to perform best with proteases, but of course one could also try the other scores though these too would suffer from issues due to minor changes of course.

Best Wishes


Hi Jason,
thanks.I think its important to have a starting baseline. What about including additional data from individual contact points from the docking results, but use them as individual input data for deriving models using gradient boosting trees (or other ML methods). Or similar to this CCDC approach by creating a hybrid docking/QSAR approach, in this case by combining the GoldScore and pKa within GoldMine. Or even additional 3D or 4D molecular descriptors? (Deriving a receptor based QSAR model using docking data; ) Either way the target could be a docking score, or the target could be an activity in a classical QSAR assay. Both can be combined in different direction to improve accuracy of either the docking score, or the QSAR outcome.


I cannot find their IC50 on moon shot ! i need the activity and active non-covalently componds

The link above I believe is the activity data that Tobias pulled together?

1 Like