filter to generate 100 molecules for each pocket in test set

[Kartinaa]

Dear author,

I try to use the command mentioned in the readme python scripts/sample.py --outdir ./outputs --config ./configs/sample/sample_MolDiff.yml --batch_size 32 --pocket_dir ./data/crossdocked --num_pharma_atoms 20 --clash_rate 0.1 to generate 100 molecules for each pocket in test set. For some receptors I could get 100 valid molecules. However, for some receptors, I only get ~ 10 valid molecule and the error message is Too many failed molecules. Stop sampling.. And some of receptors there are even no folders generated.

I plan to get 10000 molecules (100*100), but at the end, I get a total of 6314 molecules (from SMILES.txt file, if I'm correct). I also attached log file and the picture for my folder here, please check them.

Really appreciate it if you could have any suggestion on this.
log.txt

Best,
Bo

[Kartinaa]

I tried again to generate 1000 molecules in total (10 per pocket) from testset. But I'm still keep failing, seems like there are several pocket for which the algorithm is not working. I increased allowed number of failed molecules, but it couldn't give any valid molecule.

[YaelZiv]

Hi Bo,

Could you please provide more details about the targets for which the model fails to generate molecules? Do these targets have valid reference points? Are they part of the CrossDocked dataset? Could you share their ID and name?

Best,
Yael

[Kartinaa]

Hi Yael,

Thanks for reply. I just used the command you provide for sampling molecules for all pockets in the test set: python scripts/sample.py --outdir ./outputs --config ./configs/sample/sample_MolDiff.yml --batch_size 16 --pocket_dir ./data/crossdocked --num_pharma_atoms 20 --clash_rate 0.1

I do think they have valid reference point cause I downloaded the processed data from the link you provide and yes, they are part of the CrossDocked dataset. I'm not sure how to get ID and name of the target which I failed to generate valid molecules, so I'll just attach the log file from a fail run here, it did successfully generate around 830 compounds, but not 1000 which I required (10 for each).
log.txt

Please let me know if you need more information.

Best,
Bo

[YaelZiv]

Hi Bo,

Thanks again for the update! I've made some changes to the utils/reconstruct.py file. Please download the updated version from the repository and try running it again. Let me know if it works as expected now, or if you encounter any further issues.