Help Needed to Reproduce PoseBusters Bioassemblies Correctly

[schrojunzhang]

Dear AI4Science team,

Thank you very much for your excellent work and for sharing the PoseBusters dataset!

I'm currently trying to reproduce the bioassembly files for the PoseBusters dataset using the /scripts/prepare_training_data.py script together with the provided posebusters_mmcif files. However, I’ve encountered some discrepancies:

• The generated CSV file does not match the one you provided; in particular, the num_prot_chains values appear to be consistently higher.
• The generated bioassembly set only includes 302 of the 308 unique molecules from the original test set—six structures could not be reconstructed successfully.

Would it be possible for you to share more details about the process or specific parameters you used to prepare the dataset? Are there any modifications to the script or input files that I should be aware of to reproduce the results accurately?

Thank you in advance for your help!

Best regards,
Xujun

[cloverzizi]

@schrojunzhang
PoseBusters is a test dataset, and its processing differs from that of the training set. Unlike the training set, it hasn't undergone the Filter procedure, and you must select the altloc ID for specified ligands in PoseBusters. When you run the script for processing training data, the num_prot_chains is usually higher than what we provide, because PoseBusters is processed based on the asymmetric unit, rather than Assembly 1 used for the training set.

When running scripts/prepare_training_data.py, you can add the -d parameter to skip the Filter and use the Asymmetric Unit structure. This approach will yield results closer to our bioassembly_dict. To fully replicate our process, you must also provide altloc IDs for certain cases in the get_structure() method and crop the structure for 8F4J, as it contains many chains.

If you're simply curious about the differences in processed structures, the above explanation should suffice. For any other requirements, feel free to leave a message, and we’ll see how we can assist.

[schrojunzhang]

@cloverzizi,

Thank you for your kind and helpful reply!

In addition to working with PoseBusters, I’ve also tried reproducing the bioassembly files for the training set using the prepare_training_data.py script you provided, without any modifications. I’m happy to report that most of the entries in the generated CSV match your released dataset very well.

However, I did notice that a small number of entries still differ slightly. I was wondering if this might be due to differences in the environment—such as the version of biotite—or if there might be additional processing steps not fully reflected in the script?

I’d really appreciate any further insights you could share. Thank you again for your time and support!

Best regards,
Xujun

@schrojunzhang PoseBusters is a test dataset, and its processing differs from that of the training set. Unlike the training set, it hasn't undergone the Filter procedure, and you must select the altloc ID for specified ligands in PoseBusters. When you run the script for processing training data, the num_prot_chains is usually higher than what we provide, because PoseBusters is processed based on the asymmetric unit, rather than Assembly 1 used for the training set.

When running scripts/prepare_training_data.py, you can add the -d parameter to skip the Filter and use the Asymmetric Unit structure. This approach will yield results closer to our bioassembly_dict. To fully replicate our process, you must also provide altloc IDs for certain cases in the get_structure() method and crop the structure for 8F4J, as it contains many chains.

If you're simply curious about the differences in processed structures, the above explanation should suffice. For any other requirements, feel free to leave a message, and we’ll see how we can assist.