Should PDB70 version be incremented in Alphafold version 2.3.1?

[dougrenfrew]

The README section on CASP14 reproducibility with version 2.3.1 implies that one should use old versions of the databases (and gives versions), or using current PDB70 and PDB but passing the --max_template_date command line flag to reproduce the CASP14 results. The dates for the (implied older) versions of the databases for PDB and PDB70 are 2020-05-13.

The version of the PDB70 downloaded by the scripts is pdb70_from_mmcif_200401.tar.gz (2020-04-01). This has been the version downloaded by the script since the initial commit of the repo.

Is this a bug? Should a more recent version be used and set as the default in the download scripts?

Dates for readily available versions of the PDB70 are as follows

• 2020-04-01 (version currently hard coded in the download script)
• 2020-05-13 (version recommended in README for reproducing CASP14 results)
• 2020-09-16
• 2021-10-27
• 2021-11-17
• 2022-03-13

[adalal78]

I just came across this recently. I can't believe this hasn't been answered yet? @dougrenfrew, did you find the answer? Did you try using more updated versions (pdb70_from_mmcif_200916.tar.gz, or even the 100: pdb100_foldseek_230517.tar.gz) on AF2? If so, no issues so far with it? TIA

[dougrenfrew]

I have not received any answers. Using newer versions of the libraries did not break anything. I suspet additional templates would improve results, but I have not benchmarked it.

[adalal78]

Thank you for letting me know. Would u mind if I asked you about what platform you're using to do AF2 predictions? Currently, we considered some options such as on-prem, on GCP (VM and not VertexAI). We find that on-prem with even just 1 year old rtx does a good job in terms of performance. If you can, let me know and TIA. Trying to get a A100 on GCP has been a bit of a challenge, but not impossible...