google-deepmind / alphafold

Open source code for AlphaFold.
Apache License 2.0
12.64k stars 2.24k forks source link

Colab simplified version seems to have a more precise prediction? #126

Closed huanxian1995 closed 2 years ago

huanxian1995 commented 3 years ago

I‘ve predict structure via Colab version as well as full version. But compared with the full version, the simplified version is closer to the actual experimental results. Especially the stability and the expression of the protein. Is there any reason?

huanxian1995 commented 3 years ago

It depends on your setting. For example of last things https://mobile.twitter.com/sokrypton/status/1426235782747021313

But it seems that there are limited parameters in the full version of alphafold2? What can I change so that I can get results similar to the Colab simplified version?

Augustin-Zidek commented 2 years ago

You are probably lucky with this structure, but when we compared over a large dataset the full model came out as better.

To run the standalone AlphaFold in a mode similar to the Colab, you need to:

  1. Run with reduced databases (--db_preset=reduced_dbs) and disable HHBlits on UniClust (comment it out in the data pipeline).
  2. Disable templates (e.g. by setting max_template_date=1000-01-01).