AlphaFold: Usage for single sequence? about deepmind-research

[pentadotddot]

Hi, here I have found a reimplementation of alphafold in pytorch: https://github.com/Urinx/alphafold_pytorch

Here the author provided almost all the rest of the pipeline including feature generation.

Still I could not figure out how it works, but seems one can figure out something from it:)

[Kohze]

I think we should give it a try even though google is not releasing all the code yet. Many research projects would profit from it.

[huhlim]

DeepMind team only shared the AlphaFold deep learning model instead of the full pipeline for model building, so there are many features are missing including input feature generation and 3D structure generation.

If you want to build a model for your protein, you still can try typical structure prediction servers such as I-TASSER, RaptorX, and so on. They are also pretty good. In addition, some of the servers (i.e., RaptorX-Contact) have their own contact/distance prediction methods and use them to build 3D models.

If your protein sequence has close homologs with known structures, then it would be better to use template-based modeling. CASP targets usually deal with hard cases, which do not have known homolog structures or have very distant homolog structures. So, for a typical structure prediction circumstance, I hope template-based modeling would be enough to get an accurate 3D model. (I do believe that classical approaches are still better on these relatively easy cases.)

[Augustin-Zidek]

As discussed in issue #18 the released code does not include Multiple Sequence Alignment (MSA) and feature computation for arbitrary sequences, though there we give some guide as to the features used for those accustomed to computing them.

But note that in any case the AlphaFold system does not use homologous structures (templates) it is focused on "Free Modelling".