Some issues with multimer inference

[Ericzhang1]

Hi, I would like to report some issues I encountered when trying to run multimer inference:

1. in feature_pipeline.py line 81: 'num_res = int(np_example["seq_length"][0])' was giving me error and changing it to 'num_res = int(np_example["seq_length"])' fixed the problem
2. the 'make_backbone_affine' function from all_atom_multimer seems not working. 'rigid' object returned seems to have really large values. This function is called in TemplateEmbedderMultimer and is causing 'nan' errors.
3. the extra msa stack seems to cause segmentation fault which might be due to memory issues.
4. With both template and extra msa disabled, everything works fine but the model is not predicting any structures (random).

I did modify the code a bit, sicne I was trying to skip the input pipeline and directly loading the pickle file (features) from af. The example I am running is:

2RBG_1|Chains A, B|Putative uncharacterized protein ST0493|Sulfolobus tokodaii (111955) MPYKNILTLISVNNDNFENYFRKIFLDVRSSGSKKTTINVFTEIQYQELVTLIREALLENIDIGYELFLWKKNEVDIFLKNLEKSEVDGLLVYCDDENKVFMSKIVDNLPTAIKRNLIKDFCRKLS 2RBG_1|Chains A, B|Putative uncharacterized protein ST0493|Sulfolobus tokodaii (111955) MPYKNILTLISVNNDNFENYFRKIFLDVRSSGSKKTTINVFTEIQYQELVTLIREALLENIDIGYELFLWKKNEVDIFLKNLEKSEVDGLLVYCDDENKVFMSKIVDNLPTAIKRNLIKDFCRKLS

I understand that this is still under developement. Just want to help with some testing thx.

[moritzschaefer]

@Ericzhang1 did you experience the error reported in #110?

[Ericzhang1]

@Ericzhang1 did you experience the error reported in #110?

No, I did not encounter the weight importing issue. Not sure why it is happending. Did you make sure it is multimer's weight that is imported not monomer? Also multimer need to be specified in the command.

[gahdritz]

Sorry guys --- I just want to acknowledge that I've seen this and have it on my to-do list. I'm a little swamped with CASP stuff at the moment.

The multimer branch is a little rough around the edges right now. As a stopgap, I've added AlphaFold-Gap (a hack that permits prediction of complexes using the standard weights) to the main branch.