Getting nonsense predictions

[georgkempf]

First of all thanks a lot for releasing the code and model! I ran a few simple tests for small proteins, but the output is always only scattered atoms (see below).

I cannot use containerization software and installed AF in a virtual environment based on the dockerfile. I ran the tests on a RTX 2060S and RTX Quadro 4000.

The output for the example json from the Readme

{
  "name": "2PV7",
  "sequences": [
    {
      "protein": {
        "id": ["A", "B"],
        "sequence": "GMRESYANENQFGFKTINSDIHKIVIVGGYGKLGGLFARYLRASGYPISILDREDWAVAESILANADVVIVSVPINLTLETIERLKPYLTENMLLADLTSVKREPLAKMLEVHTGAVLGLHPMFGADIASMAKQVVVRCDGRFPERYEWLLEQIQIWGAKIYQTNATEHDHNMTYIQALRHFSTFANGLHLSKQPINLANLLALSSPIYRLELAMIGRLFAQDAELYADIIMDKSENLAVIETLKQTYDEALTFFENNDRQGFIDAFHKVRDWFGDYSEQFLKESRQLLQQANDLKQG"
      }
    }
  ],
  "modelSeeds": [1],
  "dialect": "alphafold3",
  "version": 1
}

looks like this:

Apart from the nonsense output I did not encounter any obvious errors in the pipeline.

[joshabramson]

We will track this and related issues in #59

[Augustin-Zidek]

Thanks for reporting!

I will close this issue so that all information is centralized and tracked in #59