Low confidence score from alphafold2_multimer_v3

Question

I am running alphafold2_multimer_v3 in ColabPro (A100 40 GB GPU) with the following parameters:

template_mode = "pdb100"
msa_mode = "mmseqs2_uniref_env"
pair_mode = "unpaired_paired"


num_recycles = "24"
num_seeds = 5
num_models = 2
recycle_early_stop_tolerance = "0.1"
use_dropout = True

max_msa = "512:1024"
pairing_strategy = "complete"

I am basically running a modified version of the colabfold notebook. My complexes are from yeast, mostly composed of 4-6 proteins with amino acid chain length of around 1200-2000. Some of the complexes holds 1 or 2 Un-charecterized proteins. The ipTM and pTM score are very low, around 0.3. Though the plDTT is around 70 for most.

Then with my configurations I tried to regenerate one results from this paper: https://www.nature.com/articles/s41586-023-06739-5

They reported a confidence score [ipTM+pTM] of 0.8 for "TCD1:TCD2:Ygr012W" complex. I got similar around 0.75 confidence score with my parameters.

Then I tried checking the alphafold 3 server, it gives slightly improved results around 0.45. But it's still in beta, so can not do an investigation there by tweaking in the parameters.

Is there any way to improve confidence score (iptm+ptm) for my complex set? Am I doing it wrong?

Answer 1

There are reports that ColabFold is slightly less accurate than native AlphaFold. So, you might be able to get a small increase in confidence when using AlphaFold. Or at least using the full BFD database that AlphaFold uses.

Otherwise, there are many papers reporting significant increases in confidence when using more appropriate MSA databases. Here is a good example. However, I suspect yeasts are well covered in the existing databases, so you might not get much improvement in your complex with current versions of AlphaFold.