Question about excessive featurising time caused by large MSA inputs and template option

[kimlab-cnu]

Hello! I am a student in Korea studying protein design, and I would like to ask about an MSA-related issue we encountered during prediction of a 3-chain complex.

We are currently predicting a tertiary complex composed of:
two designed proteins, each approximately 400 aa in length + one cyclic peptide represented as SMILES (roughly 10–20 residues in size)

To apply the template option supported by AlphaFold 3, we downloaded experimentally validated mmCIF files from the RCSB PDB, prepared template structure files for each protein chain, and added the template information to a JSON file that already contained MSA results.

However, after enabling templates, the runtime increased substantially.
Previously, one complex took about 1 hour, but with templates applied, the same job required about 6 hours.

From the log, we found the following:

Featurising data with 1 seed(s)...
Featurising data with seed 220.
Featurising data with seed 220 took 21036.24 seconds.
Featurising data with 1 seed(s) took 21046.55 seconds.

The actual model inference takes only about 1 minute, while the featurising step takes nearly 5 hours and 50 minutes.

1. I would also like to ask whether the template option can have a substantial impact on the overall runtime of AlphaFold 3, including structure inference.

We then checked the MSA sizes for each chain:

Chain A
unpairedMsa length: 5,035,677
pairedMsa length: 15,933,678

Chain C
unpairedMsa length: 4,819,630
pairedMsa length: 15,324,761

These were generated using the default jackhmmer search settings.
Since Chain B is a ligand represented by SMILES, it does not contribute much meaningful paired MSA information.

Based on this, we are currently considering the following possibilities:

A. Reducing the size of both unpairedMSA and pairedMSA
B. Since Chain B has little meaningful pairedMSA information, we suspect that the pairedMSA may fail to preserve useful inter-chain linkage and instead become overly biased toward Chains A and C. Therefore, we are also considering reducing or even removing pairedMSA

I would greatly appreciate your thoughts on whether these ideas seem reasonable.

In addition, when reducing MSA depth, I understand that there may not be a single correct answer. However, I have had difficulty finding clear guidance in the paper or related resources regarding:

2. How to decide how much to reduce the MSA?
My initial idea was to truncate the current MSA to approximately 70%, 50%, 30%, and 20% of its original size, then compare the resulting structures using metrics such as iPTM and PAE.

Would this seem like a reasonable benchmarking strategy? Any advice or practical recommendations would be greatly appreciated.

Thank you very much for your time and help.

[Augustin-Zidek]

What version of AlphaFold 3 are you running? This was fixed in 751a4b8, if you sync past that, AlphaFold 3 should crop the MSA to a maximum depth of O(10k) sequences (which is more than enough).

A. Reducing the size of both unpairedMSA and pairedMSA

Yes, the model sees the top 16k MSA sequences anyway, so anything beyond that is lost in the featurisation later anyway.

B. Since Chain B has little meaningful pairedMSA information, we suspect that the pairedMSA may fail to preserve useful inter-chain linkage and instead become overly biased toward Chains A and C. Therefore, we are also considering reducing or even removing pairedMSA

Yes, I don't think it will hurt in this case, given how many MSA hits you have anyway. Worth experimenting with though and checking AF3 prediction accuracy doesn't degrade.

2. How to decide how much to reduce the MSA?

Just apply 16k as the total limit.

[kimlab-cnu]

Hello,

Thank you for your helpful comments. I reviewed your suggestions yesterday.

It appears that our AlphaFold 3 installation was based on an older version, so we backed up the original Docker container and built a new image based on commit 751a4b8. We are currently using this updated version for the MSA step.

As you suggested, we will also test lowering the paired MSA ratio.

Our current strategy is to use experimentally validated PDB structures as templates in order to improve the structural accuracy of AlphaFold 3 predictions. I would like to ask whether this is a commonly used or recommended approach.

During testing, we found that AlphaFold 3 raised an error when templates were added before running the MSA step. In this case, would it be more appropriate to run the MSA step with all related fields left empty, and then add templates only later during the structure inference step if necessary?

We have been separating the MSA and structure inference stages using the --run_data_pipeline and --run_inference options, mainly to make GPU usage on our server more efficient.

Thank you for your support. Your advice has been very helpful.