Predicted Structures as Templates for Multimer

[universvm]

Hi there,

I am wondering whether I can use predicted structures as templates to speed up the multimer prediction.

Essentially, assume I have already predicted each chain individually and trying to see how the complex will look like,

Thanks!

[zbekkhozhin]

Hi, I am not sure if it is still needed, my guess would be that RF2 still would need to perform MSA construction, then pass it through the Evoformer-like block that is the core of the network, figuring out which residues make a contact in 3d space, so providing the predicted structures does not speed up anything. You could imagine providing the already pre-processed MSAs and Evoformer-like module outputs for prediction, and then the model would have to fill in the inter-chain pair representations, that could work, but I am not sure if anyone is interested in figuring out whether supplying the outputs from the monomers to the multimer will result in any increase in speed/accuracy

[universvm]

but I am not sure if anyone is interested in figuring out whether supplying the outputs from the monomers to the multimer will result in any increase in speed/accuracy

Thank you for your opinion but I don't think you are correct.

It's not about the speed/accuracy (although I doubt I'm the only one interested in a speed boost). It's about having already predicted my structures say A, B, C and X, Y, Z and wanting to create a multimer of all the possible combinations of set 1 and set 2 (common in docking).

I'm sure you'll realise that having to re-predict the structure each time just increases the complexity of the calculation and it is indeed useless as we already have it calculated.

Nevertheless, given that the corresponding authors have never replied to simple emails both from academic and industry contacts, I have no interest in either investing time in the codebase or use this tool.

[zbekkhozhin]

I am not sure if we are understanding each other's statements correctly, pardon me if I am saying nonsense, I just think that when you are doing a multimer prediction, you have to compute not only the attention things between amino acids in the same chain, but also all of the possible combinations of inter-chain pairings of amino acids, and it is mostly this act of the Evoformer that actually predicts what portion of a protein binds another one, as illustrated by the RoseTTAFold2 article where they show that the architecture of the Evoformer of the AlphaFold2, the FAPE loss which is super useful, these things are the most important parts of the multimer prediction, moving atoms in 3d space seems to be less important and any reasonable 3d module works ok-ish

And regarding the replies from authors, yeah, they just can not compete with DeepMind. If you can, just use AlphaFold3 server, it has absolutely enormous context length of 5000 and runs in 5 minutes. If you can not because you want the small molecules, I would guess that buying AlphaFold3 prediction service from DeepMind for your small molecules will be very cost-efficient, although I am not sure if DeepMind will be interested in a small contract (although they were paid like ~40 million from big pharma upfront, which, considering the value of AF3, is not that expensive)