How to predict homo-oligomer's and hetero-oligomer's structure?

[RodenLuo]

Hi, I'm predicting the structure for gp34, and gp37 of bacteriophage T4. Their structure are both partially solved, in 5NXH and 2XGF respectively. Both solved regions show as homotrimers and form a straight line. The straight-line feature is also shown in the first figure below. However, the AlphaFold predicted only the structure for a monomer. Some domains seem to be close to the solved structure. I attached the predicted results in screenshots below and also the corresponding ranked_0.pdbs.zip.

I can imagine that I can connect the monomer ends with some loop sequences (linkers) and feed into AlphaFold an extremely long sequence that has the original sequence 3 times plus some flexible linkers, in the hope that AlphaFold can capture the homotrimer property. But is there a cleaner or systematic way of achieving this? Similarly, the same question applies to the cases of hetero-oligomers.

Many thanks,
Roden

Credit: Hyman, Paul, and Mark van Raaij. "Bacteriophage T4 long tail fiber domains." Biophysical reviews 10.2 (2018): 463-471.

[chrisroat]

@ValZapod Is anyone involved in those notebooks willing to create a personal branch of AlphaFold with the "hacks" for homo/hetero-oligimers. I think a new command-line flag that enables that functionality in run_alphafold.py would be wonderful, and I think many people would benefit.

[NikolaTeslein]

@ValZapod Is anyone involved in those notebooks willing to create a personal branch of AlphaFold with the "hacks" for homo/hetero-oligimers. I think a new command-line flag that enables that functionality in run_alphafold.py would be wonderful, and I think many people would benefit.

If it is in that case, a dissertation can be published for the job.

[abridgland]

This repo is an implementation of the published AlphaFold system which was not used for multimeric predictions. We’re really excited by the work that’s been done in this area but since it diverges from the published system we won’t be merging those changes into our codebase.

[chrisroat]

I think AlphaFold's stance is correct, as they should only be responsible for maintaining code that they have validated and stand behind. A lot of the homo-oligomer work that I've seen are hacks/tricks that fool AlphaFold to work in a way it was not designed for or validated to do. That said, such tricks may have general use and may work.

For the people posting on this thread, please ask those people working on homo-oligomers publish their code in their own github fork. Lots of people, including myself, would be quite interested. While posting links to dissertations or other people's results is helpful, it's not necessarily the best way to be helpful. A true "open science" collaboration will only work when code is pushed to github and we can all use it and improve it.

[RodenLuo]

Thanks very much to all! Totally agree with the above two comments.

Some updates from my exploration:

1. Tracing back from @ValZapod 's post, I found this ColabFold repo. The "AlphaFold2_advanced" notebook is the solution I was looking for. It outputs a decent trimer for T4 gp36. The output is attached here, T4_gp36_rank_1_model_2_ptm_seed_0_unrelaxed.pdb.zip. But it failed at the two proteins I mentioned in this issue in the beginning. According to the readme, "ColabFold Tutorial presented at the Boston Protein Design and Modeling Club". And it is presented by Dr. Sergey Ovchinnikov who previously worked with Dr. David Baker.
2. SWISS-MODEL can model fragments decently.
3. Other than these two, I do not find other good solutions.