Alpha-fold won the CASP13 and CASP14 competitions last year and this year. It used deep learning to predict the secondary structure of a protein given the primary amino acid sequence.

Google has released its code a few weeks ago, and I was wondering how to use Alpha-fold to predict the complex of two proteins.

Naively, I'd concatenate the amino acid sequences of both proteins as if they were a single protein. Then I’d input the neural network with the concatenated primary sequence.

The network will predict the lowest energy conformation of the concatenated proteins. I am afraid that the artificial backbone connection between the two proteins might lead to a completely wrong complex.

I could do some docking of the two individual proteins, but I want to find the denovo prediction of the two-protein complex.

Edit: here is an Pytroch implementation of Alpha-fold2 on github.

  • 3
    $\begingroup$ Of course you have to do the docking! Two proteins is not one. $\endgroup$
    – Mithoron
    Commented Mar 9, 2020 at 16:11
  • 1
    $\begingroup$ And for docking you need to use a program made for docking not for single protein, just because it's good. $\endgroup$
    – Mithoron
    Commented Mar 9, 2020 at 16:14
  • 4
    $\begingroup$ @JonCuster Do a search on ChemSE for Amber, Charmm, Gromacs, or ChemDraw. You'll see lots of questions and answers. In addition, his question, and my answer, has general implication for how you can and can't use biopolymer folding prediction software (wherther the biopolymer is a protein or some other biopolymer, e.g., RNA). $\endgroup$
    – theorist
    Commented Mar 9, 2020 at 17:47

2 Answers 2


Are you asking how to take code developed to find the MFE (minimum free energy) structure of a single protein, and modify it to instead determine the structure of a two-protein complex?

Or are you asking if you can use this code, as is, to fold a concatenated combination of the two proteins, and in turn use this result as a surrogate for the predicted structure of the two-protein complex?

If you're asking the former, I would recommend contacting one of the code developers to ask them for guidance. It's possible that they have already developed a prototype option to address this.

If you're asking the latter, that approach is problematic, for two reasons:

  1. Folding the concatenated protein would constrain the last amino acid in the first protein to be next to the first amino acid in the second protein. That's not a constraint that would make sense a priori for a two-protein complex.

  2. You have to ask yourself if, in forming the two-protein complex, you envision both proteins unfolding and then refolding to do a global energy minimization— since that's what folding the concatenated protein would give you (with the additional constraint mentioned in point #1, above). If, on the other hand, you would like the proteins to each fold into their respective MFE structures, and then (while essentially retaining those structures) find the MFE configuration for the two-protein complex, then folding the concatenated protein clearly would not give you that.

  • $\begingroup$ Does Alpha-fold actually compute free energies, or is its performance based on training on known structures that are not necessarily minimum energy, particularly if experimental i.e. folded in vivo? $\endgroup$
    – Buck Thorn
    Commented Oct 3, 2020 at 9:44

From the abstract of the 2019 Nature article outlining the success of AlphaFold:

Here we show that we can train a neural network to make accurate predictions of the distances between pairs of residues, which convey more information about the structure than contact predictions. Using this information, we construct a potential of mean force4 that can accurately describe the shape of a protein. We find that the resulting potential can be optimized by a simple gradient descent algorithm to generate structures without complex sampling procedures

One implication of what is stated above is that AlphaFold provides an algorithm that maps sequence to experimental folds with high accuracy but does not attempt to provide a physically accurate multi-purpose interaction potential. It is a statistically generated potential rather than based on fine tuning terms with a physical interpretation (e.g. angular, torsional, steric, Coulombic). This results among other things in potential predictive limitations imposed by the training sequence domain. Like any model it can guarantee success only if relevant training data (or potential terms) are included in creating the model. The sequence-fold relations relevant to quaternary structure might not overlap well with those relevant to lower order folding. I would argue that a subset of the folding interactions (in particular contacts between distant residues not involved in secondary structure) are more relevant to quaternary structure. It is also not clear that the potential would be useful to perform MD simulations to probe non-equilibrium structures or sample local minima.

Despite that, if those potential roadblocks are not significant I do not see why your approach should not work, particularly if you pick an appropriate linker (which is easily tested).

This answer risks might seem opinion-based but is an interpretation based on what can be quickly inferred from information in the linked publication.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.