Why can't AlphaFold predict the consequences of point-mutations?

Question

In the literature, it specifically states that AlphaFold has "Has not been trained to predict structural consequences of point mutations". See : https://alphafold.com/faq

AlphaFold has not been validated for predicting the effect of mutations. In particular, AlphaFold is not expected to produce an unfolded protein structure given a sequence containing a destabilising point mutation

What I want to understand, is why?

If I change one amino acid, and it's not destabilising, why can't AlphaFold use its training and form a moderately confident structure? I mean, to a computer, what's the difference between a wild-type and a mutated protein? I know what you may be thinking, why don't you just try it yourself, and see if AlphaFold works. But if I don't have the experimental structure of the mutant, how on earth am I supposed to believe that this is even close to the theoretical real thing?

Maybe I'm not understanding the main premise, but it's something that's been bugging me for a while now. Any insight into this would be much appreciated.

Answer 1

Other have addressed why AlphaFold2 cannot really get the necessary signal from MSA (Multiple Sequence Alignment) and memory to properly model a variant, but I thought I'd add my observations/rants.

I have modelled in silico both engineering and clinical variants in the past, using PyRosetta, MD and other tools. For me, AlphaFold2 has really stood out but in a different way than the hype.

Variants can have complicated effects

I developed a tool Venus, https://venus.sgc.ox.ac.uk/, which collates different pieces of information to help figure out the effect a variant was having beyond a simple ∆∆G calculation. This is because often a variant is having a variety of effects, for example a stable but catalytically dead variant can still form a complex and sequester a critical component (e.g. in ref POLR2E-sequestering POR2A is more pathogenic than a simple truncation). This is not only true with human clinical variants, but also with engineering variants: for one protein I have worked on, DogCatcher, lowering the isoelectric point away from neutral pH was critical, not the stabilisation of the transition state (ref). Another example is that the entrance loops in cytochromes drive selectivity more so than the active site configuration.

The unknowns: alt conformers and complexes

I will be the first to admit that most folk want a deleterious ∆∆G value and will p-value shop for it and not catering for this is a bad idea for a publication. But in terms of mechanism, the biggest issue I face are dynamic processes involving a residue where the other active/inactive conformer is unknown or when the residue is a surface residue possible at the interface in a complex, which is known simply to exist from a coIP western. No nearby gnomAD variants or presence of post-translational sites or sequence conservation do show that something may be important but not why.

Enter AlphaFold2

AlphaFold2, via the more user-friendly ColabFold, allow the creation of wild type monomers, homo-oligomers, but also hetero-oligomers (complexes) and alternative conformations. These days I seen a cartoon of an interaction and I model it in quintuplicate. Sometimes nothing binds, but sometimes a complex is found 5 out of 5. The weirdest part is that it reveals some novel mechanism that makes sense that has probably never been seen before, but is beyond the scope of the experiment so one has the power to just put them aside —interfaces which are stronger phosphorylated, inhibitory domains binding in the same site but in a different way, unexpected complex arrangements etc. So the true revolutionary power of AlphaFold2 is predicting complexes and alternative conformations, which can be used to predict the mechanism of a variant.

WIP

Complexes do not always resolve and there is a lot of discussion in terms of massaging the MSA —e.g. boney-fish+tetrapods only etc. And there is very little cross-linking mass-spec data available for validation and/or constraining. But hopefully things will get better.

Footnote

The last paragraph sounds like an exaggeration, so I best give an example. Here is an example for the drawer from today. A target protein (catalytic domain = pale) has a regulatory terminal which inhibits in isolation (terminus = turquoise), however, when modelled with another protein (mustard) this terminal (terminus = coral) is flipped away. Thus revealing the mechanism by which the mustard protein sequesters the regulatory helix (coral), which would otherwise (turquoise) inhibit the catalytic domain (pale cyan). In the literature, this is a western blot. Where a variant of interest involved in this mechanism one state may be more energetically favourable by scoring the ∆G of the different "snapshots" (=static conformers).

Answer 2

As marcin points out in the comments: by "AphaFold (AF) has not been trained to predict structural consequences of point mutations" it is meant "AF is not able to tell you whether your point mutation is destabilizing or not, and if it is destabilizing then the structure it predicts will most probably be wrong".

It is totally possible that AF predicts the structure of a protein well when only non destabilizing mutations are introduced. As pointed out in the paragraph marcin quotes, however: "AlphaFold has not been validated for predicting the effect of mutations." And as long as we don't know, we should be warry of any of AlphaFold predictions with point mutations. Maybe it works, maybe it doesn't, or maybe it's case dependent.

Answer 3

This answer is from common knowledge rather than specific infield expertise. The question could be more expertly answered by other members.

If you really want to pursue this type of calculation, it's possible that something like INPS-MD de novo -> protein structure stability from variants will provide an answer. So maybe compare the outputs with AlphaFold?

Intuitively, I don't think it's cool, just my opinion. In my opinion, an experimentally derived structure is required as the basis for your calculation, where you want to consider the impact of an amino acid mutation.

It's important to separate de novo prediction from variant prediction. Alphafold is de novo prediction. What you are looking at is structure stability from a variant (see Summary). That's my core answer, these different calculations. To explain this more clearly ...

Good variant stability calculators are DUET (2014) and INPS-MD (2016), amongst others. Just to reiterate, a good starting point is an experimentally derived protein structure.

The webservers are here:

• http://structure.bioc.cam.ac.uk/duet
• https://inpsmd.biocomp.unibo.it/inpsSuite

I personally don't know Alphafold but understand its predecessors which accommodate convolutions by leveraging deep-learning. The basis of the predecessor's calculations firstly are de novo prediction and secondly can miss the impact of variants if they are not directly involved in the core-algorithm. It's about identifying which amino acids are interacting considering all available homologous amino acids within a database. The calculation can predict local interactions with extreme accuracy, but it's a prediction. It will also depend on whether there are single amino acid changes or two- or more amino acid changes (they could compensate each other, particularly for a natural variant).

In contrast, DUET and INPS-MD, amongst a lot of other similar programs, consider the property shift of an individual amino acid in relation to its immediate local stereo-chemical environment measured by a shift of free-energy.

Essentially, these are two sets of very different calculations.

Summary The problem I intuitively see is if de novo prediction in any form is used, okay they are very good, but they're just predictions. Variant structural predictions are also just predictions which are very widely respected. Stacking de novo prediction onto predicting protein structure stability resulting from variants has scope for error, especially if there was some hidden non-linear effect.

There could be additional issues because the de novo calculations immediately prior to Alphafold don't predict the precise local distances.

---

AlphaFold would simply respond

'it's not been trained'

They'd want a wet-lab experimental dataset with wild-type vs. variants, if these exist with sufficient abundance to prevent over-training. It would be a very skewed dataset so of limited application in my personal opinion.