In this manuscript the authors have tested the hypothesis that the MSA constructed by AlphaFold2 (AF2) contains information about the distribution of different conformational states of a protein. Whereas the current way of thinking about AF2’s MSA-predicted Cβ–Cβ distance maps focuses on their power to provide binary classifications of inter-residue contacts, the authors propose that Cβ–Cβ distances should instead be thought of as a set of collective variables that approximate a Boltzmann distribution. This is a novel hypothesis that lends AF2 the ability to decipher the conformational Boltzmann distributions of proteins. The authors test this in the contexts of protein dynamics, mutation impacts, and protein-protein interactions. They start with analyzing the correlation between AF2 contact distance and spin label distance distributions obtained from EPR spectroscopy using T4 lysozyme as a model, finding a general agreement despite broader AF2 distributions. Following this, they explore if AF2 can approximate free energy changes in systems that contain multiple biologically important minima, using EGFR KD studies for this purpose. AF2 accurately identifies altered contact distance distributions corresponding to active or inactive conformations in several mutations, indicating a sensitivity to alterations that stabilize particular conformational states. Next, they assess sensitivity to thermodynamically destabilizing mutations. AF2 was able to predict different contact distance probabilities for disruptive mutations like L198R in UBA1, but was less sensitive for milder mutations like L198A. Lastly, AF2’s sensitivity to protein-protein interactions was explored using the μ-opioid receptor (μOR). Although the helix displacement distances observed in the predicted structure of isolated and complexed μOR do not exactly match with expected values, AF2 did successfully predict differences in select contact distance distributions of active/inactive-state μOR. Demonstrating that Cβ–Cβ distance probabilities from the same AF2-learned distribution reflect distances observed in differentially behaving domains of a protein lends strong support to the hypothesis that AF2 contact distance distributions can approximate conformational distributions.
The manuscript explores the correlations and sensitivities of AF2 predicted Cβ–Cβ distances across a variety of protein contexts, giving a broad view of its capabilities and limitations. Transitions between the various sections flowed well, and overall the writing was well worded and easily comprehensible. In addition, the presentation was balanced. It doesn’t just focus on the success of AF2, but also highlights where its sensitivities might vary or fall short, providing a balanced view of its capabilities. Given limited computational resources, the conformational space explored by MD and MCMC simulations is limited by their initial states. AI methods are instead limited by how informative their system definitions (MSAs and pre-set theoretical or experimental contact distance distributions) are, allowing AI methods, such as the AF2 method outlined by the authors, to more effectively sample conformational space. This is a very fascinating implication of their work which the authors have briefly mentioned in the discussion. This (and the connection to Figure 7 in the paper) warrants a deeper discussion, but the main conclusions the authors come to are within the scope of the manuscript, and are backed up by the evidence presented.
There are a few points we would like to bring to the attention of the authors to strengthen the manuscript further.
There are some difficulties interpreting Figure 2.
(a) It is important to mark the distances between the two chosen pairs of atoms in the active and inactive state. Without this information, the purpose of Figure 2D is unclear and Figure 2D, F and G are difficult to understand.
(b) Also, what is the threshold distance to classify a state as active or inactive?
(c) Figure 2E seems confusing with different axis and ranges.
In case of DDR1, does the MD simulations reflect the peak distances (between 7.5 and 10.0 Å for DFG-in and between 16.0 and 18.0 Å for DFG-out) observed for AF2 distance distributions? Also, the probability distribution shift towards shorter distances for Y755A does not seem particularly strong at first glance. Is this why the double alanine mutant was included? Are there also MD simulations of the double mutant that show a reduced preference for the DFG-out conformation?
The overall results on EGFR mutants are striking. Many of these mutants (most notably L858R have structures deposited in the PDB (ID:2ITT and many others) that are potentially part of the overall training of AF2/OpenFold. Can you comment on how this might affect the results?
There is some ambiguity in the statement, “The central hypothesis of this manuscript is that the collective contact distance distributions predicted by AF2 contain relevant information that can approximate Boltzmann distributions provided the relevant conformational states can be adequately described by these contact distances.” We suggest adding to this such that a stronger connection is formed between the theory section and the remainder of the paper. For example, the authors could explain that the contact distances specified in each section are the set of CVs you describe earlier, “we identify a set of CVs, ξ = (ξ1, ξ2, …, ξm)...”. It would also be helpful to clarify that the distributions predicted by AF2 represent the ensemble averaged observable, as described by equation 4. Lastly, the authors mention that these distributions can approximate Boltzmann distributions, but this is somewhat vague. This could be reworded to say that AF2 distributions can approximate experimentally derived Boltzmann distributions of the same distance.
The authors are comparing Cβ–Cβ distances determined by AF2 to spin label distances from EPR. This is explained in the methods section, but the procedure for adjusting the spin label distances to facilitate a meaningful comparison between them and the AF2 distances is somewhat unclear. To make a stronger justification for why these are comparable, the authors could clarify the procedure. For example, some context from the authors’ previous paper, De Novo High-Resolution Protein Structure Determination from Sparse Spin labeling EPR Data: “[distance from spin label] dSL is a starting point for the upper estimate of dCβ, and subtracting the effective distance of 6Å twice from dSL gives a starting point for the lower estimate of dCβ” could be beneficial. Including a rank correlation coefficient, as hinted above, could also help emphasize that the results demonstrate “similar relative probabilities among the contact distances for AF2 and EPR”
In the comparison of distance distributions between AF2 predictions and EPR measures, the major peaks of the two distributions are similar but in certain cases (127CB - 154CB, 120CB - 131CB), some additional peaks are found beyond 10A. A statistical comparison of the distributions, perhaps using a KS test, will help in evaluating the significance of the similarities.
Typo in Hamiltonian Equation 1 (should be momentum squared)
In the T4 Lysozyme example, how were the six contacts between the 12 unique residues found?
In Figure 5, the fourth row could have more discussion/explanation. What does the colorbar represent? There is no label.
As mentioned earlier, the connection between the Discussion and Figure 7 is not well established. The authors could expand on their writing and/or make the figure more simplified to match the discussion better.
Jessica Flowers, Angelica Lam, Ashraya Ravikumar, James Fraser
The author declares that they have no competing interests.