Using ChimeraX to analyse AlphaFold predictions

These 3 icons are within the “Graphics tab”. This tab groups quick-access tools to change background colour and lighting effects of your window.

Don’t hesitate to play around a little with the different effects.

You will find this icon within the “Molecular display tab”. This tab groups quick-access tools to change structure granulosity (atomic, secondary structure and surface representation) as well as colouring (by heteroatom, chain, by position in the sequence, electrostatics & hydrophobicity…).

Don’t hesitate to play around a little with the different colours & representations.

In order to change the default colour palette used for bfactor colouring, you can, instead, use the command line to colour your structure and specify the alphafold colour palette:

color bfactor palette alphafold

As a reminder, pLDDT scores are a per-residue score between 0 and 100. The higher the score, the higher the confidence in the local environment of a residue.

As you can see, loops and terminal regions have a low confidence whereas well-structured regions have a high confidence i.e. AlphaFold is more confident in its prediction of the well-structured regions, and is not sure about the structure of the loop and terminal regions.

This is not at all surprising: Helices and sheets have well-defined, repeating backbone geometries and stabilising interactions. They are also often more evolutionarily conserved. All this information makes these regions easier to predict, thus AlphaFold is more confident in what it outputs.

The correct pickle file to import is result_model_3_ptm_pred_2.pkl within the afm_defaults/light_pkl/ folder.

As a reminder, the PAE plot is a square matrix with the same length & height as the number of residues in your structure. It does not show scores but estimated errors on the distance between 2 residues of the structure. It’s a value between 0 and 32 Ångström. You can see it as a “±” value that you can add to the actual distance you see in the predicted structure. Thus, the lower it is, the better (i.e. AlphaFold is more confident in the distance that it has predicted between 2 residues).

The default colour scale of the PAE matrix is explained when you click on the Help button. The colour scale tries to reflect the pLDDT colours: blue for low error values, yellow-orange for medium values and grey-white regions for high error values.

When you highlight a light area in the PAE outside of the diagonal, it will show the corresponding regions in the structure as pink (y-axis) and green (x-axis) areas.

Unsurprisingly, these light areas are seen between residues of well-structured regions and predicted loop regions with very poor pLDDT scores. Indeed, loops have generally fewer structural constraints (they are often linkers or turns), so their exact position relative to other parts of the protein is harder to predict (i.e. more prone to error).

Well-packed domains typically appear as blocks of low error along the diagonal of the plot. These blocks represent regions where residues whithin the domain have low predicted positional error relative to each other, indicating a stable well-folded structure.

ChimeraX has many inbuilt features, of which, the identification of predicted well-packed domains from a PAE plot (first button at the bottom of the PAE plot). You can then “right click” on the PAE plot to colour the PAE as the structure and keep the PAE values as a background grey scale. This is a useful tool to more easily find correspondances between the structure and the PAE plot and to help you read the PAE plot.

As you can see with the colours, ChimeraX found 3 different domains from the plot: 2 loops and the rest of the protein as a third well-packed domain. This is quite coherent with the structure.

PAE plots become particularly interesting with multi-domain and/or multi-chain proteins as they will help identify if 2 domains are likely to interact (lower PAE values) or not (higher PAE values). Unless the interaction is very strong (e.g. obligate oligomer), you should not expect the PAE values to match the intra-domain ones (the signal will be more diluted).

To help you answer this question, you can use the same trick as before: first colour the structure by pLDDT, then “right click” on the PAE to update the colouring as in the structure.

The resulting plot isn’t the most easy to read but if you look closely, it depicts the pLDDT score along the diagonal, together with the PAE error values in grey scale.

PAE values and pLDDT scores represent 2 different types of confidence levels, but as you can see, in this prediction, they are quite in agreement in the predicted domains.

NB: you can also restrict to a given selection by selecting the regions before running matchmaker, then by ticking the “Also restrict to selection” boxes.

As you can see, the structures do not superimpose that cleanly:

To get a split view, you can use the tile function in the command line:

tile columns 3 spacing_factor 1

In the above command, we specify that we want 3 columns and we reduce the default spacing between models to a factor 1. To remove tiling, you can just type: tile off

Tiling with the tile command can be useful in this case in order to see more clearly. If you set both structures side-by-side and colour the model by pLDDT, you can see that confidence score is not too bad for the region that overlaps with 4FVJ.