dgram2dmap

Convert AlphaFold distograms into distance matrices and save them into a number of formats.

The distances for a residue pair $(i, j)$ are obtained by passing the distogram logits through a softmax function, then performing a linear combination between the distogram bin edges $edge$ and the softmax output $softmax(distogram_{i,j})$ for each distance bin $b$:

$$ dist_{i,j} = \sum_{b=1}^{64} {softmax(distogram_{i,j})}_b * edge_b $$

Distances calculated this way agree quite well with the actual $C\beta$ distances extracted from a model (up until ~20Å), provided that the predicted aligned error is low.

Examples

dgram2dmap allows to plot the converted distance maps and compare them with the predicted aligned error, while highlighting distances between numbered subsets of amino acids or chain identifiers:

It also allows to compare against a 3D model:

Usage

usage: dgram2dmap.py [-h] [--maxD 20.0] [--limits i:j k:l] [--chains chain1 chain2] [--plot] [--argmax] [--rosetta] [--pdb ranked_0.pdb] in_folder

Extract and format distance constraints from AlphaFold distograms

positional arguments:
  in_folder             AlphaFold model output folder

optional arguments:
  -h, --help            show this help message and exit
  --maxD 20.0           Maximum distance (in Å) for constraints output
  --limits i:j k:l      Select a 'patch' of constraints between two subsets of residues (e.g. 0:100 200:300)
  --chains chain1 chain2
                        Extract constraints between two chains (e.g. A B)
  --plot                Plot the distances with bounding boxes
  --argmax              Use argmax to find the most likely distance instead of interpolating
  --rosetta             Export below-threshold (see maxD) distances in a Rosetta constraint files
  --pdb model.pdb       PDB model of the target protein (e.g. for comparisons against native)

Example:

If you want to extract the distances from the pickle files for CASP15 target T1105 (available here) while extracting the distances between residues 50:100 and 150:200 to a rosetta constraint file:

python dgram_to_dmap.py AF_outputs/T1105/ --chains 50:100 150:200 --plot --rosetta

Which will produce the following outputs for each pickle file (see also example/ folder):

• result_model_1_ptm_pred_0.pkl.rosetta_constraints: rosetta constraint file from selected ranges
• result_model_1_ptm_pred_0.pkl.dmap: CSV file with all distances calculated from the distogram
• result_model_1_ptm_pred_0.pkl.dmap.png: image of the calculated distances (with selection range boxes, if any). The predicted aligned error is also shown if present:
• result_model_1_ptm_pred_0.pkl.agreement.png: comparison between model (lower corner) and distogram (upper corner) distance map, scatter plot of model vs. distogram distances in range boxes (see plot above)