volkamerlab / ratar Goto Github PK

First three moments of distribution:

• Same units: 2nd moment - standard deviation; 3rd moment: 3rd root of skewness
• Scaling: none (should fingerprint/moments be normalised?)

Write class method from_path, analogous to from_molecule. Current problem: files can contain multiple molecules, thus from_path would return a list of molecule objects instead of a molecule object as in case of from_molecule.

Differing behaviour here will not work well downstream, right? Check this.

How robust are reference points in binding site?

• Within one binding site: how do different reference points affect distance distributions?
• Across similar binding sites (e.g. kinases): similar reference points?

Binding site definition/size varies between datasets:

• KLIFS (~1300 atoms)
• scPDB (~500 atoms)
• TOUGH (~150 atoms)

What size is needed for good performance of encoding method?
Can we compare performance on these datasets with each other?

Here is a list of the packages that are used in the PR refactoring #1 , queried using https://github.com/volkamerlab/ratar/search?p=1&q=import

biopandas
pandas
numpy
seaborn
pymol

Testing or env scripts:
pytest
yaml

Used but not necessary to put in conda env, since already in python (see here):

• sys
• pathlib
• datetime
• argparse
• os
• glob
• re
• shutil
• subprocess
• typing
• contextlib
• tempfile
• pickle

Molecules contain atoms belonging not to standard amino acids:

• protonated amino acids
• non-standard amino acids
• ions
• water
• other solvent molecules

For calculations including z-scales, all atoms that do not belong to a standard amino acids are removed.

What about other encoding methods such as pdbqt?

Updates needed for this code base:

December 2021

See PR #14 for details.

• Update environment file as suggested by @t-kimber in #12
• Check if CLI is still running > now it is again.
• Update README + installation + usage instructions (update tutorial!!!!)
• Fill follow-up section below

Follow-up

Methodology / encoding

• So far only full pocket encoding; we probably need subpocket encoding (overlapping patches)
• Define binding sites and their size #7
• Include non-standard amino acids #8
• Check units of 4th to 6th dimensions of reference points
• Check units and scaling of moments - First three moments of distribution:
  • Same units: 2nd moment - standard deviation; 3rd moment: 3rd root of skewness
  • Scaling: none (should fingerprint/moments be normalised?)
• We started to look into pdbqt files to be added as "physchem" properties to our fingerprint, take a look at this notebook if still of interest
• We already started to benchmark the method against similar/dissimilar pocket pairs from FuzCav, ProSPECCTs, and TOUGH-M1 (see README)
• Encoding workers fine for mol2 files; pdb files may not and need revision
• Since we probably have to move to NGLview anyways, PyMol functions have not been checked since 2019; probably they do not work anymore.

Testing and CI

• Add unit tests for similarity module
• CI: Add back Windows + MacOS support, lint package, format+lint+test docs tutorials

Code

• Check similarity module - refactoring needed?
• Address #FIXMEs and #TODOs in code (left-overs of major refactoring in PR #1); to be done after setting up unit tests
• Remove pymol dependency (not on conda-forge; currently installed from tpeulen)
• Remove flatten-dict dependency (only pip-installable)
• Add from_path class method to all ratar.encoding classes: Write class method from_path, analogous to from_molecule. Current problem: files can contain multiple molecules, thus from_path would return a list of molecule objects instead of a molecule object as in the case of from_molecule.
• We set up a logging.conf file to fine-grain our logging. Include back into the package if of interest.

Packaging

• Update ratar environment - enable conda packaging