Number of nodes and shape of node coordinates differ by 1

[diamondspark]

Hi @a-r-j

I've been observing that some times the number of nodes generated by the library differs to the coordinate data generated by it by exactly 1 node. Do you know why this happens. Following is an example code

configs = {
        "granularity": "CA",
        "keep_hets": False,
        "insertions": False,
        "verbose": False,
        "dssp_config": DSSPConfig(),
        "node_metadata_functions": [meiler_embedding,expasy_protein_scale],
        "edge_construction_functions": [add_peptide_bonds,
                                                  add_hydrogen_bond_interactions,
                                                  add_ionic_interactions,
                                                  add_aromatic_sulphur_interactions,
                                                  add_hydrophobic_interactions,
                                                  add_cation_pi_interactions]
        }
config = ProteinGraphConfig(**configs)
format_convertor = GraphFormatConvertor('nx', 'pyg', 
                                            verbose = 'gnn', 
                                            columns = None)
g = construct_graph(config=config, pdb_code='1c5y')
protdata = format_convertor(g)
print(protdata)

protdata.num_nodes == 256 ; protdata.coords[0].shape==257

Shouldn't these 2 be the same? What am I missing? Thank you!

[a-r-j]

Yep, you're right @diamondspark . There was a bug in how inserted residues are removed from the dataframe. I've pushed a fix to a pending PR. Need to write some tests but hope to get this merged in soon!

There should actually be 238 nodes in the graph as we remove the inserted residues.

[diamondspark]

Thank you for looking into this. This seems to only partially work. I have following follow up concerns

• How to get the updated code. I did pip install graphein --upgrade. That didn't change the code in protein/graphs.py

• g = construct_graph(config=config, pdb_code='1c5y') gives following error now

  
  ~/anaconda3/envs/bar/lib/python3.8/site-packages/graphein/protein/edges/distance.py in add_ionic_interactions(G, rgroup_df)
  230         condition1 = (
  231             G.nodes[r1]["residue_name"] in POS_AA
  232             and G.nodes[r2]["residue_name"] in NEG_AA
  233         )
  234 KeyError: 'residue_name'

• The shape mismatch still persist in some other calculations such as for Meiler embedding

  
  configs = {
      "granularity": "CA",
      "keep_hets": False,
      "insertions": False,
      "verbose": False,
      "dssp_config": DSSPConfig(),
      "pdb_dir":'/groups/cherkasvgrp/Student_backup/mkpandey/My_Projects/Drug_Protein_Interaction_Project1/ER_AR_project/data/PDBBind/pdbbind_v2016_refined/refined-set/',
          "pdb_dir":'./data/prot/PDB/',
      "node_metadata_functions": [meiler_embedding,expasy_protein_scale],
      "edge_construction_functions": [add_peptide_bonds,
                                                add_hydrogen_bond_interactions,
                                                add_ionic_interactions,
                                                add_aromatic_sulphur_interactions,
                                                add_hydrophobic_interactions,
                                                add_cation_pi_interactions]
      }

config = ProteinGraphConfig(**configs) format_convertor = GraphFormatConvertor('nx', 'pyg', verbose = 'all_info', columns = ['edge_index','meiler','coords','expasy','node_id','name','dist_mat','num_nodes']) g = construct_graph(config=config, pdb_code='6OGE') protdata = format_convertor(g)

yields different shape for meiler and node features (1483 and 1487)

Data(edge_index=[2, 2472], node_id=[1487], coords=[1], meiler=[1483], expasy=[1483], name=[1], dist_mat=[1], num_nodes=1487)

@a-r-j Can you please look into this again? Thank you!

[a-r-j]

Hi, @diamondspark. I'm working on this. It's a tricky problem resulting from insertions and alt_locs in the PDB files. These aren't always consinstently represented in the file so it's hard to come up with a robust way that catches all the corner cases.

While I'm figuring it out you can try pre-processing the PDBs with the excellent PDB Tools.

[a-r-j]

Hi @diamondspark I believe this is resolved in 1.1.1 - try it out & do let me know if not!