BUG in construct_graph

[kamurani]

I have found quite a severe bug in construct_graph when using the add_distance_threshold edge construction function.

Essentially, spurious edges are added to the graph that should not appear, based on the provided threshold distance in Angstroms.

I have decided to make an issue here ASAP before I can diagnose this but I intend to track down why this is happening.

This is potentially a very serious bug as using threshold_distance edges is a common protein graph construction technique and may result in people being unaware of the true structure of their data.

Version

• MacOSX
• Python 3.12.4

import graphein 
graphein.__version__
# '1.7.7'

Steps to reproduce bug

Download AlphaFold predicted structure here

uniprot_id = "P04629"

from functools import partial
from graphein.protein import construct_graph
from graphein.protein.config import ProteinGraphConfig
from graphein.protein.edges.distance import add_distance_threshold

u, v = 'A:MET:1', 'A:LEU:526'

new_edge_funcs = {"edge_construction_functions": [partial(add_distance_threshold, long_interaction_threshold=5, threshold=10.)]}
config = ProteinGraphConfig(**new_edge_funcs)
g_dist_all = construct_graph(
    path = f"{uniprot_id}.pdb",
    config=config,
)

g_dist_all.get_edge_data(u, v)
# {'kind': {'distance_threshold'}, 'distance': 48.41152358823357}

Note that this edge mysteriously appears, despite the node-node distance being well over the threshold distance specified!

[a-r-j]

Hi @kamurani thanks for flagging. I could reproduce the error. I think I've tracked down the source.

The issue is the indexing here:

graphein/graphein/protein/edges/distance.py

Line 967 in 46b935b

n1 = G.graph["pdb_df"].loc[a1, "node_id"]

For some reason the dataframe index is shifted. I believe it happens in this function, likely at this sort operation:

graphein/graphein/protein/graphs.py

Line 460 in 46b935b

protein_df = sort_dataframe(protein_df)

I suspect I didn't catch this when switching from biopandas to CPDB for parsing.

I see two options. 1. add an index reset in process_dataframe or change the indexing in the edges function to use iloc rather than loc. I prefer the first option. Would you be able to try this and make a PR?

[kamurani]

I will have a chance today to work on it, will have a go and submit.