Validate KGs against biolink schema

[redst4r]

KGs should conform to the biolink "type system". This would allow to catch systematic errors in the KG early on (either at ingestion or integration).

A few suggestions:

1. Minimally, each node's category (and edge's predicate) should be a valid biolink-class (biolink-predicate). Surprisingly that's not the case (e.g. biolink:Vitamin exists in our KG, but is NOT a valid biolink-class, similar for predicate biolink:contraindicated_for). Also abstract classes should probably not exists in the category of a node.
2. Edge types in biolink have specific domain and range, i.e. subject/object types, which we should enforce in the graph, e.g. biolink:in_taxon edges can only connect a ThingWithTaxon to a OrganismTaxon. Frequently violated at the moment.
3. Nodes often have more than one category (see all_categories). Often it's superclass-subclass relations, e.g. a node has all_categories =["Protein", "Polypeptide"], which are valid. However, certain combinations of category on the same node point to errors (e.g. a node can't really be a Disease and a Gene; the gene might be mutated in the disease, but those are still different concepts that should not be mixed up )

Comments

1. is easy, something like

import bmt
import polars as pl
B = bmt.Toolkit('https://raw.githubusercontent.com/biolink/biolink-model/refs/heads/master/biolink-model.yaml')
valid_classes = [B.get_element(el_name)['class_uri']for el_name in B.get_all_classes()]

df_nodes.with_columns(
   # not sure how do check for set-equality in polars, this one works though:
    valid_biolink=pl.col("all_categories").list.set_intersection(valid_classes).list.len() == pl.col("all_categories").list.len(),
)

2. fairly easy too. One just needs to take into account the inheritance, e.g. if an edge type has domain=="ThingWithTaxon", any subclass is a valid subject
3. Harder. Anything that adheres to the biolink class hierarchy is definitely valid, but the rest is tricky (if the node's all_categories violates the class hierarchy, it's not neccessarily wrong, e.g.:
   • genes and proteins are often mixed in a single node(i guess the type should really be GeneOrGeneProduct then)
   • A node might be both a Protein and a Drug (e.g. antibody)
   • Proteins are sometimes SmallMolecules (if its just a few AAs) ...
     We'd need to come up with a "blacklist" of category combinations that are "wrong" (.e.g Disease, Gene)

[matentzn]

@eKathleenCarter @jdr0887 I assigned you so you can decide with EC what if anything of this issue you want to implement - else you can close it if there is no interest from EC side or this is already done.