Knowledge Graph Exchange

KGX (Knowledge Graph Exchange) is a Python library and set of command line utilities for exchanging Knowledge Graphs (KGs) that conform to or are aligned to the Biolink Model.

The core datamodel is a Property Graph (PG), represented internally in Python using a networkx MultiDiGraph model.

KGX allows conversion to and from:

• RDF serializations (read/write) and SPARQL endpoints (read)
• Neo4j endpoints (read/write)
• ArangoDB endpoints (read/write)
• CSV/TSV and JSON (see associated data formats and example script to load CSV/TSV to Neo4j)
• Reasoner Standard API format
• OBOGraph JSON format

KGX will also provide validation, to ensure the KGs are conformant to the Biolink Model: making sure nodes are categorized using Biolink classes, edges are labeled using valid Biolink relationship types, and valid properties are used.

Internal representation is a property graph, specifically a networkx MultiDiGraph.

The structure of this graph is expected to conform to the Biolink Model standard, as specified in the KGX format specification.

In addition to the main code-base, KGX also provides a series of command line operations.

Example usage

Validate:

poetry run kgx validate -i tsv tests/resources/merge/test2_nodes.tsv tests/resources/merge/test2_edges.tsv

Merge:

poetry run kgx merge —merge-config tests/resources/test-merge.yaml 

Graph Summary:

poetry run kgx graph-summary -i tests/resources/graph_nodes.tsv  -o summary.txt

Transform:

poetry run kgx transform —transform-config tests/resources/test-transform-tsv-rdf.yaml

Error Detection and Reporting

Non-redundant JSON-formatted structured error logging is now provided in KGX Transformer, Validator, GraphSummary and MetaKnowledgeGraph operations. See the various unit tests for the general design pattern (using the Validator as an example here):

from kgx.validator import Validator
from kgx.transformer import Transformer

Validator.set_biolink_model("2.11.0")

# Validator assumes the currently set Biolink Release
validator = Validator()

transformer = Transformer(stream=True)

transformer.transform(
    input_args = {
        "filename": [
            "graph_nodes.tsv",
            "graph_edges.tsv",
        ],
        "format": "tsv",
    },
    output_args={
        "format": "null"
    },
    inspector=validator,
)

# Both the Validator and the Transformer can independently capture errors

# The Validator, from the overall semantics of the graph...
# Here, we just report severe Errors from the Validator (no Warnings)
validator.write_report(open("validation_errors.json", "w"), "Error")

# The Transformer, from the syntax of the input files... 
# Here, we catch *all* Errors and Warnings (by not providing a filter)
transformer.write_report(open("input_errors.json", "w"))

The JSON error outputs will look something like this:

{
    "ERROR": {
        "MISSING_EDGE_PROPERTY": {
            "Required edge property 'id' is missing": [
                "A:123->X:1",
                "B:456->Y:2"
            ],
            "Required edge property 'object' is missing": [
                "A:123->X:1"
            ],
            "Required edge property 'predicate' is missing": [
                "A:123->X:1"
            ],
            "Required edge property 'subject' is missing": [
                "A:123->X:1",
                "B:456->Y:2"
            ]
        }
    },
    "WARNING": {
        "DUPLICATE_NODE": {
          "Node 'id' duplicated in input data": [
            "MONDO:0010011",
            "REACT:R-HSA-5635838"
          ]
        }
    }
}

This system reduces the significant redundancies of earlier line-oriented KGX logging text output files, in that graph entities with the same class of error are simply aggregated in lists of names/identifiers at the leaf level of the JSON structure.

The top level JSON tags originate from the MessageLevel class and the second level tags from the ErrorType class in the error_detection module, while the third level messages are hard coded as log_error method messages in the code.

It is likely that additional error conditions within KGX can be efficiently captured and reported in the future using this general framework.

Installation

Installing from PyPI

KGX is available on PyPI and can be installed using pip as follows,

pip install kgx

To install a particular version of KGX, be sure to specify the version number,

pip install kgx==0.5.0

Installing from GitHub

Clone the GitHub repository and then install,

git clone https://github.com/biolink/kgx
cd kgx
poetry install

Setting up a testing environment for Neo4j

This release of KGX supports graph source and sink transactions with Neo4j 4.4.

KGX has a suite of tests that rely on Docker containers to run Neo4j specific tests.

To set up the required containers, first install Docker on your local machine.

Once Docker is up and running, run the following commands:

docker run -d --name kgx-neo4j-integration-test -p 7474:7474 -p 7687:7687 --env NEO4J_AUTH=neo4j/test neo4j:4.4

docker run -d --name kgx-neo4j-unit-test -p 7474:7474 -p 7687:7687 --env NEO4J_AUTH=neo4j/test neo4j:4.4

Note: The integration and unit test containers use the same ports and cannot run simultaneously. Setting up the Neo4j container is optional. If there is no container set up then the tests that rely on them are skipped.

Setting up a testing environment for ArangoDB

KGX also has a suite of tests that rely on a running ArangoDB instance. A single container serves both unit and integration tests:

docker run -d -p 8529:8529 --env ARANGO_NO_AUTH=1 arangodb/arangodb:latest

Note: Setting up the ArangoDB container is optional. If ArangoDB is not reachable on localhost:8529 then the tests that rely on it are skipped.