Knowledge graph-enhanced prediction of biomedical molecular relationships with MolNexus

Overview

• BioNexKG is a biomedical knowledge graph over compounds, proteins, genes, and metabolites.
• MolNexus provides CPI binary relationship prediction and three-class CGI/CMI prediction.
• MolNexus Skill exposes BioNexKG queries, MolNexus prediction, and pathway enrichment for use by an agent.

Dependencies

• Python >= 3.10
• PyTorch >= 2.0
• PyTorch Geometric >= 2.5
• NumPy
• Pandas
• Scikit-learn
• Matplotlib
• GSEApy

Quick Installation

git clone https://github.com/ZJUFanLab/MolNexus.git
cd MolNexus
conda env create -f environment.yml
conda activate molnexus

Usage

BioNexKG Relationship Query

python skills/MolNexus/scripts/query_kg.py \
  --compound-id 46220502 \
  --kind target \
  --kg-file demo/BioNexKG_demo/BioNexKG_demo.csv \
  --limit 100

python skills/MolNexus/scripts/query_names.py \
  --gene-symbol FYN \
  --kg-file demo/BioNexKG_demo/BioNexKG_demo.csv \
  --compound-registry demo/reference_demo/query_compounds_with_aliases.csv \
  --protein-mapping demo/reference_demo/protein_genes_idmapping.tsv \
  --limit 100

CPI Prediction

python skills/MolNexus/scripts/predict_cpi.py \
  --split DCS \
  --task-mode compound_screening \
  --compounds demo/input_demo/demo_screening_compounds.csv \
  --feat demo/input_demo/demo_screening_compounds_feat.csv \
  --candidate-file demo/input_demo/demo_screening_target.csv \
  --out-dir results/cpi/predict_cpi_demo \
  --execute

CGI Prediction

python skills/MolNexus/scripts/predict_cgi.py \
  --input-csv demo/input_demo/demo_cgi_external.csv \
  --compound-features demo/input_demo/demo_external_compounds_feat.csv \
  --model-dir model/cgi \
  --input-format triplet \
  --compound-id-col head \
  --gene-id-col tail \
  --label-col relation \
  --output-dir results/cgi/predict_cgi_demo \
  --execute

Classes: CGI-N = no significant change, CGI-D = down-regulation, CGI-U = up-regulation.

CMI Prediction

python skills/MolNexus/scripts/predict_cmi.py \
  --input-csv demo/input_demo/demo_cmi_external.csv \
  --model-dir model/cmi \
  --external-compound-features demo/input_demo/demo_external_compounds_feat.csv \
  --input-format triplet \
  --compound-id-col head \
  --metabolite-id-col tail \
  --label-col relation \
  --output-dir results/cmi/predict_cmi_demo \
  --execute

Classes: CMI-N = no significant change, CMI-D = down-regulation, CMI-U = up-regulation.

Pathway Enrichment

Pathway enrichment uses observed CGI-U and CGI-D genes from BioNexKG.

python skills/MolNexus/scripts/pathway_enrichment.py \
  --compound-id 46220502 \
  --kg-file demo/BioNexKG_demo/BioNexKG_demo.csv \
  --outdir results/pathway_enrichment

Model Training

Submit the five-fold training jobs with Slurm:

sbatch scripts/train_cpi.sh
sbatch scripts/train_cgi.sh
sbatch scripts/train_cmi.sh

MolNexus Skill

The Skill is located at skills/MolNexus/SKILL.md and supports BioNexKG relationship queries, CPI/CGI/CMI prediction, and pathway enrichment of observed CGI-U and CGI-D genes.

Example agent requests:

Use the MolNexus Skill to query known BioNexKG protein targets for abemaciclib.

Use the MolNexus Skill to query BioNexKG compounds with known CPI relationships to the protein target corresponding to gene symbol FYN.

Use the MolNexus Skill to perform pathway enrichment of observed CGI-U and CGI-D genes for compound CID 46220502.