Skip to content

05 — Making Aging & Biomedical KGs

📂 Source Code & Notebooks: pipeline/05_kg_construction

1. Purpose

This is Step 5 of the EvoAge Knowledge Graph (KG) construction pipeline. The goal is to split the unified processed triples (from Steps 02–04) into two distinct KG categories — Aging-specific and Biomedical (Generalised) — based on the kg_type column assigned during data processing. This split is performed separately for Human data and for each non-human species (using both 1:1 and 1:1∪1:M ortholog-mapped files from Step 04). Additionally, this step generates Species_AssociatedWith connection triples that link each species to the genes/entities it is associated with.

2. Overview

The pipeline has 4 runs across two phases:

Phase Runs What it does
Phase A: KG Splitting Run 1, Run 2, Run 3 Splits every processed triple file into Aging vs Biomedical buckets using the kg_type column
Phase B: Species Connections Run 4.1, Run 4.2 Creates Species_AssociatedWith triples linking each species to its entities (one file per species)

How the kg_type classification works

Every triple file from Step 02/03 carries a kg_type column with one of these values:

kg_type value Routed to Rationale
aging Aging only Triple comes from an aging-specific database (GenAge, DrugAge, CellAge, etc.)
generalised Biomedical only Triple comes from a general biomedical source (PrimeKG, DRKG, STRING, etc.)
aging::generalised Both Aging AND Biomedical Triple found in both KG's (e.g., aging genes that also appear in general KGs)

3. Pipeline — Step-by-Step

Run 1 — Split Human data by kg_type

📄 Script: Run1_split_kg_by_type_human.py

What it does:

  • Reads all CSV/Parquet files under processed_data_relation_wise_merge/generalised/ (excluding OTHER_SPECIES/)
  • For each file, classifies every row by its kg_type value into Aging and/or Biomedical
  • Saves Aging rows to Aging_specific/Human/{relation_folder}/ (as Parquet)
  • Saves Biomedical rows to Biomedical/Human/{relation_folder}/ (as Parquet)
  • The output directory structure mirrors the source relation-folder hierarchy

Input: processed_data_relation_wise_merge/generalised/{RELATION_TYPE}/ALL_{RELATION_TYPE}.csv|parquet

Output: - processed_data_relation_wise_merge/Aging_specific/Human/{RELATION_TYPE}/...parquet - processed_data_relation_wise_merge/Biomedical/Human/{RELATION_TYPE}/...parquet

Log file: [Human_KG_SPLIT_LOG.csv]

Run 2 — Split 1:1 ortholog files by kg_type (other species)

📄 Script: Run2_split_1to1_kg_by_type_otherspecies.py

What it does:

  • For each species, reads only *_ortho_1_to_1.csv files (produced by Step 04, Run 3)
  • Splits rows by kg_type into Aging / Biomedical using the same classification logic as Run 1
  • Saves to Aging_specific/{Species}/{relation}/ and Biomedical/{Species}/{relation}/ as CSV

Log file: _OTHER_SPECIES_KG_SPLIT_LOG_1to1.csv

Run 3 — Split 1:1∪1:M ortholog files by kg_type (other species)

📄 Script: Run3_split_12M_121_comb_kg_by_type_otherspecies.py

What it does:

  • Identical logic to Run 2, but reads *_ortho_1_to_one2one_plus_one2many.csv files (produced by Step 04, Run 5)
  • Produces the Aging/Biomedical split for the broader ortholog mapping Log file: _OTHER_SPECIES_KG_SPLIT_LOG.csv

Run 4 — Generate Species_AssociatedWith connection triples

This run creates a new relation type: Species_AssociatedWith, which connects each species node to every entity (gene, protein, etc.) it is associated with in the KG.

Run 4.1 — Species triples for the 1:1 ortholog KG

📄 Script: Run_4.1_make_species_triples_1_to_1.py

What it does:

  • Reads all Human generalised files + all *_ortho_1_to_1.csv files from other species
  • For each file, extracts head_species → Species_AssociatedWith → head and tail_species → Species_AssociatedWith → tail triples
  • Deduplicates across all files
  • Splits by species and saves one Parquet per species

Output: making_species_assocaitedwithconnection/1_to_1_ortholog/

Run 4.2 — Species triples for the 1:1∪1:M ortholog KG

📄 Script: Run_4.2_make_species_triples_combined121_12M.py

What it does:

  • Same logic as Run 4.1, but reads *_ortho_1_to_one2one_plus_one2many.csv for other species
  • Produces species triples for the broader 1:1∪1:M KG

Output: making_species_assocaitedwithconnection/one2one_plus_one2many_ortholog/

5. Next Step

→ [Step 06 — Final KG Building]: Merge all Aging and Biomedical splits across species into the final consolidated KGs for training and evaluation.