Paper 5b: Federated Biomedical Knowledge Graphs — ISWC 2026 In-Use Track

Title: Federated Biomedical Knowledge Graphs at Scale: A Single-Server, Single-Query Substrate for AI Agents

Venue: International Semantic Web Conference (ISWC) 2026, In-Use Track

Status: Submitted

Authors: Madhulatha Mandarapu, Sandeep Kunkunuru

Summary

Companion to the GRADES-NDA demo: a full In-Use track paper documenting the production deployment of 11 federated biomedical and public-health knowledge graphs on a single Samyama Graph server, queryable through a single OpenCypher interface and exposed to AI agents via MCP tools.

The paper covers the engineering choices that make federation tractable on commodity hardware — graph-native query planning, columnar property storage with late materialization, adjacency-aware aggregation for degree-bounded queries, and MVCC for concurrent reader/writer workloads — and reports the 500/500-pass result on the federated benchmark.

Key Results

• 11 KGs federated: ~138M nodes, 1.22B edges, single server (r7i.16xlarge spot)
• 500-query mega-benchmark: 500 / 500 pass (468 full-pass + 32 corpus-gap data-dependent, 0 errors)
• Total benchmark runtime: 78 minutes (load + query)
• BiomedQA via MCP tools: 98% accuracy

Files

Manuscript sources, build scripts, and reproduction artifacts (B3 baselines, LDBC calibration logs) live in the private samyama-research repository under papers/paper5-iswc-inuse/.