A Rust graph database loaded a 328-million-node biomedical knowledge store in 38 minutes, merging 132,511 cross-source entities from five identifier keys
AspireAI community · arXiv
Biomedical knowledge is fragmented across siloed databases including Reactome, STRING, Gene Ontology, ClinicalTrials.gov, DrugBank, DGIdb, SIDER, PubMed, ClinVar, ChEMBL and OpenTargets. The authors present an open, reproducible biomedical knowledge graph ecosystem built on Samyama, a Rust-native graph database, with three foundational graphs described end to end: Pathways KG at 118,686 nodes and 834,785 edges, Clinical Trials KG at 7,774,446 nodes and 26,973,997 edges, and Drug Interactions KG at 32,726 nodes and 191,970 edges, expanded into a 22-graph, two-cohort merged store of 328,130,739 nodes and 1,441,425,733 imported edges including 295,000 synthetic OMOP CDM patients from Synthea. They describe a reproducible ETL and portable snapshot pattern, and analyze single-store cross-KG integration through query-time property joins and load-time identifier deduplication: across all 24 snapshots, the 328.1M nodes and 1.44B edges load in 2,269.1 seconds within 366.5 GB peak RSS, about 254 bytes per edge, merging 132,511 cross-source entities across five audited keys at 87.8% of the 150,965 bound sound-identifier ceiling with zero conflicting-ID refusals. On a frozen 81-query suite 62 queries return results, with 17 data-gap empties and 2 timeouts, while an expanded 200-query oracle suite passes 195 of 200 at 97.5%. The limitation is that 19 of 81 queries returned nothing and the graph exposes no deep semantic layer of its own, so its value is as infrastructure for retrieval over biomedical entities rather than as a reasoning system.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.