How the atlas is built

Every edge carries a source, an evidence type (observed · inferred · hypothesis · contradicts) and a confidence. Nothing here is medical advice.

Representation learning — node2vec training loss

Biased second-order random walks (p = 1, q = 0.7) sample the graph; a skip-gram model with negative sampling is trained by SGD in numpy (pipeline/build_graph.py). The vectors power the "nearest in embedding space" lists and the cross-cluster inferred links. The in-browser ranking is a separate Markov chain: random walk with restart over the weighted evidence graph.

Mechanism-first clusters (Louvain on a disease projection)

Data

Seed: curated slice with OMIM / Orphanet / HPO identifiers. pipeline/fetch_live.py adds PubMed and ClinicalTrials.gov records and verifies every curated identifier; the badge on each node shows its verification state.

Close