Loading…
Disease embeddings learned from human genetic evidence and phenotype ontologies, placing rare and common conditions in one mechanistic vector space.
No providers recorded yet. Browse all providers
Deciding whether a drug that works in one disease is worth testing in another is one of the most expensive judgment calls in pharmaceutical development, and it is hardest exactly where the data is thinnest: rare disease. Computational approaches have mostly come from biomedical knowledge graphs and from language models trained on literature, both of which inherit the biases of what has already been studied. A model learning from those sources recovers well-trodden associations and fails on the sparsely annotated conditions where a prediction would change a decision.
Dis2Vec, from Alexion, AstraZeneca Rare Disease with Tina Eliassi-Rad at Northeastern University, takes a deliberately narrower input diet. It learns a vector for each disease from two signals only — human genetic association evidence and curated phenotype annotations — and grounds the similarity between any two diseases in how close their genes and symptoms sit within the Gene Ontology and Human Phenotype Ontology hierarchies. Drug and clinical trial information is excluded from training entirely and reserved for evaluation, so the benchmark tests transfer to an independent evidence modality rather than reconstruction of the training graph.
The preprint first appeared in late 2024 and has been substantially revised since, reframed around indication expansion and investment decisions. An unrelated 2016 model of the same name — a word2vec variant over a news corpus for disease taxonomy generation — shares nothing with this work.
MechDist is built by computing ontology distances for every disease pair across both modalities, thresholding at 0.2 to drop distant pairs, and aggregating with CRank. Embeddings are then learned with Node2Vec over that homogeneous, unweighted graph: 128 dimensions per disease, walk length 10, 100 walks per node, window size 10, and return and in-out parameters both set to 1. The initial pool of 6,654 diseases passing the evidence filter is reduced to the 4,710 that meet the quality threshold for a reliable representation.
Two downstream tasks reuse the embeddings unchanged. Disease-disease association learning runs them through UPNA-DDAL, a three-layer multilayer perceptron decoder in a two-shot arrangement, trained only on MechDist with all benchmark edges removed. Evaluation uses two independent networks: a positive control derived from clinical trial investment decisions, where an edge joins two diseases that entered trials against the same drug, and a negative control built by maximizing ontology distances in the Orphanet Rare Disease Ontology. Under that protocol Dis2Vec-UPNA-DDAL reached an AUROC of 0.8975, against 0.780 for a plain graph convolutional network on the same graph and 0.575 for a knowledge-graph foundation model. The second task, Gaussian mixture clustering into 50 components, produced a disease map that groups lysosomal acid lipase deficiency with its subtypes while correctly separating neurofibromatosis types 1 and 2, whose mechanisms differ.
The intended users are translational and portfolio teams asking where an existing asset might work next, particularly in rare disease where cohort evidence is too sparse for conventional analysis. Illustrative outputs include a high-confidence link between atypical hemolytic uremic syndrome and paroxysmal nocturnal hemoglobinuria — treated with the same approved antibody, and not directly connected in the training graph — and strong associations between rheumatoid arthritis and a family of rare idiopathic inflammatory myopathies. The Zenodo deposit carries the Node2Vec embeddings, the trained UPNA-DDAL decoder, and the ontology distance matrices and Open Targets tables behind them, all under CC BY 4.0.
Dis2Vec's contribution is as much methodological as it is a released artifact. Its cross-evidence
benchmark — train on ontology-derived mechanism, test on clinical trial investment decisions —
exposes how much conventional drug-repurposing evaluations reward memorizing graph topology, and
under that protocol a well-regarded knowledge graph model falls near chance while a simple
homogeneous-graph baseline does not. Its limits appear where mechanism is the wrong abstraction: the
decoder recovers only 21 of 55 indications for a drug that acts by read-through of nonsense
mutations, since those diseases share a molecular lesion rather than a pathophysiology. The paper
remains a preprint, and while the artifacts are openly licensed the pipeline behind them is not:
the repository README lists Dis2Vec_UPNA-DDAL/ and disease_definition_pipeline/ as TBD and
neither exists, so MechDist cannot be rebuilt or the decoder retrained. The work also originates
from a company with a direct commercial interest in rare disease indication expansion.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.