BioGraph-AI
A fixed Open Targets graph-method artifact for retrospective disease-target ranking and failure analysis.
BioGraph-AI connects eight brain-tumor disease concepts with known target and drug nodes from a source-hashed Open Targets snapshot. Its executable artifact combines 32-dimensional SVD node embeddings with a logistic classifier over disease-target Hadamard products.
The random edge holdout reached AUROC 0.711 and hits@10 0.194, but the more demanding Open Targets 25.03-to-current-cache holdout fell to AUROC 0.457 and hits@10 0.0084. That weak release result controls the claim boundary and blocks target-prioritization language.
The public explorer scores only known disease-target pairs from the fixed graph and surfaces cached association components, disease→drug→target paths, and a limited PubTator relation-count pilot. The output is an uncalibrated ranking score, not a probability or recommendation.
Clinical deployment note
These system pages describe technical capabilities and evaluation workflows. Any diagnostic, treatment, or patient-specific use would require appropriate validation, governance, and regulatory review.