Year: 2026
Type: conference-paper
Abstract: Every Cure's computational pharmacophenomics platform has spent two years identifying the highest quality repurposing opportunities for patients. Across our approaches, standard ranking metrics often looked strong, but performance didn't reliably translate into high-quality repurposing candidates; those which an expert human reviewer considers to be addressing a clinically recognisable, high unmet need disease, and a hypothesis that is scientifically plausible and has a feasible experimental path. Four data issues drove this gap. Training data with class imbalance produces a small set of recurring "frequent flyer" drugs, regardless of the disease queried. Evaluation sets built from known on- and off-label pairs are not representative of high-quality repurposing opportunities, so benchmark scores are a weak proxy for real-world relevance. Knowledge graphs carry both noise, which invites shortcut learning, and gaps, which block the mechanistic paths a hypothesis needs to be credible. Disease lists that aren't restricted to diagnosable, deduplicated entities generate candidates that cannot be acted on downstream. This changed how we think about the task itself. Recommendation-style approaches assume a dense space of plausible options where every option needs ranking (e.g., a list of recommended films based on last film watched). Repurposing is better thought of as a search problem, with an extremely sparse number of true opportunities. Computational methods should surface few, high-quality candidate hypotheses, and no one method will identify all repurposing opportunities. This approach typically identified 2x to 4x more candidates at our first decision gate than prior approaches. To date, we have identified 7 active candidates through this approach, with tens more in review across our pipeline. Of these 7, 4 are undergoing lab validation, 2 are moving to clinical trials, and 1 is in active dissemination.
Institution: Every Cure
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: gold
Award: G819589331
OpenAlex ID: W7204997016