Drug Repurposing: Finding New Uses for Approved Drugs¶
Industry: Pharmaceutical R&D | Role: VP Translational Science, Computational Biology Lead | Time to value: 2-3 hours
The Problem¶
There are thousands of approved drugs with well-characterized safety profiles. There are thousands of diseases with unmet therapeutic need. Somewhere in the intersection are repurposing candidates — approved drugs that could treat diseases they were never designed for. The challenge is finding them systematically, ranking them by evidence strength, and explaining why each candidate was predicted.
The Traditional Approach¶
Computational repurposing pipelines are typically assembled from separate components: literature mining to extract drug-target and target-disease associations, custom graph traversal code to walk protein-protein interaction networks, bespoke aggregation logic to combine evidence from multiple mechanistic pathways, and a structural similarity search for analogue discovery. A team of 3-4 computational biologists spends a sprint wiring these together — typically 500+ lines of Python — with no built-in explainability. When a reviewer asks "why was Drug X predicted for Disease Y?", someone traces the logic manually. Adding a new evidence source means modifying the traversal code.
With Uni¶
The notebook encodes the entire pipeline in a handful of declarative rules. Multi-hop protein-protein interaction traversal (up to 3 hops) finds indirect mechanistic paths from drug targets to disease-associated proteins. Noisy-OR aggregation (FOLD MNOR) combines evidence from independent pathways — if a drug reaches a disease through several separate mechanisms, each adds independent support. Structural analogue discovery (similar_to on molecular fingerprint vectors) identifies approved drugs with similar structure that share target diseases. Novelty filtering (IS NOT) removes known indications. Built-in EXPLAIN RULE traces exactly why each drug-disease pair was predicted — the full mechanistic derivation tree. ASSUME simulates a hypothetical new binding target and re-evaluates, then rolls back. ABDUCE searches for the minimal evidence that would support a new indication.
What You'll See¶
- 28 novel drug-disease candidates ranked by combined pathway evidence (noisy-OR scores in [0, 1])
- Structural analogue matches linking candidates to known approved drugs via molecular-fingerprint similarity
- Full mechanistic derivation trees (
EXPLAIN RULE) for a prediction — reviewable by domain experts - What-if simulation (
ASSUME) of a new protein binding showing how hypothetical evidence changes rankings, then rolled back - A minimal-evidence search (
ABDUCE) for what graph changes would make a target indication appear
Why It Matters¶
The same analysis that traditionally requires a team, a sprint, and 500+ lines of custom code is expressed in a few declarative rules that a domain expert can read, audit, and extend. That changes repurposing from a software engineering project to a scientific reasoning exercise.