Lumera

About the project

What inspired us

This project started with a simple frustration: after a rare disease diagnosis, families and clinicians often spend weeks navigating scattered papers, dense clinical language, and unclear safety details. Even when promising ideas exist, it’s hard to connect mechanism, evidence, and real-world risk into a decision-ready view. We wanted to turn “searching” into “understanding.”

What we built

We built a mechanism-aware rare disease discovery platform that produces ranked drug repurposing candidates and explains every recommendation with an evidence ledger. Users can specify disease biology (Loss of Function, Gain of Function, Dominant Negative), and the system enforces biological logic when ranking candidates. Each drug includes an uncertainty score (what’s known vs. missing) and a safety panel that highlights warnings, contraindications, pediatric risks, and key interactions.

For diseases with limited or no repurposing options, we added a molecular discovery pathway using a single-cell augmentation pipeline (scGFT) to improve target discovery and map high-confidence targets to existing FDA-approved drugs, generating testable hypotheses.

How we built it

Our system is organized into three layers:

1. Evidence ingestion and structuring
   We aggregate and normalize signals from biomedical literature, clinical references, and drug label sources into a structured representation:
   disease → mechanism → pathways/targets → drug actions → safety constraints.

2. Scoring and ranking
   Each candidate drug is scored across multiple dimensions:

   • Mechanism alignment
     
   • Evidence strength/quality
     
   • Clinical precedent
     
   • Safety risk and contraindications
     
   • Uncertainty and missing-evidence impact
     

We use a weighted scoring model (tunable by users and track requirements) to generate a transparent match score:

[ S(d) = \sum_{i=1}^{n} w_i \cdot f_i(d) ]

where ( f_i(d) ) are normalized feature scores for drug ( d ), and ( w_i ) are weights controlling importance (for example, safety can be weighted higher for pediatric use).

1. Uncertainty and safety-first outputs
   Instead of hiding ambiguity, we surface it: every recommendation includes what would most change the ranking if new evidence appeared. Safety checks run before the final output so risky options are flagged early, not after the fact.

Validation we performed

We back-tested on known rare-disease therapies to check whether our ranking pipeline surfaced established treatments near the top. We also red-teamed the safety module by cross-referencing our warnings and contraindication flags against official FDA labeling language for selected drugs to ensure the safety surfacing matched authoritative sources.

Challenges we faced

• Evidence is fragmented and inconsistent: Different papers describe the same mechanism in different terms, and clinical context is often missing. Normalizing this into a consistent structure was harder than expected.
  
• Mechanism enforcement is non-trivial: Mapping a disease mechanism (LoF/GoF/DN) to drug directionality requires careful logic to avoid recommending the right drug for the wrong reason.
  
• Safety data is dense: Safety signals live in long labels, and extracting what matters most (especially pediatric risks and severe warnings) without oversimplifying required iterative refinement.
  
• Balancing rigor with usability: The biggest design tension was making outputs scientifically grounded while still readable for non-experts.

What we learned

• Transparency matters as much as accuracy and users trust recommendations when they can see the reasoning and the limits.
  
• Safety can’t be an afterthought; it has to shape ranking, not just appear as a warning.
  
• Rare disease research needs tooling that respects uncertainty and small cohort reality while still enabling action: generating hypotheses, prioritizing follow ups, and guiding safer conversations between families and clinicians.

Built With

• javascript
• r