The Repurposing Playbook: Accelerating Therapy via Real-World Evidence

The Myth of the New Molecule

We have been lied to. For decades, the industry gospel claimed that true innovation only happens when we synthesize a brand-new molecule from scratch. It is a costly, slow, and often futile gamble. Most new drugs fail. We spend a decade and a billion dollars only to find the molecule is toxic or useless in humans. Why do we keep playing this game when the solutions are already sitting on pharmacy shelves? The real edge today isn't in discovery; it is in repurposing. By using drugs already approved for one condition to treat another, we bypass the early-stage safety nightmares that kill most pipelines.

Prerequisites: What You Will Need

You cannot repurpose drugs using a spreadsheet and a prayer. The terrain has shifted toward high-throughput causal inference. To execute this playbook, you need a specific stack of data and computational tools that can handle the noise of real-world evidence without compromising patient privacy.

• Access to decentralized health learning systems (e.g., INSIGHT or OneFlorida+).
• A federated machine learning framework to analyze data without moving it from its source.
• Causal inference algorithms capable of addressing heterogeneity in treatment assignment.
• A validation network for external cross-referencing (e.g., TriNetX global research network).
• A structured database of existing approvals and indications (e.g., OncoAID).

The Repurposing Playbook: Step-by-Step

1. Identify the therapeutic gap (e.g., Alzheimer's disease or rare cancers).
2. Aggregate decentralized real-world data across multiple health learning systems.
3. Deploy causal machine learning to scan candidate drugs for therapeutic signals.
4. Validate findings using a separate, global research network to ensure reproducibility.
5. Transition validated candidates into targeted clinical trials for the new indication.

The first two steps are about scale. You don't look at one hospital; you look at millions of patients. For example, the FLAT-WHITE algorithm recently scanned 181 candidate drugs for Alzheimer’s disease and related dementias (ADRD) by leveraging four national health learning systems, including INSIGHT (over 5.5 million patients) and OneFlorida+ (over 24 million patients) (Source: npj Health Systems, 2026). This isn't a small sample. It is a massive, real-world net that catches signals a controlled trial would miss.

Step three is where the math gets gritty. You cannot simply look for correlations; you need causal inference. The FLAT-WHITE approach specifically addresses the heterogeneity in how treatments are assigned in the real world, ensuring that the signal you see is actually caused by the drug and not by some hidden patient characteristic (Source: npj Health Systems, 2026). It is the difference between seeing that people who take a drug get better and proving that the drug made them better.

Validation is the final guardrail. A signal in one dataset is a hint; a signal across multiple global networks is a lead. The ADRD study validated its findings in an additional 339,180 patients through the TriNetX global research network (Source: npj Health Systems, 2026). This external validation removes the local bias and prepares the drug for the final leap into a clinical trial.

Mining the Oncology Web: The OncoAID Approach

Cancer is where repurposing is currently winning. The OncoAID dataset provides a clear window into this. It catalogs 657 approval entries covering 375 unique treatments across 411 distinct indications (Source: WIN Symposium, 2026). We are seeing a hard turn toward combination regimens. Instead of searching for one magic bullet, practitioners are layering existing targeted therapies to create synergy.

The data shows a marked increase in approved targeted therapies over time (Source: WIN Symposium, 2026). The goal now is to determine the predictive accuracy of treatment selection tests for off-guideline drugs. This means we are no longer just following a textbook; we are using the data to find which existing drug works for which specific patient, even if it was never intended for that cancer type.

The Invisible Friction: Where the Math Hits the Wall

Here is what the academic papers don't tell you. The math is the easy part. The real battle is fought in the corridors of insurance companies and the minds of conservative clinicians. Even when the data shows an off-guideline drug works, you hit a wall of 'physician acceptance' and 'insurance barriers' (Source: WIN Symposium, 2026). Doctors are trained to follow guidelines. When you suggest a drug approved for hypertension to treat a specific tumor, you aren't just fighting a disease; you are fighting a culture of risk-aversion.

Then there is the data friction. Every institution guards its patient data like a crown jewel. This is why federated learning is the only way forward. It allows us to query the data without ever seeing the raw patient records. The tension between data privacy and the urgency of discovery is the primary friction point in modern repurposing. If you cannot solve the trust issue, your algorithm is useless.

Common Pitfalls to Avoid

• Over-reliance on correlation: Mistaking a common co-prescription for a therapeutic effect.
• Ignoring treatment heterogeneity: Failing to account for why a doctor chose a specific drug for a specific patient.
• Underestimating insurance hurdles: Finding a drug that works but cannot be reimbursed for the new indication.
• Small-scale validation: Relying on a single health system's data, which leads to overfitting.

To avoid these traps, you must prioritize external validation. If your signal doesn't hold up in a global network like TriNetX, it is a ghost. Moreover, you must integrate operational factors—like tissue availability and testing accessibility—into your model. A drug that works in a lab but cannot be prescribed due to insurance barriers is a failure in the real world (Source: WIN Symposium, 2026).