Somewhere between "this drug might work" and "let's spend a billion dollars finding out," oncology researchers have been playing the world's most expensive guessing game. Only about 5% of cancer drugs that enter clinical trials actually make it to patients - which means 19 out of 20 promising compounds end up as very expensive lessons in humility.

A team of researchers just published something that might change how we play those odds. They call it INSPIRE, and it's basically a matchmaking algorithm for cancer treatments.

The Problem: Drug Development is a Brutal Numbers Game

Here's the uncomfortable truth about developing cancer drugs: it costs somewhere between $173 million and $2.6 billion per drug, takes about 12 years on average, and most of those drugs fail. Not some of them. Most of them. The overall clinical trial failure rate hovers around 90%, with lack of efficacy being the top culprit in 40-50% of cases.

When pharmaceutical companies pick which cancer types to test their shiny new drug on, they're often working with incomplete information, educated guesses, and whatever tumor samples happened to be available. It's like trying to find someone's soulmate by showing their photo to random people at a bus station.

INSPIRE: Teaching Machines to Read Medical Histories

The INSPIRE approach does something clever. Instead of relying purely on laboratory data about tumor biology, it learns from the messy, complicated reality of actual patient medical records. The algorithm creates mathematical representations - embeddings, if you want the technical term - based on events throughout a patient's medical journey.

Think of it like this: instead of describing a person by their height and eye color, you describe them by their entire life story compressed into a numerical fingerprint. The researchers fed their model information about cancer patients, including their diagnoses, treatments, and outcomes, then trained it to recognize patterns that predict whether certain immunotherapy drugs would work.

They focused on PD-1 inhibitors - drugs like pembrolizumab (Keytruda) and nivolumab (Opdivo) that have revolutionized cancer treatment since their first FDA approval in 2014. These drugs work by essentially taking the brakes off your immune system so it can attack tumors more effectively.

The Plot Twist: Testing the Oracle

Here's where it gets interesting. The researchers didn't just train their model on current data and call it a day. They performed what's essentially a time travel experiment.

They trained INSPIRE using only data from before PD-1 inhibitors were widely established in clinical practice. Then they asked: could the model have predicted which cancer types would eventually get FDA approval for these drugs?

The answer: 70% of the time, yes.

That's a remarkably high hit rate for a problem this complicated. The model successfully prioritized most of the indications that later received regulatory approval - before the clinical trials even happened.

Why This Actually Matters

Picking the wrong indication for a clinical trial isn't just embarrassing - it's catastrophic. Each failed Phase III trial represents years of time and hundreds of millions of dollars, not to mention the opportunity cost of not pursuing treatments that might have worked elsewhere.

If a tool like INSPIRE could help pharmaceutical companies prioritize their bets more intelligently, even modest improvements in success rates would translate to more drugs reaching patients faster. And in oncology, faster often means the difference between life and death.

The approach also sidesteps one of the persistent problems in drug development: our tendency to stick with tumor types we already understand well. Real-world data captures patterns that controlled laboratory experiments might miss entirely.

The Caveats (Because There Are Always Caveats)

This is a retrospective validation study, which means the model was tested on historical data where we already know the answers. The real test will be whether INSPIRE can prospectively guide decisions about novel drugs that haven't been tested yet.

There's also the persistent challenge of AI "black boxes" - regulatory agencies are understandably nervous about approving drugs based on models that can't explain their reasoning. The FDA has been ramping up its AI initiatives, but transparency remains a major concern.

The Bigger Picture

INSPIRE represents part of a broader trend toward using real-world data and causal machine learning to make drug development less of a casino and more of a science. Patient embeddings and representation learning have been gaining traction across healthcare, from predicting disease progression to identifying patients who might respond to specific treatments.

The core insight isn't revolutionary: patients who share similar medical journeys might respond similarly to treatments. What's new is having the computational tools to actually learn those patterns from millions of data points.

If tools like this become standard practice, the next generation of cancer drugs might find their ideal patients years earlier than they would have otherwise. And in a field where only 5% of candidates survive clinical development, improving those odds even slightly would be worth celebrating.

References

1. Eckhoff, M., Klingelschmitt, S., Van Ruijssevelt, L., et al. (2026). Finding the most promising indications for novel treatments in oncology. npj Precision Oncology. DOI: 10.1038/s41698-026-01352-x

2. Sun, D., Gao, W., Hu, H., & Zhou, S. (2022). Why 90% of clinical drug development fails and how to improve it? Acta Pharmaceutica Sinica B, 12(7), 3049-3062. PMCID: PMC9293739

3. FDA Oncology Center of Excellence. (2025). OCE Oncology Artificial Intelligence Program. Retrieved from FDA.gov

4. Papp, D. & Gálfi, L. (2025). Real-World Data and Causal Machine Learning to Enhance Drug Development. Therapeutic Innovation & Regulatory Science. PMCID: PMC12579681

5. Steinberg, E., Belthangady, C., et al. (2024). Language-model-based patient embedding using electronic health records facilitates phenotyping, disease forecasting, and progression analysis. npj Digital Medicine. PMCID: PMC11469380

6. Darvin, P., Toor, S. M., Sasidharan Nair, V., & Elkord, E. (2018). Immune checkpoint inhibitors: recent progress and potential biomarkers. Experimental & Molecular Medicine, 50, 1-11. PMCID: PMC5778665

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.