Pipeline Predictions

What the number means

Next to each drug is a single percentage: our estimate of the chance it eventually wins FDA approval. Read it as a frequency, not a promise. If a group of drugs each score around 20 percent, we expect roughly one in five of them to make it - but we cannot tell you which one.

Every score is conditional on where the drug is today. A Phase 3 score is the chance of approval from Phase 3 onward, not all the way back from the very first human trial. A drug naturally looks safer the further it has already traveled, so it would be unfair to compare a Phase 1 number directly against a Phase 3 number.

How the prediction is made

The estimate is built in two layers.

Layer one is the historical batting average. Decades of records tell us, for example, that a typical Phase 2 cancer drug is approved about 12 percent of the time. We start every drug at the published rate for its disease area and stage (the figures come from Wong et al. and related studies). On its own this starting point is honest but blunt: it gives every cancer drug in a given phase the same number, so it cannot tell two of them apart. Tested by itself, it is no better than a coin flip at picking the winner from the loser.

Layer two is a trained model that does the telling-apart. We took roughly 4,500 real drug programs whose fate is already settled - approved or failed - and let a statistical model learn how ten specific, public facts about each drug nudge its odds up or down. The model was not hand-tuned by us guessing what should matter; it read the historical record and worked out the weights itself.

A note on the facts that push down: more rigorous trials (randomized, blinded, with a comparator arm) are associated with lower approval in this historical data. This is a correlation the model found, not cause and effect. It most likely reflects that some drugs reach approval through smaller, simpler early studies, while the large confirmatory trials are exactly where many drugs fail.

How good is it, and how we know

We measure accuracy with a standard score called AUC. The idea is simple: hand the model one drug that was approved and one that failed, and ask how often it gives the approved one the higher score. A perfect model is right every time (AUC 1.00). A model with no skill is right half the time, like a coin flip (AUC 0.50). Higher is better.

Crucially, we measure this on drugs the model was not allowed to study while learning. That is the difference between memorizing answers and actually predicting. Here is how it does, stage by stage:

Two honest framings of these numbers. First, the historical batting average alone scores about 0.50 within a stage - a coin flip. The ten facts are what lift it to roughly 0.61 to 0.69, so essentially all of the real skill comes from them. Second, the scores are also well calibrated: when the model says 20 percent, drugs like that are approved about 20 percent of the time. You can see the full breakdown, including the calibration charts, on the validation page.

And to be clear about where this stands: this is version 1 of the model. It is an honest, working baseline, not a finished product. We are actively working to improve it - better data quality, richer inputs, and a sharper Phase 3 are all on the roadmap. Expect these numbers to get better over time.

What it cannot do

• These are estimates, not verdicts. The model is good on average, but it will be visibly wrong about individual drugs. Treat any single number as a starting point for your own research, never the last word.
• Phase 3 is the hardest. Its accuracy (0.61) is the lowest of the three, partly because there are fewer resolved Phase 3 programs to learn from and the outcomes are noisier.
• It learns from the past to judge the present. The model was trained on drugs that have already finished. If how trials are run keeps changing, its lessons may fit today's drugs slightly less well.
• A few drugs fall back to the batting average. About one in twenty pipeline drugs has no clearly linked trial record, so for those we show the historical base rate alone and say so.

The predictions

484 late-stage drugs (Phase 2 and Phase 3), ranked by estimated probability of approval. Click a drug for its full writeup.

Estimates are conditional on the drug's current stage - the chance of approval from here, not from the first human trial. A handful of drugs with no linked trial record fall back to the historical base rate for their disease area and stage.

Model is the model version that produced the estimate; Predictedis the date that estimate was made, at the drug's stage on that date. When a drug advances a phase we log a new, higher, dated estimate (the older one is kept in the drug's history) - so the date moves as a drug progresses. We grade ourselves only on the final outcome (approved or not), against the estimate we made before the decision; see the track record.