How AI Increases the Odds a Failed Drug Becomes a Winner

Most drug candidates die years before anyone files a patent on them, invisible to the outside world. AI is starting to mine that graveyard, though no single unpatented failure has yet been turned into a headline win.

Prepared by Richstorm.co

Key Takeaways

• Patents are typically filed just before Phase I trials begin, so a compound that fails earlier, in discovery or preclinical screening, was very likely never patented and remains invisible outside the company that made it.

• That unpatented graveyard is enormous: of every 10,000 compounds entering preclinical testing, only about 250 ever reach Phase I, meaning roughly 97.5% never make it to the point where a patent would typically be filed.

• AI is starting to mine that graveyard directly — re-scoring old rejected small molecules against modern prediction models, and for biologics, computationally redesigning failed protein and antibody structures for stability and manufacturability.

• No single named, unpatented failure has yet been publicly documented turning into an approved drug, in either small molecules or biologics — the mechanisms are real and published, but the field is too early for a clean before-and-after case study.

• AI-driven preclinical screening is reported to raise the transition rate from preclinical into Phase I from about 69% to over 75%, the closest available evidence that AI improves a pre-patent compound's odds of surviving into patented, clinical status.

The Graveyard Nobody Sees

Patent applications are typically filed just before a Phase I trial begins, timed that way specifically so the trial itself doesn't create a public-disclosure problem for the patent. That means the real dividing line between "private" and "public" isn't clinical failure versus preclinical failure — it's earlier than that. A compound only gets patented once a company commits it to the studies that lead into Phase I. Most compounds never make it that far: of every 10,000 compounds entering preclinical testing, only about 250 ever reach Phase I. The other 97.5% die in discovery or preclinical screening, before anyone ever filed a patent on them — no public record, no patent filing, nothing outside the company's own internal files.

This piece is scoped to exactly that population. A drug that reached Phase I and failed later is a different story — its structure was already disclosed through the patent, even if the trial result itself was never published. That's a partly-public case, and it's not what's covered here. What follows is specifically about compounds that died early enough to leave no public trace at all.

A companion RichStorm piece covers how pharma protects its data when working with an AI company. The same private-data mechanism described there — the reason AI reinforces certain moats instead of eroding them — is what makes this population valuable. A new entrant working only from public data cannot see any of it. A company with decades of internal discovery-stage failures can mine it directly.

How AI Mines It

Two distinct mechanisms are showing up in the pre-patent population, one for small molecules and one for biologics.

Table: How AI revisits compounds that failed before a patent was ever filed on them.

All three mechanisms work on the same population: candidates that failed early enough that no patent, and often no public record of any kind, ever existed for them.

One honest gap, in both modalities: none of these mechanisms yet has a named, publicly documented success story — a specific unpatented failure that was revived and later approved. The methods themselves are real and peer-reviewed. The finished proof point isn't public yet, in small molecules or biologics.

Does AI Actually Move the Needle?

Widely-cited industry statistics on drug repurposing and AI-selected candidate success rates almost all describe compounds that already reached clinical trials, meaning they were already patented. Those numbers sit outside the scope of this piece and are worth flagging only as general context for AI's broader effect on the industry, not as evidence for the specific pre-patent mechanism here.

The one number that does span the pre-patent boundary directly is the preclinical-to-Phase I transition rate. AI-driven computational validation performed before lab testing is reported to raise that rate from about 69% to over 75% — the closest available evidence that AI is improving a compound's odds of surviving out of the unpatented population and into patented, clinical status.

Table: The clearest available evidence on AI's effect specifically within the unpatented population.

The Economics of an Invisible Failure

Preclinical, discovery-stage R&D is estimated to cost roughly $40-120 million per eventual approved drug in out-of-pocket spending, based on separate industry cost studies that use somewhat different methodologies, so treat this as a rough order of magnitude rather than a precise figure. That spending chases thousands of compounds, nearly all of which die before ever being patented. None of it shows up in a patent filing or on ClinicalTrials.gov — it's sunk cost that's invisible to anyone outside the company that spent it, which is exactly why a company's own internal archive of these failures is worth something no outside competitor or AI-native startup can replicate.

Bottom Line

A companion piece on RichStorm covers what actually protects pharma's data when it partners with an AI company, the trust and governance side of this same private-data advantage.