↓ Skip to main content
Expert LinkedIn

AI Drug Discovery Solved the Easy Problem. Now Comes the Hard One.

#AI in Healthcare #Technology
9 min read
Dr. Sophia Patel
Dr. Sophia Patel AI in Healthcare Expert & Machine Learning Specialist

Last week, Eli Lilly wrote a $2.75 billion check to an AI drug discovery company. It may be the most consequential bet in pharmaceutical history—or the most expensive lesson that Phase I success does not equal Phase III proof.

On March 29, Lilly committed up to $2.75 billion to Insilico Medicine for access to its AI-discovered drug pipeline. The deal is structured as $115 million upfront with the remaining $2.63 billion tied to development, regulatory, and commercial milestones—which is actually significant, because it means Lilly is betting on outcomes that haven’t materialized yet. Insilico, the Hong Kong-listed AI biotech founded by Alex Zhavoronkov in 2014, has 28 drug candidates in its pipeline and nearly half already in clinical trials. Their platform can take a target from identification to preclinical candidate nomination in 12 to 18 months. Traditional pharma averages two and a half to four years for the same milestone.

This is genuinely impressive. It is also not what the industry thinks it is.

What AI Drug Discovery Has Actually Proven
#

Before I raise the cautionary flag, let me acknowledge what Insilico has actually demonstrated—because it is landmark.

Their lead compound, rentosertib (ISM001-055), is the first drug in history where an AI system identified both a novel disease target and designed the molecule to hit it—entirely from scratch. For idiopathic pulmonary fibrosis (IPF), a progressive lung disease that kills within two to four years of diagnosis, Insilico’s PandaOmics engine identified TNIK (Traf2- and Nck-interacting kinase) as a previously unrecognized regulator of fibrosis. Chemistry42, their generative molecular design platform, then created rentosertib to inhibit it. Target to Phase I in under 30 months. A Phase 2a trial completed and published in Nature Medicine in 2025 confirmed the drug was safe, well-tolerated, and showed biological activity in IPF patients who currently have no options beyond two drugs that merely slow deterioration.

That is real. The AI found something human scientists had missed and made a molecule to address it in a timeline that would have been considered impossible five years ago.

A drug development pipeline rendered as a narrowing funnel — wide and luminous at the top where AI-accelerated discovery pours in candidates, then dramatically constricting as the funnel reaches the third phase, where no AI drug has yet emerged.

And the numbers at Phase I are extraordinary. According to a December 2025 analysis by PitchBook, AI-native biotechs—companies using AI as a foundational technology, not just a productivity tool—are achieving 80 to 90 percent Phase I success rates, compared to an industry average of 40 to 65 percent. The sheer volume is matching the quality: as of early 2026, 173 AI-discovered drug programs are in clinical development, with roughly 94 in Phase I, 56 in Phase II, and 15 approaching Phase III.

The celebration in the industry is understandable. The $2.75 billion bet feels like confirmation.

Here Is What That Phase I Number Is Measuring
#

Phase I clinical trials answer one primary question: does this molecule kill people quickly? They enroll small numbers of relatively healthy patients—typically 20 to 80—and monitor safety, dosing, and pharmacokinetics. The bar for success is surviving the trial with an acceptable side-effect profile. There is no placebo comparison. There is no efficacy requirement.

AI is exceptionally good at predicting toxicity. Generative molecular design tools like Chemistry42 can screen for off-target binding, metabolic liabilities, and cardiac toxicity signals across billions of compounds in days. Molecules that make it through AI selection are cleaner—more selective, better characterized, less likely to cause harm before they cause benefit. That is valuable. That is what the 80 to 90 percent figure captures.

What it does not capture is whether the drug works.

The Phase II Reality Check
#

When you look at Phase II—where trials begin to measure efficacy in patients with the actual disease—the picture shifts. PitchBook’s same analysis shows AI-native biotechs achieving roughly a 40 percent Phase II success rate. The industry average is approximately 29 percent. AI is still better—genuinely, meaningfully better. But the improvement compressed from a nearly 2x Phase I advantage to a modest Phase II edge.

The authors of Insilico’s own Nature Medicine paper are more direct about the Phase II record: “AI-discovered drugs have experienced similar levels of phase 2 trial failure as non-AI-discovered drugs.” That is not a pessimistic interpretation of the data. That is the data.

And Phase III? As of the paper’s publication, no AI-discovered drug had completed Phase III trials. The 15 programs currently approaching pivotal trials will be the first generation to find out. The industry is placing billion-dollar bets before the most critical results are in.

Why Phase III Is Structurally Different—and Why AI Hasn’t Solved It
#

The assumption embedded in the Lilly-Insilico deal, and in most AI drug discovery coverage, is that Phase I and Phase II results are predictive of Phase III performance. Historically, they are not. The overall probability of a drug succeeding from Phase I to approval is approximately 8 percent. PitchBook projects that AI could double this to roughly 18 percent if current trends hold—a genuinely meaningful improvement that would transform pharmaceutical economics.

But “if current trends hold” is doing a great deal of work in that sentence.

Phase III trials test something qualitatively different from Phases I and II. They enroll thousands of patients across diverse populations—different ages, comorbidities, genetic backgrounds, geographic locations. They run for years, not months. They require head-to-head comparison against existing standard-of-care treatments, or against placebo in diseases where no treatment exists. And they must demonstrate clinical meaningfulness: not just that a biomarker moved, but that patients lived longer, breathed better, or suffered less.

This is where the limits of AI’s current architecture become apparent. AI systems trained on available biomedical datasets learn patterns in that data—but clinical trial datasets are famously incomplete, historically biased toward younger white male patients, and poorly representative of the populations where most disease burden lies. An AI that identified TNIK as a fibrosis regulator learned from the literature and omics data available to it. If that literature underrepresents a particular population’s fibrosis pathology, the molecule designed from it may underperform in Phase III’s diverse cohort.

I am not suggesting AI drug discovery is a dead end. I am suggesting the industry is celebrating a proof of concept at Phase I as if it constitutes a proof of Phase III viability—and those are genuinely different claims.

What I Think Lilly Is Actually Betting On
#

Here is the part of this story I think is being missed.

Lilly’s $2.75 billion commitment to Insilico is not primarily a bet on rentosertib. It is a bet on platform speed. If Insilico can generate 20 Phase I-ready drug candidates in 18 months each, and even if the overall success rate from Phase I to approval stays near current industry averages, Lilly gets dramatically more shots on goal per research dollar. The bet is not “AI solves Phase III.” The bet is “AI lets us attempt Phase III forty times more often than our competitors.”

That is a rational pharmaceutical strategy. More candidates at lower cost per candidate can beat a lower success rate per candidate on expected value grounds. The deal’s milestone structure—$115 million upfront versus $2.63 billion contingent on success—tells you exactly how confident Lilly actually is: they’re paying for pipeline optionality, not proven outcomes.

Insilico itself is signaling awareness of this gap. In a landmark paper published in ACS Central Science in February 2026, Insilico and Lilly researchers jointly described their vision of “Pharmaceutical Superintelligence”—a fully autonomous AI system that orchestrates drug discovery end-to-end from a single prompt. The paper is a vision, not a product announcement. And its authors are careful to note that current systems “still require safeguards to mitigate hallucinations, error propagation, and data-driven biases,” and that “human oversight for high-stakes decisions” remains essential. This is the same intellectual honesty I saw missing from much of the industry’s deal coverage.

What This Means for Clinical Practice
#

For those of us navigating the gap between research and patient care, the practical implications are real right now.

Patients are already asking about AI-discovered drugs. Longevity practitioners are reporting that patients arrive citing AI-discovered compounds—senolytics, fibrosis inhibitors, longevity-targeting molecules—heard about on podcasts and social media. The discovery pipeline is moving faster than clinical guidance can track. “I had not heard of it,” one longevity clinician admitted this week. That is going to become a common experience.

The framework I’d offer: Phase I completion means a molecule is probably safe in the short term. Phase II results signal whether there’s a biological effect worth pursuing. Neither is sufficient clinical evidence for prescribing decisions. The phase that actually answers the patient’s question—does this drug improve outcomes in people like me?—is Phase III, and the AI drug discovery revolution hasn’t cleared it yet.

This week’s NIH budget proposal makes this worse. The Trump administration’s request to cut $5 billion from NIH’s 2027 budget includes eliminating the National Institute on Minority Health and Health Disparities entirely. For AI drug discovery to prove its Phase III potential, we need trial populations that actually reflect disease demographics. Defunding minority health research while celebrating AI drug discovery’s billion-dollar deals is a structural contradiction that nobody in the deal announcements is discussing.

The Real Revolution Is Still Coming
#

I am not arguing against AI drug discovery. Rentosertib—the first fully AI-discovered and AI-designed molecule to complete a Phase 2a trial—is a genuine scientific achievement, and IPF patients desperately need better options. The 173 programs in clinical development represent more candidate diversity than any prior generation of pharmaceutical research. The platform speed advantage is real and will compound.

But the revolution’s proof moment is approximately 18 to 24 months away, when the first wave of AI-discovered drugs either clears or stumbles in Phase III. That outcome—not the Lilly deal, not the Phase I success rates, not the ACS Central Science vision paper—will determine whether this is a genuine disruption to pharmaceutical development or a very expensive and valuable acceleration of the first third of a pipeline that still fails most of its candidates.

The $2.75 billion bet is a bet on the former. I hope it is correct. I am watching the Phase III data with both genuine optimism and scientific caution—which is, I think, exactly the right posture for anyone who has spent time at the intersection of AI’s promises and medicine’s realities.

References
#

AI-Generated Content Notice

This article was created using artificial intelligence technology. While we strive for accuracy and provide valuable insights, readers should independently verify information and use their own judgment when making business decisions. The content may not reflect real-time market conditions or personal circumstances.

Related Articles

Tags

#AI drug discovery #clinical trials #Insilico Medicine #Eli Lilly #pharma AI