Drug discovery has always been one of the slowest, most expensive endeavors in modern science.
It costs, on average, more than $2 billion to bring a single drug to market. The process takes ten to fifteen years from initial discovery to pharmacy shelf. And despite all that investment, roughly nine out of every ten drug candidates that enter human trials will fail before they ever reach patients.
That has been the baseline reality of pharmaceutical R&D for decades. AI companies have been promising for years that they can change it.
In 2025, for the first time, there was real clinical evidence that they might be right.
A drug conceived entirely by artificial intelligence — not just screened or optimized using AI, but discovered and designed by AI from target identification through molecular design — completed a Phase IIa randomized, double-blind, placebo-controlled clinical trial and showed measurable improvement in lung function in patients with a fatal disease that has no cure.
That result, published in Nature Medicine in June 2025, is the most significant milestone the field has produced. It is also a starting point, not a finish line.
No AI-discovered drug has yet received FDA approval. The pharmaceutical industry's 90 percent clinical failure rate remains stubbornly unchanged. And the question the entire industry is now watching — whether AI can actually improve clinical success rates, not just accelerate the early stages of discovery — is still open.
That answer is coming. Multiple pivotal Phase III trials are underway in 2026. The results will either validate a decade of investment and ambition, or force a fundamental rethinking of what AI can and cannot do in medicine.
This is where things stand.
The Breakthrough That Actually Happened
Rentosertib: The First End-to-End AI Drug
The drug at the center of this story is rentosertib, formerly known as ISM001-055, developed by Insilico Medicine.
It is described as the first drug in which both the disease target and the molecular compound were discovered using generative AI — making it a true end-to-end AI-designed therapeutic.
The disease it targets is idiopathic pulmonary fibrosis (IPF), a chronic, progressive lung condition marked by irreversible scarring of lung tissue. IPF leads to a steady decline in lung function, breathlessness, persistent coughing, and reduced oxygen intake.
A few facts about IPF that put the stakes in context:
- Affects an estimated 5 million people globally
- Median survival of just 3 to 4 years after diagnosis
- Current treatments can slow progression, but none can halt or reverse it
Insilico used its generative AI platform, Pharma.AI, to identify TNIK (Traf2- and NCK-interacting kinase) as a novel target in IPF. This is a first-in-class target — no drug had ever been designed to inhibit it before. The AI then designed the molecule itself.
What makes the discovery particularly significant is the timeline.
Traditional target identification and lead optimization typically takes four to six years. Insilico identified a novel IPF target and advanced a drug candidate into preclinical trials in just 18 months, at a cost of only $150,000, excluding wet lab validation.
What the Phase IIa Trial Actually Showed
The trial (NCT05938920) was a double-blind, placebo-controlled study that enrolled 71 patients with IPF across 21 sites in China. Patients were randomized to receive either placebo, 30 mg once daily, 30 mg twice daily, or 60 mg once daily for twelve weeks.
Safety: The drug passed the primary endpoint. Drug-related adverse events were predominantly mild or moderate. The most common were diarrhea (14.8%) and abnormal liver function (14.8%).
Efficacy: This is where the data got the attention of the pulmonology community.
| Group | Mean Change in Forced Vital Capacity |
|---|---|
| Placebo | -62.3 mL (decline) |
| 60 mg once daily | +98.4 mL (improvement) |
For context: IPF patients almost never see improvements in forced vital capacity. The disease consistently moves in one direction. Measurable lung function gains over twelve weeks is the kind of signal that gets researchers paying close attention.
The study was presented at the American Thoracic Society 2025 International Conference and simultaneously published in Nature Medicine (impact factor: 58.7). Insilico has begun discussions with regulatory authorities to facilitate evaluation of rentosertib in larger patient cohorts.
A separate Phase IIa trial enrolling U.S. patients is ongoing.
The caveat: The trial enrolled only 71 patients. Efficacy signals require validation in much larger studies before they translate to clinical reality. The Phase IIa data is promising, not conclusive.
The Honest State of the Field
What AI Has Actually Changed
2025 delivered the field's first genuine clinical validation — and simultaneously exposed the limits of what AI can do at this stage.
Where AI has delivered real, documented gains:
- Early discovery timelines compressed by 30 to 40 percent
- Preclinical candidate development reduced to 13 to 18 months (vs. the traditional 3 to 4 years)
- Virtual screening of billions of compounds before any lab synthesis begins
- Target identification that reaches biology humans would have missed
Where the gains stop:
Clinical trial duration, regulatory review timelines, and manufacturing scale-up remain bound by biology, patient enrollment, and regulatory requirements. None of those are problems an algorithm can bypass.
The 90 Percent Failure Rate Is Still There
The most important number in pharmaceutical R&D is one that AI has not yet moved.
Fewer than one in ten drug candidates entering Phase I clinical trials are ultimately approved. That rate has persisted for decades, driven by the fundamental difficulty of predicting how a molecule designed in a lab will behave inside the complex, variable biology of the human body.
Here is where the clinical data stands:
| Trial Phase | AI-Discovered Drugs | Traditional Drugs |
|---|---|---|
| Phase I success rate | 80–90% | 40–65% |
| Phase II success rate | ~40% | ~30–45% |
Phase I results are genuinely encouraging. AI appears to be better at designing molecules with favorable safety and drug-like properties.
Phase II is where the story gets more complicated. This is where drugs have to demonstrate they actually work in sick patients — and that is a test that human biology, not algorithm quality, ultimately decides.
Commentary from scientific publications noted that AI-discovered compounds show clinical progression rates similar to traditionally discovered compounds. One CEO's blunt assessment: "AI has really let us all down in the last decade when it comes to drug discovery — we've just seen failure after failure."
The honest framing: AI has made the early part of drug development faster and cheaper. Whether it can make the clinical part more successful is the question 2026 is beginning to answer.
The Companies Shaping the Field
Insilico Medicine
Insilico is currently the furthest along with a fully end-to-end AI-designed therapeutic.
Rentosertib represents the company's core thesis: that generative AI can identify novel disease targets and design the drug to hit them, faster and cheaper than traditional methods. The Phase IIa results have put Insilico at the center of the field's attention. A separate Phase IIa trial in U.S. patients is ongoing and actively enrolling.
Recursion Pharmaceuticals and Exscientia
The Recursion-Exscientia merger combined two of the most prominent AI drug discovery companies, integrating Recursion's high-throughput biological imaging with Exscientia's precision molecular design tools.
Key pipeline readouts expected in 2026:
- REC-394 (C. difficile infection): Phase 2 update expected Q1 2026
- REC-1245 (solid tumors and lymphoma): Phase 1 dose-escalation data expected H1 2026
The merger came with complications. Several months after the deal closed, Recursion reworked its pipeline, deprioritizing several candidates. Integrating two organizations with different methodologies is organizationally difficult, and financial pressure to produce clinical results is significant.
Isomorphic Labs
Backed by Google's DeepMind and built on the same research that produced AlphaFold, Isomorphic Labs has attracted some of the largest pharma partnership deals in the sector.
Notable partnerships:
- Eli Lilly: $1.75 billion agreement
- Novartis: Doubled its commitment in early 2025 to pursue previously inaccessible targets
- Novo Nordisk / Valo Health: $2.76 billion deal (separate but comparable in scale)
The technical foundation is AlphaFold 3, which can now model dynamic interactions between proteins, DNA, RNA, and ligands — a capability critical for designing molecules targeting previously inaccessible biology. As of early 2026, partnerships had already transitioned from initial target identification to the generation of multiple preclinical candidates.
Isomorphic's drugs are not yet in clinical trials, but the scale of its pharma partnerships and the quality of its underlying technology mean it will be a significant presence in clinical readouts over the next two to three years.
Schrödinger
Schrödinger's physics-based AI approach — combining molecular simulation with machine learning — has produced its most advanced validation yet.
The tyrosine kinase 2 inhibitor zasocitinib (TAK-279), originally from Nimbus and developed using Schrödinger's platform, has now entered Phase III clinical trials through a partnership with Takeda. That is the furthest along any AI-assisted drug has progressed in terms of trial phase.
What AI Is Actually Doing in the Drug Discovery Process
Understanding why the field's results are mixed requires understanding where in the pipeline AI is delivering real value — and where it runs into hard limits.
Target Identification: Where AI Shines
Mining biological datasets, analyzing genetic associations, identifying proteins implicated in disease pathways — these are tasks where machine learning models can process vastly more information than any human research team.
The quality of target selection directly affects everything downstream. Insilico's identification of TNIK as a relevant target in IPF is a direct example of this capability producing a real clinical candidate.
Molecular Design: Promising but Data-Limited
Generative AI models can now design novel molecules with specified properties — binding affinity, solubility, metabolic stability — in ways that were not possible five years ago.
The challenge: these predictions are based on training data that captures what has been tested before. Novel targets in novel diseases have less historical data to work with, which limits the model's confidence.
Predicting Clinical Behavior: The Hard Wall
This is where the difficulty concentrates.
A molecule that looks excellent in a computational model and performs well in cell-based assays and animal studies still has to work inside a human being — where the variability of disease expression, genetic differences, comorbidities, concurrent medications, and the sheer complexity of human biology can produce outcomes that no model predicted.
A survey of tech executives found 68 percent cite poor data quality and governance as the main reason AI initiatives fail. High-quality, rigorously curated datasets with biological, pharmacological, and clinical annotations remain scarce due to costs, privacy regulations, and data-sharing restrictions.
The fundamental challenge is not algorithmic sophistication. It is data quality and biological unpredictability — two problems that better AI does not automatically solve.
The Regulatory Picture Is Finally Catching Up
For years, one of the significant uncertainties for companies developing AI-discovered drugs was the absence of clear regulatory guidance. That is beginning to change.
January 2025: The FDA published its first-ever draft guidance on AI in drug development — "Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products." The framework involves a seven-step risk-based credibility assessment process, from defining the question of interest through documenting validation results.
One important carve-out: The guidance explicitly excludes AI used in drug discovery itself. It applies to AI that supports regulatory decision-making during development — not to AI that was used to find the drug in the first place. That distinction matters for how companies structure regulatory submissions.
January 2026: The FDA and the European Medicines Agency jointly introduced ten Guiding Principles for Good AI Practice in Drug Development — a non-binding but significant step toward international regulatory alignment. These principles are expected to form the basis of more detailed, potentially binding requirements as both agencies develop their frameworks.
Final FDA guidance is expected Q2 2026. That clarity will matter significantly for investment decisions and submission strategies across the sector.
The Investment Landscape
The capital flowing into AI drug discovery is substantial and still growing, despite mixed clinical results.
Market size:
- 2025: $1.94 billion
- 2026 (projected): $2.6 billion
- 2034 (projected): $16.49 billion at a 27% CAGR
Major recent deals:
- Eli Lilly + Isomorphic Labs: $1.75 billion
- Novo Nordisk + Valo Health: $2.76 billion
- Bayer + Recursion: $1.5 billion
- Isomorphic Labs standalone raise: $600 million
One development worth watching: Chinese AI drug discovery companies have significantly increased their share of global biotech licensing deals, growing from 21 percent in 2023-2024 to 32 percent in Q1 2025. Insilico Medicine, which produced the rentosertib results, is headquartered in Cambridge, Massachusetts, but conducts significant research and clinical operations in China.
What 2026 Will Actually Decide
The most consequential development of 2026 will be Phase III results that determine whether AI can deliver drugs that work at scale.
Multiple AI-designed drugs are entering pivotal trials this year, with readouts expected throughout 2026. These results will provide the first large-scale test of whether AI improves clinical success rates beyond the pharmaceutical industry's persistent 90 percent failure rate.
The balanced forecast is not uniformly optimistic. Additional clinical failures are statistically inevitable given historical attrition rates. Experts predict validation and disappointment in roughly equal measure.
That is not a criticism of AI drug discovery. It is an accurate description of pharmaceutical development, which has always produced more failures than successes even for approaches that eventually prove transformative.
What has genuinely changed is the speed and cost at which those attempts can be made. If AI can reduce the time from target identification to clinical candidate from four years to eighteen months, and reduce the cost of that phase by 30 to 70 percent, then more attempts can be made in a given period. More attempts means more failures in absolute terms. It may also mean more eventual successes.
The rentosertib results established that AI-designed drugs can show clinical efficacy in human patients. That proof of concept, small and preliminary as it is, did not exist before June 2025.
The question now is whether it generalizes — whether AI-first drug discovery can produce not just one promising compound for one disease, but a consistent improvement in the odds that any given drug candidate makes it through the brutally selective process of clinical development.
The most important question for 2026 is not whether AI can accelerate preclinical timelines. We know it can. The question is whether it can improve clinical success rates.
For the patients waiting for treatments that do not yet exist, the answer to that question matters more than the investment returns, the merger valuations, or the competition between platforms. That is the real weight of what is being tested in these trials.
FAQ
What is an AI-designed drug and how is it different from a traditionally developed drug?
An AI-designed drug is one where artificial intelligence played a significant role in discovering the disease target, designing the molecular compound, or both. Traditional drug discovery relies on large-scale laboratory screening of existing compound libraries and years of manual medicinal chemistry. AI approaches use machine learning and generative models to identify promising targets in biological datasets and design novel molecules computationally before synthesis begins. The end product goes through the same clinical trial process as any other drug, but the early discovery phase is faster and, in some cases, reaches targets that traditional methods would not have identified.
Has an AI-designed drug been approved by the FDA?
No AI-discovered drug had achieved FDA approval as of December 2025. The field reached Phase IIa clinical validation for the first time with rentosertib, and multiple compounds are now in Phase III trials with pivotal readouts expected in 2026. FDA approval for the most advanced AI-designed candidates is estimated at 2026 to 2027, though this depends on Phase III outcomes and regulatory review timelines.
What was the significance of the rentosertib Phase IIa results published in Nature Medicine?
The rentosertib trial was the first time a drug with both its target and molecule discovered by generative AI completed a Phase IIa randomized, double-blind, placebo-controlled trial and showed measurable clinical efficacy. Patients receiving the highest dose experienced an average improvement in forced vital capacity of 98.4 mL, compared to an average decline of 62.3 mL in the placebo group. For a disease where lung function consistently deteriorates, an improvement signal over twelve weeks is clinically notable. The study enrolled 71 patients, so larger validation trials are needed before conclusions can be drawn at population scale.
Do AI-designed drugs have better success rates in clinical trials?
In Phase I trials, which test primarily for safety and tolerability, AI-discovered molecules show success rates of 80 to 90 percent, compared to 40 to 65 percent for traditionally discovered drugs. That is a meaningful difference and suggests AI is effective at designing molecules with favorable safety profiles. In Phase II, where drugs have to demonstrate efficacy in sick patients, AI-discovered compounds show success rates of approximately 40 percent — comparable to the historical industry average. Whether AI can improve Phase II and Phase III success rates remains the central unanswered question in the field.
Which companies are leading in AI drug discovery?
Leading companies include Insilico Medicine, which has rentosertib in Phase IIa with U.S. trials ongoing; the combined Recursion-Exscientia entity, with multiple programs approaching Phase II readouts in 2026; Schrödinger, whose physics-based AI approach has a compound in Phase III through a partnership with Takeda; and Isomorphic Labs, backed by Google DeepMind, with major partnerships with Eli Lilly and Novartis, though its compounds are not yet in clinical trials.
What is the FDA doing to regulate AI in drug development?
The FDA published its first draft guidance on AI in drug development in January 2025, establishing a risk-based credibility assessment framework for companies using AI to support regulatory decision-making on drug safety, efficacy, and quality. Final guidance is expected in Q2 2026. In January 2026, the FDA and the European Medicines Agency jointly released non-binding guiding principles for good AI practice in drug development. Notably, the guidance does not cover AI used in the drug discovery phase itself — only AI used in development and regulatory decision-making.
How does AI reduce the time and cost of drug discovery?
AI reduces development time primarily by accelerating target identification, molecular design, and lead optimization — processes that traditionally take years but can now be completed in months. AI also enables virtual screening of billions of potential compounds before any laboratory synthesis occurs, reducing dead-end experiments and wet lab costs. However, clinical trials, regulatory review, and manufacturing still follow standard timelines that AI does not meaningfully compress.
What are the biggest remaining challenges for AI drug discovery?
The three most significant challenges are data quality, biological unpredictability, and clinical validation. AI models are only as good as the data they are trained on, and high-quality annotated biological datasets remain scarce. No computational model can fully predict how a drug will behave in the complex biology of human patients, which is why clinical failure rates remain high. And until multiple AI-designed drugs have demonstrated Phase III efficacy and received regulatory approval, clinical validation remains limited to a single Phase IIa result in a small patient cohort.
Related Articles
Mark
Dani
Mark
