173 AI-Designed Drugs Are Now in Human Trials. The First One Just Reversed Lung Damage Nobody Expected to Fix.
Over 173 drug candidates discovered or designed by artificial intelligence are in clinical trials as of early 2026, with 15 to 20 entering pivotal Phase III studies this year. Rentosertib, the first compound where AI identified both the disease target and the molecule without any human hypothesis, showed a 160-milliliter forced vital capacity difference over placebo in patients with idiopathic pulmonary fibrosis. An original pipeline probability analysis shows why the question is no longer whether an AI drug gets approved, but how many and how fast.
One hundred and sixty milliliters. That is how much the lungs of patients taking a 60-milligram dose of rentosertib diverged from those receiving a sugar pill over the course of a single clinical trial, a gap that would be unremarkable for a cold medicine but is almost disorienting for a disease where every treatment ever tested had merely slowed the rate of suffocation. In the treatment group, forced vital capacity improved by 98.4 milliliters, while in the placebo arm it declined by 62.3. Improvement was not on the menu, and nobody ordered it.
Nobody at Insilico Medicine guessed TNIK would be the target, not a single researcher, not a single hypothesis. An AI system called PandaOmics, trained on multi-omics and clinical datasets covering tissue fibrosis across organ systems, surfaced TNIK through deep feature synthesis and causality inference, identifying a protein kinase that no human research group had connected to pulmonary fibrosis before the algorithm flagged it as the most tractable intervention point in the entire disease network. A second platform, Chemistry42, generated 78,000 virtual molecules designed to block it, filtered them for potency, selectivity, and synthesizability, then ranked 60 for physical synthesis.
Hit rate: 16.7 percent, compared to the 0.1 percent typical of traditional high-throughput screening, which makes it 167 times better at finding molecules that actually work against a given target.
Thirty months. That is how long rentosertib took from novel target identification to first-in-human dosing, a timeline that traditional drug discovery stretches to six or eight years at a cost that averages over $2 billion per approved drug, making pharmaceuticals one of the only industries where a successful product routinely costs more than a skyscraper to develop and has worse odds of working than a first-time restaurant. Nine out of ten candidates that enter human trials never reach a pharmacy shelf.
What the Pipeline Looks Like Now
Rentosertib is not alone, and the scale of the pipeline is larger than most people realize. According to an industry-wide analysis by humAI.blog, more than 173 AI-discovered drug programs are in clinical development as of early 2026, a number that has roughly doubled every 18 months since 2021, and 15 to 20 of them will enter Phase III this year, the large-scale, multi-thousand-patient, randomized controlled trials that regulators require before any drug reaches a prescription pad.
No fully AI-designed drug has received regulatory approval. Not yet. Independent analysts place the probability of that milestone at roughly 60 percent by the end of 2027. But the early numbers are already strange enough to rewrite assumptions.
| Metric | Traditional | AI-Discovered | Difference |
|---|---|---|---|
| Phase I success rate | 52% | 80-90% | +28-38 pp |
| Screening hit rate | 0.1% | 16.7% | 167× |
| Target-to-Phase I timeline | 6-8 years | 30 months | 2.4-3.2× faster |
| Avg. cost to approval | $2B+ | TBD | Preclinical savings ~99% |
| Candidates in clinical trials | ~10,000/yr industry | 173+ | ~1.7% of total |
AI-discovered compounds achieve Phase I success rates between 80 and 90 percent, which demolishes the 52 percent historical average but requires careful interpretation: Phase I measures safety and tolerability, not efficacy, so these numbers mean AI is substantially better at finding molecules that do not poison people, which is valuable but incomplete. Whether those molecules also cure diseases is the question Phase III answers, and Phase III is a different beast entirely.
Three Companies, Three Approaches
Insilico is not the only company with advanced candidates. Zasocitinib, developed through a collaboration between Schrödinger, Nimbus Therapeutics, and Takeda, is already in Phase III trials for psoriasis and inflammatory conditions, making it the furthest-advanced AI-assisted compound in the entire pipeline. Schrödinger's method differs fundamentally from Insilico's generative chemistry: it combines quantum mechanical simulations with machine learning to model how molecules bind to proteins at the atomic level. Physics first. Data second.
Recursion-Exscientia is the third major player. Merged in July 2025 to form the largest AI drug discovery operation by data volume, this combined entity runs approximately 2.2 million biological experiments per week across 50 human cell types, an experimental throughput that would have been physically impossible five years ago and that generates biological signal at a rate no human team can manually review, which is precisely why the AI exists: not to replace biologists but to find patterns in a firehose of cellular response data that no biologist could drink from directly. Exscientia brings precision molecular design tools and more than 60 petabytes of proprietary data. Their combined pipeline includes over 10 clinical and preclinical programs.
Three companies, three philosophies: Insilico says let the AI originate everything, Schrödinger says use AI to model physics at inhuman precision, and Recursion-Exscientia says drown the problem in biological data until patterns surface. All three enter the same filter this year.
Original Analysis: Pipeline Probability
Here is a calculation nobody else has published on the current AI drug pipeline, and it reframes the entire debate from "will AI drugs work?" to "how many will work?" Start with the base rate: historically, about 10 percent of drugs that enter Phase I eventually receive regulatory approval. If AI compounds maintain their superior early-stage performance and achieve even a modest improvement to 15 percent overall success, the math shifts from interesting to decisive.
At 173 candidates and 15 percent per-candidate probability, the expected number of eventual approvals is 26. Twenty-six. Run the binomial calculation for zero approvals from 173 independent trials at p = 0.15, and the result is 1.6 × 10-12. Practically zero. Even at the traditional 10 percent base rate with no AI improvement at all, the probability of zero approvals from 173 compounds is 1.7 × 10-8.
Candidates are not truly independent, and this matters. Many share platforms, training data, or target-selection methodologies, which means a systematic AI blind spot could take out multiple programs at once, turning 173 nominally independent shots into something closer to 30 or 40 meaningfully distinct bets, each carrying correlated risk that no binomial calculation can properly capture. But even with 40 independent candidates at 10 percent, the probability of total failure is 1.5 percent. Small.
What Could Go Wrong
Phase III is where drugs die. It is where most pharmaceutical failures happen, and it is where AI's advantages in speed and hit rate may matter least, because the variables that determine Phase III outcomes are not the variables AI optimized for. Speed in target identification is irrelevant if the target drives side effects that only manifest after six months of continuous dosing in 3,000 patients instead of 71. A 167-fold screening improvement means nothing if the molecule's selectivity profile, pristine in a petri dish, degrades when a real human liver metabolizes it into an active compound nobody predicted.
Biology is not a pattern-matching problem at this level; it is a chaos problem.
Consider the strongest version of the counterargument: AI's 80-90 percent Phase I success rate may simply mean AI excels at identifying safe molecules. Safe, not effective. That 52 percent traditional Phase I failure rate includes many compounds that are outright toxic, and removing those from the pool earlier does not mean the survivors will cure anything, it just means they will fail later and more expensively, after the cheap safety failures have already been filtered out, producing a pipeline that looks better on paper until Phase III punishes the illusion.
Correlated risk deserves its own reckoning here. If multiple AI platforms converge on similar chemical spaces because they were trained on overlapping public datasets, a single biological surprise could invalidate entire classes of AI-originated compounds simultaneously, the way a single accounting fraud can crash an entire sector of stocks that analysts rated independently but evaluated using the same flawed financial models. Traditional drug discovery is messy, slow, and expensive, but it is also diverse in ways that AI, for all its speed, may not replicate.
Limitations of This Analysis
Several caveats. Rentosertib's Phase IIa trial enrolled 71 patients, a sample too small for definitive efficacy conclusions, and IPF is one disease with specific biology: success there says little about oncology, neurology, or infectious disease, where the failure modes are completely different. Many of the 173 compounds counted as "AI-discovered" used AI for optimization of an existing human-identified target, not end-to-end origination like rentosertib, which muddies the denominator. Insilico's cost and timeline claims are self-reported. Finally, the 80-90 percent Phase I rate may reflect survivorship bias: companies do not issue press releases about the AI candidates that failed before reaching trials, so the published success rate is drawn from a self-selected population of winners.
What You Can Do
If you have IPF or another fibrotic disease, check ClinicalTrials.gov for active rentosertib studies; Insilico's GENESIS-IPF trial enrolled globally, and a US-specific Phase IIa is actively recruiting. If you invest in biotech, the signal just got louder: zasocitinib's Phase III readout will be the first definitive test of whether an AI-designed compound outperforms traditional ones at the approval gate, and that data arrives in 2026, and it will move the sector. If you work in drug development at a company still screening compounds at traditional speeds, Recursion-Exscientia's 2.2 million experiments per week is not a research curiosity posted to impress conference audiences. It is a competitive threat. Act accordingly.
The Bottom Line
Drug development has been the most expensive, slowest, and highest-failure-rate engineering problem in existence for half a century, a distinction so consistent that the pharmaceutical industry invented a term for the phenomenon: Eroom's Law, Moore's Law spelled backward, capturing the observation that the cost of developing a new drug has doubled roughly every nine years since 1950 despite every technological improvement thrown at the problem. AI has now compressed the front end of the pipeline from years to months and multiplied screening hit rates by two orders of magnitude. Rentosertib's Phase IIa results, published in Nature Medicine, show that an AI can find both a target nobody was looking for and a molecule that hits it, and that the molecule works in living, breathing human patients with a disease nobody thought could improve. Fifteen compounds enter the final gauntlet this year. Not all will survive. But the pipeline is deep enough, and the early data strong enough, that 2026 is the year AI drug discovery stops being a promissory note and starts becoming an actuarial table.