AI Drug Discovery Hits Phase 3 as Protein Design Funding Surges

Discovery is getting faster. The clinic is not.

For years, people have promised that artificial intelligence would make drug development faster. This autumn there is more to check that promise against. In September 2026, Insilico Medicine said rentosertib had become the first drug to enter Phase 3 trials with both a target found by AI and a molecule designed by AI.23 Investors are also putting large sums into companies that design proteins. One of AI's best-known biology announcements has meanwhile drawn sharp pushback from scientists. Read together, the coverage points to one conclusion: AI is shrinking the early, computer-based part of discovery, while lab validation, clinical trials and scientific credibility are becoming the real limits on speed.

The usual baseline explains why speed matters so much. Bringing a drug to market commonly takes 10 to 15 years and costs $1 billion to $2 billion, and about 90 percent of candidates fail.27 Other estimates are similar. One puts the average at about $1 billion over 10 to 15 years, with nine failures for every success.24 When reports say AI "speeds up drug development," they almost always mean the earliest steps.

What has actually been shortened

Insilico's lung-fibrosis program is the case most often cited, though reports describe its timeline slightly differently. One account says it took about 18 months to go from finding the target to picking a candidate for preclinical testing.23 Another says the company went from the start of its search to a medicine ready for animal testing in 18 months, and finished its first small safety tests in people in under 30 months.28 A third says the platform found a target, designed a molecule and moved it into clinical trials in about 30 months.27 These figures don't conflict. They measure different finish lines, and readers should keep track of which one is being quoted.

An earlier example shows a similar pattern. In 2020, DSP-1181 became one of the first AI-designed drugs to enter clinical trials. Its exploratory research phase took less than 12 months, compared with a typical average of 4.5 years.23 The shared lesson is that the steps AI speeds up most are choosing targets and making molecules. Neither saves the years spent testing in people.

The tools themselves keep getting faster. Boltz-2, released by MIT and Recursion in June 2025, predicts how a drug binds to its target and how strongly, in about 20 seconds on a single GPU. That is more than 1,000 times faster than the standard physics-based method, called free energy perturbation, and on a common benchmark it matched a popular open-source version of that method about as well.23 Chai Discovery's Chai-2 antibody model tested 20 or fewer designs for each of 52 targets and found a working antibody for half of them in one round. Its overall hit rate was 16 percent, and the whole process took under two weeks.23

The money follows de novo design

The funding figures show where investors think the value is. Companies that design new proteins, antibodies and small molecules from scratch, known as de novo design, made up 56.3 percent of AI drug-discovery deals in the first quarter of 2026, up from 35.7 percent in the second quarter of 2025.13 Isomorphic Labs raised $2.1 billion in May 2026 in a round led by Thrive Capital.12 Reuters reported that the company expected its first clinical trials by the end of 2026.17 Chai Discovery raised $400 million in July, valuing it at about $3.8 billion.19 It has signed deals with Eli Lilly, Pfizer and Bristol Myers Squibb.22

September kept up the pace. One count of 12 qualifying deals that month found more than $1 billion in disclosed financing.11 The largest included Anew Labs with $290 million, Basecamp Research with $140 million and BigHat Biosciences with $75 million.11 BigHat's round brought its total funding to $223 million. Its pipeline includes BHB810, an antibody-drug conjugate already in a Phase 1 trial.16 Another company tracker lists Generate:Biomedicines' $400 million IPO in February 2026, with its anti-TSLP antibody GB-0895 in Phase 3.18

The companies in this group are now judged differently. One industry ranking describes Xaira Therapeutics, which launched in 2024 with more than $1 billion, as still early in turning that money into a pipeline.18 Another analysis reports that Xaira revealed its first drug programs, still preclinical, in October 2026.23 The question investors face is no longer whether the models can generate designs. It is whether those designs survive lab tests and trials. On that question, rentosertib and GB-0895 count for more than any benchmark result.

Protein structure prediction: powerful, but not yet trusted by default

AlphaFold remains the foundation for most of this work. The AlphaFold Database now holds more than 260 million predicted proteins and protein complexes. A September release added predicted complex structures for more than 2,800 viruses to support pandemic preparedness.33 The partners behind that release say clearly what predictions can't do. A predicted structure doesn't show how mutations change a virus or how the virus behaves, and answering those questions requires lab experiments.33

That caveat applies across the field. A new systematic review of nine major structure-prediction models found that eight were validated mainly against standard benchmarks. Only AlphaFold 3 was checked directly against experimentally determined structures. None of the nine reported validation on independent outside datasets or in prospective lab experiments.34 Separately, a benchmark of AlphaFold 3 on modified nucleic acids found success rates above 80 percent. However, the model placed the modified parts less accurately, and it performed worse on structures released after its training cutoff.35 That second finding suggests some of its strong results may come from structures it saw during training rather than real generalization.35

These studies don't undercut AlphaFold. They show what speed alone can't fix. A team that can produce thousands of plausible designs still has to decide which ones to make and test, and that is where much of the time goes.

The lab is the new bottleneck, and companies are building around it

Parts of the industry are already treating this as a business problem. Lonza has partnered with startup Onava, which will test biologic drug candidates from its AI platform using Lonza's protein-production technology. Onava will feed the lab results back into its models to improve future designs and to make sure candidates can be manufactured at scale.29 Commentary on the deal makes the point directly: AI can design candidates in days, but checking them in the lab is still slow and expensive.29 BigHat is using its new funding partly to expand the automated systems that generate data for its protein designs.11

AI agents, systems that plan and carry out tasks on their own, are the next step in the speed story. A Stanford team reported in Science that tens of thousands of agents predicted the most promising approach to attacking B7-H3, a well-known lung cancer target. Weeks later, a drug company announced encouraging results from a drug using the same approach.26 The lab presented the work as a proof of concept. The agents still can't run experiments with cells, and it is unclear whether the method works as well on targets humans have studied less.26 A Nature review reports that one agent-based system planned chemical synthesis routes in about 16 minutes on average, compared with about 70 minutes for human chemists, at close to expert-level quality.25

The CRISPR controversy is a warning about hype

The sharpest test of AI-in-biology claims came from gene editing. Anthropic said that 950 AI agents spent 21 hours searching a DNA database, picked out 3,500 copies of an enzyme called reverse transcriptase, and flagged an unusual system in bacteriophages, viruses that infect bacteria, that resembles CRISPR.1 CEO Dario Amodei called it a possible new gene-editing mechanism, and shares of gene-editing companies fell after the announcement.2 The researchers had not worked out what the sequences actually do, and the work has not been peer-reviewed.2

Scientists outside the company quickly pushed back. Several said the finding was early and incremental, not a breakthrough.1 A University of Copenhagen researcher said his own unpublished work describes the same viral signatures. He suggested that either his material reached the model somehow or there was more human direction behind the search than the word "autonomous" implies.1 Anthropic responded that Claude isn't trained on user conversations.3 Fyodor Urnov, who coined the term "genome editing" and is working with Anthropic on a CRISPR project, said it is not yet clear the system is even an enzyme.3 The more balanced view is that the real accomplishment was searching a huge amount of data at scale. One expert described AI as very good at "grunt work when orchestrated by experts."13

Traditional gene-editing work continued alongside the controversy, at its usual pace. Arbor Biotechnologies named three ALS candidates that use compact editors delivered in a single viral vector designed to cross the blood-brain barrier. It expects to file its first applications to begin human trials in 2027.8 Researchers at WEHI in Australia used prime editing to let doctors control engineered T cells with immunosuppressant drugs hospitals already stock, and they are now working toward a first human trial.9 In South Korea, regulation is the obstacle. About 623 billion won of government gene-editing research has stalled because rules for modified organisms treat edited crops and animals the same as GMOs.6

The verdict

The evidence supports a narrower claim than the marketing does. AI is clearly shortening early discovery, from finding targets to designing molecules.2328 Investors are betting heavily that de novo design will turn into approved drugs.1319 But predictions without experimental validation remain hypotheses,34 and announcements without peer review can damage trust.23 Over the next 18 months, the AI drug-discovery companies that stand out are likely to be the ones that tie their models most closely to lab results and report clinical data, not just faster designs.