AI Drug Discovery Deals Surge as AstraZeneca Bets on insitro

Pharma keeps paying for AI, but mostly in milestones

The biggest AI drug discovery story of the week is a familiar kind of deal. AstraZeneca has signed a discovery and licensing agreement with insitro, the AI biotech led by Daphne Koller. insitro will use its Virtual Human platform to find new targets in metabolic dysfunction-associated steatohepatitis (MASH) and in cardiomyopathies.11 insitro gets $40 million upfront. It could earn about $1.46 billion more in development and commercial milestones, plus single-digit royalties, and AstraZeneca takes exclusive worldwide rights to the programmes it selects.1118 AstraZeneca will run and pay for any later clinical trials, according to Reuters reporting cited by The Financial.18

Coverage disagrees on how big the deal is. Contract Pharma and Analytics Insight round it to about $1.5 billion.1519 Crypto Briefing cites a Bloomberg figure of up to $1.86 billion.14 The disclosed terms only account for the $40 million upfront plus roughly $1.46 billion in milestones. The best reading is that $1.5 billion is the defensible number and the higher figure counts something the companies haven't itemised publicly. Either way, the headline value matters less than the structure. Less than 3% of the possible total is guaranteed. The rest depends on targets actually becoming development candidates. Analytics Insight pointed out that any resulting medicines will still need further testing before they reach patients.19

AstraZeneca's research chief, Sharon Barr, framed the deal around picking targets. She called that one of the most important decisions in R&D, and said AI could cut the number of molecules tested in the lab.1118 That focus is telling. The pitch is less that AI invents drugs and more that it helps choose the right disease biology to go after, which is where many programmes fail.

The deal also fits a pattern for insitro. A week earlier, on October 6, it said machine-learning models built with Eli Lilly were now available to some biotech companies through Lilly's TuneLab. The models predict how small molecules behave in lab tests and in living organisms, and were trained on preclinical data from hundreds of thousands of Lilly compounds.18

AI protein design draws its own deals

AI protein design had its own deal in the same week. On October 13, Zealand Pharma signed a multi-target collaboration with Generate Biomedicines to develop AI-designed protein drugs for obesity and metabolic disease.20 Zealand pays $30 million upfront, and milestones could reach about $1.2 billion plus tiered royalties.6 Generate's model is to have AI propose new protein structures and then test large numbers of them in the lab to see which ones work.6

Pairing AI design with heavy lab testing is now standard across the field. Danaher said on October 7 that it will open its first AI-powered autonomous lab, based at Abcam. AI will design antibodies and other binding reagents, and lab results will be fed back to improve the model.12 Danaher expects the lab to run at scale in early 2027 and says it could make the path from idea to validated reagent up to eight times faster.12

The same logic shows up in academic funding. AI BioDesign launched in Seattle on September 3 with $94.6 million from the Paul G. Allen estate. It is co-led by Nobel laureate David Baker and genomicist Jay Shendure.10 The institute's plan is to spend on lab infrastructure that tests millions of designed sequences, not to train ever-larger general models.10 When the field's best-known protein designer bets on lab throughput over model size, that suggests where the bottleneck now sits.

Funding is concentrated in de novo design

Venture funding backs that view. Isomorphic Labs, the Alphabet spinout run by Demis Hassabis, raised $2.1 billion in May in a round led by Thrive Capital. Alphabet and GV returned, and new backers included MGX, Temasek, CapitalG and the UK Sovereign AI Fund.1 Chai Discovery followed in July with a $400 million Series C at a valuation of about $3.8 billion. Its investors include OpenAI, and its partners include Lilly, Pfizer and Novartis.9

The money keeps coming. BigHat Biosciences closed a $75 million Series C in late September. That brought its total to $223 million and will fund BHB810, an antibody-drug conjugate already in a Phase 1 trial.39 Basecamp Research raised $140 million and Aureka Biotechnologies $100 million.9 Superluminal Medicines raised a $60 million Series B for AI-driven small-molecule work on GPCR targets.2

One analysis found that de novo design, meaning models that create new proteins, antibodies and small molecules, made up 56.3% of AI drug discovery deals in Q1 2026, up from 35.7% in Q2 2025.4 Another found that AI target-discovery companies accounted for 15% of deals but only 2.3% of capital.8 That gap matters for the insitro deal. Investors are paying a premium for molecule makers. Target finders mostly get their big money through pharma partnerships like AstraZeneca's.

Public markets are getting a test as well. Iambic Therapeutics has filed to go public at a valuation of up to $806 million. Its lead asset, the HER2 inhibitor IAM1363, is in clinical testing.29 A vendor forecast puts the AI drug discovery market at $5.09 billion in 2026 and $17.56 billion by 2031.17 Projections like that come from firms selling research and should be treated with caution.

Protein structure prediction is becoming a commodity

The field still traces back to AlphaFold. Its database covers more than 200 million predicted protein structures.21 AlphaFold also raised structural coverage of human protein residues from 48% to 76%.28 The competition has since moved past that. In February, Isomorphic described IsoDDE, a proprietary drug design engine. The company says it more than doubles AlphaFold 3's accuracy on a hard protein-ligand benchmark, predicts binding affinity better than physics-based methods, and can find binding pockets from sequence alone.2125

Two points stand out. First, IsoDDE is fully proprietary, unlike early AlphaFold releases.25 One analysis argues that open or semi-open models such as the AlphaFold 3 code, Boltz and Chai are turning structure prediction into a commodity. In that view, the value moves to the full design-make-test loop.29 A Nature commentary adds that outputs from these models shouldn't be treated as ground truth.23

Second, clinical proof is still missing. Isomorphic expects its first clinical trials by the end of 2026.2 A recent ScienceDirect review notes that no drug discovered mainly through AI has reached the market. It argues the biggest gaps are in clinical translation, not early design.28 The same review points out that AlphaFold models give rigid snapshots of proteins. Independent checks show AlphaFold 3's advantages don't reliably carry over to physically valid poses or to new targets.28

The implication is that structure prediction alone is no longer a competitive edge. Large funding rounds now go to companies that can show their molecules work in the lab and, eventually, in patients.

CRISPR gene editing news: real clinical progress, and an AI claim that got ahead of itself

Gene editing has had its own crossover with AI. In September, Anthropic said its Claude model had "autonomously" found a CRISPR-like enzyme system after 21 hours searching a DNA database. The company said it had not yet worked out what the system does.34 Shares in gene-editing companies dropped shortly after the announcement.36

Scientists were cautious. One University of Minnesota geneticist said the results don't show gene editing or explain how the system works.36 Feng Zhang called the work "genuinely intriguing," according to Anthropic's own release.36 New Scientist argued that at best it would add one more tool to a fast-growing gene-editing toolkit.38 That is the most defensible reading: interesting evidence of what AI can do as a search agent, not a new CRISPR.

The clinical news is more concrete. The FDA accepted Intellia's application for lonvo-z, an in vivo CRISPR therapy, and granted it priority review. In the company's Phase 3 trial, one dose cut hereditary angioedema attacks by 87% compared with placebo. Base editors from YolTech and Beam corrected the mutation behind alpha-1 antitrypsin deficiency in early-stage trials. UCSF researchers reported inserting a large stretch of DNA at a precise spot in human T cells without removing the cells from the body. That could eventually allow CAR-T cells to be made inside the patient.32

There are also warning signs. A Lancet editorial called on the field to reassess safety after two deaths in Chinese gene-editing trials, which appear to be linked to high doses of AAV delivery vectors.

What it adds up to

This month's news follows a consistent pattern. Pharma is paying modest upfront sums for AI platforms and backloading most of the value into milestones. Investors are putting the largest checks into design platforms. Gene editing is moving through regulators on clinical data, not on announcements. Neither AI drug discovery nor AI protein design has a marketed drug to show yet. Until one arrives, the most credible measure of progress is the lab: how quickly AI designs survive testing.