The AI Biotech That Buys Drugs Instead Of Inventing Them

Pathos AI has just agreed to pay $125 million upfront for the rights to an experimental cancer drug called JSKN016. If the drug reaches all of its milestones, the deal could eventually be worth more than $2 billion.

The drug is already in a Phase III study for triple-negative breast cancer, so Pathos is buying a fairly advanced clinical asset that another company discovered and developed.

JSKN016 is an antibody-drug conjugate, or ADC. Think of it as an antibody carrying a toxic payload, and it is designed to find cancer cells, enter them, and release the drug inside. This particular ADC can bind two proteins called TROP2 and HER3, which gives it a potentially broader way to recognize cancer cells.

At first glance, this looks like another large licensing deal in oncology, but what makes it more interesting is the type of company buying the drug.

Most AI drug discovery companies are presented as molecule factories. They use AI to find a target, design a drug, and move it toward the clinic. Pathos appears to be building something closer to a private equity firm for biotechnology.

It searches for promising drugs, buys or licenses them, places each one inside a small operating team, and then uses a common data and AI system to decide how that drug should be developed.

Pathos calls the search system Scout, the small teams are Sprint Pods, and its shared AI core is the Foundry. Pathos says each Sprint Pod operates almost like an independent biotech company inside the larger organization.

I think that structure is the most important part of the story because Pathos does not need to invent every drug it owns. It needs to be better than everyone else at deciding which drugs are worth owning, and which patients should receive them and which clinical trial should come next.

The company behind this strategy also helps explain the plan. Pathos was founded in 2022 by Eric Lefkofsky and Ryan Fukushima, both closely connected to Tempus, the healthcare data company Lefkofsky also founded.

Tempus has spent years assembling clinical, genomic, and other patient data from cancer care. Pathos is paying Tempus $200 million over three years to license a large de-identified oncology dataset. It is also covering the first $60 million of cloud-computing costs for a foundation model being developed with Tempus and AstraZeneca.

Pathos has enough capital to execute this as well. It raised $365 million in 2025 at a reported valuation of about $1.6 billion.

The private equity comparison is useful because these firms do more than buy companies. They search for overlooked assets, acquire control, install an operating plan, allocate capital, and try to make each asset more valuable.

Pathos may be attempting the same thing with drugs. JSKN016 already had encouraging clinical data and had reached Phase III. It was visible to every large oncology business-development team.

Pathos may have found something others misunderstood, or it may simply have offered enough money to win a competitive asset. From the outside, we cannot know yet.

The real evidence will come from the portfolio. Pathos will have to show that the drugs it selects succeed more often, reach answers faster, or consume less capital than drugs chosen through conventional business development.

One successful acquisition would not prove very much, but if they can repeat that success over multiple assets, that’s something else.

AI drug discovery has spent years asking whether a model can invent a better molecule. Pathos is taking a completely different route. The larger opportunity may be building a model that knows which existing molecule deserves the next $125 million.

GSK is paying Relation Therapeutics up to $110 million to generate the kind of biological data AI models usually do not have. Relation will run experiments in human cells, measure what happens after genetic and drug interventions, and use those results to train its models. The interesting part is that the lab work is not only there to validate a prediction. It is there to create better training data for the next one. The wet lab is starting to look a lot more like a data centre. [Link]

Chinese AI biotech Earendil Labs is reportedly trying to raise close to $400 million in a Hong Kong IPO at a valuation of around $5 billion. The problem is not the model or the pipeline. Chinese regulators are questioning the company’s offshore holding structure and where control of the business actually sits. AI biotechs combine models, biological data, compute, foreign capital, and drug-development capacity. Governments are starting to treat that combination as something more strategic than a normal biotech company. [Link]

More of the EU AI Act came into force this week, except for some of the parts healthcare companies have been waiting for. The main obligations for high-risk AI systems were pushed to 2027, while AI built into regulated products such as medical devices generally has until 2028. The strange part is that the law is moving faster than the standards needed to follow it. Europe has rules for trustworthy AI, but it is still working out what companies must actually show to prove their systems are trustworthy. [Link]