Merck Signs $510M AI Biologics Discovery Deal with Protillion Biosciences 🔗

Merck signed a multi-target discovery collaboration with Protillion Biosciences worth up to $510 million in milestones, centered on Protillion's Prot-MaP platform - a megascale data generation system that produces training datasets for protein-design AI at throughput that standard lab workflows cannot match. The platform tests millions of protein variants per run, generating results in 48 hours rather than months. Merck is not buying AI capability here - it is buying the data infrastructure that feeds AI capability. That distinction matters: as foundation models for biology become more accessible, the competitive moat is shifting to whoever controls the highest-quality proprietary training data. Merck's series of recent platform deals - including Google Cloud, Infinimmune, and now Protillion - follows that logic consistently.

Charles River Joins Lilly TuneLab's Federated AI Platform 🔗

Charles River Laboratories joined Lilly TuneLab, Eli Lilly's federated AI/ML drug discovery platform. TuneLab gives participating biotech companies access to models trained on decades of Lilly's internal research data; companies contribute their own data back, improving the models through federated learning. Charles River's role is to supply standardized nonclinical testing services across the TuneLab ecosystem, which matters because consistent testing protocols mean cleaner data contributions, which means better model training. This is infrastructure thinking applied to AI: the platform's predictive accuracy improves with every new participant, and every new participant gets access to a model that is collectively better than anything they could build alone.

OpenAI Builds the Benchmark for AI in Life Sciences 🔗
OpenAI released LifeSciBench, a 750-task evaluation framework developed by PhD-level scientists to assess AI performance on real biological research workflows. The tasks cover seven areas including evidence synthesis, scientific reasoning, and research communication - designed to reflect what a working researcher actually does rather than textbook recall. What makes this structurally significant is timing: as AI-generated molecules enter clinical trials and federated drug discovery models attract billion-dollar investments, the field lacked a shared standard for measuring whether the underlying AI systems are actually capable of the work being claimed. LifeSciBench does not settle that question, but it creates the instrument to start answering it.

The signal: Benchmarks matter because they set buyer expectations and create accountability. When OpenAI builds the evaluation infrastructure for life sciences AI, it is also positioning itself to define what "good" looks like in this domain. Every AI vendor in drug discovery will now be measured against a standard that OpenAI designed.

87% of Life Sciences Companies Are Using AI. Most Cannot Scale It 🔗

The KPMG Global Tech Report 2026, drawn from 124 technology leaders across life sciences, finds that 87% of organizations have integrated AI into workflows and 75% are relying on AI outputs for key decisions. The scaling problem is consistent across company size: AI works in pilots, breaks down in enterprise rollout. Separate research from Lingaro across 150 senior pharma leaders found that 60% of AI strategies lack defined ownership, creating a gap between ambition and execution that is not a technology problem, but an organizational one.

The signal: This is now the dominant conversation at every life sciences AI conference, and it has replaced the earlier debate about whether AI actually works. The market for governance tooling, change management infrastructure, and operational AI platforms is opening up faster than the market for new AI discovery capabilities. The companies positioned to capture that spend are not necessarily the ones with the best models.

Novo Nordisk's Stolen Data Is Now for Sale. The Theft Included AI Models 🔗

Novo Nordisk disclosed a breach on June 11. This week, FulcrumSec began leaking samples of the stolen data and announced it is exploring private sales after the company refused to pay a $25 million ransom. The group claims to have spent two months inside Novo Nordisk's systems, exfiltrating 1.3TB across over 700,000 files - including clinical trial records for approximately 11,500 pseudonymized patients, proprietary drug compound data, source code, and internal AI models used in drug discovery. Access was gained through credentials left exposed in client-side JavaScript and a GitHub access token with access to hundreds of private repositories.

The signal: The same week Merck committed $510 million to acquire exclusive access to proprietary molecular data for AI training, a hacker group stole Novo Nordisk's equivalent assets and is now selling them on the dark web. That parallel is not coincidental - it reflects the same underlying reality. Pharma AI's competitive edge is the data, and the data now has a quantifiable market price whether it is acquired legally or not.

EU Sharpens High-Risk AI Classification for Medtech 🔗

The European Commission released draft guidelines clarifying which AI systems in medtech and digital health qualify as high-risk under the EU AI Act. The key clarification: AI tools that use physiological signals - ECG readings, gait analysis, biometric data - to assess patient health are classified as high-risk under the biometrics use case, regardless of whether a medical device framework also applies. The guidelines are open for consultation until July 23. They do not introduce new obligations but sharpen interpretation in areas where the industry had significant uncertainty.

The signal: For any company building AI tools that touch patient physiology - including large categories of wearables, remote monitoring, and diagnostic software - the EU is clarifying the compliance path.