Zahra Khwaja presents a narrative that defies the pharmaceutical industry's entrenched culture of secrecy, arguing that Eli Lilly is betting its future on a radical premise: that sharing proprietary AI models is the only way to accelerate drug discovery. The piece is notable not for announcing a new drug, but for detailing a structural pivot where a $45 billion giant invites competitors to train on its data without ever seeing the raw numbers. This is a high-stakes gamble on the mechanics of trust, and the evidence suggests the industry may finally be ready to stop hoarding and start collaborating.
The Architecture of Trust
Khwaja anchors her analysis in the technical innovation that makes this collaboration possible: federated learning. She explains that TuneLab, launched in September 2025, does not ask partners to upload their sensitive molecular data to a central server. Instead, the platform moves the computation to the data. "The platform allows for additional privacy-enhancing techniques like differential privacy (adding mathematical noise to model updates) and k-anonymisation to further protect proprietary information," Khwaja writes. This distinction is critical. It transforms the traditional data-sharing model, which requires surrendering IP, into a system where the "global AI model is distributed... to each participating company's local node."
The author highlights the scale of the investment behind this infrastructure, noting that the models were trained on over 500,000 data points accumulated over two decades at a cost exceeding $1 billion. By opening these assets, Lilly is effectively saying that the value lies not in the data itself, but in the collective intelligence generated by the ecosystem. This approach mirrors the evolution of federated learning seen in healthcare privacy research, where the goal has long been to train robust models across disparate hospital systems without violating patient confidentiality. Here, the stakes are commercial rather than clinical, but the mathematical principle remains the same: the whole becomes greater than the sum of its parts.
"The key security principle: Lilly never hosts partner data, and partners never see each other's data. All raw molecular information stays within each organisation's infrastructure."
Critics might argue that no amount of encryption can fully eliminate the risk of model inversion attacks, where a sophisticated actor could theoretically reconstruct training data from the model updates. However, Khwaja notes that Lilly conducted an internal simulation using data from acquired companies to prove the system's viability before opening the doors. "We really proved to ourselves that it would work from both a data science perspective and a usability perspective before we opened it up to early access third-party partners," she quotes Aliza Apple, the architect of the initiative. This internal due diligence suggests the administration is not acting on faith alone, but on validated engineering.
The Strategy of Openness
The commentary shifts to the strategic rationale, where Khwaja captures the tension between corporate protectionism and the need for speed. The article details how Lilly selected specific models for the platform—small-molecule ADMET (absorption, distribution, metabolism, excretion, and toxicity) and antibody developability—because these are areas where "companies are not typically seeking to differentiate but instead are simply generating data and solving problems empirically."
This is a shrewd selection. By focusing on the "commodities" of drug discovery rather than the proprietary targets, Lilly removes the primary barrier to entry for competitors. Khwaja writes, "Since there's a lot less differentiation in the current empirical approach for ADMET and development there's a very clear measurable gain around time and dollars from a savings perspective that we can deliver if we can replace all of that work with a model." The argument is that efficiency gains in these foundational areas will free up resources for the high-risk, high-reward work that actually defines a company's competitive edge.
"We have a strong belief that the medicines of the future...are going to be discovered by AI, with the help of AI, over the next several years."
The piece also touches on the ambition to create "digital twins" of scientists, a concept that echoes the historical trajectory of computational biology. Just as the development of the first effective insulin analogs required a leap from animal extraction to recombinant DNA technology, this new era requires a leap from isolated data silos to networked intelligence. Khwaja notes that the executive team, including CEO Dave Ricks and CIO Diogo Rau, provided direct sponsorship, which was essential for overcoming the "appropriate stage gates" that usually stifle innovation in large pharma.
The Future of Agentic Workflows
Looking forward, Khwaja explores the potential for "agentic reasoning," where AI systems autonomously select and orchestrate different models. This is where the vision becomes most complex. "The core challenge is this: for an agent to make good decisions about which model to use when, we need rigorous, consistent benchmarking across all available models," Khwaja explains. Without standardized benchmarks, an AI agent cannot reliably determine which tool is best for a specific task.
The article suggests that TuneLab could become the industry's de facto benchmarking ground. If successful, this would solve a fragmentation problem that has plagued the sector for years. Khwaja quotes Apple on the potential for this evolution: "Imagine an AI agent that can intelligently select and orchestrate different models across a full drug discovery workflow." The implication is that the next breakthrough in drug discovery won't just be a new molecule, but a new way of organizing the scientific process itself.
"We're actively exploring how to incorporate agentic elements into TuneLab. The vision is compelling - imagine an AI agent that can intelligently select and orchestrate different models across a full drug discovery workflow."
A counterargument worth considering is whether this "SaaS co-creation" mindset truly levels the playing field or simply cements the dominance of the largest players who can afford the most sophisticated computing infrastructure. While the platform is open, the computational cost of running these models locally could still be a barrier for smaller biotechs. Khwaja acknowledges the potential for third-party models to be included, noting, "We recognise that there may be best in class that will not come from Lilly, and we're open to including them in TuneLab." This openness is a necessary check against the platform becoming a walled garden.
Bottom Line
Zahra Khwaja's coverage effectively reframes the pharmaceutical industry's AI strategy from a race for proprietary dominance to a collective infrastructure build. The strongest part of the argument is the technical and strategic justification for federated learning, which solves the trust deficit that has historically prevented collaboration. The biggest vulnerability remains the execution risk: can an industry built on secrecy truly adapt to a model where the value is derived from shared, rather than hoarded, intelligence? The next few years will determine if TuneLab is a fleeting experiment or the new operating system for drug discovery.