Biotech’s new moat: human data, not AI models

By late 2026, biotech and pharma leaders were publicly shifting from celebrating molecule generation to building systems that could test and correct AI in the lab. On The Prof G Pod with Scott Galloway, Lilly said its scientists still must test results and run experiments, described TuneLab as a data-sharing model-training platform, and said the task ahead was to automate the other 980 while still relying on human experimentation. The podcast agenda itself reinforced that point with 07:21 Human Proof of concept and 16:22 Translation is the moat.

That pivot was echoed across 2026 in arguments that validation required closed experimental loops, not better black boxes alone. Odd Lots framed the issue around a model that doubled the prior probability from 5% to 10%, while Gradient Dissent said drug development takes forever because someone must prove a drug is safe, and Medra positioned an AI experimentalist to plan iterative assay campaigns. The economic logic was equally clear: the pipeline doubled, but proof still had to come from real-time wet-lab feedback, and Phase II and Phase III human trials could cost anywhere from $50 million to $300 million+ per trial.

The strategic shift is happening because better models do not erase biology’s messiness or the long lag before reality answers back. BigGo Finance quotes Recursion CEO Najat Khan saying biology lacks software-like traceability, while only 4% of eligible patients enroll in trials, so Recursion mines de-identified claims and EHRs to find them. Its data factory, an autonomous lab, has generated 1 trillion iPSC-derived neuronal cells and whole-genome knockout readouts because prediction must be continuously corrected by experiment and downstream operational data. Noubar Afeyan’s Flagship Pioneering partner Armen Kahvejian said bringing a single therapeutic to market still takes over a decade and costs upwards of two billion dollars.

That is why capital is moving toward systems that connect design to validation, not standalone prediction engines. AI Drug Discovery Faces Data, Regulatory, and Outcome Challenges says headline deal values are milestone-contingent, and Jisoo Lee said Eli Lilly and NVIDIA committed up to $1 billion over five years to link AI analysis with lab experiments so models can learn from results. BigGo Finance says this closed loop helped Recursion cut target-to-clinic timelines from five years to roughly 18 months, while the broader pattern includes XTAL PI’s $6 billion, $8 billion, and $10 billion milestones and the need for integrated AI-experiment platforms that close the loop between in silico design and wet-lab testing.

The strategic moat in AI drug discovery is shifting away from model architecture alone and toward ownership of the experimental data that can correct it. As the August 3 analysis argued, “More than 90% of drugs that enter clinical trials fail… In the large majority of cases, the molecule was engineered just fine,” because the mechanism was wrong; mechanistic understanding is “a rare commodity,” and misallocation shows up in the fact that 38 targets already have over 50 programs each even as the number of novel targets advanced annually has fallen sharply.

That is why the highest-value systems are increasingly those that generate proprietary biological ground truth, not just predictions. Understanding Virtual Cells and Their Current Role in Drug Discovery said the field is “somewhere like where AI was 10 or 20 years ago,” lacking “LLMs of biology that are generally applicable problem solvers,” while a September 11 Nature paper noted that “most existing approaches lack large-scale, time-resolved perturbation proteomics data” before building on more than 38 million measurements; “Raw mass spectrometry proteomics data have been deposited in the ProteomeXchange Consortium available via iProX (IPX0007409000),” and the “PTDS protein matrix and associated resources are available for academic and” broader use, underscoring that progress depends on rich lab datasets. The clearest commercial expression of that shift is the rise of companies treating wet-lab output as the core asset that makes their models different. Security Now described Laya AI as “an AI-driven physical biochemistry laboratory” training a “frontiercale scientific reasoning AI” on “a proprietary data set of 950,000 RNA sequences that have been physically synthesized, tested in cells,” then fed back into the model, a feedback loop that shows why strategic value now sits in continuously generated, proprietary experimental results rather than in software alone.

The strongest candidate for that missing biological ground truth is no longer the animal model but the human-relevant assay. As Putman Media reported, Torch Bio CEO Jonny Sexton argued that “existing preclinical models simply aren't predictive,” with animal-to-human liver injury comparisons “only around 55 percent, barely better than a coin flip,” while policy is moving the same way: the 2022 FDA Modernization Act 2.0 eliminated the requirement for animal testing, and in 2025 the FDA published a roadmap to phase it out in favor of “new alternative methods” built from engineered tissue and organs-on-chips.

What makes those systems credible is that they generate measurable human biology, not just better theory. Putman Media described how Sexton’s team reprograms patient-derived cells into induced pluripotent stem cells, engineers “miniature liver tissue,” and uses imaging plus machine vision to score markers of liver health and injury; with acetaminophen, the platform produced an “83.6 percent drug-induced liver-injury risk score,” and across 74 known-safe and known-unsafe drugs and roughly 20 million cell observations, reached 87 percent balanced accuracy. That is why, as MTS quoted, “within the next six to 12 months” “verified tissues” and “human tested before the clinical trial” are becoming a public proof point.