On September 28, Roche presented at its Pharmaceutical Investor Day an initiative that major pharmaceutical companies have been talking about for years but few have truly undertaken. Aviv Regev, Head of Research and Early Development at Genentech, announced that the company aims to achieve 'AI independence' in drug development, establishing a self-sustaining AI laboratory that is already under construction.

Specifically, Roche plans to reallocate approximately CHF 2 billion (approximately USD 2.41 billion) in R&D savings to fund research projects and efficiency improvements. This money comes from savings generated by the old model rather than new budget allocations, indicating that AI-driven efficiency is starting to pay off.

They call this approach Lab-in-the-Loop. In simple terms, AI models generate hypotheses, robots conduct wet lab experiments, and the results are fed back into the models. After one cycle, the molecules improve, and the models become smarter. Roche began implementing this cycle in 2020, reached the delivery stage in 2023, and is now transitioning it from proof-of-concept to the main R&D workflow.

In terms of data, from the fourth quarter of 2025 to the second quarter of 2026, 40% of Roche's pipeline decisions will have traceable AI or computational contributions. In the second quarter of this year alone, this proportion reached 44%. Internally, the company uses the TargetNexus AI agent to participate in 80% of research portfolio decisions by the end of this year.

How does this cycle operate in practice? Take small molecule optimization as an example. Roche runs a round of active learning every week: generative models propose designs, potency and ADME prediction models score them, and the most promising molecules are selected for laboratory synthesis, with the data then fed back into the models.

In a prospective validation, the model achieved an AUROC of 0.728 in predicting the activity of 63 new molecules. While far from perfect, this is sufficient to answer the practical question of which molecules researchers should synthesize next week. In fact, one molecule progressed from a novel target to a clinical candidate in just 18 months, half the time of traditional processes, and welcomed its first participant this year.

Most notably, the success rate of Phase III clinical trials jumped from 65% in 2025 to over 80% so far in 2026.

Phase III clinical trials are the most costly and high-stakes phase, so this figure has significant implications for the fate of multi-billion-dollar pipelines.

The high success rate is also related to another initiative Roche undertook during the same period. The 'The Bar' framework, launched at the end of 2023, cut a large number of projects. Chief Medical Officer Levi Garraway stated that with fewer projects, the remaining portfolio has become more valuable.

Regev drew a clear line for AI's role. She said that while the company's confidence in model reliability will grow over time, AI will never replace human clinical trials.

Roche grades laboratory autonomy like autonomous driving, ranging from fully manual (L0) to fully autonomous (L5). Humans transition from operators to approvers and eventually to auditors. However, L5 is still far off. Among their partners, they have brought in Anthropic and Medra, which specializes in automated laboratories.

Notably, in March this year, Roche and NVIDIA launched an AI factory equipped with 2,176 additional Blackwell GPUs, bringing the total to over 3,500—the largest announced GPU scale in the pharmaceutical industry. The Boston Innovation Center opened on the 18th of this month in Harvard's corporate research park, signed a ten-year lease, and can accommodate 500 people, with a portion dedicated to AI and machine learning applications in drug development.