Google Cloud’s Shweta Maniar on moving life-sciences AI from research to reality

DeepMind was among the first organizations built from the ground up around artificial intelligence. Founded in 2010 and acquired by Google in 2014, the London-based lab, now known as Google DeepMind, helped catalyze the current AI wave, and two of its researchers, Demis Hassabis and John Jumper, won the 2024 Nobel Prize in Chemistry for AlphaFold, a system that made it possible to computationally predict protein structures that once took months or years of painstaking laboratory work.

That research-to-product pipeline is now central to Google’s pitch to life sciences companies. AlphaFold, once a DeepMind breakthrough, is available through tools such as AlphaFold Server for noncommercial research. AlphaFold also helped spur a wider ecosystem of open or publicly available structure-prediction systems, including OpenFold, RoseTTAFold, ESMFold, Chai-1 and the Boltz models. Newer systems from Alphabet, including AlphaGenome and AI Co-Scientist, point to the search pioneer’s ambition to use AI to help researchers interpret biology, generate hypotheses and organize scientific work around large, messy datasets.

Generative AI, a field Google researchers helped launch with the 2017 “Attention Is All You Need” paper that introduced the transformer architecture, is compressing the distance between technical capability and everyday use. Thanks to genAI, experts are faster. A computational biologist who once spent weeks on protein structure prediction can now run it in hours. The practical effect is a broader redistribution of technical work. Computational biologists can use structure-prediction systems as a faster starting point for modeling. Clinical operations teams can experiment with AI agents for administrative and data workflows. Medical writers, analysts and procurement teams can increasingly query complex datasets in natural language.

“We are now all technology people,” said Shweta Maniar, director of strategic industries for healthcare and life sciences at Google Cloud, speaking at NTT’s Upgrade 2026 conference in April. “No matter what role, what industry, what title you have.”

Maniar’s comment captured the broader pitch behind Google’s life sciences push. The company wants AI to move from specialized research systems into everyday scientific and clinical workflows, from protein-structure prediction and genome interpretation to hypothesis generation, data readiness and regulated enterprise use.

From pipetting to orchestration

The practical result, Maniar said, is that researchers in the lab, for instance, will spend less time on routine lab work and more time directing complex scientific systems. “Researchers are going to go from doing the proverbial pipetting in the lab to being the orchestrators,” she said.

Maniar also envisions that self-driving labs, which have existed as a niche technology, will grow more common. “We’re moving toward self-driving labs,” Maniar said. Just as self-driving cars went from science fiction to reality in less than two decades, she expects research facilities to follow a similar path. In this model, AI systems would manage routine experimental execution, data collection, and initial analysis. Meanwhile, researchers will be freed up to focus on defining goals, validating results and deciding which hypotheses deserve deeper exploration.

Maniar repeatedly emphasized a “human in the loop” standard, framing agentic AI in clinical trials as triage rather than autonomous decision-making. When an audience member asked about data privacy, she called its Gemini genAI platform “a closed tool” and said “our models are not being trained on any of the information that you’re putting into it.”

You can’t build a skyscraper on sand

Before organizations can reach the self-driving lab stage, Maniar said they must first address a more fundamental issue: data. “You can’t build a skyscraper on sand,” she warned. Many life sciences companies still have fragmented or poorly governed datasets, which makes it difficult to deploy even moderately complex AI tools reliably. Maniar argued that organizations should focus first on cleaning and structuring their data so it is “within an understandable reach,” even if it is not yet perfect. Only then, she said, can they begin experimenting with more advanced capabilities such as AI Co-Scientist or automated lab workflows.

Even with clean data, Maniar cautioned against starting with the most ambitious use case. The initial AI deployment that builds the most organizational trust, she said, is often unglamorous. “Administrative efficiency has seen so much success as initial success, and I’ll tell you why,” Maniar said. “It’s been able to be repeated in multiple departments or business units within an organization, and it’s a very simplified way to build trust in technology across people who may not traditionally be technology people.”

Building trust

That trust, she argued, is what unlocks everything else. And building it requires a counterintuitive approach to internal politics. Maniar said organizations should spend more energy supporting the few people already willing to experiment. “Although we tend to want to go after the naysayers, it’s actually better to focus on that one person who actually wants to implement your project,” she said. “Eventually, the naysayers come around and say, ‘I want to be that person. I want a project like that. I want that initiative in my department.'”

Ultimately, Maniar urged organizations to stop waiting for the perfect model or the perfect conditions. “Start now,” she said. “Don’t wait for a perfect model. There is no such thing as a perfect model.” Models will continue to evolve, she added, and new tools will follow. The priority, she said, is to start learning what AI can solve in a specific research, clinical or operational setting.