# How AI is rewiring scientific discovery at Google Research > Source: > Published: 2026-07-05 16:01:00+00:00 I's transformative effect on software development is well known. Next up for transformation: scientific research. [Lizzie Dorfman](https://www.linkedin.com/in/lizzie-dorfman/) leads science AI at Google Research, where her team develops AI systems to help scientists solve problems across genomics, neuroscience, epidemiology and climate science. In a conversation with The Deep View at Google I/O in May, Dorfman explained how Google's own researchers have adopted AI agents, why they can now explore hundreds of thousands of scientific ideas instead of just a handful, and why the biggest breakthroughs often come from solving smaller bottlenecks along the way. This interview has been edited for brevity and clarity. **Jason Hiner: Tell us about your background and how you ended up at Google Research working on AI for science.** **Lizzie Dorfman**: I grew up on the Stanford campus, in a house of doctors. My dad's a neurologist who once brought a brain in for show-and-tell. My mom was a pediatrician. I studied human biology and thought I'd be a doctor, then worked in a hospital surrounded by them and realized it wasn't my calling. So I joined Google in 2003 with a bunch of people who were generally intelligent and could learn what they needed to, because nobody had experience in online search or advertising. I figured I'd stay a year. I stayed five. I left because my passion was in the life sciences, went to 23andMe, then back to grad school for a PhD in public health genetics. I came back to Google in 2015, when it got much more serious about how technology could be valuable for the life sciences. That was 11 years ago. **Jason Hiner: There was so much enthusiasm around CRISPR and genomics, with the sickle cell breakthrough and the promise of personalized medicine, and then it seemed to stall. What happened?** **Lizzie Dorfman**: I'll go back to the Human Genome Project. When we announced its completion, the promise was that everything would follow: personalized medicine, the right drug for the right person at the right dose. That hasn't happened. The field moved to polygenic risk scores, then to making sure they worked equitably across populations. That also hasn't hit the mainstream. CRISPR sits on that same stepping stone. One issue is how fully we've characterized the germline mutations that cause disease. That still sits mostly in single-gene, Mendelian diseases, which is some of them, but very much not all. The second is a technological constraint: off-target effects. We can do precision-guided scissoring of the genome, but some other things happen a lot of the time, and that can't happen. The technology needs to keep advancing before it can be used at scale. We don't want change for the sake of change. We want change because it clearly confers benefit, especially in healthcare. You'd presume personalized is better, but actually a good outcome is better. **Jason Hiner: Tell me about the early days of applying deep learning to science at Google.** **Lizzie Dorfman**: When I came back in 2015, deep learning was new, and scientists were skeptical. It was a lot of "here's a big hammer, everyone's looking for nails." We pulled together people with bioinformatics backgrounds and asked what we could apply it to. The first problem was variant detection. Sequencers produce really noisy data, so there's a lot of algorithmic post-processing. What computational biologists actually did was look at the data in a genome viewer, we call it a pile-up, and eyeball it. So we thought: if human experts can see it, probably a computer vision model could too. We got a lot of pushback until the FDA started the Precision FDA competition and gave everyone an unreleased reference genome with a known right answer. Our tool, DeepVariant, won for overall accuracy, and that quieted people down. We weren't debating whether the methodology was well suited. We all agreed on the metrics, and we had the top results. That team is still going. To date, something like 2.5 million human exomes and genomes have been processed with our tools. **Jason Hiner: What's top of mind for you right now?** **Lizzie Dorfman**: We've done AI and science for a decade, and for most of that time the model was: start with a big computational problem, ask why this, why us as Google, [and] who are our collaborators. But Gemini was a real sea change. We're a microcosm of a lot of computational science teams outside Google, so we asked how we could make it useful for what we do. What came out of three years of eating our own dog food is what we call ERA: Empirical Research Assistants. And it's shockingly powerful. Different domains are at different phases of maturity. You've got early adopters, and then the super-skeptics who say, "[this is] my physics-informed, carefully crafted foundation model, thank you very much, I'm not interested in AI agents." But we're at a point where basically 100% of our team, across all these domains, this is how we do our work. We watch them go through a not-terribly-long tunnel from "okay, fine, I've heard you talking about this" to "oh, wow." I heard from someone I considered fairly grumpy and quite serious, who emailed to say he'd just gotten a transformational result he was almost done writing up. And we're mostly pretty grumpy, pessimistic people who say we'll believe it when we see it. **Jason Hiner: How does that actually change the work?** **Lizzie Dorfman**: For us it ended up being coding agents using a tree-search methodology, where you explore hundreds or thousands of different approaches to solving a problem computationally. That's what gets you extraordinarily creative and performant solutions. We have an epidemiological forecasting example with top-scoring models in the CDC competitions. We generated 200,000 models to evaluate. Most were poor and got shed immediately. On a traditional team, someone might score a couple of ideas. People describe it as: I input some ideas, went to sleep, and woke up with all these cool results. Before, it was serial and iterative. You'd go down a pathway hoping it was right, and frequently it wouldn't work. Now it's possible to explore outlandish ideas just to see, because it's trivially more expensive to you, certainly in terms of your intellectual time. **Jason Hiner: What are you doing with all that capability?** **Lizzie Dorfman**: I got obsessed with the idea of reproducing papers. We have a reproducibility crisis, partly because some work genuinely can't be reproduced, and partly under-specification, where you just didn't give all the details. So we'll take a paper, extract a short summary of its methods, and ask the system to re-implement something faithful to that description and see if we get comparable results. It's quite good at it. This isn't digging up your old code and resolving legacy dependencies. It's a couple of sentences. That's a real state change. That's part of what we're putting in more people's hands. It's being announced within the [Gemini for Science set of products](https://www.thedeepview.com/articles/google-unfolds-its-vision-for-ai-driven-science). Computational discovery is the product experiment, and ERA is part of the technology that powers it. We've done a lot of academic collaborations, but the real opportunity is getting this into more people's hands. **Jason Hiner: So does this reduce the need for human expertise?** **Lizzie Dorfman**: No. On my team, an agent will work on a single task for days. And one neat thing about our system is that every candidate is just a Python-like notebook; it's code. So we score them all, then use Gemini to annotate the changes in the code that correspond to the performance jumps. You can read a little narrative of what moved performance. But there's an example I love: solar panels are flat, but what if they weren't? I live off-grid on weekends, four miles from the power grid, so I actually have to adjust the angle of my panels. Someone wrote a paper on curved panels in 2012 and moved on with their life. We fed it in, reproduced their results, then asked the system to make it better. It did, and voila. But when we looked, it had these photovoltaic pieces that were levitating, not physically connected. It was cheating because you can maximize energy if you don't have to adhere to physics. So we added a loop that checks the solution is physically valid. You can't just shut your eyes, and it's done. It's impressive how much starting your prompt with "you are an expert scientist in blank" helps. But expertise and careful verification are still required. **Jason Hiner: Is Google Research having genuine scientific breakthroughs?** **Lizzie Dorfman**: Scientific breakthroughs [are] why I get up in the morning. Google started as a research project, so that ethos has been in the DNA from the beginning. But there's a huge amount of value, honestly, that comes from careful, boring science. A variant-detection algorithm is one step of many in producing high-quality data that someone then does something with. I don't want to poo-poo the impactful stuff that comes from working on bottleneck problems, even when they're not the breakthrough problems. That said, it's the grand-challenge problems that motivate the way. In computational neuroscience, we're mapping every cell and connection in the brains of model organisms… [For example], the next major organism is the larval zebrafish. It's transparent, so you can image its brain while it's alive, and modify it genetically so cells flash when they fire. You can do genetic modification … so that you can actually track the electrical activity of the cells in the brain in response to environmental stimuli. So you can create these loops where, from stimulus to behavior, what happened at a cellular level? … And at the end of that tunnel is understanding how a human brain works, and really understanding structurally and functionally how a human brain works. **Jason Hiner: When you take on those grand challenges, how do the breakthroughs actually come?** **Lizzie Dorfman**: There's a phase where it's "this is impossible, it's infeasible." We could map a human brain today, but the microscope time and error correction put the price tag in the hundreds of billions. So you get a methods-development phase, like the Human Genome Project, the Large Hadron Collider, the James Webb telescope, where you ask what you need to solve to get there. For brain mapping, we need to bring the computational cost down by multiple orders of magnitude. The breakthrough takes a lot of forms. It's not always something that looks like AlphaFold, where you've got 40 years of protein crystallography, the CASP competition, and then, check, done. It's frequently "oh, wow, another order of magnitude." Long-read sequencing had a persistent small-variant error problem that kept it out of the clinic, so we partnered with PacBio and built an error-correction algorithm on the instrument, called DeepConsensus, that pushed the error rate down and throughput up. One big breakthrough is often a series of obstacles overcome. In the last four years we've had something like 45 papers in Nature and Science. That's results. It's not "hey, blog post, we did a thing." **Jason Hiner: Neural networks are named after the brain. Does your team's brain-mapping work feed back into the AI models?** **Lizzie Dorfman**: The first artificial neural networks were directly inspired by our then-understanding of neurons, and they pretty immediately diverged technically. But the brain continues to be a huge source of inspiration. Consider that your brain runs on roughly a light bulb's worth of power and can do things a supercomputer can't. These fields have a bidirectional relationship: we use AI to understand neuroscience, and if we can better understand our brains, that helps motivate directions to push the frontiers of AI, which maybe helps us do even more to understand how our brains work.