Transcript

Hello, I'm Saz from the Google DeepMind science team, and I'll be sharing how we use science skills in Antigravity to accelerate the way that we do scientific research. Biology today has a workflow bottleneck. We have oceans of data and amazing predictive models, but moving from an initial observation to a physical experiment means navigating a fragmented, manual and sometimes slow process. Today, I'm going to show you how scientific agents can accelerate researchers and help us get to the bench faster. To get started I've opened my research project within Antigravity, where I've already integrated Science Skills. We're going to investigate a genetic variant. We aren't trying to magically discover a cure today. Our goal is to demonstrate how we can take an initial finding and use a variety of tools and databases to rapidly build testable hypotheses and generate the blueprints for wet lab reagents needed to test them at the bench. We'll be trying to understand the molecular consequence of this variant and how it may relate to disease. First, the agent queries standard databases and tools. It identifies that this is a missense mutation in the AK2 gene, but also flags another possible mechanism for us to investigate. Because we integrate state-of-the-art predictive models directly into this workspace, we can easily diligence this alternative before we commit time and resources in the lab. The alternative mechanism relates to splicing. So I'll ask the agent to use AlphaGenome, an AI model designed to understand genetic variants. The agent runs the model, outputs visualizations, and predicts a severe splicing defect. When we dive into the results, it looks like this variant might actually disrupt how the cell processes RNA rather than changing a single amino acid. This gives us a second distinct hypothesis to test. To understand what this predicted RNA processing error might mean for the actual protein, we need to connect our genomic data to protein data, and I can ask the agent directly to bridge that gap. The agent predicts that the cellular machinery would likely misread a splice site, eventually causing a frameshift. To see the severity of this change, I can have the agent pull the AlphaFold structure and highlight the affected region. The agent retrieves the structure and renders it. As we can see, the affected region forms a significant part of the protein. While this is just a prediction, it provides a strong structural rationale for how this variant may act. We now have enough context to warrant physical testing, and this is where the digital workflow meets the real world. Models and structures are great, but truth is determined in the wet lab. To transition to the bench, we need the exact genetic blueprints, the DNA sequences required to physically build and test both of our hypotheses. And there we go. Instead of manually looking at the transcripts and designing the fragment, the agent handles the sequence design. It used to take me an entire afternoon to do workflows like this. And now we go from an observation to hypothesis, and designing the blueprints needed to decipher the mechanism in the real world in a matter of minutes.