Over the past two weeks we\u2019ve seen a flurry of new results from Google Research, from genomics to quantum computing to geospatial understanding.<\/p>\n
These breakthroughs exemplify what I call the \u201cmagic cycle\u201d of research<\/a>: addressing global challenges and opportunities with foundational research that leads directly to real-world applications and solutions. These solutions not only benefit millions around the world, but they also uncover more important problems to solve.<\/p>\n The magic cycle is accelerating significantly, powered by stronger models, and agentic tools, like the AI co-scientist<\/a> and AI-based expert-level empirical software<\/a>. And it applies across many disciplines and domains<\/a>.<\/p>\n Research is our chance \u2014 and imperative \u2014 to improve day-to-day lives, and address societal challenges and opportunities, and it means that research and innovation is never \u201cdone.\u201d<\/p>\n Here\u2019s how Google Research\u2019s approach is helping solve challenges in three areas that affect so many people:<\/p>\n 1. Fighting cancer with AI<\/p>\n Childhood leukemias and many other cancers have incredibly complex genetic signatures, requiring tailored treatments based on their specific mutations. What if we could sequence the genomes of these cancerous cells more accurately, precisely spotting the particular variants that turned them into cancer?<\/p>\n That\u2019s what led to DeepSomatic<\/a>, our new AI-powered tool that helps scientists and doctors spot genetic variants in cancer cells. Our partners at Children\u2019s Mercy in Kansas City used DeepSomatic to identify 10 new genetic variants in samples of childhood leukemia that were missed by previous techniques. If they can pinpoint how and why a particular form of cancer is affecting a patient, they may be able to develop personalized cures.<\/p>\n Remarkably, DeepSomatic can also generalize to cancers it hasn\u2019t seen before. For example, without any training on the brain cancer glioblastoma, DeepSomatic was able to pinpoint which genetic variants cause it. This suggests it could work even on rare or new types of cancer \u2014 a big milestone that marks 10 years of genomics research at Google<\/a>.<\/p>\n In collaboration with Yale and Google DeepMind, we also introduced Cell2Sentence-Scale 27B<\/a>, a new 27-billion-parameter AI model based on Gemma that understands the language of individual cells. It generated a novel hypothesis for cancer therapy, which we validated in living cells, finding a drug combo that made cancer cells significantly more visible to the immune system in a lab setting. It\u2019s a powerful new way we\u2019re using AI to help fight cancer.<\/p>\n 2. Towards better medicines and materials with quantum computing<\/p>\n Designing better medicines and materials \u2014 like a more effective battery \u2014 requires understanding the exact behavior of atoms and molecules. But today\u2019s most powerful classical computers struggle to model these nuances because they rely on approximations and the strict binary language of 0s and 1s, and even the world\u2019s most powerful supercomputer can\u2019t capture all the nuances of how molecules behave in the wild. That\u2019s because at this tiny scale, particles don\u2019t behave \u201cclassically.\u201d Instead, they obey quantum mechanics: they can be in superposition, where they\u2019re not in one simple state, but instead \u201csmeared\u201d across a range of possibilities; and they can be entangled, where multiple atoms can behave in lockstep with each other instead of independently.<\/p>\n This is one of the most compelling reasons Google Research is building a quantum computer<\/a> \u2014 it \u201cspeaks quantum\u201d in a way that no classical computer can, and can model exactly how nature really works at a subatomic level. Our new Quantum Echoes algorithm<\/a> shows how much faster our Willow chip can be on computations that are very useful for describing how molecules behave with full precision. This is the world\u2019s first algorithm that points towards eventual practical applications of quantum computing, like designing better materials, better medicines and more. We\u2019ve begun exploring its potential in experiments in partnership with researchers at the University of California, Berkeley.<\/p>\n 3. Understanding Earth<\/p>\n The hardest and most important questions in planetary science and crisis response are never about just one kind of geospatial information \u2014 they\u2019re about pulling it all together. For example, if we want to predict which communities are most vulnerable and what infrastructure is at risk with an impending storm, it\u2019s not enough to know simply when the storm will hit, or where the buildings are. We need the whole picture: the storm’s path and severity, population density, transportation patterns, and predicted impact on vulnerable infrastructure. This comprehensive view requires synthesizing many types of geospatial data, and many models that predict different aspects of the planet \u2014 all at once.<\/p>\n