Science in the age of AI

Key Concepts

• AI for Science: Using artificial intelligence to advance scientific discovery and address societal challenges.
• AlphaFold: An AI system that predicts protein structures from their amino acid sequences.
• AlphaFold 3: An extension of AlphaFold that predicts the structures of biomolecules beyond proteins, including interactions with ligands, RNA, and DNA.
• AlphaMissense: An AI model built on AlphaFold that predicts whether missense variants (mutations) in the genome are pathogenic or benign.
• Alpha Proteio: An AI model leveraging AlphaFold 3 to design novel proteins with specific binding capabilities.
• Articulate Medical Intelligence Explorer (Amy): A large language model designed for medical diagnosis.
• Neuro-operators: AI techniques that understand physical systems and create accurate digital twins.
• Co-scientist: An AI system designed to assist scientists in their work, including literature review, hypothesis generation, and critique.
• Weather Next/Weather NextGen: AI models for weather forecasting that are more accurate and faster than conventional methods.
• Neural GCM: Global climate models based on neural networks for weather and climate simulations.
• Fireat: A satellite system for wildfire boundary prediction.
• Synthetic Data Generation: Creating artificial data to train AI models, especially when real-world data is limited.
• Medgemma: An open-weight model fine-tuned on a medical corpus.
• Quantum Computing: Using quantum mechanics to perform computations and solve scientific problems.

AI Advancing Science and Addressing Societal Challenges

Introduction

• Google focuses on AI's impact on individuals, the economy, and science, particularly addressing societal challenges.
• Teams at Google Research and DeepMind are working on hard scientific problems important for society.

AlphaFold and Structural Biology

• AlphaFold as a Milestone: AlphaFold has been a key milestone in demonstrating the potential of AI for science, solving the protein structure prediction problem.
• AlphaFold 2: Predicted the structures of proteins.
• AlphaFold 3: Expanded to predict structures of other biomolecules like protein-ligand complexes, protein complexes, and protein-nucleic acid interactions.
• Applications: Drug discovery, enzyme development for plastic decomposition, and synthetic biology.
• AlphaMissense: Predicts the pathogenicity of missense variants in the genome, impacting rare disease diagnosis.
• Alpha Proteio: Designs novel proteins for specific tasks like binding to targets, with applications in drug discovery.
• Impact: Over 2.5 million biologists in 190+ countries are using AlphaFold.

AI in Healthcare

• Amy (Articulate Medical Intelligence Explorer): A large language model for medical diagnosis, augmenting or replacing physician tasks.
• Specialty Care: Amy can perform diagnostic reasoning in cardiology and oncology, on par with or better than physicians.
• Multimodal Input: Amy can process text, uploaded documents, and potentially audio/visual data.
• Clinical Studies: A study with Beth Israel Medical Center is evaluating Amy's deployment in clinical settings under physician supervision.
• Tuberculosis Diagnosis: AI tools are being used for TB diagnosis, addressing the issue of undiagnosed cases in resource-limited areas.

AI for Understanding Physical Systems

• Neuro-operators: AI techniques for understanding physical systems and creating accurate digital twins.
• Photonic Device Design: AI can design photonic devices in the digital realm, reducing the need for lab experiments.
• Medical Catheter Design: AI optimized a medical catheter design to reduce bacterial contamination by 100-fold, requiring only one lab test.
• Multiscale Phenomena: Neural operators can learn across different scales, capturing complexities in fluid dynamics, material deformation, and quantum effects.

AI in Weather and Climate Science

• Weather Prediction: AI models can produce near-term forecasts more accurately and faster than expensive simulations.
• Democratization of Weather Prediction: Weather prediction can be done on a single chip (GPU or TPU) in minutes.
• Weather NextGen: Tracks cyclones more accurately over longer time scales.
• Neural GCM: Global climate models based on neural networks for short-term weather and long-term climate prediction.
• Wildfire Boundary Prediction: AI is used to predict wildfire boundaries in 27 countries.
• Fireat: A constellation of 53 satellites for high-resolution (5x5 meters) wildfire monitoring.

Changing the Scientific Method

• Co-scientist: An AI system designed to assist scientists in their work.
• Capabilities: Literature review, hypothesis formulation, exploration, and critique.
• Drug Repurposing and Antimicrobial Resistance: Co-scientist can explore new areas and discover insights.
• Validation: Rigorous validation is needed to ensure discoveries are novel and not due to training data contamination.
• Weather and Climate Modeling: Neural operators have achieved good accuracy in predicting typical and extreme weather events.
• Statistical Ensembles: AI enables the creation of larger statistical ensembles for events like hurricanes, improving risk assessment.

Data and New Techniques

• Data Importance: Data and experience are crucial for AI and machine learning.
• Synthetic Data: Used when real-world data is limited, e.g., in fusion research and mathematics.
• Fusion Research: AI models are trained on simulated data to stabilize plasma in tokamaks.
• Alpha Geometry: Used synthetically generated geometry problems to train the model.
• Test-Time RL (Alpha Proof): Solves hard problems by generating simpler variants and learning from the experience.

Responsibility in AI and Science

• Synthetic Data and Representation: Ensuring that training data represents people everywhere.
• Patient Actors: Using patient actors to mimic real-world scenarios and assess AI models responsibly.
• Stakeholder Engagement: Engaging with physicians, ethicists, and hospital centers to address real clinical needs.
• Pan-Genome: Drafting the first reference pan-genome with 47 genomic assemblies to represent the full genomic diversity.

Future Directions

• Physical AI: Expanding the understanding of the physical world with AI at different levels of complexity and scale.
• Approximate Data: Using approximate data to train AI models, refining with more accurate data.
• Integration of Domain-Specific AI Models and Agents: Combining domain-specific AI models with agents that can generate hypotheses and accelerate the scientific process.
• Transformative Abilities in Materials Science: Accelerating the discovery of new materials.
• Medgemma and Community Building: Encouraging the community to build on top of Medgemma to address clinical needs.
• Quantum Computing: Using quantum computing to solve scientific problems, particularly in quantum chemistry.

Lightning Round: Personal Goals for the Next Five Years

• Anema: Building physical AI in all its different levels of complexity and scale (multiscale, multiphysics).
• Joel:
  • Improving earthquake prediction.
  • Increasing the cure rate for cancers to 80-90%.
• Pushmeet:
  • Discovering a room-temperature superconductor.
  • Achieving practical fusion.
  • Solving one of the millennium prize math problems (Riemann Hypothesis).

Conclusion

AI is driving major advances in science and addressing societal challenges across various domains, including biology, healthcare, weather, and materials science. The integration of domain-specific AI models, the development of AI agents, and the use of synthetic data are accelerating scientific discovery. Responsible development and stakeholder engagement are crucial for ensuring that AI benefits everyone. Quantum computing holds promise for solving complex scientific problems in the future.