AI and Scientific Discovery Lecture Overview

Accelerating Scientific Discovery with AI

• Demis Hassabis, Google DeepMind, presented a Nobel Lecture on December 8, 2024.
• The lecture focused on using AI to accelerate scientific discovery.

Thinking about Thinking

• Images of children playing chess and early computers were displayed.
• This section emphasized the importance of understanding thought processes.

Games AI as a Critical Stepping Stone

• DeepMind was founded in 2010 as a research lab to develop Artificial General Intelligence (AGI).
• Initially, games served as a crucial testing ground for AI development.
• Go was identified as the most complex game, representing a significant challenge for AI systems.
• There are 10¹⁷⁰ possible positions in Go, far exceeding the atoms in the universe.
• In 2016, AlphaGo defeated Go master Lee Sedol 4-1 in a pivotal match.
• AlphaGo's success demonstrated novel strategic thinking.
• AlphaGo learned through self-play, gaining insights and improving its strategies.

Learning a Neural Network Model to Efficiently Guide the Search

• Diagrams illustrated a search tree approach to problem-solving.
• A neural network model was used to effectively guide the search process in complex scenarios, such as the game Go.
• The neural network learns to value different potential moves in the game.
• This allowed for a more efficient search process by prioritizing promising paths.

What Makes for a Suitable Problem for AI?

• Three crucial aspects for AI application were highlighted:
  • Massive combinatorial search space.
  • A clear objective function.
  • Lots of data and/or a precise and efficient simulator.

Proteins are the Building Blocks of Life

• This section introduced proteins as fundamental components of living organisms.

The Protein Folding Problem

• The challenge is predicting the complex 3D structure of a protein from its simpler 1D amino acid sequence.
• The thermodynamic hypothesis posits that proteins fold to minimize free energy.
• Christian Anfinsen's Nobel Prize-winning work in this area was mentioned.

The Protein Folding Problem: a 50-Year Grand Challenge in Biology

• Experimental protein structure determination is a time-consuming task.
• Levinthal's Paradox illustrates the immense number of possible protein conformations ( ~10³⁰⁰ ) to consider.
• Nature efficiently folds proteins within milliseconds, suggesting underlying principles.

Data and Benchmarks

• The Protein Data Bank (PDB) provides a substantial dataset of protein structures.
• The Critical Assessment of Protein Structure Prediction (CASP) competition is a benchmark for evaluating computational methods and improving protein structure prediction systems.
• CASP competition evaluates computational systems for accuracy in protein structure prediction, focusing on atomic-level (<1.0Å) precision.
• The challenge of accurately predicting protein structures at this level highlights the difficulty of the problem and the need for further development.

AlphaFold2 Achieved Atomic Accuracy at CASP14 (2020)

• AlphaFold2 demonstrated a remarkable improvement in protein structure prediction.
• AlphaFold2 achieved high accuracy in the CASP14 competition.
• The results showed an advancement in the field of protein structure prediction.

Innovate Architecture of AlphaFold2: A Hybrid System

• AlphaFold2's architecture is a hybrid system utilizing multiple methods, such as evolutionary and physical constraints.
• Iterative refinement through a recycling stage improves structure predictions.

AlphaFold's Iterative Steps toward Protein Structure Prediction

• AlphaFold uses an iterative approach for protein structure prediction.

Impact of AlphaFold so far

• AlphaFold has revolutionized the field.
• It's available to the scientific community and used worldwide.

AlphaFold is accelerating progress on a huge range of problems

• AlphaFold is being applied to numerous challenges, from plastic pollution to neglected diseases, drug delivery, and more.

AlphaFold 3

• AlphaFold 3's aim is to model other biomolecules in addition to proteins.
• Advanced tools can help us analyze complex interactions between biological molecules.

Implications for the Bigger Picture

• AI's impact on biology has started a new era.
• AI is transforming how we solve biological and scientific challenges.

Making Search Tractable

• Finding optimal solutions in complex problems benefits greatly from learning models.
• Applicable in a variety of fields.

Finding the best Go move / Finding the best molecule in chemical space

• Diagrams illustrate how AI models can guide the search for solutions (Go vs Chemical).

The new era of "digital biology"

• AI is providing a new language for biological processes.
• AlphaFold enables this new era by enabling the rapid and accurate prediction of protein structures.
• Isomorphic Labs seeks to reimagine the drug discovery process with AI.

What are the limits of classical systems?

• Classic Turing machines can tackle issues thought previously impossible with sufficient computational power.
• This approach provides efficient solutions in polynomial time.
• Proposed conjecture: patterns naturally occurring can be efficiently discovered and modeled with classical learning.
• Implication for complexities: potentially significant if consistent with other findings.

AI for Science, Medicine & Climate

• AI is making significant advancements across different fields.
• Examples include identifying eye disease, accelerating the discovery of new materials, and predicting weather patterns.

Advancing AI Responsibly

• Emphasizes the responsible development and use of AI.
• The need for stakeholder engagement from government, academia, and society is highlighted.

AGI - The Ultimate General-Purpose Tool

• AGI (Artificial General Intelligence) has the capability to help understand our Universe more effectively by serving as a general tool.

Thank you to the incredible AlphaFold team!

• Acknowledgements for the contributors to the successful AlphaFold project.
• Highlights the team's importance in advancing scientific knowledge.