Kendrew’s Contributions to Science:
Kendrew’s pioneering work in X-ray crystallography led to the determination of the structure of myoglobin, a protein found in muscle tissue. He was a pioneer in comparing sequences from different species, which laid the foundation for the field of bioinformatics. Kendrew was a superb science leader and organizer, helping to found EMBO and EMBL, and serving as the first director of EMBL.

Challenges in Protein Folding Problem:
The protein folding problem, determining the three-dimensional structure of proteins from their amino acid sequence, has been a grand challenge in molecular biology. Progress in solving the problem has been gradual over several decades.

DeepMind’s Breakthrough in Protein Structure Prediction:
DeepMind, a leading AI company founded by Demis Hassabis, made a significant breakthrough in protein structure prediction using machine learning. Their AlphaFold program topped the annual CASP structure prediction competition by a significant margin, demonstrating its accuracy and scope.

Demis Hassabis’ Background:
Hassabis started as a chess master at the age of 13. He took a break before university to work on designing computer games, some of which were nominated for BAFTA awards. He earned a degree in computer science from Cambridge University and later pursued a PhD in cognitive neuroscience at UCL.

Hassabis’ Research on Memory and Imagination:
Hassabis’ research at UCL focused on understanding the connection between memory and imagination. He published highly cited papers, including one showing the link between amnesia and the ability to imagine new experiences.

DeepMind’s Acquisition by Alphabet and Success in AI:
Hassabis founded DeepMind with two others in 2010. DeepMind became one of the leading AI companies in the world and was acquired by Alphabet, the parent company of Google, in 2014. DeepMind gained publicity for developing AlphaGo, a program that defeated top human players in the complex game of Go, and AlphaZero, which taught itself to play chess, go, and another game without human data.

Hassabis’ Focus on Using AI for Real-World Problems:
Despite his success in games, Hassabis’ long-term goal has been to use AI to address real-world challenges, including scientific and medical problems.

Introduction:
Demis Hassabis discusses his lecture’s theme, “Using AI to Accelerate Scientific Discovery.” He emphasizes the importance of understanding intelligence and using AI to advance science for humanity’s benefit.

Two Approaches to Building AI:
Expert systems: AI systems programmed with direct solutions by a team of engineers. Learning systems: AI systems that learn solutions from first principles through experience or trial and error.

Learning Systems:
DeepMind focuses on learning systems inspired by neuroscience and reinforcement learning. Reinforcement learning involves an agent interacting with an environment, receiving observations, and making decisions to achieve a goal. The agent builds a model of the environment based on observations and uses it to make decisions.

Key Points:
AI has the potential to solve problems that programmers or system designers may not know how to solve. Learning systems have revolutionized AI in the last decade due to their success at scale. Reinforcement learning involves an agent interacting with an environment, making decisions, and learning from its experiences. Agents in reinforcement learning build a model of the environment based on observations and use it to make decisions.

Background:
Reinforcement learning is a powerful technique where an agent learns through experience by interacting with the environment. The agent selects actions, observes their impact, and updates its internal model accordingly.

Reinforcement Learning Overview:
The agent’s goal is to maximize rewards while minimizing punishments. Reinforcement learning algorithms iteratively improve the agent’s ability to choose actions. The agent learns to map states to actions that lead to positive outcomes over time.

Generalization to Artificial General Intelligence:
The underlying concepts of reinforcement learning are crucial for developing artificial general intelligence (AGI). AGI requires the ability to learn and adapt to various environments and goals. Reinforcement learning provides a framework for AGI systems to continually learn and improve.

AlphaGo: A Prominent Example of Reinforcement Learning Success:
AlphaGo, a computer program developed by DeepMind, became the first AI to defeat a human world champion in the ancient game of Go. Go’s complexity and reliance on pattern recognition made it a challenging game for AI. AlphaGo’s victory showcased the potential of reinforcement learning to solve complex problems.

AlphaGo’s Approach to Learning Go:
AlphaGo used a neural network trained through self-play to evaluate positions and select moves. The neural network was iteratively improved through millions of games against itself. The combination of the neural network and Monte Carlo Tree Search enabled AlphaGo to evaluate a vast number of possible moves and identify the most promising ones.

AlphaGo’s Revolutionary Strategies:
AlphaGo’s victory in the million-dollar challenge match against Lee Sedol, the 18-time world champion, shocked both the Go and AI communities. AlphaGo’s strategies included original moves that had never been considered by human players before. Move 37 in game two became legendary for its unexpectedness and its critical role in AlphaGo’s victory.

The Significance of Games for AI Development:
Games provide an ideal training ground for AI methods due to their well-defined rules and structured environments. Games allow AI systems to learn through trial and error without real-world consequences. The complexity and variety of games challenge AI systems to adapt and generalize their learning.

AI for Real-World Problems:
Demis Hassabis emphasizes the potential of AI systems for tackling grand challenges, particularly in science. The focus is on finding problems with suitable criteria: massive combinatorial search spaces, clear objective functions, and sufficient data or accurate simulators.

Protein Folding: A Tractable Challenge:
Protein folding involves predicting the 3D structure of a protein from its amino acid sequence. Anfinsen’s conjecture suggests the determinability of protein structure from the amino acid sequence. Leventhal’s paradox highlights the vast number of possible conformations, making exhaustive sampling impractical. Nature folds proteins spontaneously, posing the question of a computationally tractable solution.

Hassabis’ Journey with Protein Folding:
Hassabis’ early exposure to the protein folding problem as an undergraduate at Cambridge. The Foldit game sparked his interest in citizen science and the intuition of amateur biologists in solving protein structures. Analogies between Go players’ intuition and Foldit players’ ability to make counterintuitive moves.

CASP Competition: A Benchmark for Progress:
CASP is a biennial competition assessing protein folding prediction methods. The competition provides independent benchmarks for internal progress evaluation. Historical progress in CASP was limited until AlphaFold’s participation in 2018.

AlphaFold’s Success in CASP:
AlphaFold won the CASP13 competition in 2018, introducing ML as a core component at scale. AlphaFold significantly improved accuracies, particularly in the free modeling category. AlphaFold2, entered in CASP14 in 2020, achieved atomic accuracy. AlphaFold2 outperformed other systems by a large margin, reaching a 0.96 Angstrom error.

Examples of Accurate Predictions:
AlphaFold2 generates accurate predictions for various proteins, including those from SARS-CoV-2. The side chains are often correctly positioned or close to the correct positions.

The AlphaFold 2 System:
AlphaFold 2 is a highly complex system with 32 component algorithms. It required a multidisciplinary team of machine learning experts, biologists, chemists, and physicists. The system is not end-to-end, but it includes an iterative recycling stage that allows the structure to be refined with each stage.

Advances Over CAS13 System:
AlphaFold 2 introduced several advancements over the CAS13 system. It employs attention-based neural networks to infer the implicit graph structure between residues. The system incorporates evolutionary and physics constraints into the neural network architecture without impacting the learning process.

Challenges in Machine Learning:
AlphaFold 2 highlights the challenge of incorporating prior knowledge into machine learning models without hindering learning.

Development and Resources:
The development of AlphaFold 2 was a major research effort, spanning over five years and involving approximately 20 individuals.

Focus on Applied Problem:
Unlike DeepMind’s usual focus on generality, AlphaFold 2 was tailored specifically to solve the protein folding problem. The team utilized all available knowledge and techniques, including domain knowledge, to achieve the desired outcome.

Generality and Future Steps:
AlphaFold 2’s generality will be explored in future research by identifying the necessary modules for performance. The system’s speed and scalability will be addressed in ongoing work.

Training and Optimization:
AlphaFold2 training focuses on performance, achieving error less than one angstrom, before optimizing system efficiency. Training takes approximately two weeks on eight TPU chips, equivalent to 150 GPUs, with fast inference times on a single GPU.

Broad Access to Predictions:
Instead of setting up a server for individual sequence submission and prediction, AlphaFold2 aimed to fold entire proteomes and make the results publicly available.

Human Proteome Results:
AlphaFold2 achieved the first version of the human proteome prediction, covering 36% of proteins with very high accuracy and 58% with high accuracy. The predictions included insights into unstructured or disordered proteins, expanding the understanding of the human proteome.

Confidence Metrics:
AlphaFold2 outputs a per-residue confidence metric called PLDDT to assess prediction reliability. Three thresholds are used: very high accuracy (>90), high accuracy (>70), and very low confidence (<50). Very low confidence predictions may indicate disorder or unstructured regions. Expansion to Other Organisms:
AlphaFold2 was applied to 20 model organisms and proteomes, including fruit fly, E. coli, mouse, yeast, malaria, and crop species. The project is ongoing, with plans to expand to additional organisms.

AlphaFold2: Open Access and Impact on the Scientific Community:
Demis Hassabis, an expert scholar, emphasizes the importance of open accessibility to AlphaFold2’s structural databases. A collaboration with the European Bioinformatics Institute (EMBL-EBI) ensures integration with existing biological tools and resources. Free and unrestricted access to every structure promotes scientific impact and benefits the research community. Positive feedback from experimentalists highlights the usefulness of AlphaFold2 in their work.

Future Plans for AlphaFold2 and AI in Biology:
Expanding the database to include almost every protein in UniProt, totaling over 100 million proteins. Ongoing work on protein complexes, disordered proteins, point mutations, ligand docking, and protein design. Aiming to understand protein function and dynamics through computational methods.

AI as a Tool for Understanding Biology:
Hassabis draws parallels between information processing in computer science and biology. He suggests that AI may be the ideal tool for understanding complex biological systems. AI could serve as the appropriate description language for biology, similar to mathematics for physics. AlphaFold2 is seen as a proof of concept for this idea, potentially leading to a new era of digital biology.

AlphaFold2 as a Versatile Tool for Scientific Discovery:
Hassabis emphasizes the broader applicability of AI beyond AlphaFold2. Major breakthroughs have been achieved in fields such as quantum chemistry, pure mathematics, medical diagnosis, and material design. Weather forecasting and other topics are also being explored using AI techniques. AI is viewed as a powerful tool for scientific advancement, akin to the Hubble telescope for cosmologists.

Addressing Feedback and Improving AlphaFold2:
Hassabis acknowledges instances where AlphaFold2 makes incorrect predictions, even for well-known protein domains. He compares these errors to “bugs” in the system’s knowledge rather than traditional software bugs. The team plans to create a test set of “dumb things” to identify and address these issues in future versions of AlphaFold. Collaboration with the experimental community is crucial for providing feedback and improving the accuracy of AlphaFold2.

Exploring the Physics of Proteins:
Hassabis discusses ongoing efforts to incorporate physics-based simulations into protein modeling. The goal is to understand the dynamics and interactions of proteins beyond static structure prediction. Combining AlphaFold2’s strengths with physics-based approaches could lead to advancements in understanding binding affinities and other dynamic properties.

The Hybrid Nature of DeepMind:
DeepMind is a unique organization that blends the best elements from academia, startups, and Alphabet resources. This approach combines long-term thinking, blue-sky thinking, energy, focus, organization, and computational power.

Advantages of a Commercial Setting:
Commercial settings like DeepMind offer freedom from short-term profitability concerns, which can hinder long-term research programs. The patience and resources available in a commercial setting allowed for achievements like AlphaFold and AlphaGo. Access to top-level engineers and the creation of research engineer roles enable the scaling and practical implementation of research.

Challenges in Academia:
Academia often struggles to provide permanent positions for top-level engineers, leading to a structural problem in scientific career paths. Traditional lab models may not have the funding to support the research engineer roles crucial for scaling up and implementing research.

Future Directions:
DeepMind is exploring the application of its approaches to general problems and nonlinear physics, including turbulence and shear flow.

Overall:
Demis Hassabis emphasizes the advantages of a commercial setting for advancing scientific research, highlighting DeepMind’s unique hybrid approach and the opportunities it offers for long-term, ambitious research, while acknowledging challenges in academia and exploring future research directions.

AI’s Potential Applications:
AI systems have the potential to revolutionize various fields, including plasma containment, fusion reactors, weather prediction, and climate modeling.

Rainfall Prediction:
DeepMind’s recent paper demonstrates the most accurate rainfall prediction for the next two hours, surpassing sophisticated weather models.

Output Interpretability:
While AI’s black-box nature poses challenges, useful outputs can still be valuable, as seen with AlphaFold’s contributions to protein folding.

Analytic Tools for Understanding AI:
DeepMind has a dedicated research group focused on developing analytic tools to break down and understand the inner workings of AI systems.

Ablation Studies:
Ablation studies conducted on AlphaFold revealed that all parts of the system contributed to its performance, albeit to varying degrees.

Representation Understanding:
Understanding the representations and intermediate representations built by AI systems is a key area of research, lagging behind performance engineering.

Complexity of AI Systems:
AI systems can be just as complex as biological systems, but they offer the advantage of being digital and easier to dissect.

New Analysis Techniques:
Novel analysis techniques are necessary to delve into the black boxes of AI systems and gain insights into their mechanisms.