AI summary by TwinMind
How AI Is Saving Billions of Years of Human Research Time | Max Jaderberg | TED
TED
17m 19s
Personal Background and AlphaFold Impact
- Speaker completed PhD in neural networks/CUDA kernels (computer-based science)
- Friend worked in experimental protein structure determination (messy lab science)
- Single protein structure required entire PhD's worth of work
- AlphaFold by DeepMind revolutionized protein folding (50-year challenge)
- Won Nobel Prize in Chemistry (2 weeks before talk)
- Estimated 1 billion years of research time saved since release
- PhD-level work now completed in seconds of neural network time
New AI Paradigm for Scientific Discovery
- Disagreement with "one-time event" perspective on AI breakthroughs
- Four key enablers for continued AI scientific breakthroughs:
- Neural network architectures handle any data modality
- Tried/tested recipes for incorporating signals into learning algorithms
- Engineering/infrastructure for scaling models to needed size
- Massive compute power availability
- Continuous improvement in recording/measuring real world → bigger datasets → richer models
AI Analogs Concept
- New paradigm: creating AI analogs of real messy world
- Neural networks learn to recreate world elements
- Power beyond understanding/simulation: provides virtual experimentation environment
- Enables in silico open-ended discovery loops
- Agents probe AI analogs to create new knowledge for real-world application
Drug Design Challenge
- Exponentially decreasing drug creation efficiency
- Graph shows new drugs per billion dollars R&D spending declining over time
- Becoming more expensive to create new drugs
- Simultaneous exponential increase in AI capabilities
- Moore's Law → Jensen's Law (Jensen Huang, Nvidia CEO)
- Exponential increase in GPU FLOPS powering neural networks
- Goal: reverse Eroom's law curse using AI analogs
Biology vs Physics Modeling
- Physics: can write mathematical theory, predict universe millions of light years away
- Biology: cannot write equations for complex cellular dynamics
- Can describe atomic interactions (physics level)
- Cannot simulate trillions of atoms in cells (computationally unfeasible)
- Lack coarser mathematical descriptions of complex dynamics
- Paradox: model distant universe but not fingertip cells
- AI/ML provides perfect abstraction for biological world
- Learn equations/theories/abstractions implicitly in neural network activations
Isomorphic Labs Mission
- Company name reflects belief in fundamental isomorphism/symmetry
- Between biological world and information science/ML/AI
Protein Structure and Function
- Proteins: fundamental building blocks of life
- Different functions in body; modulating function = drug creation pathway
- Composition: sequence of ~20 amino acids (depicted by letters)
- Amino acids = collections of atoms (molecules) joined in linear sequence
- Function depends on:
- Amino acid sequence
- 3D folded shape
- Thousands of proteins with unique sequences/3D shapes
- Experimental 3D structure determination: months to years of lab work
AlphaFold Evolution
- AlphaFold 2 (2020): sequence input → 3D structure prediction
- Fills gaps in known protein universe (AI analog of proteins)
- AlphaFold 3 (current year): models structure of all biomolecules together
- Unprecedented accuracy
- Input: protein sequence, DNA sequence, molecule atoms
- Neural network with transformer-based processing trunk
- "Pairformer" operates on 2D interaction grid (vs 1D sequences in LLMs)
- Explicit reasoning about every pairwise interaction
- Diffusion model conditioned on processing trunk features
- Diffuses 3D atom coordinates (vs pixels in image generation)
Drug Design Application Example
- DNA repair protein example:
- Protein clamps down on DNA, facilitates repair, releases repaired DNA
- Cancer treatment strategy: stop DNA repair protein function
- Cancerous cells lack backup DNA repair mechanism
- Stopping primary mechanism kills cancer cells, spares healthy cells
- Drug molecule design:
- Molecule enters body/cell, sticks to protein
- Glues DNA repair protein clamp shut
- Prevents effective DNA repair → cancer cell death
AI-Powered Drug Design Process
- AI analog provides malleable virtual biomolecular world
- Input changes → output structure changes observable
- Human drug designers play "game":
- Modify molecule design to increase protein interactions
- Stronger binding to target protein
- Traditional process: months for lab results per design iteration
- AI analog process: seconds per iteration
- Current reality: London-based drug designers using this system
Agent-Based Drug Design
- Extension beyond human designers
- Speaker's background: agents beating human professionals (StarCraft, Go, Capture the Flag)
- AI analog becomes game environment for training agents
- Powerful agents already operational
Scaling Potential
- Multiple computers → thousands of parallel molecule design agents
- Impact scenarios:
- Rare cancer patients: faster molecule development
- Multiple diseases tackled simultaneously
- Personalized medicine for individual mutations
- Cancer mutation complexity:
- Proteins mutations cause cancer
- Same cancer type = different mutations per patient
- One drug molecule ≠ all patients
- Personalized approach:
- Measure individual patient protein mutations
- Team of molecule-design agents per patient
- Tailored molecules for individual patients
- Demonstration: protein randomly mutating, red mutations change 3D shape, molecules generated responsive to changes
Broader Applications and Future
- Paradigm extends beyond drug design:
- Material science
- New energy forms
- Chemistry
- Core concept: real messy world → AI analogs → computer-based open-ended scientific discovery → new knowledge → world transformation
- Powerful paradigm for scientific/technological advancement wave
- Need maximum participation from ML/AI/technology professionals
