AI summary by TwinMind
DeepMind solves protein folding | AlphaFold 2
Lex Fridman
16m 41s
2020 Positive Scientific Breakthroughs
- 2020 brought challenges to civilization but also positive news
- SpaceX launching new era of space exploration
- DeepMind's AlphaFold 2 "solved" 50-year-old protein folding problem
- "Solved" means computational methods achieved prediction performance similar to experimental methods like X-ray crystallography
AlphaFold Performance Improvements
- 2018 CASP competition: AlphaFold 1 scored 58 on hardest protein class
- 2020: AlphaFold 2 scored 87 (huge improvement)
- 26 points better than closest competition
- Definite big leap forward
Significance in AI History
- Potentially biggest advancement in structural biology in 1-2 decades
- Strong case for biggest AI advancement in recent history
- Comparison to other AI breakthroughs:
- ImageNet/AlexNet moment in computer vision
- AlphaZero with reinforcement learning self-play
- GPT-3 transformers in NLP
- Tesla Autopilot deployment in safety-critical systems
- Waymo autonomous driving
- Smart speakers with natural language processing
- Boston Dynamics Spot robot
Future Nobel Prize Predictions
- At least one, potentially several Nobel prizes from derivative work
- Possible first Nobel prize where machine learning system plays major role
- Prize awarded to humans behind system, but computational approach central
- Fields: medicine, physiology, chemistry, physics
Protein Fundamentals
- Amino acids: basic building blocks of life in eukaryotes
- 21 amino acids in humans
- Proteins: chains of amino acids
- Workhorses of living organisms/cells
- Functions: structural, functional, catalysts for chemical reactions, transport
Protein Folding Process
- Goes from amino acid sequence to 3D structure
- Fascinating properties:
- Uniqueness: particular sequence usually maps 1:1 to 3D structure
- 3D structure determines protein function
- Misfolding causes many diseases
- Complexity: 10^143 possible folding combinations vs 10^80 atoms in universe
- Levinthal's paradox: explains difficulty and speed of natural folding
- Much harder than chess (10^100 possible games)
Current Dataset Status
- 200 million proteins mapped
- Only 170,000 protein 3D structures determined
- X-ray crystallography costs: $120,000 per protein, ~1 year per structure
- AlphaFold 2 could provide orders of magnitude more 3D structures
AlphaFold 1 Architecture
- Two-step process:
- Machine learning step: Convolutional neural network
- Input: amino acid residue sequences + features (multiple sequence alignment)
- Output: distance matrix with confidence distributions
- Non-learning step: gradient descent optimization
- Folds 3D structure to match distance matrix
- Machine learning step: Convolutional neural network
AlphaFold 2 Improvements (Speculative)
- No detailed paper available yet
- Key changes:
- Attention mechanisms replacing CNNs
- Transformers instead of convolutional networks
- Multiple sequence alignment now part of learning process
- Iterative information passing between sequence representation and residue distances
- "Spatial graph" terminology vs distance matrix
- Two deep learning lessons applied:
- Attention/transformers provide big boost
- Making more of problem learnable yields benefits
Breakthrough Applications
Medical/Health
- Learn unknown functions of genes encoded in DNA
- Understand disease causes from misfolded proteins
- Design proteins to fix other misfolded proteins
- Drug treatments and therapies
- Supplements for improved health and anti-aging
Other Applications
- Agriculture: insecticidal proteins, frost protective coatings
- Tissue regeneration through self-assembling proteins
- Bio materials for textiles and general materials
Long-term Future Impact
Near-term Extensions
- Multi-protein interaction prediction
- Protein complex formation (much harder problem)
- Environment incorporation into folding models
- Function changes based on environment
Physics-based Simulation Expansion
- Accurate biophysics simulation of biological systems
- Progression from proteins → cells → organs → entire organisms
- End-to-end learning for complex biological systems
Sci-fi Possibilities
- Biological and artificial life simulation
- Organisms (biologically and non-biologically based)
- Real-world systems operating outside game environments
- Similar excitement to Tesla Autopilot and robotics with ML
