AI summary by TwinMind
The Most Useful Thing AI Has Ever Done
Veritasium
24m 51s
Historical Context of Protein Structure Problem
- 60+ years: tens of thousands of biologists determined structure of 150,000 proteins
- Team of ~15 researchers determined structure of 200 million proteins in few years
- Breakthrough described as equivalent to Fermat's Last Theorem for biology
- Proteins = machines built for specific purposes (hemoglobin carries oxygen, muscle proteins contract)
Basic Protein Science
- Protein formation: string of amino acids → 20 different side chains determine amino acid type
- Peptide bonds form between amine and carboxy groups
- Electrostatic forces, hydrogen bonds, solvent interactions cause folding into 3D structure
- Shape determines function - must be in correct orientation to work
Early Experimental Methods
- X-ray crystallography: first method to determine protein structure
- John Kendrew (1962 Nobel Prize): 12 years to solve myoglobin structure
- Used whale meat from Peru for large enough crystals
- Structure looked "weird" and complex, not logical/mathematical
- Only ~100 structures resolved in next 2 decades
- Crystallization remains expensive: tens of thousands of dollars per protein
- Protein sequencing costs only ~$100
Theoretical Approach Challenges
- Linus Pauling predicted helices and sheets (secondary structure) from geometry
- Beyond helices/sheets: no reliable patterns found for final structure
- Evolution didn't design proteins systematically - added complexity over time
- Levinthal's paradox: 35 amino acid protein has astronomical folding possibilities
- Computer checking 30,000 configurations/nanosecond would take 200x age of universe
CASP Competition Origins
- Started 1994 by John Moult (University of Maryland)
- Challenge: predict protein structure from amino acid sequence
- Scoring: perfect match = 100, >90 considered solved
- First year: teams couldn't achieve scores >40
- Venue: old wooden chapel in Monterey, California with foot-tapping tradition
Rosetta Algorithm Development
- Created by David Baker (University of Washington)
- Innovation: distributed computing via Rosetta@home
- Screen saver showed protein folding calculations
- Users suggested they could do better than computer
Foldit Game Success
- Baker created video game for protein folding
- 50,000+ gamers solved HIV enzyme structure in 3 weeks
- Results confirmed by X-ray crystallography
- Gamers credited as co-authors on research paper
DeepMind's Entry
- Demis Hassabis: former chess prodigy, played Foldit, founded DeepMind
- AlphaGo made headlines beating Lee Sedol at Go (move 37 shocked champion)
- Week-long hackathon after Korea: trained AI to play Foldit
- Initiated AlphaFold project to solve protein folding
AlphaFold 1 Development
- Standard deep neural network trained on protein databank structures
- Input: amino acid sequence + evolutionary clues
- Evolutionary analysis: sequences conserved across species indicate importance
- Co-evolution: paired mutations identify amino acids close in final structure
- Output: 2D pair representation (not direct 3D structure)
- Bright intersections = close amino acids, dim = distant
- Contains distance and torsion angle information
- Separate algorithm folded amino acid string based on constraints
CASP 13 Results
- AlphaFold 1 clear winner but score of 70 insufficient
- Threshold for "solved" = 90
- Performance plateaued after CASP 8, even went downhill
- Needed breakthrough approach
AlphaFold 2 Architecture
- John Jumper recruited to lead development
- Three key improvements:
- Maximum compute power (Google's tensor processing units)
- Large diverse dataset (same data as AlphaFold 1)
- Better AI algorithms (most important factor)
Transformer Integration
- Used Transformer architecture (T in ChatGPT)
- Attention mechanism: recognizes context in sequential information
- Example: "animal didn't cross street because it was too tired" - "it" refers to animal
- Applied to amino acid sequences instead of sentences
EvoFormer Architecture
- Two towers: Biology Tower (evolutionary info) + Geometry Tower (pair representations)
- Bridge connects towers, conveys biological/geometry clues back and forth
- Biology Tower attention: identifies conserved sequences and co-occurring mutations
- Geometry Tower attention: calculates distances between amino acids
- Triangular retention: applies triangle inequality to amino acid triplets
- 48 iterations of information exchange between towers
Structure Module Innovation
- Picks 3 special atoms per amino acid to define frame
- Predicts translation/rotation to move frames to correct positions
- Doesn't explicitly encode chain constraint - treats as "bag of amino acids"
- Chain formation emerges naturally
- Explains weird non-physical movements in live folding videos
- Output recycled 3+ times through EvoFormer for deeper understanding
CASP 14 Breakthrough
- December 2020: AlphaFold 2 virtual CASP competition
- John Moult email: "performed amazingly well... congratulations"
- Many predictions "virtually indistinguishable" from actual structures
- Finally beat gold standard score of 90
- Solved 50+ year problem
Impact and Applications
- 200 million protein structures revealed in months vs 150,000 over 6 decades
- Advanced research by several decades
- Direct applications:
- Malaria vaccine development
- Breaking down antibiotic resistance enzymes
- Understanding protein mutations in diseases (schizophrenia, cancer)
- Access to proteins in endangered species
- AlphaFold 2 paper: 30,000+ citations
Nobel Prize Recognition
- 2024 Nobel Prize in Chemistry (half each):
- John Jumper & Demis Hassabis: protein structure prediction
- David Baker: designing new proteins from scratch
Protein Design Revolution
- David Baker's RF diffusion: generative AI for protein design
- Training: add random noise to known structures, AI removes noise
- Application: give random noise input, AI creates new protein for specified function
- "Cowboy biochemistry": rapid design and iteration (couple days vs traditional methods)
Synthetic Protein Applications
- Snake antivenom: human-compatible antibodies vs traditional animal-derived
- Vaccine development for cancer, autoimmune diseases
- Environmental applications: capture greenhouse gases, break down plastic, fix methane
- Manufacturing advantages: large quantities, easy transport
Broader AI Impact on Science
- Material Science: DeepMind's GNoME found 2.2 million new crystals
- 400,000+ stable materials for superconductors, batteries
- Speed improvements: 100,000x faster changes fundamental approach
- Scientists rebuild research around newly easy tasks
- Step function changes in scientific capability
Future Implications
- Root problems: solving them unlocks entire branches of discovery
- AI pushing boundaries of human knowledge at unprecedented rate
- Applications: curing diseases, novel materials, environmental restoration
- Benefits will continue for decades even if AI doesn't advance further
- Potential for previously impossible breakthroughs
