AI summary by TwinMind
Has Protein Folding Been Solved?
Sabine Hossenfelder
12m 1s
Introduction to Protein Folding Problem
- Protein folding = biggest problem in biochemistry, holy grail of drug development
- Headlines proclaimed AI solved protein folding last year
- Question: was it really solved and what does this mean?
What Are Proteins and Why They Matter
- Proteins = major building blocks of living tissue (e.g., muscles)
- Bewildering number of variants and functions throughout biology
- Key functions:
- Antibodies fighting infections
- Organ communication
- Tissue repair
- Complex molecular pumping in/out of cells
- Substance transport using walking-like motions
Protein Structure and Folding Mechanics
- Proteins = chains of smaller molecules called amino acids
- Long loose chains unstable → fold and curl into stable 3D shape
- Protein folding problem = determining stable shape(s) of proteins
- Function depends on shape
- Mutations → amino acid sequence changes → wrong folding → loss of function → severe illness
Diseases Caused by Misfolded Proteins
- Type 2 diabetes
- Alzheimer's
- Parkinson's
- ALS (Stephen Hawking's disease)
- Understanding folding essential for disease mechanisms and potential cures
Benefits of Understanding Protein Folding
- Easier synthetic protein production with desired functions
- Time and cost savings in research
- Enable development of specialized proteins
The Computational Challenge
- Hideously difficult problem due to exponential possibilities
- Amino acid chains long with many folding directions
- Cyrus Leventhal (1960s): typical protein could fold in >10^140 ways
- Number depends on protein size:
- Small proteins: ~10^50 possible foldings
- Large proteins: mind-blowing 10^300 possible foldings
- Trying all possibilities clearly not feasible
Key Scientific Discovery (1950s-1972)
- Scientists found protein shape determined by amino acid sequence under controlled conditions
- Usually true in natural environment too
- Nobel Prize for Chemistry awarded 1972 for this discovery
- Natural selection preferentially used molecules that reliably fold same way
- Exceptions exist:
- Prions (mad cow disease) have several stable shapes
- Proteins can change shape when environment changes
Approaches to Solving Protein Folding
- Two main approaches:
- Top-down: model why proteins fold one way vs another
- Bottom-up: observe large numbers of protein folding, extract regularities
- Both require examples of folded protein structures
Experimental Methods for Protein Structure
- X-ray crystallography:
- Fire X-ray beams at crystallized proteins
- Measure ray scattering patterns
- Infer 3D shape from atomic positions
- Problem: some proteins take months/years to crystallize
- Cryo-electron microscopy:
- Uses electron microscopy on deep frozen proteins
- Recently much improved situation
- Gives much better resolution
CASP Competition Overview
- Founded 1994 to track protein folding prediction progress
- Critical Assessment of Protein Structure Prediction (CASP)
- Competition among research teams every 2 years
- Teams given amino acid sequences, predict folding shapes
- Uses experimentally measured but unpublished protein structures
- Predictions compared with real shapes using Global Distance Test (GDT)
- GDT scoring: 0% = total failure, 100% = perfect score
- Each team gets average score across all predictions
CASP Competition Progress
- First 20 years: slow progress
- No one scored >40% until recent years
- AI/deep learning teams emerged (about half of teams in recent competition)
- Deep learning = neural networks trained on large datasets to recognize patterns
DeepMind's AlphaFold Breakthrough
- DeepMind = British company acquired by Google 2014, now part of Alphabet
- Same company behind AlphaGo (first to beat professional Go player 2015)
- AlphaFold protein folding program results:
- 2018: AlphaFold scored ~60% in CASP
- 2020: AlphaFold 2 reached ~90%
- Made big headlines, many outlets claimed protein folding "solved"
Criticisms and Limitations
- 90% still significant failure rate
- Struggles with most interesting cases:
- Protein complexes (oligomers)
- Cases where several amino acids interact
- General AI limitation: can only learn patterns from training data
- If entirely new functions don't appear in dataset, may remain undiscovered
- Still needs experimental data/measurements to train AI
Future Implications and Applications
- Remarkable achievement and amazing progress
- May save scientists time-consuming and expensive experiments
- Could help develop proteins with specific functions
- Wish list applications:
- Proteins stimulating immune system to fight cancer
- Universal flu vaccine
- Proteins that break down plastics
