DeepMind solves protein folding | AlphaFold 2

AlphaFold 2 Breakthrough and Performance

• DeepMind's AlphaFold 2 has solved the 50-year-old "grand challenge" of protein folding, achieving prediction performance comparable to slow, expensive experimental methods like x-ray crystallography.
• In the 2018 CASP competition, AlphaFold 2 scored 87 on the hardest protein class, a significant improvement from its 2018 score of 58.
• AlphaFold 2 outperformed the closest competition by 26 points, demonstrating a massive leap in accuracy.
• The speaker posits this as the most significant advancement in structural biology in the past decade and a major milestone in AI history comparable to the ImageNet/AlexNet moment.

Technical Methodology and Evolution

• While the full details are not yet in a published paper, the system shifts from the AlphaFold 1 architecture which used convolutional neural networks (CNNs) to a model heavily utilizing attention mechanisms and transformers.
• AlphaFold 1 Architecture:
  • Step 1: A CNN took amino acid sequences and multiple sequence alignment (MSA) as features to output a distance matrix.
  • Step 2: A non-learning gradient descent optimization calculated the 3D structure based on the distance matrix.
• AlphaFold 2 Architectural Shifts:
  • MSA is now integrated directly into the learning process rather than serving solely as pre-engineered features.
  • The system employs an iterative process where learned information is constantly passed between sequence representations and residue-to-residue distance representations.
  • The architecture appears to utilize a "spatial graph" representation rather than a simple distance or adjacency matrix.

Biological Context and Computational Challenges

• Proteins are chains of 21 amino acids that serve as the structural and functional workhorses of cells; their 3D structure determines their function.
• A single amino acid sequence typically maps one-to-one to a specific 3D structure, though misfolding is a primary cause of many diseases.
• The Computational Scale:
  • The number of possible protein folding configurations is estimated at $10^{143}$, vastly exceeding the complexity of chess ($10^{100}$).
  • Current databases contain 200 million mapped proteins but only 170,000 determined 3D structures due to the limitations of experimental methods.
  • Experimental determination via x-ray crystallography costs approximately $120,000 per protein and takes about one year.
• AlphaFold 2 aims to generate accurate 3D structures for millions of proteins, offering a several-orders-of-magnitude increase in available structural data.

Future Implications and Predictions

• Scientific Recognition: The speaker predicts at least one, potentially several, Nobel Prizes in medicine, physiology, chemistry, or physics will result from work using these computational methods.
• Medical Applications:
  • Understanding the causes of diseases rooted in protein misfolding.
  • Designing new proteins to correct misfolded proteins or alter the function of other proteins for drug development.
• Broader Industry Applications:
  • Agriculture: Engineering insecticidal proteins or frost-protective coatings.
  • Materials Science: Tissue regeneration, anti-aging supplements, and biomaterials for textiles.
• Long-Term AI Trajectory:
  • Transition from single-protein prediction to end-to-end learning of complex biological systems, such as multi-protein interactions and protein complexes.
  • Future potential for physics-based simulations of cells, organs, and entire biological organisms.
  • Expansion of machine learning from game-playing (StarCraft, Go) to real-world biological and physical systems.