The Year of the AI Nobel

2024 is being hailed as the “Year of the AI Nobel,” marking a milestone as artificial intelligence (AI) researchers received Nobel Prizes across both Physics and Chemistry. Pioneers Geoffrey Hinton and John Hopfield were recognized for their foundational work on neural networks, winning the Nobel Prize in Physics. In Chemistry, the Nobel went to David Baker, as well as to Demis Hassabis and John M. Jumper, with the latter two being members of the team behind AlphaFold, Google DeepMind’s breakthrough application of attention mechanisms in protein folding prediction. 

Apart from the celebration of individual achievements, these awards reflect the evolution of AI from purely theoretical constructs to powerful tools driving scientific discovery. This month in ScaiDigest, let’s take a closer look at the research behind this landmark year in AI, examining the original paper introducing AlphaFold2 – 

Developed by researchers at Google DeepMind, AlphaFold2 has reshaped our approach to predicting protein structures, achieving near-experimental accuracy from amino acid sequences alone. This breakthrough is pivotal for biology, as protein folding plays a central role in understanding cellular functions and disease mechanisms, with implications across drug discovery, enzyme design, and synthetic biology.

The protein-folding problem has long been a formidable challenge in science. Experimental methods, while highly accurate, are slow and expensive, leaving much of the proteome unexplored. By applying attention mechanisms, AlphaFold2 can capture intricate relationships between amino acids, learning how they influence each other’s positions even when far apart in sequence.

AlphaFold2 uses what the authors refer to as a Evoformer module, which processes large multiple sequence alignments (MSAs) from homologous proteins, thereby capturing co-evolutionary patterns and sequence dependencies. The Evoformer’s attention layers let AlphaFold2 interpret these alignments, identifying which residues are likely to interact in the folded structure. These pairwise interactions are encoded as a pair representation, where geometric relationships between amino acids are mapped, helping AlphaFold2 make highly informed spatial predictions that get iteratively refined across multiple passes, or “recycling.”

This iterative approach means that AlphaFold2 can reach extremely high accuracy, even uncovering structural details at the atomic level. In fact, AlphaFold2 excelled in the Critical Assessment of protein Structure Prediction (CASP) competition, where it rivaled traditional experimental methods in accuracy.

Now, with over 200 million protein structures publicly available through the AlphaFold Protein Structure Database, scientists have an unprecedented resource for research. From exploring new drug targets to studying misfolded proteins linked to diseases like Alzheimer’s, AlphaFold2 has opened the doors for faster, cost-effective structural biology. Its success not only underscores the adaptability of attention-based neural networks across domains but also highlights AI’s potential to solve hard challenges in biology.

Jumper, J., Evans, R., Pritzel, A. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021).