2024 Nobel Prize in Chemistry(2)
Reason for Award
Development of programs for protein structure prediction
Laureates
United Kingdom of Great Britain and Northern Ireland
United States of America
Explanation
Proteins are like long strings folded into intricate 3-D puzzles. Knowing their shapes is a key to understanding how they work, like matching a key to its lock. Dr. Hassabis and Dr. Jumper built an artificial-intelligence system that can solve this puzzle in moments. The AI learned from many known protein shapes and can now predict the shape of new proteins. Scientists use this tool every day to design medicines and therapies much faster.
Related Keywords
AlphaFold
An AI system for protein structure prediction developed by DeepMind. The first version boosted accuracy at CASP13 in 2018; AlphaFold2 reached experimental resolution in 2020. The model is transformer-based and end-to-end, outputting coordinates directly from sequence. Both code and predicted-structure databases are open access and are widely used by researchers. AlphaFold represents a watershed moment for AI applications in life science.
transformer
A neural-network architecture originating in NLP that uses self-attention to capture long-range dependencies efficiently. Protein sequences can be treated as 1-D strings, making transformers ideal for extracting co-evolutionary patterns. Adopted in AlphaFold2, ProteinBERT, and numerous bio-AI models. Suited to parallel computation, enabling large-scale training on GPUs/TPUs. Applications now include structure prediction, sequence generation, and function annotation.
multiple sequence alignment
A technique that lines up evolutionarily related protein sequences, revealing conserved residues and covariation patterns. In structure prediction, mutual information serves as a strong cue for residue contacts. AlphaFold2 feeds the MSA directly into the network and captures common features via self-attention. Accuracy drops for orphan sequences with shallow MSAs, so metagenomic data are used to improve coverage. Expanding sequence databases continues to boost prediction performance.
pLDDT
A per-residue confidence metric output by AlphaFold2, ranging from 0 to 100. Scores above 90 indicate high accuracy, while scores below 70 signal uncertainty. Color-coded visualization lets users see trustworthy and unreliable regions instantly. Benchmarks show good correlation between pLDDT and experimental accuracy. It is widely used as a filter in database searches and functional analyses.
CASP
The Critical Assessment of Protein Structure Prediction is a biennial international contest where participants predict unpublished structures and are objectively ranked. Its independent evaluation has driven advances in the field. AlphaFold2’s historic win at CASP14 was seen as solving the single-chain prediction problem. New categories now cover dynamics and complexes, spurring next-generation methods. Results are publicly released and serve as benchmarks for researchers.
conformational space sampling
The computational process of exploring all possible 3-D arrangements (conformations) of a protein chain. Traditional methods sampled only a fraction via Monte-Carlo or molecular dynamics, but AlphaFold2’s trained network directly estimates high-probability regions, greatly boosting efficiency. New generative and diffusion models can output multiple low-energy states simultaneously, capturing conformational diversity. Applications include drug binding and allosteric regulation studies.
AlphaFold-Multimer
An extension of AlphaFold for complex prediction, ingesting concatenated MSAs and inter-sequence information. Interface attention infers inter-subunit contacts, delivering higher DockQ scores than previous methods. pTM and interface-pTM provide confidence for overall and interface quality. Designed for large-scale use, it is accessible via ColabFold with shared GPUs. Applications are growing in complement formation and signaling complexes important to drug discovery.
RoseTTAFold
A three-track neural network from the University of Washington that simultaneously processes sequence, 2-D distance maps, and 3-D coordinates. Its lightweight architecture differs from AlphaFold, offering faster predictions and easy integration with design workflows. While the initial version handled single chains, the latest All-Atom edition addresses ligand-containing complexes. Open-source distribution leads to broad adoption and ensemble use for accuracy gains. Lightweight models are also supplied for educational purposes.
biopharmaceutical development
With structure prediction AI, detailed 3-D information on targets and antibodies is obtained instantly, speeding pocket design and mutation scanning. Using AlphaFold pLDDT to identify stabilizing loop mutations has enhanced antibody affinity in reported cases. Drug-discovery costs and timelines shrink, accelerating medicines for rare diseases and pandemic response. Regulators are drafting guidelines for reviewing in-silico data. The technology underpins the emerging era of AI-driven therapeutics.