The one-sentence version
Proteins consist of a string of amino acids that fold into specific three-dimensional shapes, and these shapes determine their function. Predicting the 3D structure from an amino acid sequence has been a "protein folding problem" that has troubled biology for half a century. In 2020, DeepMind's AlphaFold 2 achieved accuracy near experimental determination in the CASP14 international structure prediction competition, essentially solving the single-chain protein structure prediction problem. The 2024 Nobel Prize in Chemistry was awarded to AlphaFold's Demis Hassabis and John Jumper, as well as David Baker, who made pioneering work in computational protein design. Since then, the field has moved from "predicting existing proteins in nature" toward "designing proteins that do not exist in nature."
Regarding Timeliness
This article is compiled based on public information available at the time of writing and has not been cross-checked online item by item.
Layer 1: The Intuition
1.1 A Rope That Knots Itself
Imagine a long rope made of 20 different colored beads. If you throw it into water, it will automatically curl and fold into a specific shape, the same every single time. This shape determines what it can do: whether it acts as a "pincer" to grab a certain molecule, or "scissors" to cut a specific chemical bond.
The question is: by looking only at the sequence of bead colors, can we calculate what shape it will eventually fold into?
1.2 Why It Is So Difficult
- A medium-length protein has a theoretically astronomical number of possible conformations ("Levinthal's paradox");
- Experimental structure determination (X-ray crystallography, NMR, cryo-EM) is expensive and time-consuming; a single structure can take months or even years;
- Before AlphaFold, decades of effort had determined about 200,000 structures, while there are hundreds of millions of known protein sequences.
Layer 2: Milestones
| Time | Event |
|---|---|
| Since 1994 | CASP Competition: Held every two years, using unpublished experimental structures to blind test various prediction methods |
| 2018 | AlphaFold 1 ranked first in CASP13, but its accuracy was still far from practical use |
| 2020 | AlphaFold 2 achieved a median accuracy (GDT) of approximately 92 in CASP14, approaching experimental accuracy, and is widely considered to have essentially solved single-chain structure prediction |
| 2021 | AlphaFold 2 paper published in Nature and open-sourced; the same year, Baker lab at University of Washington released RoseTTAFold; DeepMind and EMBL established the AlphaFold Protein Structure Database |
| 2022 | The database expanded to approximately 200 million predicted structures, covering almost all known protein sequences |
| 2022–2023 | Meta's ESMFold uses protein language models to predict structures directly from a single sequence, with faster speed |
| 2023 | Baker lab's RFdiffusion: Uses diffusion models to design protein backbones from scratch |
| 2024 | AlphaFold 3 is released, capable of predicting complexes of proteins with DNA, RNA, small molecule ligands, ions, etc.; Nobel Prize in Chemistry awarded to Baker (computational protein design), Hassabis, and Jumper (protein structure prediction) |
Layer 3: Principles, Impacts, and Limitations
3.1 Key ideas of AlphaFold 2
| Component | Role |
|---|---|
| Multiple Sequence Alignment (MSA) | Collects a large number of sequences homologous to the target sequence. If two positions co-evolve, they are likely to be close to each other in the 3D structure (co-evolutionary signal) |
| Evoformer | An attention network that repeatedly exchanges information between "sequence × sequence" and "residue × residue" representations |
| Structure Module | Directly outputs the position and orientation of each amino acid in 3D space, designed to be equivariant to overall rotation and translation (see Geometric Deep Learning) |
| Recycling | Feeds the prediction results back into the network multiple times for refinement |
| Confidence Estimation (pLDDT) | Provides a confidence score for each position, telling the user which parts are reliable |
AlphaFold 3 uses a more general architecture and employs diffusion models to directly generate atomic coordinates, allowing it to handle combinations of various molecules.
3.2 Impact
- Accelerating Research: Numerous laboratories use predicted structures to design experiments, resolve cryo-EM data, and understand disease mutations;
- Drug Discovery: Understanding target structures and predicting how drug molecules bind (complex prediction in AlphaFold 3);
- Protein Design: Moving from "reading" to "writing": designing new enzymes, vaccine antigens, binding proteins, and biosensors. The Baker lab has designed many proteins that do not exist in nature and have been experimentally verified to be functional;
- Research Paradigm: Demonstrated the power of "deep learning + high-quality public data (decades of accumulation in the Protein Data Bank, PDB)" in solving fundamental scientific problems.
3.3 Limitations
| Limitation | Description |
|---|---|
| Dynamics | It predicts a (or a few) static structure(s); however, many protein functions depend on conformational changes |
| Disordered Regions | Intrinsically disordered protein regions are difficult to predict (low pLDDT) |
| Mutation Effects | Limited predictive capability regarding how single-point mutations change structure and stability |
| Training Data Bias | Relies on existing structure types in the PDB; performs poorly on rare categories |
| Complexes and Ligands | AlphaFold 3 has significantly improved this, but accuracy for certain interactions remains limited |
| Availability | The manner in which code and weights were released upon the launch of AlphaFold 3 sparked academic debate; a version for academic use was later provided |
Summary
| Layer | Key points |
|---|---|
| Problem | Predicting 3D structure from amino acid sequences, a half-century-old challenge |
| Breakthrough | AlphaFold 2 (2020) approaches experimental accuracy; AlphaFold 3 (2024) extends to complexes |
| Honor | 2024 Nobel Prize in Chemistry: Hassabis, Jumper (prediction), Baker (design) |
| Direction | From prediction to design; dynamic structures and mutation effects remain to be solved |