AI solves a complex problem well when the problem is precisely defined, enough solved examples exist to learn from, and an independent test can check the answer. The clearest case is protein structure prediction: AlphaFold, from Google DeepMind, was judged by the CASP14 organizers in 2020 to have reached accuracy comparable to laboratory experiments for about two thirds of the test proteins, and the work shared the 2024 Nobel Prize in Chemistry. The same case also shows the limits, because the problem was a narrow one and the predictions still need careful reading.
Key takeaways
- Protein structure prediction was a 50-year-old open problem before AlphaFold, and measuring progress required a blind, independent contest called CASP.
- In CASP14, announced November 30, 2020, AlphaFold scored above 90 out of 100 on about two thirds of the target proteins.
- The AlphaFold Protein Structure Database now offers more than 200 million predictions, compared with roughly 180,000 structures determined by experiment.
- Each predicted residue carries a confidence score, and low-scoring regions should not be treated as reliable.
- The CASP14 targets were single proteins or domains, not protein complexes, so success on one task does not transfer automatically.
The problem: from amino acid sequence to three-dimensional shape
Proteins are built from amino acids strung together in sequence, and once folded, their three-dimensional form largely governs their biological role. Researchers have tried since the 1970s to predict the shape from the sequence alone. Determining one structure in a laboratory can take months or years and costly equipment, according to the CASP14 organizers and to the Nature paper that describes AlphaFold. EMBL-EBI reports that experiments have so far produced atomic-detail structures for over 180,000 proteins, a small fraction of the proteins known to science.
That gap makes the task a fair test of "complex problem solving." The problem resisted decades of effort, yet the correct answer for many proteins is known from experiments, so a program can be graded.
How the result was verified
The grading came from CASP, the Critical Assessment of Structure Prediction, a biennial challenge in which teams receive amino acid sequences for around 100 proteins and submit predicted shapes, which independent assessors then compare against structures determined in laboratories. In the 14th round, announced on November 30, 2020, AlphaFold produced models for about two thirds of the targets with a global distance test score above 90 out of 100. The organizers explain that above that threshold, remaining differences from the experimental structure are small and of the size expected from experimental error.
The peer-reviewed description followed in Nature on July 15, 2021. Its authors state that the system can regularly predict structures with atomic accuracy even when no similar structure is known, and that the deep learning model was designed around physical and biological insight into how proteins are built.
Checklist: when does AI suit a complex problem?
The table uses AlphaFold as the worked example. The right-hand column is a general rule of thumb drawn from this one case, not a law.
| Question | Protein structure answer | Why it matters |
|---|---|---|
| Is the target precisely defined? | Input is a sequence, output is atomic coordinates | Vague goals cannot be scored |
| Are solved examples available? | Experimental structures built up over decades | A learning system needs verified answers to learn from |
| Can an outsider check the output? | CASP compares predictions with laboratory-determined structures | Claims of success need a blind test |
| Is a confidence estimate available? | A per-residue score is reported for each prediction | Users can tell which parts to trust |
| Is the task narrow? | Single proteins or domains, not complexes, in CASP14 | Results rarely transfer to neighboring tasks without new work |
Efficiency, and what changed after the contest
The efficiency gain came from reuse. Google DeepMind and EMBL-EBI released the AlphaFold Protein Structure Database, which offers free access to predictions with broad coverage of UniProt, the standard repository of protein sequences. The Nobel Committee reported in October 2024 that AlphaFold2 had gained over two million users across 190 countries, with applications such as studying antibiotic resistance and imaging enzymes that can break down plastic. A researcher can now look up a predicted shape before deciding what to test in the lab.
The database stores a confidence value called pLDDT for each residue. Values of 90 or above are classed as very high confidence, 70 to 90 as confident, 50 to 70 as low, and below 50 as very low, and very low scores correlate with regions that tend to be intrinsically disordered. Treating a floppy, low-confidence loop as a precise structure is a typical misuse.
What this case does not show
- A prediction is a model, not a measurement. EMBL-EBI describes it as a complement to experiments, not a replacement.
- Success at protein folding does not mean AI can solve any complex problem. Many problems lack a blind benchmark, enough solved examples, or a clear definition of correct.
- Knowing a structure is not the same as having a drug. The Nobel announcement speaks of potential, and further research is needed between a shape and a therapy.
- The CASP14 organizers named protein complexes as a next frontier, which marks where the demonstrated result stops.
FAQ
Is protein folding now a closed problem?
It solved the prediction of shapes for most single proteins to a high level of accuracy, as judged in CASP14. The CASP14 organizers named protein complexes as the next frontier.
What does the Nobel Prize in Chemistry 2024 recognize?
David Baker received half for designing proteins computationally, and Demis Hassabis and John Jumper shared the rest for predicting protein structure. The announcement was made on October 9, 2024.
Can anyone use the predictions?
The AlphaFold Protein Structure Database is run by EMBL-EBI with Google DeepMind and offers its predictions freely to the scientific community.
How should the confidence scores be read?
Higher pLDDT means higher expected reliability for that residue. Scores below 50 are classed as very low confidence and should not be used to draw conclusions about local shape.
Related reading
- How AI approaches problem solving
- What makes AI problems hard
- Constraint satisfaction problems
- Opportunities and obstacles in complex tasks
More in AI Basics.
Last verified: October 2026
Sources
- Royal Swedish Academy of Sciences, press release, The Nobel Prize in Chemistry 2024, published October 9, 2024
- CASP14 press release, Protein Structure Prediction Center, November 30, 2020
- Jumper J et al., Nature 596:583-589, July 15, 2021, Highly accurate protein structure prediction with AlphaFold
- Varadi M et al., AlphaFold Protein Structure Database, Nucleic Acids Research 50(D1):D439-D444, 2022
- AlphaFold Protein Structure Database, About, EMBL-EBI, accessed October 2026