The folding problem
A peptide or protein is first a sequence, but its function depends on how it folds in three dimensions. For decades, predicting the folded shape from the sequence alone was one of biology's grand challenges. Even a modest chain can adopt an astronomical number of shapes.
A breakthrough
In 2020 and 2021, a deep-learning system called AlphaFold predicted protein structures with accuracy close to experimental methods for many targets. The work was recognised with a share of the 2024 Nobel Prize in Chemistry, alongside David Baker's pioneering work on computational protein design. Today, predicted structures for hundreds of millions of proteins are freely searchable.
What AI does for peptides
- Screening at scale. Models can score millions of candidate sequences for traits like binding strength, stability or predicted toxicity, then hand a short list to the lab.
- Generative design. Like language models that write text, generative models can propose new sequences with desired properties, including peptides that bind hard-to-reach targets.
- Antimicrobial discovery. Research groups have trained models on known antimicrobial peptides and used them to mine genomes and design new candidates, a few of which have worked in laboratory and animal tests.
- Optimising drugs. Algorithms suggest small changes that make an existing peptide last longer or bind better.
Where the limits are
- Predicted is not proven. A model can say a peptide should bind, but only a physical experiment can confirm it.
- Data bias. Models learn from existing data, which favours well-studied families.
- Safety and delivery. Whether a molecule survives in the body, reaches its target and is safe is still determined by lab work and clinical trials, which AI shortens but cannot skip.
- Peptides are flexible. Short chains are less rigid than large proteins, so structure prediction is harder and less reliable for them.
Why AI designed is not a quality label
The phrase is increasingly used in marketing. A computer-generated candidate has no more evidence behind it than a human-designed one until it has been tested. Ask the same questions: what was shown, in whom, and by whom?
What to watch
Expect faster early-stage discovery, more candidates reaching trials and, over time, new classes of peptide medicines. Expect failures too: biology is stubborn, and the fraction of candidates that succeed in humans has improved only slowly.
The takeaway
AI is a powerful new tool in the peptide workshop. It speeds up the first steps of a long road, and the rest of the road still has to be walked.
Sources
- Nobel Prize in Chemistry 2024: press release
- PubMed, PMID 34265844: Jumper et al. 2021, protein structure prediction with AlphaFold (Nature)
- PubMed, PMID 37933859: Varadi et al. 2024, AlphaFold Protein Structure Database (Nucleic Acids Res)
- PubMed, PMID 38843834: Santos-Junior et al. 2024, antimicrobial peptides in the global microbiome by machine learning (Cell)