AI in structural biology and protein science

How AI is used

Proteins are chains of amino acids that fold into particular three-dimensional shapes, and the shape largely decides what the protein does. Structural biology works out those shapes, usually with X-ray crystallography, cryo-electron microscopy or nuclear magnetic resonance, all of which give indirect signals that must be turned into a model. Protein science then asks how changing the chain changes the behaviour: how stable the protein is, how well it works, whether it binds something. Both tasks involve mapping between sequence, shape and function, and that mapping is where the learned models are put to work.

The jobs fall into a few groups. Some models predict a structure straight from a sequence, and the predicted model is then used as a working object: trimming floppy ends before an experiment, or supplying measures of how buried each residue is. Some run the other way, designing a sequence likely to fold into a given backbone. Others sit inside the experimental pipeline, interpreting raw spectra or merging diffraction measurements. And some model the measurements themselves, learning from a study's own results which sequence to make next.

Why researchers use it

The field's old bottleneck is that structures are slow and expensive to determine, while sequences are cheap and abundant. A model that predicts a plausible shape from a sequence stands in for a structure nobody has solved, and it can be used in bulk, across every candidate in a database rather than the handful an experiment could cover. Similar reasoning applies inside experiments: spectra and diffraction images need many judgement calls, traditionally made by an experienced person over days, and a learned component can take those calls on so the whole pipeline runs without stopping for a human.

In protein engineering the problem is different. The space of possible sequences is far larger than any laboratory can test, and most variants are simply broken. Here the model stands in for exhaustive testing. It is fitted to the measurements already made, predicts which untested sequences look promising, and the laboratory tests those, so each round of experiments narrows the search. The design side works the same way in reverse: a sequence model proposes chains for a desired shape, replacing hand-crafted energy calculations.

In the record

Limits

A predicted structure is a guess, not a measurement, and confidence scores tell you where the model is unsure but not where it is confidently wrong. Predictions tend to show one rigid shape, while real proteins move, shift between states and behave differently in a crowd. Sequence design models and variant predictors learn from what has already been solved or measured, so they favour the familiar and mislead on unusual folds, disordered regions and small changes with large effects. Laboratory testing remains the arbiter.

Written by AIxSci: the background from general knowledge, the examples from checked records.

What the AI was for

• Classification
• Structure determination
• Property prediction
• Candidate generation
• Detection
• Experimental design
• Segmentation
• Simulation surrogate
• Denoising

Model families

• Transformers
• Protein language models
• Multilayer perceptrons
• Convolutional neural networks
• Graph neural networks
• Linear models
• Gradient-boosted trees
• Random forests
• Support vector machines
• Recurrent neural networks
• Diffusion models
• Clustering
• Probabilistic graphical models
• Gaussian process
• Large language models
• Autoencoders
• Generative adversarial networks

Named models

• AlphaFold
• ESM-2
• ProtT5
• ColabFold
• AlphaFold 3
• ProteinMPNN
• ESMFold
• XGBoost
• crYOLO
• ProtBert
• PONDR
• ChatGPT and GPT models
• Boltz
• ESM-1v
• ESM-IF1
• ModelAngelo
• GEMME
• IUPred
• LightGBM
• RoseTTAFold
• RFdiffusion
• Chai-1
• CatBoost

Every article