In plain language
is the broad field. is an approach within it: learning patterns from data rather than specifying every rule.
Protein language models, unfolded
Attention is all you need to fold proteins.
For beginners in protein design, and for anyone who forgets the basics. This is the closest we have come to what nature does in seconds, and what nature does is still the better-kept secret. Start with the mechanism, then read the models that use it.
A map of machine intelligence
Select an idea to look insideOne family of architectures. More than one language.
FieldsArchitectureApplications
01 to 06 / Select to discoverThe model catalogue
Architectures, weights and sourcesYour first discoveries
Start anywhere. Follow your curiosity.Inside a neuron
Build a decision, one weight at a time.
Step insidePaying attention
Watch tokens exchange information.
Step insideThe language of life
Discover what proteins can teach a model.
Step insideThe technical idea
is a subset of based on multilayer neural networks. are one architecture; language modeling is an objective. These are related categories, not synonyms.
From data to proteins
apply sequence-learning objectives to . Not every protein model predicts a 3D structure, and not every language model is a .
Look deeper: mathematics, methods & limitations
AI ⊃ ML ⊃ DL
Classical includes logistic regression, decision trees, random forests and support machines. “Classical” does not necessarily mean linear, simple or obsolete.
A () is a language model at large scale; there is no universal threshold. A (, sometimes called a ) models amino-acid sequences. Here always means , not simply pretrained language model.
Follow the numbered galleries for a guided visit, or select any in the map. Hover, click, touch and keyboard focus reveal the same information. All live experiments use compact educational models, not hosted foundation models.