Three ways in.

You know some biology. The machine learning words are the unfamiliar part.

1. 01Train a classifier on 100 pointsBefore any protein appears, watch a model learn a boundary from examples. Everything after this is the same loop with more parameters.
2. 02Look inside one neuronWeights, bias, activation. Three ideas that reappear in every architecture in the catalogue.
3. 03Fill in a masked wordThis is the exact objective that produced ESM-2. Run it on words first, where you can judge the answer yourself.
4. 04Read ESM-2Now the same objective on amino acids, at a scale where the internal states start to encode structure.
5. 05Follow it to a structureWhat has to be added to a language model before coordinates come out the other end.

You have a sequence and you need a model of it, plus a way to judge what you get.

1. 01Start with AlphaFold 2The alignment-based route, still the reference for a single chain with known relatives.
2. 02See what AlphaFold 3 changedIf your question involves a ligand, a nucleic acid or a complex, the answer is a different model.
3. 03Use ESMFold when speed winsNo alignment search. Worth it for orphan sequences and for anything you need thousands of.
4. 04Check where the line is nowThe 2026 preprint, with the caveats that come with reading a preprint.

You want to build something that does not exist yet, and you want to know which model does which step.

1. 01Generate a backboneDiffusion produces the shape. Constraints are how you say what the shape is for.
2. 02Give it a sequenceA backbone with no sequence cannot be ordered. Inverse folding fills that in.
3. 03Add the ligand to the pictureIf your design holds a small molecule, a metal or DNA, the protein-only model is blind to half the problem.
4. 04Check before you orderFold the designed sequence and see whether it returns to the shape you asked for. Most designs do not.
5. 05Or prompt one model insteadThe multimodal route: describe the motif and the function, let one model write the rest.