Learn/Predict a structure
Choose a model
The folding models on Phi fall into three families, and the family decides most of what matters: what the model can take, how long it runs, and where it is strong. The times below are for the 639-residue ectodomain on one day; a longer chain, an alignment to build, or a busy queue moves them, so read them against each other rather than as promises.
Single sequence
Protein chainsA protein language model reads the one sequence and writes the structure. No alignment is built, so it is fast, and it is what designed proteins want.
| Model | Time on this sequence | pLDDT | pTM |
|---|---|---|---|
| ESMFold | 3 min | 73.56 | 0.8944 |
| ESMFold 2, fast | 7 min | 93.4 | 0.9718 |
Alignment first
Protein chains, templatesAn alignment of related sequences is built, then the structure is read out of their covariation. The slowest, and the reference method for natural proteins with many relatives.
| Model | Time on this sequence | pLDDT | pTM |
|---|---|---|---|
| AlphaFold 2 | 4 min | 92.67 | 0.92 |
All-atom diffusion
Protein, DNA, RNA, ligandsAn alignment or a language model conditions a diffusion model that places every atom, so ligands and nucleic acids fold with the protein.
| Model | Time on this sequence | pLDDT | pTM |
|---|---|---|---|
| Boltz-2 | 4 min | 90.81 | |
| Chai-1 | 6 min | 85.19 | 0.9214 |
| OpenFold 3 | 8 min | 88.89 | 0.9315 |
| Protenix v2 | 4 min | 91.12 | 0.9547 |
| OpenDDE v1 | 9 min | 93.27 | 0.9605 |
| ESMFold 2 | 6 min | 93.9 | 0.9777 |
| Boltz-1 | 7 min | 91.21 |
What Auto picks
- One protein chain
- ESMFold 2 folds a single chain from sequence alone with the full ESMFold 2 model.
- Several chains
- AlphaFold 2 predicts protein complexes and binder-target interfaces.
- A ligand, DNA or RNA
- Boltz-2 co-folds proteins with ligands, DNA or RNA and can predict binding affinity.
Expert mode lists every model the input allows and lets several run at once, so the choice can be compared rather than made.
Beyond proteins
The diffusion models place a small molecule in its pocket and fold DNA and RNA alongside the protein. Boltz-2 also estimates a small molecule's binding affinity. Treat that number as a screen that separates likely binders from unlikely ones, not as a free energy to report: it is unreliable for molecules above about fifty heavy atoms, ignores cofactors and ions, and says nothing about protein–protein binding.
Read more: ColabFold · Single-sequence methods · Boltz-2