SwitchCraft
commit 67ffc0dDesign proteins that change structure when a ligand binds.
Most protein design asks for one structure. A biosensor or a molecular switch needs two: the protein sits in one shape on its own and snaps into another when a small molecule binds. SwitchCraft designs for that directly. It optimizes one sequence by gradient descent through the Boltz-1 structure predictor, scoring the sequence against every state at once, so that the same protein is predicted to adopt a different structure in each state.
ProteinIQ runs SwitchCraft with its five built-in tasks, described by Jing and colleagues in the 2026 SwitchCraft paper. Each task fixes the states and the loss terms, and the job asks for the ligand or ligands and a few size settings. The results are the designed sequences, Boltz-1 structures for every state in PDB and mmCIF, and the Boltz-1 confidence metrics of every sample. For a single static structure, use BoltzGen or another design tool from the protein design page.
Custom loss functions, custom state definitions, protein, DNA, and RNA binding partners, and user-supplied motif files are not offered. Motifs come from the 27 motif files that ship with SwitchCraft.
SwitchCraft is billed by the minute of runtime at 15 credits per minute on an NVIDIA T4 GPU, with a minimum of 15 credits. Optimization runs for 240 iterations per design, so a job takes tens of minutes to several hours. ProteinIQ estimates the runtime from the task, the scaffold size, and the settings, and shows the estimate as the quote before submission. The charge follows the actual runtime.
| Task | Scaffold | Estimate, 1 design |
|---|---|---|
Positive allostery | motif 4zyp, 50 residues | 163 credits |
Positive allostery | motif 1prw, 105 residues | 427 credits |
Positive allostery | motif 1cll, 144 residues | 701 credits |
Induced binding | Scaffold length 60 | 427 credits |
Induced binding | Scaffold length 100 | 709 credits |
Induced binding | Scaffold length 150 | 1,170 credits |
Ligand discrimination | Scaffold length 100 | 649 credits |
Ligand discrimination | Scaffold length 150 | 1,205 credits |
Motif switching | 1prw then 1cll | 1,802 credits |
These estimates use the default Recycles of 0 and no LigandMPNN redesign. Cost grows with the size of the states, because every state is predicted at every iteration. Motif tasks use the maximum scaffold length of the chosen motif file. Negative allostery costs the same as Positive allostery for the same motif.
Designs multiplies the cost, so three 1prw designs estimate at 1,229 credits. Recycles 1 adds about 2.4 times the optimization time, which moves the 1prw single-design quote from 427 to 989 credits. LigandMPNN sequences adds one redesign run and one structure prediction per sequence, and 4 sequences move the same job to 528 credits. Ligand size is priced at an average, because a CCD code or SMILES string is not measured before submission. If the credit budget runs out during a run, the run stops and the designs and files finished so far are returned.
| Input | Accepted forms | Used by |
|---|---|---|
Ligand | CCD code or SMILES text, a PubChem lookup, or a .ccd, .smiles, .smi, or .txt file | Positive allostery, Negative allostery, Motif switching |
Target ligand | Same forms | Induced binding |
Effector ligand | Same forms | Induced binding |
Ligand 1 | Same forms | Ligand discrimination |
Ligand 2 | Same forms | Ligand discrimination |
Job name | Text | Optional label for the run. |
Each ligand is one small molecule, entered as a Chemical Component Dictionary code such as OQO or FAD, or as a SMILES string. Files are limited to 10 MB. SwitchCraft needs no protein input, because it designs the protein from scratch.
| Parameter | Type | Default | Description |
|---|---|---|---|
Task | enum | Positive allostery | Positive allostery, Negative allostery, Induced binding, Ligand discrimination, or Motif switching. |
Positive allostery designs a protein in which a motif is absent without the ligand and forms when it binds. Negative allostery does the reverse: the motif is present without the ligand and is disrupted when it binds. Induced binding designs a protein that binds a target ligand only when an effector ligand is also bound. Ligand discrimination uses three states, with no ligand, Ligand 1, and Ligand 2, each with a distinct conformation. Motif switching activates one motif without the ligand and a second motif when the ligand binds.
| Parameter | Type | Default | Description |
|---|---|---|---|
Motif | enum | 1prw | The motif the scaffold is built around, one of 27 bundled motif files, named by PDB entry with a segment variant where one exists. For Motif switching this is motif 1, active without the ligand. Used by the allostery tasks and Motif switching. |
Second motif | enum | required | Motif 2, active when the ligand is bound. Required for Motif switching. |
Scaffold length | integer | 100 | Residues in the designed protein, 1 to 150. Used by Induced binding and Ligand discrimination, which have no motif. |
With a motif, SwitchCraft chooses the scaffold length itself within the range the motif file allows, so the designed protein can be shorter than the file's maximum.
| Parameter | Type | Default | Description |
|---|---|---|---|
Designs | integer | 1 | Independent designs, 1 to 5. They are optimized one after another, so runtime and cost grow with this number. |
Recycles | integer | 0 | Boltz-1 recycling steps during optimization, 0 or 1. The final structures are always predicted with 3 recycles. |
LigandMPNN sequences | integer | 0 | When above 0, LigandMPNN redesigns each design into this many sequences, 1 to 8, and Boltz-1 predicts their structures. |
The results open on a Designs table, a Model metrics table, a Structures viewer, and a Files tab.
| Output | Contents |
|---|---|
Designs table | One row per design: Design, Residues, and the sequence read from chain A of state 0, sample 0. |
Model metrics table | One row per design, state, and sample with the Boltz-1 confidence metrics. |
| Predicted structures | Five samples for every state of every design, as PDB and mmCIF files. |
| Raw model outputs | One Boltz-1 output file per state, with per-residue pLDDT and the PAE and PDE matrices. |
config.yaml | The task configuration that was passed to SwitchCraft. |
run.log | The log of the run. |
With LigandMPNN sequences above 0, the redesigned sequences and their predicted structures are returned with the other files SwitchCraft writes.
States are numbered from 0. In the three single-ligand tasks, state 0 is the protein alone and state 1 is the protein with the ligand. For Induced binding, state 0 contains the target ligand and state 1 contains the target ligand and the effector. For Ligand discrimination, state 0 is the protein alone, state 1 has Ligand 1, and state 2 has Ligand 2. A successful design is one where the predicted structures differ between states in the way the task asks, for example a motif that is formed only in state 1.
The Model metrics table reports Confidence score, pTM, ipTM, Ligand ipTM, Protein ipTM, Complex pLDDT, Complex ipLDDT, Complex PDE, and Complex iPDE for each sample. The pTM and ipTM guide and the pLDDT guide explain how to read these. A low Ligand ipTM in a bound state means Boltz-1 is not confident that the ligand sits in the intended place, and a low Complex pLDDT means the fold itself is uncertain. Compare samples within a state first, then compare states.
Boltz-1 predicts the structures and also drives the optimization, so these metrics measure how convincing a sequence is to the same model that shaped it. They are a filter, not proof that the protein switches in a cell or a tube. Open the structures of each state in the viewer, superimpose them, and check that the motif or binding site changes as intended, and that the ligand has plausible contacts in the bound state. Designs that pass should be tested with independent predictors and, for any real use, in the lab. The guide to structure prediction confidence covers independent checks.
The designed sequence in the Designs table is read from the structure files and is not written by SwitchCraft itself, so it always matches chain A of state 0, sample 0.
Each task is a set of states, and each state is a protein plus a list of ligands. SwitchCraft optimizes one sequence for all states together: at every iteration Boltz-1 predicts the structure of each state, loss terms on those predictions score whether the task is met, and gradient descent updates the sequence. The run lasts 240 iterations per design, and Boltz-1 then predicts five samples for each state.
All-atom SE(3)-equivariant diffusion for unconditional generation, motif scaffolding, and binder design.
Generate proteins with the ESM3 multimodal language model.
Family-conditioned protein sequence generation from FASTA, A2M, or A3M input
Generate de novo protein sequences from plain-text prompts with sampling controls.
Design all-atom protein binders with controlled target sites, lengths, and interaction parameters.
Design proteins from a text description
Generate novel drug-like molecules in protein binding pockets using AI-powered structure-based design.
Generate docking poses, drug-like molecules, and peptides inside protein binding pockets with PocketXMol.
Design antibody CDRs with Proteo-R1 reasoning, diffusion, and optional framework inpainting.
Generate protein sequences de novo, scaffold fixed motifs, or inpaint selected regions.