SwitchCraft: Programmatic Multi-State Protein Design

• SwitchCraft is a programmatic framework that designs state-switching proteins by optimizing a single sequence across multiple conformational states.
• It employs a compositional constraint system using loss functions such as motif, binding, and conformational change losses to enforce state-specific designs.
• In silico experiments validate its potential for engineering allostery, induced binding, ligand discrimination, and biosensor-like switching behaviors.

SwitchCraft is a programmatic framework for designing state-switching proteins: proteins whose function depends on which molecular state they occupy, such as whether a ligand is bound, which ligand is bound, or what conformation is adopted. It is formulated as a multistate optimization system in which a single sequence is optimized against multiple state-specific structural and interaction constraints, with losses evaluated through the structure prediction model Boltz-1 and gradients propagated back to sequence parameters. In contrast to protein design workflows centered on one static fold, SwitchCraft targets multistate mechanisms including allostery, induced binding, ligand-driven conformational rearrangement, ligand discrimination, and biosensor-like switching (Jing et al., 29 May 2026).

1. Problem formulation and scientific context

The framework is motivated by the observation that many natural protein functions are intrinsically multistate. The examples explicitly cited include motor proteins, ATP synthase, flagella, polymerases, and allosteric receptors and enzymes. In these systems, function depends on state-dependent structural changes and state-specific interactions rather than on a single folded structure alone (Jing et al., 29 May 2026).

SwitchCraft is positioned against a background in which existing deep learning methods are effective at generating sequences from known protein families, designing binders to a target structure, and scaffolding static motifs, but are not built to specify and optimize multiple coupled states simultaneously. The paper argues that there is no large-scale dataset of sequence/structure pairs annotated with rich multistate behaviors, so a purely data-driven generative solution is not yet feasible. The resulting design gap is addressed by recasting multistate protein design as an explicit optimization problem over compositional constraints evaluated with a structure predictor rather than as direct generation from a multistate training distribution (Jing et al., 29 May 2026).

Within this framing, the central design objective is to find a sequence that satisfies a set of state-specific requirements at once. This shifts the abstraction from “design a protein with structure $X$” to “design a protein whose behavior changes appropriately across states.” A plausible implication is that the framework is intended less as a replacement for single-state design tools than as a higher-level control layer for functions that require conditional behavior.

2. Programmatic specification of multistate design

A SwitchCraft design is specified as a collection of states. Each state has a folding context $\mathcal{C}_s$, one or more loss functions, a residue design mask, and optionally fixed motif residues. The folding context may include ligands, DNA, metal ions, or other molecules. Losses are computed from Boltz-1 predictions for the designed sequence in the relevant context, and the global objective is the sum of the state-specific terms (Jing et al., 29 May 2026).

This design language is explicitly compositional. State definitions can combine motif scaffolding in one state, motif suppression in another, binding in a holo state, anti-binding in an apo state, and conformation-separation constraints across states. The framework therefore treats multistate behavior as a constrained program over protein states rather than as a monolithic end-to-end objective.

Sequence optimization is performed in a logit-parameterized space: $\mathbf{z} \in \mathbb{R}^{20 \times L}.$ The soft sequence is defined as

$$\mathbf{z}_{\text{soft}} = \mathrm{softmax}(\mathbf{z}/\tau),$$

and the hard sequence as

$$\mathbf{z}_{\text{hard}} = \mathrm{onehot}(\arg\max(\mathbf{z})).$$

Because $\arg\max$ is non-differentiable, SwitchCraft uses a straight-through estimator so that the forward pass behaves discretely while the backward pass uses gradients through the soft relaxation. The optimization update is given as

$$\mathbf{z} \gets \mathbf{z} - \alpha \nabla_{\mathbf{z}} \mathcal{L}(\mathrm{Boltz\text{-}1}(\mathbf{z}_\text{pseudo})).$$

When a motif mask is present, motif residues are clamped to one-hot identities and only scaffold positions are optimized (Jing et al., 29 May 2026).

The optimization schedule consists of 240 steps in four stages: 30 steps with $\alpha = 0.2$, $\beta = 0$, $\gamma = 1$, $\mathcal{C}_s$0; 100 steps with $\mathcal{C}_s$1, $\mathcal{C}_s$2, $\mathcal{C}_s$3 interpolated from 0 to 1, $\mathcal{C}_s$4; 100 steps with $\mathcal{C}_s$5, $\mathcal{C}_s$6, $\mathcal{C}_s$7, and $\mathcal{C}_s$8 annealed from 0.5 to 0.005; and a final 10 steps with $\mathcal{C}_s$9, $\mathbf{z} \in \mathbb{R}^{20 \times L}.$0, $\mathbf{z} \in \mathbb{R}^{20 \times L}.$1, $\mathbf{z} \in \mathbb{R}^{20 \times L}.$2. This schedule progressively moves from exploration toward a nearly discrete sequence (Jing et al., 29 May 2026).

3. Constraint primitives and loss construction

SwitchCraft builds multistate behavior from a small set of reusable loss primitives. These include motif loss, anti-motif loss, binding loss, anti-binding loss, conformational change loss, and contact loss (Jing et al., 29 May 2026).

The motif loss is used to scaffold a motif in a chosen state by matching pairwise motif distances under the predicted distogram. The anti-motif loss is defined as $\mathbf{z} \in \mathbb{R}^{20 \times L}.$3, so it actively discourages motif formation in a designated state. This pairing allows the same structural element to be presented in one state and suppressed in another.

The binding loss is adapted from BoltzDesign1 and uses a thresholded entropy of predicted protein-ligand distance distributions. In the description provided, it encourages high-confidence contacts, rewards contact probability within a cutoff, and aggregates the strongest contacts across ligand tokens; the number of considered tokens varies during optimization, with $\mathbf{z} \in \mathbb{R}^{20 \times L}.$4 changing from 8 to 12. The anti-binding loss is likewise defined as $\mathbf{z} \in \mathbb{R}^{20 \times L}.$5, enabling explicit discouragement of interaction in selected states (Jing et al., 29 May 2026).

To induce structural divergence across states, SwitchCraft uses a Jensen-Shannon-divergence-based conformational change objective that pushes residue-pair distance distributions to differ between two contexts. In effect, it does not merely require each state to be internally plausible; it also requires them to be distinguishable from one another. Contact loss provides a complementary regularizer by encouraging each residue to have at least one confident long-range contact, helping maintain well-folded and confidently predicted states.

Taken together, these losses define a library of “functional primitives.” The framework’s key methodological claim is that higher-order switching behaviors can be composed from these primitives without requiring multistate supervision. This suggests a compiler-like view of protein behavior design: the target mechanism is written as a combination of state-local and cross-state constraints, then optimized through a common predictor.

4. In silico demonstrations of switching behavior

The paper reports six in silico validation settings spanning several classes of multistate function. These include positive and negative allostery, motif switching, ligand modification, induced binding, ligand discrimination, and biosensor-oriented switching (Jing et al., 29 May 2026).

In the allostery setting, the goal was either to promote motif formation upon ligand binding or to disrupt motif formation upon ligand binding. The experiments used 24 motifs from the RFDiffusion benchmark and five ligands: OQO, FAD, Zn$\mathbf{z} \in \mathbb{R}^{20 \times L}.$6, Mg$\mathbf{z} \in \mathbb{R}^{20 \times L}.$7, and dsDNA (GAATTC). The study generated 100 designs per problem and found 11 motifs with at least one successful design. The success criteria were motif RMSD $\mathbf{z} \in \mathbb{R}^{20 \times L}.$8 Å in the active state, motif RMSD $\mathbf{z} \in \mathbb{R}^{20 \times L}.$9 Å in the inactive state, low intra-state variability, and meaningful state difference (Jing et al., 29 May 2026).

Motif switching extended this logic to reciprocal presentation of two motifs. Using motifs 3IXT and 1YCR with OQO as effector, 3 of 100 designs satisfied the full reciprocal switching specification. Ligand modification tested whether the same protein could respond differently to different chemical forms of a ligand; in the heme versus oxygenated-heme example, 10 of 558 designs showed the desired change, and one example exhibited a 3.8 Å conformational shift due to histidine displacement by oxygen (Jing et al., 29 May 2026).

Induced binding targeted effector-gated recognition. In the reported example, a 50-aa protein was designed to bind a 16-aa Top7 fragment only in the presence of Ca$$\mathbf{z}_{\text{soft}} = \mathrm{softmax}(\mathbf{z}/\tau),$$0. Among 940 designs, 8 showed significantly higher interface confidence in the calcium-bound state, and one example exhibited a 12.5 Å conformational change while forming the interface only in the bound state. Ligand discrimination further generalized the framework to more than two states: among 465 designs for unbound, OQO-bound, and Ca$$\mathbf{z}_{\text{soft}} = \mathrm{softmax}(\mathbf{z}/\tau),$$1-bound conditions, 12 were successful. The highlighted design formed a salt bridge in the apo state, rearranged to a hydrophobic pocket in the OQO state, and shifted further to coordinate calcium in the third state (Jing et al., 29 May 2026).

These experiments are notable not because of uniformly high success fractions, which the paper does not claim, but because they span qualitatively different multistate mechanisms within one optimization framework. A plausible implication is that the compositional objective library is expressive enough to encode several archetypal regulatory behaviors without changing the underlying optimizer.

5. De novo fluorescent biosensor design

One of the principal application domains emphasized for SwitchCraft is fluorescent biosensor design. The reported strategy combines a designed conformation-switching module with cpGFP, following the general biosensor pattern in which ligand binding perturbs the environment of a fluorescent chromophore and changes fluorescence (Jing et al., 29 May 2026).

The framework was used to design switchers for SAM, cGMP, and ATP. The specification combined ContactLoss in apo and holo states, BindingLoss in the holo state, and ConfChangeLoss across states. The study generated 13,858 total designs of length 150–200, of which 89 passed strict criteria: effector iPTM $$\mathbf{z}_{\text{soft}} = \mathrm{softmax}(\mathbf{z}/\tau),$$2, intraRMSD $$\mathbf{z}_{\text{soft}} = \mathrm{softmax}(\mathbf{z}/\tau),$$3 Å, average pLDDT $$\mathbf{z}_{\text{soft}} = \mathrm{softmax}(\mathbf{z}/\tau),$$4, crossRMSD $$\mathbf{z}_{\text{soft}} = \mathrm{softmax}(\mathbf{z}/\tau),$$5 Å, and compact fold (Jing et al., 29 May 2026).

For cpGFP insertion, the workflow identified insertion sites at residues with the largest backbone dihedral changes, enforced spatial separation among insertion sites, inserted cpGFP, co-folded the fusion construct, and screened for changes in chromophore contacts. Using this procedure, 44 designs passed the chromophore modulation screen. The paper highlights a SwitchCraft-designed SAM biosensor that reproduces the same “unquenching” mechanism observed in a known nicotine biosensor: ligand binding shifts a glutamate away from the chromophore, enabling fluorescence (Jing et al., 29 May 2026).

This biosensor application is significant because existing genetically encoded fluorescent biosensors often depend on a natural switch protein already evolved for the ligand of interest. SwitchCraft instead treats the switch module itself as a design target. This suggests a route toward analyte-specific biosensors that are not restricted to naturally occurring sensing domains, although the evidence reported is primarily computational.

6. Limitations, assumptions, and methodological significance

The paper is explicit that many reported success rates are modest in absolute terms. It also states that the framework relies on Boltz-1 predictions, so practical success depends on how faithfully the predictor reflects real physical behavior. In addition, kinetic plausibility is not guaranteed: some designed structural changes may be valid as predicted states but mechanistically unrealistic if they would require unfolding or otherwise inaccessible pathways. Evaluation is described as largely in silico, although the appendix includes preliminary wet-lab validation for zinc-induced PD-L1 binders (Jing et al., 29 May 2026).

A further limitation is representational. Because SwitchCraft works by specifying states and losses, success depends on whether the intended behavior can be adequately expressed through the available compositional constraints. This is a strength insofar as it provides explicit control, but also a constraint insofar as behaviors outside the expressible loss vocabulary are not directly targeted.

Within those boundaries, the framework’s significance lies in reframing protein design as programmable state-dependent behavior design. The functional primitives reported include ligand-activated motif scaffolding, ligand-inactivated motif scaffolding, motif switching, ligand-modified conformational change, induced binding, ligand discrimination, and biosensor-ready conformational switching. The broader claim is not that all such designs are solved, but that a general-purpose, compositional, gradient-based paradigm for multistate design is feasible (Jing et al., 29 May 2026).

The paper also points to likely extensions, including atomic-motif constraints, more complex multistate specifications, experimental validation, and mechanisms incorporating kinetic plausibility. This suggests an emerging design program in which proteins are specified less as static objects than as conditional dynamical systems.

7. Terminological scope and unrelated uses of the name

The name “SwitchCraft” has been used for unrelated computational systems outside protein design. In agentic AI, “Switchcraft” denotes a model router for agentic tool calling that selects the cheapest predicted-correct model for a structured function-calling request (Agarwal et al., 8 May 2026). In generative video modeling, “SwitchCraft” denotes a training-free framework for multi-event text-to-video generation based on Event-Aligned Query Steering and the Auto-Balance Strength Solver (Xu et al., 27 Feb 2026).

These usages are distinct in domain, method, and objective from the protein-design framework described above. In the protein-design context, SwitchCraft specifically refers to multistate protein optimization through compositional constraints and backpropagation through Boltz-1 (Jing et al., 29 May 2026).