Protein-based biosensors have important roles in synthetic biology and clinical applications, but the design of biosensors has so far been mostly limited to reengineering natural proteins1. Finding specific analyte-binding domains that undergo conformational changes upon binding is challenging, and even when available, extensive protein engineering efforts are generally required to effectively couple them to a reporter domain5,6. It is therefore desirable to construct modular biosensor platforms that can be easily repurposed to detect different protein targets of interest. Modular systems have been developed to detect antibodies7,8,9 and small molecules10,11, but general protein sensors are a bigger challenge given the great diversity of protein structures, sizes and oligomerization states, and approaches such as semisynthetic protein platforms12,13,14, or calmodulin switches15,16, usually require considerable screening to find potential candidates owing to limited predictability17.

A protein biosensor can be constructed from a system with two nearly isoenergetic states, the equilibrium between which is modulated by the analyte being sensed. Desirable properties in such a sensor are: (i) the conformational change triggered by an analyte should be independent of the details of the analyte, so the same overall system can be used to sense many different targets; (ii) the system should be tunable so that analytes with different binding energies and different typical concentrations can be detected over a large dynamic range; and (iii) the conformational change should be coupled to a sensitive output. We hypothesized that these attributes could be attained by inverting the information flow in de novo designed protein switches in which binding to a target protein of interest is controlled by the presence of a peptide actuator2. We developed a system that consisted of two protein components: first, a ‘lucCage’ that comprises a cage domain and a latch domain that contains a target-binding motif and a split luciferase fragment (small BiT (SmBiT) 114)18; and second, a ‘lucKey’ that contains a key peptide that binds to the open state of lucCage and the complementary split luciferase fragment (large BiT (LgBit) 11S)18 (Fig. 1a). lucCage has two states: a closed state, in which the cage domain binds to the latch and sterically occludes the binding motif from binding the target and SmBiT from combining with LgBit to reconstitute luciferase activity, and an open state, in which these binding interactions are not blocked and lucKey can bind to the cage domain. The association of lucKey with lucCage results in the reconstitution of luciferase activity (Fig. 1a, right). The thermodynamics of the system are tuned such that the binding free energy of lucKey to lucCage (ΔGCK) is insufficient to overcome the free energy cost of lucCage opening (ΔGopen) in the absence of target (ΔGopen − ΔGCK >> 0), but in the presence of the target, the additional binding free energy of the latch to the target (ΔGLT) drives latch opening and luciferase reconstitution (ΔGopen − ΔGCK − ΔGLT << 0) (Fig. 1b, c). This system satisfies properties (i) and (ii) above, as a wide range of binding activities can be caged, and because the switch is thermodynamically controlled, the lucKey and target binding energies can be adjusted to achieve activation at the relevant target concentrations. Because lucKey and lucCage are always the same, the system is modular—the same molecular association can be coupled to the binding of many different targets. Bioluminescence provides a rapid and sensitive readout of the analyte-driven lucCage–lucKey association, satisfying property (iii).

Fig. 1: De novo design of multi-state biosensors.
figure 1

a, Sensor schematic mechanism. The closed form of lucCage (left) cannot bind to lucKey, thus preventing the split luciferase SmBiT fragment from interacting with LgBit. The open form (right) can bind to both the target and the key, enabling the reconstitution of SmBiT and LgBiT for luciferase activity. b, Thermodynamics of biosensor activation. The free energy cost (ΔGopen) of the transition from closed cage (species 1) to open cage (species 2) disfavours the association of key (species 5) and reconstitution of luciferase activity (species 6) in the absence of target. In the presence of the target, the combined free energies of target binding (2→3; ΔGLT), key binding (3→4; ΔGCK), and SmBiT–LgBiT association (4→7; ΔGR) overcome the unfavourable ΔGopen, driving the opening of the lucCage and reconstitution of luciferase activity. c, Thermodynamics of biosensor design. The designable parameters are ΔGopen and ΔGCK; ΔGR is the same for all targets, and ΔGLT is pre-specified for each target. For sensitive but low background analyte detection, ΔGopen and ΔGCK must be tuned such that the closed state (species 1) is substantially lower in free energy than the open state (species 6) in the absence of target, but higher in free energy than the open state in the presence of target (species 7).

The states of this biosensor system are in thermodynamic equilibrium, with the tunable parameters ΔGopen and ΔGCK governing the populations of the possible species, along with the free energy of association of the analyte to the binding domain ΔGLT (Fig. 1b). We simulated the dependence of the sensor system on ΔGopen (Extended Data Fig. 1a), ΔGLT (Extended Data Fig. 1b), and the concentration of analyte and the sensor components (Extended Data Fig. 1c, d). The sensitivity of analyte detection is a function of ΔGLT, with a lower limit of roughly one-tenth of the dissociation constant (Kd) for analyte binding (Extended Data Fig. 1b). Above this lower limit, varying the concentration of lucCage and lucKey enables the system to respond to different ranges of target concentration (Extended Data Fig. 1c, d). Sensitivity can be further modulated by tuning the strength of the intramolecular cage–latch interaction and the intermolecular cage–key interaction (ΔGopen and ΔGCK, respectively); for example, too tight cage–latch interaction results in a low signal in the presence of target, and too weak an interaction results in a high background signal in the absence of target (Extended Data Fig. 1a, e). Our design strategy aims to find this balance by modulating ΔGopen and ΔGCK by varying the length of the latch (and key) helix and by introducing either favourable hydrophobic or unfavourable buried polar interactions at the cage–latch or cage–key interfaces2 (Extended Data Fig. 1f, g).

Designing tunable lucCage sensors

To design sensors based on these principles, we developed a ‘GraftSwitchMover’ Rosetta-based method to identify placements of target binding peptides within the latch such that the resulting protein is stable in the closed state and the interactions with the target are blocked (Supplementary Methods). As a first test, we grafted the SmBiT peptide and the BIM peptide in the closed state of the previously described optimized asymmetric LOCKR switch2 (Extended Data Fig. 2). SmBiT adopts a β-strand conformation within the luciferase holoenzyme, but we assumed that it could adopt a helical secondary structure in the context of the helical bundle scaffold, because secondary structure can be context-dependent19. We sampled different placements for the two peptide sequences across the latch, selected the lowest energy solutions (Extended Data Fig. 2a) and expressed 12 designs in Escherichia coli. We mixed the designs with lucKey in a 1:1 ratio, then added BCL-2, which binds to BIM with nanomolar affinity20, and observed a rapid increase in luminescence (Extended Data Fig. 2b, f; we refer to the best of these as ‘lucCageBIM’), which demonstrates that the LOCKR actuator2 operated in reverse can function as a biosensor. The detection range of the analyte could be tuned by varying the concentration of the sensor (lucCage plus lucKey) (Extended Data Fig. 2g), as anticipated in our model simulations (Extended Data Fig. 1c). lucCageBIM has SmBiT at position 312 in the latch (SmBiT312) (Extended Data Fig. 2d); the cage with this placement (lucCage) was used as the base scaffold for the biosensors described below.

lucCage sensors with miniprotein sensing domains

We next investigated the incorporation of a range of binding modalities for analytes of interest within lucCage by developing methods for computationally caging target-binding proteins, rather than peptides, in the closed state (Supplementary Methods). As a test case, we caged the de novo designed influenza A H1 haemagglutinin (HA)21 binding protein HB1.9549.2 into a shortened version of the LOCKR switch22 (sCage), optimized to improve stability and facilitate crystallization efforts (Fig. 2a). Two out of the five designs were functional, and bound HA in the presence but not the absence of key (Extended Data Fig. 3b). The crystal structure of the best design, sCageHA_267-1S, determined to 2.0 Å resolution (Supplementary Table 1, Protein Data Bank (PDB) code 7CBC), showed that all HA-binding interface residues except one (Phe273) interact with the cage domain (blocking binding of the latch to the target) as intended by design (Fig. 2a, Extended Data Fig. 3a–c).

Fig. 2: Design and characterization of de novo biosensors incorporating small proteins as sensing domains.
figure 2

a, Structural validation of sCageHA_267-1S, which cages a designed influenza binding protein inside a LOCKR switch. Left, design model of the de novo binder HB1.9549.2 (cyan ribbon) bound to the stem region of influenza haemagglutinin (HA, green ribbon)21. Right, crystal structure (PDB code 7CBC) of sCageHA_267_1S, comprising HB1.9549.2 (cyan) grafted into a shortened and stabilized version of the LOCKR switch22 (sCage, yellow ribbon). Middle, all residues of HB1.9549.2 involved in binding to HA (magenta, top) except for F273 are buried in the closed state of the switch (bottom). The magenta labels indicate the same set of amino acids in the two panels (for example, F2 in the top panel corresponds to F273 in the bottom panel). bd, Functional characterization of lucCageBot (b), lucCageProA (c) and lucCageHER2 (d). Left, structural models incorporate a de novo designed binder for BoNT/B (Bot.671.2)21 (b) the C domain of protein A (SpA C)23 (c) or a HER2-binding affibody24 (d) into lucCage (blue ribbon) with caged SmBiT fragment (gold ribbon). Middle, measurement of luminescence intensity after the addition of 50 nM of analyte (BoNT/B (b), IgG Fc (c) or HER2 (d)) to a mixture of 10 nM of each lucCage and 10 nM of lucKey. Right, detection over a wide range of analyte concentrations by changing the biosensor (lucCage plus lucKey) concentration (coloured lines). All experiments were performed in triplicate, representative data are shown, and data are mean ± s.d.

With this structural validation of the design concept, we next sought to develop sensors for Botulinum neurotoxin B (BoNT/B), the immunoglobulin Fc domain and the HER2 receptor. We grafted a de novo designed binder for Botulinum neurotoxin (Bot.0671.2)21, the C domain of the generic antibody-binding protein A23 and a HER2-binding affibody24 into lucCage. After screening a few designs for each target (Extended Data Figs. 4, 5), we obtained highly sensitive lucCages (lucCageBot, lucCageProA and lucCageHER2) that can detect BoNT/B (Fig. 2b, Extended Data Fig. 4), human IgG Fc domain (Fig. 2c, Extended Data Fig. 5a–d), and HER2 receptor (Fig. 2d, Extended Data Fig. 5e–h), respectively, demonstrating the modularity of the platform. The designed sensors respond within minutes after the addition of the target, and their sensitivity can be tuned by changing the concentration of lucCage and lucKey (Fig. 2). With further development, these sensors could enable the rapid and low-cost detection of botulinum neurotoxins in the food industry25, and detection of serological levels of soluble HER2 (>15 ng ml−1; within the detection range of lucCageHER2) associated with metastatic breast cancer26.

lucCage sensor for cardiac troponin

We next designed sensors for cardiac troponin I, which is the standard early diagnostic biomarker for acute myocardial infarction27. We took advantage of the high-affinity interactions between cardiac troponins T, C and I (cTnT, cTnC and cTnI, respectively) (Fig. 3a) and designed 11 biosensor candidates by inserting 6 truncated cTnT sequences at different latch positions (Extended Data Fig. 6a). The best candidate, lucCageTrop627, was able to detect cTnI but not at sufficiently low levels for clinical use as the rule-in and rule-out levels of cTnI assay for the diagnosis of patients with acute myocardial infarction are in the low picomolar range27. Because the limit of detection (LOD) of our sensor platform is about 0.1 × Kd of the latch–target affinity (KLT), we sought to improve the sensitivity of lucCageTrop627 by increasing the cTnI binding affinity. We fused cTnC to the C terminus of the sensor to take advantage of the high-affinity interaction between the three cardiac troponins (Extended Data Fig. 6b–d). The resulting sensor, lucCageTrop, has a single-digit picomolar LOD that is suitable for the quantification of clinical samples (Fig. 3b, Extended Data Fig. 6e, f).

Fig. 3: Design and characterization of biosensors for cTnI and an anti-HBV antibody.
figure 3

a, Design of cTnI sensor. Left, structure of cardiac troponin (PDB code 4Y99). cTnT, cTnC and cTnI are shown in cyan, green and magenta, respectively. Right, design model of lucCageTrop. b, Left, luminescence signal increases after the addition of 1 nM cTnI to 0.1 nM lucCageTrop plus lucKey. Right, wide detection range accessible by changing the concentration of the sensor components (coloured lines). Grey area indicates the cTnI concentration range relevant to the diagnosis of acute myocardial infarction27; the dotted line indicates the clinical cut-off for acute myocardial infarction defined by the World Health Organization (WHO) (0.6 ng ml−1, 25 pM). c, HBV sensor design models (gold, SmBiT; grey, linker; magenta, HBV preS1 epitope). d, lucCageHBVα with two epitope copies has higher affinity by biolayer interferometry for the anti-HBV antibody HzKR127-3.2 (Kd = 0.68 nM) than lucCageHBV (Kd = 20 nM). e, Left, luminescence signal increases after the addition of 50 nM anti-HBV antibody to 1 nM lucCageHBVα plus lucKey. Right, sensitive anti-HBV antibody detection over a wide concentration range. f, Mechanism for the detection of preS1 using lucCageHBV. g, Kinetics of bioluminescence after the addition of the anti-HBV antibody (‘1’) and subsequently preS1 (‘2’), which decreases bioluminescence by competing with the sensor for the antibody. h, The detection of preS1 can be achieved over the relevant post-HBV infection concentration levels (grey area) by varying the concentration of antibody (indicated by coloured labels). All experiments were performed in triplicate, representative data are shown, and data are mean ± s.d.

lucCage sensors for HBV and SARS-CoV-2 antibodies

The detection of specific antibodies is important for monitoring the spread of a pathogen in a population28, the success of vaccination29, and levels of therapeutic antibodies9. To adapt our system for serological antibody analyses, we sought to incorporate linear epitopes recognized by the antibodies of interest into lucCage. We first developed a sensor for antibodies against the preS1 domain of the hepatitis B virus (HBV) surface protein L30. The best of eight designs tested, termed ‘lucCageHBV’, had an approximately 150% increase in luciferase activity after the addition of the anti-HBV antibody HzKR127-3.231 (Extended Data Fig. 7a–d). To further improve the dynamic range and LOD of lucCageHBV (Extended Data Fig. 7e), we introduced a second copy of the peptide at the end of the latch to increase latch affinity with the bivalent antibody (KLT) (Fig. 3c, d). The resulting design, termed ‘lucCageHBVα’, had a LOD of 260 pM and a dynamic range of 225% (Fig. 3e, Extended Data Fig. 7g–i), with a luminescence intensity easily detectable with a camera (Extended Data Fig. 7j). Because the concentrations of most therapeutic antibodies in serum are in the low micromolar to nanomolar range9, this platform should be useful for monitoring the concentrations of therapeutic antibodies in circulation32.

We next sought to use the lucCageHBV sensor to detect HBV surface antigen. Because our sensors are under thermodynamic control, we hypothesized that the pre-assembled sensor–antibody complex would re-equilibrate in the presence of the target HBV surface antigen protein preS1, with antibody redistributing to bind free preS1 instead of the epitope on lucCageHBV (Fig. 3f). The luminescence of the lucCageHBV plus HzKR127-3.2 mixture decreased shortly after the addition of the preS1 domain (Fig. 3g); the sensitivity of this readout enabled quantification of the preS1 concentration in the clinically relevant range33 (Fig. 3h, Extended Data Fig. 7f).

The COVID-19 pandemic has created an urgent need for diagnostic tools for both the SARS-CoV-2 virus and antiviral antibodies3. To design sensors for anti-SARS-CoV-2 antibodies, we first identified from the literature highly immunogenic linear epitopes in the proteomes of SARS-CoV34,35 and SARS-CoV-236 that are not present in ‘common’ strains of Coronaviridae. Among these, we focused on two epitopes in the membrane (M) and nucleocapsid (N) proteins that are recognized by sera from patients with SARS and COVID-1935,36 and for which cross-reactive animal-derived antibodies are commercially available (Methods). We designed sensors for each epitope and identified designs that specifically responded to anti-membrane and anti-nucleocapsid antibodies (Extended Data Fig. 8a, b). These sensors reached full signal in 2–5 min and had an approximately 50–70% dynamic range in response to low nanomolar amounts of antibodies (Fig. 4a, b, Extended Data Fig. 8c, d).

Fig. 4: Design of high-specificity biosensors for anti-SARS-CoV-2 antibodies and SARS-CoV-2 viral proteins.
figure 4

a, Left, lucCageSARS2-M sensor incorporates two copies of the SARS-CoV-2 membrane protein 1–17 epitope (red) connected with a flexible spacer. Middle, kinetics of luminescence activation of 50 nM lucCageSARS2-M plus lucKey after the addition of 100 nM anti-SARS-CoV-1-M rabbit polyclonal antibodies (pAb) that cross-react with residues 1–17 of the SARS-CoV-2 membrane protein. Right, response of 5 nM lucCageSARS2-M plus lucKey to varying concentrations of target anti-M polyclonal antibody. b, Left, lucCageSARS2-N incorporates two copies of the SARS-CoV-2 nucleocapsid protein 369–382 epitope (light blue). Middle, kinetics of luminescence activation of 50 nM lucCageSARS2-N plus lucKey after the addition of 100 nM anti-SARS-CoV-1-N mouse monoclonal antibody (clone 18F629.1) that recognizes the epitope. Right, response of 50 nM lucCageSARS2-N plus lucKey to varying concentration of anti-N monoclonal antibody. c, Left, lucCageRBD incorporates a de novo SARS-CoV-2 RBD binder4 (LCB1, magenta). Middle, luminescence intensities increase after the addition of 16.7 nM SARS-CoV-2 RBD or trimeric spike protein to a mixture of 1 nM lucCageRBD plus lucKey. Right, detection over a range of analyte concentrations in buffer, 10% synthetic nasal matrix38 or 10% serum. d, Biosensor specificity. Each sensor at 1 nM was incubated with 50 nM of its cognate target (magenta lines) and the targets for the other biosensors (grey lines). Targets are BCL-2, BoNT/B, human IgG Fc, HER2, cTnI, anti-HBV antibody (HzKR127-3.2), anti-SARS-CoV-1-M polyclonal antibody and SARS-CoV-2 RBD. All experiments were performed in triplicate, representative data are shown, and data are mean ± s.d.

lucCage sensors for SARS-CoV-2 spike protein

To create sensors that can detect SARS-CoV-2 viral particles directly, we integrated a de novo designed picomolar affinity binder to the receptor-binding domain (RBD) of the SARS-CoV-2 spike protein named LCB14 into the lucCage format (Fig. 4c). Of 13 candidates tested, the best, which we refer to as ‘lucCageRBD’, could detect both monomeric RBD and the full trimeric SARS-CoV2 spike protein37 with 15 pM and 47 pM LOD, respectively, and a more than 1,700% dynamic range for the RBD detection (Fig. 4c, Extended Data Fig. 9). We further increased the dynamic range of lucCageRBD to 5,300% by tuning the cage–key affinity (KCK) through shortening lucKey (Extended Data Fig. 10a–c). In addition to virus detection, the RBD sensor could also be used to monitor antibody generation in response to vaccination using a competition format analogous to that described above for the detection of HBV antibodies (Fig. 3f)—the ability to quantify responses over a wide dynamic range and to distinguish neutralizing antibodies binding at the ACE2-binding  site on the RBD (and hence competing with LCB1 in the sensor) from non-neutralizing antibodies binding elsewhere are potential advantages over lateral flow assays.

To evaluate the ability of our sensor platform to function in complex biological matrices, we compared RBD detection by lucCageRBD in buffer, simulated nasal matrix38, and human serum, and observed only a minor reduction in the latter two conditions (Fig. 4c). Following a suggestion by M. Merkx39, we controlled for variation in absolute luminescence signal in spiked serum samples from four different donors and spiked simulated nasal matrix using a BRET reference40 for internal calibration, and found that with such calibration the RBD could be accurately quantified without compromising the sensor dynamic range (Extended Data Fig. 11). These results suggest that the lucCage system could be used in point-of-care diagnostic devices.

Sensor specificity

To test the specificity of the designed biosensors, we measured the activation kinetics of each lucCage in response to each of the targets one at a time. Each sensor responded rapidly and sensitively to its cognate target, but not to any of the others (Fig. 4d). For the most part, the actual sensors (Supplementary Tables 2, 3) performed as predicted by the simple thermodynamic model; for example, experiments at varying key and sensor concentrations suggest little coupling between parameters. However, there is considerable variation between different sensors in the level of activation at saturating target concentrations or high lucKey concentrations, which for most is lower than that expected for the complete luciferase reconstitution predicted by the model (Extended Data Fig. 10d–g, Supplementary Table 4). This may be a consequence of steric interference between target binding to the latch and luciferase reconstitution as the target binding motif and the luciferase SmBiT are adjacent to each other in the latch; such interference could be resolved by increasing the separation between the two in the switch. The potential of the lucCage system is illustrated by the high dynamic range (5,300%) and picomolar sensitivity of the lucCageRBD sensor: the near optimal Kopen value results in a very low background in the absence of target without compromising the extent of activation at low target concentrations.


It is instructive to put our sensors in the context of the many protein-based biosensor platforms that have been developed over the years with considerable success (Supplementary Discussion, Supplementary Table 5). Our sensor platform is based on the thermodynamic coupling between defined closed and open states of the system, and thus, its sensitivity depends on the free energy change that occurs after the sensing domain binds to the target but not the specific geometry of the binding interaction (the semi-synthetic small molecule sensors10,11 also have this property). This enables the incorporation of various binding modalities—including small peptides, globular mini-proteins, antibody epitopes and de novo designed binders—to generate sensitive sensors for a wide range of protein targets with little or no optimization. For point of care applications, our system, similar to other bioluminescence-based protein biosensor platforms8, has the advantages of being homogeneous, no-wash, and a nearly instantaneous readout; the quantification of luminescence can be performed with inexpensive and accessible devices such as the camera of a mobile phone8. In hospital settings, the ability to modularly design sensors with identical readouts for diverse targets could enable the quick readout of large numbers of different compounds using an array of hundreds of different sensors.

Until recently, the focus of de novo protein design was the design of proteins with new structures that correspond to single deep free energy minima; our results highlight the progress in the field that now enables more complex multistate systems to be readily generated. Similar to other de novo designed proteins, our sensors are expressed at high levels in cells and are very stable41, which should considerably facilitate their manufacturing and distribution. As highlighted by the outstanding performance of the lucCageRBD sensor, there is a strong synergy between the general ‘molecular device’ architecture of our platform and de novo designed high-affinity mini-protein binders4,21 (these de novo mini-proteins are also effective with other platforms42). As the power of computational design continues to increase, it should become possible to detect an ever wider range of targets with higher sensitivity using lucCage sensors. Beyond biosensors, our results highlight the potential of de novo protein design to create more general solutions for current day challenges than can be achieved by repurposing native proteins that have evolved to solve completely different challenges.


Design of the sensor system (lucCage and lucKey)

The low affinity SmBiT 114 (VTGYRLFEEIL)18 was grafted into the latch of the asymmetric LOCKR switch previously described2 using GraftSwitchMover, a RosettaScripts-based protein design algorithm (see Supplementary Methods for details). The grafting sampling range was assigned between residues 300 and 330. The resulting designs were energy-minimized, visually inspected and selected for subsequent gene synthesis, protein production and biochemical analyses. The best SmBit position on the latch was experimentally determined to be an insertion at residue 312, as described in Extended Data Fig. 2. This design was named lucCage. lucKey was assembled by genetically fusing the LgBit of NanoLuc18 to the key peptide previously described2. All protein sequences are listed in Supplementary Table 6.

Computational grafting of sensing domains into lucCage

For peptides and epitopes, the amino acid sequence for each sensing domain was grafted using Rosettascripts43 GraftSwitchMover into all α-helical registers between residues 325 and 359 of lucCage. In the cases in which the desired sequence to be inserted exceeded the length of the lucCage latch, we made use of Rosetta Remodel44 to model the C terminus extension of lucCage (see Supplementary Methods for details). The resulting lucCages were energy-minimized using Rosetta fast relax45, visually inspected and typically less than ten designs were selected for subsequent protein production and biochemical characterization.

For protein domains, the main secondary structure element segment forming the interface of the binding protein domain with the target was identified. The amino acid sequence was extracted and grafted into lucCage using the GraftSwitchMover or Rosetta Remodel as described above. Then, we used MergePDBMover and Pymol 2.0 to align, model and visualize the full-length binding domain in the context of the switch (see Supplementary Methods for details). The designs were energy-minimized using Rosetta fast relax and visually inspected for selection.

Synthetic gene construction

The designed protein sequences were codon-optimized for E. coli expression and ordered as synthetic genes in pET21b+ or pET29b+ E. coli expression vectors. The synthetic gene was inserted at the NdeI and XhoI sites of each vector, including an N-terminal hexahistidine tag followed by a TEV protease cleavage site and a stop codon was added at the C terminus.

General procedures for bacterial protein production and purification

The E. coli Lemo21(DE3) strain (NEB) was transformed with a pET21b+ or pET29b+ plasmid encoding the synthesized gene of interest. Cells were grown for 24 h in LB medium supplemented with carbenicillin or kanamycin. Cells were inoculated at a 1:50 ml ratio in the Studier TBM-5052 autoinduction medium supplemented with carbenicillin or kanamycin, grown at 37 °C for 2–4 h, and then grown at 18 °C for an additional 18 h. Cells were collected by centrifugation at 4,000g at 4 °C for 15 min and resuspended in 30 ml lysis buffer (20 mM Tris-HCl pH 8.0, 300 mM NaCl, 30 mM imidazole, 1 mM PMSF, 0.02 mg ml−1 DNase). Cell resuspensions were lysed by sonication for 2.5 min (5 s cycles). Lysates were clarified by centrifugation at 24,000g at 4 °C for 20 min and passed through 2 ml of Ni-NTA nickel resin (Qiagen, 30250) pre-equilibrated with wash buffer (20 mM Tris-HCl pH 8.0, 300 mM NaCl, 30 mM imidazole). The resin was washed twice with 10 column volumes (CV) of wash buffer, and then eluted with 3 CV of elution buffer (20 mM Tris-HCl pH 8.0, 300 mM NaCl, 300 mM imidazole). The eluted proteins were concentrated using Ultra-15 Centrifugal Filter Units (Amicon) and further purified by using a Superdex 75 Increase 10/300 GL (GE Healthcare) size exclusion column in TBS (25 mM Tris-HCl pH 8.0, 150 mM NaCl). Fractions containing monomeric protein were pooled, concentrated, and snap-frozen in liquid nitrogen and stored at −80 °C.

In vitro bioluminescence characterization

A Synergy Neo2 Microplate Reader (BioTek) was used for all in vitro bioluminescence measurements. Assays were performed in 50% DPBS with calcium (Gibco) plus 50% Nano-Glo (Promega) assay buffer for cTnI sensors and 50% HBS-EP (GE Healthcare Life Sciences) plus 50% Nano-Glo assay buffer was used for other sensors. 10× lucCage, 10× lucKey and 10× target proteins of desired concentrations were first prepared from stock solutions. For each well of a white opaque 96-well plate, 10 μl of 10× lucCage, 10 μl of 10× lucKey and 20 μl of buffer were mixed to reach the indicated concentration and ratio. The lucCage and lucKey components were incubated for 60 min at room temperature to enable pre-equilibration. The plate was centrifuged at 1,000g for 1 min and incubated at room temperature for a further 10 min. Then, 50 μl of 50× diluted furimazine (Nano-Glo luciferase assay reagent, Promega) was added to each well. For assays containing serum or simulated nasal matrix (110 mM NaCl, 1% (w/v) mucin, 10 μg ml−1 human genomic DNA38), buffer composition was replaced by the biological matrix. Bioluminescence measurements in the absence of target were taken every 1 min after injection (0.1 s integration and 10 s shaking during intervals). After around 15 min, 10 μl of serially diluted 10× target protein plus a blank was injected and bioluminescence kinetic acquisition continued for a total of 2 h. To derive half-maximal effective concentration (EC50) values from the bioluminescence-to-analyte plot, the top three peak bioluminescence intensities at individual analyte concentrations were averaged, subtracted from blank, and used to fit the sigmoidal 4PL curve. To calculate the LOD, the linear region of bioluminescence responses of sensors to its analyte was extracted and a linear regression curve was obtained. It was used to derive the standard deviation of the response and the slope of the calibration curve (S). The LOD was determined as: 3 × (s.d./S).

Detection of spiked RBD in human serum specimens

Serum specimens were derived from excess plasma or sera from adults (>18 years) of both genders provided by the Director of the Clinical Chemistry Division, the hospital of University Washington. All anonymized donor specimens were provided de-identified. Because the donors consented to have their excess specimens be used for other experimental studies, they could be transferred to our study without additional consent. All samples were passed through 0.22-μm filters before use. Ten microlitres of 10× serial diluted monomeric RBD (167–0.69 nM), 5 μl of 20× lucCage (20 nM), 5 μl of 20× lucKey (20 nM), 5 μl of 20× Antares2 (2 nM), and 10, 20, 25 or 50 μl of human donor serum or simulated nasal matrix were mixed with 1:1 HBS:Nano-Glo assay buffer to reach a total volume of 75 μl. The plate was centrifuged at 1,000g for 1 min. Then, 25 μl of 25× diluted furimazine in buffer was added to each well. Bioluminescence signals were recorded from both 470/40 nm and 590/35 nm channels every 1 min for a total of 1 h. The ratio at each time point was calculated by the equation described in Extended Data Fig. 11b. Monomeric SARS-CoV-2 RBD was expressed and purified as previously described46.

Biolayer interferometry

Protein–protein interactions were measured by using an Octet RED96 System (ForteBio) using streptavidin-coated biosensors (ForteBio). Each well contained 200 μl of solution, and the assay buffer was HBS-EP+ buffer (GE Healthcare Life Sciences, 10 mM HEPES pH 7.4, 150 mM NaCl, 3 mM EDTA, 0.05% (v/v) surfactant P20) plus 0.5% non-fat dry milk blotting grade blocker (BioRad). The biosensor tips were loaded with analyte peptide or protein at 20 μg ml−1 for 300 s (threshold of 0.5 nm response), incubated in HBS-EP+ buffer for 60 s to acquire the baseline measurement, dipped into the solution containing cage and/or key for 600 s (association step) and dipped into the HBS-EP+ buffer for 600 s (dissociation steps). The binding data were analysed with the ForteBio Data Analysis Software version

Design and characterization of lucCageBIM

The BIM peptide sequence (EIWIAQELRRIGDEFNAYYA) was threaded into the lucCage scaffold as described in ‘Computational grafting of sensing domains into lucCage’. The selected designs were expressed in E. coli, purified and characterized for luminescence activation. The bioluminescence detection signal was measured for each design lucCage at 20 nM mixed with lucKey at 20 nM, in the presence or absence of target BCL-2 protein at 200 nM. Recombinant BCL-2 was produced as previously described47.

Design and characterization of lucCageHER2, lucCageProA, lucCageBot and lucCageRBD

The main binding motifs of the Bot.0671.2 de novo binder, Staphylococcus aureus protein A domain C (SpaC), the HER2 antibody and the de novo RBD binder LCB1 were threaded into lucCage as described in ‘Computational grafting of sensing domains into lucCage’ (see Supplementary Tables 3 and 6 for sequences). The selected designs were expressed in E. coli, purified and characterized for luminescence activation. The designs were screened by measuring bioluminescence signal for each design lucCage at 20 nM mixed with lucKey at 20 nM, in the presence or absence of 200 nM target protein. The target proteins used were: Botulinum neurotoxin B HcB expressed as previously described48, human IgG1 Fc-HisTag (AcroBiosystems, IG1-H5225) and human HER2-HisTag (AcroBiosystems, HE2-H5225). Monomeric SARS-CoV-2 RBD and the trimeric SARS-CoV-2 spike protein (Hexapro pre-stabilized version37) were expressed and purified as previously described46.

Design and characterization of lucCageTrop

The cTnT binding motif sequence was truncated into fragments of different length (Extended Data Fig. 6) and threaded into the lucCage scaffold as described in ‘Computational grafting of sensing domains into lucCage’. The selected designs were expressed in E. coli, purified and characterized for luminescence activation. The designs were screened by measuring bioluminescence signal for each design lucCage at 20 nM mixed with lucKey at 20 nM in the presence or absence of 100 nM cTnI (Genscript, Z03320-50). Subsequently, lucCageTrop, an improved version by fusion to cTnC, was created by genetically fusing the following sequence to the C terminus of lucCageTrop627.

Design and characterization of lucCageHBV and lucCageHBVα

The binding motif (GANSNNPDWDFN) of the preS1 domain was threaded into the lucCage scaffold at every position after residues 336 using the Rosetta GraftSwitchMover. Following the Rosetta FastRelax protocol, eight designs were selected for protein production. The designs were screened by measuring bioluminescence signal for each design lucCage (20 nM) and lucKey (20 nM) in the presence or absence of the anti-HVB antibody HzKR127-3.2 (100 nM) to select lucCageHBV. Subsequently, lucCageHBVα was constructed by genetically fusing a sequence containing a second antigenic motif (GGSGGGSSGFGANSNNPDWDFNPN) to lucCageHBV.

Design and characterization of lucCageSARS2-M and lucCageSARS2-N

Antigenic epitopes of the SARS-CoV-2 membrane protein (amino acids 1–31, 1–17 and 8–24) and the nucleocapsid protein (amino acids 368–388 and 369–382) were computationally grafted into lucCage as described in ‘Computational grafting of sensing domains into lucCage’. The selected designs were expressed in E. coli, purified and characterized for luminescence activation. All designs at 50 nM were mixed with 50 nM lucKey and experimentally screened for an increase in luminescence in the presence of rabbit anti-SARS-CoV membrane polyclonal antibodies (ProSci, 3527) at 100 nM or mouse anti-SARS-CoV nucleocapsid monoclonal antibody (clone 18F629.1, NovusBio NBP2-24745) at 100 nM.

Design and characterization of sCageHA variants

HB1.9549.2 was embedded into the parental six-helix bundle for sCage design at different positions along the latch helix of the scaffold. To promote more favourable intramolecular interactions, three consecutive residues on the latch were intentionally substituted with glycine to allow for conformational freedom. The five designs were produced in E. coli. Biolayer interferometry analysis was performed with purified cages (1 μM) and biotinylated influenza A H1 HA21 loaded onto streptavidin-coated biosensor tips (ForteBio) in the presence or absence of the key (2 μM) using an Octet instrument (ForteBio).

Production and purification of HzKR127-3.2

The synthetic VH and VL DNA fragments were subcloned into the pdCMV-dhfrC-cA10A3 plasmid containing the human Cγ1 and Cγ DNA sequences. The vector was introduced into HEK 293F cells using Lipofectamine (Invitrogen), and the cells were grown in FreeStyle 293 (GIBCO) in 5% CO2 in a 37 °C humidified incubator. The culture supernatant was loaded onto a protein A-sepharose column (Millipore), and the bound antibody was eluted by the addition of 0.2 M glycine-HCl (pH 2.7), followed by immediate neutralization with 1 M Tris–HCl (pH 8.0). The solution was dialysed against 10 mM HEPES-NaOH (pH 7.4), and the purity of the protein was analysed by SDS–PAGE.

Production and purification of the preS1 domain

The DNA fragment encoding the preS1 domain (residues 1–56) was cloned into the pGEX-2T (GE Healthcare) plasmid, and the protein was produced in the E. coli BL21(DE3) strain (NEB) at 18 °C as a fusion protein with glutathion-S-transferase (GST) at the N terminus. The cell lysates were prepared in a buffer solution (25 mM Tris-HCl pH 8.0, 300 mM NaCl), and clarified supernatant was loaded onto GSTBind Resin (Novagen). The GST–preS1 domain was eluted with the same buffer containing additional 10 mM reduced glutathione, further purified using a Superdex 75 Increase 10/300 GL (GE Healthcare) size-exclusion column, and concentrated to 34 μM.

Production of SCageHA_267-1S and its variants

sCageHA_267-1S and sCageHA_267-1S(E99Y/T144Y) were expressed at 18 °C in the E. coli LEMO21(DE3) strain (NEB) as a fusion protein containing a (His)10-tagged cysteine protease domain (CPD) derived from Vibrio cholerae49 at the C terminus. The protein was purified using HisPur nickel resin (Thermo), a HiTrap Q anion exchange column (GE Healthcare) and a HiLoad 26/60 Superdex 75 gel filtration column (GE Healthcare). For selenomethionine (SelMet)-labelling, an I30M mutation was introduced additionally to generate a sCageHA_267-1S(E99Y/T144Y/I30M) variant. This protein was expressed in the E. coli B834 (DE3) RIL strain (Novagen) in the minimal medium containing SeMet, and purified according to the same procedure for purifying the other variants.

Crystallization and structure determination of sCageHA_267-1S

Two point mutations (Glu99Tyr and Thr144Tyr) were introduced in an attempt to induce favourable crystal packing interactions. Good-quality single crystals of sCageHA_267-1S(E99Y/T144Y/I30M) were obtained in a hanging-drop vapour-diffusion setting by micro-seeding in a solution containing 11% (v/v) ethanol, 0.25 M NaCl, 0.1 M Tris-HCl (pH 8.5). The crystals required strict maintenance of the temperature at 25 °C. For cryoprotection, the crystals were soaked briefly in the crystallization solution supplemented with 15% 2,3-butanediol and flash-cooled in the liquid nitrogen. A single-wavelength anomalous dispersion dataset was collected at the Se absorption peak and processed with HKL200050. Se positions and initial electron density map were calculated using the AutoSol module in PHENIX51. The model building and structure refinement were performed by using COOT52 and PHENIX.

Statistical analysis

No statistical methods were used to pre-determine the sample size. No sample was excluded from data analysis, and no blinding was used. De-identified clinical serum samples were randomly used for spiking in target proteins. Results were successfully reproduced using different batches of pure proteins on different days. Unless otherwise indicated, data are shown as mean ± s.d., and error bars in figures represent s.d. of technical triplicate. Biolayer interferometry data were analysed using ForteBio Data Analysis Software version All data were analysed and plotted using GraphPad Prism 8.

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this paper.