De novo design of luciferases using deep learning

1Bioluminescent light produced by the enzymatic oxidation of a luciferin substrate by luciferases is widely used for bioassays and imaging in biomedical research. Because no excitation light source is needed, luminescent photons are produced in the dark; this results in higher sensitivity than fluorescence imaging in live animal models and in biological samples in which autofluorescence or phototoxicity is a concern,. However, the development of luciferases as molecular probes has lagged behind that of well-developed fluorescent protein toolkits for a number of reasons: (i) very few native luciferases have been identified,; (ii) many of those that have been identified require multiple disulfide bonds to stabilize the structure and are therefore prone to misfolding in mammalian cells; (iii) most native luciferases do not recognize synthetic luciferins with more desirable photophysical properties; and (iv) multiplexed imaging to follow multiple processes in parallel using mutually orthogonal luciferase–luciferin pairs has been limited by the low substrate specificity of native luciferases,.

2We sought to use de novo protein design to create luciferases that are small, highly stable, well-expressed in cells, specific for one substrate and need no cofactors to function. We chose a synthetic luciferin, diphenylterazine (DTZ), as the target substrate because of its high quantum yield, red-shifted emission, favourable in vivo pharmacokinetics, and lack of required cofactors for light emission. Previous computational enzyme design efforts have primarily repurposed native protein scaffolds in the Protein Data Bank (PDB),, but there are few native structures with binding pockets appropriate for DTZ, and the effects of sequence changes on native proteins can be unpredictable (designed helical bundles have also been used as enzyme scaffolds–, but these are limited in number and most do not have pockets that are suitable for DTZ binding). To circumvent these limitations, we set out to generate large numbers of small and stable protein scaffolds with pockets of the appropriate size and shape for DTZ, and with clear sequence–structure relationships to facilitate subsequent active-site incorporation. To identify protein folds that are capable of hosting such pockets, we first docked DTZ into 4,000 native small-molecule-binding proteins. We found that many nuclear transport factor 2 (NTF2)-like folds have binding pockets with appropriate shape complementarity and size for DTZ placement (pink dashes in Fig. 1e), and hence selected the NTF2-like superfamily as the target topology.

3Native NTF2 structures have a range of pocket sizes and shapes but also contain features that are not ideal, such as long loops that compromise stability. To create large numbers of ideal NTF2-like structures, we developed a deep-learning-based ‘family-wide hallucination’ approach that integrates unconstrained de novo design, and Rosetta sequence-design approaches to enable the generation of an essentially unlimited number of proteins that have a desired fold (Fig. 1a). The family-wide hallucination approach used the de novo sequence and structure discovery capability of unconstrained protein hallucination, for loop and variable regions, and structure-guided sequence optimization for core regions. We used the trRosetta structure prediction neural network, which is effective in identifying experimentally successful de-novo-designed proteins and hallucinating new globular proteins of diverse topologies. Starting from the sequences of 2,000 naturally occurring NTF2s, we carried out Monte Carlo searches in sequence space, at each step making a sequence change and predicting the structure using trRosetta. As the loss function guiding search, we used the confidence of the neural network in the predicted structure (as in our previous free hallucination study) supplemented with a topology-specific loss function over core residue pair geometries (see Supplementary Methods); in the loop regions, we also allowed the number of residues to vary, which resulted in short near ideal loops. To further encode structural specificity, we incorporated buried, long-range hydrogen-bonding networks. The resulting 1,615 family-wide hallucinated NTF2 scaffolds provided more shape-complementary binding pockets for DTZ than did native small-molecule-binding proteins (Fig. 1e). This method samples protein backbones that are closer to native NTF2-like proteins (Fig. 1f) and that have better scaffold quality metrics than those produced in a previous non-deep-learning energy-based approach (Fig. 1g).

4Computational enzyme design generally starts from an ideal active site or theozyme consisting of protein functional groups surrounding the reaction transition state that is then matched into a set of existing scaffolds,. However, the detailed catalytic geometry of native marine luciferases is not well understood because only a handful of apo structures and no holo structures with luciferin substrates have been solved (at the time of this study)–. Both quantum chemistry calculations, and experimental data, suggest that the chemiluminescent reaction proceeds through an anionic species and that the polarity of the surroundings can substantially alter the free energy of the subsequent single-electron transfer (SET) process with triplet molecular oxygen (3O2). Guided by these data (Extended Data Fig. 1), we sought to design a shape-complementary catalytic site that stabilizes the anionic state of DTZ and lowers the SET energy barrier, assuming that the downstream dioxetane light emitter thermolysis steps are spontaneous. To stabilize the anionic state, we focused on the placement of the positively charged guanidinium group of an arginine residue to stabilize the developing negative charge on the imidazopyrazinone group.

5To computationally design such active sites into large numbers of hallucinated NTF2 scaffolds, we first generated an ensemble of anionic DTZ conformers (Fig. 1b). Next, around each conformer, we used the RifGen method, to enumerate rotamer interaction fields (RIFs) on three-dimensional grids consisting of millions of placements of amino acid side chains making hydrogen-bonding and nonpolar interactions with DTZ (Fig. 1c). An arginine guanidinium group was placed adjacent to the N1 atom of the imidazopyrazinone group to stabilize the negative charge. RifDock was then used to dock each DTZ conformer and associated RIF in the central cavity of each scaffold to maximize protein–DTZ interactions. An average of eight side-chain rotamers, including an arginine residue to stabilize the anionic imidazopyrazinone core, were positioned in each pocket (Supplementary Fig. 2a). For the top 50,000 docks with the most favourable side chain–DTZ interactions, we optimized the remainder of the sequence using RosettaDesign (Fig. 1d) for high-affinity binding to DTZ with a bias towards the naturally observed sequence variation to ensure foldability. During the design process, pre-defined hydrogen-bond networks (HBNets) in the scaffolds were kept intact for structural specificity and stability, and interactions of these HBNet side chains with DTZ were explicitly required in the RifDock step to ensure the preorganization of residues that are essential for catalysis. In the first sequence-design step, the identities of all RIF and HBNet residues were kept fixed, and the surrounding residues were optimized to hold the side chain–DTZ interactions in place and maintain structural specificity. In the second sequence-design step, the RIF residue identities (except the arginine) were also allowed to vary, as Rosetta can identify apolar and aromatic packing interactions that were missed in the RIF owing to binning effects. During sequence design, the scaffold backbone, side chains and DTZ substrate were allowed to relax in Cartesian space. After sequence optimization, the designs were filtered on the basis of ligand-binding energy, protein–ligand hydrogen bonds, shape complementarity and contact molecular surface, and 7,648 designs were selected and ordered as pooled oligos for experimental screening.

6Oligonucleotides encoding the two halves of each design were assembled into full-length genes and cloned into an Escherichia coli expression vector (see Supplementary Methods). A colony-based screening method was used to directly image active luciferase colonies from the library and the activities of selected clones were confirmed using a 96-well plate expression (Extended Data Fig. 2). Three active designs were identified; we refer to the most active of these as LuxSit (from the Latin lux sit, ‘let light exist’), which at 117 residues (13.9 kDa) is, to our knowledge, smaller than any previously described luciferase. Biochemical analysis, including SDS–PAGE and size-exclusion chromatography (Fig. 2a,b and Extended Data Fig. 3), indicated that LuxSit is highly expressed in E. coli, soluble and monomeric. Circular dichroism (CD) spectroscopy showed a strong far-ultraviolet CD signature, suggesting an organized α-β structure. CD melting experiments showed that the protein is not fully unfolded at 95 °C, and that the full structure is regained when the temperature is dropped (Fig. 2c). Incubation of LuxSit with DTZ resulted in luminescence with an emission peak at around 480 nm (Fig. 2d), consistent with the DTZ chemiluminescence spectrum. Although we were not able to determine the crystal structure of LuxSit, the structure predicted by AlphaFold2 (ref. ) is very close to the design model at the backbone level (root-mean-square deviation (RMSD) = 1.35 Å) and over the side chains interacting with the substrate (Fig. 2e). The designed LuxSit active site contains Tyr14–His98 and Asp18–Arg65 dyads, with the imidazole nitrogen atoms of His98 making hydrogen-bond interactions with Tyr14 and the O1 atom of DTZ (Fig. 2f). The centre of the Arg65 guanidinium cation is 4.2 Å from the N1 atom of DTZ and Asp18 forms a bidentate hydrogen bond to the guanidinium group and backbone N–H of Arg65 (Fig. 2g).

7We next sought to apply the knowledge gained from designing LuxSit to create 2-deoxycoelenterazine (h-CTZ)-specific luciferases. Because the molecular shape of h-CTZ is different from that of DTZ, we created an additional set of NTF2 superfamily scaffolds (see Supplementary Methods) with matching pocket shapes and high model confidence (AlphaFold2-predicted local-distance difference test (pLDDT) > 92). We then installed catalytic sites in these scaffolds and designed the first shell-protein side chain–h-CTZ interactions using the histidine and arginine substrate interaction geometries that were most successful in the first round for DTZ. To design the remainder of the sequence, we used ProteinMPNN, which can result in better stability, solubility and accuracy than RosettaDesign. After filtering on the basis of the AlphaFold2-predicted pLDDT, Cα RMSD, contact molecular surface and Rosetta-computed binding energies (see Supplementary Methods), we selected and experimentally expressed 46 designs in E. coli and identified 2 (HTZ3-D2 and HTZ3-G4) that had luciferase activity with the h-CTZ luciferin substrate. Both designs were highly soluble, monodisperse and monomeric, and the luciferase activities were of the same order of magnitude as LuxSit (Extended Data Fig. 4). The success rate increased from 3/7,648 to 2/46 sequences in the second round, probably owing to the knowledge of active-site geometry from the first round and the increased robustness of the ProteinMPNN method of sequence design.

8To better understand the contributions to the catalysis of LuxSit, the most active of our designs, we constructed a site-saturation mutagenesis (SSM) library in which each residue in the substrate-binding pocket was mutated to every other amino acid one at a time (see Supplementary Methods), and determined the effect of each mutation onluciferase activity. Figure 2f–i shows the amino acid preferences at key positions. Arg65 is highly conserved (Fig. 2g), and its dyad partner Asp18 can only be mutated to Glu (which reduces activity), suggesting that the carboxylate–Arg65 hydrogen bond is important for luciferase activity. In the Tyr14–His98 dyad (Fig. 2f), Tyr14 can be substituted with Asp and Glu, and His98 can be replaced with Asn. As all active variants had hydrogen-bond donors and acceptors at these positions, the dyads might help to mediate the electron and proton transfer required for luminescence. Hydrophobic (Fig. 2i) and π-stacking (Fig. 2h) residues at the binding interface tolerate other aromatic or aliphatic substitutions and generally prefer the amino acid in the original design, consistent with model-based affinity predictions of mutational effects (Extended Data Fig. 5). The A96M and M110V mutants (highlighted in pink) increase activity by 16-fold and 19-fold, respectively, over LuxSit (Supplementary Table 1). Optimization guided by these results yielded LuxSit-f (A96M/M110V), with a flash-type emission kinetic, and LuxSit-i (R60S/A96L/M110V), with a photon flux more than 100-fold higher than that of LuxSit (Extended Data Fig. 6). Overall, the active-site-saturation mutagenesis results support the design model, with the Tyr14–His98 and Asp18–Arg65 dyads having key roles in catalysis and the substrate-binding pocket largely conserved.

9The most active catalysts, LuxSit-i (Extended Data Fig. 3b,e,h) and LuxSit-f (Extended Data Fig. 3c,f,i), were both expressed solubly in E. coli at high levels and are monomeric (some dimerization was observed at the high protein concentration; Extended Data Fig. 3l) and thermostable (Extended Data Fig. 3j,k). Similar to native luciferases that use CTZ, the apparent Michaelis constants (Km) of both LuxSit-i and LuxSit-f are in the low-micromolar range (Fig. 3a) and the luminescent signal decays over time owing to fast catalytic turnover (Extended Data Fig. 7a). LuxSit-i is a very efficient enzyme, with a catalytic efficiency (kcat/Km) of 106 M−1 s−1. The luminescence signal is readily visible to the naked eye (Fig. 3b), and the photon flux (photons per second) is 38% greater than that of the native Renilla reniformis luciferase (RLuc) (Supplementary Table 2). The DTZ luminescent reaction catalysed by LuxSit-i is pH-dependent (Extended Data Fig. 7b), consistent with the proposed mechanism. We used a combination of density functional theory (DFT) calculations and molecular dynamics (MD) simulations to investigate the basis for LuxSit activity in more detail; the results support the anion-stabilization mechanism (Extended Data Fig. 8a and Supplementary Fig. 3a) and suggest that LuxSit-i provides better DTZ transition-state charge stabilization than LuxSit (Extended Data Fig. 8b).

10As luciferases are commonly used genetic tags and reporters for cell biological studies, we evaluated the expression and function of LuxSit-i in live mammalian cells. HEK293T cells expressing LuxSit-i-mTagBFP2 showed DTZ-specific luminescence (Fig. 3c), which was maintained after targeting of LuxSit-i-mTagBFP2 to the nucleus, membrane and mitochondria (Extended Data Fig. 9). Native and previously engineered luciferases are quite promiscuous, with activity on many luciferin substrates (Fig. 4ac and Supplementary Fig. 4); this is possibly a result of their large and open pockets (a luciferase with high specificity to one luciferin substrate has been difficult to control even with extensive directed evolution,). By contrast, LuxSit-i exhibited exquisite specificity for its target luciferin, with 50-fold selectivity for DTZ over bis-CTZ (which differs only in one benzylic carbon; MD simulations suggest that this arises from greater transition-state shape complementarity (Extended Data Fig. 8b,c and Supplementary Fig. 3b,c)), 28-fold selectivity over 8pyDTZ (differing only in one nitrogen atom) and more than 100-fold selectivity over other luciferin substrates (Fig. 4b). One of our active design for h-CTZ (HTZ3-G4) was also highly specific for its target substrate (Fig. 4c and Extended Data Fig. 4d). Overall, the specificity of our designed luciferases is much greater than that of native luciferases, or previously engineered luciferases (Supplementary Table 5).

11We reasoned that the high substrate specificity of LuxSit-i could allow the multiplexing of luminescent reporters through substrate-specific or spectrally resolved luminescent signals (Fig. 4d and Extended Data Fig. 10a,b). To investigate this possibility, we placed LuxSit-i downstream of the NF-κB response element and RLuc downstream of the cAMP response element (Fig. 4e). The addition of activators (TNF) of the NF-κB signaling pathway resulted in luminescence when cells were incubated with DTZ, while the luminescence of PP-CTZ (the substrate of RLuc) was observed only when the cAMP–PKA pathway was activated (Fig. 4f). Because DTZ and PP-CTZ emit luminescence at different wavelengths, they can in principle be combined and the two signals can be deconvoluted through spectral analysis. Indeed, we observed that activating the NF-κB signaling resulted in luminescence at the DTZ wavelength, while the addition of cAMP–PKA pathway activators (FSK) generated luminescence at the PP-CTZ wavelength, allowing us to simultaneously assess the activation of the two signaling pathways in the same sample with either cell lysates (Fig. 4g) or intact HEK293T cells (Extended Data Fig. 10c–e) by providing both substrates together. Thus, the high substrate specificity of LuxSit-i enables multiplexed reporting of diverse cellular responses.

12Computational enzyme design has been constrained by the number of available scaffolds, which limits the extent to which catalytic configurations and enzyme–substrate shape complementarity can be achieved–. The use of deep learning to produce large numbers of de-novo-designed scaffolds here eliminates this restriction, and the more accurate RoseTTAfold (ref. ) and AlphaFold2 (ref. ) should enable protein scaffolds to be generated even more effectively through family-wide hallucination and other approaches,. The diversity of shapes and sizes of scaffold pockets enabled us to consider a range of catalytic geometries and to maximize reaction intermediate–enzyme shape complementarity; to our knowledge, no native luciferases have folds similar to LuxSit, and the enzyme has high specificity for a fully synthetic luciferin substrate that does not exist in nature. With the incorporation of three substitutions that provide a more complementary pocket to stabilize the transition state, LuxSit-i has higher activity than any previous de-novo-designed enzyme, with a kcat/Km (106 M−1 s−1) in the range of native luciferases. This is a notable advance for computational enzyme design, as tens of rounds of directed evolution were required to obtain catalytic efficiencies in this range for a designed retroaldolase, and the structure was remodelled considerably; by contrast, the predicted differences in ligand–side-chain interactions between LuxSit and LuxSit-i are very subtle (Supplementary Fig. 2b; achieving such high activities directly from the computer remains a challenge in computational enzyme design). The small size, stability and robust folding of LuxSit-i makes it well-suited in luciferase fusions to proteins of interest and as a genetic tag for capacity-limited viral vectors. On the basic science side, the small size, simplicity and high activity of LuxSit-i make it an excellent model system for computational and experimental studies of luciferase catalytic mechanism. Extending the approach used here to create similarly specific luciferases for synthetic luciferin substrates beyond DTZ and h-CTZ would considerably extend the multiplexing opportunities illustrated in Fig. 4 (particularly with the recent advances in microscopy), and enable a new generation of multiplexed luminescent toolkits. More generally, our family-wide hallucination method opens up an almost unlimited number of scaffold possibilities for substrate binding and catalytic residue placement, which is particularly important when the reaction mechanism and how to promote it are not completely understood: many alternative structural and catalytic hypotheses can be readily enumerated with shape and chemically complementary binding pockets but different catalytic residue placements. Although luciferases are unique in catalysing the emission of light, the chemical transformation of substrates into products is common to all enzymes, and the approach developed here should be readily applicable to a wide variety of chemical reactions.

13Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

14Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41586-023-05696-3.