Unusually broad-spectrum small-molecule sensing using a single protein scaffold

Significance

Small-molecule biosensors are critical tools for monitoring environments, controlling cellular behavior, and enabling new biotechnologies; yet, they remain challenging to design. The plant hormone receptor Pyrabactin Resistance 1 (PYR1) has emerged as a new scaffold for creating small-molecule sensors. To define the chemical scope accessible with PYR1 and empower sensor development, we screened libraries of PYR1 variants against thousands of compounds. We identified several hundred new sensors for diverse targets, including drugs, plant metabolites, explosives, and persistent “forever” per- and polyfluoroalkyl substances (PFAS). The ease, scale, and breadth of sensed molecules position PYR1 as a privileged scaffold for creating customizable biosensors with antibody-like simplicity.

Abstract

Small-molecule sensing in plants is dominated by chemical-induced dimerization modules. In the abscisic acid (ABA) system, allosteric receptors recruit phosphatase effectors and achieve nM in vivo responses from µM receptor–ligand interactions. This sensitivity amplification could enable ABA receptors to serve as generic scaffolds for designing small-molecule sensors. To test this, we screened collections of mutant ABA-receptors against 2,726 drugs and other ligands and identified 553 sensors for 6.6% of these ligands. The mutational patterns indicate strong selection for ligand-specific binding pockets. We used these data to develop a sensor design pipeline and isolated sensors for multiple plant natural products, 2,4,6-trinitrotoluene (TNT), and “forever” per- and polyfluoroalkyl substances (PFAS). Thus, the ABA sensor system enables design and isolation of small-molecule sensors with broad chemical scope and antibody-like simplicity.

The ability to create protein binders through antibodies and designed proteins has transformed biotechnology research by making essentially any protein a targetable entity (1). The success of this strategy is partly due to the simplicity of using a single protein domain and engineering pipelines for isolating binders. Developing protein scaffolds that can be programmed to recognize small molecules as easily as antibodies can recognize antigens would empower many biotechnology and synthetic biology applications. In principle, numerous biological parts can be reprogrammed to develop small-molecule sensors, including ligand-induced transcription factors (for example, lacI) (2, 3), cell surface receptors (for example, GPCRs and DREADDs) (4), and chemical-induced dimerization (CID) systems (for example, rapamycin/FKBP/FRB) (5). Of these, CID systems are attractive because they enable facile control of many outputs via induced proximity mechanisms (5–9).

Genetic dissection of phytohormone signaling in land plants has uncovered many naturally occurring CID modules that could potentially be harnessed to engineer new sense–response capacities (10–16). We recently introduced the plant ABA receptor PYR1 (Pyrabactin resistance 1) as an engineerable scaffold for sensor design and demonstrated the ability to reprogram binding for several cannabinoid and pesticide ligands (17–20). Part of the success of the PYR1 receptor as a scaffold is due to the separation of ligand-recognition and binding from its coreceptor, HAB1 (Homolog of ABA Insensitive 1). This simplifies engineering efforts by restricting designs to a single molecular recognition surface. PYR1 also benefits from intrinsic sensitivity amplification. The native ligand, ABA, binds to the receptor with micromolar affinity, but the activated receptor conformer binds with nanomolar affinity to HAB1 in an orientation that blocks ligand dissociation, leading to an up to 100-fold increase in affinity to the ligand (21), which facilitates the identification of low-affinity binders (22). Moreover, its components are soluble and function in diverse prokaryotic and eukaryotic hosts (13, 17, 19, 23). These features coalesce to make the PYR1 CID module an effective sensor scaffold and a candidate for designing small-molecule binders akin to antibodies for protein antigens.

Here, we set out to define the chemical scope of the PYR1 binding pocket and its potential as an antibody-like scaffold for sensing small molecules. To do this, we established a miniaturized sensor-isolation pipeline that capitalizes on ligand-induced growth triggered by the physical interaction between PYR1 and HAB1 using yeast growth-based selections (Fig. 1). Using this approach in thousands of parallelized assays, we screened receptor variants against a collection of 2,726 small molecules. Together, these yielded sensors for 181 molecules, 6.6% of the ligands screened. We show that this large dataset of ligand–receptor interactions can be harnessed to isolate high-affinity sensors in a single step with antibody-like simplicity.

Fig. 1.

Large-scale sensor isolation to enable data-driven sensor design. (A) The PYR1 sensor isolation pipeline. PYR1 binds to its effector protein HAB1 in response to ABA, its native ligand. This interaction can be measured in yeast growth assays by activating a genetic circuit that rescues uracil auxotrophy. Mutant PYR1 variants that recognize new ligands are identified from receptor library pools using selection experiments (responders grow in the absence of uracil). (B) A high-throughput version of the receptor isolation approach was developed. In this, we plate ~150,000 cells onto media containing a test chemical and select colonies from the wells for retesting and subsequent sequencing. These efforts identified 553 receptors that recognize 181 unique molecules. (C) Chemoinformatic analysis of the data is used to build maps of sequence–ligand interactions. (D) The sequence–ligand maps developed can be used to drive the design of sensors. Libraries of receptors targeted to specific molecule classes are constructed using sequence profiles to design oligonucleotide pools that can be assembled using Golden Gate cloning; subsequent growth-based screens enable one-step isolation of high-affinity sensors.

To profile the binding scope of PYR1, we assembled a collection of diverse small molecules from a commercially available collection of FDA-approved drugs and natural products, and a curated set of agrochemicals and other research ligands (Dataset S1). Since commercial screening libraries often contain redundancy—for example, the same parent molecule present as different salts—we used ChemmineR (24, 25) to define the nonredundant set of molecules screened. We mapped them to the ChEMBL database (26) to identify redundancies and extract molecular properties and approval statuses. These analyses showed that the screening collection contains 2,726 small molecules. This collection was screened in duplicate against previously and newly designed PYR1 variant libraries harboring double and triple mutations localized to 19 ligand-contacting residues (see Materials and Methods and SI Appendix, Fig. S1 for details) (17). Candidate sensor strains for a given molecule were screened in duplicate at 100 µM and retested at multiple concentrations of target molecules to eliminate false positives and profile strain sensitivity; retest dose–response image data are provided in Dataset S2. Together, these screens yielded 553 distinct sensor sequences and strains that respond to 181 small molecules (Dataset S3), including 81 FDA-approved ligands, 30 plant natural products, 17 agrochemicals, and 53 research molecules, representing an overall success rate of 6.6% of the ligands screened (Fig. 2A).

Fig. 2.

The PYR1 binding pocket can be mutated to sense small molecules of broad chemical scope. (A) Summary of ligands sensed by receptors isolated from a 2,726-member small molecule library screen. The hits obtained include FDA-approved molecules, natural products, agrochemicals, and a collection of research ligands (see Dataset S1 for the ligands screened). (B) The molecules sensed by PYR1-derived sensors are structurally diverse. The heatmap shows hierarchical clustering based on the pairwise chemical similarities of the 181 hit molecules. Pairwise ligand-similarity, measured using Tanimoto scores calculated from atom-pair descriptors in ChemmineR (21), is shown by the heatmap (blue = identical). Selected clusters of similar ligands are highlighted. (C). Selected chemical ligands from each cluster identified in part B. (D) PYR1-derived receptors sense drug-like molecules. Comparisons of the physicochemical properties of ligand hits (n = 181) and the nonhits (n = 2,545). Molecular weight (MW), computationally calculated octanol–water partition coefficient (cLogP), topological polar surface area, and a quantitative estimate of drug-likeness (24) values were extracted from ChEMBL (23). The median for each population is shown, along with the p-value from a two-sided Wilcoxon rank-sum test. R code is available at https://github.com/cutlersr/PYR1_sensor_screen.

Chemical clustering of the hit ligands based on their pairwise Tanimoto distances, calculated using atom-pair fingerprints, reveals a broad distribution of structurally dissimilar hits, with several small clusters of similar molecules (Fig. 2 B and C). Thus, there is broad structural diversity among the collection of sensed ligands. Several important pharmacological agents and agrochemicals are sensed by the collection, including steroids, nonsteroidal anti-inflammatory drugs (NSAIDs), dihydropyridine calcium channel antagonists, Transient Receptor Potential Vanilloid 1 (TRPV1) receptor agonists, and antifungal strobilurins (Fig. 2C). The physicochemical properties of the 181 ligands sensed are drug-like, as 173 (96%) comply with Lipinski’s rule of 5, compared to 76% for the library. Moreover, they score highly on the quantitative estimate of drug-likeness (QED) metric, a unitless measure of drug-likeness (med. QED hits = 0.69, QED nonhits = 0.55) (27). Compared to the nonhit members of the collection screened, the ligands sensed are enriched for higher cLogP (computed log₁₀octanol/water partition coefficient) and lower tPSA (topological polar surface area; Fig. 2D). These trends are likely driven partly by the cell-based nature of our sensor isolation method, which enriches for molecules with good membrane permeability and bioavailability. The size of ligands sensed is relatively smaller than that of the nonhit collection (median MW_hits = 276 Da; MW_non-hits = 324 Da) but includes molecules as small as 90 Da and as large as 748 Da (Fig. 2D). We note that the two largest ligands sensed are the poly-iodinated thyroxins tetrac (MW 748) and lyothyronine (MW 651). Blotz 2 (28) docking of these ligands suggests that they can adopt binding conformations in PYR1^TETRAC– and PYR1^LYOTHYRONINE–HAB1 complexes that enable HAB1’s Trp385 lock residue to insert into the 4’ pore of the closed-conformer receptors, a critical interaction for stabilizing the receptor complex (SI Appendix, Fig. S2). Collectively, these data demonstrate that the chemical space accessible with PYR1-derived sensors is structurally diverse and drug-like, which defines PYR1 as an unusually versatile and malleable scaffold for developing sensors for drugs and other bioactive ligands.

The screen identified an average of 4 distinct receptor-binding pocket sequences per hit ligand, providing a rich dataset with which to probe sequence–ligand relationships. Analysis of the binding pocket dataset reveals that only 26 receptors (~5%) were isolated in three or more screens, suggesting that a small number of promiscuous binders did not dominate the screening outcomes (Fig. 3A). Consistent with this observation, the mutations obtained are distributed broadly across the ligand-binding pocket (Fig. 3B). However, F159, which resides at the top of the pocket and contacts both the ligand and HAB1, was preferentially mutated and may be important for stabilizing PYR1–HAB1 interactions. The mutational patterns obtained also suggest strong selection for specific ligand–receptor interactions (Fig. 3C and Dataset S3). For example, we isolated six sensors for the anticonvulsant valproic acid, and all harbor the mutation I110R, suggesting our screens selected for a stabilizing salt bridge between valproic acid’s carboxylate and the mutant receptors’ basic arginine side chains. Similarly, 11 receptors were identified for the amine-containing ligand cinacalcet, used to treat hyperthyroidism, and 10 of these harbored K59D/E, suggesting selection for a stabilizing salt bridge. These PYR1 sensors suggest a strong selection bias toward distinct binding-pocket mutations (Fig. 3B).

Fig. 3.

Mutant PYR1 receptor–ligand interactions can be mapped using large-scale sensor screens. (A) Screening outcomes: ligand hits per receptor sequence, number of receptors identified per hit ligand, and the minimum dose–response of receptor hits during validation (each sensor strain was tested for ligand-dependent growth on 1, 10, 100 µM ligand). (B) Sensor mutations are broadly distributed throughout the PYR1 binding pocket; the top panel shows mutations per binding pocket residue, and the bottom panel maps the frequency of specific mutations per residue for the 553 receptors isolated. (C) Selected binding-pocket sequences for ligands that yielded ≥ 4 receptor hits. WT residues are shown at the top; colored boxes denote basic (blue), hydrophobic (green), polar (orange), and aromatic (gray) substitutions. (D) Sensor-binding-pocket mutations are tailored to their target ligands. Selection for mutational bias in each sensor–ligand set was quantified by calculating the Shannon diversity of mutation counts per position across receptors, with lower values indicating more positionally clustered mutational patterns (i.e., sensors for a ligand tend to have mutations in the same binding pocket locations). The analysis was performed for the 66 ligands with four or more sensors isolated (n = 502 receptor–ligand interactions). The null model shows Shannon diversity values calculated from randomly permuted receptor–ligand assignments (P < 6.4 × 10⁻¹³; two-sided Wilcoxon rank-sum test). R code is available at https://github.com/cutlersr/PYR1_sensor_screen.

To quantify positional mutational patterns, we computed Shannon diversity across pocket positions for each ligand, using a shuffled dataset as a null comparator. This null dataset preserved the overall mutation frequency at each pocket position but randomized receptor–ligand pairings, thereby disrupting ligand-specific positional structure. We limited our analysis to the 60 ligands for which four or more receptors were isolated (502 total ligand–receptor combinations). Shannon diversity values (Fig. 3D) were significantly lower in the sensors dataset than in the shuffled null set (Wilcoxon rank-sum, P < 6.4 × 10^-13), indicating that a ligand’s binding pocket mutations tend to cluster in specific residues, rather than being randomly distributed. Finally, we conducted yeast surface display assays to probe direct ligand–receptor interactions and observed nM to µM limits of detection for a subset of receptors that functioned well in the YSD assay (SI Appendix, Fig. S3). These experiments demonstrate direct physical interactions without a cell-based growth assay. Collectively, these analyses show strong global selection for ligand-specific mutational patterns, with mutations broadly distributed across the binding pocket.

Our large dataset of ligand–receptor interactions provides maps of binding-pocket residues that, in principle, could be harnessed to improve receptors. We took two complementary approaches to test this. First, we mined our data for the coumarin scaffold—the largest cluster of plant-derived natural products in our screening library (selected examples shown in Fig. 4A). Of the 14 coumarins in the library, we isolated sensors for eight, which harbored mutations at 11 of the 19 binding-pocket sites mutagenized. We derived a coumarin-binder sequence profile from these data and encoded combinations of up to nine of the most frequent coumarin-associated mutations (Fig. 4B and Dataset S4). The library contained 77,327 members and was constructed by the combinatorial assembly of 218 oligonucleotides (Dataset S4). It was then transformed into S. cerevisiae and screened against 12 coumarins (Dataset S5). This yielded sensors for three ligands not identified in the original screens (5, 6, and 5,7-dihydroxy-4-methylcoumarin) and between 10- and 400-fold sensitivity gains for the eight positive coumarin hits, including (1–3) (Fig. 4C and SI Appendix, Fig. S4 and Dataset S7). As a second approach, we asked whether profiles derived from receptors that recognize single-ring phenyl derivatives (6–9) could guide the discovery of sensors for the chemically similar molecules, TNT (11) and three of its degradation products (12, 13, and 4-amino-2,6-dinitrotoluene) (Fig. 4D and SI Appendix, Fig. S5); none of these were present in the screening collection and DSM/TSM library screens did not yield hits for TNT. A 506,229-member receptor library built using a sequence profile derived from receptors activated by (6–9) yielded initial sensors for the four targets (Dataset S4). Sequence profiles derived from the first-round sensors were used to design a second-round library, which ultimately yielded sensors for TNT and three degradation products, DNT, 2ADNT, and 4ADNT, with low µM EC₅₀ ligand responsiveness, as measured using reporter and growth assays (Fig. 4F and SI Appendix, Fig. S5 and Dataset S8). Together, these data demonstrate that our large dataset of ligand–PYR1 interactions can be harnessed to yield receptors with improved affinity and target recognition.

Fig. 4.

Ligand–receptor interaction maps and sequence profile-guided design enable the isolation of high-affinity sensors. (A) Selected chemical structures of coumarin hits used to generate the sequence profile of the coumarin-target library: osthole (1), imperatorin (2), and angelicin (3). Receptors for 4-methylumbelliferone (4) and 7-methoxycoumarin (5) were not identified in the initial screen; sensors for these compounds were isolated from the coumarin-target library. (B) Sequence profile of the coumarin targeted library and binding pocket mutations of PYR1 receptors activated by compounds (1–5). (C) Yeast-hybrid β-galactosidase colony assays comparing PYR1 sensor response of the initial sensor and high-affinity sensors isolated from the coumarin-targeted library. The targeted library yielded PYR1 receptors for (4 and 5), which were not identified in the initial data generation screen. The minimum dose–responses are shown in comparison to the mock condition. SI Appendix, Fig. S2 shows additional growth and reporter dose–response assays for the coumarin sensors. (D) Chemical structures of selected ligands used to design the sequence profile of a TNT target library: 2,4-dinitrophenol (6), chloroxylenol (7), 2,4,6-trihydroxybenzoic acid (8), 2,6-dihydroxybenzoic acid (9). Hits for trinitrotoluene (TNT; 10), dinitrotoluene (DNT; 11), and 2-amino-dinitrotoluene (2ADNT; 12) were not identified using the screening libraries. Sensors for these were isolated from the TNT target library. (E) Sequence profile of the TNT target library (TNTv2) and binding pocket mutations for TNT, DNT, and 2ADNT receptors isolated from the library. The sequence profile used for library construction additionally includes 117L/W and 160A/V (Dataset S4). (F) Yeast-hybrid β-galactosidase colony and cell-based titration assays for PYR1^TNT, PYR1^DNT, and PYR1^2ADNT for their on-target ligands. Triplicate data points are shown. The inset data indicate the EC₅₀ and range.

Our data establish a simple approach for creating new sensors for small molecules: 1) screen receptor libraries with relatively few binding pocket mutations to identify initial sensors, 2) use the data generated to derive sequence profiles, and then 3) design and screen affinity maturation libraries to yield high-affinity sensors. To test this pipeline on a class of molecules dissimilar to members of our initial screening library, we focused on “forever” PFAS molecules, which are critical biosensing targets due to their persistence, bioaccumulation, and associated health and environmental risks (29). To do this, we constructed and screened an improved double-site mutant library designed to enhance coverage in HTS assays by reducing autoactivating and wild-type receptors from the original DSM library (DSM-Hao; see Materials and Methods). We screened this against a panel of 103 different PFAS (Dataset S6), including six of major concern that targeted for regulation in drinking water by the EPA: perfluorononanoic acid (PFNA; 13), perfluorooctanesulfonic acid (PFOS; 14), perfluorooctanoic acid (PFOA; 15), perfluorohexanesulfonic acid (PFHxS; 16), hexafluoropropylene oxide dimer acid (HFPO-DA), and perfluorobutanesulfonic acid (PFBS) (Fig. 5A). The initial library screens yielded 86 sensors targeting 18 targets, providing a dataset for creating a PFAS-targeted library to isolate high-affinity sensors (Fig. 5B; and Dataset S9). Screens using this targeted library expanded the scope of PFAS sensing, increasing the total number of PFAS molecules sensed to 25, and yielded sensors with improved on-target sensitivity. PYR1^PFAS, isolated in a screen against PFOA (13), shows a strong response to low µM concentrations of four of the six molecules targeted for regulation (Fig. 5D and SI Appendix, Fig. S6) and provides a new sensor for the most problematic environmental “forever” molecules. These experiments further validate the data-driven sensor approach and its ability to develop new sensors.

Fig. 5.

Data-driven engineering of PYR1 receptors for PFAS detection. (A) Chemical structures of selected PFAS molecules targeted for sensor design. PFNA (13), PFOS (14), PFOA (15), and PFHxS (16) are four of six PFAS molecules regulated by the US Environmental Protection Agency in drinking water. (B) PFAS-library sequence profile and binding pocket mutations in PYR1^PFAS, a pan-PFAS receptor strongly activated by any of (13–16). The library was screened against 103 unique PFAS molecules (Dataset S5). (C) Yeast-hybrid β-galactosidase colony assays show that PFNA, PFOS, PFOA, and PFHxS strongly activate PYR1^PFAS. (D) Dose-dependent response of PYR1^PFAS to four on-target PFAS molecules in cell-based β-galactosidase assays. Triplicate data points are shown. EC₅₀ values and range are shown as insets.

Discussion

The ligand-binding pocket of PYR1 is unusually malleable; it can be reshaped to recognize a surprising number of structurally unrelated ligands. We isolated and validated sensors for 6.6% of the nearly 3,000 molecules screened and isolated and validated over a dozen more using data-driven, targeted libraries. The majority of the ligands sensed are recognized by independently isolated sequence variants. Sequence analyses of the receptors demonstrate binding pocket mutational patterns tailored to each ligand, mutations distributed broadly across the binding pocket, and a small frequency of promiscuous receptors. Collectively, these trends indicate a robust pipeline for sensor isolation.

It has been estimated that drug-like space encompasses approximately 10³³ molecules (30). Even without improvements in hit rates, which should be possible, our system provides access to a large chemical space that can be harnessed to build new biochemical, cellular, and whole-organism control systems. Our sensors recognize drugs and drug-like molecules, opening many possibilities. For example, NSAID-controlled switches, of which we isolated several, could be valuable in clinical applications. Our sensors can also be orthogonalized for deployment in plants using the PYR1*–HAB1* module, which enables two-channel control systems and control of plant phenotypes without perturbing native ABA signaling (19). Thus, PYR1’s extreme malleability, ease of sensor design, and induced dimerization mechanism converge to make PYR1 a privileged scaffold for sensor design across biological kingdoms.

Why is PYR1 imbued with such unusual malleability? PYR1 belongs to the ligand-binding START superfamily. These are characterized by a helix-grip motif, which consists of a seven‐stranded antiparallel β‐sheet that wraps around a C-terminal α‐helix to create a central ligand-binding pocket. START proteins are present in all domains of life and, like our PYR1 variants, bind structurally diverse molecules – from hydrophobic steroids to polar flavonoid-glucosides (31–33). This domain is frequently involved in transport and catalysis. However, PYR1 and its relatives link ligand recognition to effector binding to control signaling. PYR1, therefore, couples an evolutionarily malleable binding pocket with a biochemical mechanism well-suited for engineering signaling systems. Given the success of computational design for CID sensors (9, 34) and the new methods developed for PYR1 (35), we anticipate that PYR1’s properties can be broadly harnessed to design sense–response systems controlled by user-specified molecules with antibody-like simplicity.

Plant hormone signaling employs a suite of sensor modules that, like PYR1, involve hormone-mediated stabilization of protein–protein interactions (14, 16, 36–38). For example, the gibberellic acid receptor is a member of the large GDSL lipase family and, like the ABA system, its module employs induced dimerization with an effector that greatly stabilizes the activated receptor complex (39–41). GID1 and other plant hormone sensors should provide a suite of scaffolds for building new sensor systems with the antibody-like simplicity and scope of the PYR1 system.

Supplementary Material

Appendix 01 (PDF)

Dataset S01 (XLSX)

Dataset S02 (PDF)

Dataset S03 (XLSX)

Dataset S04 (XLSX)

Dataset S05 (XLSX)

Dataset S06 (XLSX)

Dataset S07 (XLSX)

Dataset S08 (XLSX)

Dataset S09 (XLSX)

Dataset S10 (XLSX)

Dataset S11 (XLSX)

Dataset S12 (XLSX)

Data, Materials, and Software Availability

All study data are included in the article and/or supporting information.