Rhobot-Screen: an integrated robotic platform for functional screening of rhodopsin variants

Abstract

Background

Rhodopsins are photoreceptive membrane proteins widely used as optogenetic tools in basic research and medical applications, and extensive mutational studies have been performed to improve or modify their functional properties. Recently, in the broader field of protein engineering, data-driven strategies based on machine learning have attracted increasing attention, as they enable efficient exploration of vast mutational spaces with a reduced number of experiments. Such approaches require large, consistent datasets that link predefined mutations to quantitative functional properties, which necessitates systematic construction and characterization of targeted variants rather than random mutagenesis. For rhodopsins, however, generating these datasets remains challenging due to operator-dependent, non-integrated workflows that are difficult to scale and standardize.

Results

To address this limitation, we developed an automated screening platform termed Rhobot-Screen, based on a robotic liquid-handling workstation, which integrates multiple experimental steps into a standardized workflow with reduced dependence on operator-specific expertise. This platform performs site-directed mutagenesis, plasmid preparation, protein expression in bacterial and mammalian cultured cells, and functional characterization in a 96-well format through automated liquid-handling operations. For spectroscopic characterization, we established a 96-well plate-based hydroxylamine bleaching assay that determines absorption maximum wavelengths without protein purification. As a demonstration of the platform, we comprehensively mutated three established color-tuning residues in Gloeobacter rhodopsin, generating 57 single-point variants. Using Rhobot-Screen, the absorption maxima of 46 variants were successfully determined. The resulting dataset revealed position-dependent relationships between spectral shifts and amino acid physicochemical properties, with clear correlations between absorption wavelength and side-chain volume at positions 129 and 256, but not at position 226. The platform was further extended to mammalian cell-based assays for functional characterization of animal rhodopsins.

Conclusions

Rhobot-Screen provides an integrated workflow in which all liquid-handling steps for systematic construction and spectroscopic characterization of rhodopsin variants are automated in a 96-well plate format under standardized conditions. By automating and standardizing multiple operator-dependent steps, the platform provides a reproducible framework for acquiring quantitative sequence–function data from predefined rhodopsin variants. This framework should support both mechanistic studies of rhodopsins and future data-driven engineering of rhodopsin functions.

Graphical Abstract

Supplementary Information

The online version contains supplementary material available at 10.1186/s12915-026-02691-8.

Background

Protein engineering aims to design and optimize proteins with improved or novel functions for applications in basic research, industrial biotechnology, and medicine. Recent advances in machine learning (ML) have begun to transform this field by enabling data-driven exploration of protein sequence–function relationships [1–3]. In parallel, major progress has been achieved in protein structure prediction from amino acid sequences through ML, as exemplified by AlphaFold and related models, largely driven by the availability of extensive and high-quality structural datasets such as those accumulated in the Protein Data Bank [4–8]. In contrast, the prediction of protein function directly from sequence remains substantially more challenging, particularly for quantitative functional properties. Deep learning approaches, including protein language models, have shown promise in functional annotation tasks and in predicting mutation effects through transfer learning and zero-shot inference [9–13]. However, these models remain limited in modeling fine-grained, quantitative properties within narrowly defined families, such as enzyme activity, stability, or spectroscopic characteristics [2, 11, 14]. Progress in this area has been further constrained by the limited availability of well-curated experimental datasets with reliable quantitative labels [14–16]. Consequently, protein engineering efforts targeting such properties cannot rely on direct prediction alone and instead require strategies that iteratively explore mutational space through experimental evaluation.

In these approaches, ML-based predictions are used to guide variant selection, thereby substantially reducing the number of experiments required compared with exhaustive or unguided screening [1, 3, 16]. The effectiveness of such strategies stems from iterative cycles in which candidate variants are proposed by the model, experimentally evaluated, and incorporated into subsequent rounds of model refinement, progressively improving exploration efficiency [3, 16]. Implementing this iterative strategy in practice requires experimental platforms that can systematically generate predefined variants through site-directed mutagenesis and quantitatively characterize their functional properties in a high-throughput and standardized manner, in contrast to random mutagenesis-based library screening [2]. However, conventional workflows for site-directed mutagenesis and functional assays involve multiple manual steps and are difficult to scale. Moreover, their performance often depends on operator-specific expertise, leading to variability in throughput and data quality. In general, automation of experimental workflows can mitigate such variability by standardizing procedures, thereby improving precision and reproducibility [17–19]. Therefore, scalable and automated experimental platforms are essential for enabling data-driven modeling of quantitative protein function and for efficient ML-guided exploration of vast mutational spaces in protein engineering.

Rhodopsins have been used, including in our own work, as a model for ML-based prediction of absorption maxima wavelengths (λ_max) [20–23]. Rhodopsins are photoreceptive seven-transmembrane proteins that contain a retinal chromophore and mediate a wide range of light-driven ion transport and signaling functions [24, 25]. Owing to their relatively simple architecture, well-defined photochemical behavior, functional diversity, and robust expression in heterologous systems, they are particularly well suited for systematic mutational studies of structure–function relationships. Furthermore, rhodopsins are widely employed as optogenetic tools for manipulating neural activity and monitoring membrane potential, with applications extending beyond neuroscience and basic biology into emerging areas of biomedical research and therapeutic development [26–29]. For such applications, engineering wavelength properties tailored to specific functional needs is critical, and AI-based wavelength prediction can greatly facilitate the design of optimized tools. Our predictive model, trained on a curated dataset of ~ 800 wild-type and mutant microbial rhodopsins, achieved an average prediction error of ± 7.8 nm and enabled the discovery of long-wavelength-shifted variants with potential utility in optogenetics [21, 30]. Beyond its standalone predictive performance, this model serves as a foundation for iterative AI-guided protein engineering approaches, in which model predictions and experimental measurements are combined in a cyclic manner to guide variant selection. Realizing such AI-guided approaches in practice requires experimental workflows that overcome the limitations of manual experimentation by enabling high-throughput and reproducible construction and spectroscopic characterization of predefined rhodopsin variants.

To meet these experimental requirements, we developed an integrated robotic platform termed “Rhobot-Screen,” which enables systematic construction and functional characterization of predefined rhodopsin variants in a 96-well format (Fig. 1). As a test case, we used Rhobot-Screen to perform site-directed mutagenesis and characterization of a light-driven outward H⁺-pumping microbial rhodopsin, Gloeobacter rhodopsin (GR): each of three color-tuning residues at positions 129, 226, and 256 in GR was substituted with each of 19 amino acids. The resulting variants exhibited λ_max distinct from that of the wild-type (WT) protein. Furthermore, the λ_max of the GR variants showed clear correlations with amino acid volume at positions 129 and 256. We also extended the workflow to a mammalian cell-based fluorescence assay for a G-protein-coupled receptor (GPCR)-type animal rhodopsin, illustrating its modularity beyond microbial rhodopsin spectroscopy. These results demonstrate the utility of Rhobot-Screen as a reproducible and scalable framework for generating standardized datasets that may support systematic studies of rhodopsin functions and future data-driven engineering of rhodopsins.

Fig. 1.

Integrated experimental workflow for systematic mutagenesis and functional screening of rhodopsins. The workflow for a DNA-to-function automation pipeline enables parallel processing of rhodopsin mutants in a 96-well format. The integrated system combines site-directed mutagenesis, cloning, plasmid preparation, protein expression, and functional characterization using the Tecan Fluent® liquid-handling platform

Results

Strategy for establishing an integrated 96-well workflow

To establish an integrated workflow that enables systematic construction and evaluation of predefined variants, we designed all experimental steps to be carried out using a 96-well plate format, allowing explicit specification of individual mutations and consistent acquisition of quantitative functional data across variants. All liquid-handling steps were performed using a Tecan Fluent® robotic liquid-handling workstation (Fig. 2). A particular focus was placed on optimizing the plasmid cloning step following site-directed mutagenesis. Conventional protocols often involve transformation onto agar plates and colony isolation by robotic picking, which introduces complexity in hardware requirements and scalability. To circumvent this, we adopted a limiting dilution approach that is more compatible with automation [31]. In our system, plasmid transformation is followed by serial dilution in liquid medium, and clonal outgrowth is detected by measuring OD₆₀₀ using an integrated microplate reader, allowing automated identification of wells containing clonal Escherichia coli cultures. For functional assessment of rhodopsin variants, we established a 96-well plate-based hydroxylamine bleaching assay, which enables rapid evaluation of absorption spectra without requiring protein purification [21]. Together, these components constitute an integrated platform in which all liquid-handling steps—from mutagenesis to spectral measurement—are performed automatically within the same plate format. We refer to this fully automated experimental workflow as “Rhobot-Screen” (Fig. 1).

Fig. 2.

The configuration and workspace layout of the automated liquid-handling system. A Photograph of the integrated robotic workstation used for systematic mutagenesis and screening. The robotic gripper arm is equipped with exchangeable gripper fingers to transport microplates, including access to the plate reader located below the workdeck. The pipette arm operates with an 8-channel pipette head capable of aspirating, dispensing, and mixing liquids with high precision using liquid-level detection. The barcode reader links barcoded microplates to downstream data, enabling reliable traceability. The HEPA-filtered workspace is maintained clean by UV sterilization. B Schematic layout of the working deck showing the spatial arrangement of labware carriers and modules. The corresponding components are defined in the legend below

As a test case, we performed comprehensive mutagenesis of three color-tuning residues in GR—a light-driven outward H⁺-pump from the cyanobacterium Gloeobacter violaceus. GR has been well characterized through spectroscopic, structural, and mechanistic studies [32–37], and serves as a tunable spectral scaffold, with mutants exhibiting diverse absorption and photochemical properties upon mutation [38, 39]. In addition to its role in fundamental biophysical studies, GR has also been explored in medical and biotechnological contexts [40–45]. Several amino acid positions in GR are of particular interest because they correspond to well-known “color-switch” sites in microbial rhodopsins, such as the L/Q switch at Leu129, the G/P switch at Pro226, and the A/TS switch at Ala256 (Fig. 3) [22]. For this platform demonstration, we selected these three conserved color-switch residues in GR as representative targets. These residues have been shown to cause red- or blue-shifted spectral changes upon specific substitutions in GR and other rhodopsins [22, 46–49]. In addition, they are located at distinct regions surrounding the retinal chromophore (Fig. 3), providing a structural rationale for examining and comparing their roles in spectral tuning. Using the automated Rhobot-Screen platform described above, we constructed and characterized all 57 single-point GR mutants, in which each of the three color-switch residues was substituted with each of the other 19 amino acids (GR L129X, P226X, and A256X).

Fig. 3.

Spatial arrangement of the three residues for site-directed mutagenesis. The 3D structure of GR (PDB ID: 6NWD [35]) in the vicinity of the retinal chromophore is shown from different orientations. PSB, protonated Schiff base

Site-directed mutagenesis

Mutations were carried out using the QuikChange site-directed mutagenesis. DNA oligonucleotide primers were annealed to the target gene to synthesize the desired mutant DNA. The template DNA was subsequently digested using the restriction enzyme DpnI, and the successful synthesis of mutant DNA was confirmed by agarose gel electrophoresis. As a representative example, Fig. S1 (Additional file 1) shows the electrophoretic analysis of 16 individual mutants at position 256 after the mutagenesis reaction. Clear DNA bands were detected in all lanes, confirming successful synthesis of mutant DNA in each case. However, variations in band intensity were observed, suggesting that the introduced mutations influenced amplification efficiency. The remaining mutant DNA was introduced into E. coli cells by transformation for subsequent DNA sequencing.

E. coli transformation and clone isolation by limiting dilution

When constructing mutants manually, transformants are typically spread onto antibiotic-containing agar plates, and colonies are selected after colony formation is confirmed. However, this conventional approach becomes increasingly time- and space-consuming as the number of samples increases, making it unsuitable to execute in automated systems. To address this, we adopted the limiting dilution method as a more automation-friendly alternative. In this method, transformation mixtures were serially diluted to achieve a cell density of less than one cell per well, and aliquots were dispensed into 96-well plates. After incubation for a defined period, only wells containing single transformants exhibited bacterial growth that were confirmed by recording the OD₆₀₀ values using an integrated microplate reader. Cultures from these wells were collected and used for subsequent plasmid extraction experiments.

Figure 4 shows a representative example of transformant selection using the limiting dilution method. Following the approach of Yehezkel et al. [31], we empirically determined the appropriate dilution rate to achieve bacterial growth in approximately 25% of replicate wells. Under this condition, Poisson statistics predict that multiple cells may occasionally be present in a single well; the probability of such events is estimated to be approximately 3.4%, corresponding to roughly 14% of positive wells potentially containing more than one clone. Increasing the dilution factor would further reduce the probability of mixed populations and improve the likelihood of obtaining strictly clonal cultures. However, because the number of wells that can be processed simultaneously in the Fluent® system is limited by the available deck space, increasing the dilution factor would also reduce the effective throughput of the screening workflow. The chosen condition of approximately 25% positive wells therefore represents a practical balance between clonality and throughput in our automated setup. Although the presence of mixed populations cannot be completely excluded at this stage, the plasmid sequence of each selected clone was subsequently verified by Sanger sequencing within the coding region, as described later, showing single nucleotide peaks at the mutation sites without detectable heterogeneity and thereby minimizing the risk that heterogeneous plasmid populations would affect the expressed protein. Sixteen replicates were prepared for 12 mutants, and those with OD₆₀₀ values between 0.08 and 1.0 were selected. Cultures derived from a 1:525 dilution (3.5-fold serial dilution performed over five steps) yielded 1–5 candidate clones per 16 replicas from each mutant. Heterogeneity was observed in the number of clones obtained for different mutants. One possible factor contributing to this variation is the efficiency of plasmid amplification during the former site-directed mutagenesis step.

Fig. 4.

Transformant screening by limiting dilution method in a 96-well plate format. A Schematic diagram of limiting dilution and replica preparation. The optical density (OD₆₀₀) of E. coli cultures, diluted 525-fold and incubated for 17 h after dispensing into a 96-well microplate, was measured. Aliquots of each suspension were dispensed into two adjacent columns (16 wells). B Representative results of the limiting dilution method. The upper panels present OD₆₀₀ values as 3D bar plots, while the lower panels highlight wells with OD₆₀₀ values between 0.08 and 1.0 in red. Wells marked in red were selected as positive clones for subsequent experiments

DNA purification and sequencing

Selected positive clones were inoculated in LB medium with ampicillin and cultured overnight. Plasmid extraction from the overnight cultures was automated using a magnetic bead-based method. This procedure yielded 13–138 ng/μL of plasmid DNA from 400 μL of culture (Fig. 5). The purified plasmids exhibited A260/280 ratios ranging from 1.63 to 1.86 (Fig. 5), which was slightly lower than the manufacturer’s specified value (typical A260/280 index ≥ 1.8). However, Sanger sequencing confirmed that 34 out of 37 clones produced readable sequencing results, indicating that the plasmid quality was sufficient for sequence verification. Furthermore, the intended mutations were successfully introduced in 30 out of 34 verified clones (88%). For all plasmids subjected to sequencing, only single nucleotide peaks were observed at the mutation sites, suggesting that the purified plasmids originated from single dominant clones at least within the coding region. These results indicate that the selected cultures were dominated by single plasmid variants within the sequenced region.

Fig. 5.

Representative results of plasmid extraction using an automated liquid-handling system. Bars show the concentration of purified plasmid DNA (ng/μL). Dark-blue bars represent clones with correct sequencing results, pale-blue bars indicate clones where the target mutation was not introduced, and pale-orange bars represent clones with low-quality sequencing results. Dark-orange dots indicate the A260/280 ratios for each sample. Source data are provided in Additional file 2

Protein expression and spectroscopic characterization

After sequence verification, the obtained plasmids encoding the 57 single-point GR mutants were transformed into the E. coli strain C43(DE3) for protein expression using an automated system. The cells were pre-cultured overnight in 2 × YT medium and diluted 1:100 into 1 mL of fresh 2 × YT medium, followed by incubation for 4 h before induction. Protein expression was induced by adding IPTG, and the cells were incubated for an additional 4 h in the presence of all-trans-retinal. These induction conditions with fixed incubation times were pre-optimized for GR expression to ensure reproducible pigment production within the automated workflow. Since the timing of induction was not determined based on optical density measurements, induction parameters, including incubation time and IPTG concentration, may require pre-optimization for different proteins when applying this automated system.

The E. coli cells expressing the target proteins exhibited coloration ranging from orange to pink and purple (Fig. 6A). These color differences reflect variations in the absorption spectra of the individual mutants. The absorption maxima of the mutant proteins were determined using the hydroxylamine (HA) bleaching method. In this method, HA is added to a solution containing rhodopsin proteins, and the sample is illuminated by > 500-nm light. HA hydrolyzes the C = N double bond of the retinal Schiff base, producing retinal oxime. This reaction causes a decrease in rhodopsin-derived absorbance accompanied by an increase in retinal oxime absorbance. By calculating the difference spectrum by subtracting the spectrum recorded before illumination from that recorded after illumination, the absorption changes attributable to rhodopsin can be extracted, allowing determination of the absorption maximum [21]. Because this method selectively detects absorbance changes specific to proteins reactive with HA, it is suitable for measurements using crude extracts prepared by solubilizing E. coli cells expressing the target protein. This makes the method well-suited for high-throughput studies. To further streamline the determination of λ_max, we established an automated HA-bleaching assay by integrating a liquid-handling system with a microplate reader. Lysozyme treatment and solubilization of the cells with DDM, which served as sample pretreatment for the measurements, were performed using an automated liquid-handling system. After centrifugation to remove the insoluble fraction as a pellet, the supernatant was dispensed into each well of a microplate using the automated system again. This procedure effectively reduced light scattering that would otherwise interfere with absorption spectroscopy.

Fig. 6.

λ_max determination by HA bleaching of rhodopsin expressed in E. coli cells. A E. coli cell pellets expressing various GR mutants. Circles represent representative colors of the pellets. B Absorption spectra of the solubilized membrane fraction from E. coli cells expressing GR WT. C Light-induced difference spectra obtained by subtracting the spectrum recorded before the addition of HA (gray line in A) from each spectrum in A. The sample was kept in the dark and subsequently illuminated for the indicated durations. Source data are provided in Additional file 2

When the colored protein is successfully expressed, the absorption spectrum of the solubilized E. coli membrane fraction showed a peak derived from the retinal chromophore in the 500–600 nm region (Fig. 6B). Upon addition of HA to the dispensed samples in the dark, the absorption of the retinal decreased, while a new peak representing the formation of retinal oxime at ~ 360 nm increased during the incubation in the dark, followed by illumination to accelerate the reaction. Difference spectra were calculated by subtracting each absorption spectrum from the spectrum obtained before the addition of HA so that a positive and a negative peak exhibit the absorption of the retinal chromophore in the original protein and the produced retinal oxime, respectively (Fig. 6C). From the position of the positive peak, we determined the λ_max of GR WT from three biological replicates to be 538.7 ± 0.4 nm (SEM, n = 3). The reproducibility of our measurements and the agreement of our measured value with previously reported λ_max values of purified GR WT at neutral pH (538–544 nm) [39, 50, 51] suggest that our method can provide reliable λ_max data under these conditions. In addition, we confirmed that we obtained comparable results of protein expression and HA bleaching signals using another E. coli strain, BL21(DE3), under the same experimental conditions (Additional file 1: Fig. S2). Although the expression conditions were optimized using the C43(DE3) strain, these results suggest that the experimental system is not strictly limited to C43(DE3).

Systematic mutagenesis of color-switch sites in GR

We used the automated platform described above to construct and characterize all 57 possible single-point GR mutants in which each of the three known color-switch positions (Fig. 3) was individually substituted with all other amino acids (L129X, P226X, and A256X). Absorption spectra of GR WT and the 57 single-point mutants were examined by a hydroxylamine bleach assay (Fig. 7A). Among them, 11 mutants were not expressed enough to determine their λ_max values.

Fig. 7.

Spectral shifts of GR mutants. A The relative changes in the λ_max relative to GR WT. The red- and blue-shifted changes are indicated by red and blue bars, respectively. Each bar represents the mean value for the absorption change in the three biological replicates, and error bars indicate the standard error of the mean. B Correlation between the absorption light energy of each mutant and the side-chain volume of the substituted amino acid at positions 129 (left), 226 (middle), or 256 (right) in GR. The vertical axis shows the difference in the absorption light energy relative to the wild type (= 1/λ_max; cm⁻¹), with positive and negative values indicating spectral blue- and red-shifts, respectively. Each plotted point represents the mean absorption light energy change from three biological replicates, and error bars denote the standard error of the mean. The colored dashed lines indicate the best-fit lines from linear least-squares regression. Source data are provided in Additional file 2

Figures 7B and S3 (Additional file 1) present plots of absorption-energy shifts versus physicochemical properties of amino acids. While the absorption-energy shifts did not consistently depend on the hydropathy index of amino acids (Additional file 1: Fig. S3), clear correlations were observed with amino-acid volume for L129X (R² = 0.4194) and A256X (R² = 0.7747) mutants in GR (Fig. 7B). By contrast, P226X mutants showed little correlation between the absorption-energy shifts vs amino acid volume. Most P226X mutants exhibited red-shifted absorptions compared to the WT. This trend is attributable to the differences in the orientation of the main-chain dipole moment between proline and other amino acids with a secondary and primary amine in their main chain, respectively, as proposed for mutants of Krokinobacter rhodopsin 2 [22].

Extension of the Rhobot-Screen platform to animal rhodopsins

Because the Rhobot-Screen platform consists of modular experimental workflows, it can be extended to different classes of rhodopsins. To demonstrate this flexibility, we developed an additional screening module for animal rhodopsins, which are light-sensitive GPCRs widely used in optogenetic applications. Unlike microbial rhodopsins, animal rhodopsins cannot be functionally expressed in E. coli; therefore, we incorporated a mammalian cell-based expression pipeline into the Rhobot-Screen workflow (Fig. 1). As a proof-of-concept, we applied this module to an animal rhodopsin, SpiRh1. SpiRh1 is the only bistable-type animal rhodopsin for which three-dimensional structures of both inactive and active states have been determined [52–54]. In addition, it is sufficiently stable to tolerate mutations [55] and can activate a broad range of G-protein subtypes [56], highlighting its potential as a template for engineering GPCR-type optogenetic tools [57, 58].

SpiRh1 was transiently expressed in COS-1 cells together with the calcium indicator GCaMP6s using automated liquid handling with the Fluent® system. SpiRh1 activates G_q/11-type G proteins to evoke intracellular Ca²⁺ elevation, which was detected by fluorescence monitoring of GCaMP6s using the microplate reader (Fig. 8A). Fluorescence signals corresponding to Ca²⁺ elevation were consistently observed across eight replicates and exhibited a dose-dependent response to the amount of transfected SpiRh1 DNA (Fig. 8B and Additional file 1: Fig. S4). These results demonstrate that our system is compatible with mammalian cell expression systems and GPCR-type rhodopsins, highlighting its potential applicability to other proteins including eukaryotic microbial rhodopsins and non-rhodopsin GPCRs.

Fig. 8.

GPCR signaling assay. A Schematic diagram of the putative intracellular signaling pathway underlying SpiRh1-mediated Ca²⁺ elevation and its detection using the fluorescent Ca²⁺ indicator GCaMP6s in mammalian cultured cells. PLC, phospholipase C; PIP₂, phosphatidylinositol 4,5-bisphosphate; IP₃, inositol 1,4,5-trisphosphate. B Fluorescence-based Ca²⁺ monitoring in cells transiently transfected with different amounts of the SpiRh1 expression DNA plasmid. Error bars represent the standard error of the mean (SEM; n = 8). Source data are provided in Additional file 2

Discussion

In this study, we established Rhobot-Screen, an integrated robotic platform in which all liquid-handling steps required for systematic screening of rhodopsin variants are performed within a consistent 96-well plate format on a single liquid-handling system. To validate the platform’s utility in a practical setting, we performed comprehensive mutagenesis at three color-tuning sites in GR, generating a total of 57 distinct mutant DNA constructs. Among these, the λ_max of 46 variants were successfully determined using our automated workflow. In addition, we demonstrated the flexibility of Rhobot-Screen with different expression systems (two different E. coli strains and mammalian cultured cells), different proteins (ion-transporting and GPCR-type rhodopsins), and different measurement methods (UV–Vis spectroscopy and fluorescence-based intracellular Ca²⁺ monitoring). These results demonstrate that all experimental steps, from site-directed mutagenesis to functional characterization, can be conducted entirely within the 96-well plate format. To our knowledge, this is the first report in rhodopsin research of an end-to-end workflow that integrates variant construction and functional characterization on an automated liquid-handling platform. Taken together, these findings establish Rhobot-Screen as an effective and scalable platform for the functional screening of rhodopsin variants.

The key innovation of Rhobot-Screen lies not in the individual experimental steps, many of which are already widely used in 96-well plate-based workflows [39, 59–62], but in its ability to decouple experimental performance from operator-specific expertise. Once the liquid-handling and measurement procedures are established, the platform enables consistent and high-precision execution of complex, multi-step experiments spanning mutagenesis to functional characterization, without requiring specialized training in each individual experiment. This reduces variability associated with operator-dependent techniques and allows screening to be performed in a highly standardized manner, helping to maintain stable throughput and consistent data quality across different users, instruments, and laboratories, as well as over extended periods, thereby enabling parallelization and long-term continuity of screening. Such characteristics, together with the ability to generate predefined variants through targeted rather than random mutagenesis, are particularly advantageous for iterative, data-driven protein engineering workflows, in which candidate variants are repeatedly designed, experimentally evaluated, and incorporated into subsequent rounds of model refinement [3, 16]. In addition, the modular architecture further supports the incorporation of different expression systems and assay types, as demonstrated here by E. coli-based spectroscopic assays for a microbial rhodopsin and mammalian cell-based fluorescence assays for a GPCR-type animal rhodopsin. Together, these features position Rhobot-Screen as a framework for scalable and reproducible data generation, complementing existing experimental and computational approaches in protein engineering.

Despite the advantages of Rhobot-Screen, certain limitations remain to be addressed. First, prioritizing assay throughput by omitting protein purification results in a lower signal-to-noise ratio compared to conventional spectrophotometric measurements using purified samples. This reduced sensitivity can make it challenging to accurately determine the λ_max for variants with extremely low expression levels. In addition, application of this assay to rhodopsins with λ_max values in the blue or shorter-wavelength region may require careful validation, because measurements in the shorter-wavelength region may be more susceptible to scattering and other background contributions. Second, although the platform can incorporate different protein-expression modules, including the E. coli and mammalian cell systems demonstrated here, its applicability has so far been validated using only a limited number of rhodopsins. While many upstream steps in the pipeline, such as mutagenesis, plasmid preparation, and sequencing, are broadly applicable to other proteins, downstream modules, particularly protein expression systems and functional measurements, may require modification depending on the properties of the target proteins. Third, the current measurement modules are limited to λ_max characterization and intracellular Ca²⁺ monitoring. Several studies have utilized microplate reader-based experimental systems for photoreceptor characterization [63, 64], as well as microplate-based automated electrophysiology for rhodopsin engineering [65]. Because these assays are compatible with multiwell-plate formats, similar approaches could be incorporated into Rhobot-Screen as additional measurement modules. Expanding the system to include multi-parametric assays will be a key direction for future development to provide a more comprehensive functional landscape for rhodopsin engineering.

In parallel with the development of automated experimental platforms, multiple efforts have been made to improve different layers of the optogenetic protein engineering workflow. At the level of data acquisition, robotic and automated screening systems have been developed to increase experimental throughput, including fluorescence-based screening platforms [66] and the Rhobot-Screen system described here. At the level of experimental design and construct generation, vector libraries streamline the production of modular fusion constructs [67]. In addition, efforts to improve data accessibility and reuse have led to the development of curated databases and resource platforms for optogenetic tools, such as OptoBase [68] and OPTICS [69]. These complementary advances across data acquisition, experimental design, and data integration highlight the importance of connecting different layers of the workflow. Integrating automated screening platforms such as Rhobot-Screen with standardized resources and curated databases will be critical for enabling efficient, data-driven cycles of protein engineering and accelerating AI-assisted optimization of photoreceptor functions.

Our systematic mutational analysis revealed that spectral tuning at different color-tuning sites is governed by distinct, position-dependent mechanisms. Using the Rhobot-Screen platform, we successfully performed the first comprehensive mutational analysis of three color-tuning switches in microbial rhodopsin, with each mutant independently replicated three times. The results of our systematic mutational screen across three known color-tuning sites, residues 129, 226, and 256 (Fig. 3), revealed position-dependent relationships between amino acid type and spectral shift, suggesting that each site influences retinal absorption through distinct mechanisms. At positions 129 and 256, significant correlations between spectral shift and side-chain volume were observed (Fig. 7B), implying that steric effects near the chromophore may partly explain the tuning behavior at these positions. A similar correlation is known for L105 in green-absorbing proteorhodopsin, the residue equivalent to L129 in GR [49]. The slope of the correlation of position 129 in GR (− 6.5 ± 1.9 mol cm⁻⁴) has a smaller absolute value than that reported for position 105 in GPR (− 12.2 mol cm⁻⁴). Notably, position 256 exhibited a larger volume dependence (− 22 ± 3 mol cm⁻⁴), with the five mutations involving the bulkiest side chains (Y, F, L, I, and M) showing the largest red-shifts among the mutants expressed in this study. A256 is located near the 13-methyl group of retinal (Fig. 3) and in proximity to Y225 in TM6, which interacts with the counterion D253 in TM7. One possible factor contributing to the large red-shifts is modulation of the retinal torsion angles by bulky side chains at position 256. Another possible factor is alteration of the interaction between the retinal Schiff base and the counterion, potentially mediated by Y225. Consistent with the latter possibility, substitution of A215 in bacteriorhodopsin (BR), which corresponds to GR A256, with a threonine residue (BR A215T) alters the interaction between Y185 and the counterion D212, which correspond to GR Y225 and D253, respectively [70]. Furthermore, the distribution of water molecules is also affected by this mutation. Therefore, the change in the volume of the residue at this position in GR may alter the electrostatic interaction of the retinal chromophore with D253, as well as the electrostatic environment around the Schiff base region, leading to the λ_max shift of the retinal chromophore. In contrast, position 226 showed minimal correlation with physicochemical properties of substituted amino acids. Taken together, our results highlight the complexity of spectral tuning in rhodopsins, which emerges from site-specific mechanisms likely involving both direct chromophore interactions and more distributed, context-dependent structural effects.

These position-dependent behaviors indicate that spectral tuning in rhodopsins cannot be explained by simple physicochemical descriptors and instead reflects the influence of local structural factors, including the spatial arrangement of neighboring residues, local electrostatics, and hydrogen-bonding networks, which vary across mutation sites. From a modeling perspective, this suggests that the determinants of spectral tuning are not fully captured by conventional residue-level descriptors, but instead reside in nonlinear, high-dimensional spaces defined by site-specific biophysical constraints. These findings highlight the importance of training ML models for wavelength prediction on large, systematically generated datasets, as enabled by platforms like Rhobot-Screen.

Conclusions

In this study, we developed Rhobot-Screen, an integrated experimental platform that enables systematic construction and quantitative characterization of predefined rhodopsin variants within a consistent 96-well plate format using a single liquid-handling system. By integrating site-directed mutagenesis, liquid-based cloning, protein expression, and microplate-based assays, Rhobot-Screen provides a reproducible and scalable framework for generating quantitative sequence–function datasets with reduced operator dependence, which are difficult to obtain using conventional manual workflows. Application of the platform to comprehensive mutagenesis of three color-tuning sites in GR demonstrated its ability to reliably produce standardized spectral data across dozens of variants. In addition, extension of the platform to mammalian cell-based assays enabled functional characterization of a GPCR-type rhodopsin through light-dependent intracellular signaling. Although the present study focused on absorption wavelength and GPCR signaling as model phenotypes, the modular design of Rhobot-Screen allows straightforward adaptation of the functional readout, making it suitable for AI-guided protein engineering studies targeting other rhodopsin properties, such as ion transport activity or fluorescence.

Methods

System configuration

A custom-configured Fluent® 780 liquid-handling system (Tecan, Switzerland) was used to carry out all automated dispensing and associated procedures (Fig. 1). The system includes a pipette arm, Flexible Channel Arm (FCA), with eight air-displacement pipetting channels, each capable of handling volumes between 1.0 and 1000 μL using disposable conductive tips. Liquid level detection is based on pressure and capacitive sensing. The arm accommodates standard Society for Biomolecular Screening (SBS)-format plates (96- and 384-well) as well as 13-mm and 16-mm tubes, with adjustable spacing between channels. The system was operated with three types of conductive disposable tips: 50 μL (DITI LIHA 50 μL CONDU.FIL.STE., Tecan, Switzerland), 200 μL (DITI LIHA 200 μL CONDU.FIL.STE., Tecan, Switzerland), and 1000 μL (DITI LIHA 1000 μL CONDU.FIL.STE., Tecan, Switzerland). Plate transfer across the workdeck (Fig. 2) is performed by a robotic gripper arm, which is equipped with an automatic finger exchange mechanism to adapt to different labware formats. The workdeck is enclosed within a front cover constructed from UV-resistant material and is maintained under positive-pressure HEPA filtration. UV illumination is included for decontamination. A multimode plate reader (Infinite200F, Tecan, Switzerland) is installed beneath the deck for absorbance and fluorescence measurements. The unit is equipped with optical filters for absorbance at 260, 280, and 600 nm, and for fluorescence with excitation at 485 nm and emission at 535 nm. An on-deck thermal cycler (ODTC; INHECO, Germany) is integrated into the system for temperature-controlled reactions. The system is operated through a dedicated industrial-grade personal computer running FluentControl (version 3.3 or later) and Fluent Gx software (Tecan, Switzerland).

DNA constructs and site-directed mutagenesis

The coding region of the GR gene (GenBank accession number: WP_011140202), tagged with a 6 × His sequence, was codon-optimized for Escherichia coli expression and cloned into the pET-21a(+) vector (Merck KGaA, Germany). Site-directed mutagenesis was performed using the QuikChange protocol (Agilent Technologies, CA) following the manufacturer’s instructions. Mutation-specific oligonucleotide primers were designed using the Agilent Technologies primer design tool (Additional file 1: Table S1). A thermal cycling reaction for mutagenic primer extension was conducted using PfuUltra™ II Fusion HS DNA Polymerase, and the resulting products were digested using DpnI at 37 °C for 1 h.

Transformation and clone isolation by limiting dilution

Transformation and selection of transformants by the limiting dilution method were performed using a Fluent® automated liquid-handling system. The resulting plasmids were mixed with E. coli JM109 (Takara, Japan) using chemically competent cells prepared according to a previously reported method [71]. The transformed cells were incubated on ice for 30 min, followed by heat shock at 42 °C for 45 s and an additional incubation on ice for 2 min. The cells were mixed in SOC medium in a 96-deep-well plate and incubated at 37 °C for 1 h for recovery. Transformed cells were subjected to limiting dilution for clone isolation. The recovered cells were serially diluted using LB medium supplemented with ampicillin (dilution factor: 3), and the 3.5-times dilution step was repeated five times to a final dilution of 1:525. The diluted suspension was dispensed into 96-well microplates at 16 replicates per sample. After overnight incubation at 37 °C, cell growth in each well was assessed by measuring the absorbance at 600 nm (OD₆₀₀) using Infinite200F (Tecan, Switzerland). Replicates that reached an OD₆₀₀ of 0.08–1.0 were designated as positive clones and used for subsequent plasmid purification.

DNA purification and sequencing

Plasmid DNA was purified using the Zyppy-96 Plasmid MagBead Miniprep kit (Zymo Research, CA) according to the manufacturer’s instructions. To verify the correct introduction of mutations and ensure the absence of off-target mutations in the rhodopsin coding region, plasmid sequences were confirmed by Sanger sequencing. Sequencing was performed by a commercial service (Eurofins Japan) using the T7 promoter primer (5′-TAATACGACTCACTATAGGG-3′) and T7 terminator primer (5′-ATGCTAGTTATTGCTCAGCGG-3′).

Protein expression

Protein expression of GR was conducted with slight modifications to previously described methods [38]. Plasmids were transformed into E. coli C43(DE3) (Lucigen, WI) or BL21(DE3) (NIPPON GENE, Japan), and cells were cultured in 800 μL of 2 × YT medium supplemented with 50 μg mL⁻¹ ampicillin in a 96-deep-well plate. Preincubation was performed at 37 °C and 1400 rpm for 12 h using a plate shaker (M-BR-032P, TAITECH, Japan). The preculture was diluted 1:100 into 1 mL of fresh 2 × YT medium containing 50 μg mL⁻¹ ampicillin and incubated under the same conditions. After 4 h, protein expression was induced by the addition of 1 mM isopropyl β-ᴅ-1-thiogalactopyranoside (IPTG) and 10 μM all-trans-retinal, followed by a 4-h incubation.

Solubilization and spectroscopy

To determine the λ_max of the generated mutants, a hydroxylamine (HA) bleaching assay was performed as described previously [21]. Cell pellets containing expressed rhodopsins were washed with bleaching buffer (133 mM NaCl, 66 mM Na₂HPO₄ (pH 7)) and resuspended in bleaching buffer containing 1 mM lysozyme. The cell suspensions were incubated for 1 h at 25 °C with shaking to digest their cell wall, and then solubilized with 3% (w/v) n-dodecyl-β-ᴅ-maltoside (DDM) in a 96-deep-well plate. After incubation for 1 h at 25 °C with shaking, the samples were centrifuged at 3500 × g for 15 min at 25 °C, and the supernatant was transferred to a 96-well microplate. After addition of HA (final conc. 125 mM), absorption spectra were recorded using a microplate reader equipped with a UV–visible spectrometer (FLUOstar Omega, BMG Labtech, Germany). Illumination was carried out using a white LED device equipped with a color filter (> 500 nm). Difference spectra were obtained by subtracting each spectrum from the pre-illumination spectrum. Data were omitted when the difference spectra did not exhibit a clear positive peak in the visible region with a spectral shape consistent with rhodopsin-derived absorption. Baseline shifts in the difference spectra, primarily arising from scattering, were corrected using a third-degree polynomial function of wavelength [72], fitted with Igor Pro software (WaveMetrics, OR) over a wavelength range in which GR shows no absorption. The λ_max values of visible-region peaks were determined by Gaussian fitting using Igor Pro software.

Protein expression and intracellular Ca²⁺ monitoring in mammalian cultured cells

COS-1 cells were cultured as described previously [73]. Cells were maintained in Dulbecco’s Modified Eagle’s Medium (DMEM) supplemented with 10% fetal bovine serum in a CO₂ incubator (5% CO₂, 37 °C). The coding sequences of SpiRh1 (GenBank accession number: BAG14330) and GCaMP6s [74] were cloned into the pcDNA3.1(+) vector (Thermo Fisher Scientific, Waltham, MA). The SpiRh1 construct carried a C-terminal Rho1D4 epitope tag (TETSQVAPA). Prior to transfection, COS-1 cells were manually seeded into 96-well plates at a density of 10,000 cells per well and incubated overnight. Subsequent transfection procedures were performed using the Fluent® system. Transient transfection was carried out using Lipofectamine® 2000 (Thermo Fisher Scientific, MA) according to the manufacturer’s instructions. To vary the expression level of SpiRh1, a dilution series of the SpiRh1 construct was prepared and used for transfection. In addition to 50 ng of the GCaMP6s construct, the SpiRh1 construct and blank vector DNA were mixed so that their total amount was 50 ng for each well. Cells were then transfected with this DNA mixture using Lipofectamine. Following transfection, cells were cultured for approximately 48 h at 37 °C in 5% CO₂. All subsequent procedures were conducted in the dark or under dim red light to avoid activation of SpiRh1. Before measurement, the culture medium was replaced with a CO₂-independent, phenol red-free medium (pH 7.0): DMEM (D2902, Merck KGaA, Germany) supplemented with 4.5 g L⁻¹ glucose, 10 mM HEPES, and 1 µM 11-cis-retinal. After incubation at room temperature for 3 h to allow pigment reconstitution, fluorescence signals from GCaMP6s were measured using the FLUOstar microplate reader (BMG LABTECH GmbH, Germany) with a GFP filter set at room temperature. Under these conditions, the excitation light used for fluorescence measurements was sufficient to activate animal rhodopsins without an additional light source, enabling detection of light-induced increases in intracellular Ca²⁺ levels, consistent with previous reports [62]. The fluorescence signal was normalized using the following calculation:

$$\mathrm{\Delta }F/F_0=\left(F,-,F_0\right)/F_0$$

where F is the fluorescence intensity at each time point and F₀ is the initial fluorescence value.

Abbreviations

AI

Artificial intelligence

DDM

n-Dodecyl-β-ᴅ-maltoside

DMEM

Dulbecco’s Modified Eagle’s Medium

GPCR

G-protein-coupled receptor

GR

Gloeobacter rhodopsin

HA

Hydroxylamine

HEPA

High efficiency particulate air

IP₃

Inositol 1,4,5-trisphosphate

IPTG

Isopropyl β-ᴅ-1-thiogalactopyranoside

λ_max

Absorption maximum wavelength

ML

Machine learning

PIP₂

Phosphatidylinositol 4,5-bisphosphate

PLC

Phospholipase C

PSB

Protonated Schiff base

SBS

The Society of Biomolecular Screening

UV

Ultraviolet

WT

Wild-type

Authors’ contributions

Methodology development and design were carried out by TN, MK, DRH, and KI. KI conceived and designed the study. TN, MK, and DRH performed experiments. TN and MK analyzed the data. The manuscript was written through the contributions of all authors. All authors have read and approved the final version of the manuscript.

Funding

This work was supported by JST CERST (Grant Number: JPMJCR22N2 to K.I.), JSPS KAKENHI Grants-in-Aid (Grant Number: JP23H04863/JP25K09698 to T.N., JP23K05007/26H00462 to M.K., JP24H02268/JP25H00424 to K.I.), the Research Foundation for Opto-Science and Technology (to M.K.), and MEXT Promotion of Development of a Joint Usage/Research System Project: Coalition of Universities for Research Excellence Program (CURE) (Grant Number: JPMXP1323015482 to K.I.).

Data availability

All data generated or analyzed during this study are included in this published article and its supplementary information files (Additional file 1 and Additional file 2).

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.