Data-driven construction of an imine reductase library capable of broad-scope reductive amination at equimolar substrate concentrations Authors: Sarah A Berger, Christopher Grimm, Marco Cespugli, Andreas Krassnigg, Tobias Schopper, Irene Marzuoli, Isabel Oroz-Guinea, Stephan Vrabl, Lukas Roemer, Melanie A Weber, Fabian M Kulier, Yuliya Orel, Bettina M Nestl, Christian C Gruber, Georg Steinkellner, Serena Bisagni, Francis Gosselin, Hans Iding, Kurt Puentener, Dennis Wetzl, Wolfgang Kroutil, Joerg H Schrittwieser Published: 2026 ABSTRACT AbstractReductive amination catalysed by imine reductases and reductive aminases has emerged as a powerful method for the asymmetric synthesis of chiral amines, yet the generality and practical limitations of these enzymes remain poorly understood. Here, we combine large-scale, iterative activity screening, sequence analysis and machine-learning-guided exploration to systematically profile reductive amination performance across 175 enzymes and structurally diverse substrate panels. We show that asymmetric reductive amination at equimolar substrate concentrations is widespread within this enzyme family and assemble catalysts that display broad scope, high specific activity and excellent stereoselectivity. Sequence-level analysis identifies characteristic features associated with broad-scope reductive aminases and pinpoints key residues governing stereoselectivity, while machine-learning models trained on the screening data enable successful prediction of enzyme activity for previously untested substrate combinations. Selected reactions were scaled to a preparative batch size, affording gram quantities of optically pure amine products.Subject terms: Biotechnology, Computational biology and bioinformatics, Biocatalysis, Synthetic chemistry methodology Introduction Chiral amines are a wide-spread structural motif in pharmaceuticals, agrochemicals, and natural products1. Over 90% of the 200 top-selling small-molecule drugs in 2025 contain at least one nitrogen atom, with 47% containing a non-racemic α-chiral amine or amino acid moiety2,3. The most versatile method for synthesising chiral amines is asymmetric reductive amination, which proceeds through the condensation of a carbonyl compound and a primary or secondary amine to give an imine or iminium intermediate, followed by asymmetric reduction to the target chiral amine. With two variable residues on each substrate, reductive amination offers a broad scope, theoretically allowing the convergent synthesis of almost any α-chiral amine (cf. Figure 1a). Visual outline of the present study. a General reaction scheme of IRED-catalysed reductive amination. Unless residues R1 and R2 are identical, the formed amine product is chiral. b Schematic visualisation of the three-dimensional screening space. Representative coverage by non-exhaustive screening depends on the balanced selection of enzymes and substrates. c Summary of computational and experimental workflows and overall outcome of the study. Substrate and enzyme panels were selected based on chemo- and bioinformatic analyses. Experimental screening was performed in two phases (Overview Screening and Diversification Screening) and the resulting data were used to train AI models with the aim to establish structure−activity relationships in the enzyme and substrate dimensions. The models were used to suggest additional IREDs for experimental testing, either on known substrate combinations (AI Activity Prediction) or on substrate combinations outside of the training data set (Deep-Learning Pipeline). Arrows represent the flow of information between the different parts of the study. Recent years have witnessed impressive advances in transition metal-catalysed, organocatalytic, and biocatalytic asymmetric reductive amination4–7. Metal-catalysed approaches typically use chiral iridium(I), rhodium(I), or ruthenium(II) complexes at elevated hydrogen pressure ( > 10 bar) and have demonstrated broad ketone and amine scope. Organocatalytic methods predominantly rely on axially chiral phosphoric acids as catalysts, Hantzsch esters, silanes or boranes as reducing agents, and aniline derivatives as amine substrates. Enzymatic reductive amination has long been limited to the preparation of α-amino acids from α-keto acids using amino acid dehydrogenases and the synthesis of α-chiral primary amines from ketones using transaminases (the latter being a reversible group transfer process)7–11. Recently, however, the scope of biocatalytic reductive amination has been significantly expanded with the discovery that imine reductases (IREDs)—and in particular a sub-class of this enzyme family termed reductive aminases (RedAms)—are able to perform this reaction with a broad range of amine and carbonyl substrates using NADPH as the reducing agent 12–14. Following the initial discovery of asymmetric reductive amination by IREDs and RedAms15–17, sequence-based and metagenomic enzyme discovery efforts have rapidly generated extensive libraries of these biocatalysts18–22, and protein engineering has been used to tailor them to specific target reactions23–31. Within a decade, IRED-catalysed reductive amination has progressed from proof of concept to industrial application13, including the recent syntheses of a lysine-specific demethylase-1 (LSD1) inhibitor23, a Janus kinase 1 (JAK1) inhibitor24, and a cyclin-dependent kinase inhibitor on a kilogram–ton scale27,28. Despite these advances, the scope and limitations of asymmetric reductive amination catalysed by IREDs and RedAms need to be further explored. Many screening campaigns have focused on a narrow selection of model substrates, particularly in the amine dimension. Additionally, many studies use limited concentrations of the carbonyl compound and a large excess of the amine substrate (commonly 5–10 mM and 10–50 eq., respectively)17,18,20,22, which helps to achieve faster biotransformations but negatively impacts the resource efficiency and preparative applicability of these reactions. Screening multiple-substrate catalytic reactions for diversity is intrinsically challenging, as multiplicative growth of the parameter space to be covered results in a large experimental workload, a problem known in data science as the “curse of dimensionality”32. Literature-reported screenings of IRED-catalysed reductive amination have included over 150 carbonyl compounds and more than 50 amines, and the number of currently identified IRED sequences is in the thousands (the Imine Reductase Engineering Database, maintained by Pleiss and co-workers at the University of Stuttgart and available online at https://ired.biocatnet.de/, currently lists 1,409 sequences). Covering this search space exhaustively would hence require the analysis of more than 10 million individual substrate–enzyme combinations, a number and chemical diversity that appear overwhelming even for modern high-throughput experimentation33–35. Non-exhaustive screening, however, depends on the selection of representative substrate and enzyme panels and still requires assay methods that offer high throughput and broad scope. Even if these prerequisites are met, non-exhaustive screening leaves a significant part of the search space unexplored, which poses challenges for systematic data analysis and predictive modelling36. These fundamental problems have been addressed in various chemical contexts and in diverse ways, ranging from combinatorial screening strategies that increase experimental throughput37–40, through predictive modelling in the substrate and catalyst dimensions41–44, to fully autonomous robotic experimentation (“self-driving labs”)45–48. In this work, we present a general strategy for the iterative, multi-dimensional screening of catalytic reactions, which we have used to broadly explore the IRED sequence space for reductive amination activity at equimolar substrate concentrations. Our approach involves swift elimination of underperforming enzymes early in the screening campaign, with an emphasis on activity and substrate scope over stereoselectivity. Additionally, we leverage machine learning for performance prediction, and employ analysis methods of moderately high throughput that should be accessible to many catalysis laboratories. The screening data collected in this study provide a representative fingerprint of reductive amination performance in the IRED/RedAm enzyme family and have pinpointed IREDs that catalyse synthetically interesting reductive amination reactions with high activity and stereoselectivity. Results Screening strategy Our screening strategy is iterative in all three dimensions (carbonyl, amine, and IRED; cf. Figure 1b) and comprises two phases: First, we rapidly evaluated the reductive amination performance of candidate enzymes using a limited substrate panel that is covered exhaustively (Overview Screening, 8 carbonyls × 7 amines; Fig. 1c). We eliminated poorly performing IREDs from the screening panel and replaced them with new candidates identified using a machine-learning model trained on the collected activity data. Next, we explored the substrate scope of the best-performing IREDs more broadly using a split screening that covers a diverse range of substrates non-exhaustively (Diversification Screening, 24 carbonyls × 4 amines and 6 carbonyls × 16 amines; Fig. 1c). Although combinatorial reactions of multiple substrates or pooling of reactions before analysis can expedite screening37–40,49, we chose to carry out and analyse reactions individually to generate a data set for machine learning that is as unaffected by inhibition phenomena and analytical artifacts as possible. Table S27 (Supplementary Information) provides a tabular overview of the computational and experimental contributions to the overall screening workflow. In-silico analysis In the first stage of the screening process, we had to find representative panels of substrates and enzymes. To enable a rational and balanced selection, we compiled large collections of candidates, parametrised their structural properties, and applied dimensionality reduction by principal component analysis (PCA) and multidimensional scaling (MDS) to identify candidates that cover the search space relatively evenly. Amines and carbonyl compounds included in the analysis comprised 136 literature-known IRED screening substrates and 55 additional structures relevant for medicinal chemistry applications (137 carbonyls and 54 amines in total; see Table S1). Ten descriptors based on two- and three-dimensional structure were computed for the amine and carbonyl substrates to represent their molecular geometry as well as their chemical characteristics (Table S28). Principal component analysis revealed clustering based mainly on the size and diameter of the compounds, the presence of halogens and, in case of carbonyls, of ester groups (Figs. S2, S3). Multidimensional scaling confirmed these findings (Figs. S4, S5). We selected eight carbonyl compounds and seven amines for experimental testing in the Overview Screening, and 30 carbonyl compounds and 19 amines for the Diversification Screening (see Table of Screening Substrates in the Supplementary Information). These selections were made to thoroughly cover the majority of the principal component space, deliberately excluding regions of minor interest, such as small carbonyl compounds like acetone, 2-butanone, and butanal. IRED sequences were gathered from various literature sources (mainly refs. 18–20,22,50, and 51; see Table S30 for complete literature references) and from the BioCatNet IRED database (v. 3, accessed in January 2021). C = N-reducing enzymes outside the canonical IRED/RedAm family52 were not considered, as we deemed a relatively homogeneous sequence basis an important prerequisite for extracting generalisable insights from the screening data. The initial dataset was expanded with additional sequences identified through BLAST searches53 on the NCBI nr (non-redundant) database, using as queries nine IRED sequences reasonably distant from each other on a phylogenetic tree (see Figure S6). Redundancy removal at an identity threshold of 99% resulted in a dataset of 1,688 sequences. Further refinement was necessary to remove 68 β-hydroxyacid dehydrogenases (β-HADs), which share remarkable sequence similarity with IREDs (around 25% pairwise identity) but do not possess reductive amination activity. The filtering criterion was the presence of a catalytic lysine residue in the active site instead of the aspartic acid or tyrosine commonly found in IREDs54,55. Homology models of the resulting 1620 sequences were generated based on suitable templates (Figure S7 and Table S29), and 65 further enzymes were removed based on model quality. Thus, the final in-silico panel comprised 1555 curated sequences (available in CSV format at https://doi.org/10.17632/jn4sfsrz8x.2). From these, 115 IREDs (110 of which could be expressed in soluble form) were initially selected for the Overview Screening and later supplemented with batches of 45 and 20 additional enzymes based on machine-learning (ML) predictions of their reductive amination performance. Overview screening The Overview Screening required the analysis of 9800 individual biotransformations (8 carbonyls × 7 amines × 175 enzymes), a number that exceeds the capacities of standard chromatographic methods. As alternatives, we considered a colorimetric screen (called IREDy-to-go) developed by Turner and co-workers, which detects the formation of NADPH in the reverse reaction (i.e., oxidative deamination)22, and the photometric quantification of NADPH consumption in the reductive amination reaction, the feasibility of which has been demonstrated by Höhne and co-workers56. We chose the latter option because it provides a more direct measure of the desired activity and implemented it in a 96-well microtiter plate format that required two plates per enzyme for the triplicate analysis of all 56 substrate combinations and a ‘no amine’ negative-control condition (see Photometric IRED Activity Assay: Overview Screening, Supplementary Information). All carbonyl and amine substrates, except ammonia (500 mM, 10 eq.), were screened at equimolar concentrations of 50 mM each, and NADPH was supplied at 2 mM. Under these conditions, two substrate combinations (benzaldehyde–benzylamine and benzaldehyde–aniline) gave turbid assay mixtures, which prevented the photometric monitoring of NADPH consumption. All other combinations could be assayed without interfering absorption of light by substrates, imine, or product at a wavelength of 370 nm (ϵNADPH = 2216 M–1 cm–1). Although all investigated IREDs were expressed with an N-terminal His-tag, we sought to avoid the effort of purifying all 175 proteins by establishing the assay using crude, lyophilised cell-free extracts. To enable the reporting of specific activities, we normalised the apparent rates for IRED expression level, determined using the Bradford protein assay and SDS-PAGE pixel densitometry (see Determination of Expression Level, Supplementary Information). Previous studies suggest that the use of non-purified enzymes could result in pronounced background NADPH depletion; for instance, due to competing carbonyl reduction by endogenous E. coli ketoreductases15,56. Gratifyingly, we found that our initially chosen assay conditions (bicine–NaOH buffer, 100 mM, pH 8.0, 10% v/v DMSO; 2 mg/mL crude IRED, corresponding to a substrate/enzyme ratio [w/w] of approx. 2–4) resulted in an acceptable background NADPH depletion rate ( < 5 mAU/min) in negative control reactions performed in the absence of amine, even for easily reduced carbonyl substrates such as benzaldehyde (c1) and cyclohexanone (c2). This low background, combined with a statistical criterion for hit identification that takes the standard deviation of the triplicate biotransformations into account (Equation 3, Supplementary Information), allowed us to reliably detect reductive amination activities as low as 5–10 mU/mg of IRED. The results obtained in the photometric assay were validated by re-analysing selected substrate–enzyme combinations chromatographically. Biotransformations on a 500-µL scale were set up using the appropriate substrates and IRED combined with phosphite dehydrogenase (PtDH) for cofactor regeneration, incubated for 24 h, and analysed by GC–MS or HPLC–MS. Three strategies guided the selection of substrate–enzyme combinations for re-analysis: (i) 50 combinations (25 positives, 25 negatives) were randomly selected to estimate the overall rate of false-positive and false-negative results, (ii) each positive substrate pair was re-analysed at least once, and (iii) low-level positives were given precedence for re-analysis, in particular if the respective substrate pair had yielded no high-level positives. These data clean-up measures not only confirmed the overall validity of the photometric assay (Tables S14, S15), but also contributed to the elimination of low-level false positives (Tables S16, S17). The consolidated activity data demonstrate that reductive amination at equimolar substrate concentrations is more common across the entire IRED enzyme family than previously thought. Out of the 9,450 interpretable substrate–enzyme combinations, 1437 (15%) were positives, and 36 of the 54 possible products (67%) could be accessed by at least one enzyme (Fig. 2). Six of the 18 products that were inaccessible derive from α-ketoester c5, which was found to undergo competing ester hydrolysis under the assay conditions, a side reaction that severely interferes with the desired reductive amination process. Summary of positives identified in the Overview Screening. The coloured rectangles (tiles) show the products resulting from the 56 tested substrate combinations, along with the number of active IREDs for the respective combination given in the bottom-left corner of each tile. Activity was assayed spectrophotometrically by following the consumption of the IREDs’ cofactor, NADPH, and confirmed by re-screening of selected positives using GC–MS or HPLC–MS analysis. The colour scale represents the number of confirmed hits found for each substrate combination. Two reactions that could not be assayed spectrophotometrically due to sample turbidity (c1-a6, c1-a7) are shown as grey tiles. a Dibenzylamine (Bn2NH) was formed alongside the expected benzylamine, and the latter was present in the EtOAc-extracted samples as the benzylidene imine (N-benzylidenebenzylamine). Source data are provided as a Source Data file. Aggregating the determined activities for each substrate across all coupling partners reveals general reactivity trends (Fig. 3a, b): In line with previous findings, benzaldehyde (c1) and cyclohexanone (c2) were the most reactive carbonyl substrates, but different from earlier reports17,20,21, the aromatic ketone acetophenone (c3) was accepted more readily than the aryl-aliphatic phenylacetone (c8). Among the tested amines, propargylamine (a4) and methylamine (a2) were accepted best, while the secondary amine piperidine (a5) proved to be a slightly more challenging substrate than the α-branched but primary isopropylamine (a3). Analysis of data collected in the Overview Screening . a, b Box plots of non-zero activities (scaled logarithmically) of each carbonyl (a) and amine (b) substrate, showing median (centre band), 25th and 75th percentiles (lower and upper hinge of box), and data range (whiskers). Individual data points are shown as semi-transparent circles. n = 875 individual combinations, aggregated (mean) over three technical replicates each, for substrate c1, 1225 for substrates c2–c8, 1400 for substrates a1–a5, 1225 for substrates a6, a7. c Scatter plot of maximum activity vs. number of positive substrate combinations for each of the 175 tested IREDs. The labelled enzymes along the top-right front are compared to each other in Figure S15 and Table S5 (Supplementary Information). d Box plot of non-zero activities (scaled logarithmically) grouped by enzyme panel, showing median (centre band), 25th and 75th percentiles (lower and upper hinge of box), and data range (whiskers). Extreme values (data points farther than 1.5 times the inter-quartile range from the hinges) are shown as semi-transparent circles. n = 5940 individual combinations, aggregated (mean) over three technical replicates each, for the starting panel, 2430 for the first additions, 1080 for the second additions. Source data are provided as a Source Data file. To compare the performance of different enzymes, we found it helpful to visualise the collected data along two aggregated dimensions: the maximum activity found across all tested substrate combinations (activity score), and the number of positive substrate combinations (generality score). This allows a rough categorisation of IREDs into generalists vs. specialists and fast vs. slow enzymes. As shown in Fig. 3c, fast generalists (top-right quadrant) are notably absent, as the IREDs that reached the highest maximum activities are only moderately general, and vice versa, suggesting an inherent trade-off between substrate scope and catalytic efficiency in any single reaction. The enzymes located closest to the top-right quadrant (labelled in Fig. 3c) share an average sequence identity of only 46% (with pairwise identities ranging from 26% to 96%) and originate from diverse phylogenetic backgrounds. However, most of them possess key active-site residues (Asn93, Asp169, Tyr177; AspRedAm numbering) that have been assigned a catalytic role in the proposed reaction mechanism for fungal RedAms (Fig. S15 and Table S5)12,17,57. In general, we found the presence of these key residues to be a useful predictor for reductive amination performance in the Overview Screening. As discussed above (see Screening Strategy, In-Silico Analysis), the 175 tested IREDs were screened iteratively in three stages. Initially, a panel of 110 enzymes was tested, and the collected data were used to train a machine-learning (ML) model that correlated the determined activities to differences in physico-chemical descriptors of the enzymes’ active sites, formulated as dissimilarity scores among the corresponding set of Catalophore™ point clouds (for details, see Activity Prediction Using Machine-Learning Models, Supplementary Information). Results from this model were used to predict two additional panels of 45 and 20 IREDs expected to possess good reductive amination activity for the substrate set of the Overview Screening. As shown in Fig. 3c, the most general IREDs were found in the starting panel, while the enzymes recruited based on ML predictions tended towards moderate substrate scope but high maximum activity. This is reflected in an increase in the median non-zero activity across all substrate combinations from 74 mU/mg in the starting panel to 150 mU/mg (2-fold increase) in the first addition panel and to 253 mU/mg (3.4-fold increase) in the second addition panel (Fig. 3d). We attribute the observed trend towards high maximum activity and moderate generality to characteristics of the k-nearest-neighbour model and the clustering and selection process (for detailed discussion, see Activity Prediction Using Machine-Learning Models, Supplementary Information). In summary, the ML model predicted high-activity enzyme candidates based on three-dimensional representations of the active sites and an experimental activity data set of 5940 points (960 non-zero points). As most of the investigated reductive amination reactions lead to chiral amine products, stereoselectivity is another important measure of an enzyme’s performance next to activity and substrate scope. To gather representative stereoselectivity data, we performed a total of 128 small-scale biotransformations covering 25 positive substrate combinations and determined conversion and enantiomeric excess of the products by GC and/or HPLC analysis. In agreement with previous studies on IRED-catalysed reductive amination17,20,58, stereoselectivity varied widely across substrates even for the same enzyme, but among the investigated biotransformations were many that proceeded with high conversions and excellent enantio- and/or diastereoselectivity (Table 1; see Table S21 for complete data). Most reactions afforded the (R)-enantiomer of the product, but in eight cases both enantiomers could be accessed (e.g., c6-a2, Table 1, entries 10–12; c6-a4, Table 1, entries 13 − 15). The reductive amination reactions of hexane-2,5-dione (c4) predominantly proceeded with high enantio- but moderate diastereoselectivity (Table 1, entries 4 − 9) and were accompanied by the formation of the corresponding pyrroles through a Paal−Knorr cyclisation that was also observed in negative control reactions performed in the absence of IRED. This outcome is largely consistent with the findings of Cosgrove, Turner, and co-workers, who used similar 1,4-diketones as substrates for IRED-catalysed reductive amination59. The amphetamine derivative c8-a4 was formed with moderate conversion and high, but not exclusive, selectivity for the (R)-enantiomer by several enzymes. (R)-c8-a4 is a precursor to selegiline, a pharmaceutical drug that can be obtained from c8-a4 by N-alkylation and that is used for the treatment of Parkinson’s disease and major depression 31,60. Table 1Conversions and stereoselectivities (enantiomeric excess, ee; diastereomeric ratio, dr) of selected IRED-catalysed reductive amination reactions of the Overview ScreeningEntryCarbonylAmineIREDProductConv. (%)ee (%)dr1c3a4IR00411685 (R)2c3a4IR0043499 (R)3c3a4IR0059198 (S)4c4a4IR003521>99 (B)a85:155c4a4IR201010>99 (B)a90:106c4a4IR00667>99 (B)a23:777c4a6IR214713>99 (A)a9:918c4a6IR04998>99 (A)a7:939c4a6IR0015491 (B)a21:7910c6a2IR20215695 (R)11c6a2IR00684899 (R)12c6a2IR185539>99 (S)13c6a4IR006893>99 (R)14c6a4IR176085>99 (R)15c6a4IR03953699 (S)16c6a6IR14536599 (R)17c6a6IR214758>99 (R)18c6a6IR146043>99 (R)19c6a7IR0032332 (R)20c6a7IR00411>99 (R)21c7a2IR00205083 (R)22c7a2IR003210>99 (S)23c7a4IR00687997 (R)24c7a4IR023346>99 (R)25c7a4IR0032498 (S)26c7a6IR21479195 (R)27c7a6IR18556894 (R)28c8a2IR00522068 (R)29c8a2IR0035998 (R)30c8a4IR00415895 (R)31c8a4IR00355597 (R)32c8a4IR185517>99 (S)Reaction conditions: Carbonyl compound (50 mM), amine (50 mM), NADP+ (1 mM), Na2HPO3 · 5 H2O (100 mM), IRED preparation (4 mg/mL lyophilised cell lysate), PtDH preparation (4 mg/mL lyophilised cell lysate), bicine–NaOH buffer (100 mM, pH 8.0), DMSO (10% v/v), reaction volume 500 µL, 30 °C, 800 rpm, 24 h. a (A) and (B) refer to the first-eluting and second-eluting enantiomer, respectively, where absolute configurations are unknown. Conversions and stereoselectivities (enantiomeric excess, ee; diastereomeric ratio, dr) of selected IRED-catalysed reductive amination reactions of the Overview Screening Reaction conditions: Carbonyl compound (50 mM), amine (50 mM), NADP+ (1 mM), Na2HPO3 · 5 H2O (100 mM), IRED preparation (4 mg/mL lyophilised cell lysate), PtDH preparation (4 mg/mL lyophilised cell lysate), bicine–NaOH buffer (100 mM, pH 8.0), DMSO (10% v/v), reaction volume 500 µL, 30 °C, 800 rpm, 24 h. a (A) and (B) refer to the first-eluting and second-eluting enantiomer, respectively, where absolute configurations are unknown. Diversification screening We advanced the 100 best-performing IREDs (55 from the initial panel; 30 and 15 from the first and second addition panel, respectively) into a second screening phase to explore their reductive amination scope using a broader set of substrates, including larger and more functionalised compounds commonly used in medicinal chemistry. In this Diversification Screening, a total of 30 carbonyls and 19 amines were investigated, resulting in 57,000 individual substrate–enzyme combinations for exhaustive coverage. As this large number exceeded our screening capacity, we decided to explore the substrate space non-exhaustively by splitting the screening into two parts: the carbonyl scope (24 carbonyls × 4 amines) and amine scope (6 carbonyls × 16 amines), each implemented in a 96-well plate format (Figs. S31, S32). While this reduces the number of individual combinations to 19,200, we still considered triplicating every reaction (as was done in the Overview Screening) too experimentally costly. Therefore, only the negative controls—performed in the same plate layouts but using cell-free extract not containing any IRED (“empty-vector controls”)—were multiplicated 9–10 times, while the enzymatic assays were carried out only once for each IRED. This makes the experimental replicates an additive rather than a multiplicative effort, but still provides an estimate of global random error that allows the identification of hits based on a statistical criterion (Eqs. 5 and 7, Supplementary Information). The photometric NADPH consumption assay used in the Overview Screening could be readily adapted for the amine scope plate, with only six of 96 substrate combinations showing interfering turbidity. However, about a third of the combinations on the carbonyl scope plate proved unsuitable for photometric analysis due to turbidity or elevated background absorbance, making an alternative assay method necessary. Our recently developed flow-injection analysis mass spectrometry (FIA–MS) method61, which quantifies product formation by relating the mass signal (M + H+) of the target amine to the mass signal of the bicine buffer, worked well for all combinations except one, in which the molecular masses of product and buffer coincide (c29-a2, m/z = 164 for M + H+). FIA–MS thus allowed us to screen a diverse set of carbonyl substrates at a substantial concentration (50 mM), even when the reaction mixtures were not fully homogeneous. While the Overview Screening had been set up manually using multi-channel pipettes, the two assay setups required for the Diversification Screening were automated using a liquid-handling robotic platform. This reduced the required hands-on time and improved throughput, as eight amine scope plates could be conveniently prepared and analysed in a single overnight run without any user interference. However, the robotic pipetting resulted in photometric reads of slightly lower quality compared to the manual preparation of plates, possibly due to changing the order of reagent addition to minimise the time required for the last pipetting operation and transfer of the plate into the reader (for details, see Photometric IRED Activity Assay: Diversification Screening, Supplementary Information). Nevertheless, 8916 of the 9000 analysed assay reactions (99%) gave interpretable absorbance curves. For the carbonyl scope, samples were prepared autonomously by the robot and manually transferred to the autosampler for FIA–MS analysis. Hits were identified from the FIA–MS integral data by comparing the target/buffer ratio in enzymatic reactions to those in the multiplicate empty-vector controls, with no need for calibrating the target response (for details, see FIA–MS Activity Assay: Diversification Screening, Supplementary Information). As was the case for the Overview Screening (see above), we used chromatographic re-analysis of selected biotransformations to consolidate the results of the photometric and FIA–MS assays (Tables S18, S19). This quality assurance step was particularly important for the carbonyl scope data, as the high sensitivity of single-ion monitoring MS resulted in product mass integrals significantly above the empty-vector background even for reactions that formed only minute levels of product. A cut-off for the target/buffer integral ratio was hence introduced to narrow down the identified hits to those that produced detectable amounts of product according to GC–MS or HPLC–MS. Applying this threshold reduced the FIA–MS hit rate from 32% to 25% (see Validation Experiments for the Diversification Screening, Supplementary Information). The complete results of the Diversification Screening are shown in Figs. 4, 5. To create these visualisations, the carbonyl scope and amine scope data, which are based on different performance metrics (specific activity in mU/mgIRED for the amine scope, and target/buffer integral ratio as a dimensionless number for the carbonyl scope), had to be mapped onto a common scale. We found that both data sets, when plotted logarithmically, are reasonably well-described by Gaussian distributions (see Data Transformation for Generating z-Score Heatmaps, Supplementary Information), enabling their transformation into z-scores (distance from mean in units of standard deviation). The aggregated data confirm our earlier conclusion that IRED-catalysed reductive amination at equimolar substrate concentrations is more common than previously believed. Overall, 27% of the interpretable assay reactions were positive (2422 of 9500 = 25% for the carbonyl scope, 2529 of 8916 = 28% for the amine scope; coloured tiles in Figs. 4, 5), and 118 out of 185 possible reductive amination products (64%) could be produced by at least one of the investigated IREDs (Figs. S24, S25). Nevertheless, certain substrate motifs proved challenging for all tested enzymes, such as tert-butylamine (a14), which was not accepted at all, secondary amines, the heterocyclic primary amines a17 and a18, and bulky, aromatic ketones (e.g., c32–c35; Fig. 5). At the other end of the reactivity spectrum are amines like allylamine (a9) or cyclopropylamine (a10) and carbonyl compounds like 2-(benzyloxy)cyclohexanone (c18), 3-phenylcyclohexanone (c19), or tert-butyl 3-oxopyrrolidine-1-carboxylate (c16), which were accepted in 46–90% of the assayed reactions. The encouraging results with bulky, alicyclic carbonyls suggest that the poor acceptance of aromatic substrates has electronic rather than steric origins, a conclusion we could support by protein–ligand docking and by correlating z-scores with chemoinformatic descriptors of reactivity (see Protein Structure–Activity and Structure–Selectivity Analysis, Supplementary Information). Heatmap representation of all amine scope data collected in the Diversification Screening . Each of the 100 enzymes was screened against 6 carbonyl compounds × 16 amines, and the collected specific activities were logarithmised and converted to z-scores (distance from mean in units of standard deviation) as described in the Supplementary Information. The z-scores are visualised using a diverging purple–orange colour scale with white as the centre point (mean, z = 0). Grey tiles represent inactive substrate–enzyme combinations, and black tiles indicate combinations for which no data are available. Enzymes are grouped according to the panel from which they were sourced, sorted within each panel by the sum of activities (highest on top), and labelled alternatingly on both sides for better readability. The twenty enzymes that showed the broadest substrate scope overall (amine and carbonyl scope combined) are underlined. Heatmap representation of all carbonyl scope data collected in the Diversification Screening . Each of the 100 enzymes was screened against 24 carbonyl compounds × 4 amines, and the collected target/buffer integral ratios were logarithmised and converted to z-scores (distance from mean in units of standard deviation) as described in the Supplementary Information. The z-scores are visualised using a diverging purple–orange colour scale with white as the centre point (mean, z = 0). Grey tiles represent inactive substrate–enzyme combinations, and black tiles indicate combinations for which no data are available. Enzymes are grouped according to the panel from which they were sourced, sorted within each panel by the sum of activities (highest on top), and labelled alternatingly on both sides for better readability. The twenty enzymes that showed the broadest substrate scope overall (amine and carbonyl scope combined) are underlined. The data collected in the Diversification Screening do not indicate any trade-off between the substrate scope of an IRED and its maximum activity for any single substrate pair, in contrast to the findings of the Overview Screening (cf. Figure 3c). In fact, the enzymes with the broadest substrate acceptance are also among those with the highest individual activities, irrespective of whether the amine and carbonyl scope data are analysed separately or combined after z-score transformation (Fig. S18, Table S7). The twenty top-performing IREDs (underlined in Figs. 4, 5 and highlighted in Fig. S18) are active on more than one-third (34 − 44%) of the investigated substrate combinations and display maximum activities of 1.0 − 27.7 U/mgIRED for the most favourable amine−carbonyl pair. Sequence conservation analysis (Table S8) identified 12 positions—most of them outside the active site and with a putative role in general structure stabilisation (Figs. S19, S20)—that are conservationally enriched in these 20 broad-scope enzymes. Chromatographic re-investigation of representative positives with a focus on the top-20 IREDs revealed high conversions and excellent enantiomeric purities in many cases (Fig. 6a; see Table S22 for complete data). Access to both product enantiomers was achieved for 15 of the 22 chromatographically analysed substrate pairs, enabled by the presence of IREDs in our enzyme panel whose stereopreference is opposite to that of established fungal RedAms62. Notably, we were able to use protein–ligand docking (Table S10 and Fig. S23) and multiple-sequence alignments (Table S11) to establish a robust link between three active-site key residues and the observed enantiopreference of the enzymes. Hence, our iterative screening approach has succeeded in identifying IREDs that catalyse reductive amination at equimolar substrate concentrations with broad scope, high activity, as well as high and complementary stereoselectivity, while also providing a basis for generalisable protein structure–function analysis. Results of selected IRED-catalysed asymmetric reductive amination reactions. Conversion (c), enantiomeric excess (ee), and diastereomeric ratio (dr, where applicable) were determined by GC–FID and HPLC–DAD analysis. Yield refers to isolated yield after purification by column chromatography. a Biotransformations on analytical scale (0.5 mL) using substrate combinations covered in the Diversification Screening. b Biotransformations on analytical scale (0.5 mL) using substrates of the Overview and Diversification Screenings in novel combinations. IREDs were selected for testing using a deep-learning (DL) pipeline trained on the available screening data. c Biotransformations on preparative scale (200 mL). General reaction conditions: Carbonyl compound (50 mM), amine (50 mM), NADP+ (1 mM), Na2HPO3 · 5 H2O (100 mM), IRED preparation (4 mg/mL lyophilised cell lysate), PtDH preparation (4 mg/mL lyophilised cell lysate), bicine–NaOH buffer (100 mM, pH 8.0), DMSO (10% v/v), reaction volume 500 µL (a, b), 200 mL (c), 30 °C, 800 rpm, 24 h. a (A) and (B) refer to the first-eluting and second-eluting enantiomer, respectively, where absolute configurations are unknown. b Conversion given as sum of lactam and open-chain amino ester products. c Increased amine substrate concentration (100 mM, 2 eq.) used. d Increased IRED loading (8 mg/mL lyophilised cell lysate) used. Despite this gratifying outcome, it is essential to acknowledge the limitations of our approach to the Diversification Screening, which only covered 32% of all potential substrate combinations experimentally (18,416 of 57,000, considering only interpretable assay reactions). To address this limitation, we explored whether a machine-learning model trained on the available screening data could make useful activity predictions for amine−carbonyl pairs outside of the training set. We chose a gradient-boosted random forest for this task, using physico-chemical descriptors of the substrates, deep learning-derived representations of the IREDs’ active-site point clouds, and z-transformed screening data as inputs (for details, see Activity Prediction Using Machine-Learning Models, Supplementary Information). To assess the predictive performance of the model, we generated ranked lists of IREDs for four novel combinations of substrates (c16-a6, c16-a15, c23-a10, c24-a6) and picked the top 15 and bottom 5 candidates in each list for experimental testing. As shown in Tables S23–S26, the top-15 picks comprised positives with substantial conversion and ee in all cases, while the 5 lowest-ranked IREDs consistently showed little to no activity. The 20 top-performing IREDs of the Diversification Screening (cf. Table S7) generally rank high in the DL suggestions, but the model showed no tendency to overrepresent them, identifying also less obvious candidates for experimental testing. In fact, in two cases these turned out to afford the best results in terms of conversion and enantioselectivity (IR0464 for c16-a15, IR2147 for c24-a6, see Fig. 6b). This indicates that the DL model possesses predictive power for substrate pairs outside its training set that goes beyond pure face validity. To demonstrate the synthetic applicability of the investigated biocatalytic reductive amination reactions, we scaled five of them to a preparative batch size (10 mmol carbonyl substrate, 200 mL reaction volume). Near-complete conversions (93% or higher) were achieved in all cases, and the desired products were isolated in 63 − 89% yield and high enantiomeric purity (98% to >99% ee, see Fig. 6c). In line with the results of analytical-scale biotransformations, the reductive amination of ethyl levulinate (c10) and cyclopropylamine (a10) afforded a mixture of lactam and open-chain amino-ester products, which could be readily cyclised by treatment with a commercial lipase (Novozym® 435, now marketed as Sustine® 110 IM). Discussion In summary, we have successfully used a data-driven, iterative screening approach to assemble a panel of imine reductases that perform asymmetric reductive amination at equimolar concentrations of the amine and carbonyl substrates with high activity, broad scope, and good stereoselectivity. Instrumental in this process were the balanced, rational selection of enzymes and substrates for experimental testing based on systematic in-silico analysis, the use of photometric and mass spectrometry-based screening methods that combine sufficient throughput with broad applicability and a low rate of false positives and negatives, and the feedback of machine-learning models trained on the screening data, which guided our recruiting of additional IREDs and allowed us to extrapolate structure−activity relationships from the experimental data set to substrate combinations that were not covered by our screening. The abundance of activity data this study has generated (27,866 total data points) is made openly accessible in a machine-readable format for re-analysis and re-use by other researchers. It represents a balanced fingerprint of reductive amination performance across the entire IRED/RedAm sequence space, without bias towards positive results63 (77% of the tested substrate−enzyme combinations are non-hits). The data have been consolidated through experimental replicates, negative controls, and partial chromatographic re-analysis. They will serve as a valuable resource for the training and validation of machine-learning models that predict reductive amination outcome either from experimental data or from in-silico analysis of substrate and enzyme structure directly (e.g., IREDFisher)43. We also note that the z-score transformation introduced in our study provides a framework for integrating performance data from other sources (e.g., previous literature or future screening campaigns) into the present data set. Overall, our screening demonstrates that reductive amination at equimolar substrate concentrations is feasible with a substantial share of IRED enzymes and across a broad range of carbonyl and amine coupling partners. Out of the nearly 28,000 assay reactions conducted, a notable 23% (6,388 substrate−enzyme combinations) were positives, and 64% of the investigated amine products (154 out of 239) could be formed by at least one enzyme. The most proficient IREDs identified in our study combine a broad substrate scope (70 − 100 total positive combinations for Overview Screening and Diversification Screening) with high specific activities for the most favourable substrate pair (1.7 − 27.7 U/mgIRED) and high stereoselectivity, in several cases with a stereopreference opposite to that of fungal RedAms. This positions them as promising starting points for engineering towards a specific target reaction64,65. While several of these top-performing biocatalysts are literature-known from earlier screening campaigns18–20,22,58, others were experimentally characterised for the first time in the course of this work, highlighting both the remarkable proficiency of established IREDs and RedAms and the continued potential for discovery in this enzyme family. Correlating our broad-scope screening data with enzyme sequence information (RedAm key active-site residues reported in the literature)17,57 revealed a significant performance advantage of reductive aminases compared to “general” IREDs. However, our data also indicate that this difference is gradual rather than categorical, as the specific activity values observed for both groups across the investigated substrate range span several orders of magnitude and overlap considerably. This finding supports the notion that RedAms may catalyse imine formation only for a subset of the amine and carbonyl substrates they are able to process66. Additional sequence- and structure-based analyses also identified residues outside the active site that are linked to high reductive amination performance as well as three positions in the substrate-binding pocket that are robust predictors for (R)- vs. (S)-enantiopreference. Several of the reductive amination reactions investigated in this study were efficiently scaled up to gram scale. The desired amine products were isolated in good yields (63−89%) and with excellent enantiomeric purity (98% to >99% ee), demonstrating the preparative applicability of IRED-catalysed reductive amination at equimolar substrate concentrations. Therefore, we are confident that our results will actively encourage the wider use of imine reductases for biotransformations at low or no excess of amine, both in academic research and in industrial route development and production. Methods General methods and materials 1H and 13C NMR spectra were recorded on 300 and 500 MHz instruments (Bruker Avance III HD, Bruker Avance NEO 500). Chemical shifts are given in parts per million (ppm) relative to tetramethylsilane (δ = 0 ppm) and coupling constants (J) are reported in Hertz (Hz). Thin-layer chromatography (TLC) was used to follow the progress of synthetic reactions and to analyse fractions collected during preparative column chromatography. TLC was carried out on silica gel 60 F254 plates (Merck) and compounds were visualised either by dipping into cerium ammonium molybdate (CAM) reagent [100 g/L (NH4)6Mo7O24 · 4 H2O, 4 g/L Ce(SO4)2 · 4 H2O, in 10% aq. H2SO4], by dipping into basic permanganate reagent (10 g/L KMnO4, 50 g/L Na2CO3, 0.85 g/L NaOH, in H2O), by spraying with ninhydrin solution (2 g/L ninhydrin in EtOH), or by UV light. Preparative chromatography was carried out on silica gel 60, particle size 40–63 µm (230–400 mesh; Merck Millipore), either using traditional glassware or using an automated flash chromatograph (Biotage Selekt). Unit resolution GC–MS analyses were performed using electron impact (EI) ionisation at 70 eV and quadrupole mass selection. High-resolution MS analyses were performed using electrospray ionisation (ESI) and time-of-flight mass selection. Optical rotation values [α]D20 were measured on an electronic polarimeter (Anton Paar MCP 5100) at 589 nm (Na D-line) and 20 °C using a cuvette of 1 dm path length. Chemicals Unless otherwise noted, all reagents and biotransformation substrates were obtained from commercial suppliers in reagent grade quality and used without further purification. Aniline used for reference synthesis or as biotransformation substrate was purified by short-path distillation (Kugelrohr) before use. Product reference compounds were either obtained from commercial suppliers in reagent-grade quality (c1-a1, c1-a2, c1-a3, c1-a5, c1-a6, c1-a7, c2-a1, c2-a2, c2-a5, c2-a6, c2-a7, c2-a9, c2-a11, c2-a15, c2-a19, c2-a21, c2-a22, c3-a1, c3-a2, c6-a1, c6-a2, c7-a1, c11-a19, c13-a2, c16-a2, c20-a2, c27-a2, c28-a2) or synthesised as described in the Supplementary Information. Enzymes The imine reductases and phosphite dehydrogenase used in this study were produced by heterologous expression in E. coli as described in the Supplementary Information. Whole-cell preparations of transaminases used for the synthesis of non-racemic reference compounds were prepared as described in previous publications from our group67–69. Novozym® 435 (lipase B from Candida antarctica immobilised on acrylic resin; Novozymes A/S, Bagsvaerd, Denmark) was obtained from Pointner & Rothschädl GmbH (Baden, Austria). Biotransformations Reductive amination on analytical scale (general procedure) Stock solutions of the amine substrates (100 mM) in bicine–NaOH buffer (100 mM, pH 8.0) and of the carbonyl substrates (500 mM) in DMSO were prepared and stored as described in the Supplementary Information. A stock solution (2.5×) of NADP+ sodium salt (1.9 mg/mL, 2.5 mM), disodium hydrogen phosphite pentahydrate (54 mg/mL, 250 mM) and PtDH preparation (10 mg/mL) was freshly prepared for each series of biotransformations in the required quantity. IRED preparations (2.0 ± 0.2 mg) were weighed into microcentrifuge tubes (2.0 mL) and dissolved in NADP–phosphite–PtDH stock (200 µL), followed by addition of amine stock (250 µL) and carbonyl stock (50 µL) to give final concentrations of 50 mM carbonyl compound, 50 mM amine, 1 mM NADP+, 100 mM phosphite, 4 mg/mL lyophilised IRED preparation (cell-free extract), 4 mg/mL lyophilised PtDH preparation (cell-free extract), and 10% (v/v) DMSO in a total volume of 500 µL. Alternatively, individual stock solutions of NADP+ sodium salt (15.3 mg/mL, 20 mM), disodium hydrogen phosphite pentahydrate (216 mg/mL, 1.0 M), and PtDH preparation (40 mg/mL) could be prepared and used to achieve the same final concentrations and total reaction volume. The biotransformations were incubated at 30 °C and 120 rpm (horizontal placement of microcentrifuge tubes) in an incubator shaker or at 30 °C and 800 rpm (vertical placement of microcentrifuge tubes) in a benchtop thermoshaker for 24 h. Afterwards, the samples were subjected to one of the three work-up procedures described below. Work-up A (extraction for GC): The reaction mixtures were basified (pH ≥12) by addition of saturated aqueous Na2CO3 solution (200 µL) and extracted with EtOAc (2 × 500 µL, cont. 10 mM n-decane or n-dodecane as internal standard), whereby phase separation was accelerated by centrifugation (13,000 rpm, 16,250 × g, 1 min). The organic phases were combined in another microcentrifuge tube (1.5 mL), dried over Na2SO4, and centrifuged again to pellet the solid. The supernatant was transferred to a crimp-cap glass vial for GC–MS or GC–FID analysis. Work-up B (extraction and acetylation for GC): The reaction mixtures were basified (pH ≥ 12) by addition of saturated aqueous Na2CO3 solution (200 µL), and EtOAc (1.0 mL, cont. 10 mM n-decane or n-dodecane as internal standard) as well as acetic anhydride (10 µL) were added. The samples were shaken by hand for 1 min, followed by centrifugation (13,000 rpm, 16,250 × g, 1 min) to separate the liquid phases. The organic phase was transferred to another microcentrifuge tube (1.5 mL), dried over Na2SO4 and K2CO3, and centrifuged again to pellet the solids. The supernatant was transferred to a crimp-cap glass vial for GC–MS or GC–FID analysis. Work-up C (dilution and protein removal for HPLC–MS): Aliquots (100 µL) of the reaction mixtures were transferred to microcentrifuge tubes (1.5 mL) and diluted with HPLC-grade acetonitrile (900 µL). Precipitated protein was pelleted by centrifugation (13,000 rpm, 16,250 × g, 1 min), and the supernatant was filtered through a layer of cotton in a pipette tip and transferred to a crimp-cap glass vial for HPLC–MS analysis. Reductive amination on preparative scale (general procedure) A screw-cap Erlenmeyer flask (250 mL) was charged with a solution of the carbonyl substrate (10 mmol, 1 eq.; final concentration: 50 mM) in DMSO (20 mL) and a solution of the amine substrate (10–20 mmol, 1–2 eq.; final concentration: 50–100 mM) in bicine–NaOH buffer (180 mL; 100 mM, pH 8.0). Then, NADP+ sodium salt (153 mg, 0.2 mmol; final concentration: 1 mM), disodium hydrogen phosphite pentahydrate (4.32 g, 20 mmol, 2 eq.; final concentration: 100 mM), IRED preparation (0.8–1.6 g lyophilised cell lysate, final concentration 4–8 mg/mL), and PtDH preparation (0.8 g lyophilised cell lysate, final concentration 4 mg/mL) were added, and the flask was closed and placed in an incubator shaker at 30 °C and 120 rpm for 24 h. After this time, a sample (500 µL) was taken, worked up as described above (see Reductive amination on analytical scale, Work-up A), and analysed by GC–MS, indicating near-complete conversion in each case. The reaction mixture was basified (pH ≥12) by addition of half-saturated aqueous Na2CO3 solution (50 mL), transferred to polypropylene tubes (8 × 50 mL), and extracted with EtOAc (4 × 20 mL per tube), whereby phase separation was accelerated by centrifugation (4,500 rpm, 3,760 × g, 5 min). The combined organic phases were dried over Na2SO4, filtered, and the solvent was removed under reduced pressure to give the crude amine products, which were purified by column chromatography. (S)-N-(1-Phenylpropan-2-yl)cyclopropanamine, (S)-c8-a10 Biotransformation using phenylacetone (c8; 1.36 g, 10.1 mmol), cyclopropylamine (a10; 0.73 mL, 0.60 g, 10.5 mmol), and IR1855 (0.8 g), followed by the general work-up, gave 4.2 g of a yellow liquid. Column chromatography using silica gel 60 (100 g, 15 × 5 cm) and cyclohexane/EtOAc = 1:2 afforded the title compound (1.28 g, 7.3 mmol, 72%) as a pale-yellow liquid. TLC (silica gel 60, cyclohexane/EtOAc = 1:2, KMnO4 staining): Rf = 0.35. ee > 99% (HPLC). [α]D20 = +17.9 (c 1.00, CHCl3), +15.9 (c 1.11, MeOH); lit.70 [α]D26 = +16.1 (c = 0.56 mM, MeOH). 1H NMR (300 MHz, CDCl3): δ [ppm] = 7.33–7.25 (2H, m, Ar-H), 7.25–7.16 (3H, m, Ar-H), 3.04 (1H, sext, J = 6.5 Hz, CH2-CH-CH3), 2.80 (1H, dd, J = 13.3, 6.9 Hz, Ar-CH2-CH), 2.59 (1H, dd, J = 13.3, 6.7 Hz, Ar-CH2-CH), 2.06 (1H, tt, J = 6.6, 3.7 Hz, cyclopropyl CH), 1.82 (1H, br s, NH), 1.10 (3H, d, J = 6.3 Hz, CH3), 0.53–0.25 (4H, m, cyclopropyl CH2). 13C NMR (75 MHz, CDCl3): δ [ppm] = 139.7, 129.4, 128.5, 126.2, 55.6, 43.8, 28.9, 20.7, 7.3, 6.0. GC–MS (EI, 70 eV): m/z = 175 (M+, 2), 160 (6), 146 (7), 118 (15), 115 (5), 91 (38), 84 (100), 65 (8), 56 (6), 41 (15). HRMS calcd for C12H18N+ [M + H+]: 176.1434, found 176.1445. (R)-N-(Prop-2-yn-1-yl)−1,2,3,4-tetrahydronaphthalen-2-amine, (R)-c24-a4 Biotransformation using 2-tetralone (c24; 1.49 g, 10.2 mmol), propargylamine (a4; 0.67 mL, 0.58 g, 10.4 mmol), and IR2021 (0.8 g), followed by the general work-up, gave 3.6 g of a brown liquid. Column chromatography using silica gel 60 (70 g, 11 × 5 cm) and cyclohexane/EtOAc = 1:1 afforded the title compound (1.61 g, 8.7 mmol, 85%) as a dark-brown liquid. TLC (silica gel 60, cyclohexane/EtOAc = 1:1, KMnO4 staining): Rf = 0.38. ee 98% (HPLC). [α]D20 = +85.6 (c 1.03, CHCl3). 1H NMR (300 MHz, CDCl3): δ [ppm] = 7.16–7.02 (4H, m, Ar-H), 3.55 (2H, dd, J = 2.5, 0.8 Hz, NH-CH2-C ≡ CH), 3.28–3.13 (1H, m, NH-CH), 3.01 (1H, ddd, J = 16.0, 5.0, 1.6 Hz, CH2), 2.96–2.76 (2H, m, CH2), 2.60 (1H, dd, J = 16.0, 9.2 Hz, CH2), 2.21 (1H, td, J = 2.4, 0.8 Hz, C ≡ CH), 2.11–1.95 (1H, m, CH2), 1.72–1.53 (1H, m, CH2), 1.48 (1H, br s, NH). 13C NMR (75 MHz, CDCl3): δ [ppm] = 136.3, 135.0, 129.5, 128.8, 126.0, 125.8, 82.3, 71.5, 51.8, 36.3, 35.5, 29.2, 27.9. GC–MS (EI, 70 eV): m/z = 185 (M+, 28), 184 (96), 170 (73), 155 (23), 144 (42), 130 (100), 115 (68), 105 (24) 104 (85), 91 (43), 80 (88), 78 (37), 65 (14), 51 (12), 39 (29). HRMS calcd for C13H16N+ [M + H+]: 186.1277, found 186.1298. (R)-N-Methyl-1,2,3,4-tetrahydronaphthalen-2-amine, (R)-c24-a2 Biotransformation using 2-tetralone (c24; 1.49 g, 10.2 mmol), methylamine (a2; 1.73 mL of a 40 wt-% solution in water, 20.0 mmol, 2 eq.), and IR0035 (1.6 g), followed by the general work-up, gave 4.4 g of a brown liquid. Column chromatography using silica gel 60 (100 g, 15 × 5 cm) and MTBE/MeOH/NH4OH = 90:9:1 afforded product that was still contaminated with residual DMSO. This material was taken up in MTBE (50 mL), washed with a 10% (w/v) aq. LiCl solution (20 mL), and the organic phase was dried over Na2SO4 and concentrated under reduced pressure to give the title compound (1.46 g, 9.1 mmol, 89%) as a brown liquid. TLC (silica gel 60, MeOH, KMnO4 staining): Rf = 0.17. ee 99% (HPLC). [α]D20 = +51.5 (c 1.18, CHCl3). 1H NMR (300 MHz, CDCl3): δ [ppm] = 7.15–7.02 (4H, m, Ar-H), 3.03 (1H, ddd, J = 15.8, 4.9, 1.7 Hz, CH2), 2.97–2.73 (3H, m, CH2 and NH overlap), 2.60 (1H, (ddd, J = 16.0, 9.3, 1.2 Hz, CH2), 2.52 (3H, s, CH3), 2.15–2.00 (1H, m, CH2), 1.69–1.50 (2H, m, CH2). 13C NMR (75 MHz, CDCl3): δ [ppm] = 136.3, 135.3, 129.5, 128.8, 125.9, 125.8, 55.4, 36.4, 33.9, 29.2, 28.0. GC–MS (EI, 70 eV): m/z = 161 (M+, 100), 146 (29), 130 (56), 115 (21), 104 (36), 91 (11), 78 (13), 70 (11), 57 (10), 42 (9). HRMS calcd for C11H16N+ [M + H+]: 162.1277, found 162.1286. (R)-Ethyl 3-(prop-2-yn-1-ylamino)butanoate, (R)-c22-a4 Biotransformation using ethyl acetoacetate (c22; 1.31 g, 10.1 mmol), propargylamine (a4; 0.67 mL, 0.58 g, 10.4 mmol), and IR0068 (1.6 g), followed by the general work-up, gave 3.3 g of a yellow liquid. Column chromatography using silica gel 60 (100 g, 15 × 5 cm) and cyclohexane/EtOAc = 1:1 afforded the title compound (1.28 g, 7.6 mmol, 75%) as a pale-yellow liquid. TLC (silica gel 60, cyclohexane/EtOAc = 1:1, KMnO4 staining): Rf = 0.33. ee > 99% (GC). [α]D20 = –2.2 (c 1.93, CHCl3). 1H NMR (300 MHz, CDCl3): δ [ppm] = 4.14 (2H, q, J = 7.2 Hz, CH2-CH3), 3.53 (1H, dd, J = 17.1, 2.5 Hz, NH-CH2-C ≡ CH), 3.45 (1H, dd, J = 17.1, 2.5 Hz, NH-CH2-C ≡ CH), 3.37 (sext, J = 6.4 Hz, NH-CH), 2.94 (1H, br s, NH), 2.52 (1H, dd, J = 15.6, 6.8 Hz, CH2-CO2Et), 2.40 (1H, dd, J = 15.6, 6.0 Hz, CH2-CO2Et), 2.24 (1H, t, J = 2.5 Hz, C ≡ CH), 1.25 (3H, t, J = 7.1 Hz, CH2-CH3), 1.16 (3H, d, J = 6.4 Hz, 3H, CH-CH3). 13C NMR (75 MHz, CDCl3): δ [ppm] = 172.0, 81.0, 72.3, 60.7, 48.9, 41.1, 35.5, 19.6, 14.3. GC–MS (EI, 70 eV): m/z = 169 (M+, < 1), 154 (11), 130 (11), 108 (12), 96 (9), 82 (100), 66 (8), 54 (16), 42 (11). HRMS calcd for C9H16NO2+ [M + H+]: 170.1176, found 170.1175. (S)-1-Cyclopropyl-5-methylpyrrolidin-2-one, (S)-c10-a10 Biotransformation using ethyl levulinate (c10; 1.46 g, 10.1 mmol), cyclopropylamine (a10; 1.46 mL, 1.20 g, 21.1 mmol), and IR1855 (0.8 g), followed by the general work-up, afforded a mixture of the desired lactam along with the open-chain amino ester in a ratio of approx. 10:90. This material was taken up in MTBE (100 mL), washed with a 10% (w/v) aq. LiCl solution (20 mL) to remove residual DMSO, and the organic phase was dried over Na2SO4 and concentrated under reduced pressure to give 2.1 g of a yellow liquid. The residue was dissolved in cyclohexane (100 mL) and the solution transferred to a screw-cap Erlenmeyer flask (250 mL), Novozym® 435 (100 mg) was added, and the flask was closed and placed in an incubator shaker at 30 °C and 120 rpm for 16 h. Afterwards, the immobilised enzyme was removed by filtration through a glass frit and the filtrate was concentrated under reduced pressure to give the crude lactam (1.2 g) as a yellow liquid. Column chromatography using silica gel 60 (70 g, 10 × 5 cm) and EtOAc/MeOH = 5:1 afforded the title compound (0.89 g, 6.4 mmol, 63%) as a yellow liquid. TLC (silica gel 60, EtOAc/MeOH = 5:1, KMnO4 staining): Rf = 0.55. ee > 99% (GC). [α]D20 = –90.4 (c 1.07, CHCl3). 1H NMR (300 MHz, CDCl3): δ [ppm] = 3.57 (1H, dqd, J = 7.7, 6.3, 4.6 Hz, CH-CH3), 2.50–2.23 (3H, m, CH2), 2.09 (1H, dddd, J = 12.7, 9.6, 7.8, 7.1 Hz, CH2), 1.57 (1H, dddd, J = 12.7, 9.5, 6.0, 4.5 Hz, cyclopropyl CH), 1.24 (3H, d, J = 6.3 Hz, CH3), 0.93 (1H, dtd, J = 9.7, 7.0, 5.4 Hz, cyclopropyl CH2), 0.79 (1H, dddd, J = 10.4, 6.7, 5.4, 3.9 Hz, cyclopropyl CH2), 0.65 (1H, dtd, J = 9.8, 7.0, 5.2 Hz, cyclopropyl CH2), 0.51 (1H, dddd, J = 10.3, 7.0, 5.2, 4.0 Hz, cyclopropyl CH2). 13C NMR (75 MHz, CDCl3): δ [ppm] = 176.1, 55.2, 30.8, 26.4, 23.2, 20.1, 7.5, 4.3. GC–MS (EI, 70 eV): m/z = 139 (M+, 53), 124 (100), 96 (14), 84 (36), 68 (17), 57 (5), 55 (20), 41 (25), 39 (12). HRMS calcd for C8H14NO+ [M + H+]: 140.1070, found 140.1072. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Reporting Summary Transparent Peer Review file Source data Source Data Acknowledgements The authors thank Markus Köck (University of Graz) for HRMS measurements, Bernd Werner (University of Graz) for recording NMR spectra, Lisa M. Berger (University of Graz) for help with data analysis, and Verena Resch (luminous-lab.com) for support in designing Fig. 1. Author contributions Conception of the project: J.H.S., D.W., H.I., C.C.G. and G.S. Acquisition of funding: J.H.S., D.W., H.I., C.C.G., G.S. and W.K. Experiments: S.A.B. (enzyme expression, screening, method development on robotic platform, synthesis of reference compounds, biotransformations on analytical scale, biotransformations on preparative scale, development of analytical methods); C.G. (enzyme expression, screening, method development on robotic platform, synthesis of reference compounds, biotransformations on analytical scale, development of analytical methods); I.O.-G. (method development on robotic platform); J.H.S. (synthesis of reference compounds, biotransformations on analytical scale, development of analytical methods); S.V., L.R., M.W., F.M.K. and Y.O. (enzyme expression, synthesis of reference compounds, biotransformations on analytical scale). Computation: S.A.B. (script-based data processing); M.C. (bioinformatic analysis and selection of enzyme panel, protein–ligand docking, sequence conservation analysis); B.M.N. (bioinformatic analysis and selection of enzyme panel; I.M. (chemoinformatic analysis of substrate panel); A.K. and T.S. (AI analysis of screening results); J.H.S. (explorative data analysis and z-score transformation, stereoselectivity key residue analysis). Supervision of the project: J.H.S., B.M.N., S.B., H.I., F.G., K.P. and D.W. Preparation of manuscript: S.A.B., J.H.S., I.M., A.K. and M.C. (first draft, preparation of tables and figures); S.A.B., S.V., C.G., B.M.N., S.B., H.I., J.H.S., W.K., F.G. and K.P. (editing and proof-reading) Peer review Peer review information Nature Communications thanks Yu-Fei Ao, Godwin Aleku and the other anonymous reviewer for their contribution to the peer review of this work. A peer review file is available. Funding This work was funded by F. Hoffmann–La Roche Ltd. (Basel, Switzerland) and Genentech Inc. (San Francisco, USA) via the companies’ Technology Innovation & Science programme. Financial support by the University of Graz and the Field of Excellence ‘BioHealth’ is gratefully acknowledged. Data availability The data generated in this study (consolidated screening data, NMR spectra, protein and DNA sequences, chemoinformatic descriptors of the substrates) are openly available in the Mendeley Data repository (https://data.mendeley.com/) at https://doi.org/10.17632/jn4sfsrz8x.2.71 Source data for Figs. 2, 3, S7, S9, S14, S15, S16, S17, S18, S22, S24, and S25 are provided with the paper. All other figures can be directly reproduced from the deposited data set. Code availability The Python scripts used for data collection and processing are available on GitHub (https://github.com/saberger/Screening_for_Generality) and have also been archived in the Zenodo data repository (https://zenodo.org/) at https://doi.org/10.5281/zenodo.21285045.72 Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Sarah A. Berger, Christopher Grimm. Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-76869-7. References Associated Data Supplementary Materials Supplementary Information Reporting Summary Transparent Peer Review file Source Data Data Availability Statement The data generated in this study (consolidated screening data, NMR spectra, protein and DNA sequences, chemoinformatic descriptors of the substrates) are openly available in the Mendeley Data repository (https://data.mendeley.com/) at https://doi.org/10.17632/jn4sfsrz8x.2.71 Source data for Figs. 2, 3, S7, S9, S14, S15, S16, S17, S18, S22, S24, and S25 are provided with the paper. All other figures can be directly reproduced from the deposited data set. The Python scripts used for data collection and processing are available on GitHub (https://github.com/saberger/Screening_for_Generality) and have also been archived in the Zenodo data repository (https://zenodo.org/) at https://doi.org/10.5281/zenodo.21285045.72