Underexplored Catalysts as General Structures: Application of Machine Learning Techniques for Reaction‐Specific Datasets

ABSTRACT

General catalysts are usually identified through broad experimental screening to find structures that perform reliably across many substrates and reaction classes. In secondary amine organocatalysis, historical reporting is strongly skewed toward a small set of standard catalysts, leaving many plausible scaffolds underexplored and difficult to evaluate objectively. Here, we apply a bias‐aware machine learning workflow designed for small, uneven datasets to prioritize candidate general catalysts from limited historical data. Within the iminium‐based reaction space used to construct the curated and virtually balanced dataset, this analysis surfaced several high‐performing candidates, including a rarely studied imidazolidinone bearing a benzyl‐protected indole substituent. Despite minimal precedent, this scaffold performed competitively in experimental benchmarking and external transferability tests. In a retrospective analysis restricted to pre‐2005 examples, the same workflow prioritized catalyst families that later became widely adopted (e.g., diarylprolinol silyl ethers and imidazolidinones) among its top candidates, consistent with earlier prioritization from the literature available at the time. Together, these results show how bias‐aware modeling can highlight overlooked scaffolds and reduce the experimental burden required to identify broadly useful catalysts. Pairing targeted experiments with data‐driven prioritization provides a practical route to expanding the set of reliable secondary‐amine catalysts beyond the structures that dominate current practice.

Bias‐aware machine learning identified an underexplored imidazolidinone catalyst with broad competitive performance in iminium‐based reactions. Experimental benchmarking shows how data‐driven prioritization can uncover overlooked catalyst scaffolds from sparse and historically biased literature data.

1. Introduction

Reaction‐specific optimization of asymmetric organocatalysis has been instrumental in developing small organic catalysts that deliver precise stereochemical control [1]. A key advantage of many organocatalyst classes is that their structures can be tuned iteratively to improve performance for a specific transformation, and trial‐and‐error optimization can often converge on high selectivity under a defined set of conditions. While highly effective, this approach naturally favors catalysts optimized for narrow substrate classes or tightly defined reaction parameters. In parallel, there exists a distinct category of broadly effective, “privileged” [2] catalysts that may not be uniquely optimal for any single transformation, but perform reliably across diverse reactions and substrates. These general catalysts serve an important role in synthetic chemistry. Specifically, they enable rapid reaction development, substrate scouting, and method exploration when mechanistic understanding is incomplete or when time and material are limited. At the same time, highly specialized catalysts remain essential for targeted synthesis and industrial deployment. Together, specialized and general catalysts address complementary needs in chemical synthesis.

Motivated by both the maturity of the field and the practical limitations of exhaustive empirical screening [3, 4, 5, 6, 7, 8, 9], we initiated a program to retrospectively study catalyst generality in asymmetric organocatalysis using data science and machine learning (ML) tools [10, 11, 12]. This approach allows us to systematically analyze data from specific reaction or substrate classes, identify key structural features important for broad enantioselectivity outcomes, and apply statistical models to assign catalyst generality values. Importantly, ML can be useful even when historical data are sparse and unevenly distributed, provided that bias and uncertainty are handled explicitly (Figure 1A). In our recent work, we applied this approach to asymmetric nucleophilic additions to imines catalyzed by BINOL‐derived chiral phosphoric acids [10]. In that setting, strong structural priors exist (for example, a rigid C₂‐symmetric scaffold with modular 3,3′ substitution) [13, 14, 15, 16], and the central question was largely which substituent patterns generalize best. A more stringent and widely relevant question is whether similar workflows remain informative when historical data are extremely biased and only a small number of catalysts dominate the literature.

FIGURE 1.

Conceptual framework for identifying general catalysts from sparse, biased data. (A) Bias‐aware ML models are used to prioritize catalysts and predict enantioselectivity for untested catalyst–reaction combinations. (B) Schematic adoption curves illustrate how catalysts differ in historical usage, reflecting both practical constraints and performance. The bottom panel shows the six imidazolidinone catalysts (A–F) evaluated in this study.

Secondary amine organocatalysts provide a prototypical test for this problem. These catalysts are historically significant and have been studied for decades [17, 18], yet practical use remains concentrated in a small number of canonical structures. This concentration likely reflects a combination of early visibility, synthetic accessibility, commercial availability, and genuine performance, but it also means that many plausible scaffolds have been evaluated far less frequently than their structural potential might suggest. Here, we ask whether bias‐aware data‐driven modeling can help prioritize underexplored secondary amine scaffolds for generality testing using only the information already present in the literature, while explicitly accounting for uneven historical coverage (Figure 1B).

Rather than treating historical adoption as a proxy for quality, we explicitly test whether underused catalysts can deliver performance comparable to widely used benchmarks when evaluated across multiple reaction classes. Our working hypothesis is that data‐driven prioritization can identify overlooked catalysts with broad performance, and that historical adoption reflects not only intrinsic selectivity but also practical constraints that shape which catalysts are repeatedly deployed. The results outlined herein demonstrate how data‐driven prioritization can surface underexplored catalyst scaffolds and focus experimental effort on the most informative tests, rather than relying solely on historical popularity. We envision a continuous workflow in which models are updated as new reactions are reported, enabling progressively improved generality estimates across reaction space. Even in the presence of unavoidable sparsity and reporting bias, existing literature data can still support actionable catalyst shortlists when paired with targeted experimental validation. In this study, we further ask whether such prioritization remains stable across reasonable choices in clustering, success threshold, and related workflow assumptions, rather than arising from a single tuned analysis setting.

2. Results and Discussion

2.1. Catalyst Chemotype Selection

This study centers on secondary amines, an early and enduring class of organocatalysts whose practical use remains dominated by a small number of well‐established structures [19]. To evaluate whether ML can identify overlooked but general secondary amine catalysts, we constructed a curated dataset capturing both their historical use and performance across diverse transformations. In more detail, we analyzed 24 representative secondary amines based on their use as either reagents or catalysts across a broad range of reaction types, including both polar and single‐electron activation mechanisms, as classified in SciFinder (November 15, 2024). This dataset tracks catalyst deployment from the date of first publication and highlights temporal trends in usage. A summary of these findings, with reaction frequency represented as bubble sizes, is shown in Figure 2A.

FIGURE 2.

Catalyst deployment in the literature and composition of training and virtual sets. (A) Structural comparison of proline‐, diarylprolinol‐, and imidazolidinone‐based catalysts, highlighting underrepresented scaffolds. (B) Distribution of secondary amine usage across the literature and the training dataset. (C) UMAP projection comparing virtual and training reaction distributions.

Our analysis reveals a striking concentration of usage within a small subset of catalysts. Proline emerges as the dominant structure, accounting for 41% of all reactions, reflecting both its early adoption in the 1970s [20, 21] and the decades of follow‐up study that preceded the modern surge in organocatalysis research [18]. The Hayashi–Jørgensen diarylprolinol silyl ethers, which leverage steric‐blocking motifs to induce selectivity, represent the second most prevalent group [22, 23, 24]. Within this class, derivatives bearing either phenyl or 3,5‐bis(trifluoromethyl)aryl substituents collectively account for an additional 48% of reactions. The MacMillan imidazolidinone catalysts [25], characterized by a benzyl arm and steric features such as gem‐dimethyl or tert‐butyl groups, comprise only 5% of the total reactions analyzed, despite being published earlier than the Hayashi‐Jørgensen catalysts. Remarkably, only six catalysts collectively account for over 95% of the 40 602 reactions analyzed, leaving many other secondary amine structures comparatively underutilized. This skewed usage motivates a central question addressed in this work: can bias‐aware modeling help identify underexplored secondary‐amine scaffolds that perform broadly well when evaluated head‐to‐head against the dominant catalysts?

2.2. Generality Analysis

To evaluate the generality of underexplored catalyst structures, we utilized a previously published database on secondary amine organocatalysis [26, 27]. This dataset comprises 452 reactions curated from diverse sources, including peer‐reviewed literature and high‐quality but unpublished PhD theses. The latter often contain rigorously characterized data from reactions that failed to deliver high enantioselectivity (i.e., low ee), making them especially valuable for model development. The dataset spans a wide range of reaction components, including catalyst structures such as diphenylprolinol ethers, imidazolidinones, and hydrogen‐bonding motifs like proline. The carbonyl electrophiles are exceptionally diverse, covering aldehydes, enals, branched enals, ketones, and both acyclic and cyclic enones. Nucleophiles also vary substantially, ranging from electron‐rich aromatics to Hantzsch esters and dienes. Additional variation is found in solvent systems and amine salt additives, further enhancing the dataset's breadth and relevance for assessing catalyst generality. While the dataset of 452 reactions represents a small fraction of the total reported reactions shown in Figure 2A, it is important to note that the distribution of catalysts is largely mirrored between the literature and the training sets, enabling accurate comparisons (Figure 2B).

Because both the literature and curated datasets are strongly imbalanced across catalyst types, direct generality comparisons can be confounded by uneven sampling. More broadly, reaction prediction models can inadvertently reproduce reporting bias unless corrective strategies are applied [28, 29]. We therefore combine bias‐aware modeling with model‐guided balancing by imputing ΔΔG^‡ for unmeasured but mechanistically valid catalyst–substrate pairs. This enables generality estimates to better reflect performance consistency across a common substrate space rather than historical reporting frequency.

The first part of this corrective effort concentrated on determining a training set that would ensure meaningful comparisons within a mechanistically unified, well‐studied space. Thus, we focused on reactions involving α,β‐unsaturated carbonyl compounds activated through iminium formation [30, 31]. We further narrowed the scope to catalysts that induce enantioselectivity primarily through steric repulsion rather than attractive hydrogen‐bonding control. This constraint ensures that catalysts under evaluation operate through similar activation modes and stereocontrol features, enabling confident assessment of relative generality within this chemical space. After applying these mechanistic filters, the focused dataset comprised 195 reactions (see Supporting Information, p. 7).

Next, regression models were trained on the curated data to impute enantioselectivity outcomes (reported as ΔΔG ^‡) for unmeasured catalyst–substrate combinations. Critically, imputation was restricted to combinations expected to form stereocenters (e.g., predictions on Hantzsch esters paired with unbranched enals were not performed), so the virtual expansion respected the same outcome constraints used in model development. As shown in Figure 2C, the curated training subset (purple) and the corresponding imputed virtual set (white) occupy similar regions of chemical space, with virtual combinations filling gaps between experimentally reported points. This alignment supports the use of the virtual set as a tractable proxy for balanced generality scoring under a shared substrate landscape.

We then applied our previously established generality workflow to explore the secondary amine catalyst space. Generality was assessed by first mapping the reaction landscape using nonlinear dimensionality reduction (uniform manifold approximation and projection, UMAP) [32] and clustering reactions via k‐means clustering. Reactions were described based on nucleophile and electrophile properties, with UMAP reducing dimensionality to group similar reactions spatially. These groupings were then labeled using k‐means clustering, with the optimal number of clusters determined via elbow and silhouette methods. Catalyst generality was quantified as the proportion of clusters where a given catalyst demonstrated performance above a user‐defined threshold, optimized to maximize variance between catalyst performances (Figure 3A). For a detailed description of this workflow, see ref [10]. This metric is intended to capture reliability across diverse substrate clusters, rather than identify the single most selective catalyst for any one reaction. To discriminate among top‐performing catalysts, generality rankings in Figure 3B were computed using a stringent success cutoff (≥ 95% ee); at lower cutoffs (e.g., 90% ee) many catalysts converge to similar scores and the rankings become less informative (see Supporting Information: Figures S11–S13). Importantly, the prioritization of catalyst B was stable across reasonable variations in the workflow: when the analysis was repeated over k = 20–40 clusters and success thresholds of 90%–98% ee, catalyst B remained top‐ranked, and the resulting rank orderings were highly correlated across most settings (see Supporting Information, p. 19–21). Taken together with the sensitivity analysis, imputation perturbations, and the additional bootstrap and leave‐one‐out diagnostics reported in the Supporting Information, these results indicate that catalyst B is robustly prioritized within the iminium‐based reaction space studied and that the k  =  30, 95% ee setting used in the main text is a practical operating point within a broader stable regime rather than a uniquely tuned choice.

FIGURE 3.

Statistical modeling workflow for general catalyst identification and secondary analysis of catalyst generality in iminium reactions. (A) Schematic of the data‐driven workflow used to identify general catalysts, consisting of: (1) clustering the reaction space, (2) evaluating catalyst performance within each cluster, and (3) quantifying generality as the average success across clusters. (B) Application to secondary amine catalysts for iminium‐based reactions. Generality scores are shown below each structure, with larger circles indicating higher generality. A benzyl‐protected indole‐substituted imidazolidinone was identified as the top‐ranked catalyst within the iminium‐based dataset despite its rarity in the literature and training data.

Applying this workflow to the secondary‐amine catalyst space yielded generality rankings consistent with expectations for widely used catalyst families, with many top candidates belonging to Hayashi–Jørgensen diarylprolinol silyl ethers. However, the highest‐ranked catalyst in our analysis was a rarely explored imidazolidinone bearing a benzyl‐protected indole substituent (catalyst B, Figures 1B and 3B). While this structure retains elements of the MacMillan‐type imidazolidinone architecture, it is rarely used and is not commercially available, consistent with its limited historical visibility (two studies [33, 34] and 26 reactions, representing under 0.1% of the literature data analyzed). Under the ≥ 95% ee criterion, catalyst B ranks above the canonical MacMillan catalyst A (benzyl, t‐Bu) (G = 0.90 vs. 0.00), and the separation decreases as the threshold is relaxed (90% ee: G = 1.00 vs. 0.43). Because underrepresented catalysts may lack sufficient counterexamples in the training data, absolute generality values for rare scaffolds can be inflated if sparse appearances occur primarily in favorable contexts. We therefore interpret these values as prioritization‐oriented relative rankings, and we experimentally benchmark catalyst B directly against imidazolidinone catalyst A. Supporting this interpretation, repeated out‐of‐fold and bootstrap analyses did not indicate that catalyst B was associated with unusually optimistic residuals or inflated predictive uncertainty relative to the represented catalysts in the filtered iminium dataset; the strongest sparsity effects were instead observed for singleton or zero‐count cases (see Supporting Information, p. 26–29). At the same time, imputation‐sensitivity analysis shows that the margin of the absolute top‐rank assignment is finite: under a deliberately conservative catalyst‐B‐specific penalty, catalyst B remained uniquely top‐ranked after a 1 percentage‐point downward shift, remained ahead of catalyst A after a 2 percentage‐point shift, and became tied with catalyst A at larger penalties, while remaining rank 1 under noise, shrinkage, and borderline‐exclusion perturbations (see Supporting Information: Figure S16). We also expand this analysis to additional catalyst structures C–F across multiple reaction classes to test whether the workflow reliably distinguishes broadly effective from non‐general structures (vide infra; Figure 1B).

2.3. Reaction Selection

To evaluate whether catalyst B could be considered general, we directly compared it to the leading imidazolidinone catalyst A under standardized experimental conditions (Figure 1B). This head‐to‐head design ensures that any performance differences between the imidazolidinone catalysts were attributable to structural features rather than differences in reaction setup. All reactions were conducted at room temperature or 4°C using a shared solvent system and Brønsted acid co‐catalyst compatible with both scaffolds. These conditions were selected to be broadly competent for both scaffolds rather than individually optimized for either catalyst, since our goal was to compare trend‐level generality rather than to maximize ee on a case‐by‐case basis. Accordingly, we treat these experiments as a direct test of robustness under standardized conditions—that is, whether the relative trends predicted by the generality analysis persist without bespoke optimization. We emphasize that this design does not constitute a comprehensive optimization study across all secondary‐amine‐catalyzed manifolds; rather, it provides a controlled comparison within a shared operational window chosen to make direct A/B benchmarking possible. Moreover, this approach allowed for a streamlined comparison without relying on specialized equipment or extended reaction times at very low temperatures. Enantiomeric excesses were quantified by Supercritical Fluid Chromatography (SFC) equipped with a Photodiode Array (PDA) detector, using isolated compounds obtained via column chromatography.

To ensure that all products could be detected, cinnamaldehyde was chosen as the base aldehyde, as it contains a chromophore regardless of the nucleophile structure. However, since aliphatic aldehydes are also effective substrates for these reactions, we included them where possible and specifically in cases where the nucleophile contained a chromophore. This strategy enabled a broad assessment of substrate diversity. The only challenge encountered was with Diels–Alder reactions, which we overcame by performing additional structural modifications post‐reaction (see Supporting Information).

To streamline the experimental evaluation of catalyst generality, we implemented a two‐layered workflow shown in Figure 4. The first layer focuses on benchmark reactions involving α,β‐unsaturated aldehydes and nucleophiles for which catalyst A is commonly reported to deliver high enantioselectivity. These reactions provide a high‐confidence foundation for assessing whether the underexplored catalyst B can match benchmark performance. Notably, catalyst B had not been studied in any of these transformations, underscoring its potential as a latent general structure. Reactions known to pose intrinsic challenges for catalyst A, such as enones where it consistently underperforms or fails to induce enantioselectivity, were excluded to avoid confounding comparisons [35, 36]. In addition, due to differences in reactivity and distinct requirements for other reaction parameters, reactions that were reported only with Hayashi‐Jørgensen diarylprolinol silyl ethers were not selected for evaluation [37, 38].

FIGURE 4.

A workflow for evaluating predicted general catalysts. This structure‐guided expansion enabled the prioritization of reactions for experimental validation in tested and untested regions of chemical space.

This design deliberately decouples performance from historical frequency by combining benchmark reactions (where A is established) with model‐suggested pairs that probe underexplored regions of the substrate landscape. Although virtual screening has been widely used to guide reaction discovery, including by our own group, it typically stops at ranking or prioritizing structures, without testing enumerated catalyst–substrate combinations in the lab. Thus, in the second phase of our study, we deliberately focused experimental efforts on virtually predicted catalyst–substrate pairs. Rather than limiting Schemes 1, 2, 3, 4 to a few illustrative examples, we chose to present a broader range of results to better capture the scope of our analysis. This includes both nucleophile variation and virtual pair testing, presented together to capture both substrate‐driven trends and model‐guided pair testing within a single, coherent evaluation. The remainder of the experimental outcomes are summarized in Figure 5. Moreover, if successful, this approach serves as a proof of concept for validating ML workflows beyond in silico predictions and highlights the potential of closer integration between computational prioritization and experimental execution. In this context, the principal value of the ML workflow is not to replace chemical intuition, but to reduce popularity bias, compare catalysts across a shared model‐balanced substrate landscape, and prioritize the most informative validation experiments among many plausible but underexplored scaffolds. This added value is also supported by comparison with simpler baselines (see Supporting Information, p. 30–32). These include popularity‐based, observed‐mean, penalized‐mean, and observed‐only clustered analyses, which are each confounded by historical coverage or sparse sampling in different ways, whereas the model‐derived workflow evaluates catalysts across a shared virtual reaction landscape and retains catalyst B as the top‐ranked candidate at high selectivity.

SCHEME 1.

Various Friedel‐Crafts alkylation of indoles tested with MacMillan catalysts A and B. Reaction conditions: indole (0.2 mmol), aldehyde (0.6 mmol), catalyst (20 mol%), CF₃COOH(TFA) (20 mol%), DCM/isopropanol (85/15, 0.5 mL), r.t. for specified time. Isolated yields given for the corresponding alcohol after reduction with excess NaBH₄ in methanol. Enantiomeric excess (ee) was measured by chiral SFC. Stereochemistry is assigned in analogy to reported examples [39].

SCHEME 2.

Various Friedel‐Crafts alkylation of pyrroles screened against MacMillan catalysts A and B. Reaction conditions: pyrrole (1 mmol), aldehyde (0.2 mmol), catalyst (20 mol%), TFA (20 mol%), THF/H₂O (20/3, 0.5 mL). Isolated yields given for the corresponding alcohol after reduction with excess NaBH₄ in methanol. Enantiomeric excess (ee) was measured by chiral SFC. Stereochemistry is assigned in analogy to the reported examples [40].

SCHEME 3.

A Diels‐Alder reaction examined with MacMillan catalysts A and B. Reaction conditions: cyclopentadiene (0.6 mmol), aldehyde (0.2 mmol), catalyst (10 mol%), HCl (10 mol%), MeOH/H₂O (95/5, 0.2 mL), r.t., 16 h. Isolated yields. Diastereomeric ratios (exo/endo) were determined by ¹H NMR and chiral SFC. Enantiomeric excess (ee) was measured by chiral SFC. Opposite ee values indicate the different configurations produced as the major product by the two catalysts. Stereochemistry is assigned in analogy to reported examples [17].

SCHEME 4.

Various Friedel–Crafts alkylation of anilines tested with MacMillan catalysts A and B. Reaction conditions: Aniline (0.2 mmol), aldehyde (0.6 mmol), catalyst (20 mol%), HCl (20 mol%), DCM (0.2 mL), r.t., 12 h. Isolated yields given for the corresponding alcohol. Enantiomeric excess (ee) was measured by chiral SFC. Stereochemistry is assigned in analogy to reported examples [41]. ^aUse Aniline (0.6 mmol), aldehyde (0.2 mmol) and stir at r.t. for 48 h instead. ^bYields refer to the aldehyde and enantioselectivities were determined using the corresponding alcohol.

FIGURE 5.

Head‐to‐head comparison of measured enantioselectivity (ΔΔG ^‡) between imidazolidinone catalysts A and B across a range of substrate pairs. Negative ΔΔG ^‡ differences (blue, below zero) indicate reactions where catalyst B is measured to outperform catalyst A, while positive values (red, above zero) indicate reactions where catalyst A is favored. (A) Relative differences. (B) Absolute values displayed.

As we transition to the experimental results, it is important to define what constitutes a meaningful outcome. Put simply, if catalyst B is truly as general as catalyst A, we would expect it to exhibit similar enantioselectivity trends across a diverse range of substrates. That is, when catalyst A delivers high selectivity, catalyst B should also perform well; when A performs poorly, a corresponding drop in B's performance is acceptable. This type of trend‐level correlation, rather than outright superiority, offers compelling evidence for generality, highlighting consistent behavior across mechanistically distinct transformations.

2.4. Experimental Validation of General Catalyst B

The Friedel‐Crafts alkylation of indoles was first examined in this study (Scheme 1) [39]. Overall, catalyst B provided a marginal improvement under standardized reaction conditions, aligning well with the generality analysis. For cinnamaldehyde derivatives, this transformation showed slightly better enantioselectivity with catalyst B compared to catalyst A. This corresponds to an average improvement of 0.23 kcal/mol at room temperature. The differences are less obvious for alkyl‐substituted α,β‐unsaturated aldehydes, but the trend is maintained.

These promising results prompted further evaluation using other nucleophiles within the same reaction class. In the Friedel–Crafts alkylation of pyrroles [40], we observed a range of enantioselectivities spanning 0.03–0.15 kcal/mol in ΔΔG ^‡ across different substitutions. Overall, catalyst B again exhibited improved selectivity compared to A.

The Diels–Alder reaction [17] was also tested, where the tryptophan‐derived catalyst B performed marginally better than the phenylalanine‐derived counterpart (ΔΔG ^‡ difference = 0.42 kcal/mol). A slightly higher exo/endo ratio was observed compared to the optimal results, while the endo isomer was obtained with diminished enantioselectivity. This deviation indicates a different steric or electronic influence at play and highlights the importance of substrate‐catalyst interactions in fine‐tuning selectivity beyond general trends.

From this test set, the only transformation where catalyst A outperformed B was the Friedel–Crafts alkylation of anilines (Scheme 4) [41]. Across a panel of nucleophiles bearing varying electronic properties, catalyst B consistently gave slightly lower enantioselectivity, although the differences were minor. This suggests that the electronic matching between nucleophile and catalyst scaffold may be a limiting factor in this subclass.

Finally, we evaluated the hydrogenation of branched enals, a reaction class known to be difficult for secondary amine catalysts [42]. Literature‐reported optimal catalysts for this transformation differ substantially from those used in our study, often bearing minimal steric bulk and lacking aromatic substituents. Consistent with this, the tryptophan‐substituted catalyst B, with its extended aromatic surface, exhibited lower enantioselectivity than the benzyl‐substituted variant (Scheme 5). This trend was consistent across additional branched substrates, suggesting that the extended aromatic substituent perturbs the relative energies of competing transition states in this subclass. Although we do not assign a specific interaction model here. Given the limited performance and clear mechanistic constraints, we deprioritized further investigation of this transformation.

SCHEME 5.

Various transfer hydrogenations of branched enals evaluated with MacMillan catalysts A and B. Reaction conditions: Aldehyde (0.2 mmol), Hantzsch ester (0.24 mmol), catalyst (20 mol%), TFA (20 mol%), toluene (1 mL), 4°C for 2 h. Isolated yields given for the corresponding alcohol. Enantiomeric excess (ee) was measured by chiral SFC. Stereochemistry is assigned in analogy to the reported examples [42].

Our analysis showed that reactions selected through initial screening were reasonably well‐distributed across the mapped space. To further refine the experimental set, we prioritized diversity in electrophile structures and ensured balanced representation across the five major reaction classes. This enabled a more systematic exploration of key structural and mechanistic trends. A key advantage of integrating chemical space analysis into reaction selection is the ability to navigate the reaction landscape directly—using the virtual matrix to identify unexplored substrate pairs and guide testing of reactions absent from historical literature.

The resulting head‐to‐head ΔΔG ^‡ profiles (Figure 5A,B) provide a compact view of generality across structurally varied reactions. Across the full panel, catalyst B matches or exceeds catalyst A in a majority of cases (negative ΔΔG ^‡ differences), consistent with its high generality ranking and supporting the conclusion that this underused scaffold can deliver benchmark‐level performance under standardized conditions. At the same time, catalyst A shows sharper “local maxima” where it remains superior, consistent with a catalyst that has been historically optimized for specific substrate subclasses. Together, these patterns reinforce the central point of the generality analysis: rather than producing a single universally dominant catalyst, the workflow identifies broadly competent scaffolds with complementary selectivity profiles—exactly the behavior expected for general catalysts. In this sense, the goal of the workflow is not to identify a universally dominant catalyst, but to distinguish catalysts that remain reliably competitive across a broad substrate landscape from those whose performance is more narrowly localized. Importantly, the main implication is not that catalyst B replaces catalyst A, but that a second, comparably general imidazolidinone has remained comparatively underutilized for nearly two decades. This disconnect between historical usage and broad performance underscores why popularity is an imperfect proxy for generality, and why bias‐aware, data‐driven prioritization can surface viable catalysts that empirical screening may overlook. Finally, while we avoid assigning a detailed stereochemical model here, the consistent subclass‐dependent differences (for example, within the cyclopentadiene Diels–Alder set) suggest that catalyst structure modulates selectivity in ways that can be probed in future mechanistic studies without changing the core conclusion that model‐guided workflows can accelerate identification of overlooked, general catalysts.

2.5. Probing More Difficult Substrates With Catalyst B

To more clearly define the practical boundary of catalyst B under standardized conditions, we also examined several deliberately more demanding cases (Scheme 6). Within this additional iminium‐focused set, we selected substrate combinations that were more challenging than the original benchmark systems, while avoiding cases so extreme that reactivity would be completely suppressed and direct A/B comparison would no longer be informative. We therefore adopted a local, modular approach, making small but meaningful changes to either the electrophile or the nucleophile. These included examples with increased steric bulk adjacent to the reacting center of the nucleophile, as well as enals bearing more strongly electron‐withdrawing or electron‐donating substituents. This expansion included not only additional virtually suggested catalyst–substrate pairs that broaden the comparison beyond standard literature‐precedented model cases, but also substrates not represented in the initial training set, including newly prepared compounds. Because these experiments sit somewhat outside the core reaction‐selection workflow summarized in Figure 4, we discuss them separately here rather than incorporating them directly into the Figure 5 summary. In selected difficult cases, meaningful A/B comparisons could still be obtained, including examples where stereocontrol remained high despite diminished conversion. For example, Friedel–Crafts reactions of indoles bearing additional steric bulk adjacent to the site of alkylation proceeded cleanly under the standardized conditions, giving high yields and moderate enantioselectivities (Scheme 6A). Expanding the enal component to include both electron‐donating and electron‐withdrawing substituents gave comparably informative results. In most of these cases, catalyst B performed better than catalyst A, with one case giving essentially equivalent selectivity. Notably, the nitro‐substituted cinnamaldehyde example furnished 93% ee with catalyst B at room temperature, which is notable for a substrate class often evaluated under more forcing or cryogenic conditions. Scheme 6B shows that changing indole for pyrrole showed an appreciable advantage for catalyst B over catalyst A. Finally, inclusion of a Michael addition involving a silyl enol ether showed that even in a lower‐reactivity setting, high stereocontrol could still be achieved (Scheme 6C) [43]. At the same time, several more classical difficult reactions proved too challenging for either catalyst under the matched conditions and therefore were not informative for direct comparison (see Supporting Information for details). This broader set of outcomes, spanning both successful and low‐reactivity cases, reduces concern that the validation set is limited to favorable examples and helps define the practical boundary of the benchmark space studied here.

SCHEME 6.

Additional tests of substrates and reactions in iminium catalysis. (A) Friedel‐Crafts alkylation of indoles for sterically demanding indoles and enals with different electronic effects. Reaction conditions follow Scheme 1. (B) Friedel–Crafts alkylation of pyrroles on an electron‐deficient enal. Reaction conditions follow Scheme 2. (C) Mukaiyama–Michael addition. Conditions: aldehyde (0.2 mmol), silyl enol ether (1 mmol), catalyst (22 mol%), 2,4‐dinitrobenzenesulfonic acid (20 mol%), ^t BuOH/ ⁱ PrOH = 5/1 (0.5 mL). Isolated yields. Enantiomeric excess (ee) was measured by chiral SFC. Stereochemistry is assigned in analogy to the reported examples [43].

2.6. Experimental Validation of Less General and Prospective Catalysts

A useful generality workflow should do more than nominate a top candidate—it should also (i) deprioritize catalysts predicted to be less general, including scaffolds with sparse historical coverage, and (ii) show transferability to structures beyond the training set. Although low‐generality catalysts are less attractive as practical starting points, including them in the validation set provides a stringent test of whether the model has learned portable structure–selectivity relationships rather than reproducing historical usage patterns.

To probe these two points, we evaluated four additional imidazolidinone catalysts across representative substrates. Catalysts C (CH₂‐iPr, gem‐dimethyl, G = 0) and E (CH₂‐indole, t‐Bu, G = 0.37) were included in the training dataset to test whether the model correctly deprioritizes lower‐generality members of the same family. In contrast, catalysts D (phenyl, t‐Bu) and F (benzyl‐protected CH₂–indole, gem‐dimethyl) are prospective analogs not present in the training set; they were introduced to test whether trends inferred from the curated data extrapolate to new structures assembled by swapping a single aromatic or alkyl element.

When tested across three representative substrates spanning distinct reaction classes (protocols as in Schemes 1, 2, and 4), all four catalysts delivered substantially lower enantioselectivities than catalyst B under identical conditions. Specifically, catalyst C gave 15%, 28%, and 2% ee; catalyst D gave 40%, 32%, and 27% ee; catalyst E gave 68%, 46%, and 80% ee; and catalyst F gave 46%, 73%, and 38% ee (Figure 6). The broad spread in performance across closely related imidazolidinones highlights that small structural changes can translate into large selectivity differences even under standardized conditions.

FIGURE 6.

Experimental evaluation of catalysts C‐F. These structures were tested following the protocols outlined in Schemes 1, 2, and 4.

Although this panel is intentionally compact, the ordering is consistent with two qualitative determinants that emerge from the model‐guided prioritization. First, catalysts bearing t‐Bu at the alkyl position outperform the corresponding gem‐dimethyl variants across the substrates tested. Second, increased aromatic surface area on the CH₂–aryl side correlates with improved performance (CH₂–indole/benzyl‐type motifs outperform phenyl, which outperforms a purely aliphatic substituent). Consistent with these expectations, catalysts A and B (both t‐Bu and extended aromatic character) deliver the highest selectivities; catalyst E (CH₂–indole, t‐Bu) is intermediate; and C (CH₂–i‐Pr, gem‐dimethyl) is consistently the poorest. Notably, the prospective catalysts D and F fall between these extremes, supporting the idea that the workflow captures structure–selectivity trends that remain informative beyond the specific catalysts represented in the training set.

Overall, this expanded benchmarking supports three practical uses of the workflow: identifying an underutilized high‐performing scaffold (B), deprioritizing low‐performing analogs within the same family (e.g., C), and guiding the placement of prospective variants (D, F) without requiring historical adoption as a proxy for quality. Notably, these qualitative structure–performance trends are clearest only after the ML‐guided prioritization has identified which overlooked scaffolds warrant direct comparison; in that sense, the workflow organizes experimental attention rather than replacing conventional chemical interpretation. These qualitative structure–performance trends are also consistent with the descriptor‐contribution analysis reported in the Supporting Information, which indicates that catalyst‐local steric, electronic, and polarizability‐related features make major contributions to the model, consistent with the idea that larger, more polarizable aromatic substituents can be beneficial within this scaffold family.

2.7. Additional Experimental Tests

Encouraged by the strong performance of catalyst B in the benchmark reaction panels, we next sought to evaluate its applicability in transformations that were mechanistically related yet absent from the original generality dataset. To do so, we carefully selected a set of reactions that (i) have precedent with MacMillan catalysts, (ii) test different demands on the catalyst (complex product formation [44], enamine catalysis [45, 46, 47], and synergistic multicatalysis [48]), and (iii) can be evaluated under matched conditions for catalysts A and B. Because these reactions fall outside the scope of the generality metric, we treat them as external transferability tests rather than additional inputs to generality scoring. They are therefore used to probe whether the scaffold‐level trends identified in the iminium‐based analysis persist in related but mechanistically distinct settings, not to redefine the scope of the formal generality model.

First, we tested catalyst B in a modified Friedel–Crafts alkylation of tryptophol, designed to generate complex multicyclic products (Scheme 7A). Despite the increased structural complexity, which would be expected to favor specific, tailored catalyst–substrate interactions, catalyst B delivered enantioselectivity comparable to catalyst A.

SCHEME 7.

These transformations represent a carefully chosen reaction set used to probe whether catalysts A and B can retain benchmark‐level behavior across mechanistically diverse but related substrates. (A) A molecule generated from iminium catalysis with increased complexity. (B) Reactions of enamine catalysis. (C) A synergistic reaction involving a secondary amine and copper catalyst. Reaction conditions: (A) N‐benzyltryptophol (0.2 mmol), acrolein (1 mmol), catalyst (20 mol%), TFA (20 mol%), DCM/IPA (85/15, 0.5 mL), r.t., 1 h. (B) (1) benzyloxyacetaldehyde (1.2 mmol, for a 0.4 mmol scale reaction), catalyst (10 mol%), Cl₃CCOOH (10 mol%), Et₂O (0.4 mL), r.t., 48 h. (2) diarylmethanol (0.2 mmol), propionaldehyde (0.6 mmol), catalyst (20 mol%), TFA (20 mol%), Et₂O (2 mL), r.t., 24 h. (3) aldehyde (0.2 mmol), NFSI (0.24 mmol), catalyst (20 mol%), DMF (0.5 mL). (C) phenylpropyl aldehyde (0.2 mmol), TEMPO (0.24 mmol), catalyst (20 mol%), TFA (20 mol%), CuCl₂ (10 mol%), DMF (0.5 mL), r.t., 20 h. Isolated yields given. Enantiomeric excess (ee) and diastereomeric ratio (dr) were measured by chiral SFC. Stereochemistry is assigned in analogy to reported examples [44, 45, 46, 47, 48].

Second, recognizing that our generality metric had been developed using only iminium‐based pathways, we challenged catalyst B in a mechanistically distinct transformation proceeding via an enamine intermediate (Scheme 7B). To probe diversity within this out‐of‐domain setting, we examined an aldehyde self‐aldol reaction [45], a nucleophilic substitution [46], and two fluorination reactions [47] spanning differentiated substrate demand, including electronically varied and more challenging cases. Overall, catalyst A was somewhat more selective in several of these examples, although catalyst B remained competitive and, in selected cases, matched or exceeded the performance of A. Most importantly, the expanded enamine panel shows that the two scaffolds continue to display related behavior outside the original iminium‐based training space, rather than diverging completely once the activation mode changes. Even in the more difficult cases, meaningful A/B comparisons could still be obtained, and in one example, useful stereocontrol was retained despite diminished reactivity.

Finally, to further probe transferability, we examined a synergistic catalytic system involving a secondary amine and copper co‐catalyst (Scheme 7C). This dual‐activation mechanism requires precise cooperation between catalytic cycles. While neither catalyst achieves high enantioselectivity, catalyst B nevertheless outperformed catalyst A under these complex conditions, indicating a degree of compatibility with cooperative catalysis and providing a foundation for further fine‐tuning. Most notably, the trend observed throughout this study, that B is consistently as selective or more selective than A, is maintained even in this challenging dual‐activation context.

We further explored several SOMO and related synergistic catalytic transformations under matched conditions. In these cases, no product was obtained for either catalyst, and these experiments, therefore, did not enable a meaningful A/B comparison for inclusion in the main text (see Supporting Information). Nonetheless, these negative results are useful in delineating the current performance limits of both scaffolds under especially demanding conditions.

Overall, these experiments extend the transferability analysis beyond the original training domain and provide a more balanced picture of scaffold behavior. They do not imply universal generality across all secondary‐amine‐catalyzed manifolds, but they do show that catalyst B can retain useful performance in selected mechanistically distinct settings while helping define the operational boundaries of its applicability.

2.8. Retrospective Insight: Time‐Sliced Forecasting of Catalyst Generality

Motivated by the identification and experimental validation of an underused imidazolidinone scaffold (catalyst B), we asked whether generality signals can be detected early from incomplete literature coverage (Figure 7A). To test this, we performed a retrospective time‐sliced analysis in which models were trained using only reactions reported up to 2005 (Figure 7B), and catalyst generality was then evaluated under the same virtual substrate landscape and scoring protocol used throughout this study (methods and controls in the Supporting Information). Despite sparse early representation of scaffolds that later became widely adopted, the time‐sliced models prioritized diarylprolinol silyl ethers and imidazolidinones among the most general structures, while several catalysts that were prominent in the early literature were not predicted to be broadly general [49, 50, 51, 52, 53]. Importantly, this analysis is not presented as a claim that ML “would have” discovered these catalysts historically; rather, it provides a retrospective demonstration that data‐driven generality scoring can prioritize catalyst families that subsequently proved broadly useful, even when early data are limited. Together with the prospective experimental validation presented above, these results support the idea that generality‐oriented modeling can help decouple catalyst prioritization from historical popularity and identify scaffold families that merit earlier experimental attention.

FIGURE 7.

Trajectory and early activity of proline and related amine catalysts in the literature. (A) Cumulative literature reports from 1999–2023 for proline and six structurally distinct amine catalysts (A, C‐F), showing divergent long‐term adoption. Proline shows rapid early growth but plateaus, while AmineE and AmineF display sustained uptake over time. (B). Number of literature reports for each catalyst in 2005, highlighting their comparable early activity despite divergent long‐term generality. Catalyst structures are shown above their respective bars.

3. Conclusion

This study presents an ML framework for identifying general catalysts from historically sparse and biased datasets, a setting where conventional screening and popularity‐driven heuristics often fall short. By combining bias‐aware modeling, a shared virtual substrate landscape, and cluster‐based generality scoring, we show that catalyst prioritization can be decoupled from historical frequency of use. The identification and experimental validation of a rarely studied imidazolidinone scaffold—one that is not widely commercialized yet performs competitively under standardized conditions—highlights the ability of ML‐guided workflows to surface effective but overlooked catalysts.

Importantly, the formal generality claims made here are restricted to the iminium‐based reaction space used to construct the curated and virtually balanced dataset; the additional enamine and synergistic examples are best interpreted as transferability tests that help define both the reach and the present limits of scaffold B. Time‐sliced retrospective analyses further suggest that generality‐oriented ranking signals are detectable even from incomplete early literature coverage. Using only reactions reported up to 2005, our framework prioritized catalyst families that later became widely adopted, including diarylprolinol silyl ethers and imidazolidinones, while several scaffolds prominent in the early literature were not similarly prioritized under the same metric. While such retrospective tests cannot recreate the full decision context of the field at the time, they support the central premise that data‐driven generality scoring can help focus experimental effort earlier than is possible through empirical accumulation alone.

Beyond catalyst nomination, this work underscores the importance of explicitly addressing dataset bias and maintaining interpretability when applying ML to chemical discovery. The outputs of the framework are intended to be experimentally actionable—guiding head‐to‐head benchmarking, deprioritizing low‐promise candidates, and enabling structure‐guided hypothesis generation—rather than serving as black‐box predictions.

Looking forward, the same workflow can be extended to other catalyst classes and to settings where data are even more limited or costly to generate. More broadly, pairing mechanistically informed representations with bias‐aware prioritization provides a practical foundation for discovery strategies that are faster, more systematic, and less constrained by historical reporting patterns. Future extensions to computationally proposed scaffolds, moving beyond literature‐precedented structures, could further accelerate discovery by enabling de novo design explicitly optimized for generality.

Author Contributions

Jiajing Li: writing – original draft, writing – review and editing, investigation, validation, methodology. Isaiah O. Betinol: investigation, writing – original draft, writing – review and editing, validation, methodology. Junshan Lai: investigation, validation. Soresu Juyo: investigation. Jolene P. Reid: conceptualization, funding acquisition, writing – original draft, writing – review and editing, project administration, resources, supervision.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Supporting File 1: anie73014‐sup‐0001‐SuppMat.pdf.

Supporting File 2: anie73014‐sup‐0002‐DataS1.zip.

Data Availability Statement

The data that support the findings of this study are available in the Supporting Information of this article.