Data‐Driven Approach to Understanding Tetrabutylammonium Decatungstate‐Catalyzed C(sp³)–H Functionalization Selectivity

ABSTRACT

Photocatalytic hydrogen atom transfer (HAT) has emerged as a powerful strategy for selective C(sp³)–C bond formation. Among these systems, photocatalyst tetrabutylammonium decatungstate (TBADT) has drawn attention for its ability to functionalize electron‐rich C–H bonds via polar and steric control, often achieving highly regioselective outcomes. While these principles explain outcomes in simple scaffolds, translating them to understand regioselectivity in more complex molecules remains a central challenge. Here, we integrate electronic and steric descriptors to analyze the factors governing regioselectivity. The strong association of radical‐based descriptors with observed selectivity trends motivates targeted mechanistic experiments to probe the extent to which catalyst‐controlled hydrogen atom abstraction (HAA) versus radical‐controlled addition governs regioselectivity. Deuterium‐labeling studies reinforce that TBADT's HAA accessibility is broad across electronically similar C–H bonds. Kinetic analyses further establish that SOMO energy directly correlates with radical addition (RA) rates, showing that product distributions are funneled through frontier orbital control. Together, these results define the kinetic principles of radical sampling in TBADT‐mediated functionalization, where HAA accessibility is broad, and RA kinetics further shape regioselective outcomes. Logistic regression analysis further demonstrates the value of mechanistically rich features in predicting site selectivity across complex molecular scaffolds.

This work leverages electronic and steric descriptors to both justify and predict complex regioselectivity outcomes in TBADT‐mediated hydrogen atom transfer. Radical‐based features show strong trends with reactivity, inspiring targeted mechanistic studies which reveal that while TBADT abstracts many C–H bonds, radical addition kinetics ultimately shape functionalization outcomes.

1. Introduction

The ubiquity of C(sp³)–H bonds across chemical matter makes them an attractive platform for rapid late‐stage diversification, and their direct functionalization has emerged as a powerful strategy for forging new C(sp³)–C bonds to access more potent and selective therapeutics [1, 2, 3]. Among the approaches to achieve this, photocatalytic hydrogen atom transfer (HAT) has emerged as particularly valuable, enabling the functionalization of strong C(sp³)–H bonds under mild conditions [4]. Photocatalytic HAT mechanisms use light as an energy source to generate reactive open‐shell species, overcoming otherwise prohibitive energy barriers and granting access to a wide range of previously inaccessible substrates, including unactivated aliphatic C(sp³)–H bonds [5].

Given the importance of selective molecular transformation in synthesis, one of the major challenges of C(sp³)–H functionalization is achieving precise regioselectivity. Regioselectivity is essential for the generation of pharmaceuticals, materials, and imaging agents with the desired purity and function [6]. The current understanding of regioselectivity outcomes is supported by empirical patterns as well as general mechanistic rationale. Bond dissociation energies (BDEs) are often invoked in discussions of site selectivity, with regioselectivity thought to arise from abstraction of the C–H bond with the lowest BDE [7]. While BDEs often provide a useful framework for rationalizing selectivity outcomes, they are not the sole determinant of regioselectivity; polarity‐matching effects can also play an important role [8, 9]. Polarity‐matching describes the principle that electron‐deficient species preferentially react with electron‐rich partners, and vice versa [10]. This alignment of electronic character can occur between a catalyst and substrate, or between two substrates directly. Thus, regioselectivity often occurs at the site where this electronic complementarity is maximized, a selectivity‐determining factor across many radical cross‐coupling systems [11, 12]. Steric effects are another important dimension in controlling regioselectivity in HAT reactions. Catalysts with bulky frameworks may be sterically excluded from approaching hindered sites that might otherwise be electronically favorable [6]. Likewise, steric congestion at the substrate can inhibit radical approach to the reaction site, disfavoring the subsequent product‐forming step [13]. Furthermore, under reversible HAT conditions, the radical addition (RA) step can become the selectivity‐determining event between electronically similar sites, with regioselectivity dictated by the relative stabilities and polarity matching of the incipient radical–acceptor interactions [14].

In simple molecules with few competing sites, weighing the interplay of these effects to rationalize regioselectivity or overall reactivity can be straightforward. Extending these trends to more complex molecular scaffolds, however, becomes increasingly difficult when multiple electronically or sterically viable sites compete for functionalization. For instance, the widely studied C–H partner ibuprofen contains benzylic, tertiary, and α‐ester hydrogens—each of which has documented reactivity toward electrophilic hydrogen abstractors [15, 16]. Sclareolide, a diterpene lactone, provides an even more challenging case, where overlapping electronic and steric influences create ambiguity in site outcomes. In such cases, it is difficult to distinguish whether regioselectivity stems from steric congestion, subtle electronic differences, or steric effects imposed by the coupling partner itself [17]. These dilemmas highlight the limitations of using empirical patterns for predicting regioselectivity outcomes and highlight the need for more rigorous frameworks. Bridging this gap between qualitative intuition and accurate prediction has the potential to achieve both improved reactivity and regioselectivity outcomes in photocatalysis‐mediated HAT.

A more robust, quantitative understanding of the factors governing selectivity offers a path forward. Computational analysis is especially well‐suited to this challenge, as it enables predictive models that generalize across chemical space while minimizing experimental input [18]. For example, Milo, Reisman, and co‐workers developed a machine‐learning framework for accurate prediction of C(sp³)–H oxidation [19]. Their strategy employed lower‐cost alternatives to traditional DFT‐based physiochemical descriptors, leveraging semi‐empirical methods, machine‐learning prediction tools, and structural features (Figure 1a). These streamlined descriptors proved effective for accurate rank‐based identification of reactive C(sp³)–H sites within complex scaffolds, highlighting the power of computational methods to systematically capture and generalize selectivity patterns. Still, the interplay of thermodynamic, polar, and steric factors remains difficult to disentangle, making it challenging to extend predictive models beyond well‐characterized systems such as radical arene C–H borylation. This highlights the need for a broader quantitative framework capable of distinguishing the contributions of these overlapping effects. In this work, we address this need by combining DFT‐based electronic and steric descriptors with a logistic regression modeling framework applied to photocatalytic direct HAT substrates. Logistic regression was selected based on its simplicity and transparency, which enables direct interpretation of feature coefficients [20]. This approach enables not only a predictive framework for selectivity but also reveals an underlying mechanistic architecture that governs selectivity across structurally diverse substrates.

FIGURE 1.

(a) Milo and Reisman et al.’s use of low‐cost featurization methods enable a machine‐learning framework for accurate prediction of C(sp³)–H oxidation regioselectivity in radical arene C–H borylation [19]. (b) Ravelli et al. establishes TBADT's steric and polar control over product outcomes [17]. (c) The present work integrates qualitative precedents with quantitative computational descriptors to define steric and electronic thresholds, thereby providing mechanistic insight into site‐selectivity across TBADT‐mediated HAT reactions.

This work focuses on regioselectivity outcomes mediated by the photocatalyst tetrabutylammonium decatungstate (TBADT), which belongs to the broader class of transition metal polyoxometalate (POM) photocatalysts, valued for their strong molar absorptivities and long‐lived triplet states [7, 21]. The selectivity of POM catalysts is widely attributed to the mechanism of catalyst activation, where photoirradiation generates oxyl radicals with partial electrophilic character that display high selectivity for nucleophilic, electron‐rich C–H bonds [22, 23]. Similarly, aromatic ketone and aldehyde organic photocatalysts achieve regioselectivity through oxygen‐centered radicals with electrophilic character and thus the same preference for polarity‐matched, nucleophilic hydrogens [24, 25, 26]. This shared mechanistic feature defines a general electrophilic paradigm, in which electron‐deficient oxygen‐centered radicals display strong selectivity toward electron‐rich C–H sites.

Within this broad class of electrophilic HAT catalysts, TBADT was chosen as a model system due to its photostability [27], commercial availability, high functional group compatibility [28], and ability to abstract exceptionally strong C(sp³)–H bonds—approximately 105 kcal mol^{− 1}, such as those in methane [29]. This high reactivity is attributable to the strong thermodynamic driving force of forming a tungsten–oxygen–hydrogen (W–O–H) bond in the decatungstate framework, found to be approximately 113.4 kcal mol⁻¹ [30]. These properties make TBADT an ideal benchmark for probing general selectivity trends in electrophilic catalyst systems. Additionally, TBADT‐mediated HAT chemistry represents a broad and actively developed area of research, providing a substantial body of experimental precedent for characterizing reactivity and establishing a strong foundation for computational analysis [31, 32, 33, 34, 35]. Ravelli and co‐workers demonstrate that, like other POM‐based photocatalysts, TBADT's photoexcited‐state oxyl ligands favor abstraction of polarity‐matched, nucleophilic C–H bonds (Figure 1b) [17]. The influence of steric effects on TBADT's selectivity has also been postulated, arising from the catalyst's bulky nature [16]. As a result, certain C–H sites that would otherwise be electronically favored can become disfavored once steric constraints are introduced.

This investigation further focuses on reactions involving olefins as radical acceptors, chosen for their ubiquity and mechanistic simplicity [36]. The addition of radicals to olefins follows a well‐studied mechanism [37, 38, 39, 40], making it possible to distinguish which features influence the HAA step versus the RA step. This separation reduces confounding variability and enhances the interpretability of computational analysis, particularly in identifying which descriptors are most strongly associated with each mechanistic contribution. Olefins further serve as ideal acceptors because they enable installation of multiple new C(sp³) centers in a single step, rapidly increasing the saturation and three‐dimensionality of a substrate [41]. Their advantageous properties, combined with extensive precedent in radical chemistry, have made olefins among the most widely employed radical acceptors, ensuring that insights are broadly generalizable and effective for establishing quantitative trends.

Thus, computational analysis restricted to TBADT‐mediated RA to olefins provides an effective platform for deconvoluting the contributions of electronic and steric effects and distinguishing between the selectivity‐determining features of the hydrogen abstraction step versus those of the subsequent product‐forming RA step (Figure 1c). By aligning feature selection with both general reactivity principles and the specific mechanistic context of these transformations, this approach enables direct testing of current mechanistic hypotheses and provides a more rigorous, quantitative framework for understanding regioselectivity in electrophilic HAT systems. BDEs, long recognized as key thermodynamic descriptors of C–H activation, were included to benchmark their importance. Steric descriptors—for both the HAA step and olefin addition step—were incorporated to establish steric thresholds for reactivity. Of course, because TBADT's high site selectivity is widely attributed to polarity‐matching of the C–H abstraction step [17], various methods for calculating hydrogen atomic charges were included to probe this hypothesis. Finally, features describing the RA step, such as radical philicity, singly occupied molecular orbital (SOMO) energy, and electronic and steric descriptors of the reaction partner, were incorporated to assess the influence of the olefin addition step on site‐selectivity.

Surprisingly, features reflecting polarity‐matching of the C–H abstraction step proved to be less indicative of reactivity when evaluated in a relative context, while descriptors of the subsequent RA step emerged as far more reliable. This conclusion is supported by deuterium‐labeling experiments, which reinforce that TBADT engages a broad range of nucleophilic C–H bonds within a single molecule, even as product distributions remain highly selective. These findings support the hypothesis that site selectivity arises not from the initial abstraction event alone but is strongly influenced by the characteristics of the RA step [14]. Among the surveyed descriptors, radical philicity and SOMO energy proved especially indicative of site selectivity, while steric descriptors revealed key thresholds governing reactivity across both mechanistic steps. The emergence of SOMO energy and radical philicity as important parameters inspired targeted kinetic experiments, providing evidence for the relevance of frontier orbital interactions to regioselectivity. This evidence directly relates the selectivity of modern photocatalyst systems to classical Giese addition principles, describing the relevance SOMO‐LUMO interactions as a unifying determinant that transcends catalyst identity. Finally, to integrate takeaways and descriptors in a physically meaningful manner, a logistic regression model was trained on the literature‐curated dataset of small molecules, then applied to unseen, complex scaffolds to demonstrate the predictive capabilities of the surveyed descriptors.

This work therefore provides two complementary advances. First, the mechanistic insights obtained from descriptor analysis and kinetic experiments help rationalize previously ambiguous selectivity patterns. Second, the logistic regression framework provides a predictive tool for identifying reactive sites in complex scaffolds. Together, these advances create opportunities to expand reactivity to new scaffolds, access previously inaccessible C–H sites, and fine‐tune regioselectivity through subtle electronic and steric modifications.

2. Methods

One of the most challenging aspects of building comprehensive chemical datasets is the design of negative examples [42, 43]. For understanding TBADT‐mediated HAT site selectivity, simply including molecules that are entirely unreactive toward functionalization would be misleading. Unreactivity may arise from confounding factors unrelated to intrinsic molecular properties. Therefore, complete molecular unreactivity should not be conflated with site‐specific unreactivity.

Instead, this work augments negative examples by treating experimentally observed reactive sites as positive examples and all other sites on the same molecule, where no reactivity was recorded, as negative examples [44]. This intramolecular comparison provides a more controlled and chemically meaningful context for evaluating feature influence on reactivity. That said, binary assignment carries inherent limitations. First, it assumes that the experimentally observed site is the only reactive site, whereas in practice, side reactivity is often present but goes unrecorded. This reductive labeling risks oversimplifying the data and obscuring minor but real contributions. To assess the validity of this assumption, benchmarking was conducted using substrates with reported product mixtures, finding that the same descriptor trends were maintained (Supporting Information: Section I). Second, defining reactivity solely by product outcome conflates the two mechanistic steps of the reaction—hydrogen abstraction and olefin addition. Since these steps can exert distinct influences on product distribution, more granular experimental data would be needed to fully disentangle their contributions. Ultimately, however, the dataset is constrained by experimental feasibility, making this binary approach a pragmatic compromise for extracting meaningful trends.

Additional assumptions arise from the experimental conditions represented in the literature dataset. Solvent conditions were restricted to acetonitrile to minimize solvent‐specific artifacts, while light source, reactant concentrations, and catalyst loading vary modestly across the curated reports. These considerations, along with a detailed summary of reaction parameters across the curated literature set, are provided in Supporting Information: Section I.

Binary labels for site reactivity were assigned to each C(sp³)–H position across 39 literature‐curated molecules drawn from 9 sources [16, 45, 46, 47, 48, 49, 50, 51, 52], encompassing 72 reactions with 15 different olefin partners. In total, this yielded 232 data points, comprising 72 reactive sites labeled as Reactive and 159 unreactive sites labeled as Unreactive (Figure 2a). A more in‐depth description of the molecules, sources, and reaction outcomes is available in Supporting Information: Section I.

FIGURE 2.

(a) Binary labeling approach used in dataset curation. (b) General overview of hydrogen and radical features.

Feature selection was guided by mechanistic considerations specific to TBADT‐mediated reactions as well as broadly relevant chemical principles such as thermodynamics and steric effects (Figure 2b). Given the strong thermodynamic driving force provided by W–O–H bond formation, machine‐learning predicted C–H BDEs were calculated for all sites to assess the contribution of thermodynamic control to selectivity [53]. To evaluate the catalyst‐substrate polarity‐matching principle, several atomic charge descriptors were considered, spanning the two major classes of atom‐in‐molecule definitions [54]. NBO [55, 56], Mulliken [57], Löwdin [58], and Hirshfeld [59] charges were computed for both reactive and unreactive hydrogens. This range of approaches allows for benchmarking how different charge analysis schemes capture electronic differences and, ultimately, how well they predict site selectivity.

To evaluate the steric influence on reactivity, DBSTEP parameter buried volume (%V_bur) was used to quantify the steric profile at each site [60]. While %V_bur is a well‐established metric for assessing the steric bulk of organometallic ligands [61, 62, 63], it has also been successfully applied to organic substrates [64]. In this work, %V_bur was calculated for both the hydrogen abstraction site and the resulting radical center of the lowest energy conformer, enabling differentiation between steric effects operative in the HAA step versus those that influence the olefin addition step.

To characterize the electronics of the RA step, a distinct set of descriptors was employed: radical philicity, the Hirshfeld charge of the radical carbon, and SOMO energy. Polarity matching between radical and acceptor is well recognized as a governing factor in many radical cross‐coupling systems [8, 11]. Accordingly, radical philicity and radical carbon Hirshfeld charge were included to probe how polarity differences distinguish reactive from unreactive sites, thereby evaluating the contribution of addition‐step electronics to selectivity. SOMO energy was also incorporated due to its central role in frontier orbital interactions; SOMO–LUMO or SOMO–HOMO alignments dictate whether a radical exhibits nucleophilic or electrophilic character [65]. SOMO energy thus serves as a direct probe of the intrinsic electronic character of the radical and its compatibility with the olefin partner.

Features describing the olefin reaction partner were also incorporated to assess their role in influencing site selectivity. Orbital energies, including LUMO and HOMO values, were recorded to evaluate the boundaries for orbital complementarity and assess important coupling partner parameters. Additionally, %V_bur of the olefin at the RA carbon was calculated to capture steric effects and enable comparisons across different olefin partners.

A more extensive description of all the features used in this study as well as how they were calculated, atomic radii selection for %V_bur analysis, and basis set benchmarking can be found in Supporting Information: Sections II–IV.

3. Results and Discussion

Several descriptors exhibited clear and consistent relationships with site selectivity. Density plots provide an effective method for visualizing these relationships, highlighting the distinct value distributions for reactive versus unreactive hydrogens and enabling direct comparison of their respective ranges (Figure 3a). Density plots use kernel density estimation (KDE) to generate a continuous curve that represents the probability density of the data [66, 67]. Each point along the curve represents an estimate of the probability density at that descriptor value, thus the plots highlight not only where values are clustered but also the relative likelihood of encountering sites with specific properties. Sharp peaks suggest strong convergence of feature values, while broader plateaus reveal a larger distribution of outcomes.

FIGURE 3.

(a) Density plots reveal general trends for reactive and unreactive C‐H sites. (b) Accuracy of hydrogen and radical descriptors when evaluated on a relative, per‐molecule basis. (c) Logistic regression analysis applied to previously unseen, complex scaffolds.

Given the significant overlap between the density curves for reactive and unreactive hydrogens, BDE values did not reliably differentiate reactivity. One important factor to consider, however, is that arene C(sp²)–H bonds were excluded from featurization in this study, as TBADT has not been shown to abstract them. These bonds are characterized by significantly higher BDE values, and if they had been included as negative examples, the BDE density plot would almost certainly display a more distinct separation. That said, based on the C(sp³)–H featurized sites with relatively similar BDE values, the absence of a clear correlation supports the conclusion that thermodynamics play a background role but do not dominate selectivity when multiple C–H bonds of similar strength are present. Indeed, TBADT can abstract a wide range of strong C–H bonds, including those up to 99.2 kcal mol^{− 1}—the upper limit established in this study.

Atomic charge features, particularly NBO and Mulliken charges, showed a strong relationship with reactivity based on their density plots: reactive sites had a consistently more nucleophilic probability distribution than unreactive sites. This observation reinforces the polarity‐matching principle, whereby a favorable electronic alignment between the C–H bond and the photoexcited catalyst stabilizes the hydrogen‐abstraction transition state, rendering electron‐rich C–H bonds more susceptible to abstraction [8]. By contrast, Löwdin and Hirshfeld charges were comparatively poor predictors, given that their density plots showed little differentiation between reactive and unreactive hydrogens, shown in Supporting Information: Section V.

Reactive and unreactive sites also showed clear differentiation when evaluated by radical descriptors, particularly radical philicity and SOMO energy. Reactive sites consistently exhibited more nucleophilic philicities and higher SOMO energies when compared to their unreactive counterparts. These findings suggest that the electronic characteristics of radical intermediates play a key role in governing regioselectivity.

Although density plots reveal no absolute threshold separating reactive from unreactive sites, given that there was significant overlap in their value ranges, several descriptors provide clearer differentiation when evaluated in a relative context (Figure 3b). Under this relative framework (detailed in Supporting Information: Section V), features were evaluated against their expected reactivity trends. For example, by polarity‐matching, functionalization is expected at the hydrogen with the lowest (most negative) Mulliken charge; indeed, for 32 of 39 C–H partners (82.1%), the site with the lowest Mulliken charge matched the experimental site of functionalization. Applying the same ranking protocol, SOMO energy correctly identified the experimentally functionalized site in 92.3% of cases, radical philicity in 94.1%, and NBO charge in 76.9%. The strong consistency of multiple descriptors with the observed outcomes indicates that site selectivity is not dictated by fixed global cutoffs but instead arises from intramolecular competition, where subtle electronic differences between sites bias the reaction pathway toward the most favorable site.

While these descriptors exhibit strong reactivity trends within the literature‐curated dataset, the interplay between hydrogen‐ and radical‐centered steric and electronic effects becomes increasingly difficult to disentangle in structurally complex scaffolds. Logistic regression provides a simple yet powerful statistical framework for integrating these descriptors in a transparent and interpretable manner (Figure 3c). A logistic regression model was trained on the electronic, steric, and thermodynamic features of the literature dataset and subsequently applied to previously unseen complex scaffolds to generate site‐specific functionalization probabilities for ibuprofen, sclareolide, and ambroxide. In all three cases, the experimentally observed site of functionalization was assigned the highest predicted probability. These results demonstrate how mechanistically informed, highly accurate descriptors can often be as effective as larger datasets and more sophisticated models.

Logistic regression further enables quantitative interpretation of feature contributions within the modeling framework. Analysis of grouped feature importances highlights hydrogen‐centered steric descriptors as a dominant contributor to the model. However, hydrogen and radical steric descriptors were found to be strongly correlated (Supporting Information: Section VII), reflecting the limited structural reorganization typically observed between closed‐ and open‐shell species. Accordingly, the disparity in their individual feature weights should not be interpreted as definitive evidence of fundamentally different steric effects. Rather, the model may preferentially assign weight to one of two highly collinear variables that encode largely overlapping steric information.

Nevertheless, the prominence of steric descriptors overall underscores the critical interplay between steric accessibility and electronic factors in governing site selectivity. Radical electronic descriptors also emerged as significant contributors, consistent with the earlier identification of SOMO energy and radical philicity as key parameters in relative reactivity trends. A more comprehensive correlation and multicollinearity analysis within the context of the logistic regression model is provided in Supporting Information: Section VII.

Steric hindrance restricts access to several tertiary positions on sclareolide that would otherwise be highly reactive based on their electronic profiles (Figure 4a, Supporting Information: Section VI). Thus, these positions serve as an effective probe for evaluating steric thresholds across both the hydrogen abstraction and Giese addition steps. In contrast, the tertiary positions on cis‐decalin—a similarly fused bicyclic system—do exhibit reactivity (Supporting Information: Section XI). This comparison allows for a steric boundary to be inferred: inaccessibility occurs between 85.4% and 90.7% radical buried volume. Similarly, analysis of the substrate with the highest hydrogen % buried volume in the small‐molecule dataset establishes a steric threshold for the hydrogen abstraction step: between 52.5 and 60.8 hydrogen % buried volume.

FIGURE 4.

(a) Steric thresholds define regions where reactivity switches from accessible to inaccessible. (b) Application of these thresholds to unreactive substrates validates the predictive framework [17]. (c) Quantitative descriptors distinguish electronic and steric contributions to hydrogen abstraction, clarifying how subtle changes bias reactivity.

Interestingly, the tertiary site of cis‐decalin has previously been shown to undergo HAT under TBADT‐mediated conditions, isomerizing from cis‐ to trans‐decalin [68]. However, in the presence of an alkene coupling partner, tertiary‐site reactivity is diminished, producing a mixture of regioisomeric products from both tertiary and secondary positions. This outcome further supports the conclusion that the olefin addition step exerts a steric influence on the regioisomeric distribution.

As discussed previously, a limitation of binary labeling based on product distributions is that it conflates the hydrogen atom abstraction and olefin addition steps into a single outcome. Consequently, the steric threshold inferred for HAA may be underestimated, since the olefin addition step also introduces steric constraints that shape the final product profile. That said, the established steric boundaries still provide a quantitative framework that supports why certain substrates, classified as sterically unreactive, fail to undergo functionalization (Figure 4b) [17]. The consistency of these trends across multiple scaffolds reinforces the validity of the established thresholds. Notably, the unreactive molecules, Substrates 1 and 2, exceeded the threshold by only a few percentage points of buried volume: the hydrogen %V_bur threshold was established to be 52.5, and the unreactive substrates both had 53.6 hydrogen %V_bur values; the radical %V_bur threshold was established to be 85.4, and Substrate 2 had a radical %V_bur value of 87.8. Thus, even small increases in steric bulk can have a decisive impact on reactivity.

These trends can also be useful for identifying when steric or electronics bias reactivity or selectivity outcomes (Figure 4c). Substrate 3, previously deemed unreactive due to steric hindrance [17], was found instead to be electronically disfavored by comparing its atomic charges and radical feature values to both the density plots and steric thresholds. In contrast, Substrate 4, which has a comparable steric profile but enhanced electronic properties, exhibited reactivity. These findings reveal that the steric tolerance of TBADT‐mediated HAT is substantially broader than previously assumed, effectively expanding the available reactive chemical space.

Three molecules deviated from the general SOMO and radical philicity trends (Figure 5). Substrate 5, which reacts with dimethyl maleate to give a 3% yield at the secondary site, highlights the role of thermodynamic factors in governing reactivity [45]. While BDE values had little influence when comparing sites of similar radical stability, the pronounced stability difference between secondary and primary radicals in this case likely governs the observed outcome. Substrates 6 and 7 illustrate a different scenario: although this study emphasizes the role of RA in dictating selectivity, TBADT still exerts strong polar control when hydrogens differ substantially in polarity. Both the aldehyde and formamide‐containing substrates feature hydrogens with markedly lower atomic charges compared to other sites, making these positions disproportionately favored for abstraction, independent of their radical descriptors. The NBO charge density plot further highlights their distinct electronic character: the pronounced peak near the 0.1 charge region reflects the strong polar effects of aldehydes and formamides, which enhance their susceptibility to abstraction and explain their observed reactivity. Thus, while RA governs selectivity in the majority of cases where C–H bonds exhibit comparable hydridic character and strength, the extremes of this parameter space reveal important nuances: when C–H bonds are either unusually weak or strongly polarized, the kinetics of hydrogen abstraction can compete with or even outpace functionalization, shifting the selectivity‐determining step back toward the HAA process. Although the relative SOMO energy and radical philicity paradigm did not correctly identify the preferred site of reactivity in these cases, the logistic regression model successfully predicted the reactive site for all three substrates (Supporting Information: Section VI). This outcome further emphasizes that regioselectivity cannot be attributed to a single descriptor alone. Instead, accurate prediction of reactivity in complex scaffolds requires consideration of multiple interacting steric and electronic features that collectively govern site selectivity.

FIGURE 5.

The three molecules which deviated from the relative SOMO energy or radical philicity trend.

Building on the comparison of feature performance, the finding that radical descriptors outperformed hydrogen polarity descriptors has important mechanistic implications. It reinforces TBADT's weak preference for nucleophilic C–H bonds, and that the electronic properties of the resulting radicals exert an additional influence on site selectivity. Deuterium‐labeling experiments with sclareolide and ibuprofen reinforce this conclusion [69]. Within 4 h multiple electron‐rich C–H bonds undergo abstraction (Figure 6a). In ibuprofen, both the benzylic and tertiary sites are comparably abstractable, yet the ultimate product distribution favors the tertiary site, showing that the olefin addition step funnels reactivity toward that position. Likewise, several A‐ring sites in sclareolide were efficiently labeled, yet subsequent functionalization at H₄ dominates the product outcome, again reflecting bias from the olefin addition step. While slight changes in the coupling partner and reaction conditions does not necessarily translate perfectly to olefin addition selectivity, it further demonstrates that hydrogen abstraction is a considerably less selective process. Once C–H bonds fall within an accessible range of nucleophilic polarities, the features of subsequent steps—here, olefin addition—play a decisive role in shaping product distributions. That the coupling partner exerts an influence over the product distribution is further evidenced by ibuprofen fluorination, where a different regioisomeric distribution is achieved simply by changing the coupling partner (Supporting Information: Section X).

FIGURE 6.

(a) Deuterium‐labeling experiments show that the initial HAA step is more permissive than previously emphasized. (b) Kinetic experiments by Giese et al. demonstrate that olefins bearing more electron‐withdrawing substituents exhibit lower LUMO energies, which enhance SOMO–LUMO overlap and accelerate reaction rates [36]. Complementary kinetic studies further reveal a direct relationship between C–H partner SOMO energy and rate, where higher SOMOs promote more efficient orbital overlap and faster bond formation.

The mechanistic focus is then shifted to RA as a key selectivity‐determining step. Although rarely emphasized in discussions of TBADT, the influence of SOMO energy and radical philicity on reactivity is well established in the context of Giese addition. Seminal studies on C–C bond formation via RA to alkenes have shown that reaction rates are strongly governed by olefin electronic properties, quantified through Hammett parameters. These trends are attributed to frontier molecular orbital (FMO) interactions, in which electron‐withdrawing substituents lower the olefin LUMO energy, improving overlap with the radical SOMO and accelerating bond formation [70]. Direct calculations of olefin LUMO energies confirm this hypothesis, illustrating the interplay between electron density, FMO energetics, and RA rates (Figure 6b).

By examining simple scaffolds with minimal competing sites but varying electronic properties, we demonstrate that radical characteristics—particularly SOMO energy—correlate strongly with reaction rates, extending this well‐known principle to TBADT‐mediated HAT [71]. Electrophilicity trends are consistent with this relationship: more electron‐rich systems raise the SOMO energy of the corresponding radical, thereby enhancing reactivity [65]. While SOMO energy and radical philicity are strongly correlated in these kinetic experiments, detailed correlation analysis (Supporting Information: Section VII) reveals systematic deviations between these descriptors, indicating that they capture related but not identical electronic information.

These findings highlight the importance of radical characteristics in predicting site selectivity and extend the general principles of Giese‐type additions to the olefin addition step of TBADT‐mediated C(sp³)–H functionalization. The observation that this classical radical correlation between SOMO energy and reactivity translates so clearly into a photocatalytic HAT system emphasizes the broad mechanistic continuity between traditional radical chemistry and modern photocatalysis.

4. Conclusion

This work defines a computational framework that supports the concept of radical sampling in TBADT‐mediated C(sp³)–H functionalization selectivity, quantitatively showing that when the initial HAA step is permissive, the subsequent Giese addition step further refines the product distribution based on radical intermediate characteristics. Quantitative thresholds and features were identified, demonstrating strong trends with reactivity across the literature‐curated dataset: Mulliken charge correctly identified the experimental site of functionalization with 82.1% accuracy, SOMO energy with 92.3%, and radical philicity with 94.1%. To assess the robustness of these descriptors, a logistic regression model trained on the literature dataset was applied to more complex scaffolds. The successful prediction of selectivity in unseen substrates by a simple linear model highlights the value of carefully selected high‐accuracy descriptors for chemical modeling.

Beyond prediction, this work also supports a mechanistic understanding that explains why product distributions in TBADT‐mediated photocatalysis appear more selective than deuterium‐labeling experiments suggest. While TBADT abstracts hydrogens across multiple electron‐rich sites, not all intermediates are funneled into productive pathways. Instead, selectivity reflects the reactivity of the radical intermediate toward the coupling partner, consistent with the broader principles governing Giese‐type additions in radical chemistry. This mechanistic picture helps rationalize why the descriptor trends identified here are likely to extend beyond TBADT to related electrophilic HAT systems. This broad abstraction ability also raises an important consideration for substrates containing stereogenic centers, as non‐productive abstraction can lead to epimerization—a potential limitation for late‐stage functionalization. Similarly, it helps clarify why selective isotopic incorporation remains a challenge in these systems. On the other hand, the recognition that product selectivity is strongly shaped by the coupling partner highlights an opportunity: by tuning steric or electronic properties, product distributions can be directed toward desired outcomes (Supporting Information: Section X). In this way, the mechanistic insights gained here not only define the limits of current applications but also provide a foundation for new strategies to expand the reactivity and selectivity of electrophilic photocatalytic HAT.

Author Contributions

Mimi Lavin: conceptualization, methodology, investigation, validation, writing – review and editing, writing – original draft, data curation, formal analysis, visualization. Makeda A. Tekle‐Smith: conceptualization, funding acquisition, supervision, writing – review and editing, writing – original draft, formal analysis, visualization.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Datasets and workflows are available at github.com/teklesmithlab. Experimental procedures and computational details are provided in the Supporting Information. Supporting File 1: anie72803‐sup‐0001‐SuppMat.pdf.

Data Availability Statement

The data that support the findings of this study are openly available in github at https://github.com/teklesmithlab.