On the Origin of Scission Selectivity for Phosphoranyl Radicals in Photoredox Catalysis

Abstract

Phosphoranyl radicals are versatile synthetic intermediates that can form downstream radicals through α- and β-scission. However, the unclear origin of scission selectivity hinders the rational design of new transformations using these elementary steps. Here we report a comprehensive study of phosphoranyl radical scission mechanisms for sulfonamide nucleophiles. Using a combination of experimental, computational, and machine learning-driven strategies, the primary factors influencing selectivity in these transformations were uncovered. Computational and experimental studies revealed an unexpected Curtin–Hammett scenario and demonstrated the impact of reaction stoichiometry on selectivity outcomes. Statistical modeling enabled selectivity prediction via an interpretable linear regression model, which revealed that small, electron poor phosphines favor β-scission. Taken together, these studies expose the interplay of reaction conditions, phosphine properties, and alkene properties which can be used to control selectivity in phosphoranyl radical-mediated reactions of sulfonamides. These insights were harnessed to develop a photoredox-catalyzed phosphine-mediated hydrosulfonylation reaction of alkenes, the development of which makes primary sulfonamides the first example of a nucleophile with switchable scission selectivity controlled only by phosphine and reaction conditions for phosphoranyl radical-mediated reactions.

Graphical Abstract

INTRODUCTION

Organic radicals are ubiquitous intermediates for diverse transformations. While radical chemistry has experienced a renaissance with the rise of photo-,^1–3 electro-,^4–7 and transition metal catalysis,^8–11 limitations remain with respect to controlled generation and capture of radical species. Radical generation from unfunctionalized starting materials is particularly valuable, as this strategy provides direct entry into desired transformations from abundant compounds. One approach to generate radical intermediates from readily available nucleophilic functional groups is through fragmentation, or scission, of phosphoranyl radicals.^12–16 While phosphoranyl radicals are traditionally accessed through the addition of a radical intermediate to a phosphine,^12–16 photoredox catalysis has enabled the formation of these species under mild photochemical conditions from simple nucleophiles^16–21 including alcohols,^22,23 carboxylic acids,^22,24,25 oximes,^26,27 sulfoxides,²⁸ and sulfonamides^29,30 (Fig. 1A). This approach has been utilized in various synthetic transformations including deoxygenation,^{20–22,24,25,31,32} alkene hydrogenation,^33,34 and alkene hydroamination²⁹ or hydroazolation.^35,36

Figure 1.

A. Phosphoranyl radical formation and scission pathways in photoredox-catalyzed reactions, including selected nucleophiles reported for phosphoranyl radical-mediated transformations. B. Previous studies of scission selectivity and kinetics, and representative transition states for scission selectivity of the same nucleophilic phosphoranyl radical substituent. C. This work – an experimental and data-driven study of phosphoranyl radical scission selectivity in photoredox-mediated reactions, including understanding and development of phosphine-driven selectivity control, kinetic origins of selectivity, and a predictive selectivity model.

While phosphoranyl radical-mediated reactions have proven synthetic potential, there remains a general lack of understanding surrounding control of selectivity in these transformations, specifically regarding the scission processes through which organic radical intermediates are generated from the phosphoranyl radical (3) species (Fig. 1A). In these photoredox catalyzed reactions, P(III) phosphine 1 is oxidized to a phosphine radical cation (2), which is trapped through addition of a nucleophile to form P(IV) phosphoranyl radical 3.^16,17 To form downstream radicals, intermediate 3 can fragment via α-scission, wherein the P‒X bond undergoes homolytic cleavage to form a X-centered radical on the attacking nucleophile (4) and phosphine 1, or β-scission, wherein the X‒Y bond of the nucleophile homolytically cleaves to generate a Y-centered radical (5) and a P(V) byproduct (6) (Fig. 1A).¹⁶ The ability of phosphoranyl radicals (3) to fragment through either α- or β-scission provides the potential to develop tunable reactivity, wherein divergent transformations can be achieved from the same starting materials based on control of the reaction conditions. While this divergent reactivity would be highly desirable, there is no known phosphoranyl radical-mediated method to activate the same nucleophile for different reaction outcomes based on the choice of phosphine reagent.

To rationally develop new transformations that achieve tunable selectivity through α- or β-scission, a comprehensive understanding of selectivity determination in phosphoranyl radical mediated reactions is required. Prior studies of scission selectivity have focused on phosphoranyl radicals generated through addition of a radical (i.e. X–Y, Fig. 1B) to a P(III) center (1),^12–14,30 inherently limiting study of α-scission of the incoming group, as α-scission is the microscopic reverse of radical addition, making the origin of any product arising directly from the incoming radical unclear (Fig. 1B). As such, early studies focused on comparing scission of two differing groups in a phosphoranyl radical intermediate (3): β-scission of the incoming radical (–O–R group in 7), versus α-scission of one of the phosphine substituents (–X in 7).^37–41 These studies, conducted through both EPR spectroscopy and analysis of phosphine products, demonstrated that relative bond strengths of the two groups involved in scission were the primary driver of selectivity.^{12–14,37–41} This conclusion was later supported by work from Bietti and coworkers, who reported β-scission kinetics for a variety of alkoxy radicals reacting with trialkyl phosphites (via 8), and determined that the rate of β-scission was influenced primarily by the stability of the resulting radical.⁴² A recent study from Schmidt and coworkers similarly explored kinetics for sulfur atom transfer (SAT) to phosphines from thiyl radicals, a process involving β-scission, and found that the relative rate of sulfur atom transfer (SAT) to the phosphine was influenced by phosphine identity, with phosphines with increased electron-donating inductive effects favoring faster rates of SAT.⁴³ While insights regarding the influence of bond strength on α-scission have been used to design a hydroalkylation reaction,⁴⁴ prior mechanistic studies primarily focus on reaction outcomes with respect to phosphine-based species, and as such are not generally transferable to selectivity in synthetic methods, where control over both scission processes of the incoming nucleophile (9 vs 10, Fig 1B) is desired and additional elementary steps leading to product formation could influence selectivity outcomes.

To enable development of tunable selectivity for phosphoranyl radical-mediated transformations, a comprehensive understanding of how phosphine structure influences selectivity under synthetic conditions is required. Given our groups’ long-standing interest in using data science to understand and predict reaction outcomes,^45–47 we envisioned that a data-driven approach leveraging existing phosphine DFT feature datasets (kraken)⁴⁸ and photoredox-mediated, synthetically relevant, conditions would allow us to understand and predict scission selectivity as a function of phosphine structure, enabling rational design of scission-mediated reactions. Here we report a combination of experimental, computational, and machine learning-driven studies that uncover key features controlling scission selectivity for sulfonamide-derived phosphoranyl radicals (Fig. 1C). These insights, which encompass the impacts of phosphine structure, reaction conditions, and alkene identity, provide a new understanding of scission selectivity, which we expect will enable the development of new and/or improved reactions that leverage phosphoranyl radical scission processes.

RESULTS AND DISCUSSION

Selection and optimization of a model system:

We sought to develop a model system that could read-out scission selectivity with a measurable reaction outcome, such as the ratio of α-scission product (product A) and β-scission product (product B), while being synthetically tractable and suitable for collecting a dataset of approximately 50 data points to enable predictive modeling.

As a starting point, we examined phosphoranyl radical mediated reactions of sulfonamide derivatives with varied selectivity outcomes. Previous work from the Doyle group demonstrated that primary sulfonamides (11) can undergo hydroamination reactions of alkenes (12) to form anti-Markovnikov products (13A) via α-scission of phosphoranyl radical intermediates derived from catalytic PCy₃ (Fig. 2A).²⁹ Phosphoranyl radicals derived from N-amidopyridinium salts of secondary sulfonamides (14) have also been shown by Hong and coworkers to undergo β-scission, enabling a variety of transformations including alkene hydrosulfonylation (13B).³⁰ Studies of selectivity for a related alkene difunctionalized reaction found that electron-poor triaryl phosphines with smaller cone angles led to higher yields, attributed to increased radical delocalization in the β-scission transition state.³⁰ However, since the mechanism of this reaction proceeds through addition of an N-centered radical (NCR) arising from fragmentation of the N‒N bond of 14 to the phosphine (analogous to the prior study design shown in Fig. 1B), it is unclear if any NCR-derived products arise from α-scission, precluding conclusions about the influence of the phosphine on both α- and β-scission processes.

Figure 2.

Identification and optimization of a model system. A. Prior art for phosphoranyl radical activation of sulfonamide derivatives via α- and β-scission. B. Design of a model system for studying scission selectivity. C. Optimization of a model system and validation with varied phosphines. Table entries run under optimal conditions unless otherwise noted. ^a 5 mol% PCy₃, otherwise conducted under optimal Doyle conditions shown in Fig. 2A; ^b with 1.8 equiv. P(4-F-Ph)₃, otherwise conducted under optimal Doyle conditions shown in Fig. 2A. Yield reported for 23B only. ^c 10 mol% TRIP–SH, 23–24 h. Ir1, [Ir(dFMeppy)₂(dtbbpy)]PF₆; TRIP–SH, 2,4,6-triisopropylbenzenethiol.

Based on these prior studies, we envisioned a model reaction wherein a phosphoranyl radical 15 would be generated from 1 and 11 under photoredox conditions without requiring a radical precursor that could interfere with reaction selectivity outcomes (Fig. 2B). Phosphoranyl radical 15 would then fragment via α- (16) or β-scission (17), generating N-centered radical (NCR) 18 or S-centered radical (SCR) 19 respectively. From this point, addition of either radical to alkene 12 would form intermediate 20, ultimately providing hydroamination product 13A (product A) or hydrosulfonylation product 13B (product B) following hydrogen atom transfer (HAT). Assuming the two radicals arising from scission, 18 and 19, are captured for the formation of their respective products, we hypothesized that the ratio of reaction products B:A (13B:13A) could be used to estimate scission selectivity (17:16).

While the proposed mechanism for the formation of 13A is precedented in the Doyle hydroamination methodology,²⁹ the formation of 13B via the proposed pathway has not been previously reported. As such, we sought to determine if β-scission reactivity for the formation of 13B was possible under photoredox-mediated conditions. Performing the desired reaction of aryl sulfonamide 21 and 1-hexene (22) under modified hydroamination conditions (Fig. 2C, entry 1) afforded a high yield of 23A, and a small amount of 23B (~5% yield), indicating both α-scission and β-scission of primary sulfonamides is possible in the same pot. Seeking to optimize formation of 23B, we identified a triaryl phosphine (P(4-F-Ph)₃) with precedent for β-scission reactivity,³⁰ and found that when used in excess (1.8 equiv.), this phosphine led to an increased 11% yield of 23B under photoredox conditions (entry 2). To balance the requirements of β-scission, which is stoichiometric in phosphine, and α-scission, which has been reported to perform less well under higher phosphine loadings,²⁹ we elected to use phosphine as the limiting reagent, which, after optimization, ultimately raised the yield of 23B to 47% while still allowing formation of measurable 23A (B/A = 16.1, entry 3) (see SI for optimization details). To validate this model system for the study of scission selectivity with respect to phosphine identity, we performed the reaction with a few phosphines spanning from triaryl phosphine P(4-F-Ph)₃ (entry 3), to trialkyl (PCy₃, entry 6), and found that the B:A ratio varied, indicating selectivity is controlled by phosphine identity. Furthermore, moderate yields of 23A (40%) were still accessible with PCy₃ under our optimized conditions, highlighting the successful balance of conditions favoring α- and β-scission processes.

Mechanistic Studies of the Model System:

With our model system in hand, we sought to understand how the experimentally observed B:A selectivity related to the reaction energy profile of β- and α-scission. We began our investigations with density functional theory (DFT) calculations of the free energy surface for P(4-F-Ph)₃ (P1), a B-selective phosphine from our preliminary screen. Starting from the phosphoranyl radical (P1-rad), we computed transition states and related intermediates for β-scission (24), S-centered radical (SCR) (25) addition to hexene (22) (not shown), and N-centered radical (NCR) (28) addition to hexene (transition state 29) (Fig. 3A). While the transition state structure for α-scission (27) could not be located, we expect that this transition state is kinetically favorable.^{12–14,29,30} Interestingly, we find that while β-scission (24) is an irreversible selectivity determining transition state for the formation of product 23B, based on an assumed low barrier for α-scission, the selectivity determining step for the formation of 23A is addition of the NCR (28) to hexene (transition state 29) (see SI for further discussion of selectivity-versus rate-determining steps in this system). While the computed energy surface qualitatively aligns with our experimental results, the computed difference in energy (ΔΔG^‡) between 24 and 29 is slightly higher than we would expect based on the measured B:A selectivity (see SI for details). Additionally, an energy surface computed for an A-selective phosphine, PAd₂Bn (P4), showed the same kinetic relationship between the transition states (see SI section 4.2.2). These results suggest that selectivity is controlled not by a simple competition between α- and β-scission (‘intrinsic’ selectivity), but by a Curtin–Hammett-like^49–51 scenario wherein an α-scission equilibrium between P1-rad and P1 + 28 results in selectivity determination through a competition between β-scission (24) and NCR-addition (29), as described by equation 1. Importantly, the Curtin–Hammett scenario controlling this reaction involves competition between two differing elementary steps, meaning selectivity is influenced by the relative concentrations of phosphine (P) and alkene (Alk) as well as the ΔΔG^‡ between the rate determining steps.

Figure 3.

Computational and experimental mechanistic studies of the model system. A. DFT-computed free energy surface ((U)ωB97X-D/def2-TZVP (SMD, toluene)//(U)ωB97X-D/def2-SVP). Ar₁ = 4-F-Ph; Ar₂ = 4-t-Bu-Ph. Radicals in transition state structures shown in their ultimate position for clarity. B. Experimental comparison of B:A selectivity over a range of hexene concentrations. For this plot [B]/[A] represents 23B/23A, [Alk] represents the hexene concentration. C. Kinetic selectivity relationship under experimental and saturation conditions.

$$\frac{[\mathbf{\text{B}}]}{[\mathbf{\text{A}}]}\approx \frac{\frac{d[\mathbf{\text{B}}]}{\mathit{\text{dt}}}}{\frac{d[\mathbf{\text{A}}]}{\mathit{\text{dt}}}}=\left(e^{\frac{-\Delta \Delta G^{‡}}{\mathit{\text{RT}}}}\right)\frac{[\mathbf{\text{P}}]}{[\mathbf{\text{Alk}}]}$$ (equation 1)

To experimentally validate this kinetic scenario, we tested varied [P]:[Alk] ratios in the model reaction with a set of phosphines with differing selectivity (P1–P4) (Fig. 3B). Based on equation 1, increasing alkene concentration ([Alk]) would decrease the phosphine to alkene ratio, leading to a shift in selectivity towards product A. The relationship between B:A and [P]:[Alk] should be linear, with a slope dependent on ΔΔG^‡ of the selectivity determining steps. As anticipated, increasing the concentration of hexene in our model reaction leads to decreased β-scission selectivity for all phosphines, supporting the existence of a low energy α-scission transition state and the proposed Curtin-Hammett scenario (Fig. 3B). Interestingly, while the two B-selective phosphines (P1, P2) display a good linear relationship, the two A-selective phosphines (P3, P4) exhibit saturation behavior at increased hexene concentrations, with the concentration of hexene required to reach saturation being phosphine dependent. The initiation of saturation behavior can also be observed at high equivalents of hexene for the moderately B-selective phosphine P2. Notably, at these high hexene concentrations, this phosphine (P2) exhibits slight selectivity for product 23A, despite being initially B-selective.

These results reveal an unexpected selectivity scenario with synthetic implications. Rather than arising from direct competition between α- and β-scission (‘intrinsic’ selectivity), selectivity arises from a Curtin–Hammett-like scenario. Experimental results are consistent with this kinetic relationship, with a linear relationship between [P]:[Alk] and selectivity (B:A) at synthetically relevant concentrations of hexene. As the Curtin-Hammett regime is operative under these conditions, we consider this as the ‘experimental’ selectivity, determined by the ΔΔG^‡ of 24 and 29 and reaction stoichiometry (equation 1) (Fig. 3C). However, at increased alkene concentrations, we observe the breakdown of this linear Curtin-Hammett relationship, and the appearance of saturation behavior. As alkene equivalents increase, the rate of NCR addition (29) increases to become closer to the rate of α-scission (27), violating the pre-equilibrium assumption of the Curtin-Hammett scenario, and causing selectivity determination to deviate from the derived equation 1. Given the observation that the onset of saturation is impacted by both phosphine identity and alkene identity (See SI Section 2.3), saturation represents a move towards an ‘intrinsic’ selectivity regime but does not represent a direct competition between the two scission steps. Rather, the onset of saturation behavior is likely related to rates of α-scission and NCR addition, and possibly by the reaction to form 23A approaching diffusion-limited rates. Interestingly, the saturation-dependent selectivity limit is generally reached with moderate to high A-selectivity, though this value is phosphine dependent.

Comparing phosphines P1–P4 under this lens, we can assess selectivity under the ‘experimental’ selectivity regime as a function of the transition state barrier for β-scission (24), rather than as a competition of β-scission and α-scission. Since phosphine identity should not influence the NCR-addition transition state barrier, changes in selectivity between phosphines will be driven by changes in β-scission transition state energy. With respect to the saturation-dependent selectivity limit, these studies reveal that A-selective phosphines generally reach saturation with lower B/A values than B-selective phosphines, despite saturation generally occurring in an A-selective regime. Furthermore, phosphines that are B-selective under experimental conditions have lower selectivity for product A as they approach saturation, with highly A-selective phosphine P4 reaching saturation at a B:A ratio of ~0.05:1.0, moderately A-selective phosphine P3 reaching saturation at a B:A ratio of ~0.3:1.0, and B-selective phosphine P2 beginning to approach saturation at a B:A ratio of ~0.8:1.0. These data indicate that while ‘experimental’ selectivity differs considerably from selectivity under saturation conditions, these values seem to be related with respect to the relative selectivity within the phosphine series.

Taken together, these studies provide essential insight into the origin of phosphoranyl radical scission selectivity under synthetic conditions. Rather than being controlled by direct competition between β-scission and α-scission, the ratio of products B and A is determined through a complex interplay of transition state barriers and reaction stoichiometry. Importantly, for all cases where saturation is reached, the saturation-dependent selectivity limit favors α-scission and product A, indicating that β-scission (B) selective reactions require a balance of low energy β-scission transition states and slower NCR radical addition.

Factors Influencing Reaction Selectivity:

Based on the above, B:A selectivity in these phosphoranyl radical-mediated reactions is influenced by three major factors:

1. Reaction Stoichiometry. Specifically, the ratio of phosphine and alkene in the reaction, where increasing the concentration of the alkene relative to the phosphine shifts selectivity towards α-scission product A.

2. Phosphine identity. Wherein modulation of the β-scission transition state based on phosphine structure is expected to control scission selectivity.

3. Alkene identity. Wherein the identity of the alkene will influence selectivity via changes in the energy of the NCR addition transition state.

While equation 1 provides the relationship between reaction conditions and selectivity, the impact of phosphine structure and alkene identity remained unknown.

Modeling Selectivity as a Function of Phosphine Structure:

Given the challenges encountered using DFT to computationally study these systems (see SI Section 4 for details), we hypothesized that a machine learning-driven approach would be well suited to study phosphine structure-selectivity relationships. Through the generation of predictive models, machine learning can uncover underlying trends in a dataset with respect to phosphine structure and selectivity, facilitating in silico screening of phosphine candidates for future applications.

To train predictive models, we envisioned collecting selectivity data using a representative dataset of monophosphines. This data would then be modeled as a function of phosphine structural features extracted from the kraken phosphine database which contains DFT-level stereoelectronic properties of >1600 monophosphines.⁴⁸ Based on the ‘experimental’ selectivity regime, the rate of β-scission will change with phosphine identity while the rate of NCR-addition remains constant, resulting in a dataset representative of the relationship between phosphine structure and β-scission transition state energy.

To select a dataset of phosphines covering the breadth of chemical space, we employed a dimensionality reduction and clustering approach, wherein the kraken PC₃-type phosphine space was grouped based on similarity.⁴⁸ We utilized all DFT-computed features in kraken to conduct principal component analysis (PCA) and clustering of the points with a k-nearest neighbors (k-NN) algorithm (Fig. 4). Points near to the center of each cluster were selected to ensure a dataset representative of chemical space.^48,52 We selected 24 phosphines as an initial dataset, and found that in the model reaction, 20/24 gave measurable B:A selectivity. The majority of these 20 phosphines were selective for product B, leaving only 3 A-selective phosphines. To train effective machine learning models, a more well-distributed dataset was required. However, examination of the reaction outcomes in the kraken chemical space revealed that A-selective phosphines were not located in the same vicinity, making identification of additional A-selective phosphines difficult (Fig. 4, step 1). To overcome this challenge, we used a ‘chemist-in-the-loop’ active learning strategy,⁵³ wherein the initial 24-point dataset was analyzed using a combination of data science approaches (step 2) to identify features most important to selectivity outcomes (step 3). These features were then used to generate an alternative chemical space map defined by only four of the kraken features (Fig. 4, step 4). By limiting the features to only those identified to be related to selectivity, the refined selectivity space is better divided with respect to selectivity outcomes. As such, the A-selective phosphines from the first dataset were located in the same region, allowing oversampling of the nearby clusters where phosphines were expected to provide A-selectivity. Collecting additional data for this second round of phosphine selections resulted in a more balanced selectivity distribution (~40% A-selective phosphines), providing a dataset representative of both phosphine chemical space and reactivity space (Fig. 4, right plot). Collecting data for 7 additional phosphines for model validation completed our selectivity dataset (total 46 phosphines, ~40% A-selective).

Figure 4.

Workflow for dataset selection from phosphine chemical space. Step 1. Experimental results from the first dataset on kraken chemical space (left). Steps 2. and 3. preliminary modeling and feature selection. Step 4. Experimental results from both datasets on refined selectivity 4-feature chemical space. Definitions: vmin, minimum of the electrostatic potential at P (Boltzmann average); nbo_bds, average P–C antibonding orbital energy (Boltzmann average); vbur, buried volume (radius 3.5 Å, Boltzmann average); E_reduction, ionization potential (Boltzmann average); pyr, minimum value of pyramidalization.

Using experimental B:A ratios from our dataset (as ln(B/A)) and phosphine features from the kraken database, we trained models for classification and regression tasks and found that a 4-feature linear regression model was the highest performing for prediction (train MAE = 0.437, R² = 0.889; validation MAE = 0.429, R² = 0.930; external test MAE = 0.802, R² = 0.633) (Fig. 5A). The features in the final model describe a combination of stereo- and electronic phosphine properties, with the phosphine HOMO energy (fmo_e_homo) and P‒C antibonding orbital energy (nbo_bds) representing electronic features, pyramidalization (pyr) representing a stereoelectronic feature, and buried volume (%vbur) representing a steric feature. These features were also observed in highly ranked models throughout our modeling efforts, providing additional confidence that they represent true structure-selectivity relationships. The successful use of kraken features of the parent phosphine demonstrates the utility of the kraken database for predictive modelling and provides the opportunity to use this model for prediction on all PC₃-phosphines in the kraken dataset without any additional calculations, enabling future applications.

Figure 5.

A. 4-Feature linear regression model for ln(B/A) prediction. B/A represents 23B/23A. B. Feature interpretation showing trends with increasing B-selectivity. C. DFT study of a phosphoranyl radical subset and relation of phosphoranyl radical features to selectivity. Phosphoranyl radical structures derived from model sulfonamide 4-tert-butylbenzene sulfonamide and the relevant phosphine. DFT: ((U)ωB97X-D/def2-TZVP (SMD, toluene)//(U)ωB97X-D/def2-SVP). Definitions: fmo_e_homo, energy of the HOMO (Boltzmann average); nbo_bds, average P–C antibonding orbital energy (Boltzmann average); pyr, minimum value of pyramidalization; vbur, buried volume (radius 3.5 Å, Boltzmann average); MAE, mean absolute error.

Given the interpretability of linear regression models and the chemically meaningful features in the kraken database, we hypothesized that the features included in the predictive MLR model could also be used to understand structure-selectivity relationships, despite not representing transition state or phosphoranyl radical intermediate structures. Prior to feature interpretation, we analyzed the collinearity among the four descriptors in the MLR model and found that although none of the features were strongly correlated, there was a moderate correlation between pyramidalization (pyr) and buried volume (vbur) (R² = 0.52), arising due to their shared sensitivity to steric environments. Analyzing each feature, we find that the B:A ratio increases as HOMO energy decreases, and as P–C antibonding orbital energy and pyramidalization increase (Figure 5B). Based on the collinearity analysis, we hypothesize that pyramidalization is capturing the dominant stereoelectronic trend associated with phosphine geometry, including steric influences, while buried volume provides additional resolution for phosphines where pyramidalization is similar despite steric differences (see SI section 3.2.2). Furthermore, energy decomposition analysis indicates that the buried volume feature correlates with dispersion energy, suggesting that it may serve as a proxy for capturing non-covalent interactions that aid in the stabilization of the phosphoranyl radical intermediates. Thus, the buried volume feature may be functioning as more of a context-dependent term in the model than an independent monotonic descriptor of selectivity. Overall, our analysis suggests that more electron-poor, and pyramidalized phosphines favor β-scission and product 23B.

While the kraken structural features cannot be directly related to the phosphoranyl radical or transition state structure, we expect that these general properties transfer from the parent phosphine structures in kraken to represent phosphoranyl radicals that can better stabilize the β-scission transition state, leading to higher B-selectivity overall. To examine this relationship, we used DFT to compute optimized β-scission transition states for a set of 7 phosphines (P1, P2, P4–P8) and extracted a set of atomic and geometric features (i.e., atomic charge, spin density, bond length) (Fig. 5C). These transition state features, along with features from a phosphoranyl radical for each selected phosphine, were used to assess how the structures of these intermediates relate to selectivity and to the properties identified from the kraken features.

Overall, phosphoranyl radical features provided the strongest correlations with selectivity (ln(B/A)). Of these features, the length of the P–N bond in the phosphoranyl radical has the strongest correlation to B:A selectivity, where phosphoranyl radicals with short P–N bonds favor B-selectivity. This feature could be represented in the linear regression model by both steric and electronic kraken features, as the P–N bond length is influenced by the favored phosphoranyl radical geometry, which is controlled by a variety of factors. For this small phosphine subset, it was found that the B-selective phosphines have accessible and low-energy tetrahedral geometries (P1, P2, P5), compared to unselective (P6) and A-selective phosphines (P4, P7, P8), for which only trigonal bipyramidal (TBP) structures were located computationally (Fig. 5C). These differences in geometry are consistent with the correlated phosphoranyl radical features, particularly P‒N bond length, with TBP phosphoranyl radicals generally having longer P‒N bonds than tetrahedral phosphoranyl radicals.

Selectivity (ln(B/A)) is also correlated with the charge on the P-atom, aligned with the inclusion of the kraken fmo_e_homo feature in the linear regression model, and indicating that phosphoranyl radicals with more electron-poor phosphine atoms favor formation of product B. Additionally, the spin on the P-atom in the phosphoranyl radical is moderately correlated to B-selectivity (Spearman ranked correlation = ‒0.64), with phosphoranyl radicals that have greater delocalization of the unpaired spin having greater selectivity for product B. This result is consistent with the conclusions of Hong and coworkers,³⁰ demonstrating the ability of our approach to identify important features describing selectivity for both known and previously unknown features influencing selectivity control in these complex systems.

Based on the above correlations, we conclude that more electron poor phosphoranyl radicals and phosphoranyl radicals with shorter P–N bonds favor B-selectivity through formation of more stabilized β-scission transition states. As the β-scission transition states tend to have shorter P–N bonds and a more tetrahedral-like structure than the parent phosphoranyl radicals, phosphoranyl radicals which can stabilize these types of geometries can be expected to result in more facile β-scission. While the relationships between phosphoranyl radical structure and features and the kraken features used for modeling cannot be comprehensively characterized with our small dataset, this analysis reveals key connections between each feature set, and that structural information from both the kraken phosphines and the phosphoranyl radicals can provide handles for understanding and controlling B:A selectivity.

Impact of Alkene and Sulfonamide Identity on Reactivity:

Having determined how phosphine structure controls selectivity, we turned to exploring the impact of alkene and sulfonamide identity on the B:A ratio. While the sulfonamide starting material is involved in both selectivity determining steps, the alkene was expected to exclusively impact the rate of NCR addition (29, Fig. 3A). Preliminary computational assessment of the model reaction with various alkenes and aryl-substituted sulfonamides revealed that changes in sulfonamide identity had minimal impact on ΔΔG^‡, while changes in alkene identity changed the magnitude of ΔΔG^‡ substantially (see SI for details). Based on these computational studies, we hypothesized that changing the alkene identity would impact the magnitude of the B:A selectivity for a specific phosphine, but that overall, the relative order of selectivity among phosphines would remain the same.

Based on our preliminary calculations, we selected 5 alkenes and 3 sulfonamides with which to measure experimental selectivity for a subset of phosphines selected to represent the larger dataset (Fig. 6A). The identity of the aryl-substituted sulfonamide had a small impact on selectivity for two of the sulfonamides tested (21, 31), while sulfonamide 32 showed a slight difference in selectivity profile, with phosphines P1 and P4 showing complete selectivity for product B and product A respectively. Though product yields were low for sulfonamide 32, this interesting example represents a complete switch in selectivity based solely on phosphine identity. Notably, our model reaction was found to be limited in scope to aryl-substituted sulfonamides, precluding comparison of more diverse sulfonamides that may be expected to have a larger impact on selectivity.

Figure 6.

A. Relationship of sulfonamide and alkene identity and selectivity outcomes. Ar₁ = 4-tert-butylbenzene. B. Computed features of the alkenes used in this study. DFT features computed using: BDFE, (U)ωB97X-D/def2-TZVP//(U)ωB97X-D/def2-SVP; Hirshfeld charge, (U)ωB97X-D/def2-SVP. BDFE, bond dissociation free energy.

As anticipated, the alkene identity has a large impact on selectivity, with many alkenes (33–35) shifting the reaction toward A-selectivity. Specifically, for 1,1-disubstituted alkene 33 and vinyl ether 34, even the most B-selective phosphine P1 achieves only very modest selectivity towards product B (B/A = 1.3 or 1.2 respectively for phosphine P1). For the vinyl pyrrolidine substrate 35, no phosphines favored product B, with the most B-selective phosphine P1 having a modest favorability towards product A (B/A = 0.8). For styrene (36), the selectivity values were more comparable to those of hexene, although phosphine P3, which is generally moderately A-selective, slightly favored product B (B/A = 1.2).⁵⁴ Crucially, in all cases, the relative selectivity ranking for the four phosphines remains the same, indicating that while alkene identity is an important factor in determining selectivity, the selectivity rankings provided by the MLR model are still suitable for predicting relative phosphine selectivity for substrates beyond sulfonamide 21 and hexene 22.

Initial assessment of the five alkenes used for this study reveals that electron rich alkenes shift selectivity towards A-selective reaction outcomes (33–35). To provide a more quantitative assessment of the alkenes, we extracted features related to the alkene carbon atoms (C1, terminal carbon, C2, internal carbon) from DFT-optimized alkene structures (Fig. 6B). We found that the features that best categorize the five alkenes with respect to selectivity are the molecular dipole moment and the bond dipole moment for the C1=C2 bond, with the more B-selective alkenes 22 and 36 having a lower bond dipole moment than more A-selective alkenes 33–35. For the more polarized alkenes, we conclude that selectivity-determining step for the formation of product A, NCR-addition, is favored due to the increased polarization and electron density on the terminal carbon (C1) of the alkene. Visual assessment of the HOMO orbitals of these alkenes (Fig. 6B), also demonstrates that for the nonaryl containing alkenes, the more A-selective alkenes 33–35 have less symmetry in their π-orbitals than B-selective alkene 22.

We also expected that generation of a more stabilized radical (relative to that for hexene) during NCR-addition could lead to an increase in A-selectivity through analogous stabilization of the partial spin in the NCR addition transition state. To probe this, we computed the bond dissociation free energy (BDFE) for a C–H bond on C2 of an alkane model system corresponding to each alkene substrate. As the radical corresponding to the dissociation of the specified C–H bond is in the same position as the radical in the post NCR-addition intermediate (30), we used the computed BDFE values as an estimate of stabilization of this species, which we expect to correspond to potential stabilization of the NCR addition transition state (29). For the aliphatic alkenes (22, 33–35), more B-selective alkenes correspond to lower BDFE values in their respective alkanes. While ethyl benzene (representing styrene 36) has the lowest BDFE, it generally has higher selectivity for product B, similar to hexene. Given potential differences between aliphatic and conjugated alkenes, we cannot rule out that radical stability has influence on selectivity along with bond polarization in these systems.

Control of Scission Selectivity:

With our new understanding of the impact of phosphine structure and alkene identity on selectivity, the framework for selectivity control in these phosphoranyl radical-mediated transformations allows for general guidelines to be discussed:

1. Reaction Stoichiometry. Selectivity is influenced by the ratio of phosphine and alkene in the reaction, where increasing the ratio of [P]/[Alk] increases the B:A ratio.

2. Phosphine identity. Phosphine identity has a strong influence over selectivity, with electron poor, pyramidalized phosphines favoring formation of product B and higher B:A ratios. While overall selectivity depends on a variety of factors, our studies indicate that the selectivity ranking for phosphines remains the same across a series of alkenes and reaction conditions. Predicted selectivity rankings of kraken phosphines⁴⁸ are available as part of an interactive Jupyter notebook included with the supporting information for this work.

3. Alkene identity. For a given phosphine, alkene identity can influence selectivity outcomes, with more polarized alkenes providing greater selectivity for product A and lower B:A ratios. Overall, alkenes for which N-centered radical addition is more favored should be expected to have increased rates of formation for product A, and decreased B:A ratios compared to simple unactivated alkenes such as hexene.

Development of a hydrosulfonylation reaction from primary sulfonamides:

With this framework for selectivity outcomes for phosphoranyl radical-mediated reactions, we aimed to develop a new hydrosulfonylation reaction of alkenes with primary sulfonamides, which would represent the first example of switchable scission reactivity outcomes for a single nucleophile class based on phosphine identity. Using our MLR model, we selected phosphines with high predicted B/A ratios and screened them under reaction conditions analogous to those for hydroamination. We found that of the phosphines evaluated, P(4-F-Ph)₃ (P1) represented the best combination of yield and selectivity; some phosphines that were predicted to be more selective provided no observable yield in this transformation, likely due to increased phosphine oxidation potentials. While this shows that our selectivity prediction model cannot account for reaction yield, it demonstrates how the model can be used to select a targeted screen of phosphines for reaction discovery.

Aiming to develop a method for selective formation of the hydrosulfonylation product (B) for a variety of alkenes, we recognized that we would need to harness both phosphine identity and reaction stoichiometry to achieve B-selectivity for more electron rich alkenes such as 34 and 35. As such, optimization efforts focused on changing the phosphine and alkene ratio for improved B-selectivity. Ultimately, we found that combining the B-selective phosphine P1 with reaction conditions in which the alkene is the limiting reagent led to good yield and excellent B:A selectivity for our model system forming 23B (64% isolated yield, Fig. 7). Importantly, the development of this transformation under photoredox-mediated conditions enables the direct use of primary sulfonamides for hydrosulfonylation and demonstrates a complementary transformation to the previously reported phosphine-catalyzed hydroamination reaction:²⁹ divergent reactivity is observed from the same starting materials dependent on phosphine identity.

Figure 7.

Scope of the developed hydrosulfonylation reaction of primary sulfonamides and comparison to previously reported hydroamination reactions of the same substrates.²⁹ The results from the hydroamination reaction are reported for 1.0 equiv. sulfonamide and 1.0 equiv. alkene unless otherwise noted.²⁹ The hydrosulfonylation reaction was performed with 1.0 equiv. alkene and 1.5 equiv. sulfonamide. Yields reported as isolated yield (assay yield). Assay yield determined by GC-FID using 1,3,5-trimethoxybenzene as standard unless otherwise noted. B/A determined by GC-FID for examples where GC-FID was used to determine assay yield. ^a Assay yield determined by ¹H NMR; ^b 5.0 mol% PCy₃, 2.0 equiv. alkene; ^c (B/A) determined by ¹H NMR; ^d Published yields;^{29 e} 4.0 mol% PCy₃. Ir1, [Ir(dFMeppy)₂(dtbbpy)]PF₆; TRIP–SH, 2,4,6-triisopropylbenzenethiol; Boc, tert-butoxycarbonyl.

With optimized conditions in hand, the hydrosulfonylation reaction was demonstrated for a variety of alkenes and primary sulfonamides (Fig. 7). For the alkene scope, products of the new hydrosulfonylation method (23B, 37B–41B) were compared to reported products from the corresponding hydroamination method (23A, 37A–41A).²⁹ Comparing the model hydrosulfonylation substrate, 23B, to its hydroamination counterpart 23A, it can be seen that divergent reactivity is now accessible with good yields (64% 23B, 75% 23A) and high selectivity (17.9:1 B:A for 23B; 0.07:1 B:A for 23A) in each case. Interestingly, across the scope of alkenes, yields followed similar trends for both products, with 22, 33, and 34 being well-performing alkenes in both methods (products 23, 37, 39), and 35 and 36 being lower yielding alkenes for both products (40, 41). This comparison showcases sulfonamides as the first nucleophile with tunable selectivity in phosphoranyl radical mediated reactions.

Among the hydrosulfonylation products, we found that while hexene provides high B:A selectivity of 17.9:1 for 23B, 1,1-disubstituted methylene cyclopropane (product 37B), ethyl vinyl ether (product 39B), and vinyl pyrrolidone (product 40B) provide lower B:A selectivities of 4.3:1, 3.0:1, and 3.4:1 respectively. These results are consistent with our earlier alkene screen, however, the increased selectivity imparted by our carefully optimized conditions allows for isolation of these products in yields of 13–58%. This highlights the importance of understanding the origin of selectivity in these phosphoranyl radical mediated processes for the design of improved synthetic methodologies. A 1,1-disubstituted alkene containing a piperidine scaffold was also tolerated in the reaction, providing 38B in 41% isolated yield, and a conjugated alkene, styrene, provided excellent selectivity of >50:1, albeit with a decreased yield of 23% for product 41B.

A variety of para-substituted sulfonamides also were suitable for this transformation. The model para-tert-butyl sulfonamide (forming 23B) and para-fluoro-sulfonamide (forming 43B) provided similar yields (64% and 58% respectively) and selectivities (17.9:1 and 12.4:1 respectively), aligned with their performance under screening conditions. Product 42B, arising from para-methyl sulfonamide was formed in a slightly decreased yield of 47%. The more electron-rich para-methoxy-sulfonamide yielded the hydrosulfonylation product 44B with excellent selectivity (35.8:1), aligned with the complete selectivity for product 44B seen under the screening conditions. The lower yield of 44B is partially attributed to a lack of solubility for this sulfonamide under the reaction conditions. Finally, the small molecule pain reliever Celecoxib was a suitable substrate for this reaction, providing product 45B in 20% yield, demonstrating the potential of using this transformation for small-molecule drug modification.

CONCLUSIONS

This work reports a comprehensive study of scission selectivity for phosphoranyl radicals with sulfonamide nucleophiles. Experimental, computational, and data-driven study of a photoredox- and phosphine-mediated model reaction reveals the kinetic origin of selectivity under experimental and saturation conditions, demonstrating that a Curtin–Hammett relationship controls selectivity between β-scission and α-scission products. Based on this relationship, we establish how reaction stoichiometry, phosphine stereoelectronic features, and alkene identity control selectivity under experimental conditions, providing essential insights for the development of new phosphoranyl radical-mediated transformations. Furthermore, we report a practical multivariable linear regression model for predicting phosphines for β-scission or α-scission selective reactions and demonstrate a hydrosulfonylation reaction of primary sulfonamides, making these nucleophiles the first reported to react with phosphine-controlled tunable selectivity.

Overall, this work expands our understanding of phosphoranyl radical scission mechanisms, which we believe will enable further rational design of phosphoranyl radical mediated transformations.

Supplementary Material

Supporting Information: Experimental procedures, description of workflow and dataset collection using kraken, description of machine learning procedures and outcomes, computational details, optimization details, characterization for the hydrosulfonylation reaction. This material is available free of charge via the Internet at http://pubs.acs.org.

Interactive phosphine selectivity predictions: ‘phosphine_modeling_rev.zip’ and at: https://github.com/SigmanGroup/Phosphoranyl-Radical-Scission

DFT computed geometries: ‘DFT_xyz_coords.zip’

Computed DFT energies: ‘DFT_energies_SI.xlsx’

EDA computed energies: ‘EDA_energies.xlsx’

EDA output files: ‘EDA_calculations.zip’

Experimental Selectivity Data: ‘experimental_data_and_kraken_descriptors.xlsx’

DFT computed geometries for the phosphoranyl radical feature set: ‘phosphoranyl_radical_xyz.zip’