Introduction

Gating the release and the transport of molecules underpins a range of essential biological functions, especially in advanced organisms. For instance, exocytosis of neurotransmitters into a synapse is mediated by the fusion of secretory vesicles and porosomes1,2, whereas the membrane potential of a cell is controlled by the gating of ion fluxes through its ion channels3. In enzymes, the access of solvent molecules and substrates to the active site can be restricted through conformational gating4 and salt concentration5. The gating of substrates and products can even serve the purpose of controlling and synchronizing reaction chains, as is the case in ammonia-transferring enzymes6. While these biological processes are often mediated by complex chemical networks, the gating of molecular release is less common in simple synthetic host-guest systems.

Cucurbit[n]urils (CBn, n = 5–8,10) are synthetic macrocycles composed of repeating glycoluril units. These organic macrocyclic compounds possess carbonyl-rich portals and a hydrophobic cavity. They are able to form extremely stable complexes with guests, with binding affinities surpassing that of the avidin-biotin complex7,8. Despite their deceptively simple symmetric structure, cucurbiturils have proved to be extremely versatile molecules. Cucurbiturils have been shown to promote the stabilization of degradation-prone species9 and short-lived electronic states10. In catalysis, CB6 was found to dramatically enhance the 1,3 dipolar cycloaddition between azides and acetylenes, or as it is known today, in situ click chemistry11, while a selective retro-Diels-Alder reaction was demonstrated in the gas phase12. Recently, CB6 has also been suggested to assist the isomerization of xylene13. A complete review of the role of CBn in catalysis can be found elsewhere14. A variety of guests have been reported to form host-guest complexes with CBn with a range of binding affinities15. Building on these successes, CBn have been included in several supramolecular architectures16 such as hydrogels17, micelles18,19, hybrid nanoparticles20,21,22, and nanomachines23,24.

In 2004, Kim leveraged CB5 and CB6 as part of egg yolk L-α-phosphatidylcholine (EYPC) membranes to create artificial selective ion channels able to respond to pH variations25. Additional improvements on the earliest passive ion channels based on cyclodextrins26 involved the design of voltage-, light-, mechano- and ligand-gating mechanisms27. Ligand-based gating, in particular, is especially interesting given that its prevalence in natural systems28 has yet to be matched in artificial constructs. In a bio-inspired example of ligand-gated ion channel, a synthetic chiral receptor embedded in a lipid membrane can allow ion transport to occur across a channel when the addition of a ligand induces a conformational change. The addition of β-cyclodextrin would scavenge the ligand and return the channel to a non-conductive state29.

Previous works have highlighted that the cavity of CB6 can be gated through modulation of pH30 and salt concentrations31 but the gated release of a neutral molecular species from CB7 has not been reported to the best of our knowledge due to its larger portal. Here we design a de novo approach to highlight strong ligands for the portal of CB7 that are able to hinder the entry and exit of the guest from the cavity of the macrocycle. Machine Learning is used to explore the chemical space and complex potential energy surfaces32,33,34. In particular, ligands with a strong binding affinity with CB7’s portal (up to −41.85 kcal mol−1) are able to boost a guest residence time by over 7 orders of magnitude, to 350 s, thanks to π-orbital-driven halogen bonds. We believe that these ligands will stand as a valuable addition to the toolkit of designers of CB7-based channels and hope our de novo design approach will inspire other designs of supramolecular gating systems.

Results and discussion

Overview of design procedure

The general workflow used to design strong ligands for CB7 is illustrated in Fig. 1. To keep computations tractable, a binary host-guest complex composed of adamantanone (ADA) encapsulated inside of CB7 is considered. From preliminary studies of known ligands for the portal of CB7, a correlation between the binding affinity of the ligands with the portal and their ability to prevent the escape of the ADA guest is observed. Further investigations revealed that this correlation is a Linear Free-Energy Relationship (LFER) that can be used as a heuristic to quickly estimate the ‘gating power’, or the ability to prevent a guest from escaping, of a given ligand candidate. Instead of systematically screening a library of molecular compounds, a generative algorithm (a Monte Carlo Tree Search with a recurrent neural network used in the roll-out phase to speed up the exploration, based on ChemTS33) trained on chemical data is set up to suggest ligand candidates with up to 12 non-hydrogen atoms (see the reduction of the model loss during its training in Supplementary Fig. 3). The candidates are quickly evaluated via their computed binding affinity \({{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{1}}})}\), given as:

$$\begin{array}{lll}{{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{1}}})}=G[\,{{\mathrm{ADA}}}\,@{{{\bf{CB7}}}}+\,{{\mathrm{1}}\, {\mathrm{ligand}}}]-G[{{\mathrm{ADA@CB7}}}]\\ \qquad\qquad\quad-\ G[{{\mathrm{1}}\, {\mathrm{ligand}}}],\end{array}$$
(1)

with the CB7 portal thanks to the cheap and accurate GFN2-xTB tight-binding method35. This binding affinity is used to improve the algorithm’s subsequent guesses as part of a fitness score J(S) written as,

$$J(S)={({{{\rm{portal}}}}^{\prime} {{\mathrm{s}}\, {\mathrm{occluded}}\,{\mathrm{fraction}}})}^{2}* {{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{1}}})}$$
(2)
Fig. 1: Overview of the design procedure for CB7 strong ligands.
figure 1

Overview of the design procedure for CB7 strong ligands. In (a), the Monte Carlo Tree Search generative algorithm suggests fragments of ligand candidates in the SMILES format. This fragment of SMILES is expanded in (b) thanks to a recurrent neural network until a complete SMILES candidate is obtained. If the SMILES candidate is deemed valid by the RDKit in (c), a 3D structure of the ligand candidate (CAP) is generated. In (d), the ligand candidate is docked on CB7 with adamantanone (ADA) as an inclusion guest, then the ternary complex is optimized and its binding free energy is computed. In step (d), the fraction of the portal occluded by the ligand is also estimated to obtain a fitness score J(S) (see Supplementary Fig. 6 and Supplementary Methods 2.3 for details). Using the fitness score of the ligand candidate, the tree nodes are updated in (e) and a new iteration of the tree exploration can begin. After several iterations of the a → e procedure, promising candidates are further evaluated in (f) using infrequent metadynamics, the ligand residence time t1/2 is estimated, with stronger ligands expected to show a longer residence time.

Full size image

Eventually, strong ligand candidates resulting from this iterative process and their commercially available derivatives have their average residence time t1/2 computed through biased molecular dynamics and are finally evaluated experimentally. Additional details on the computation of the fitness function is provided in Supplementary Methods 2.3 and illustrated in Supplementary Fig. 6. The structure of the [ADA@CB7] binary complex is illustrated in Supplementary Fig. 5 while the ability of ADA to enter CB7’s cavity is shown in Supplementary Fig. 10 through 1H NMR data. The ADA guest is chosen as it almost completely fills the CB7’s cavity and therefore favors ligand interactions with the CB7’s portals instead of the inner cavity.

Linear free-energy relationship

Linear free-energy relationships (LFER) connect the equilibrium constants and reaction rates of related reactions. As such, a LFER can be seen as a correlation between thermodynamics and kinetics which, while empirical in nature, has found useful applications such as in the Hammet equation36 or the Brønsted catalysis equation37. LFER have been described in the context of supramolecular host-guest binding before38, through the relation between the number of salt bridges to a host-guest binding affinity but a LFER between a host-guest binding kinetics and the binding affinity of a separate constituent of the system (the ligand in the present case) has, to the best of our knowledge, not been reported.

Here we highlight a LFER between the binding affinity of a ligand with CB7’s portal and the decomplexation kinetics of an [ADA@CB7] complex on which two identical ligands are bonded on CB7’s portals. The [ADA@CB7+2 ligands] quaternary complex is seen as a simple model to probe gating processes involving CB7. In this case, a modulation of the ADA release time (t1/2) from an ultrashort channel, consisting of the CB7 cavity only, is used as a metric to quantify the ‘gating power’ of a ligand on the portals of CB7. Intuitively, a correlation between \({{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{1}}})}\) (from equation (1)) and \({{\mathrm{log}}}\,\,\)t1/2 is expected as a large guest such as ADA cannot escape CB7’s cavity if both of its portals are occluded by ligands. Such a relationship can help power a search of the chemical space to spot strong ligands as \({{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{1}}})}\) is much cheaper to evaluate computationally than t1/2. The correlation between t1/2 (representing the kinetics) and \({{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{full}}})}\) (representing the thermodynamics), which is written as,

$$\begin{array}{lll}{{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{full}}})}=G[\,{{\mathrm{ADA}}}\,@{{{\bf{CB7}}}}+\,{{2} \,{\mathrm{ligands}}}]-G[{{\mathrm{ADA}}}]\\ \qquad\qquad\quad-\ G[{{\bf{CB7}}}]-2* G[{{1} \,{\mathrm{ligand}}}],\end{array}$$
(3)

is evaluated on a selection of organic cations. Indeed, experimental reports suggest that organic cations such as guanidinium39 and ammonium40 are able to interact with CB7’s portal, likely through electrostatic interaction with the electron-rich carbonyl groups of CB7’s portals. Therefore, 10 organic cations were selected from the ChEBI41 database for benchmarking purposes. Table 1 shows the log t1/2, \({{\Delta }}{{\mathrm{G}}}_{{\mathrm{bind}}}^{({\mathrm{full}})}\) and \({{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{1}}})}\) values of the manually selected ligands. The binding affinity of ADA inside the capped cavity, \({{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{ADA}}})}\), is also given as

$$\begin{array}{lll}{{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{ADA}}})}=G[{{\mathrm{ADA}}}\,@{{{\bf{CB7}}}}+2 \,{\mathrm{ligands}}]\\ \qquad\qquad\qquad-G[{{\bf{CB7}}+2 \,{\mathrm{ligands}}}]-G[{{\mathrm{ADA}}}].\end{array}$$
(4)

It is expected that t1/2 is a reliable estimator of a ligand’s relative ability to modify a guest’s complexation kinetics due to the extensive sampling that takes place during its computation. Given that the evaluation procedure of t1/2 does not take into account the discrete nature of water, which is known to be important in the descriptions of systems involving CBn42, t1/2 values are considered to provide relative information on the ligands rather than to be taken as quantitative estimates. In Fig. 2, it can be seen that t1/2 correlates relatively well with \({{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{full}}})}\) which is an encouraging result. Notably, a direct evaluation of \({{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{full}}})}\) might take 20-30 minutes on a single CPU using the tight-binding method GFN2-xTB43 while the evaluation of t1/2 using the same method and infrequent metadynamics may take hours or days on 50 CPU. In addition, \({{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{full}}})}\) correlates well with \({{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{1}}})}\) as shown in Supplementary Fig. 1, which allows to estimate t1/2 through \({{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{1}}})}\) and further reduces the computation time per ligand evaluated.

Table 1 Overview of log t1/2 and \({{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{x}}})}\) values for the manually selected ligands arranged in increasing order of log t1/2 (all \({{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{x}}})}\) values are given in kcal mol−1).
Full size table
Fig. 2: Relationship between log t1/2 and \({{\Delta}}{\,\mathrm{G}}_{{{\mathrm{bind}}}}^{({{\mathrm{full}}})}\) for selected organic cations.
figure 2

Relationship between log t1/2 and \({{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{full}}})}\) for a set of manually selected organic cations which are expected to strongly bind to CB7’s portal. The dashed trend line has an equation of f(x) = − 6.36x − 44.6 while the data points' labels are given in Table 1. The R2 value is 0.408.

Full size image

Interestingly, there does not seem to be a correlation between \({{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{ADA}}})}\) and t1/2 as shown in Supplementary Figure 2. All ligands tend to lower the binding affinity of ADA (\({{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{ADA}}}\,)}\)) inside the cavity even though most are seen to increase ADA’s residence time in the cavity as shown in Fig. 2. They are therefore expected to extend t1/2 purely through a kinetic effect by providing an additional energy barrier to the escape event rather than by stabilizing ADA in the cavity, which is consistent with experimental observations of a CB5 derivative capped by metal cations44. This kinetic barrier effect can be understood as an extension of the portal’s ‘constrictive binding’ often described in CB745.

The control case of adamantanone with no ligands (labeled as ‘no ligand’) appears with an estimated average residence time of 16 μs. This is larger than that estimated for xylene in CB7 using a similar procedure13, which is consistent with the fact that adamantanone is a bulkier guest for CB7 (68% packing coefficient in CB7 vs 57% for p-xylene). It is worth noting that guanidinium and pyridinium capping of the [ADA@CB7] complex yields a lower t1/2 than in the uncapped case. This is unsurprising given that while guanidinium was successfully used in the gas phase to trap benzene-sized guests in CB7, slightly bigger guests like xylene were not observed due to size exclusion. As ADA is large, it is likely that it clashes with the tight-fitting guanidinium ligands which would therefore undermine the complex kinetic stability. Confident in the ability of \({{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{1}}})}\) to spot strong ligands, the Monte Carlo Tree Search is initiated using a fitness function that takes \({{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{1}}})}\) into account.

Hunting for strong ligands using Monte Carlo Tree Search

The Monte Carlo Tree Search procedure is run 20 times in parallel, each using 4 cores for 3 days and evaluated a total of 5568 molecules. Additional details on the MCTS implementation are provided in Supplementary Methods 2.2 and 2.3. We chose a binding free energy threshold of −20 kcal mol−1 to isolate the top ligand candidates, in light of the very stable (in the fM range) biotin/avidin complex46 itself possessing a binding affinity of −20.4 kcal mol−1 and the ~2 kcal mol−1 mean absolute error in binding energy expected from the GFN2-xTB43 method. Such threshold should ensure only very strong ligands qualify and are further considered. Of the 122 molecules passing the selection criterion, only 48 were chemically stable and possessed at most one positive charge. Computations of binding affinities however can be subject to substantial inaccuracies and binding affinities of ligands larger than −20 kcal mol−1 are probably overestimates. Nevertheless, with the GFN2-xTB model used, the strong neutral guest amantadine (SMILES: C1C2CC3CC1CC(C2)(C3)N), has a predicted binding affinity of −11.14 kcal mol−1 (Kd = 7.5 nM) for CB7’s inner cavity, qualitatively comparable with the experimental value of −15.9 kcal mol−115, which supports the view that GFN2-xTB is able to spot strong interactions. As can be seen in Supplementary Fig. 4, several candidates such as the numerous phosphonium derivatives are either unstable in water or unlikely to be synthetically available. Additional discussions on the manual sorting of the top-ranked ligands suggested by the MCTS are provided in Supplementary Discussion 3.1. Common to almost all of the top-ranked candidates, is a presence of halogens and in particular iodine and bromine. Benzene rings extensively substituted with bromine and iodine make up almost a quarter of the 50 best candidates and half of the 10 best candidates.

Halogenated benzenes likely rank highly as they appear to be a natural fit for the CB7’s carbonyl-rich portal as carbonyls are known to be able to form halogen bonds47. Iodine-carbonyl bonds, in particular, have been shown to be central to design strongly binding inhibitors for certain enzymes48,49. Lower down the ranking, in the −10 to −20 kcal mol−1, chlorine and charged nitrogen atoms become more abundant. Diazonium groups also show up with high frequency.

A sample of promising, commercially available, ligands is selected and considered for further analysis using metadynamics. In particular, three organic cations labeled as numbers 13, 15, and 16 and three halogenated benzene compounds corresponding to numbers 25, 26, and 32, as recorded in Table 2, are further investigated using metadynamics.

Table 2 Overview of log t1/2 and \({{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{x}}})}\) values for a selection of the best ligands suggested by the MCTS (All \({{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{x}}})}\) values are given in kcal mol−1 and the compounds are ranked by increasing t1/2).
Full size table

Investigation of the best leads

The residence time t1/2 of the promising ligands suggested by the MCTS (numbers 13, 15, 16, 25, 26, and 32 in Table 2) is computed and reveals that the ‘gating power’ of halogenated ligands dramatically exceeds that of the benchmarking organic cations, sometimes by as much as 7 orders of magnitude.

As CB7 is water-soluble and water is a non-toxic and popular solvent, the insolubility of the promising halogenated benzenes is not optimal. The PubChem database was therefore queried for other halogenated benzenes that might be water-soluble. The search highlighted a class of X-ray radiocontrast agents using 1,3,5-triiodobenzene (20) as their absorbing core which has been optimized for water-solubility and biocompatibility decades ago already50. Four water-soluble iodinated contrast agents (27, 28, 29, and 31) were further considered and displayed large t1/2 values. This large predicted t1/2 for water-soluble derivatives of 1,3,5-triiodobenzene is encouraging and might hopefully lead to drug delivery applications in vivo thanks to their modest toxicity51.

Drawing from the observation that halogen-, and in particular iodine-, substituted benzenes are able to bring about strong ligand-portal interactions, several halogenated benzene rings are further characterized using infrequent metadynamics with the objective to obtain a spectrum of ligand-binding affinities. The results for iodobenzene (12), diiodobenzene (14, 17, 18), triiodobenzene (19, 20, 21), tetraiodobenzene (22, 23, 24) and pentaiodobenzene (30) are listed in Table 2.

Increasing the number of halogen substituents on a benzene ring tends to lower its aromaticity, decrease its HOMO-LUMO gap and increase its polarizability52. Atomic polarizabilities are responsible for dispersion interactions which are known to be important in halogen bonds53. As shown in Table 2, periodinated compounds are successful in modulating the predicted host-guest binding kinetics over almost 7 orders of magnitude.

Interestingly, ligands with an explicit negative charge due to a carboxylate group at neutral pH such as diatrizoate (31) and acetrizoate (29) still display a large enhancement in t1/2. Following a line of thought similar to our assumption that cationic ligands should bind CB7’s portals strongly, negatively charged ligands were expected to show a poorer binding to CB7’s portals. In the context of the alpb solvation model as implemented by xtb 6.3.3, this does not seem to be the case. Iopamidol (27) stands out by strongly destabilizing the [ADA@CB7] complex (with \({{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{ADA}}})}\, >\, 0\)) despite still extending ADA’s residence time in the cavity. The 1,3,5-triiodobenzene (20) framework itself is outperformed by almost 4 orders of magnitude, as measured by t1/2, by the four ligands based on it (27, 28, 29 and 31). This discrepancy originates in the side chains of the contrast agents as they are rich in groups capable of forming hydrogen bonds with the portals such as alcohols and secondary amines that can be seen forming during the MD trajectory.

One outlier is iodoxamic acid (28) which, despite showing low binding affinities between the ligand and portal, displays a remarkably high guest residence time t1/2. This phenomenon is attributed to the structure of iodoxamic acid which consists in two ligands which are, unlike all other ligands, linked together by a molecular chain. This effect is expected to boost the binding affinity as iodoxamic acid behaves as a multivalent ligand where one ligand binding the portal improves the binding affinity of the second ligand54.

All promising ligands were further displayed in Fig. 3 and shown to follow the hypothesized LFER between \({{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{full}}})}\) and t1/2 over a range of t1/2 spanning 7 orders of magnitude. The well-defined LFER in the case of capped CB7 with a large guest (ADA) begs the question whether such a relationship would hold for other host-guest complexes. We argue that for a ligand to create a kinetic barrier to the decomplexation of a guest, it only needs to bind somewhere on the macrocycle where it hinders the escape path of the guest over the energetically accessible conformational space of the capped host-guest complex. However, the LFER is only expected to become apparent if the different ligands considered bind through a similar mechanism. For example, a smaller guest exiting a hexaiodobenzene (32) capped CB7 would be required to break several halogen bonds to exit the cavity while in the case of a 1,2,3 triiodobenzene (21) ligand, the ligand might be able to just shift without requiring a breakup of as many halogen bonds. For this hypothetical small guest, the LFER between \({{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{full}}})}\) and t1/2 might not be apparent. In the case of ADA, the LFER is well-defined as the whole ligand needs to be removed from the portal to allow the bulky guest to exit the cavity and therefore \({{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{full}}})}\) correlates well with t1/2. Interestingly, cyclohexanone, a guest smaller than ADA with a weaker affinity for CB7 (−11 kcal mol−1 at the GFN2-xTB level), still displays the LFER as illustrated in Supplementary Fig. 7.

Fig. 3: Relationship between log t1/2 and \({{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{full}}})}\) for ligands suggested by the MCTS.
figure 3

Relationship between log t1/2 and \({{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{full}}})}\) for a range of ligands suggested by the MCTS and manually selected derivatives in the case of [ADA@CB7]. Organic cations (13, 15, 16) are labeled ‘+’ while the ‘no ligand’ reference is labeled □. Periodinated benzenes are labeled ○ while water-soluble commercial derivatives (27, 28, 29, 31) are labeled •. Other halogenated benzenes (25, 26) are finally labeled as Δ. There is a strong linear relationship between the free energy \({{\Delta }}{\,{{\mathrm{G}}}}_{{{\mathrm{bind}}}}^{({{\mathrm{full}}})}\) and t1/2 that holds over 8 orders of magnitude of t1/2. Water-soluble derivatives yield large values of t1/2 which is encouraging for water-based applications of the capped CB7. The periodinated compounds with one to six substituted iodine atoms show a log t1/2 that spans a large −4.18 to 2.55 range which hints at its ability to effectively modulate host-guest decomplexation kinetics. The dashed trend line has an equation of f(x) = −10.1x − 66.3. The R2 value is 0.832.

Full size image

Halogen bonds

As the strong ligands involve halogens in close proximity to the portal’s carbonyls, the presence of halogen bonds as the source of the ligand-portal interaction is checked. Despite the fact that halogen bonds are usually understood as highly directional and linear, there is evidence that halogen bonds can occur over a large range of angles from linear to lower than 90∘, including in CBs55.

In the case of halogen bonds involving amide carbonyl groups, π-orbitals predominantly act as donors in contrast to the more frequent halogen bond involving the oxygen’s lone electron pair. These π-orbitals-driven bonds can display N-C=O ⋅ ⋅ ⋅ I angles much lower than in the common case where only n-orbitals are involved53. For example, iodobenzene (12) in water forms an angle C-I ⋅ ⋅ ⋅ O of 179.3∘ and an angle N-C=O ⋅ ⋅ ⋅ I of 123.2∘ with a single glycoluril unit as optimized at the wB97XD/CPCM/Def2-SVPP level. To explore the possible occurrence of such halogen bonds in the case of the hexaiodobenzene ligand, the N-C=O ⋅ ⋅ ⋅ I and C-I ⋅ ⋅ ⋅ O angles of a [ADA@CB7 + 2 hexaiodobenzene] complex were recorded at every timestep when the O ⋅ ⋅ ⋅ I distance appeared below 3.5 Å (see Supplementary Figs. 8 and 9) over a molecular dynamics trajectory. Interestingly, a high number of bonds (up to 5 bonds per ligand per snapshot on average) is able to form. The range of both angles closely matches that of O ⋅ ⋅ ⋅ I bonds found along with carbonyl-bearing amino acids in biological molecules56. Indeed the C-I ⋅ ⋅ ⋅ O distribution shows a significant peak in the 80∘ to 120∘ while it peaks at around 75∘ in the histogram shown in Supplementary Fig. 9. Similarly the N-C=O ⋅ ⋅ ⋅ I distribution peaks at 140∘ in the histogram in Supplementary Figure 8 while it peaks at 115∘ in the reference data. Due to the fact that ligands lying flat on top of CB7’s portal would naturally exhibit a 120∘ angle with the pre-organized portal’s carbonyl groups, a ligand rich in halogens and able to be involved in halogen bonds involving π-orbitals would be expected to bind strongly to the portal while occluding it.

In addition, a Symmetry-adapted perturbation theory (SAPT) energy decomposition on a model system consisting of a single hexaiodobenzene (32) ligand docked on CB7 and optimized at the GFN2-xTB level is shown in Fig. 4. Each glycoluril unit is considered separately by deleting the other units and replacing the four connecting carbon atoms by hydrogen atoms. These newly added four hydrogen atoms are further optimized with all other atomic positions kept frozen. The SAPT energy decomposition is performed using Psi457 with sSAPT0 and the Def2-TVZP basis set with RI and JK approximations. Results from Table 3 indicate that all glycoluril units are able to sustain strong dispersion interactions with the polarizable iodine atoms of hexaiodobenzene (32). Strikingly, however, there is a large variability in Eelst and Eind. Electrostatic, induction (under the form of charge transfer), and dispersion interactions are all known to be driving forces for the formation of halogen bonds53. It is most interesting that the interaction of glycoluril unit 4 in hexaiodobenzene (32) with CB7’s portal yields the most negative EsSAPT0 of all glycoluril units thanks to strong Eelst and Eind interactions which are compatible with the energetics of a halogen bond. For glycoluril unit 4, the angle C-I ⋅ ⋅ ⋅ O is 163∘ while the angle N-C=O ⋅ ⋅ ⋅ I is 139.4∘. The staggered configuration of the hexaiodobenzene (32) ligand on CB7’s portal, therefore, maximizes electrostatic interactions with at least an iodine-oxygen couple while still allowing for dispersion interactions to take place between the other iodine and oxygen atoms. In light of the results of Table 3, the interaction of periodinated benzenes with the carbonyl portal appears driven by both Eelst and Edisp with the probable formations of halogen bonds.

Fig. 4: Hexaiodobenzene (32) binding to the CB7’s portal as optimised using GFN2-xTB.
figure 4

Hexaiodobenzene (32) binding to the CB7’s portal as optimised using GFN2-xTB. Glycoluril units are numbered from 1 to 7. The staggered configuration of the ligand appears to maximise electrostatic interactions of the ligand with the glycoluril unit 4.

Full size image
Table 3 SAPT energy decomposition of the interaction between hexaiodobenzene (32) and each of the seven glycoluril units in CB7 at the sSAPT0/Def2-TZVP level in kcal mol−1 (the sum of all energy contributions: electrostatic Eelst, exchange Eexch, induction Eind and dispersion Edisp is labeled EsSAPT0).
Full size table

Experimental validation

Experiments were carried out with two objectives. The first was to verify the ability of the ligands to modify the host-guest exchange rate by using 1H NMR, and the second was to provide evidence that the best ligands predicted by the infrequent metadynamics investigation are able to form quaternary complexes with CB7 with an encapsulated guest by using DOSY-NMR.

An observation that the ligands are able to induce a fast to slow shift in host-guest exchange kinetics on the 1H NMR timescale would be a strong clue that they are able to influence the ingression and egression kinetics of a guest, presumably by binding to CB7’s portal. The ADA guest used in the in silico ligand exploration is a known strong guest of CB7 exhibiting slow exchange kinetics (as shown in Supplementary Fig. 21) and is therefore inadequate to highlight a fast to slow shift in host-guest exchange kinetics.

Cyclohexanone is a small molecule able to form inclusion complexes with CB7 with a fast exchange kinetics, with respect to the measurement time scale of a 500 MHz instrument, as do other cyclohexane derivatives58. At the bottom of Supplementary Fig. 11, it can be seen that the 1H NMR peaks of free cyclohexanone in solution at 1.6, 1.75, and 2.25 ppm shift to 0.85, 0.95, and 1.4 ppm upon addition of CB7, exhibiting a fast host-guest exchange kinetics on the 500 MHz 1H NMR timescale. When ligands such as hexabromobenzene, hexaiodobenzene, iopamidol, diatrizoate, and acetrizoate are added to the [cyclohexanone@CB7] solution, the host-guest exchange kinetics switches to slow, in 500 MHz 1H NMR time scale, as the ligands hinder the ingression and egression of the cyclohexanone guest (see Supplementary Fig. 11). Under slow host-guest exchange kinetics, the 1H NMR peaks of both the free and encapsulated cyclohexanone appear on the same spectrum for each ligand probed. The fast to slow exchange kinetics switch is a qualitative indication that the ligands are able to influence the host-guest complexation process. Such modulation of the exchange kinetics has been reported before by the means of pH30 and salt concentration variations31 but not using neutral molecular ligands.

DOSY-NMR experiments further reveal the existence of quaternary complexes involving CB7, two ligands and a cyclohexanone inclusion guest. In the case of iopamidol (27) in Supplementary Figure 20, the ligands’ protons clearly appear near the same diffusion coefficient as CB7 and cyclohexanone at a diffusion coefficient of 1.36 × 10−10 m2 s−1. This represents almost a 50% reduction in diffusion coefficient going from the [cyclohexanone@CB7] binary complex (shown in Supplementary Fig. 17) to the quaternary complex involving two additional iopamidol ligands. The isolated cyclohexanone diffusion coefficient is measured to be 6.36 × 10−10 m2 s−1 in Supplementary Fig. 13 while the diffusion coefficients of isolated CB7 and iopamidol are only 2.71 × 10−10 m2 s−1 as shown in Supplementary Figs. 12 and 16 respectively. The presence of hexaiodobenzene (32) and hexabromobenzene (26) also causes the diffusion coefficient of the [cyclohexanone@CB7] to drop by about 15% as shown in Supplementary Fig. 18 and Supplementary Figure 19 respectively. As neither of these perhalogenated compounds possesses protons, they do not yield 1H NMR peaks and do not appear on the DOSY-NMR plots (Supplementary Figs. 14 and 15) and therefore only provide indirect evidence of the formation of a quaternary complex.

Discussion

A range of ligands able to modulate the guest release kinetics from the cavity of CB7 thanks to interactions with the portals of the macrocycle is reported. The strong ligands were designed using an efficient exploration of the chemical space through the Monte Carlo Tree Search generative model which is able to improve its suggestions in a closed-loop, iterative process. The search is enabled by the discovery of a Linear Free-Energy Relationship between the computationally cheap binding affinity of a ligand and its ability to modulate the residence time of a guest inside of the cavity of CB7. This residence time, quantifying the host-guest decomplexation kinetics, is itself used as a proxy for the ‘gating power’ of a ligand. Among the ligands investigated, organic cations were found to mildly extend or shorten a guest’s residence time while periodinated ligands are able to boost it by up to 7 orders of magnitude. The high binding affinity of the perhalogenated ligands with CB7’s carbonyl-rich portals is found to originate from strong halogen bonds with a non-linear geometry involving the π-orbital of the carbonyl groups. The best ligands are shown to experimentally form quaternary complexes with CB7 and a guest, hinting at their ‘gating power’. Their ability to influence the host-guest exchange dynamics was confirmed using 1H NMR and DOSY-NMR. It is expected that these ligands will provide flexible tools for the study of ion fluxes in synthetic, CB7-based, ionic channels. In addition, the [CB7+ 2 ligands] systems studied here might be of interest in the context of drug delivery and nanocontainers design. These encouraging results obtained in the context of a moderate-size supramolecular system shed light on the exciting potential of generative design models to assist the construction of other supramolecular gating systems.

Methods

Estimation of a guest’s decomplexation kinetics

The ability of a ligand to act as a capping agent for CB7 is investigated using metadynamics59 as implemented by xtb 6.3.335. The ligands are docked at both portals and optimized with a convergence criteria of 10−6 Ha using the alpb implicit solvation model for water. The RMSD of CB7 and adamantanone is chosen as the default collective variable while the ligands are not included. The system is then propagated in time for up to 1 ns in 30 independent trajectories. The mass of the hydrogen atoms is set to 4 u and the time step is set to 4 fs. Only covalent bonds involving hydrogen atoms are maintained using the SHAKE algorithm. The temperature is set to 298 K and controlled using xtb’s default thermostat. The kpush parameter is set to 0.05 Ha while the Gaussian width α is set to 0.6 Bohr−1. An estimate of the unbiased time to dissociation eventually built from the distribution of biased escape times using infrequent metadynamics60. Additional details are given in Supplementary Methods 2.1.

Monte Carlo Tree Search

This model is adapted from ChemTS33. ChemTS is retrained to generate small molecules by retraining its roll-out neural network on molecules with at most 12 non-hydrogen atoms and a maximum SMILES length of 31 characters. To serve as training dataset, the first 250.000 SMILES from the PubChem database with at most 12 heavy atoms from the set of elements C, N, O, S, P, F, Cl, Br, and I are chosen. Entries with isotopic information were discarded before inclusion as well as entries containing SMILES characters not considered in the original implementation of ChemTS (’\n’, ’&’, ’C’, ’(’, ’)’, ’c’, ’1’, ’2’, ’o’, ’=’, ’O’, ’N’, ’3’, ’F’, ’[C@@H]’, ’n’, ’-’, ’#’, ’S’, ’Cl’, ’[O-]’, ’[C@H]’, ’[NH+]’, ’[C@]’, ’s’, ’Br’, ’/’, ’[nH]’, ’[NH3+]’, ’4’, ’[NH2+]’, ’[C@@]’, ’[N+]’, ’[nH+]’, ’\\’, ’[S@]’, ’5’, ’[N-]’, ’[n+]’, ’[S@@]’, ’[S-]’, ’6’, ’7’, ’I’, ’[n-]’, ’P’, ’[OH+]’, ’[NH-]’, ’[P@@H]’, ’[P@@]’, ’[PH2]’, ’[P@]’, ’[P+]’, ’[S+]’, ’[o+]’, ’[CH2-]’, ’[CH-]’, ’[SH+]’, ’[O+]’, ’[s+]’, ’[PH+]’, ’[PH]’, ’8’, ’[S@@+]’). The SMILES were further canonicalized using Open Babel61. The rest of the code was adapted to work with a SMILES length of 31 characters.

The number of Gated Recurrent Units was set to 512 while the learning rate was set to 0.001, the network was trained for 10 epochs with a batch size of 512 and the training loss can be seen in Supplementary Fig. 3. Character prediction accuracy reaches 83.8% for the training set and 82.8% for the validation set. Total training time for the 10 epochs was 4 h on 10 cores from a Intel(R) Xeon(R) Gold 6140 node (2.3 GHz).

The rewards function for a molecule S with a score J(S) as computed by the fitness function of the original implementation of ChemTS which is written as:

$$r(S)=\left\{\begin{array}{ll}\frac{-aJ(S)}{1\,+\,a| J(S)| }&\,{{\mathrm{valid}}\, {\mathrm{SMILES}}}\,\\ -1&\,{{\mathrm{otherwise}}}\,\end{array}\right.$$
(5)

where a = 0.8. Given that the fitness function proposed here tends to return values in the [−50;0] range which is larger than the default ChemTS one, the parameter a is set to 0.05. Note that since high docking scores, which are taken into account in the fitness function, are negative, a strongly binding molecule will lead to a positive reward. Additional details on the SMILES preprocessing and calculation of a ligand’s fitness score J(S) can be found in Supplementary Methods 2.2 and Supplementary Methods 2.3.

Experimental details

Cyclohexanone was purchased from Sigma-Aldrich with 99% purity. Hexaiodobenzene and hexabromobenzene were purchased from Sigma-Aldrich with 98% purity. Iopamidol was purchased from Sigma-Aldrich with a 99.9% purity. Sodium acetrizoate hydrate and sodium diatrizoate hydrate were purchased from Sigma-Aldrich. All reagents were used as received without further purification steps. CB7 was synthesized by our group using the classical method developed by Day62 and Kim63. Both 1H NMR and DOSY-NMR studies were carried out on a 500 MHz JNM-ECZR Research NMR Spectrometer by JEOL. For fast to slow host-guest exchange kinetics studies using 1H NMR, a 6:1:6 cyclohexanone:CB7:ligand ratio was used while a 1:1:2 cyclohexanone:CB7:ligand ratio was used for all DOSY-NMR studies. For DOSY-NMR, the temperature was set to a constant of 25∘ C, and sample rotation was disconnected to aid reproducible scanning of the sample at a different pulsed-field gradient strength ranging from 0.3 mT m−1 to 0.5 μT m−1.