id
stringlengths 24
24
| idx
int64 0
402
| paragraph
stringlengths 106
17.2k
|
|---|---|---|
67d02e0a6dde43c908694072 | 4 | This work was financially supported by the KAKENHI Grant-in-Aid for Transfomative Research Areas (A) "Supra-ceramics" [grant numbers JP23H04629 (TM) and JP22H05145 (KS)] and KAKENHI [grant number JP24K01497 (TM)] from the Japan Society for the Promotion of Science (JSPS). Synchrotron radiation experiments were performed at BL02B2 of SPring-8 with the approval of the Japan Synchrotron Radiation Research Institute (JASRI) (Proposal Nos. 2024A1551 and 2024B1687). |
669a52265101a2ffa8ed7c5a | 0 | The metal-electrolyte interface is prototypical in electrochemistry and highly relevant for energy storage, conversion and corrosion processes. In the boundary region between metal electrode and electrolyte solution, an electric double layer (EDL) is formed when the excess electronic charge on the metal surface is compensated by equal and opposite ionic charge in electrolyte solution. This happens when the system is outside of so-called the potential of zero charge (PZC). Despite of decades of efforts, EDLs at the metal-electrolyte interfaces are still full of surprises. This is evidenced by recent experimental reports of a strong deviation from the Gouy-Chapman theory for the diffuse double layer, an unexpected high capacitance of the metal nanoparticle-water interfaces, and different ionic responses from Na + and Cl -shown in the absorption THz spectra. Understanding the molecular structure of EDLs at the metal-electrolyte interface requires inputs from theoretical investigations. Molecular modelling built on the principle of quantum mechanics and statistical mechanics are most suitable to unveil the factors that contribute to the double-layer structures and their effects on electrochemical activity. Significant progress in this area has been made recently, which points out the importance of water dynamics, the chemisorption of water, and specific ion adsorptions. Nevertheless, a direct computation of the Helmholtz capacitance of electrified metal-electrolyte interfaces which is a key experimental observable is not fully available from first-principles-based molecular dynamics simulations. |
669a52265101a2ffa8ed7c5a | 1 | This gap between experimental and theoretical investigation of EDLs at metal-electrolyte interfaces is partly due to the challenges of molecular modelling of electrified interfaces. There are in general three different categories of approaches to model electrified interfaces within density functional theory (DFT) or density functional theory-based MD (DFTMD) simulations: grand-canonical, counter-ion 10,21-24 /pseudo-atom and finite-field. Grandcanonical approaches are usually parameterized with implicit solvation models where the Helmholtz capacitance (commonly quoted as 20 µF/cm 2 ) is an input quantity rather than a prediction. In the counter-ion/pseudo-atom approaches, ions in electrolyte solution are either out of their equilibrium positions or simply neglected. This renders the difficulty of being certain about the computed capacitance from DFTMD simulations with these methods and make it less useful to guide semi-classical and quantum-mechanical/molecular mechanical (QM/MM) types of approaches. On the other hand, finite-field methods, in particular with constant electric displacement D Hamiltonian, have been shown to be rather useful in investigating the dielectric response of polar fluids, bulk electrolyte solutions, polar surfaces, and the protonic double-layer at insulating metal-oxide/electrolyte interfaces. Despite of that, its implementation for electrified metallic electrodes immersed in electrolyte solution beyond a semi-classical approach is long awaited. In the finite-field treatment of electrified interfaces under periodic boundary conditions, there is only one metal slab, which is polarized by the D field (see Fig. ). The metal slab as a whole remains neutral and so does the electrolyte slab. Since the metal slab is an electronic conductor while the electrolyte slab is an electronic insulator, the (electronic) surface charge density on the metal σ m is related to the applied D field by D = 4πσ m . The resulting potential across the supercell of length L is ∆ϕ = -(D -4πP )L, as guaranteed by the modern theory of polarizatioin 41 P and the Stengel-Spaldin-Vanderbilt constant D Hamiltonian. For electrified interfaces, this (cell) potential ∆ϕ equals to the sum of the potential over two double layers ∆ϕ EDL because the electrolyte is an ionic conductor and the Maxwell electric field over the electrolyte vanishes (on statistical average). |
669a52265101a2ffa8ed7c5a | 2 | In this work, we have implemented DFTMD simulations for modelling electrified interfaces of Au/NaCl(aq) under finite electric displacement D, which allows us to directly compute the Helmholtz capacitance for both Au(100) and Au(111) surfaces with ionic and electronic degrees of freedom treated on an equal footing. The Helmholtz capacitance that we obtained is about 60 µF/cm 2 , in excellent agreement with (recent) measurements for single crystal Au electrodes with non-specific ion adsorption. The high value of this capacitance we found in our simulations can be understood by studying the electronic response of the Au electrodes and verified by tuning the position of image plane in the semi-classical models for exactly the same chemical composition of the system and electric boundary condition. Unexpectedly, we observed fast Cl adsorption on both Au(111) at PZC and Au(100) with positive free charge within the time scale of DFTMD simulations. This leads to a charge inversion on the anodic polarized Au(100) which produces almost identical water orientational distribution as that found on cathodic polarized Au(100) without Cl adsorption. These observations suggest a non-monotonic charging behavior of Au(100) with positive bias. |
669a52265101a2ffa8ed7c5a | 3 | DFTMD simulations of Au/NaCl(aq) under finite D To begin with, both Au(100)/NaCl(aq) and Au(111)/NaCl(aq) system (Figs. and) were well equilibrated with semi-classical MD simulations under the same electric boundary conditions before porting the system into DFTMD simulations (See Supporting Information for details). This has been shown to be very useful to speed up the convergence of capacitance calculations. The Helmholtz capacitance can be computed using the formula shown below (Eq. 1), similar to the one used in the case of insulting oxide with proton charge. However, a key distinction here is that DFTMD simulations have been carried out under finite D and the imposed surface charge (σ m = D/4π) is electronic (free) charge on the metal surface instead of adsorbed proton charge due to the acid-base chemistry at different pH values. 44 |
669a52265101a2ffa8ed7c5a | 4 | where ∆ϕ is the sum of the potential differences at two sides of the same metal electrode (therefore a factor 2 for the averaged Helmholtz capacitance), P z is the Berry phase polarization (see Supporting Information for further explanation) and L z is the length of simulation box in the direction perpendicular to the interface. |
669a52265101a2ffa8ed7c5a | 5 | Figure shows the cumulative average of the cell potential ∆ϕ (Eq. 1) over 35 ps of DFTMD under applied field D = 0.018 au (σ m = 8.2 µC/cm 2 ). They are averages of 3 independent trajectories for Au(100)/NaCl(aq) as well as for Au(111)/NaCl(aq). The resulting capacitance value for Au(100) is higher (66 µF/cm 2 ) than the value for Au(111) (52 µF/cm 2 ), however, the standard deviations of the potential are overlapping in these two cases which makes them indistinguishable in practice. These trajectories are free from any specific ion adsorption (see Figure in the Supporting Information for the distance of the nearest counter-ions to the polarized surface). To put these numbers into perspective, the Helmholtz capacitance of the Au(100) single crystal in the absence of specific anion adsorption and at comparable surface charge density is about 65 µF/cm 2 from impedance spectroscopy and that of the Au(111) single crystal is about 72 µF/cm 2 from cyclic voltammetry with the Parsons-Zobel analysis. Therefore, the agreement between theoretical and experimental values is excellent for Au electrodes without specific ion adsorption by considering the uncertainties involved both simulations and measurements. The orientation of surface water molecules is known to shift in response to the surface charge of metal electrodes. Although the vibrational density of states of water bilayer look quite similar between Au(111) and Au(100), the response in a single water molecule to electric field was weaker for the surface with higher Au-O adsorption energy. However, how surface water molecules reorient themselves at electrified Au surfaces immersed in electrolyte solutions is less clear and conclusive. In Figure in the Supporting Information). Therefore, our results confirm previous studies of water orientation at PZC for Au(111) but show a clear shift in the peak positions for electrified surfaces in addition to the pronounced intensity. Moreover, we find that water at Au(100) is more responsive to the negative bias while water at Au(111) is more responsive to the positive bias. |
669a52265101a2ffa8ed7c5a | 6 | The high capacitance seen at metal/water interfaces has to do with the contribution from the electronic response of the metal electrode. The origin of this negative contribution was attributed either to the response of electronic spillover by Schmickler using the jellium model for the sp metals or to the response of contact potential between adsorbed water molecules and metal electrode by Cheng and co-workers using DFTMD and counter-ion methods. However, what the response of electronic spillover looks like for a realistic d-block metal using modern electronic structure calculations and how that changes upon the contact with electrolyte solution is not known. These are precisely the types of questions that finite-field methods can answer. In particular, we can leverage on the freedom to switch between the full DFT Hamiltonian implemented in the CP2K code and the semi-classical Hamiltonian implemented in the MetalWalls code for the same chemical composition of the system under the same electric boundary conditions. |
669a52265101a2ffa8ed7c5a | 7 | where δρ(z)| D = ρ D (z) -ρ D=0 (z) is the response charge density. It is found that the image plane position z im (relative to the atomic plane z a ) is 1.22 Å for Au(100) and 1.49 Å for Before discussing the impact of the image plane on the total capacitance, one may wonder how to validate the physical significance of z im computed using Eq. 2. Lang and Kohn showed that the definition of image plane according to Eq. 2 recovers the classical expression for the image potential. Here we looked into this question by thinking how dielectric response works for metal. A metal should completely screen the electric field. This means the following relation must hold true: |
669a52265101a2ffa8ed7c5a | 8 | where C sol is the solution capacitance and C M is the metal capacitance. We assume C sol as the capacitance with z im -z a = 0 from the semi-classical simulations, i.e. 5.95 and 6.09 µF/cm 2 for Au(111) and Au(100) respectively. These baseline values are comparable with those obtained in previous studies with similar types of models. Therefore, Eq. 5 provides a theoretical prediction on the effect of image plane on the total capacitance, which can be checked with simulations. |
669a52265101a2ffa8ed7c5a | 9 | As shown in Fig. , we do see a significant increment in the total capacitance when increasing the position of image plane in the semi-classical simulations. The relationship between the image plane and the Gaussian width parameter in semi-classical models for both Au(100) and Au(111) can be found in Fig. and S3 in the Supporting Information. Results from these simulations agree quite well with those predicted using Eq. 5. Interestingly, when the total capacitance from the semi-classical simulations for Au(100) reaches the value of capacitance computed from finite-field DFTMD simulations, the corresponding image plane z im -z a of 1.29 Å is close to that we found for Au(100) slab in vacuum with DFT (z im -z a = 1.22 Å). However, the corresponding value for Au(111) obtained from DFT appears to be too large and go beyond the singularity point (-C M = C sol ). This is likely to do with the overestimation of the work function for our Au(111) slab (see Table in the Supporting Information). Instead, taking the cross-over point between the theoretical line using Eq. 5 |
669a52265101a2ffa8ed7c5a | 10 | and the capacitance value obtained from the finite-field DFTMD simulations, one gets z im -z a for Au(111) is about 1.35 Å instead. This value is almost the same as the assumed position of image plane for Au(111). In addition, we have also seen the adsorption of Na ions in the semi-classical simulations with z im -z a = 1.22 Å (Fig. in the Supporting Information), similar to what was reported previously. This suggests a reparameterization of force fields for Na ions or the introducing of explicit polarizability is probably needed to restore the correct molecular structure at the electrified interface in semi-classical MD simulations. Nevertheless, it is clear now that the origin of this enhancement in the total capacitance have little to do with the adsorption of Na ions but come from the electronic response of gold electrode. Therefore, what is shown here provides a stepping stone for understanding the unexpected high capacitance of single Au nanoparticle reported experimentally. We have also observed very rapid Cl adsorption on the Au(111) surface at PZC (D = 0) and at polarized Au(100) (D = 0.018 au). Fig. shows the change in position of the Cl ions over the DFTMD trajectories. The Cl starts a few Å from the surface and adsorbs within a few ps and becomes stationary at the surface for the rest of the simulation. This comes in line with the "immeasurably high" adsorption rate deduced from the impedance spectroscopy. The observed fast Cl adsorption on the Au(111) surface at PZC instead of the Au(100) at PZC is likely due to a higher surface potential (therefore work function) of Au(111) as compared to that of Au(100). The opposite happened to polarized Au(111) and Au(100) surfaces are related to the interfacial water structure (see below). |
669a52265101a2ffa8ed7c5a | 11 | The migration of Cl ions correlates well with the change in the potential difference across the simulation cell (Fig. ). The movement of a negatively charged ion towards the surface results in a more positive potential from the increase in polarization. The correlation between the Cl migration towards the positive Au(100) surface and the increase in potential can be clearly seen from the corresponding trajectories. For one of the trajectories of Au(111) at D = 0, the Cl moves towards the other side of gold electrode, which leads to a more negative potential instead. |
669a52265101a2ffa8ed7c5a | 12 | For Au(111) at PZC, the distribution of water orientation on the side of metal slab with Cl adsorption show a blue-shift in the angle between water bisection and the surface normal and a clear diminishment in the peak around 60 degrees (Fig. ). This resembles the distribution of water orientation at 130 degrees as observed in the anodic polarized Au(111) surface at D = 0.018 au but with enhanced intensity. |
669a52265101a2ffa8ed7c5a | 13 | and the anodic polarized Au(100) with Cl adsorption look almost the same. Given the size of cross-section of our simulation cell, an integer charge of -1e will lead to a surface charge density of about -20 µC/cm 2 . This means when one Cl ion adsorbed on an anodic polarized Au(100) with free charge of +8.2 µC/cm 2 (D = 0.018 au), one would expect to see a charge inversion with almost the same amount but an opposite type of charge. This |
669a52265101a2ffa8ed7c5a | 14 | is exactly what we have seen for the case of Au(100) (Fig. ). The corresponding oxygen and hydrogen density distributions confirming this point are shown in Figures S7 in the Supporting Information. The charge inversion with adsorbed Cl -is expected to increase significantly the interfacial capacitance, as reported from the cyclic voltammetry. This signifies the ability of finite-field DFTMD to simultaneously control the free charge via the electric displacement D and dynamically include the ionic charge due to surface chemistry. |
669a52265101a2ffa8ed7c5a | 15 | It also does not escape our notice that this charge inversion due to specific adsorption of Cl -at the anodic polarized Au(100) electrode provides an alternative explanation for the non-monotonic change in the absorption THz spectra reported in recent experiments for gold electrode under positive bias. Furthermore, as mentioned before (Fig. ), the 120 degrees peak of water orientation is clearly visible for anodic polarized Au(100) but completely disappears for anodic polarized Au(111). Upon the adsorption of Cl -, H-bond network needs to response where the 130 degrees peak of water orientation shows up as a feature for both Au(111) and Au(100) (Figs. and). This means the H-bond networks on the anodic polarized Au(100) and the Cl-adsorbed Au(100) look more similar than those on the anodic polarized Au(111) and the Cl-adsorbed Au(111). Therefore, this structural feature of H-bond network on the anodic polarized Au(100) is expected to facilitate the Cl adsorption as compared to the case of anodic polarized Au(111). |
669a52265101a2ffa8ed7c5a | 16 | To gain further insights into the charge state of adsorbed Cl ions, we have also analyzed their electronic structures on the Au electrode and in the electrolyte solution. It is found that the 2p orbital of Cl is fully occupied when Cl -is solvated in the electrolyte solution but it becomes split upon the adsorption at both Au(100) and Au(111) surfaces (Figs. and). |
669a52265101a2ffa8ed7c5a | 17 | This indicates the nature of a strong covalent bond between Cl and Au electrode and the splitting is stronger for Au(100) than that for Au(111). These results by analyzing the density of states are further supported by the charge analysis using the restrained electrostatic potential (RESP) method for periodic systems. Despite that the atomic charge itself is not quite meaningful, the change in atomic charge does reflect the change in chemical bonding. |
669a52265101a2ffa8ed7c5a | 18 | It is found that the change in RESP charges before and after adsorption is more significant in the case of Au(100) as compared to Au(111) (Fig. ). The percentage of partial charge transfer is about 50% for adsorbed Cl -on Au(100) and 20% for that on Au(111), and the latter can be related to the experimentally measured electrosorption valency of 0.4e for Cl - adsorption at gold electrode. |
62fdb3de60751b5b1b086bd5 | 0 | RDX (1,3,5-Trinitroperhydro-1,3,5-triazine) is a common high-energy dense oxidizer in many energetic materials (EMs), which is a representative example of the caged polynitroamines, such as HMX and CL-20 . RDX has been used for more than 70 years, and there is a long history of scientific research on the material . Thermal decomposition is a fundamental process for any EMs exposed to external stimuli. It relates to the ignition of explosives, the subsequent detonation performance, and their sensitivity from mechanical stimuli to direct heating . |
62fdb3de60751b5b1b086bd5 | 1 | The decomposition mechanism of RDX crystal is widely investigated in abundance by experiments and computational simulations. Wight and Botcher explored the initial products of RDX pyrolysis by FTIR spectrum; they found NO and N2O4 are the main gas-phase products in the initial stage, illustrating that the N-N bond scission is the first step in the thermal decomposition of RDX crystal. Zhao et al. studied the infrared multiphoton dissociation (IRMPD) of RDX in a molecular beam using time-of-flight velocity spectra (TOFVS). They detected the major products from laser photolysis (mass range from 22 to 120 amu). It was concluded that the dominant channel in the RDX decomposition is a concerted symmetric triple fission of the ring in RDX. Khichar et al. conducted thermal analysis on RDX using a coupled TGA/DSC-FTIR system. N2O, NO2, CH2O, NO, HCN, CO2, CO, and H2O were identified as the major decomposition products. The combustion characteristics of RDX crystal were studied by Yan et al. with optical diagnosis methods. The reactions of NO and HCN (the products from RDX decomposition) dominate the heat release in the flame, which are further converted to N2, H2O, and CO in the downstream flame. Although many efforts are made to understand the thermal decomposition of RDX, the overall chemical mechanism during thermolysis remains obscure due to the ultrafast and complex reactions involved, which can hardly be captured in experiments. |
62fdb3de60751b5b1b086bd5 | 2 | The quantum mechanics (QM) approach is employed in the investigations of the thermal decomposition of RDX molecules, especially for the initial decomposition reaction of single RDX molecule. The first theoretical analysis of the decomposition mechanism for RDX was conducted by Melius and Binkley 8 at the MP4 level of theory. They found that the N-N homolytic cleavage in an RDX molecule is the primary decomposition pathway with a dissociation energy of ~48 kcal/mol. Harris and Lammertsma 9 calculated the potential energy surface at the B3LYP/6-31G* level of theory and identified that the N-NO2 and C-H bonds in RDX are fragile. A mechanism is proposed to initiate the decomposition by N-N bond cleavage and propagate by H atom transfer. A detailed decomposition mechanism was established by Patidar et al. , including HONO elimination, N-NO2 homolysis, reactions with NO, autocatalytic decomposition via HONO and ONNO2 addition, and hydrogen abstraction via NO. In this mechanism, RDX gradually decomposes into N2O, NO2, NO, HCN, CH2O, CO, CO2, H2O, and other small gas-phase products. Recently, Zhang et al. proposed a kinetic model for RDX decomposition. They claimed that the N-NO2 homolysis to form RDR radical is the dominant decomposition pathway of RDX, which subsequently undergoes C-H �-scission and ring-opening reaction. Although the accurate DFT calculations reveal the decomposition of RDX molecules, the current analysis is all derived from single RDX molecules and the potential intermolecular reactions in the condensed phase requires further investigation. |
62fdb3de60751b5b1b086bd5 | 3 | In the past two decades, ab initio molecular dynamics (AIMD) simulations were performed to investigate the initial decomposition pathways of HMX , CL-20 , and NTO 14 molecules. However, these works were limited to a small system (i.e., <300 atoms), and a short reaction time (i.e., <50 ps). Despite impressive progress in computing hardware and software in recent decades, AIMD calculations on the complete reaction process of RDX crystal are still challenging due to the high computational cost. To alleviate the demand for computing power, empirical potentials (or force fields) were developed to construct the potential energy surface (PES) from DFT calculations. Such empirical potentials, including ReaxFF and REBO , trade accuracy for a lower computational expense, making it possible to extend simulation scales to orders of magnitude beyond AIMD methods. ReaxFF is a bond order-based force field that describes reactive systems without prior knowledge of the predefined reactive sites. It is a powerful tool for studying kinetic mechanisms for large molecules and complex reactions . However, the computational accuracy of the ReaxFF model is relatively limited due to the underlying functional forms. Recently, machine learning-based tools, especially neural networks (NNs), have been applied to construct PES models in an entirely data-driven manner, where the PES is abstracted from a well-selected training dataset using suitable functional expressions automatically . NN models constitute a very flexible class of mathematical functions, enabling the development of PES models with the efficiency of the empirical potentials and the accuracy of the DFT method. A few successful NN-based potentials (NNPs) are proposed for materials and biomolecules , which can accurately reproduce the interatomic forces and energies predicted by ab initio calculations in condensed matters. In particular, Cao et al. implemented an NNP to reveal the mechanisms of CL-20/TNT co-crystal. They found that the TNT molecules in the co-crystal act as a buffer to slow down the chain reactions triggered by nitrogen dioxide, which is more significant at lower temperatures. |
62fdb3de60751b5b1b086bd5 | 4 | In this work, we develop an NNP to explore the decomposition mechanism of RDX crystals. The NNP is first trained and validated against the DFT database with bulk RDX molecules under a wide range of thermodynamic states. The lattice constant, the equation of state, decomposition rate, and reaction pathways are calculated with the NNP to evaluate its accuracy. Finally, the decomposition mechanism of bulk RDX molecules is proposed for the first time at the ab initio level of accuracy. |
62fdb3de60751b5b1b086bd5 | 5 | The NNP is constructed following the Deep Potential (DP) scheme developed by Zhang et al. . In the DP scheme, PES is represented by a deep neural network model that interprets the atomic coordination (R) into interatomic forces (F) and energies (E). The deep neural network contains a filter (embedding) network with three layers (25, 50, and 100 nodes/layer) and a fitting net with three layers (240 nodes/layer). The loss function (L) is defined as, |
62fdb3de60751b5b1b086bd5 | 6 | where pe and pf are the weight for the energy and force terms, respectively. N represents the number of atoms in the structure. Similar to a classical neural network, the DP scheme trains the model by computing the gradient of the loss function using the back-propagation algorithm . The NNP is trained for 1.0 × 10 6 iterations with an exponentially decaying learning rate from 1.0 × 10 -3 to 5.0× 10 -8 . , 3000, and 4000 K using the ReaxFF forcefield . One thousand configurations are randomly selected from the above trajectories to build the initial dataset, and high-level DFT calculations are further conducted to obtain accurate energies and forces. DFT calculations are performed using the CP2K package . |
62fdb3de60751b5b1b086bd5 | 7 | Core electrons are treated using Goedecker-Teter-Hutter (GTH) pseudopotentials and the Perdew Burke Ernzerhof (PBE) generalized gradient approximation method . The Grimme DFT-D3 method is used to account for dispersion interactions. A double-zeta Gaussian basis set plus polarization (DZVP-MOLOPT) is considered. In addition to MD simulations, further configurations are obtained using active learning sampling implemented in the DP-GEN package . In the stage of active learning, four NNPs are firstly trained with different random seeds. Then, MD simulations with a temperature range from 300 to 4000 K are performed using one of the NNPs. The MD trajectories are evaluated by the other three NNPs to obtain the deviation of atomic forces, which is used as a criterion to identify the new configurations. The configurations with relative deviation in the range of 0.05-0.15 are added to the training set. In total, we perform 20 iterations of active learning to derive an accurate NNP for the decomposition of RDX crystal (Table ). The iterations 1-13 investigate the RDX decomposition process under an NVT ensemble with a 2x2x2 supercell. The temperature gradually increases from 300 K to 4000 |
62fdb3de60751b5b1b086bd5 | 8 | The RDX decomposition process is investigated with the final NNP. The simulation is performed under the NVT ensemble with a 2x2x2 supercell. The equations of motion are integrated by the velocity Verlet method using periodic boundary conditions. A Nose-Hoover thermostat is applied with an equilibrium temperature of 1000, 1250, 1500, 1750, 2000, 2250, 2500 K and a dump parameter of 20 fs. A 100-ps MD simulation is performed with a time step of 0.l fs. Three parallel simulations are performed for each NVT MD simulation to ensure the statistical significance of the simulated results. The ReacNetGenerator is used to extract species and reactions from the MD trajectories. |
62fdb3de60751b5b1b086bd5 | 9 | The overall performance of the NN potential is tested against the ab initio database. Figure shows the prediction of DFT energies and forces using the NNP trained by the DP scheme. The results predicted from the ReaxFF model are also included for comparison. The energies predicted from the NNP cluster on the diagonal in We also evaluate the computational costs of NNP, ReaxFF, and DFT methods on the bulk RDX systems with 100 to 100,000 atoms (Figure ). The neural network potential-based molecular dynamics simulations (NNP-MDs) are performed on an NVIDIA V100 GPU. The ReaxFF and DFT calculations are solved on a 64-processor server with two AMD EPYC 7452 CPUs. The NNP shows a linear scaling rule in the computational cost, which is 27 times faster than the ReaxFF model. Compared to the AIMD method, the simulations using NNP are faster by four orders of magnitude. This significant improvement in the computational costs is consistent with previous works using the NNPs . This can be attributed to the implementation in the neural network combining the-state-of art GPU computing. The computing efficiency of the NNP in small systems with hundreds of atoms deviates from other significant cases. This is expected as the implementation of the NNP is optimized for large systems. Also, it is worth noting that the ReaxFF model exhibits an O(N) scaling rule rather than an O(NlogN) scaling rule, different from previous works in the low-density gas . The better performance of the NNP enables the exploration of a system with tens of thousands or even millions of atoms at the ab initio accuracy, providing a feasible approach to investigate the complex reaction network of RDX crystal from an atomic perspective. |
62fdb3de60751b5b1b086bd5 | 10 | The rate constants are calculated using the species evolution of RDX molecules in 10 ps. The rate constant derived from the NNP is in the same order as the ReaxFF model but higher by a factor of two. From the Arrhenius plot in Fig. , Arrhenius parameters for RDX decomposition are determined by a direct fitting. The activation energies predicted by NNP and ReaxFF are 25.24 and 26.01 kcal/mol, respectively. The results from both models are in good agreement with the experimental value of 28.6 kcal/mol using the Kissinger method . The above discussion confirms that the NNP accurately describes the kinetics of RDX decomposition. In the following section, we discuss the detailed decomposition mechanism of RDX crystal from condensed phase into gas-phase intermediates. |
62fdb3de60751b5b1b086bd5 | 11 | produced from the homolytic fission of the N-NO2 bond in RDX molecules. This finding is consistent with previous works as the breakage of N-NO2 bond triggers the RDX decomposition. Chakraborty et al. identified the first bond-breaking event as N-NO2 bonds in a single RDX molecule compared to the HONO elimination and concerted ring break. The corresponding dissociation energy is the lowest among the other two potential reactions computed by the B3LYP method. The NO2 molecules are subsequently consumed in a rapid manner within 10 ps, followed by the production of NO and H2O molecules. In the later stage, the production of N2 and CO2 are observed, and the overall reactions reach an equilibrium after 100 ps. The production of NO2, NO, HNO2, H2O, and CO2 is also reported in a recent FTIR experiment .The RDX decomposes into NO2, NO and H2O first, and then NO2 and NO starts to decrease after a certain time, while H2O continues to increase. Although the experiments are performed at a lower temperature (e.g. 538K), the product evolution agrees with our NNP simulations qualitatively. The predicted species evolution using the ReaxFF model is included in Figure . The overall evolution is similar with the NNP results, where NO2 is the major intermediate, and N2, H2O, and CO2 are the major products. Compared with the results of ReaxFF model, the NNP predicts more NO molecules, a faster H2O production rate, and a higher N2 concentration at the equilibrium. To validate the accuracy of the NNP in species evolution, the evolution of NO2, NO, and H2O are also compared in Fig. , indicating that the NNP outperforms the ReaxFF model in the prediction of the decomposition process. In addition, the NNP predicts that N2 is the most abundant final product, followed by H2O and CO2. This finding agrees with the experiments where the final products of RDX decomposition is 37% N2, 31% H2O, 18% CO2, and 14% CO. It should be noted that the NNP predicts a lower concentration of CO, because PBE method underestimates the stability of CO . The current NNP inherits this issue to some extent, and further improvement can be carried on to add more accurate DFT calculations of CO-involved reactions in the training set. |
62fdb3de60751b5b1b086bd5 | 12 | The MD trajectories using the NNP are further analyzed by ReacNetGenerator to reveal the complex reaction network. The primary reaction pathways of RDX decomposition and the formation of final products are constructed in Fig. , where the arrow width represents the observed number of reactions. The major channel of RDX decomposition starts with N-N homolysis to produce NO2 molecules and RDR radicals, which agree with previous DFT studies . Then, the RDR radicals mainly decompose through three pathways: In our simulations using the NNP, the N-N homolysis is more critical compared to the ring-opening reactions and the H abstraction/addition reactions in the RDR decomposition as the corresponding frequency number is 49, 16 and 11. This finding here disagrees with the DFT calculations by Chakraborty et al. ; they reported that the RDR prefers to decompose through the ring-opening reactions rather than N-N homolysis. The reason might be attributed to the effects of neighbor molecules. Chakraborty et al. studied the reaction dynamic of an isolated molecule. In contrast, our NNP-MD simulations resolve the decomposition of the bulk RDX crystal. The neighbor molecules around the RDR might impose intense steric effects on the ring-opening reactions prohibiting the progress of ring opening, and this effect are discussed latter. In previous DFT studies , the HONO pathway is another channel for RDX decomposition, where H atoms migrate to -NO2 in RDX and undergoes a direct elimination of HNO2. This pathway is observed in our NNP-MD simulations, but its frequency is very low (e.g. In addition, we further perform MD simulations on RDX in the gas phase using NNP, where the densities are selected as 0.5 and 0.1 g/cm . The simulations are performed under an NVT ensemble at 2500 K. The results from RDX crystals (1.8 g/cm 3 ) are also included for comparison. Figure shows the flux of N-N homolysis, H abstraction, and ring-opening reactions of RDX and RDR molecules. As discussed above, RDX crystals are mainly consumed through the N-N homolysis, while the proportion of H abstraction and ring-opening reaction increase for RDR molecules. As the density decreases, the proportion of ring-opening reaction increases significantly (from 6% to 26% for RDX, 17% to 34% for RDR). The proportion of N-N homolysis, and H abstraction reactions decrease with the system density. Such behaviors are attributed to the effects of neighbor molecules. At high-density conditions, the ring-opening reaction is hindered by the neighbor molecules due to the steric effects. |
62fdb3de60751b5b1b086bd5 | 13 | Decreasing the density, RDX decomposition is less affected by the neighbor molecule, which is close to the isolated molecule in the previous DFT study . In summary, the effect of the neighbor molecule should be taken into consideration when studying the reaction mechanism of RDX and other EM in the condensed phase in practice. The present work develops a NN model for RDX at the level of ab initio calculations. Our NN model considers both atomic energy and force information from high-level DFT calculations. We demonstrate the accuracy of our NNP outperforms the widely used ReaxFF model. In particular, the ReaxFF model predicts unphysical atomic force compared to DFT calculations. This issue makes it unsuitable for the exploration of the processes in the reaction dynamics. Our NN model also exhibits a great computational efficiency, which is 10000 times faster than the DFT method, and 27 times faster than the ReaxFF method, allowing the possibility of investigating the complex reaction process of bulk RDX crystal from an atomic perspective. Thus, this study opens new opportunities for complex reactive systems to build reaction kinetics models. |
66fa3c1812ff75c3a1e99030 | 0 | Diabetes is a long-term metabolic disease marked by high blood glucose levels . It is sometimes referred to as diabetes mellitus . Over time, diabetes can cause severe damage to the kidneys, nerves, eyes, heart, and blood vessels . The body's inability to produce enough insulin results in type 1 diabetes, but Type 2 diabetes is caused by inefficient insulin utilization in the body . Moreover, gestational diabetes raises the mother's and the child's future chance of developing Type 2 diabetes during pregnancy . Diabetes is becoming increasingly common worldwide and is becoming more severe in middle-and low-income nations . This presents a serious public health concern and emphasizes the need for quick decision-making and coordinated efforts to prevent and treat this crippling condition . |
66fa3c1812ff75c3a1e99030 | 1 | Because it is involved in several physiological processes frequently dysregulated in diabetic settings, clusterin, also known as apolipoprotein J, has potential as a biomarker in diabetes . Clusterin is a secreted glycoprotein made inside cells and transported outside the cell . It is formed in organs such as the liver, kidneys, brain, and adipose tissue . When cellular stress, inflammation, or tissue damage-all prevalent in diabetes-clusterin production and release can be increased, which helps shield cells from harm, lessen inflammation, and encourage tissue repair . Diabetes is characterized by metabolic dysregulation and chronic inflammation, which have an impact on clusterin expression and release . Inflammatory cytokines, oxidative stress, and elevated glucose levels all promote the synthesis of clusterin . |
66fa3c1812ff75c3a1e99030 | 2 | Measuring blood levels of clusterin may help in early diagnosis and monitoring of diabetes complications, especially diabetic nephropathy, and cardiovascular illnesses, which are linked to elevated blood levels of this protein . Research has demonstrated a correlation between the severity of a disease and clusterin levels; higher levels may signify a more advanced disease or more significant tissue damage . Monitoring clusterin levels may reveal information about the efficacy of treatment measures and act as a predictive biomarker for the onset of problems in diabetes patients . To increase diagnostic accuracy and prediction value for diabetes complications, measurement methodologies must be standardized, large-scale validation studies must be conducted, and clustering must be integrated with other biomarkers for clinical use . |
66fa3c1812ff75c3a1e99030 | 3 | A short, single-stranded DNA or RNA molecule known as an aptamer has a high affinity and selectivity for binding to particular target molecules . The systematic procedure known as SELEX is used to select these compounds. Because of their comparable binding properties to antibodies, aptamers are frequently used interchangeably, but they also have unique benefits such as simpler manufacturing, less immunogenicity, and increased thermal stability. Aptamers are used in many contexts, such as biosensors, medicinal agents, and diagnostic instruments. |
66fa3c1812ff75c3a1e99030 | 4 | They can be designed to target unhealthy cells, such as cancer cells, or to obstruct particular biological pathways in therapies. Aptamers can be diagnostic tools to find proteins, small compounds, or other disease-related biomarkers. They are instrumental in research and medical contexts due to their excellent specificity in binding to various targets. Molecular docking is a computational method for predicting how a molecule will bind to another to form a stable compound . It plays a crucial role in drug development by simulating interactions between small compounds, like potential therapeutic candidates, and their biological targets, typically proteins. By identifying these molecules' binding affinity and specificity, molecular docking helps design more effective and targeted medications. This method relies on algorithms that explore various binding poses and rank them based on their stability, aiding in understanding molecular interactions and creating new drugs. |
66fa3c1812ff75c3a1e99030 | 5 | Since clusterin protein is expressed in beta cells, its expression level could be correlated with the beta cell biomass. Increased clusterin expression in beta cells is related to beta cell oxidative stress and could eventually lead to Type 1 diabetes. Since the clusterin protein is released into the bloodstream, it is easy to measure its plasma level, which could indicate the beta cell mass or cell stress level. Based on this theory, Simaeys' et al. recently designed an RNA-based aptamer targeting clusterin protein as a pre-diabetic biomarker . We hypothesize that the binding of aptamers should be due to their unique shape and sequence. Therefore, we have performed computational simulations to understand the aptamer-biomarker protein interactions. Specifically, we designed aptamers' 3D structures and later docked them on clusterin protein. |
66fa3c1812ff75c3a1e99030 | 6 | We have used graph neural networks using the GrASP web server to forecast this site is druggable . In Figure , the druggable site is displayed. This indicates that the druggable and binding sites are yellow and green. Understanding the surface properties of the clusterin is crucial in validating the clusterin-aptamer docking results. Vfold2D was utilized to estimate the secondary structure of aptamers . Aptamer 2D is vital in defining the three-dimensional aptamer shape and provides information on base pair interactions (hydrogen bonds). A popular method for representing the secondary structure of nucleic acids, including aptamers, is dot and bracket notation, Figure . Several symbols, such as dots and brackets, are used to denote paired and unpaired bases in this notation. In this notation, paired bases in a stem are represented by parenthesis( ) and unpaired by dots (.). A pairing begins with an opening parenthesis, "(," and ends with a closing parenthesis, ")." The dot-bracket notation is shown in Table , and the 2D image is shown in Figure . |
66fa3c1812ff75c3a1e99030 | 7 | Additionally, we used CrustalW, a bioinformatics computer tool for multiple sequence alignment, to carry out the sequence alignment . The sequences were organized using sequence alignment to identify common portions across aptamer sequences that might suggest structural and functional similarity. A phylogenetic tree is called a branching diagram or "tree" that illustrates the connections between the various aptamer sequences. Aptamer sequences are aligned during this process to arrange similar or identical nucleotides in the same column. From SELEX, 39 aptamers were selected for further evaluation and sequence alignment was divided into a set of two. Therefore, we have chosen only two of these two sets' aptamers. A phylogenetic tree analysis was used to confirm this further, as shown in Figure . |
66fa3c1812ff75c3a1e99030 | 8 | Figure illustrates the tertiary structure of aptamers, which results from the secondary structure's folding. The function of aptamer depends on these three-dimensional structures, which also control how well an aptamer interacts with a protein. Per our forecast, every aptamer possesses a distinct three-dimensional structure, which is evident in how they bind together in docking simulations (explained subsequently). Figure and b displays the clusterin-aptamer docked structures. The figure shows that the clusterin-aptamer interaction happens to the same binding spot predicted by the graph neural network. In addition, we have also obtained the clusterin-aptamer binding interaction using AlphaFold 3, which also shows a similar binding site. |
66fa3c1812ff75c3a1e99030 | 9 | Aptamers have been used in diabetes detection before. For example, Lee et al. developed ssDNA-based aptamers that bind to the retinol-binding protein for the early detection of diabetes . These aptamers were attached to a gold chip, and SPR was used to detect the aptamerprotein specificity. They have also stated that these aptamers are more sensitive than ELISA assay . In a different study, Apiwat et al. have developed graphene-based aptasensor diagnosing diabetes mellitus . Moreover, aptamer has been used to identify SecinH3 in human liver cells . Apart from diabetes mellitus, aptamer has been used in various metabolic diseases such as aptamer targeting adipose tissues could have an impact on obesity and metabolic syndrome . Till now, aptamer targeting is as follows: HbA1c, GHSA, RBP4, adipocyte cell line, and visceral adipose tissue-derived serpin . Noninvasive aptamermediated therapy targeting pancreatic B cells will be a valuable tool and pave the way for treating diabetes and its complications. We hope future research will focus on the aptamer-mediated inhibitory effect of T cells and macrophages in diabetes complications . |
66fa3c1812ff75c3a1e99030 | 10 | First, the quality of the models and algorithms employed significantly impacts the overall accuracy of the molecular docking, and poor models can and often do produce false positives. In the future, we hope to incorporate the behavior of the protein in the presence of water surrounding its ligand through molecular dynamics simulations. MD simulations will also reveal the strength of the interaction and whether the ligands remain with the protein during the simulation. Moreover, ITC or SPR should be used to validate the computational research after further analysis further. These assays will prove that the suggested ligands bind to the protein in the anticipated manner. |
66fa3c1812ff75c3a1e99030 | 11 | In this work, we investigated the potential of aptamers as diabetes diagnostic agents by studying their interactions with the biomarker clusterin protein, which is implicated in cellular stress and inflammation associated with diabetes. This aptamer can bind to the β cells and could be used in the β cell mass for the early detection of Type 1 diabetes. We found particular binding sites on clusterin where aptamers produced stable complexes characterized by numerous hydrogen bonds and salt bridges. Our results show that aptamers can specifically target and bind to clusterin, indicating its potential use in the early detection and tracking of problems associated with diabetes. The accuracy and effectiveness of aptamer-based diagnostics were demonstrated by applying graph neural networks (GNNs), which improved the identification of druggable locations on the protein. The research highlights the potential of aptamers in developing tailored treatments and diagnostic instruments for diabetes, opening the door to more precise and efficient treatment. To combine the clusterin protein with other biomarkers for better diagnostic and predictive potential, future studies should concentrate on standardizing measurement techniques and carrying out extensive validation studies. |
66fa3c1812ff75c3a1e99030 | 12 | AlphaFold 3 predicted the protein structure of clusterin protein . The models were acquired, and their structural accuracy was confirmed. The Vfold2D and Vfold3D servers received the aptamer sequences and produced 2D and 3D structural predictions, respectively . The stability and possible binding conformations of the secondary and tertiary structures were examined. Using molecular docking software, the 3D structures of the aptamers were positioned onto the anticipated protein structures. The binding locations and interaction strengths were determined by analyzing the docking results. Chimera software was used to visualize the individual structures and the docked complexes (23) . For additional investigation, figures showing the structural alignments and interactions were created. An aptamer binding site on the proteins was predicted using a graph neural network (GNN). Known aptamer-protein interactions were used to train and validate the GNN model. The protein sequences were subjected to phylogenetic analysis using MEGA software . Phylogenetic trees were generated to determine links and evaluate the conservation of the aptamer binding sites. |
639f5730e8047a4e4aed64a8 | 0 | The use of curved supports for inducing local strain is well established for generating exotic properties from conventional layered materials 1 . An excellent example is graphene, which already exhibits remarkable properties in its planar configuration. Straining graphene can modify its electronic structure to create polarized carrier puddles, induce pseudomagnetic fields, and alter surface properties . In MoS2, modifying the supporting glass sphere diameter induces curvature in the MoS2, which permits precise bandgap tuning of MoS2 in a continuous range as large as 360 meV . Curved MoS2 conformally coated on Au nanocone arrays is a promising catalyst for the hydrogen evolution reaction (HER). The improved HER activity is attributed to sulfur-vacancies and reduction of the bandgap under strain . Graphene and MoS2 are good candidates for strain engineering because they are relatively large (mm) in size. However, strain induction for sub-2 nm materials is challenging due to the mismatch of lateral size with supports and difficulties in material transfer. |
639f5730e8047a4e4aed64a8 | 1 | Molecular complexes such as metalloporphyrins and metallophthalocyanines are efficient CO2 reduction reaction (CO2RR) catalysts because of their electronic structures and the tunable ligand environments surrounding the active sites . However, these molecules along with most other molecular catalysts primarily reduce CO2 to CO; catalysts that selectively reduce CO2 beyond CO are scarcely reported . An early study of cobalt phthalocyanine (CoPc) in 1984 showed a methanol (MeOH) Faradaic efficiency (FE) of <5%. This work did not receive much attention until recently when the Wang and Robert 10 groups separately reported improved FEMeOH with CoPc deposited on multiwalled carbon nanotubes (MWCNTs). In the paper from the Robert group, it was emphasized that only small amount of methanol could be obtained from CO2RR, while larger amount could be generated from CO reduction reaction (CORR). In stark contrast, the Wang group claimed that a FEMeOH as high as 44% could be obtained with CoPc deposited on MWCNT. However, most studies have been unable to obtain the improved FEMeOH observed by the Wang group, instead finding CO to be the prevailing product even with monodispersed CoPc . Numerous discrepancies in reaction selectivity and kinetics amongst CoPc studies make it difficult to determine the ability of CoPc to catalyze CO2RR beyond the prominent CO product. |
639f5730e8047a4e4aed64a8 | 2 | CNTs are exceptional support materials for heterogeneous catalysis. Their large specific surface areas readily disperse nanoparticles to avoid agglomeration, and their high electronic conductivities make them promising for electrochemical applications. Herein, we present new insights into the role of CNTs in heterogeneous catalysis. Thanks to recent advancements in synthesis and purification, the diameters of CNTs can be controlled from 2 nm to >50 nm, making them ideal supports for inducing strain in sub-2 nm planar molecules. Depending on the local interactions between curved CNTs and the overlayer, molecules may undergo controllable distortion to alleviate strain (Fig. ). Assuming the overlayer is fully elastic and the interlayer distance is 0.3 nm, the bending angle can range from ~96 o (1 nm-diameter CNT) to ~1.5 o (100 nmdiameter CNT) (See Supplementary Methods 1 for estimates). We report X-ray spectroscopic studies and other spectra to assess the structure of molecular CoPc before and after monodispersion on various CNTs. We observe significant molecular distortion and strong molecule-CNT interactions for SWCNTs. This is supported by our DFT calculations that find increased distortion strengthens Co-CO binding, which improves selectivity towards MeOH. As a result, the distorted CoPc/SWCNT exhibits a 385% improvement in FEMeOH compared to CoPc/MWCNT. We also extend these findings to ORR and CO2RR studies on SWCNTs, which also exhibit straindependent catalytic activity. |
639f5730e8047a4e4aed64a8 | 3 | Metallophthalocyanines deposited on different CNTs are denoted as MPc/X, where M is the metal and X refers to SWCNTs or the average diameter (in nm) of the MWCNTs (Supplementary Fig. and). Scanning TEM energy-dispersive X-ray spectroscopy (STEM-EDS) mapping shows a uniform distribution of N and Co along the SWCNT surface (Fig. ). In aberration-corrected HAADF-STEM (Fig. ), the bright spots marked with red circles further verify the uniformly dispersed Co sites. The Raman spectra for CoPc and various bare and CoPc-decorated CNTs can be found in Supplementary Fig. and . |
639f5730e8047a4e4aed64a8 | 4 | We sought to elucidate how interactions of CoPc with CNTs might affect catalysis. Due to π-π stacking and through-space orbital interactions, the electron delocalization should increase the absorption wavelength . This is validated by the UV-vis spectra of the catalysts, showing a red shift at the Q band of the CoPc molecules after deposition on CNTs (Supplementary Fig. ) . X-ray photoelectron spectroscopy (XPS) for both the Co 2p1/2 and Co 2p3/2 of CoPc/CNTs shift to higher binding energies (Fig. ), reaching a maximum difference of 1.2 eV for CoPc/SWCNT. The signal evolution is even more prominent for the N 1s peak (Fig. ). Previous reports studied the interactions of MPc with various substrates such as TiO2, Au, and other semiconductors . It was thought that the strong molecule-substrate interaction would induce a peak splitting, while a weak interaction would cause only a peak shift. Indeed, we observe a new peak at >400 eV for all the CoPc/CNT samples. CoPc/SWCNT shows the most intense new peak at ~400.5 eV, implying strong CoPc-SWCNT interactions. The peak splitting of CoPc/SWCNT is ~1.3 eV, attributed to the deviated N 1s energies according to theoretical calculations discussed below. To understand the CoPc/CNTs' local structures, we performed X-ray absorption near-edge structure (XANES) and extended X-ray absorption fine structure (EXAFS) analyses of Co. As shown in Fig. , the Co K-edge XANES spectra exhibits an obvious peak at 7715 eV; this 1s→4pz transition is indicative of the Co-N4 structure . A decline in the 1s→4pz transition is observed when CNTs are introduced, likely due to the decreased symmetry (D4h to C4v). The peaks near 7725 eV result from the 1s→4px,y transition, with the exact peak position dependent on Co's valency . The peak position of CoPc/CNT shifts slightly to lower energy compared to CoPc, indicating charge transfer between CoPc and the CNT. The Fourier transform (FT) of the EXAFS spectra shows the coordination environment around Co sites (Fig. and Supplementary Fig. ). The signal can be cataloged into three groups: . In addition, the peak intensities increase with smallerdiameter CNTs; this increased Co coordination number is a result of stronger interaction of CoPc with the CNT, as inferred from signal fitting. The fitted EXAFS signals (Fig. and Supplementary Fig. and, Supplementary Table ) imply molecular bending around the CNT. The high degree of molecular bending for CoPc/SWCNT leads to elongated Co-N1 and Co-C1 ) compared to the 1.9 mA cm -2 for CoPc/15 and 1.5 mA cm -2 for CoPc/50. To confirm that MeOH is derived from CO2 rather than other impurities, isotopic-labelling experiments were conducted in 13 CO2-saturated 0.5 M KH 13 CO3 while under continuous 13 CO2 flow. The peak splitting of 1 H NMR at 3.32 ppm and the obvious C NMR peak at 49.11 ppm verifies that the produced MeOH originates from the CO2 input (Supplementary Fig. ). The bare CoPc, SWCNT, and MWCNT/15 samples show negligible electrocatalytic CO2RR activity with little CO produced throughout the potential window (Supplementary Fig. ). In the durability test, the CoPc/SWCNT maintained ~30% FEMeOH for 10 h at an operating total current density of 16 mA cm -2 (Supplementary Fig. ). Supplementary Fig. shows the relation of FEMeOH with CoPc loading on SWCNT. Increasing the CoPc:SWCNT ratio from 1:50 to 1:3 and 1:1 did not increase FEMeOH. This could be explained by CoPc-CoPc stacking at high loadings that weakens the support effect. High CoPc loading also leads to catalyst leaching during chronoamperometry. |
639f5730e8047a4e4aed64a8 | 5 | We next explored how FEMeOH varies with CNT diameter at -0.93 V (Supplementary Fig. and). The FEMeOH decreases from 53.2% for CoPc/SWCNT to 13.9 % for CoPc/50. The conversion of CO2 to MeOH involves six electron transfers, so that high charge transfer resistance will negatively affect FEMeOH. Electrochemical impedance spectroscopy (EIS) at -0.93 V shows that CoPc/SWCNT has less charge transfer resistance than CoPc/15 and CoPc/50, indicating more facile substrate reduction with CoPc/SWCNT (Supplementary Fig. ). |
639f5730e8047a4e4aed64a8 | 6 | The CO2RR current density is limited in the H-cell by low CO2 solubility and mass transport in aqueous electrolyte. In order to achieve higher jMeOH, we constructed a flow cell featuring a gas diffusion electrode (Supplementary Fig. ). In the flow cell, 0.1 M KOH + 3 M KCl (instead of KHCO3) was used as the catholyte to improve jMeOH and suppress HER at high potential . The total current density in the flow cell is substantially higher (Supplementary Fig. and) than that of the H-cell at all applied potentials. The maximum jMeOH of CoPc/SWCNT reaches 66.8 mA cm -2 with a 31.3% FEMeOH at -0.9 V vs. RHE, which is 7.6 times that of jMeOH in the H-cell (Fig. ). The maximum jMeOH of CoPc/15 and CoPc/50 are 21.7 mA cm -2 (15.6% FEMeOH) and 9.3 mA cm -2 (9.3% FEMeOH) at -0.9 V vs. RHE, respectively. Additionally, CoPc/SWCNT exhibits a stable total current density with no loss for 20 min at various potentials (Supplementary Fig. ). Even at 200 mA cm -2 (-0.9 V vs. RHE), the FEMeOH is maintained at ~26% for 10 h without decay (Supplementary Fig. and). |
639f5730e8047a4e4aed64a8 | 7 | Competition with *CO2 adsorption can be resolved by a tandem reaction. Specifically, CO2 can be first reduced to CO with >95% FE, and then the produced CO can be further reduced to MeOH in a second electrolyzer free of CO2 in alkaline media. We conducted direct CO reduction in a flow cell identical to the parent CO2RR process, except that CO was used as the feed gas to achieve higher FEMeOH and jMeOH. For CoPc/SWCNT, jMeOH reaches 62.1 mA cm -2 at -0.8 VRHE with a corresponding FEMeOH of 50.5% (Fig. and Supplementary Fig. ). This corresponds to a FEMeOH of 60.1% for the tandem reaction, assuming a 95% FECO from CO2. CoPc/15 and CoPc/50 achieve jMeOHs of only 20.5 mA cm -2 and 15.5 mA cm -2 with FEMeOHs of 21.8% and 16.6%, respectively (Fig. and Supplementary Fig. ). Moreover, CORR with CoPc/SWCNT maintains a total current density of 100 mA cm -2 at a FEMeOH of ~50% for 10 h (Fig. and Supplementary Fig. ). |
639f5730e8047a4e4aed64a8 | 8 | An extremely high applied potential would over-reduce the Pc ligand, resulting in deactivation via demetallation . To probe for demetallation, we measured the UV-vis absorption during long-term stability tests for the H-cell, the CO2RR flow cell, and the CORR flow cell (Supplementary Fig. ). Compared to the pristine UV-vis spectrum (Supplementary Fig. ), post-electrolysis CoPc/SWCNT displayed no new absorption peaks, indicating no demetallation. |
639f5730e8047a4e4aed64a8 | 9 | We find that binding the CoPc catalyst to the CNT walls changes the geometric and electronic structure. To further investigate this, we employed molecular dynamics (MD) calculations using the Universal Force Field (UFF) . These MD simulations used the polarizable charge equilibration (PQEq) scheme for electrostatics and the RexPoN universal nonbond (UNB) van der Waals interactions (UFF is used solely for valence terms). Using UFF + PQEq + UNB allows us to obtain molecular geometries at the accuracy level of high-quality quantum mechanics while at the cost of classical mechanics. |
639f5730e8047a4e4aed64a8 | 10 | We estimate that the smallest SWCNTs in our experiments have a diameter of ~2 nm. To model the SWCNT, we use a large C116H28 graphitic sheet with curvature matching that of a SWCNT with 2 nm diameter (Supplementary Fig. and); the 116 C atoms provide ample surface area for coordinating CoPc. We then placed a flat CoPc catalyst within van der Waals distance of the outside of the SWCNT. With the H atoms of the SWCNT frozen, we optimized the geometry of the CoPc when near the outside of the SWCNT wall. The MD simulation shows that the CoPc curves around the C116H28 such that the curvature of the catalyst and the CNT are equal. Additionally, the curved plane of the catalyst lies ~3.11 Å from the wall of the SWCNT, indicating significant p-p stacking between the catalyst and the SWCNT. We also explored how CoPc interacts with a large, flat sheet of carbon, which represents the MWCNT. In this case, the CoPc plane remained parallel to the flat C116H28 sheet with the distance between the two planes being ~3.37 Å. The predicted structures agree well with our XANES and EXAFS analyses. |
639f5730e8047a4e4aed64a8 | 11 | In addition to the geometric distortion of CoPc upon binding to the SWCNT, we observe a shift in the N 1s indicating electronic structure distortion. DFT predicts 4 degenerate N 1s orbitals at -381.16 eV below the vacuum energy. Forcing the complex into the curved geometry breaks the degeneracy of the N 1s orbitals. Specifically, the 4 orbitals now range from -381.44 eV to -381.36 eV (0.08 eV variation). The N 1s orbitals are also stabilized in the curved complex by an average of -0.24 eV, indicating increased binding energy as observed in experiment (Fig. ). |
639f5730e8047a4e4aed64a8 | 12 | The modulated selectivity for CO2RR towards MeOH over CO is likely due to a change in the CO absorption step. To explore this, we applied Grand Canonical DFT to examine the energetics as a function of chemical potential. Our Grand Canonical Potential Kinetics (GCPK) method keeps the applied potential constant for the initial and product states of each reaction step, just as in experiment (see Methods for more details). For CoPc on the SWCNT, we use the curved CoPc generated from the MD simulation as the starting point. For CoPc on the MWCNT, we start with the typical CoPc catalyst with D4h symmetry. Liao and coworkers recently suggested that at -1.0 V vs. RHE (-1.4 V vs. SHE at pH=6.8), the CoPc catalyst is spontaneously reduced to CoPcH4 in water. Specifically, the 4 outward N of the Pc ligand are hydrogenated (hence PcH4) with a free energy change of -2.92 eV relative to CoPc at 298.15 K. Thus, for our calculations we hydrogenate the 4 outward N of the Pc ligand for both curved and flat CoPcH4. These hydrogenations are not likely to cause Co demetallization since the reductions do not occur at the Co-N sites. The CO absorption energy as a function of applied potential is shown in Fig. . We see that for the calculated potential window, the curved CoPc has stronger CO binding than the flat CoPc. At In Fig. , the curved CoPcH4 is higher in energy than the flat analog due to the induced strain. However, the curved CoPcH4(CO) is lower in energy than the flat analog, because the strain makes for favorable CO binding. These two factors lead towards improved CO binding for curved CoPc, which ultimately improves MeOH selectivity. |
639f5730e8047a4e4aed64a8 | 13 | The distortion of the square pyramidal CoPcH4(CO) geometry on the SWCNT is likely due to strong spin-orbit coupling in the d 7 configuration . The 4-coordinate planar Co (II) ion has degenerate dxz and dyz orbitals which lie below the singly occupied dxy. Upon binding to the SWCNT, the Co distorts out of the N basal plane, swapping the dxz and dyz orbitals with the dxy orbital, so that dxy becomes doubly occupied lying below the doubly occupied dxz and singly occupied dyz. |
639f5730e8047a4e4aed64a8 | 14 | We evaluated the geometries of the curved and flat CoPcH4(CO) intermediates at varying charge to understand how CO binding affinity changes with potential. We investigated the N-Co-N angle, the Co-C distance, and the C-O distance (Fig. ). At all charges, the curved CoPcH4(CO) maintains a smaller N-Co-N angle than the flat analog, indicating more out-of-plane distortion (Fig. ). For both cases, the angle decreases as additional electrons are introduced (potential becomes more negative). This is likely because electrons are occupying the dyz orbital, and to minimize overlap with the in-plane p orbitals of the ligand, the Co distorts axially out of plane. |
639f5730e8047a4e4aed64a8 | 15 | The curved CoPcH4(CO) maintains a shorter Co-C distance for all charges until n-n0 = 1.5, where the Co-C distance is 1.72 Å for both flat and curved cases (Fig. ). As electrons are added, the Co-C distance decreases, indicating stronger binding of Co to C. In this potential range, the first orbitals filled are dyz and d ! ! , both of which can participate in bonding to CO. Because Co is more distorted in the curved case, these d orbitals are more easily occupied due to decreased overlap with the PcH4 in-plane p orbitals, leading to more facile binding to CO. The curved CoPcH4(CO) maintains a longer C-O distance for all charges except for n-n0 = 1.5, where the C-O distance is 1.20 Å for both the curved and flat CoPcH4(CO). As bonding between Co and C increases, the C=O bond becomes activated, increasing the bond distance. We consider this C=O bond activation desirable since it makes it easier to reduce CO, increasing selectivity towards MeOH. Because the C-O distance is longer in the curved CoPcH4(CO), CO reduction and further conversion to MeOH will be more facile. |
639f5730e8047a4e4aed64a8 | 16 | We extended the concept of CNT-induced molecular distortion to enhance catalytic performance of other molecular systems. We investigated the oxygen reduction reaction (ORR) activity of FePc/SWCNT in O2-saturated 0.1 M KOH via measurements on a rotating disk electrode (RDE). Linear sweep voltammetry (LSV, Fig. ) indicates that FePc/SWCNT achieves higher activity than FePc/MWCNT, with a more positive onset (Eonset) and half-wave potential (E1/2) than all other samples. The E1/2 of FePc/SWCNT is 0.93 V, which is 40 mV more positive than FePc/15, FePc/50, and Pt/C. We calculated the average electron transfer number (n) from the LSV curves with different rotation rates using Koutecky-Levich (K-L) plots (Supplementary Fig. ) . We find n = 3.98 for FePc/SWCNT, close to the theoretical limit of 4.00 for a four-electron reduction process. |
639f5730e8047a4e4aed64a8 | 17 | We also measured the CO2RR performance of NiPc with three different diameter CNTs in an Hcell containing CO2-saturated 0.5 M KHCO3. All samples lead only to gas products with similar FECO (Supplementary Fig. ). However, as shown in Fig. and, the NiPc/SWCNT shows higher current density and CO turnover frequency (TOFCO) in CO2RR performance than NiPc/15 and NiPc/50. |
639f5730e8047a4e4aed64a8 | 18 | SWCNT (XFNANO, Co., Ltd.) was pretreated in 6 mol L -1 HCl solution for 12 h to remove any metal impurities. After that, the SWCNT sample was filtrated, washed with deionized water and freeze-dried. Then 20 mg of the purified SWCNT was subsequently dispersed in 20 ml of DMF using sonication. Then, an appropriate amount of CoPc dissolved in 5 ml DMF was added to the SWCNT suspension. The mixture was sonicated for 30 min to obtain a well-mixed suspension, which was further stirred at room temperature for 24 h. Subsequently, the mixture was centrifuged, and the precipitate was washed with DMF, ethanol and DI water. Finally, the precipitate was lyophilized to yield the final product. |
639f5730e8047a4e4aed64a8 | 19 | The morphology of samples was characterized using transmission electron microscopy (TEM, Philips Technai 12) equipped with energy dispersive X-ray spectroscopy. ICP-atomic emission spectroscopy (ICP-OED) measurements were conducted on Optima 8000 spectrometer. Samples were digested in hot concentrated HNO3 for 1 h and diluted to desired concentrations. UV-vis spectrum was performed on a Shimadzu 1700 spectrophotometer in ethanol solution with a concentration of 1×10 -5 mol mL -1 . The X-ray photoelectron spectroscopy data were collected on a Thermo ESCALAB 250Xi spectrometer equipped with a monochromatic AlK radiation source (1486.6 eV, pass energy 20.0 eV). The data were calibrated with C 1s 284.6 eV. Raman spectra were collected using a LabRAM HR800 laser confocal micro-Raman spectrometer with a laser wavelength of 514.5 nm. Scanning transmission electron microscopy was characterized on a double spherical-aberration-corrected FEI Themis Z microscope at 60 kV. |
639f5730e8047a4e4aed64a8 | 20 | X-ray absorption fine spectroscopy (XAFS) measurements were performed in the fluorescence mode using a Lytle detector at beamline 01C1 of National Synchrotron Radiation Research Center (NSRRC) in Taiwan. The electron storage ring was operated at 1.5 GeV with a constant current of ~ 360 mA. A Si (111) Double Crystal Monochromator (DCM) was used to scan the photon energy. XANES analysis were conducted using the Athena software based on the IFEFFIT program to determine the structural environment of Co atoms. Averaged XAS spectra were first normalized to the absorption edge height, and the background was removed using the automatic background subtraction routine AUTOBK implemented in the Athena software . A reference foil of Co foil was used for energy calibration of the monochromator, which was applied to all spectra. The Co K-edge calibration was set to the first inflection point of the reference foil, set at 7709 eV for easy comparison with other work . Quantitative information on the radial distribution of neighboring atoms surrounding Co atoms was derived from the extended absorption fine structure (EXAFS) data. An established data reduction method was used to extract the EXAFS χ-functions from the raw experimental data using the IFEFFIT software. |
639f5730e8047a4e4aed64a8 | 21 | H-cell, catalyst ink was prepared by dispersing 2 mg of catalyst in 1 mL of ethanol with 20 μL 5 wt.% Nafion solution (Sigma Aldrich, Nafion 117, 5 wt.%) and sonicated for 1 h. Then 200 μL of the ink was drop-casted on the glassy carbon working electrode and subsequently dried naturally. The loading was 0.4 mg/cm 2 . The electrochemical performance was carried out in a customized glass H-cell. A platinum and Ag/AgCl were used as the counter and reference electrodes, respectively. The working electrode was separated from the counter electrode by the Nafion-117 membrane (Fuel Cell Store). Before use, the Ag/AgCl reference was calibrated as reported . All potentials in this study were converted to the reversible hydrogen electrode (RHE) according to the Nernst equation (E (vs. RHE) = E (vs. Ag/AgCl) + 0.231+0.0592×pH). The 10 mL of 0.5 M KHCO3 solution electrolyte was added into the working and counter compartment, respectively. The cell was purged with high-purity CO2 gas (Linde, 99.999 %, 20 sccm) for 30 min prior to and throughout the duration of all electrochemical measurements. The electrochemical measurements were controlled and recorded with a CHI 650E potentiostat. The automatic iR (85%) compensation was used. Gas-phase products were quantified by an online gas chromatograph (Ruimin GC 2060, Shanghai) equipped with a methanizer, a Hayesep-D capillary column, a flame ionization detector (FID) for CO, and a thermal conductivity detector (TCD) for H2. The CO2 flow rate was controlled at 3 sccm using a standard series mass flow controller (Alicat Scientific mc-50 sccm). The liquid products were quantified after electrocatalysis using 1 H NMR spectroscopy with solvent (H2O) suppression. 450 μL of electrolyte was mixed with 50 μL of a solution of 10 mM dimethyl sulfoxide (DMSO) in D2O as internal standards for the 1 H NMR analysis. The concentration of MeOH was calculated using the ratio of the area of the MeOH peak (at a chemical shift of 3.32 ppm) to that of the DMSO internal standard. Electrolyte of 13 CO2 in 0.5 M KH 13 CO3 were prepared by bubbling 13 CO2 into 0.5 M KOH for more than 30 minutes. |
639f5730e8047a4e4aed64a8 | 22 | For CO2RR flow electrolysis, in order to get a good FEMeOH in the flow cell, CoPc/SWCNT with high CoPc loading catalysts was prepared using a CVD-type procedure according to our previous work . Then 5 mg catalysts mixed with 40 μL Nafion solution were deposited in 2 mL ethanol and sonicated for 1 h to form uniform ink and then drop-casting on 1×2.5 cm 2 GDL (Sigracet-28BC) (mass loading of the sample: 1 mg/cm 2 ). GDL with catalysts as a CO2RR cathode. A platinum and Ag/AgCl were used as the counter and reference electrodes, respectively. The cathode chamber and anode chamber were separated by Nafion-117 membrane (Fuel Cell Store). The CO2 gas flow with 10 sccm flow rate was conducted on the cathode side while 0.1 M KOH+3 M KCl and 1 M KOH electrolyte at 5 mL/min flow rate was circulated in cathode and anode chamber, respectively. The cathode electrolyte was collected in a flask with an ice bath for NMR testing. |
639f5730e8047a4e4aed64a8 | 23 | For FePc/CNT ORR measurement, 2 mg of catalyst was dispersed in 1 mL of solution containing 0.882 ml of ethanol and 0.098 ml of water and 20 μL of 5 wt.% Nafion solution, which was sonicated for 1 h to form a homogeneous catalyst ink. All the catalysts were cast onto the RDE (diameter 5 cm) with a loading amount of 0.2 mg/cm 2 . |
639f5730e8047a4e4aed64a8 | 24 | RDE tests were performed in O2 saturated 0.1 mol/L KOH solution with a scan rate of 10 mV/s between 1.1 V and 0.2 V at different rotating rates using PINE 636 rotating-disk electrode system and CHI650 workstation. Ag/AgCl and Pt were used as reference and counter electrodes, respectively. All potentials were converted to the reversible hydrogen electrode (RHE). |
639f5730e8047a4e4aed64a8 | 25 | where J is the measured current density, JK and JL are the kinetic and limiting current densities, ω is the angular velocity of the disk, n is the overall number of electrons transferred in oxygen reduction, F is the Faraday constant (96485 C/mol), C0 is the bulk concentration of O2 (1.2 × 10-6 mol cm-3), D0 is the diffusion coefficient of O2 in 0.1 M KOH (1.9 × 10 -5 cm 2 s -1 ), and V is the kinematic viscosity of the electrolyte (0.01 cm 2 s -1 ). |
639f5730e8047a4e4aed64a8 | 26 | For NiPc/CNT CO2RR measurement, 2 mg of catalyst in 1 mL of ethanol with 20 μL 5 wt.% Nafion solution and sonicated 1 h to form a homogeneous catalyst ink. All the catalysts were dropcast onto carbon paper (Toray, TGP-H-060, Fuel Cell Store) (diameter of 0.5 inch) with a 0.4 mg/cm 2 loading. The electrochemical performance was carried out in a customized threecompartment cell, as previously reported . A platinum foil and Ag/AgCl leak-free reference (LF-2, Innovative Instrument Inc.) were used as the counter and reference electrode, respectively. The working electrode was separated from the counter electrode by the Nafion-117 membrane. |
639f5730e8047a4e4aed64a8 | 27 | Molecular dynamics simulations were performed using the LAMMPS software . For these simulations we used the Universal Force Field (UFF) for valence interactions (bond, angle, and dihedral terms) combined with the RexPoN universal nonbond potentials (UNB) to describe van der Waals interactions, and the polarizable charge equilibrium (PQEq) scheme for electrostatics. |
639f5730e8047a4e4aed64a8 | 28 | Density Functional Theory geometry optimizations were performed using VASP 5.4.4 with the solvation module . For the curved species, the edge H atoms of the Pc ligand were fixed to maintain the curvature. Spin polarization was allowed during optimization. We used the PBE functional with the D3 45 empirical correction for London dispersion forces. The kinetic energy cutoff was set to 500 eV, the wavefunction cutoff was set to 1E-5 eV, and the force cutoff was set to 0.03 eV/Å. All VASP optimizations were in a 20 Å 3 box with a 1x1x1 K point Monkhorst Pack grid. |
639f5730e8047a4e4aed64a8 | 29 | To obtain the energy as a function of applied potential, we performed single point energy calculations using JDFTx with the CANDLE 47 solvation model. Because our systems are finite (non-periodic), we were able to perform vibrational frequency calculations using Jaguar 48 v10.9, to obtain mode-dependent entropies, zero point energies, and enthalpies at 298.15 K. |
639f5730e8047a4e4aed64a8 | 30 | In the Grand Canonical Potential (GCP) method, we first calculate the free energy (F) as a function of the number of electrons (n). F(n) includes the librational and vibrational contributions to the zero point energy, entropy, and enthalpy at 298.15 K. F(n) has a quadratic form, which we write as Equation : |
639f5730e8047a4e4aed64a8 | 31 | where the a, b, and c parameters are fitted to the QM calculations. Here a should be positive to obtain a stable system (minima as opposed to maxima at n = n0) and n0 is the number of explicit electrons for a neutral system (explicit because we utilize pseudopotentials). The quadratic form of F(n) is strictly verified in our calculations. |
60c749e0469df4860ef43c54 | 0 | The human ether-a-go-go-related gene (hERG) encodes for the poreforming alpha-subunit of voltage-gated potassium ion channels. The hERG channel regulates the efflux of potassium ions in cardiac myocytes and thereby plays a key role in coordination of heartbeat. Literature indicates that blockade of hERG channel leads to prolonged QT interval of the action potential which can result in fatal cardiac arrhythmia (Torsade de pointes). Several marketed drugs, including antiarrhythmic agents, were withdrawn after being reported to trigger cardiac arrhythmia that sometimes led to sudden death. Consequently, hERG channel emerged as an important off-target, marking early assessment of hERG liability an essential step in drug discovery. The gold-standard in vitro and in vivo assays that facilitate screening of hERG channel inhibition are expensive and provide low throughput. Meanwhile, in silico methods emerged as an alternative for early assessment of pharmacological and toxicological effects of chemical substances. Multiple studies reported in silico models for predicting hERG channel inhibition over the past several years. Tropsha et al. provided an overview of quantitative structureactivity relationship (QSAR) studies from the literature that were reported before 2014. More recently, several other studies reported models based on simple methods like read across and machine learning (ML) methods , including deep neural networks (DNNs). Many models are based on proprietary or in-house datasets which restricts the use of this data to build newer models for academic drug discovery. It is suspected that hERG channel can bind a wide variety of chemotypes and a major limitation to developing robust prediction models using publicly-domain hERG activity data is the fairly limited chemical diversity of available training data. Although different combinations of ML algorithms and molecular descriptors have been tested, there is no combination of choice that performs well on unseen data. For instance, a consensus of Support Vector Machines, Random Forests, Gradient Boosting Model and Tree Bagging provided better performance in comparison to individual models that performed very similar to each other when used with different descriptors. Recent studies suggest that neural networks based on learnable representations offer broadly a better performance than classical ML algorithms. Latent descriptors that are derived directly from the neural network architecture are gaining popularity in molecular property prediction. Furthermore, descriptor-free QSAR models that are based on recurrent neural networks (RNNs) and molecular representations like SMILES were reported to demonstrate superior generalization capabilities on out-of-domain test data. Despite this body of literature, no one has attempted to compare these methods against each other in a prospective validation study. In this study, we enriched the public domain hERG data with a dataset comprising bioactive compounds screened in a homogeneous high-throughput assay of hERG channel inhibition to provide the community with a large reference dataset of high integrity. The primary goals of the study are to develop prediction models representing both classical and novel AI developments, validate the best models on prospectively screened compounds, as well as on recently approved FDA drugs with hERG liability data. |
60c749e0469df4860ef43c54 | 1 | Public Domain hERG Bioactivity Data. ChEMBL has been a major repository in the public domain for compound activity data extracted from scientific literature. The most recently updated ChEMBL database (version 25, accessed 28 October, 2019) provides more than 20000 activity records for hERG channel (UniProt accession: Q12809). The activity records were preprocessed, as previously described in literature, to generate a high-confidence bioactivity dataset. Briefly, only the potency and affinity values reported as IC50, EC50 |
60c749e0469df4860ef43c54 | 2 | Thallium Flux Assay Data. A high-throughput ion channel screen was developed and validated at National Center for Advancing Translational Sciences -NCATS (formerly known as the NIH Chemical Genomics Center) as a modified version of the FluxOR TM thallium flux assay that detects inhibition of the hERG channel by measuring flow of a surrogate ion, thallium. The study compared the activities of 10 common hERG inhibitors measured in thallium flux and patch-clamp experiments and concluded that the homogeneous high-throughput assay can be used as a cost-effective alternative to patch-clamp technique. More recently, NCATS reported a collection of 4,323 compounds screened in thallium flux assay used for generating support vector classification models of hERG channel inhibition. In the present study, we aimed to merge this dataset with the high-confidence activity data obtained from ChEMBL database for development and critical assessment of modern machine learning approaches. First, we analyzed the correlation between flux assay data and ChEMBL data that originated from multiple assay types (e.g. electrophysiology, ion flux, radioligand binding, fluorescence, etc.). A subset of 86 approved drugs with activity data from both sources was identified for this purpose. However, since ChEMBL data spans multiple assay types, the correlation analysis was performed twice, first with patch-clamp data alone (Figure ) and next with data obtained after merging activities from different assay types (Figure ). We noticed that the outcomes from different assay types correlated well with the patch-clamp data, suggesting that the data could be used together. Furthermore, given the concordance between ChEMBL data and flux assay data, we decided to merge the two datasets to generate a combined dataset after removal of duplicates. for training the classification models. Additionally, a collection of 840 compounds was selected from an in-house library to generate a test dataset for prospective validation of the classification models. These compounds were measured in the same thallium flux assay to generate IC50 data. We did not use a single activity threshold to discriminate blockers from non-blockers in ChEMBL whole-cell patch clamp data. Previous studies using a binary threshold (1𝜇M and 10 𝜇M) provided superior performance as compared to using a single threshold (1𝜇M or 10 𝜇M). For the thallium-flux assay, a threshold of 30𝜇M was used considering the average fold difference in the activity for compounds with data available from both sources. Finally, whole-cell patch clamp hERG data was manually extracted for 177 FDA approved drugs (2012 to 2018) from their pharmacological and safety reviews. Many previously published studies do not include such data into their training or validation sets. We believe that validation on this data offers a more realistic evaluation setting for our models. Since it is hard to determine the hERG liability of a drug based on the IC50 value alone, without the knowledge of the peak serum concentration unbound to plasma proteins, we use both activity thresholds (1𝜇M and 10 𝜇M) to classify the drugs as blockers and nonblockers. Table summarizes the datasets used for modeling and validation. Five different sets of molecular descriptors were calculated and used in combination with different methods in this study. Morgan fingerprints (1024 bits) and RDKit descriptors were calculated using the RDKit toolkit. Recent progress in deep learning facilitates the development of the different molecular representations such as latent vectors that are used as descriptors for modeling molecular properties. In this study we utilized several different approaches to generate fixedlength latent vector representations for our modeling sets. The detailed methodology involved in generation of latent vectors is described in the next section. RNNs that learn on sequential data such as SMILES as input have been employed for molecular property prediction. Three variations of SMILES including the commonly known canonical SMILES, randomized SMILES and DeepSMILES were employed as input for RNN models in this study. While the RDKit toolkit was used to generate canonical SMILES, we relied on their original implementations for randomized SMILES and DeepSMILES. |
60c749e0469df4860ef43c54 | 3 | Autoencoder and Adversarial Autoencoder Models. RNNs that encode SMILES and decode them back to SMILES constitute the most recent generation of methods for ligand-based de novo design. In order to generate new structures, an RNN learns to predict the probability of the next character in a SMILES string, given the previous characters. An autoencoder (AE) constitutes a special architecture that generates a compressed representation of the provided input that could be used to reconstruct the chemical structures. The code layer (i.e., the compressed representation) produced by an autoencoder is a fixed-length vector of descriptors increasingly referred to as 'latent descriptors' that have been used for molecular property prediction. Variations of the original autoencoder architecture have been proposed that translate between different string representations of molecules (e.g., canonical SMILES, InChI, etc.). One such variation focused on the ability of AEs to generate new samples which resulted in variational autoencoders (VAEs). an open-source ML library was used to train and validate the models. The second pipeline consisted of two types of neural networks: feed-forward neural networks based on the four descriptors and RNNs based on SMILES. |
60c749e0469df4860ef43c54 | 4 | We implemented two types of data splits: random split and scaffold split, both adapted from the DeepChem 51 library. The training set was partitioned into internal training (80%) and test (20%) sets. Parameter optimization was performed on these partitions in a five-fold crossvalidation format. For this purpose, the split was performed five times each for both split types. Finally, models based on the best parameters were built using the unpartitioned training set and evaluated on the prospective validation set and FDA approved drugs. All learning methods are briefly explained below. |
60c749e0469df4860ef43c54 | 5 | Random forest (RF) is an ensemble of decision trees that are fitted on various subsamples of the data and uses averaging to restrict overfitting and improve prediction accuracy. The 'RandomForestClassifier' method from Scikit-learn was used to build the model. The number of estimators was set to 300 and random state was set to an integer. The rest of the parameters were set to default values. |
60c749e0469df4860ef43c54 | 6 | Gradient Boosting eXtreme Gradient Boosting (XGBoost) is an ML method that allows both regression and classification. It is based on the Gradient Boosting Decision Tree technique and has been widely applied in the field of data mining. Due to its recently gained popularity over RF in cheminformatics, we used XGBoost as the second baseline method. Similar to RF, XGBoost models were used with a total of 300 estimators and random state set to an integer. The remaining parameters were set to default values. |
60c749e0469df4860ef43c54 | 7 | Artificial neural networks (ANNs) have been applied for a wide range of QSAR tasks. Increase in the use of RF and Support Vector Machines for classification and regression in cheminformatics led to a decline in the use of ANNs. Eventually, the ANNs have evolved into DNNs. Unlike ANNs, a DNN consists of multiple fully connected layers with two or more hidden (or intermediate) layers between the input and output layers. In a feedforward neural network (referred simply as DNN in the rest of the study), the information passed through the input layer flows in forward direction through the hidden layers to the output layer. A number of parameters are available for tuning a DNN such as the number of hidden layers, number of epochs, activation function, optimizer and its learning rate. Hyperparameter optimization is essential to improve the performance of DNN and avoid overfitting on training data. This is detailed in the Model Optimization section of the results. |
60c749e0469df4860ef43c54 | 8 | Long Short-Term Memory (LSTM) networks are RNNs that can be used to model sequence data such as natural language. Previous studies reported the use of LSTMs to learn directly from SMILES which led the community towards descriptor-free QSAR models. The LSTM networks built in this study were fed with canonical SMILES that are first encoded into onehot vectors and then passed to the computing cell which performs as many computations as the length of the input SMILES in a loop. At each step, one character of SMILES is taken as input and the computed activation value is passed to the next step which takes the next character as input. In this way, the information from previous characters is persisted while the next characters are being processed. Finally, the network produces a prediction probability between 0 and 1. These values can be used to obtain the binary classification labels. |
60c749e0469df4860ef43c54 | 9 | Furthermore, we investigated attention-based modeling in which the neural network architecture is extended to search for parts of the input sequences that are relevant to the target variable. In case of LSTM networks, the attention mechanism gives importance to certain parts of the sequence (i.e., SMILES) rather than considering the whole sequence as important. For this purpose, we implemented Multiplicative Attention from Keras Self-Attention library with regularization and without any attention bias. |
60c749e0469df4860ef43c54 | 10 | Performance Assessment. The performance of the models was mainly accessed using the area under the curve (AUC) from the receiver operating characteristic (ROC) curves. A ROC curve plots the true positive rate against the false positive rate and thus provides an estimate of the performance of a binary classifier. In addition to AUC, the following metrics were estimated: |
60c749e0469df4860ef43c54 | 11 | Model Optimization. The baseline methods Random Forests and XGBoost are robust and do not require extensive parameter optimization. The quality of deep learning models is more dependent on the number of descriptors, hyperparameters and computational capabilities (e.g., use of GPU). We first report parameter optimization performed using one of the five-fold crossvalidation training sets. For the DNN, a series of 260 models were built using different combinations of optimizer learning rate, activation function, number of epochs and batch size for each descriptor type. The same dense layer architecture was maintained for the first round of grid search and once the best parameters were obtained, we tried to find an optimal dense layer structure for each descriptor. The DNNs in this study typically consisted of three to five layers with decreasing number of neurons as it moves forward that resulted in a pyramidal network structure, previously reported as an optimal setting for DNNs. The number of units in the input layer was defined based on the shape of the incoming descriptors and was reduced in the hidden layers and finally, the output layer consists of only one unit which uses a sigmoid activation function to return the output. Different combinations of the number of hidden layers and the number of neurons per hidden layer were examined and the best performing architecture was retained for both five-fold cross-validation and final validation. In the case of LSTM, along with the parameters considered for DNN, we also investigated the number of LSTM units. The grid search results for DNN are provided in the supporting information (S3). |
60c749e0469df4860ef43c54 | 12 | While ) and BACC (Table ) with RDKit descriptors. However, these two methods provided the worst performance (Sensitivity < 0.6) when used together with the latent descriptors. In contrast, the DNN models provided better performance (BACC > 0. For each method-descriptor combination, the standard deviation of the average of performance for different folds (N=5) is presented as an error bar. |
60c749e0469df4860ef43c54 | 13 | The prospective validation set containing 839 compounds was used to evaluate the models. A nearest-neighbor analysis with the training set revealed that a majority (>80%) of these compounds are below a Tanimoto similarity threshold of 0.6 (supporting information, S5). The optimal settings from cross-validation were retained for DNN and LSTM models. A performance trend similar to cross-validation was observed ( In particular, the latent descriptors derived from encoder-decoder architectures performed very well on validation set and emphasize their applicability in prediction of molecular properties and biological activity. However, classical ML methods such as XGBoost and RF are still in the league of best performing models, in agreement with previous studies. The original implementation of this VAE model, built on a subset of ZINC database, was used to generate latent descriptors of length 192 bits for both training and validation set compounds. DNNs were used to evaluate the utility of these descriptors for prediction of hERG channel blockade. Hyperparameter optimization was performed similar to other descriptors. This model did not perform as well as the DNN models based on our latent descriptors (Table ). This could be due to the fact that the latent space of the VAE model was originally shaped for predicting specific molecular properties such as the water-octanol partition coefficient. We also noticed that the reconstruction rate of the encoder-decoder model can influence the QSAR model performance. An inverse correlation was observed between the reconstruction rate of the encoder-decoder models and the improvement in performance of QSAR models using the latent descriptors. Considering this into account, we trained our AE and AAE models in a small number of epochs to limit the reconstruction rate and obtain optimal performance using the latent descriptors. However, a detailed investigation to arrive at the best reconstruction rate could not be performed due to the huge computational costs involved development of these models. DeepSMILES that could be used instead of the conventional SMILES representations in building generative neural networks. They tried to address the syntactical limitations of SMILES that could be a reason behind the poor validity of the newly generated structures. In another benchmark, canonical SMILES and DeepSMILES were compared to 'Randomized SMILES' for the development of generative RNN models. Randomized SMILES were earlier proposed as a data augmentation technique to improve the performance of QSAR models. Further, they were also shown to improve the relevance of latent descriptors for QSAR when used in generation of autoencoder models. In this study, we developed LSTM models using these two SMILES adaptations and compared the performance with our best LSTM model based on canonical SMILES. In the case of Randomized SMILES, different enumeration factors (e = 2, 3, 4, 5) were considered i.e. in case of e = 5, five unique randomized SMILES were generated for each molecule in the training set. In all cases, the LSTM model started to provide higher Sensitivity although the overall performance declined. Similarly, DeepSMILES did not perform as well as the canonical SMILES (see Table ). In order to evaluate DeepSMILES on a larger dataset, the AE model developed in this study was rebuilt using the same ChEMBL data but this time using DeepSMILES. Again, the AE model based on canonical SMILES resulted in a higher reconstruction performance and the latent descriptors derived from the same model provided better QSAR performance. . With more stringent activity criteria, the consensus model achieved a BACC of 0.79. Similar to the validation set, a majority (>80%) of these drugs were found to be below a Tanimoto similarity threshold of 0.6 (supporting information, S5). Thus, we demonstrated the ability of our models to provide robust predictions on unseen chemical space. At the same time, it is clear that the activity threshold used to separate blockers from non-blockers can result in a completely different dataset and model performance. While 10𝜇M is the generally accepted threshold, in the case of this dataset, we believe that 1𝜇M offers a realistic composition with more non-blockers than blockers. Furthermore, no clear trend was noticed in the evaluated time period (2012 to 2018) for the potential of newly approved drugs to inhibit hERG (see Figure ), while the expectation was to observe a decrease in the inhibitory potential over the time. This should draw the attention of the community to the question -is hERG blockade still a concern for drug discovery? and Latent2). The performance obtained using latent descriptors from AE and AAE models was comparable to that obtained using fingerprints and other descriptors only when employed with DNNs. The poor performance of RF and XGBoost models with latent descriptors could be attributed to the continuous distribution of the compounds in the low-dimensional space. |
60c749e0469df4860ef43c54 | 14 | Overall, MorganFP performed the best among all numerical descriptors. The PCA plots in Figure indicates that the blockers could be better discriminated from the non-blockers by MorganFP. The continuous distribution of compounds in the latent space explains the poor ability of simple classifiers such as RF and XGBoost to distinguish blockers from nonblockers. Previously, these representations have been shown to provide improvements over baseline models based on molecular fingerprints. It is also worth noting that these representations are not only useful in reconstruction of molecules but also in capturing properties of molecules that include biological activity. Generating compounds without hERG liabilities. The recently introduced sequence-tosequence based models that rely on the encoded representation (i.e. latent space) of molecules facilitates exploration of new chemical space. Apart from its novel applicability in QSAR modeling and virtual screening , the encoded representation has been explored to generate focused chemical libraries with molecular properties of interest. A number of key factors such as validity, novelty, diversity and synthetic feasibility of the sampled molecules have been addressed. Such models have been recently reported to identify promising drug candidates. In this context, the AAE architecture was used to sample new compounds using hERG blockers and non-blockers as separate starting points. Distribution of the prediction probabilities for the newly generated compounds (using the consensus model) revealed that most compounds generated around non-blockers were predicted as non-blockers by the consensus model (Figure ). Similarly, a majority of new structures sampled from the blockers were predicted as blockers (Figure and Figure ). The activities of new compounds were predicted using the consensus model. |
60c749e0469df4860ef43c54 | 15 | Although synthesizability of the generated structures is a bottleneck for generative models, it was recently demonstrated that the fraction of synthesizable molecules is comparable to that of training set used to derive the new compounds. Since our training set compounds originate from ChEMBL database that reports bioactivities for already synthesized compounds and in-house high-throughput assay, it is our expectation that the newly generated compounds have similar rate of synthesizability. Furthermore, the generated chemical structures are fairly diverse and not completely similar to the original training set subsets used for sampling (Figure ; supporting information, S8). These findings emphasize the potential of generative models in designing new chemical libraries with desired properties (or poor toxic liabilities), particularly in combination with predictive models. CONCLUSIONS. Modeling hERG channel inhibition has been important ever since the recall of marketed drugs due to fatal cardiac arrhythmias. To date, several computational modeling approaches have been proposed for early assessment of hERG liability and several in silico models have been reported in the recent years. In this study, both classical and modern learning approaches were evaluated and compared for their ability to predict hERG liabilities of small molecules. Both feed-forward neural networks (DNN models) and recurrent neural networks (LSTM models) performed on par with classical machine learning methods. It was also demonstrated that novel representations derived from the latent space of chemical autoencoders offer an alternative to traditional descriptors in structure-activity and structure-property modeling. Particularly, the DNNs provided a significantly better performance using these novel descriptors. Further, the utility of generative models to derive a new chemical space with a property of interest has been demonstrated. In addition, we also provide a high-quality reference dataset obtained by combining public domain hERG activity data with experimental data generated in a high-throughput thallium flux assay as well as hERG activity data for small molecules approved between 2012 and 2018. The validation data from this study can be used to evaluate hERG models proposed in future studies. |
624f04d08d07c63a2e829977 | 0 | All MEPA compounds were synthesized by Knoevenagel condensation of appropriate benzaldehydes with 2-methoxyethyl cyanoacetate, catalyzed by base, piperidine (Scheme 1). The preparation procedure was essentially the same for all the MEPA compounds. In a typical synthesis, equimolar amounts of 2-methoxyethyl cyanoacetate and an appropriate benzaldehyde were mixed in equimolar ratio in a 20 mL vial. A few drops of piperidine were added with stirring. The product of the reaction was isolated by filtration and purified by crystallization from 2-propanol. The condensation reaction proceeded smoothly, yielding products, which were purified by conventional techniques. The compounds were characterized by IR, 1 H and 13 C NMR spectroscopies. No stereochemical analysis of the novel alkoxy ring-substituted MEPA was performed since no stereoisomers (E or/and Z) of known configuration were available. |
624f04d08d07c63a2e829977 | 1 | Copolymers of the ST and the MEPA compounds, P(ST-co-MEPA) were prepared in 25-mL glass screw cap vials at ST/MEPA = 3 (mol) the monomer feed using 0.12 mol/L of ABCN at an overall monomer concentration 2.44 mol/L in 10 mL of toluene. The copolymerization was conducted at 70ºC. After a predetermined time, the mixture was cooled to room temperature, and precipitated dropwise in methanol. The composition of the copolymers was determined based on the nitrogen content (cyano group in MEPA monomers). The novel synthesized MEPA compounds copolymerized readily with ST under free-radical conditions (Scheme 2) forming white flaky precipitates when their solutions were poured into methanol. The conversion of the copolymers was kept between 10 and 20% to minimize compositional drift (Table ). Nitrogen elemental analysis showed that between 17.2 and 29.6 mol% of MEPA is present in the copolymers prepared at ST/MEPA = 3 (mol), which is indicative of relatively high reactivity of the MEPA monomers towards ST radical which is typical of halogen ring-substituted phenylcyanoacrylates. Since MEPA monomers do not homopolymerize, the most likely structure of the copolymers would be isolated MEPA monomer units alternating with short ST sequences (Scheme 2). |
65f04fd966c1381729d79a16 | 0 | Halohydrin dehalogenases (HHDHs) have recently distinguished themselves as powerful enzymes for the asymmetric diversification of oxyfunctionalized synthons. In nature, these bacterial lyases catalyze the reversible dehalogenation of β-haloalcohols through formation of the corresponding epoxides. More importantly, they are capable of accepting a number of anionic C-, N-, O-, and S-nucleophiles in the reverse reaction, i.e. epoxide ring opening, giving access to a large repertoire of valuable products. For instance, recent impressive biocatalytic examples for the application of HHDHs in asymmetric synthesis include the preparation of enantiopure βnitroalcohols , and thiiranes , as well as the desymmetrization of 2-substituted-1,3-dichloro-2-propanols with subsequent cyanate-mediated ring opening to afford optically pure epoxides and oxazolidinones among others. 7,13,14 HHDHs share significant homology/similarity with short-chain dehydrogenases and reductases (SDRs) on the sequence, structural and mechanistic level as the result of a close phylogenetic relationship, although they catalyze entirely different chemical reactions. Previously, this similarity has largely impeded a fast discrimination of HHDHs and SDR enzymes solely based on given protein sequences. With the discovery of HHDH-specific sequence motifs in 2014, database mining approaches facilitated the identification of novel HHDHs and yielded a plethora of new enzymes with partially unprecedented catalytic characteristics. These HHDH sequence fingerprints include the N-terminal motif T-X4-F/Y-X-G (motif 1), lining the nucleophile binding pocket of halohydrin dehalogenases, as well as the motif S-X12-Y-X3-R (motif 2), covering the catalytic residues serine, tyrosine and arginine. For comparison, the corresponding sequence motifs of SDRs are T-G/A-X3-G/A-X-G (motif 1) and S-X10-14-Y-X3-K (motif 2), respectively. In contrast to HHDHs, SDR motif 1 represents the typical glycinerich nucleotide-binding sequence required for nicotinamide cofactor binding in those enzymes. While the mechanistic role of the three conserved catalytic residues serine, tyrosine and arginine (as part of motif 2) of HHDHs has already been elucidated using HheC from Agrobacterium radiobacter AD1 as a model enzyme, the functional role and mutability of the conserved residues in motif 1 remains largely unexplored. Moreover, we hypothesized that HHDH variants with substantially improved biocatalytic performance could be accessed through engineering of motif 1 residues, as they should impact nucleophile binding. Crystal structures of HheC in complex with a bromide or chloride ion (PDB IDs 1PWX and 1PWZ, respectively) have demonstrated that the side chain of the central aromatic residue F12 in motif 1 forms a direct interaction with the negatively charged halide, which is tightly wedged between aromatic residues. Such a direct interaction with the nucleophile, however, was not present in the crystal structure of HheB from Corynebacterium sp. strain N-1074 in complex with chloride (PDB ID 4ZD6), which carries a tyrosine at position 19. Moreover, in a study by Wu et al. on the thermostabilization of HheC by combinatorial directed evolution, mutation F12Y was found to increase the enzyme's thermostability and yielded a 1.5-fold gain in specific activity in the dehalogenation of 1,3-dichloropropanol. Similarly, mutations at the structurally equivalent position Y15 in AbHheG from Acidimicrobiia bacterium yielded variants with improved enantioselectivity in the ring opening of styrene oxide with cyanate quite recently. Both examples directly hint at the hitherto underexplored possibility to steer the catalytic properties of HHDHs via targeted mutagenesis of sequence motif 1. |
65f04fd966c1381729d79a16 | 1 | Building on these precedents, we therefore set out to systematically engineer the conserved motif 1 residues in two representative HHDHs, HheC from A. radiobacter as well as HheG from Ilumatobacter coccineus. We selected these HHDHs based on their biocatalytic relevance, the fact that both enzymes offer highresolution crystal structures, as well as their highly dissimilar catalytic properties. HheC is the by far best studied member of the HHDH family, generally displaying high catalytic activity and enantioselectivity in the dehalogenation of various substrates as well as in the ring opening of terminal epoxides. Moreover, numerous protein engineering studies of HheC have been published aiming either at an increased activity, enantioselectivity or stability of the enzyme. In contrast, HheG is rather thermolabile and much less selective. However, this HHDH was the first reported enzyme with relevant ring-opening activity towards sterically more demanding internal epoxides (cyclic as well as acyclic). Even though less protein engineering data are available for HheG so far, the enzyme exhibits distinct structural differences compared to HheC. This includes a much broader and open active site, an additional α-helix in the nucleophile binding site loop, as well as a long and highly flexible loop in the N-terminal part of the protein, modulating substrate access to but possibly also substrate binding in the active site. In this work, we experimentally examined all possible single mutants of HheC and HheG with defined amino acid exchanges at the three conserved motif 1 residues, namely threonine, phenylalanine/tyrosine and glycine (Figure ), to investigate the impact of each residue on enzyme activity, selectivity and stability. In this regard, our in-depth characterization revealed HheC F12Y to be enhanced considerably and towards multiple parameters. Complementary molecular dynamics (MD) and quantum mechanics (QM) simulations highlighted the formation of additional hydrogen bonds in mutant F12Y compared to HheC wild type, resulting in a better preorganization of the active site and a lower activation barrier for the epoxide ring-opening reaction. In addition, a quantitative workflow for spectrophotometric activity determination in epoxide ring opening reactions has been developed, enabling also fast and reliable kinetic measurements. |
65f04fd966c1381729d79a16 | 2 | To study the impact of motif 1 mutations on the activity, selectivity and stability of HHDHs, each of the three conserved residues of this motif (T-X4-F/Y-X-G) in HheC and HheG was individually replaced by the corresponding other 19 proteinogenic amino acids using either a Golden Gate-based or MEGAWHOP mutagenesis strategy. This resulted in a total of 112 defined single mutants, of which all could be generated successfully, except for mutants F12M and F12R of HheC. Subsequent heterologous production of all 110 mutants in E. coli BL21(DE3) in 15 mL scale and partial purification via N-terminal His-tag using gravity flow revealed that only few variants per position, usually carrying exchanges to structurally similar amino acids, still yielded observable amounts of soluble HHDH (Figure and Figure ). Hence, this result suggests a possible direct impact of motif 1 residues on protein folding and stability. As an exception, position F12 of HheC permitted more diverse mutations as almost all variants could be obtained as soluble proteins (Figure and Figure ). |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.