id
stringlengths 24
24
| idx
int64 0
402
| paragraph
stringlengths 106
17.2k
|
|---|---|---|
630f3a8858843b843ca424ff | 33 | Whilst the graph algorithm exhibits rapid convergence with grid size, the convergence is quite noisy -as can be seen in detail in Fig. . However, this noise is on the order of (or smaller than, in the case of the kinetic energy) the difference between the lunatic and extra-lunatic Multiwfn grids, using 4 orders of magnitude fewer grid points than the latter. One potential method to reduce this convergence noise would be to assign fractional basin weights from each grid point to each basin, as illustrated in Fig. . Given that we already construct the Delaunay triangulation, which allows for rapid calculation of the Voronoi tiling, the weighting method proposed in Ref. would be particularly suitable. |
630f3a8858843b843ca424ff | 34 | As noted in Ref. for uniform grids, topographical properties (such as Bader charges) can be affected by a grid bias, whereby a systematic error arises due to geometrical properties of the arrangement of grid points. This error is due to steepest-ascent paths on the graph diverging from the true gradient path of the function, and, remarkably, persists even in the infinite-grid-density limit. In particular, gradient paths which are nearly, but not quite, aligned with move directions on the graph will lead to gradually-diverging steepest-ascent paths, as can FIG. . The divergence of steepest on-graph paths from the true gradient βf for different grids. At each step of the steepest-ascent path the gradient is followed as closely as possible on the graph, but small errors at each step accumulate and the paths eventually diverge. Note that the diagonal moves introduced into a uniform grid via triangulation help (more so in 3D), but do not remove the problem. The easiest way to see that this problem is scale independent is by considering the upper-left uniform grid case -the steepest ascent path will always be "upward" (the horizontal moves will never be taken), regardless of the grid spacing. |
630f3a8858843b843ca424ff | 35 | Ref. provides a solution to the grid bias problem by allowing the trajectory of ascent paths to go "off-grid". For the graphs employed in this work, an analogous "offgraph" method can be straightforwardly implemented by allowing our ascent path to move freely in R N , following the nearest neighbour gradient g N N (x) given by βf |
630f3a8858843b843ca424ff | 36 | The nearest-neighbour lookup x β G (see Eq. 11) can be implemented efficiently as a KD-tree . An example of the resulting off-graph gradient paths is shown in Fig. (bottom). In Fig. , it is clear that this technique corrects the grid bias for a DFT grid so that it agrees with the (also corrected) uniform result. The uniform grid shows significantly less grid bias in Fig. due to the inclusion of diagonal moves by the Delaunay triangulation (the DFT grid also has such diagonal moves, but they are less helpful as a significant radial bias remains). These diagonal moves were not present in previous uniform-gridbased approaches , which therefore exhibited significantly larger grid bias than the present method. We note that the methods developed by RodrΓguez et. al. are inherently off-grid and so do not suffer from grid-bias, at the cost of requiring it to be possible to evaluate f and βf freely. In contrast, Eq. 21 requires no evaluations of f beyond those given as input (Eq. 1), at the expense of a KD-tree lookup. |
630f3a8858843b843ca424ff | 37 | The rate-limiting step in performing topological analysis via a graph over grid points is the construction of the Delaunay triangulation which scales as O(N log(N )) in the number of grid points N . This scaling is reflected in our calculation times (see Fig. ). Note that we use a largely unoptimized python code, so the absolute time shown on the y axis of Fig. could be improved relatively easily if desired, but the scaling will remain O(N log(N )). |
630f3a8858843b843ca424ff | 38 | A method has been presented to extract topographical and topological properties of a function defined on an arbitrary set of points in space, strictly via postprocessing with no additional function evaluations. By connecting the points with a neighbourhood graph, welldefined and robust algorithms were developed that allow for identification of local and global maxima (both pointlike and spatially-extended), saddle points, critical paths (and their critical points) and basins of attraction. By simple transformations of the function, one can also identify local and global minima, isosurfaces and stagnation graphs. Applications of the analysis was demonstrated for a few problems in quantum chemistry including Bader charge and bond analysis, identification of valence shells and their charge concentration, location of lone pairs via the electron localization function or the Laplacian of the electron density and identification of stagnation graphs of magnetically-induced currents. All of these investigations were carried out directly on a real-space integration grid used in density functional theory (DFT) calculations, allowing the results to be easily, efficiently and directly incorporated into DFT calculations. The analysis was found to scale as O(N log(N )) where N is the number of grid points and quantities of interest were found to converge rapidly N , requiring orders of tude fewer grid points than uniform-grid methods. Topographical results calculated using such DFT grids were found to exhibit a significant "grid bias" when the algorithm was constrained to stay on the graph. The source of this bias was analysed and found to be removed by allowing off-graph moves. Grant topDFT (grant agreement no. 772259). We are grateful for access to the University of Nottingham High Performance Computing facility. |
6655a628418a5379b0782c70 | 0 | These are cross-linked by nucleophilic attack from amino groups in the hardener, forming rigid bonds that prevent crystallisation, and instead form a complex, amorphous, 'glassy' structure with a high glass transition temperature (π π ). The yield behaviour of this thermoset (below its π π ) is identical to that of uncross-linked glassy thermoplastics . |
6655a628418a5379b0782c70 | 1 | However, its inability to similarly undergo crazing reduces the amount of deformation post-yield, resulting in brittle behaviour . Fracturing this highly cross-linked network requires the breaking of primary (covalent) bonds, since the molecules are chemically and physically prevented from sliding past each other. As such failure will originate at, and propagate from, defect sites in the polymer network, where local bonds are under increased stress . A secondary 'filler' phase can be introduced into the epoxy to help dissipate energy, delaying crack propagation, through a transfer of stress from the matrix. |
6655a628418a5379b0782c70 | 2 | The use of particulate nanomaterials as fillers has enabled very large interphases to be created at low loadings . Facilitating significant property transfer that can not only improve the mechanical properties of the resultant nanocomposite, but a variety of other functions too, e.g., electrical, thermal and tribological . Nanocarbon based fillers boast strong mechanical properties, as well as high thermal and electrical conductivities. Making them attractive prospects for multifunctional improvements to epoxy resins. They can be synthesised in various dimensions, e.g., 0D fullerenes, 1D carbon nanotubes (CNTs), and 2D graphene nanoplatelets (GNPs), with each generating a unique interphase geometry. However, their graphitic surfaces adhere poorly to the surrounding epoxy matrix resulting in poor interfacial stress transfer and high phonon and electron scatter . Furthermore, these weak interactions mean the fillers preferentially interact with one another, leading to fullerenes aggregating, CNTs bundling, and GNPs restacking. To counter these issues, there has been extensive work on functionalising the nanocarbon surfaces, to promote a better interphase through chemical compatibilities. |
6655a628418a5379b0782c70 | 3 | A variety of functionalisation strategies have been proposed. Oxidation can promote esterification, forming covalent bonds between carboxylic acids on the filler and epoxies on the polymer . While other oxygen groups, such as hydroxyls and carbonyls, better interact with the matrix through non-covalent hydrogen bonding. Fluorination can result in amino functionalisation, through the substitution of reactive fluorine atoms by N-alkylidene amino groups in the hardener . While direct amination enables nucleophilic attack from the nitrogen moieties onto epoxide groups . These schemes were developed on CNTs, and later applied to GNP epoxy nanocomposites as summarised in Table . The larger aspect ratio of a graphene sheet's geometry relative to 0D or 1D nanocarbons, and graphene manufacturer's capability to mass produce high lateral size (tens of microns) multi-layer flakes, has made GNPs attractive particulate fillers for industrial polymer nanocomposite applications . However, there is still scope for structured analysis of the multifunctional effects of different GNP surface chemistries. |
6655a628418a5379b0782c70 | 4 | In this work we carry out a multifunctional property study on functionalised GNP-epoxy nanocomposites. We use commercial grade GNPs, functionalised using scalable and effluent-free plasma treatment, to prove industrial suitability. We plasma treat the GNPs using two different process gases, CF4 and NH3. Then characterise these functionalised GNPs to evaluate their new surface chemistries and any changes in particulate morphology. These are used, alongside the baseline material, to produce GNP epoxy nanocomposites. The mechanical, thermal, and electrical properties of the different nanocomposites are evaluated across a range of filler loadings, and the dispersion of GNPs in the matrix examined using SEM. The effects of the different filler surface chemistries on the multifunctional properties of GNP epoxy nanocomposites are then discussed. |
6655a628418a5379b0782c70 | 5 | The GNP system studied was First Graphene's PureGraph PG-10 Graphene, a graphitic, platelet material with typical quoted lateral size of 10 Β΅m. The Haydale HT60 Plasma Reactor was used to carry out plasma treatment. It enables the functionalisation of lab scale quantities of nanomaterial powder; with larger reactors capable of treating commercial quantities. The treatment was carried out on 50 g batches of GNP, at low pressures (< 10 mbar) and temperatures (< 100 Β°C) using CF4 and NH3 process gases for fluorination and amination respectively. |
6655a628418a5379b0782c70 | 6 | Electrical Conductivity. The electrical conductivity of these pellets was measured using a Jandel Model RM3000+ and four-point probe rig, with 1 mm spacing between probes. A 100 mA current (πΌ) was applied between the outer probes, and the voltage drop (Ξπ) read between the inner probes. The conductivity (π) was found using: |
6655a628418a5379b0782c70 | 7 | Where B is the peak FWHM, ΞΈ is the peak position, π is the x-ray wavelength and K is a shape factor, typically taken to be 0.89 for carbon . Values of πΎ β² = 2 and π = 16 A Μ were used as proposed by Lim et. al . |
6655a628418a5379b0782c70 | 8 | BET. theory was used to find the surface area and estimate layer numbers, it was carried out on a Quadrasorb EVO, the GNPs are outgassed (< 1 mbar, 150 C) for 24 hours prior to beginning the test, to remove residual moisture from the surface. Helium gas was used to calculate the dead volume and nitrogen gas as the adsorbate. The degree of exfoliation can be estimated by finding the 'layer number' (π), from the specific surface area (SSA) using the formula: |
6655a628418a5379b0782c70 | 9 | Raman. Raman spectroscopy was carried out on a Renishaw InVia, using a 532 nm laser, 1200 mm -1 grating, and 50x lens. Samples were prepared by vacuum filtering 6 mL of a 1 mg ml -1 IPA, water GNP dispersion onto a hydrophobic PVDF membrane support. The StreamHR function was used to map 1024 points over a 100 Β΅m 2 area . |
6655a628418a5379b0782c70 | 10 | XPS. The powders were deposited on carbon tape and X-ray photoelectron spectroscopy (XPS) performed using an ESCA2SR spectrometer (ScientaOmicron GmbH) using monochromated Al KΞ± radiation (1486.6 eV, 20 mA emission at 300 W, 1 mm spot size) with a base vacuum pressure of ~1Γ10 -9 mbar. Charge neutralisation was achieved using a low energy electron flood source (FS40A, PreVac). Binding energy scale calibration was performed using C-C in the C1s photoelectron peak at 284.4 eV. Three points were measured for each powder. XPS survey scans were used to establish the relative atomic contribution, the areas under the following peaks: C1s, O1s, F1s and N1s at 285, 530, 690, 400 eV binding energy respectively, were calculated and corrected using relative sensitivity factors from the Scofield library in CasaXPS , and a standard deviation provided from the three repeats. |
6655a628418a5379b0782c70 | 11 | The hot curing Araldite LY 564 (Bisphenol-A resin), Aradur 2954 (cycloaliphatic polyamine hardener) epoxy matrix system was sourced from Huntsman Advanced Materials. The GNP epoxy nanocomposites were manufactured at five weight loading specifications, 1, 2.5, 5, 7.5 and 10 wt.% with respect to the overall mass of the nanocomposite. The GNPs were dispersed in the resin using a Synergy Speedmixer DAC 4000 under 5 mbar vacuum at 800 rpm for 2.5 minutes. The mixture was degassed for 10 minutes at 40 Β°C. The hardener was subsequently added (100:35 resin, hardener mass ratio) and mixed under 5 mbar vacuum at 800 rpm for 3 minutes, then 2350 rpm for 2 minutes. The mixture was degassed for 25 minutes at 40 Β°C and poured into a silicone mould, itself manufactured using CS25 condensation cure silicone rubber from Easy Composites, and 3D printed shape cut outs. The mixture was further degassed at 40 Β°C until bubble free. A manufacturer recommended cycle of 1 hour at 80 Β°C and 4 hours at 160 Β°C was used to cure the nanocomposites. |
6655a628418a5379b0782c70 | 12 | TGA. Thermogravimetric Analysis (TGA) was carried out using a Mettler Toledo TGA/DSC 1, under a N2 atmosphere between 30 -600 Β°C with a 10 K min ramp. Samples were prepared by grinding down to a powder and weighing out 10 mg for analysis. The residual mass of GNP was found by subtracting the mass at 600 Β°C for the neat epoxy sample from the nanocomposite samples. |
6655a628418a5379b0782c70 | 13 | Tensile Testing. Tensile testing was carried out on an Instron 3365 with a 5 kN load cell, using an AVE2 video extensometer to measure strain. Dumbbell samples were prepared to ASTM D638 Type IV specifications, and specimen measurements read off before testing using digital callipers. The procedure for evaluating the tensile modulus is described in the SI. A rule of mixtures (RoM) was fit to the modulus (πΈ), and vol.% (π) data: |
6655a628418a5379b0782c70 | 14 | Where π π was found using the Archimedes principle as described below, and π π was taken from graphite to be 2.267 g cm -3 . For SEM imaging cryo-fracture surfaces were prepared by immersing the specimen in liquid nitrogen for 5 minutes before snapping, then coating with a 5 nm 80:20 Au, Pd layer using a Quorum Q150T ES sputter coater. |
6655a628418a5379b0782c70 | 15 | Electrical Conductivity. To measure the through plane electrical conductivity, wires were attached to each side of the composites (cut to 5 mm x 3.5 mm x 10 mm cuboids), using silver ink and epoxy silver paste to minimise the contact resistance, as shown in Figure (a). These were analysed using an N4L NumetriQ Phase Sensitive Multimeter and Impedance Analysis Interface over a 1 -10 6 Hz frequency range. The measured complex impedance (π * ) values were used to calculate the specific conductivity (π) using: |
6655a628418a5379b0782c70 | 16 | They were sanded down to achieve a smooth surface using grade P320 silicon carbide abrasive paper, with a polishing wheel. Then dried in an oven overnight under vacuum at 50 C. Final dimensions were read off using digital callipers. The in-plane thermal diffusivity measurements were carried out on the above samples using the Angstrom method , and the experimental setup described by Steiner et. al , detailed in the SI. |
6655a628418a5379b0782c70 | 17 | The density (π π ) of the nanocomposites was calculated using the Archimedes principle. The nanocomposites were immersed in distilled water on a Sartorius YDK03 hydrostatic balance, and the buoyancy (πΊ) determined using πΊ = π π -π ππ where π is the weight of the nanocomposite in air (π) and water (ππ) respectively. The temperature of the water was measured using a thermometer and the density calculated using: |
6655a628418a5379b0782c70 | 18 | Heat Capacity. The specific heat capacity was measured using the Modulated DSC (MDSC) mode on the TA Q100 DSC . The correction factor was initially found using a sapphire disc calibrant at the midpoint temperature (50 C). Following that the samples, (weighing approx. 15 mg) underwent a modulated ramp at 3 K min -1 , with an amplitude of 1 K and period of 120s, as recommended for crimped hermetic aluminium pans, between 0 -100 C. Three non-concurrent measurements were carried out for each sample. The use of modulation enables the reversible and non-reversible components of heat capacity to be discerned. A typical heat capacity, temperature plot is shown in Figure . The thermal diffusivity (π), density (π), and reversible heat capacity values (π π ) were used to calculate the in plane thermal conductivity (πΌ) using π = πΌππ π . |
6655a628418a5379b0782c70 | 19 | for PG-CF and PG-NH respectively. As does the Raman analysis: The average Raman spectra are shown in Figure (e). Three characteristic modes, D, G and 2D can be observed at 1347, 1578 and 2700 cm -1 . The defect density, found from disentangling 0D and 1D defects as detailed in Figure (b,c), increases from 2.03 x 10 -3 (Β±1.10 x 10 -3 ) nm 2 |
6655a628418a5379b0782c70 | 20 | The ATR-FTIR spectra are shown in Figure In the latter a hydroxyl O-H stretching peak at 3725 cm -1 [65] is also observed. The elemental composition of the GNPs was found from the XPS survey spectra displayed in Figure (g) and summarised in Table . The PG-10 and PG-CF elemental composition values have been reported previously . The C1s and O1s peaks can be seen in all three GNPs at 284 eV and 532 eV respectively, while a strong F1s peak can be observed in PG-CF at 688 eV, and a weak N1s peak in PG-NH at 399 eV. The oxygen content in PG-10 derives from oxygen passivated defects on the graphitic lattice. Exposure to CF4 plasma, leads to a ~5 at.% increase in fluorine content, at the expense of carbon content, while the oxygen content increases very slightly (+0.3 at.%). On the other hand, exposure to NH3 plasma more than doubles the oxygen content, while showing only a modest increase in nitrogen content. This suggests that the ionized NH3 gas is ineffective at adhering N modalities onto the carbon lattice or displacing existing oxygen groups. |
6655a628418a5379b0782c70 | 21 | The mechanical, thermal, and electrical properties of the composites are summarised in Figure . The loading values were corrected from the target wt.% to the experimentally measured value by TGA, as shown in Figure (d). The tensile stress-strain data analysis has been detailed in Figure , and the resultant tensile modulus values plotted against filler loading (vol.%) in Figure ). |
6655a628418a5379b0782c70 | 22 | The measured πΈ π value of 2.48 GPa sits within the expected range of 2.45 -2.55 GPa from the manufacturer. The highest πΈ π achieved was 5.01 GPa for 6.4 wt.% PG-NH loading, an ~100% increase on πΈ π . This percentage change in modulus outperforms all the papers cited in a recent review on 2D material epoxy nanocomposites [67], except for a Ti3CN epoxy nanocomposite with a significantly higher loading of 90 wt.% . The πΈ π increases with loading for all GNPs, apart from for the highest loading PG-NH specimen. This failure to further stiffen the composite, can be attributed to the agglomeration of PG-NH flakes, reducing their effective surface area in the matrix . This is prone to happen at higher loadings as there is more opportunity for flake-flake van der Waals interactions [70], |
6655a628418a5379b0782c70 | 23 | particularly for the more oxygenated PG-NH surface. Prior to this point, the PG-NH nanocomposites exhibit a ~75% increase in πΈ π over the baseline PG-10, while the PG-CF nanocomposites exhibit a ~50% decrease. A higher πΈ π enables improved mechanical reinforcement at lower loadings, e.g., a 50% increase in πΈ π can be achieved with only 3 wt.% PG-NH loading, whereas this requires loadings of 6 wt.% PG-10 and 12 wt.% PG-CF. |
6655a628418a5379b0782c70 | 24 | The in plane and through plane thermal conductivity values are shown in Figure ), the capacity, density and diffusivity values that make up the former are shown in Figure . There is a close correlation in results in either direction for the nanocomposite specimens, despite the different experimental approaches used for measurement. The discrepancy for the neat epoxy specimens can be accounted for, as the through plane thermal diffusivity measurements are less accurate over a transparent surface. Since more laser light is lost to the sample rather than reflected at the thermal camera. Some anisotropic behaviour is observed for higher loadings of PG-CF (β₯ 5 wt.%), with higher through plane than in plane conductivity. The thermal conductivity of the samples increases with increasing loading, the highest increase achieved (in the through plane direction) was 0.50 W m -1 K -1 for 10 wt.% PG-CF loading, a ~120% increase on neat epoxy. |
6655a628418a5379b0782c70 | 25 | The slight reduction in GNP electrical conductivity post functionalisation (detailed in 0) has no discernible effect on the conductivity of the composite. The neat epoxy electrical conductivity is 6.1 x 10 11 S m -1 and no significant change is observed with the addition of any filler. This indicates that no percolative network is formed, meaning that higher loadings than those tested are required to achieve an electrically conductive composite. Having considered how the different filler surface chemistries lead to different fillermatrix dispersions, the effect of these can be linked to the multifunctional composite properties. The mechanical reinforcement from a nanoplatelet filler to a nanocomposite under tensile stress can be understood using shear-lag theory . Whereby the deformation of the surrounding matrix generates an axial shear stress on the filler particle, that builds from zero at its edges to a maximum at its centre. This stress-transfer requires a large enough platelet over which the filler and matrix axial stresses can equilibrate, and a high matrix shear modulus (and thus πΈ π ) to efficiently transfer the force. The former requires graphene platelets with a lateral size of at least 3 Β΅m , and the latter means that stiffer matrices unlock higher πΈ π values . The overall reinforcing effect is then dependent on the amount of matrix each filler platelet can affect, i.e., the shear-lag unit. Maximising this requires well dispersed and distributed, high aspect ratio flakes . The PG-10 and PG-NH retain their platelet shape in the matrix, and the lateral size of the flake stacks appear to follow the distribution found for the powder. However, the stacks of the former are thicker than the latter, resulting in a lower aspect ratio and thus lower effective surface area and weaker reinforcement. The PG-CF flakes do not preserve their platelet shape, and instead form amorphous aggregates, with a higher density of low lateral size (< 2 Β΅m) particulates. These are not long enough to bear the tensile load from the surrounding matrix, and as such are ineffective at mechanical reinforcement, leading to the lowest πΈ π values. |
6655a628418a5379b0782c70 | 26 | Despite the clear differences in mechanical performance, the thermal and electrical conductivity data appear less susceptible to the different filler-matrix dispersions. This is due to the relatively low loadings used, leading to large interparticle gaps, and thus no percolative networks. The high density distribution of PG-CF particulates should favour the creation of connected filler networks, and this effect can be observed as its thermal conductivity values tend to outperform the other GNPs as the loadings increase. |
6655a628418a5379b0782c70 | 27 | Fully leveraging the outstanding mechanical, thermal, and electrical properties of nanocarbon fillers in polymer nanocomposites remains a formidable challenge for material scientists. Advances in the mass production of GNPs are unlocking larger lateral size and higher aspect ratio flakes. However, without surface modification, their inert carbon surface will continue to be prone to agglomeration, poor dispersion, and a weak interphase. In this study we aim to link changes in GNP surface chemistry, with their dispersion and morphology in an epoxy matrix, and subsequently their multifunctional properties. |
6655a628418a5379b0782c70 | 28 | We use scalable, effluent-free plasma functionalisation techniques to fluorinate and aminate a commercial GNP. We find that plasma fluorination adheres ~5 at.% fluorine, with minimal damage to the graphitic plane. Whereas the plasma amination adds only a modest amount of nitrogen (~ 1 at.%), but more significantly damages the GNP surface and the passivation of these new defects in air leads to a doubling in oxygen content (to ~10 at.%). However, the characterisation of these GNP is only half the story. Rather, it is the interactions between the surface chemistry and the chosen matrix that determine the ultimate morphology. |
6655a628418a5379b0782c70 | 29 | The PG-NH and PG-CF nanocomposites exhibit a ~75% increase and ~50% decrease in filler modulus respectively, compared to PG-10 nanocomposites. The homogeneous dispersion of high aspect ratio PG-NH flakes enables an exceptional ~100% increase in tensile modulus to 5.01 GPa at 6.4 wt.% loading. Whereas the high density of particulates in the PG-CF nanocomposites, which are ineffective for stress transfer, increase the thermal conductivity of the epoxy by 120% to 0.50 W m -1 K -1 at 10 wt.% loading. While the loadings tested (β€ 10 wt.%) did not enable electrical percolative networks to form for any GNP. |
6655a628418a5379b0782c70 | 30 | This study demonstrates how crucial interphase engineering is for maximising the transfer of filler properties in epoxy nanocomposites, and that the chemical interaction between filler and matrix is key in determining dispersed filler distribution and morphology. However, the problem of different material functions requiring different functionalisation strategies remains. Overcoming this to produce polymer nanocomposites with multiple high performance functions, is likely to require hybrid filler combinations, with optimised loading, dimensionality, and surface chemistry. GNP Fluorine functionalised (gas). |
6655a628418a5379b0782c70 | 31 | Multilayer GNPs have a higher graphitic volume to surface area than single and few-layer graphenes, reducing the sensitivity of spectroscopy techniques to changes in the surface chemistry. Furthermore, the heterogeneity of flakes sampled by the Raman laser, leads to significant spectral variance that can be lost when observing an average spectrum. In Box plots for peak fit parameters associated with the D, G and 2D modes, as well as their intensity and amplitude ratios. |
6655a628418a5379b0782c70 | 32 | shows curves typical of a brittle plastic, high modulus, high yield strength, and low elongation. The challenge with this tensile behaviour is correctly identifying the linear region used to fit the modulus. The automatic peak fitting algorithm developed by was used to systematically analyse the curves, a breakdown of which is shown in Figure ). This algorithm works to iteratively minimise the coefficient of variation for a linear regression across a section of the stress-strain curve, as described in ASTM D638 standards. Two material dependent values are used; the initial (π = 1) maximum stress factor π π = 0.9, and the stress ratio tolerance π max = 0.3, found using trialand-error. The yield strength (π YS ) is calculated by constructing a line with the same gradient as the modulus fit, and a 0.002 mm/mm x-intercept offset, then finding the data intercept. The algorithm iterates by decreasing the value of π π until the condition: |
646fb56cbe16ad5c57e86d39 | 0 | Click chemistry is a powerful and efficient method for rapidly connecting chemical fragments, enabling the modification of biologically active molecules . Among the various types of click chemistry, SuFEx (Sulfur(VI) Fluoride Exchange) reactions have emerged as a reliable resource for drug discovery , chemical biology , polymer chemistry , and surface modifications . In particular, the unique properties of the sulfur(VI)-fluorine bond make SuFEx reactions highly versatile, allowing for the formation of covalent bonds under mild conditions . The S(VI)-F bond is highly stable, allowing it to withstand harsh conditions, yet readily cleavable in the presence of a suitable activator or reaction partner . Consequently, the versatility and efficiency of SuFEx reactions make them a valuable tool for researchers in a wide range of fields, including synthetic chemistry, drug discovery, and materials science . |
646fb56cbe16ad5c57e86d39 | 1 | Among the various SuFEx hubs, the -SO2F moiety has received significant attention due to its unique biophysical properties and its potential as a versatile connector between nucleophilic entities . One method for installing this moiety is the use of gaseous sulfuryl fluoride (SO2F2), which has been demonstrated for various organic molecules by the group of Sharpless . However, despite its potential as an economic and traceless compound, its mild toxicity and difficulty to handle have motivated researchers to seek for more practical alternatives. While in-situ generation of SO2F2 from 1,1'-sulfonyldiimidazole (SDI, Figure ) or the use of solid reagents such as phenyl]imidodisulfuryl difluoride (AISF) or 1-(fluorosulfonyl)-2,3-dimethyl-1H-imidazol-3-ium trifluoromethanesulfonate (FDIT) have been explored, these alternatives still require the use of SO2F2 in their preparation and produce unnecessary by-products. Hence, it is clear that these methods go against the principles of click chemistry and represent a challenge to date . Addressing these challenges by developing a new strategy for producing and dosing SO2F2 in a safe and controlled manner from simple chemicals would represent a significant advance in the field of SuFEx chemistry. Furthermore, such a process would significantly reduce the number of synthetic steps needed for SuFEx handle installation and therefore streamline the overall process. |
646fb56cbe16ad5c57e86d39 | 2 | In order to address the challenges associated with using sulfuryl fluoride (SO2F2) as a reagent, we considered using cheap commodity chemicals such as sulfuryl chloride (SO2Cl2) and potassium fluoride (KF) to generate SO2F2 in-situ. This approach was motivated by the greater thermodynamic stability of the S(VI)-F bond (~90 kcal/mol) compared to the S(VI)-Cl bond (~46 kcal/mol), which suggests that this exchange should be achievable . Our initial batch experiments using SO2Cl2 and KF in CH3CN confirmed the feasibility of this approach, demonstrating succesful conversion of first SO2Cl2 to SO2FCl and later to SO2F2 was achieved within two hours (see Supplementary Materials). |
646fb56cbe16ad5c57e86d39 | 3 | Due to the slow kinetics observed in batch reactions and the mixture of SO2FCl and SO2F2 obtained, we turned to flow technology as a tool for more effectively generating and controlling the delivery of this reactive gas . Hereto, we designed a modular system (Figure ) using microfluidic technology to greatly enhance the safety and scalability of the overall process (25). Our modular system consists of two interconnected flow reactors. The first reactor is a packed-bed reactor filled with KF that generates SO2F2 on demand via chlorine-fluorine exchange. The second reactor is where the generated gaseous SO2F2 is mixed with the nucleophilic partner, ultimately yielding the desired SuFEx product. Since the first reactor generates SO2F2 on demand, the reagent remains contained and is subsequently mixed with the nucleophilic partner in the second reactor. By immediately reacting away the toxic SO2F2 in the SuFEx module, our modular system effectively eliminates the safety and practical concerns associated with the handling of this reagent, while generating only the required quantities. In addition, we anticipated that our use of flow technology would also reduce the time required for halogen substitution reactions, thanks to enhanced liquid-to-solid contact in the first Cl-F exchange module and excellent gas-to-liquid mass transfer in the SuFEx module . Our experiments confirmed the effectiveness of our modular system: when we directed a solution of sulfuryl chloride over the packed-bed reactor filled with a mixture of KF and glass beads (Figure , see Supplementary Materials for further details), we observed a rapid and selective formation of SO2F2. We found that varying the flow rate was crucial for the selectivity of the transformation, and that 0.5 mL/min was the optimal flow rate in terms of conversion, selectivity, and time (7 min reaction time), effectively avoiding the presence of undesired SO2FCl. Under optimized conditions, the packed-bed reactor was able to produce ~18 mmol of SO2F2 starting from a ~80 mmol KF bed (see Supplementary Materials for further details). Having obtained promising results from our SO2F2 generator, we proceeded to integrate it with the SuFEx module to enable the reaction with nucleophilic partners. By introducing the appropriate nucleophiles with an excess of a base, we were able to obtain a diverse range of SuFExed products with excellent isolated yields in just two minutes of residence time (Figure ). This short residence time can be attributed to the intimate contact between gas and liquid phase in flow (i.e., enhanced gas-liquid mass transfer) , which should allow for the generation of large libraries of SuFExed compounds with minimal effort and time. Notably, a variety of phenols could be cleanly converted to their corresponding fluorosulfates regardless of the position or electronic nature of the substituents (Figure , compounds 1-6). While using large quantities of gaseous SO2F2 in batch reactions can be challenging, our flow protocol overcomes this issue, allowing for a gram-scale synthesis of fluorosulfate 1 by simply increasing the amount of starting materials pumped through the reactor assembly . |
646fb56cbe16ad5c57e86d39 | 4 | We further found that our flow protocol was not limited to the synthesis of simple phenol-based fluorosulfates. In fact, we observed that a wide variety of natural products, drugs, and fluorescent tracers could be successfully reacted using this approach. For instance, N-Bocprotected tyrosine, broxyquinoline, fluorescein, Ξ±-tocopherol, estrone, and unprotected amoxicillin were all cleanly converted in just 2 minutes of reaction time (Figure , compounds 7-12), demonstrating the excellent functional group tolerance of our method. Similarly, the challenging class of nitrogen-based nucleophiles could also be subjected to our protocol delivering the corresponding sulfamoyl fluorides in excellent isolated yields. For example, 4to 7-membered ring secondary amines (Figure , compounds 13-17), heterocyclic derivatives (Figure , compounds 18-20), and an aniline compound (Figure , compound 21) were effectively reacted with SO2F2. Our flow protocol was also effective in synthesizing analogs of several pharmaceutically relevant molecules, such as stanozolol, paroxetine, desloratadine, amoxapine, and olanzapine, which rapidly yielded the desired SuFExed products (Figure , compounds 22-26). Additionally, nucleosides such as adenosine and cytidine derivatives (Figure , compounds 27-29) could be used as competent reaction partners. Finally, we were able to use bidentate nucleophiles, such as BINOL, catechol, and salicylamide (Figure , compounds 30-32), to produce the corresponding sulfates and sulfamates. generation: SO2Cl2 (2 equiv, 0.2 M in CH3CN) passed through a 3.8 mL cartridge filled with a 1:1 mixture of KF and glass beads. Standard conditions for the second step: [a] Nucleophile (1 equiv, 0.2 M in CH3CN), Et3N (2.5 equiv). [b] Nucleophile (1 equiv, 0.2 M in DMF), Et3N (2.5 equiv). [c] The compound has been isolated after an acetylation step. [d] Nucleophile (1 equiv, 0.2 M in CH3CN), DBU (4.0 equiv). [e] Nucleophile (1 equiv, 0.2 M in DMF), DBU (4.0 equiv). Next, we capitalized on the modular nature of our setup and developed a multistep process that orchestrates several reactions in a sequential fashion (Figure ). For instance, we streamlined a three-module flow setup to synthesize sulfate derivative 33 (67%) uninterrupted. After generating sulfuryl fluoride in the packed-bed reactor and trapping it with estrone in the SuFEx capillary reactor, the resulting fluorosulfate was reacted with (4-iodophenoxy)trimethylsilane in a second SuFEx capillary. Our flow approach carefully balances the stoichiometry of the gaseous reagent, preventing any remaining SO2F2 from unproductively consuming the silyl ether in the second SuFEx step . Furthermore, after a solvent switch, the reaction crude can also immediately be used to carry out a base-promoted hydrolysis of ethylenacetal-protected estrone to obtain bisulfate derivative 34 (80% yield) or a palladium-catalyzed Suzuki-Miyauratype cross coupling to yield product 35 in 55% yield. Our flow approach has demonstrated excellent functional group tolerance and selectivity in producing SuFExed products from various small molecules. This is due to the favorable reaction conditions, including high mass transfer and excellent dosing of gaseous reagents, provided by the flow protocol that allows for reduced reaction times of just two minutes, thereby avoiding the formation of deleterious byproducts. We next sought to explore the potential of our microfluidic strategy for the late-stage modification of unprotected peptides . Our aim was to directly install the sulfur-centered electrophilic handle within the peptide core, allowing for later SuFEx-enabled derivatization opportunities . Based on our previous experience with small molecules, we anticipated that the reaction would be kinetically favored for tyrosine residues, which should enable site-selective modification within the complex peptidic framework. After a minimal re-optimization of the reaction conditions (see Supplementary Materials), we investigated the reactivity of different nucleophilic amino acid residues within different pentapeptides. We found that tyrosine-containing peptides were successfully converted in a site-selective fashion in just two minutes (Figure , compounds 36 and 37). Our evaluation also showed limited modification at lysine and histidine (Figure , compounds 38 and 39), while other nucleophilic residues, such as tryptophan and cysteine, were not reactive under our reaction conditions (see Supplementary Materials). Encouraged by these results, we subjected various complex and therapeutically valuable peptides to our flow protocol. Cyclic peptides (Figure , compounds 40 and 41) and therapeutic drugs such as Angiotensin II and Bivalirudin (Figure , compounds 42 and 43) were selectively modified at the tyrosine residue in good to excellent results. Even natural peptides, such as Ξ±-Endorphin and Ξ²-Amyloid (1-28), were effectively converted into the corresponding fluorosulfates with good to excellent conversions (Figure , compounds 44 and 45). |
646fb56cbe16ad5c57e86d39 | 5 | As the ultimate test for our SuFEx ligation protocol, we focused on the direct modification of proteins in flow. By minimizing lysine competition (see Supplementary Information), we managed to exclusively install the electrophilic SO2F handle on tyrosine residues in just 1.5 minutes. To the best of our knowledge, this is one of the fastest methods for direct protein modifications reported to date . For instance, we merged a solution of Ξ²-Casein with the SO2F2-containing stream and observed predominantly single, chemoselective functionalization at different tyrosine residues with a Y180/Y193/Y114 = 10.7/2.4/1 regioselectivity ratio (Figure , compound 46). Notably, Myoglobin was obtained as a single Y103-modified adduct (Figure , compound 47), without observing denaturation or loss of the heme group, demonstrating the mild nature of the protocol. Remarkably, when we attempted to perform the same experiments in fed-batch mode, only a 16% conversion was obtained, while using a batch H-type reactor only yielded a complex mixture of products (Figure , bottom right). These results demonstrate that the enhanced mass transfer and confined access to the gaseous and hydrophobic SO2F2 observed in capillary flow reactors are critical to enable efficient SuFEx hub installation into complex macromolecular systems. Conversions reported as ratios of areas under the peak of product and starting compound obtained by LC/MS analysis. Peptides: Standard conditions for the SO2F2 generation: SO2Cl2 (40 equiv, 0.2 M in CH3CN) passed through a 3.8 mL cartridge filled with a 1:1 mixture of KF and glass beads at f.r. 0.5 mL/min. Peptide (1 equiv, 10 mM in CH3CN:H2O 1:1), Et3N (6 equiv), f.r. 0.25 mL/min, res. time 2 min at room temperature. [a] Ξ±-Endorphin (6 mM), SO2F2 (67 equiv). [b] Ξ²-Amyloid (3 mM), SO2F2 (133 equiv). [c] A picture of a general protein was chosen to represent Ξ²-Casein as no crystal structure is reported. Proteins: Standard conditions for the SO2F2 generation: SO2Cl2 (0.1 M in CH3CN) passed through a 3.8 mL cartridge filled with a 1:1 mixture of KF and glass beads at f.r. 0.1 mL/min. Ξ²-Casein (1 equiv, 5 mM in Tris buffer pH = 7.7), TMG (10 equiv), f.r. 0.9 mL/min, res. time 1.5 min at room temperature, SO2F2 (2.2 equiv). Myoglobin (1 equiv, 1 mM in 10 mM acetate buffer pH = 5), TMG (1 equiv), f.r. 0.9 mL/min, res. time 1.5 min at room temperature SO2F2 (11 equiv). TMG: 1,1,3,3-Tetramethylguanidine. |
646fb56cbe16ad5c57e86d39 | 6 | The practical flow protocol presented in this study enables the safe and efficient generation of the coveted gaseous SO2F2 reagent, as well as high reaction rates of the subsequent SuFEx ligation, with wide applicability to various substrates, including therapeutically relevant small molecules, peptides, and proteins. Based on these findings, we believe that this protocol opens up new opportunities in the field of SuFEx click chemistry. In particular, the use of this flow process makes sulfuryl fluoride a viable reagent for installing the -SO2F handle on a variety of phenol and amino functionalities. |
62da45fb27b1e4961e4722da | 0 | Modeling the interaction of proteins with small ligands, in particularly, for the goals of drug design is traditionally performed using molecular mechanics or classical molecular dynamics methods with force field parameters. Application of quantum-based simulation tools is also growing, but is still far from the routine use. In this work, we describe quantumbased simulations for a molecular system, which can hardly be handled at the classical level. Specifically, we characterize a reactant complex of a flavin-dependent enzyme RutA from Escherichia coli with the molecular oxygen and uracil. |
62da45fb27b1e4961e4722da | 1 | The RutA enzyme catalyzes a first step of uracil degradation in hydroxypropionate, ammonia, and carbon dioxide, allowing bacteria to use pyrimidine rings as nitrogen source and belongs to a large family of flavoprotein monooxygenases (FMOs) widely spread across all kingdoms of life. Mammalian FMOs are indispensable as xenobiotic metabolizing enzymes with a prominent role in drug metabolism. Growing evidence points to implication of FMOs in aging, several deceases and metabolic pathways. FMOs activate molecular oxygen and insert a single oxygen atom in small organic substrates containing heteroatoms and convert the second oxygen atom in a water molecule. The chemical formulae of the reactants in the monooxygenation reaction of uracil by RutA are shown in Fig. . According to the current knowledge, the reduced form of flavin mononucleotide (FMN), which is non-covalently bound to the RutA macromolecule, activates the dioxygen molecule by transferring an electron from FMN. The reactive oxygen is capable to bind chemically to one of the atoms of the isoalloxazine ring of FMN, and this intermediate flavin-based species is involved in the chain of transformations of the uracil molecule. There are unresolved puzzles in this plausible mechanism; one of them is specification of a target atom in the isoalloxazine ring (C4a, N5, C6) for the attack of the reactive oxygen. Detailed consideration of the reaction mechanism is far beyond the scope of this work; here, we focus on computational characterization of reactants, namely, on how the ligands, FMN, uracil and O2, are accommodated inside the protein favoring possible reaction pathways. The work of Matthews et al. reported atomic coordinates of the RutA-FMN-O2uracil complex deposited to the Protein Data Bank (the structure PDB ID 6TEG). This complex was obtained as the O2-pressurized crystal at 15 bars O2; however, low occupancies of the ligands were observed. From the theoretical side, it should be noted that methods of describing protein-ligand interaction using customary force field parameters may fail in such project, even if some hints are provided by results of crystallography studies. It is not an essential problem to accommodate FMN in the protein cavity, despite location of this ligand close to the solvent accessible area. A small organic molecule, uracil, can also be localized in the active site; however, with more uncertainties, given the disparity in the FMN and uracil sizes. The greatest problem presents the dioxygen molecule because no reliable classical force field parameters are available to describe polarization and a partial charge transfer to this species when it approaches the negative charged FMN and the protein groups. Needless to remind that the localization of dioxygen in protein cavities is a nontrivial experimental task. Thus, quantum-based computer simulations are the indispensable tools to characterize complexes of proteins with small organic molecules and dioxygen. |
62da45fb27b1e4961e4722da | 2 | To prepare molecular model systems we utilized coordinates of heavy atoms from the structures PDB ID 6TEG and PDB ID 6SGG . The first structure refers to the RutA-FMN-O2-uracil complex, which we sometimes call below for brevity as the ROU complex. The second structure corresponds to the complex without uracil, RutA-FMN-O2 (the RO complex). Hydrogen atoms were added assuming the conventional protonation states of the polar residues at neutral pH -Arg, Lys (positively charged); Glu, Asp (negatively charged); the histidine residues were assumed in the neutral state. Water molecules resolved in the crystal structure were kept in the model system, and solvation water box was built using the visual molecular dynamics (VMD) program. The CHARMM36 force field topology and parameters were employed along with the FMN parameters in the reduced form (total charge -3) taken from Ref. and those for the molecular oxygen taken from Ref. , while water molecules were treated as TIP3P. Classical MD trajectories were simulated with the NAMD 3.0 software package. The isothermal-isobaric (NPT) ensemble at P = 1 atm and T = 300 K using the NosΓ©-Hoover Langevin piston pressure control and the Langevin dynamics were employed. Periodic boundary conditions and the particle mesh Ewald algorithm to account for the longrange electrostatic interactions were applied, whereas the non-bonded interaction cut-off parameter was set to 12 Γ
; integration step was set to 1 fs. A harmonic constraint potential of 1 kcal/Γ
2 was applied to the CA atoms of the protein, except for the residues #290-310 of the region that poorly resolved in the crystal structures. In sum, over 500 ns of combined trajectories were produced. The QM/MM MD trajectories were computed with the TeraChem/NAMD via an appropriate interface. Calculations were initiated from selected frames of classical MD. The energies and forces in QM parts were computed with the density functional theory (DFT) with the range-separated Οb97X functional and the D3 dispersion correction. The 6-31G** basis set was employed for carbon and hydrogen atoms and 6-31+G** for nitrogen and oxygen atoms. The MM part was treated by the CHARMM36 force field, water molecules were described by the TIP3P model. |
62da45fb27b1e4961e4722da | 3 | Figure illustrates an important result of this work revealed in classical molecular dynamics simulations. According to these simulations either without or with the uracil molecule in the active site, the oxygen molecule can be trapped in several binding pockets in the protein; five of them are schematically shown in Fig. . The retention time of oxygen in these pockets varies from hundreds of picoseconds to tens of nanoseconds. It is important to note that the overall protein structure is stable along trajectories, as shown in Fig. for the protein backbone. The RMSD fluctuations through the graph are mainly due to the region of amino acids #290-310, which is not properly resolved in the crystal structures. The FMN coordination and position is very stable for more than 400 ns trajectory (Fig. ), while uracil position fluctuates during the trajectory (Fig. ). The hump in the uracil RMSD chart (Fig. ) during the ~200-250 ns window is associated with oxygen entering the pocket-5 (Fig. ) and consequently pushing off the uracil molecule. We remind that the starting coordinates for trajectory runs were selected from the crystal structures PDB ID 6SGG and PDB ID 6TEG. The identified pocket-1 roughly corresponds to the structural motifs in the crystal structures. In this site, the oxygen molecule fluctuates in the region between the flavin moiety near atom N5 of flavin (see Fig. ), polar residues Asn134, Thr105 and a hydrophobic cluster of Leu65, Met67 and Phe25 sidechains (see Fig. ). Here, the oxygen and uracil molecule are located at different sides of the plane of the isoalloxazine ring (see Fig. ). In pocket-5, the O2 molecule is surrounded by water molecules and resides closer to the C4a atom of the flavin molecule. Here, the oxygen and uracil molecule are located at the same side of the plane of the isoalloxazine ring. In other pockets, the oxygen molecule stays farther from the FMN ligand. Pocket-2 is mostly hydrophobic; it is formed by the sidechains of Asn134, Phe224, Ala206, Leu65 and flavin. Pocket-3 consists of the sidechains Phe222, Phe224, Tyr257, Ile204, Phe63; it is close to the protein surface, showing one of the obvious ways connecting the pocket-1 to the bulk. Pocket-4 is formed by the FMN phosphate group, the Tyr160, Cys205 sidechains and nearby main chain of beta-strands. Pockets 1 and 5 are the most perspective sites for the uracil oxygenation reaction in RutA. Therefore, selected frames along classical MD trajectories in the areas corresponding to pocket-1 and pocket-5 were taken to initiate trajectories with the QM/MM potentials. |
62da45fb27b1e4961e4722da | 4 | In QM/MM MD calculations for the RutA-FMN-O2-uracil system, with the oxygen molecule in pocket-1, the QM part included the FMN, O2 and uracil moieties and the Asn134, Thr105 sidechains. Simulations were carried out for triplet spin state of the system due to the ground electronic state of O2. Correspondingly, the unrestricted DFT approach was used in QM/MM calculations of energies and forces. |
62da45fb27b1e4961e4722da | 5 | Figure shows a general view of the RutA-FMN-O2-uracil model system. Models of the reactants are presented in right side of Fig. . Fig. reproduces the features in the crystal structure PDB ID 6TEG. According to the reported results, the oxygen molecule stands closely to the N5 atom of flavin, the distance between N5 and the nearest oxygen atom is 2.09 Γ
. The authors of Ref. advocate the reaction mechanism of uracil oxygenation by RutA via formation of a temporary adduct with the O(oxygen)-N5(flavin) chemical bond. Correspondingly, a short distance of 2.09 Γ
in the reactants resolved in the crystal structure benefits this hypothesis. However, comparison of Fig. and Fig. shows a strong discrepancy between computational (3.71 Β± 0.51 Γ
) and crystallography (2.09 Γ
) data for the distance between the oxygen molecule and the N5 atom of flavin. We comment that the reported value 2.09 Γ
for the crystallography structure PDB ID 6TEG seems to be too short for the hydrogen bond, if N5 is protonated in the reduced from of FMN, and too long for a covalent bond if one assumes a covalent binding of oxygen to the deprotonated nitrogen in FMN. We remind that location of a small molecule O2 in protein cavities is a difficult experimental task; therefore, the results of computational approaches should be taken seriously. |
62da45fb27b1e4961e4722da | 6 | As shown in Fig. , the oxygen molecule is localized almost equally close to the N5 and C6 atoms of flavin. Fig. illustrates that the O2 molecule is stably located in this pocket; this conclusion holds for both complexes, ROU and RO, that is with or without uracil molecule present in the system. At this point, we speculate about subsequent chemical transformations in the system. As mentioned above, when the oxygen molecule comes close enough to the isoalloxazine ring, the electron transfer (coupled with the proton transfer) might occur, resulting in creation of the reactive oxygen species capable to interact with flavin. We model this step by switching the trajectory at some QM/MM MD frame to the singlet spin state of the system. In the case of the unrestricted singlet state DFT approximation, the resulting electronic structure of the system can describe the radical pair FMN and OOH, as shown in Fig. . This structure remains stable along QM/MM MD trajectories, and the Asn134 amino acid residue is very important to stabilize the HOO radical. Interestingly, the average N5(flavin)-O(O2) distance is very short (2.64 Γ
) and exhibits low variation (Β±0.09 Γ
), We note that in this singlet radical pair configuration, the N5-O distance is somewhat close (2.64 Γ
) to the value 2.09 Γ
, suggested upon the interpretation of experimental density in the PDB ID 6TEG and 6SGG structures. |
62da45fb27b1e4961e4722da | 7 | If the restricted singlet state DFT approximation is applied instead of the unrestricted version, the electronic structure cannot describe radical pairs. Figure illustrates the consequences of such modeling showing that the covalently bound adduct is formed after spontaneous binding oxygen to the C6 atom of flavin. It is not evident how the reactive species shown in Fig. and Fig. react with the uracil molecule given its position on the other side of the isoalloxazine ring; further cumbersome calculations are required to clarify the full reaction pathway. We only note that the main difference between the ROU and RO systems are due to uracil blocking water access to the C6 and N5 atoms of the flavin and to the oxygen atoms of O2 or subsequent HOO radical/group. Now we turn to the results of simulations in the area corresponding to the pocket-5 oxygen binding site. In this case, the QM subsystem included FMN, O2, uracil, the side chains of Trp139, Lys39, Glu143, Asn134 and several water molecules. The results for the QM/MM MD trajectory (Fig. Analyzing these results, we observe significant charge transfer to the oxygen molecule at many steps of the simulation. Such structures are characterized by: (i) close distances of oxygen to the flavin, (ii) close position of the Lys69 sidechain, (iii) proper position of the uracil that allows for hydrogen bonding network to stabilize the charge transfer. All these issues are observed in the 60-80 ps window (see Fig. , Fig. ). Here, we also model a possible reaction step toward the intermediate formed by flavin and oxygen. We again switched the trajectory at some frame to the restricted singlet spin state of the system. Figure shows that the covalently bound adduct is formed after binding oxygen to the C4a atom of flavin. Subsequent reactions may lead to the oxidation of uracil. |
62da45fb27b1e4961e4722da | 8 | The results of the simulations described in this work allow us to conclude that the use of the QM/MM MD approaches considerably expands the repertoire of theoretical tools to model complex protein-ligand interaction. The treatment of protein-dioxygen complexes can hardly be handled at the level of molecular mechanics with conventional force fields. As shown in this work, quantum-based approaches may greatly assist in interpretation of crystallography data; in this case, accommodation of the oxygen molecule near the flavin moiety is improved. The QM/MM MD calculations enable us to speculate on the reaction mechanisms when analyzing trajectories near the reactant configurations. |
64e6056a79853bbd78421e53 | 0 | Historically, natural products (NPs) have been the biggest source of bioactive compounds for medicinal chemistry. For instance, in cancer research, in the lapse of time 1946 to 1980, seventy-five small molecules were approved worldwide, of which 53% were unaltered NPs or natural product (NP) derivatives. Moreover, from 1981 to 2019, of the 185 small molecules approved to treat cancer, 64.9% were unaltered NPs and synthetic drugs with a NP pharmacophore . Another example is the actual development of new promising antibiotics against drug-resistant bacteria from NPs . Furthermore, in a recent review it was shown that 697 natural steroidal alkaloids were isolated and characterized with various biological activities, from 1926 to 2021 . The bioactive compounds encompass marine , fungal , bacteria , plants and endogenous substances produced by humans and animals, sources , including venoms and poisons produced by different animals . Even, as recently reviewed, the fruit peels are a source of bioactive compounds that in many instances display better biological and pharmacological applications than the compounds of other sections of the fruit . |
64e6056a79853bbd78421e53 | 1 | To date, the discovery process of more than seventy commercialized drugs has included the rational use of at least a computational method . Computer-aided drug design (CADD) has the potential to reduce the billionaire cost and decrease the time through the drug design process, e.g., the hit identification rate for highthroughput screening (HTS) to discover novel inhibitors for the enzyme protein tyrosine phosphatase-1B is only 0.021% and the one for molecular docking is 34.8% . A crucial resource in CADD are the databases of chemical compounds including NP databases. From the compound databases, it is possible to identify potential hit molecules through several virtual screening (VS) techniques , including the training of artificial intelligence (AI) algorithms . When the compound databases are annotated with biological activity (or other property of relevance), it is possible to use the data to perform structure-activity (property) relationships and develop predictive models. From 2003 to 2018, 104 research articles reported the identification of potential drug candidates from NP databases by using computational tools . |
64e6056a79853bbd78421e53 | 2 | Between 2000 and 2019, one-hundred twenty-three commercial and public NP databases have been published. Among them, ninety-eight are still somehow accessible (online or under request access), ninetytwo are free access, and only fifty contain molecular structures that can be retrieved for a chemoinformatic analysis . Examples of the most representative open-access NP databases include: The Collection of Open Natural Products (COCONUT) which is a major repository containing more than 411,000 NPs collected from 50 open access NP databases. The Universal Natural Product Database is a compilation that tries to gather all the known NPs; it has more than 229,000 NPs. It is not yet accessible through the link in the original publication, nevertheless, it is contained and maintained on the ISDB website . SuperNatural β
‘ database contains over 325,000 NPs, nonetheless, it does not provide a bulk download. ZINC database has over 80,000 NPs, approximately 48,000 purchasable. Moreover, it contains some NP databases that are no longer accessible through the link provided in the original publication, e.g., Herbal |
64e6056a79853bbd78421e53 | 3 | Ingredient Targets and Herbal Ingredients in vivo Metabolism database , which contain NPs mostly from Chinese plants. Moreover, there are NP databases which contain compounds isolated and characterized in certain geographical areas. That is the case of China, where have been published multiple compound databases containing only NP of this country , nevertheless, TCM@Taiwan is the largest, which contains 58,000 compounds. There are two databases of NPs from India, IMPPAT composed of approximately 10,000 phytochemicals extracted from 1,700 medicinal plants, and MedPServer , containing 1,124 NPs. Regarding NPs from Africa, there are several NP databases , nonetheless, AfroDB is the most extensive, containing over one thousand NPs. Recently, was published Phyto4Health , a NP database with 3,128 NPs isolated from medicinal plants of Russia. |
64e6056a79853bbd78421e53 | 4 | Latin America contains at least a third of the global biodiversity , in fact, half of the countries have been classified as megadiverse: Bolivia, Brazil, Colombia, Costa Rica, Ecuador, Mexico, Peru and Venezuela) . Therefore, Latin America represents a large source of bioactive molecules and potential drug candidates (Figure ). There have been published databases containing NPs from some Latin American countries such as NaturAr (Argentina), NuBBEDB , SistematX , UEFS (Brasil), CIFPMA (Panama), PeruNPDB , (Peru), UNIIQUIM and BIOFACQUIM (Mexico). Recently, it was reviewed the present state of the art in developing Latin American NP databases and their practical applications to the drug discovery area . Multiple drug candidates have been identified from the Latin American NP databases as therapeutic agents for diseases caused by infectious agents (Chagas disease , tuberculosis , Leishmaniasis , schistosomiasis , coronavirus disease , human immunodeficiency virus infection and acquired immunodeficiency syndrome, hepatitis B and C) , pain , obesity, diabetes, hyperlipoproteinemia, cancer, and age-related diseases . Active compounds of representative medicinal plants from Latin American countries and some of its therapeutic effects described in the literature. Brazil (gallic acid and casearin x ), Costa Rica (acetogenin squamocin ) El Salvador (Genkwanin ), Mexico (3Ξ±-hydroxy masticadienoic acid and mescaline ), Panama (furoquinoline alkaloid ) and Peru (mitraphylline and benzylglucosinolate ). |
64e6056a79853bbd78421e53 | 5 | The long-term goal of the project is to collect, unify, and standardize the Latin American NP collections available in the public domain into one public database. In this study, we report significant advances towards this goal by the assembly of the first version of the unified database herein called Latin American Natural Products Database (LANaPD). We report its curation, standardization, and a comprehensive analysis of nine compound databases, totaling 12,959 unique molecules. As part of this study, analyzed the structural content (scaffolds and ring systems), structural diversity, and complexity of the compounds in LANaPD. We also represent coverage in the chemical space of compounds in LANaPD using the concept of chemical multiverse . |
64e6056a79853bbd78421e53 | 6 | The compounds were classified in a total of seven different pathways, fifty-three superclasses, and 336 classes (Figure ). The three predominant pathways are terpenoids 63.2%, shikimates and phenylpropanoids 18% and alkaloids 11.8%. The main superclasses are diterpenoids 34.3%, sesquiterpenoids 17.6% and flavonoids 10.3%. The prevalent classes are kaurane and phyllocladane diterpenoids 6.99%, colensane and clerodane diterpenoids 5.91% and germacrane sesquiterpenoids 5.36%. The results are in accordance with expectations because the terpenoids are the most diverse group of secondary metabolites derived from natural sources . |
64e6056a79853bbd78421e53 | 7 | Veber's rules , GlaxoSmithKline's (GSK) 4/400 rule and Pfizer 3/75 rule (Table ). Having physicochemical properties in the limits of either Lipinski's, Veber's or GSK rules is usually related with a good NPs contain complex structures and are large and diverse, therefore, compared with synthetic drugs, it is not easy that they satisfy most of the criteria of Lipinski's Ro5 or the other drug-likeness parameters mentioned above. Nevertheless, it is shown in the violin plots that a broad range of the LaNaPDB compounds satisfy most of the rules of thumb of Table for the physicochemical properties of pharmaceutical interest. |
64e6056a79853bbd78421e53 | 8 | The LaNaPDB and COCONUT compound distribution of the physicochemical properties is in general similar (Figure ). Also as expected, COCONUT covers the broadest area of the chemical space, because it is the largest database (411,000 compounds) (Figure ). Many compounds of LaNaPDB fulfill the rules of thumb associated with drug-likeness (Figure ) and part of the LaNaPDB chemical space overlaps with the chemical space comprised by the approved drugs (Figure ). The distribution of the physicochemical properties of the NPs in the countries with more compounds (Brazil and Mexico) is in general, more focused in certain regions, compared with the NPs from countries with less compounds (Costa Rica, El Salvador, Panama, and Peru) which is broader (e.g. SlogP, Brazil vs Peru from Figure ). The chemical space represented by the six physicochemical properties is overlapped among the NPs from the six Latin American countries (Figure ). In the principal component analysis (PCA), the first two principal components are enough to represent most of the explained variance percentage: 89.3% in the LaNaPDB, COCONUT and approved drugs comparison and 84.6% in the Latin American countries' comparison (Table ). Moreover, TPSA, MW, HBD and HBA are the descriptors with more contribution to the principal component 1. The descriptors with more contribution to the principal component 2 are SlogP and Rb. |
64e6056a79853bbd78421e53 | 9 | Figures and show the visual representation of the chemical multiverse of LANaPDB generated with tdistributed stochastic neighbor embedding (t-SNE) and tree MAP (TMAP) and two fingerprints of different design: MACCS keys (166-bits) (Figure and Figure ) and MAP4 (Figure and Figure ). As discussed recently, the chemical multiverse can be defined as a group of chemical spaces, each generated with a diverse set of descriptors . A chemical multiverse is a natural extension of the concept of chemical space and its advantage is that it provides a more complete description of the chemical space of a set of compounds as opposed to using only one representation. Based on the visual representation of the chemical multiverse it is concluded that t-SNE has a better performance with MACCS keys (166-bits) fingerprint over MAP4 fingerprint, separating the NPs on clusters according to the structural features (Figure ). The efficacy of TMAP to separate compounds in clusters from MACCS keys (166-bits) and MAP4 fingerprints is similar |
64e6056a79853bbd78421e53 | 10 | with both fingerprints (Figure ). Moreover, TMAP performed better than t-SNE in the NPs cluster creation with both fingerprints. A interactive version of the scatter plot created with TMAP from MAP4 fingerprints (Figure ) is freely available at . To open the interactive map, download the file and open it in web explorer. Since TMAP performed better than t-SNE, and MACCS keys (166-bits) and MAP4 fingerprints showed a similar efficacy in the TMAP, the comparison of LaNaPDB with the reference databases was made with TMAP and MACCS keys (166-bits) fingerprint (Figure ). It can be observed that LANaPDB overlaps with COCONUT in well-defined areas, nevertheless, the approved drugs are more dispersed and some of them overlap with compound in LANaPDB (Figure ). |
64e6056a79853bbd78421e53 | 11 | The visual representations of the different chemical spaces that consider, either physicochemical properties or molecular fingerprints were illustrated with scatter plots (Figures ). Every point in the scatter plots represents a unique compound. The scatter plots were created in the python programming language (version 3.10.7), employing the seaborn module (0.12.2) . |
64e6056a79853bbd78421e53 | 12 | The Latin American NP databases of Table were used to construct the unified NP database LANaPDB. The process was carried out in the python programming language (version 3.10.7), employing the RDKit (version 2022.03.5) and MolVS (version 0.1.1) modules. The standardization process of MolVS was applied, which consist in the remotion of explicit hydrogens, disconnection of covalent bonds between metals and organic atoms, application of normalization rules (transformations to correct common drawing errors and standardization of functional groups), reionization (ensure the strongest acid groups protonate first in partially ionized molecules) and recalculation of the stereochemistry. The salts were removed, keeping the largest fragment, which was neutralized, and the remaining partially ionized fragments were reionized. The canonical tautomer was determined, and, from the InChIKey strings of the canonical tautomer, the duplicate compounds were removed. |
64e6056a79853bbd78421e53 | 13 | Compounds in LANaPD were classified with NPClassifier which is a freely available deep neural networkbased structural classification tool for NPs. NPClassifier establishes a classification system based on the literature from the specialized metabolism of plants, marine organisms, fungi, and microorganisms. The categories used in NPClassifier are defined at three hierarchical levels: Pathway (nature of the biosynthetic pathway), Superclass (chemical properties or chemotaxonomic information), and Class (structural details). |
64e6056a79853bbd78421e53 | 14 | Employing the software KNIME version 4.7.1, with the RDKit nodes, six physicochemical properties of pharmaceutical interest were calculated: SlogP , MW, TPSA , Rb, HBA, HBD. Violin plots were constructed to summarize the distribution of each property individually. In each violin plot we highlighted the limit of drug-like rules of thumb (Table ). To generate a visual representation of the chemical space of the compound libraries based on the six properties, we reduced the data dimensionality to two dimensions employing PCA and t-SNE with the python module Scikit-learn version 1.2.2 . PCA: principal component |
64e6056a79853bbd78421e53 | 15 | Here LANaPDB is part of one of the strategic actions to contribute to the further development of chemoinformatics and related disciplines in Latin America and strengthen the interactions between Latin America and other geographical regions . We encourage the community to visit the websites where the individual NP databases of the different Latin American countries are reported (Table ). |
64e6056a79853bbd78421e53 | 16 | We anticipate that LANaPDB will continue growing and evolving with the update of more compounds from each existing database plus the addition of databases from other Latin American countries. One of the first steps in this direction is the integration of a larger set of NAPRORE-CR and the incorporation of natural products database NPDB-EjeCol from Colombia. Another perspective is the implementation of the database in a free-web server. Likewise, LANaPD could be integrated with other large public databases of natural products such as COCONUT or LOTUS. |
619f74ac568d334004462b5c | 0 | Hydrogen is a vital ingredient in chemical (majorly in synthesis of ammonia), metallurgical, glass, semiconductor, and pharmaceuticals industries . The calculated demand for hydrogen in fossil fuels free future is expected to increase to more than two gigaton per year . The majority of all hydrogen produced globally comes from natural gas (50%), primarily through steam methane reforming (reaction of steam and natural gas) and the rest is generated from oil (30%), coal (16%), and electrolysis of water (4%). The by-products of the major synthesis processes are complex mixtures of condensable hydrocarbons and CO2. This causes emission of 830 Mt of CO2 per year, which is detrimental to the environment. Therefore, the instantaneous hydrogen production from water splitting will unleash the full environmental benefits. |
619f74ac568d334004462b5c | 1 | Alkali metals like sodium, aluminum, magnesium, and potassium react with water to produce H2 gas , However, these metals cannot be recycled, therefore they have limited scope in hydrogen production. Moreover, these reactions are highly exothermic and extremely difficult to control . Recently, several high-entropy alloy and metal-hydrides have been explored to generate H2 by controlled water splitting . However, most of these materials have high activation barriers for H2 production. The state-of-the-art electrocatalyst consists of platinum and 2D materials hybrids that require external energy to overcome the activation barrier . |
619f74ac568d334004462b5c | 2 | Here we demonstrate instantaneous (without any external energy) generation of highly pure hydrogen using Gadolinium Telluride (GdTe) crystal from water splitting (Figure ). The GdTe is produced by a simple and easily scalable induction melting (see Method section in supplementary material). The catalyst is recyclable without loss of any activity. Post reaction analysis of catalysis confirms the formation of non-uniform oxide layers (Gd2O3) which assists the reaction. Theoretical calculations indicate that the charge transfer from Te to Gd enhances the chemical activity which results in ultralow overpotential for instantaneous hydrogen generation. |
619f74ac568d334004462b5c | 3 | GdTe crystals produced by induction melting method, (using Gd and Te in 1:1 atomic ratio see ESI) crystallizes in the well-known NaCl-type fcc structure (SG: Fm-3m) (Figure ) with lattice parameters of a = b = c = 6.25 Γ
and Ξ± = Ξ² = Ξ³ = 90ΒΊ. This matches with the optimized lattice parameter of GdTe obtained from the DFT calculations and previously reported values (Figure ). Synthesized sample shows four major peaks corresponding to E 1 g, Ag, B 2 g, and E 2 g modes in the Raman spectrum (Figure ) which confirms its purity . Also, Energy dispersive X-Ray analysis (EDAX) shows the presence of Gd and Te after melting (Figure ). |
619f74ac568d334004462b5c | 4 | To understand the catalytic activity of produced GdTe crystal, its ability of splitting the water to generate hydrogen without external energy is demonstrated. A large number of hydrogen bubbles were generated immediately as GdTe reacts with water at room temperature (Figure and movie S1). In the beginning, a steady HER rate of ~0.022 ΞΌmol min -1 gm -1 was observed for 5 min which then increases rapidly after 10 min where eventually ~1.0 g of GdTe produced more than ~1.87 Β΅mol of H2 within 30 min (Figure ). Along with water, GdTe also reacts with dry methanol and produced H2 with slower rate (Figure ). Moreover, GdTe shows the most favorable onset potential when compared with state-of-the-art HER catalyst (Figure ) . |
619f74ac568d334004462b5c | 5 | Such high activity was further quantified and compared by calculating the free energies of HER for different closed packed planes of GdTe, namely (100), (110), (111) and (211), found from the experimental XRD using Computational hydrogen electrode (CHE) model by Peterson et al. . Computed free energies of HER are compared in (Figure ). H* intermediate is most strongly adsorbed on the Gd atom in the terrace site of (211) plane, with adsorption energy of -1.03 eV where formation of H2 is highly endothermic and promotes poisoning of the surface. |
619f74ac568d334004462b5c | 6 | However, low stability of the intermediate on Te atom of (110) surface with adsorption energy 1.78 eV also decreases the reactivity of GdTe. Therefore, considering all possible different adsorption sites, H* is optimally stable on Gd atom of (100) plane having free energy of hydrogen adsorption, βG = 0.06 eV. This outperforms Platinum (Pt), which otherwise showed prominent capability for HER with βG = -0.09 eV. Similarly, the Gd atoms in (110) plane and corner site of (211) plane have substantial potential for HER with βG = 0.14 eV and 0.21 eV, respectively. Further, according to calculated overall surface activities, (100) plane shows best potential for HER reaction followed by (110), ( ) and ( ) surface. Eventually this ensures uninterrupted formation of large number of H2 as (100) plane has lowest surface energy of 0.02 eV/Γ
2 . |
619f74ac568d334004462b5c | 7 | After 30 min of the reaction, the H2 generation reaches saturation (Figures ) and the color of the water changes from transparent to blackish brown (Inset of Figure and Figure ). The reacted particles separated from the water after reaction saturation was found in the nanoscale range (a few nanometers) under microscopic analyses (Figure ). Thus, nanostructuring of GdTe is possible by the simple addition of water without any external energy and the GdTe surface morphology changes with time and forms black-colored nanoparticles in water (Figure ). This surface reaction is shown by in-situ optical microcopy (Figure and GdTe crystal along with the formation of Gd-Ox. The de-convoluted O 1s spectrum shows two peaks positioned at 530.2 eV and 531.6 eV, which could be ascribed to oxygen-metal bonds and surface adsorbed oxygen, respectively. The core level spectra of Gd 3d reveal its characteristics peaks at 1186.5 and 1218.6 eV by virtue of 3d5/2, and 3d3/2, respectively which originates due to the spin-orbital coupling on a doublet. Absorption spectra also confirms the regeneration of GdTe after reaction and melting (Figure ). Therefore, due to the presence of GdTe in GT-II sample, it regained its ability to produce H2 when reacted with water at equivalent rate. Although the amount of generated H2 decreases from freshly prepared GdTe, Further, The Te 3e XPS profile can be assigned to Te 0 (572.9 eV), O-Te3d3/2 (575.9 eV), C-Te3d5/2 (583.3 eV), and O-Te3d5/2 (586.2 eV) peaks (Figure ), indicating the formation of elemental Te and oxides of Te along with Gd2O3 and Gd2Te3 . |
619f74ac568d334004462b5c | 8 | To understand the role of produced by-product on the reaction kinetics of GdTe in water, Thermodynamics calculations were performed at 25β by varying the mole fraction of reacting reagents (Figure ). Both the production of H2 and Gd2O3 increase with increase in mole fraction of GdTe that reacts with 1 mole of H2O. However, their production saturates at around 0.66 mole fraction of GdTe where the amount of Gd2Te3 starts increasing. Further, for all increasing mole fraction of H2O that reacts with 1 mole of GdTe, production of both H2 and Gd2O3 increases. However, the GdTe is fully consumed at around 0.5 mole fraction H2O and the amount of Te starts increasing. This indicates that the formation of Gd2O3 helps in formation of . This causes Gd to atoms to react with water and form Gd2O3. However, 3-coordinated Te atom in (111) surface is more reactive than 6-coordinated Gd atom and hence, forms respective oxide. Further, the reaction of Gd in water shows less activity (0.0011 ΞΌmol min -1 gm -1 ) for HER in comparison with GdTe (Figure ). As, d-band centre of Gd atom is at -0.35 eV for Gd (100) surface, which is less appropriate than the Gd atom on (100) surface of GdTe for the adsorption of hydrogen. |
619f74ac568d334004462b5c | 9 | Moreover, according to the DOS plot of Gd (100), the filled states near fermi level is contributed by the filled dx2-dy2 and dxy orbital. However, in GdTe (100) surface, the filled states near fermi level are mainly contributed by the dxz orbital (Figure ). This is because in GdTe the Gd atoms interact with the Te atoms by their dx2-y2 orbitals as confirmed by the electron accumulated blue region around Te atoms that is spread along the bond (Figure ). |
619f74ac568d334004462b5c | 10 | Correspondingly, the electron density in the dx2-y2 orbital in Gd decreases as Te atoms being more electronegative pulls electron density from Gd. Now, H atom approaching the surface from the z-direction interacts by its s-orbital with the orbitals of the surface atoms having more zdirected orientation (dxz, dyz, dz2). This in turn proves the superiority of GdTe bimetallic surface over Gd surface. |
619f74ac568d334004462b5c | 11 | Stoichiometric amounts of 99.99% pure gadolinium and tellurium were taken in a quartz tube. The metals were melted at a temperature of 1500 β under vacuum conditions and cooled by quenching to room temperature using Argon. Although the melting point of GdTe is 1825 β, a lower temperature was sufficient because the high vacuum conditions of 1 x 10 -5 mbar inside the furnace reduces the melting point of the metals and the compound. Hydrogen evolution reaction started immediately after contact with the atmosphere. Thus, the samples were stored under a vacuum. |
619f74ac568d334004462b5c | 12 | The GdTe polycrystals obtained were powdered and XRD patterns were obtained using a Bruker D8 Advance Diffractometer with Cu-KΞ± (Ξ» = 1.5406 Γ
) radiation and characterized based on the Pearson's Crystal Database. Room temperature Raman spectroscopy was done using WITec UHTS Raman Spectrometer (UHTS 300 VIS, Germany) at a laser excitation wavelength of 532 nm. UV-Visible Spectroscopy was used to study the optical absorbance. |
619f74ac568d334004462b5c | 13 | The catalyst was enclosed in a glass vial, and gas inside the vial was taken as a blank gas sample. Deionized water (1 ml) was added to the glass vial, and gas evolution was analyzed in a specified time interval. GdTe reacted with water immediately at room temperature and produced gas bubbles on the surface of the GdTe sample in the glass vial. Quantitative analyses of the product were carried out to corroborate the hypothesis of Hydrogen production and to determine considering the first moment of d-states for Gd atoms and p-states for Te atoms. All the calculations were spin-polarized. |
619f74ac568d334004462b5c | 14 | Posterior to the electronic and lattice optimization of the bulk GdTe structure, taken from Materials Project database (3), different surfaces, namely (100), ( ), ( ) and (211) were created each of them having 4 layers. Afterwards, 2Γ2Γ1 supercell of each considered surfaces was created followed by electronic optimization. ( The adsorption energy of the H* intermediate involved in HER was calculated from the DFT using the following expression (27, 28): |
619f74ac568d334004462b5c | 15 | where, βZPE is the difference in zero-point energies, βS is the change in entropy due to vibrational contributions, and βGU = -eU, where U is the applied electrode potential. The values of (βZPE-TβS) for the reference molecule (H2 gas) is 0.1328 eV. Calculated values of adsorption energies, (βZPE-TβS), free energies of hydrogen adsorption are given in Table . |
66fdb15251558a15efe0557a | 0 | While the design space of possible TMCs is immense, computational strategies provide a means to accelerate the discovery process by rapidly evaluating large samples of proposed complexes. Such high-throughput virtual screening (HTVS) campaigns typically rely on density functional theory (DFT) to solve for the approximate electronic structure of a given complex. Different TMCs may be compared in terms of DFT-calculated molecular properties, including their geometries, relative energies, and frontier orbital energies, to name a few. Although DFT alone is too slow to enable exhaustive sampling of the TMC design space, it may be combined with more rapid machine learning (ML) approaches to accelerate the process of rapidly screening TMCs and identifying promising candidates to guide experimental synthesis. A key challenge with this integrated HTVS approach is that DFT calculations are reliant on accurate three-dimensional initial structures of the systems to be analyzed, while ML models are often sensitive to the two-dimensional connectivity between the metals and ligands, both of which are difficult to determine a priori. Experimental datasets can provide meaningful insights regarding trends in structures and properties of synthetically-accessible TMCs. The Cambridge Structural Database (CSD) is a digital repository of crystal structures from X-ray and neutron diffraction of experimentally synthesized complexes, and is one such large dataset that has been used extensively to study TMCs. Large datasets of transition metal complexes and their constituent ligands have been curated, aided by the development of software designed for enhanced structure generation and parsing of crystallographic data. Such datasets have been used to train ML models for accurately predicting electronic properties of TMCs in the hopes of generalizing well to unseen complexes and guiding future design efforts. While prior work has involved dataset curation and model development for property prediction, these efforts have largely been focused on TMCs with known metal-ligand connectivity. One exception is accurate models that were developed based on CSD structures to identify when a ligand may exhibit multiple coordination modes. Most efforts centered exclusively on predicting ligand properties typically assume a single known metal-ligand coordination, leaving researchers with no accurate way to determine how a novel, unseen ligand will coordinate in a TMC without resorting to expensive and time-consuming experimental characterization techniques. |
66fdb15251558a15efe0557a | 1 | In this work, we leverage graph neural networks to answer the question of ligand coordination to transition metals with high accuracy and precision. We curate mononuclear TMCs from the CSD and extract their coordinated ligands to curate a large dataset of unique ligands with known connectivity to transition metals. This dataset provides valuable insight into chemical trends in ligands of particular coordination modes and reveals under-sampled regions of chemical space in ligand design. We train graph neural networks on the curated dataset to accurately predict coordinating atoms of unseen ligands directly from their SMILES representations. We interpret the trained models through analysis of the learned molecular representations and uncertainty quantification, demonstrating strategies for synergistic use of models in cases of high prediction uncertainty. Finally, we leverage the trained models by integrating them with the high-throughput structure generation code molSimplify. We illustrate this functionality through a case study in TMCs exhibiting novel metal-ligand combinations, validating the generated structures with DFT calculations to demonstrate our strategy to generate TMCs with de novo combinations of transition metals and ligands in physically realistic coordination. |
66fdb15251558a15efe0557a | 2 | To curate a dataset of synthetically accessible ligands with known connectivity in TMCs, we queried the CSD for all entries containing a single transition metal atom in any coordination geometry with high resolution (see Methods). From this set, we used molecular graphs to identify 299,035 complexes with unique connectivity, and we extracted ligands from these complexes using molSimplify. This process resulted in a total of 95,385 unique ligands (SI Appendix, Table ). We filtered these unique ligands in several steps. In terms of elemental filters, we kept only those that consisted of atoms frequently present in organic ligands (i.e., H, B, C, N, O, F, Si, P, S, Cl, Br, I), and we required that the metal-coordinating atoms were common (i.e., C, N, O, Si, P, S, Figure and SI Appendix, Table ). We also restricted our ligands to those with between one and six total coordinating atoms as defined by the mol2 file representing each CSD entry. |
66fdb15251558a15efe0557a | 3 | Experimentally characterized ligands are typically classified according to their denticity (i.e., number of distinct electron-donating groups) and hapticity (i.e., number of electron-donating atoms in a continuous series). However, such distinctions become difficult to maintain with respect to TMC connectivity processed from recorded structures (i.e., in mol2 files) in the CSD. In a connectivity record associated with a structure file, the bonds act as proxies for spatial relationships rather than as descriptions of the nature of the chemical bond. Because our goal is to accurately predict metal-ligand coordination and generate 3D structures of the resulting TMCs, we opt to classify ligands by the total number and identity of coordinating atoms, avoiding the distinction between denticity and hapticity. We also required that the ligand consist of more than one atom (Figure and SI Appendix, Table ). The decision to include only ligands with up to six coordinating atoms was motivated by the observation that ligands with seven or more coordinating atoms were extremely rare in the dataset: at the point of applying the filter on the number of coordinating atoms, only 337 ligands with seven or greater coordinating atoms were removed (SI Appendix, Table ). Finally, we only retained ligands for which a valid and unique SMILES could be written by RDKit, 75 where failures were typically due to kekulization failure in aromatic systems (see Methods and SI Appendix, Table ). Hemilabile ligands exhibiting multiple unique coordination modes were removed to ensure the dataset contains only ligands of unambiguous coordination that also do not exhibit coordination preferences that depend on metal identity. |
66fdb15251558a15efe0557a | 4 | Following additional filtering steps to ensure computational compatibility with a high-throughput screening workflow, we obtained a total of 70,069 unique, non-hemilabile ligands (see Methods and SI Appendix, Table ). Due to the inherent difficulty in identifying confident negative labels for hemilability, a set of 6,031 hemilabile bidentate, tridentate, and tetradentate ligands we previously curated with semi-supervised learning was used in the present work for training models to predict hemilability (see Methods and SI Appendix, Table ). Structures of the most frequently occurring ligands in the dataset prior to removing duplicates are shown in the top inset for each number of coordinating atoms, with the coordinating atoms circled in blue. Distributions of coordinating atom elements, number of coordinating atoms, and transition metals to which ligands coordinated are also shown for the curated dataset of unique ligands. Atoms are colored as follows: H in white, C in gray, N in blue, O in red, S in yellow, and Sb in purple. |
66fdb15251558a15efe0557a | 5 | To investigate the diversity of ligands in the curated dataset, we evaluated the distribution of coordinating atom counts from one-coordinate to six-coordinate as well as the most frequently occurring ligands. Bidentate (i.e., two-coordinate) ligands are the most well-represented in the curated dataset, accounting for 38% of the data (Figure ). This observation is to be expected, as bidentate ligands are widely used in TMC synthesis due to the high stability of the resulting complexes, which is attributable to the chelate effect and their low ring strain when compared to higher-denticity ligands. Ligands with only one (22%), three (19%), and four (15%) coordinating atoms are also well-represented in the dataset, while the rarely occurring ligands with six coordinating atoms (4%) are only slightly more represented than five-coordinate ligands (3%) (Figure ). The top five ligands that were the most represented before removing duplicates are all monodentate ligands: carbonyl, trimethylphosphine, tetrahydrofuran, nitrosyl, and methanol, all of which are well-known and widely used ligands or solvents in inorganic chemistry. The most common two-coordinate ligand is 1,2-dicyanoethene-1,2-dithiolate (Figure ). The most common four-coordinate ligand is the well-known tetraphenyl porphyrin, and other common ligands include three-coordinate 2,6-di(1H-pyrazol-3-yl)pyridine, five-coordinate N-benzyl-N,N',N'-tris(2pyridylmethyl)-1,2-diaminoethane, and six-coordinate p-cymene (Figure ). These ligands include various multidentate and haptic coordination modes with unique coordinating atoms, representing the diversity of the curated dataset. |
66fdb15251558a15efe0557a | 6 | We next analyzed trends in common coordinating atoms and metals to gain insight into common metal-ligand environments represented in the curated dataset. The most common coordinating atom in the ligand dataset by far is nitrogen, representing over 52% of the total number of coordinating atoms across all ligands (Figure ). The 2p coordinating atoms carbon (19%) and oxygen (15%) are the next most common, while the 3p coordinating atoms phosphorus (7%), sulfur (6%), and silicon (0.4%) are less abundant (Figure ). We include silicon in our dataset despite its low occurrence because it is a known coordinating atom used in ligand design for TMCs and occurs far more frequently than any other coordinating atom element removed during filtering (SI Appendix, Figure ). Due to the imbalanced distribution of coordinating atom elements in the overall dataset, we anticipate that machine learning models will perform better on ligands containing the more abundant 2p coordinating atoms carbon, nitrogen, and oxygen rather than 3p coordinating atoms silicon, phosphorus, and sulfur. When analyzing the coordinating atom frequency for each specific denticity individually, we identify distinct trends from the global frequencies (SI Appendix, Figures ). For example, nitrogen is the most common coordinating atom overall, but carbon is the most common in the individual analysis of ligands with six coordinating atoms, which is indicative of the high presence of benzene coordinating motifs in six-coordinate ligands (SI Appendix, Figure ). We also assessed trends by ligand coordination number in the most frequent metal identities. While ruthenium is the most common metal overall and for one or six coordinating atoms, the sets of ligands with two, three, four, and five coordinating atoms each favor different metals (SI Appendix, Figures and). The abundance of one-coordinate ligands observed coordinated to ruthenium is in agreement with the known chemical relevance of Grubbs catalysts, which typically contain monodentate ligands. Six-coordinate ligands coordinated to ruthenium are similarly well-represented likely due to their use in ruthenium piano-stool complexes widely studied for their cytotoxic activity. Fivecoordinate ligands are most frequently observed coordinated to iron, consistent with the abundance of ferrocene structures in the CSD. While the cyclopentadienyl binding motifs characteristic of ferrocenes are an example of ligands with five coordinating atoms but different denticities and hapticities, many five-coordinate, pentadentate ligands consisting of bridged pyridine and pyrazole substructures are also observed. On the other hand, four-coordinate ligands being most frequently observed coordinated to nickel is in line with the known significance of bis(cyclooctadiene) nickel and related low-valent nickel species, which include four-coordinate octadiene and other haptic cycloalkene ligands. Two-and three-coordinate ligands are most often coordinated to copper, the second-most common transition metal center observed in the overall dataset, although their occurrence with ruthenium is still high. Despite these differences in metal preference among ligands with different numbers of coordinating atoms, nearly all metals are observed at least once in each category. The only exception to this is gold, which is not observed with five-or sixcoordinate ligands (SI Appendix, Figures ). |
66fdb15251558a15efe0557a | 7 | We next endeavored to determine whether it was possible to learn the preferred ligand denticity and identify of coordinating atoms based on the experimental ligand structures we curated from the CSD. To avoid data leakage, we separated 95% of the data into training, validation, and test sets based on Tanimoto similarity of the RDKit fingerprints 75 of each ligand (see Methods). |
66fdb15251558a15efe0557a | 8 | In addition, we set aside a final holdout set (5%) also selected based on Tanimoto similarity for analysis of model generalization. Using Tanimoto similarity as a metric, we observe that many ligands are dissimilar from others in the dataset, suggesting that our separate train/test sets are sufficiently distinct but indicating potential challenges in training models accurate across the entire dataset. To predict the number of coordinating atoms in a ligand, we treat this as a graph-level multiclass classification problem (i.e., to select one of six possible classes). To identify individual coordinating atoms, we treat this as a node-level binary classification task, with the target represented as a length N vector where N is the number of atoms in the molecular graph and each atom indicated as being either coordinating (1) or non-coordinating (0) (see Methods). Given the distinct nature of the way the models are trained, the models for predicting the count and identity of coordinating atoms, we will later show that they provide synergistic information that can also be used to assess each model's reliability. |
66fdb15251558a15efe0557a | 9 | To train low-cost models that generalize well across unique ligands, we opted to start from a string-based (i.e., SMILES ) representation to train message-passing neural networks. Specifically, we independently trained both models using a combined directed message-passing neural network (D-MPNN) and feed-forward neural network (FFN) architecture. We used the validation data for hyperparameter tuning and model training (see Methods and SI Appendix, Table ). First, we trained the multi-class model to predict the total number of coordinating atoms in each ligand. This SMILES-based model achieves 88.5% accuracy and its high receiver operating characteristic area under the curve (ROC-AUC) score of 0.98 indicates effective distinguishment between the multiple classes (Figure ). Nearly identical precision and recall scores (0.82) suggest the model is balanced and has no bias toward false positives nor false negatives (Figure ). While evaluating accuracy over each individual class is misleading due to class imbalance, comparing precision and recall reveals better model performance among the ligands well-represented in our data set (i.e., one coordinating atom -0.89 precision and 0.92 recall, and two coordinating atoms (i.e., five coordinating atoms -0.57 precision and 0.60 recall, and six coordinating atoms -0.80 precision and 0.72 recall, Figure and SI Appendix, Table ). The poor performance on higherdenticity ligands is expected due to the rarity of samples in these classes in the overall dataset. |
66fdb15251558a15efe0557a | 10 | Analysis of the confusion matrix demonstrates that incorrect predictions are typically off by only one class (Figure ). Overall, the good accuracy of the model combined with the low cost and complexity for generation of the SMILES string suggests this model should provide a good approach to identification of likely ligand coordination number. While the overall accuracy of the model trained to predict the number of coordinating atoms is high, identifying the specific metal-coordinating atoms is essential for automating the generation of transition metal complex structures. For the model to predict coordinating atom identities, we again trained a D-MPNN and FNN combined model, using a similar training and hyperparameter tuning strategy to that used in the model trained to predict the total number of coordinating atoms (see Methods and SI Appendix, Table ). In the case of coordinating atoms, the model predictions are made individually on each atom within a ligand. For select cases where resonance applies, two or more atoms may be equivalently correct but one might be identified as the sole "true" correct coordinating atom, and we automate the identification of these cases to ensure they are identified as correct (SI Appendix, Text S1 and Figure ). The trained model reaches a very high overall accuracy of 98.8% and ROC-AUC score of 1.00, with similar precision and recall scores of approximately 0.97 (Figure ). Nevertheless, these metrics alone are misleading due to the inherent class imbalance between coordinating and non-coordinating atoms, and alternative measures are needed, even beyond the balanced accuracy (96.9%) that is still very high. Ideally, to accurately predict how a ligand will coordinate to a transition metal, we require that all predictions of coordinating atoms within that ligand be correct. We thus define our most stringent measure, molecular accuracy, in terms of the percentage of predictions where every individual atom-level prediction in a given ligand is correct. The trained model reaches a molecular accuracy of 84.8% when evaluated on the test set (Table ). Grouping coordinating atom predictions by the known total number of coordinating atoms in each ligand, we observe the highest molecular accuracies on two-coordinate (87.6%) and three-coordinate (86.2%) ligands (SI Appendix, Table ). Interestingly, one-coordinate, monodentate ligands have lower molecular accuracy (82.8%) in comparison to the relatively small four-coordinate set (87.8%), which also corresponds to lower balanced accuracy (SI Appendix, Table ). One reason for this lower molecular accuracy on monodentate ligands could be that the chemical structure is more constrained, and thus learnable, in the higher-denticity ligands. For example, an imidazole, triazole, or tetrazole-derived ligand with multiple nitrogen atoms in a monodentate structure could lead to only a single correct answer from a range of options, whereas two-coordinate or threecoordinate ligands with two or more nitrogen atoms are easier to predict. Finally, predictions made on five-coordinate (67.8%) and six-coordinate (67.3%) ligands again have lower molecular accuracy due to their smaller numbers in the curated dataset. The disparity between the balanced and molecular accuracy across the entire dataset highlights difficulties in accurately representing the chemical interactions across an entire ligand. |
66fdb15251558a15efe0557a | 11 | Table . Accuracy of models trained to predict coordinating atom identities as evaluated on test set (6,616 ligands). Overall accuracy considers predictions made on all atoms for all ligands in the dataset. Positive accuracy considers only predictions made on true coordinating atoms, while negative accuracy considers only predictions made on true non-coordinating atoms. Balanced accuracy is the average of positive and negative accuracy. To determine if we could improve upon the models trained on a learned representation by incorporating domain-specific chemical information, we considered alternative descriptors. First, since canonical SMILES strings do not explicitly account for hydrogens, we considered whether adding explicit hydrogen atoms would aid model predictions. We anticipated that this additional information could enhance model predictions derived from greater information about the local environment of each potential coordinating atom. We also considered whether adding quantum mechanical information could aid our models. Specifically, we used reactivity descriptors (i.e., Hirshfeld partial charges, Fukui indices, NMR shielding constants, bond length and order) predicted from literature-reported machine learning models to mitigate computational cost of featurization. These features were generated and added to the models to predict the number and identity of coordinating atoms (see Methods and SI Appendix, Tables ). Finally, to attempt to account for the class imbalance in coordinating atoms, we also trained models with an increased weight applied to positive predictions. Despite the expectation that quantum mechanical information or explicit hydrogens should inform the models, we observed comparable performance of these new models (SI Appendix, Tables ). Specifically, adding explicit hydrogens to denticity prediction lowered model accuracy slightly (88.5% to 87.6%), with a similar reduction in accuracy (to 87.8%) observed after adding QM descriptors (SI Appendix, Table ). Adding both explicit hydrogens and QM descriptors together resulted in a model with only marginally increased accuracy (89.0%), possibly due to the synergistic benefit of these two models to resolve subtle differences in ligands with similar heavy atoms but distinct saturation (SI Appendix, Table ). For coordinating atom prediction, positive weighting worsened molecular accuracy substantially (to 70.7% versus 84.8%, SI Appendix, Table ). For addition of explicit hydrogen atoms and QM descriptors, either individually or in combination, we observed essentially unchanged (i.e., less than 1% difference) overall/positive/negative/balanced accuracy as well as molecular accuracy (SI Appendix, Table ). These observations are consistent with previous work suggesting the addition of QM descriptors to D-MPNN architectures has little benefit for datasets as large as the one considered here. Given the additional complexity and cost of these alternative models and the desire to avoid using an overparameterized model less likely to generalize well, we retained our original models and used the learned molecular representations for further analysis. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.