id
stringlengths 24
24
| idx
int64 0
402
| paragraph
stringlengths 106
17.2k
|
|---|---|---|
6578b127bec7913d2774045d | 21 | The fluorescence images were recorded using an Andor Zyla sCMOS camera from Andor Oxford Instruments using the Andor Solis software. Videos were sampled at 2000 frames per second with an image dimension of 45 × 200 pixels. The analogue-to-digital-units (ADU) fluorescence intensity was converted to the number of photons by calibrating the camera detector gain from a series of gradient images (Figure ). |
647ad85ebe16ad5c57636b44 | 0 | Molecular generation is highly significant for its applications in novel drug discovery, material design, and the exploration of chemical space. It enables the efficient search and identification of new drug candidates, speeding up the process of drug development . In material design, it allows for the creation of materials with tailored properties, contributing to advancements in various industries . Furthermore, molecular generation aids in the systematic exploration of chemical space, uncovering novel compounds with unique properties and expanding our understanding of molecular diversity . Overall, it has the potential to revolutionize the fields of drug discovery, material science, and chemical research for accelerating scientific innovation . |
647ad85ebe16ad5c57636b44 | 1 | Whereas it has been estimated as there are theoretically more than 10 60 small organic molecule chemical structures , the number of molecules actually explored in drug discovery is limited to about 10 8 at most . To efficiently propose new molecules with desirable physicochemical properties from a wide chemical space, an artificial intelligent (AI) technique called molecular generative model has been studied in recent years . As an input of the AI model, the chemical structure is represented in two ways: SMILES and a molecular graph . In general, molecular graphs are more robust and precise to represent the molecular features than SMILES, because graph representations can capture the molecular similarity and can consider chemical checks, such as protecting the number of valence electrons, unlike SMILES representations . |
647ad85ebe16ad5c57636b44 | 2 | The two main learning methods for molecular generative models are deep learning and reinforcement learning . Since deep learning-based models, which learn molecular features of known compounds , tend to generate molecules similar to the learned compounds, ability of generating structurally new compounds is fundamentally limitted . On the other hand, reinforcement learning-based models, which learn molecular features from scratch without prior learning of known compounds, is superior in generating molecules with structures distinct from known compounds . However, the generated molecules fundamentally lack drug-like properties due to the algorithm exploring chemical space distinct from the existing compounds. |
647ad85ebe16ad5c57636b44 | 3 | In this study, we proposed a new molecular generative model that can explore chemical spaces unreachable by previous molecular generative models and discover new molecules with druglike properties by combining deep learning and reinforcement learning based on a molecular graph representation. Specifically, the proposed method uses chemical features which learned physicochemical properties of known compounds using the Variational Graph Auto-Encoder (VGAE) and generates molecules with desirable properties through reinforce learning with Monte Carlo Tree Search (MCTS) . Evaluation of the generated molecules demonstrated that the validity and novelty of the chemical and the optimization of physicochemical properties was equivalent to or better than the previous methods. Furthermore, investigating the chemical structure diversity showed that the generated molecules are distributed in chemical space that was not well explored by the previous methods. The proposed method is expected to be useful for efficiently discovering and designing new molecules in the drug development. |
647ad85ebe16ad5c57636b44 | 4 | We have developed a new molecular design model that combines a deep learning model, Variational Graph Auto-Encoder (VGAE), and a reinforcement learning model, Monte Carlo Tree Search (MCTS). Our developed model (called VGAE-MCTS) is divided into three parts: a part for preparation of input data, a part for training of VGAE, and a part for molecular generation using MCTS (Figure ). Details of each part are described in Materials and Methods section. The basic performance of molecular generation using VGAE-MCTS, namely validity, uniqueness, novelty, Kullback-Leibler divergence (KL divergence), and Fréchet ChemNet Distance (FCD) was evaluated with Distribution-Learning Benchmarks in the GuacaMol framework (Table ). We compared the performance of VGAE-MCTS among the previous models that are Graph MCTS and VGAE . |
647ad85ebe16ad5c57636b44 | 5 | The ability of VGAE-MCTS to generate molecules was evaluated when the physicochemical properties were optimized. The physicochemical properties to be optimized are the Quantitative Estimate of Drug-likeness (QED) and penalized logP. QED is a quantitative measure of druglikeness and an evaluation of the ability to optimize single properties. Penalized logP is an index that combines three physicochemical properties: liposolubility, synthetic accessibility score, and penalty for large rings, and is an evaluation of the ability to optimize multi-properties. Both indices are commonly used physicochemical properties of drug discovery in the evaluation of molecular generation models . These indices range from 0 to 1, with molecules closer to 1 indicating that they are better molecules for drug discovery. We compared the performance of VGAE-MCTS among the prior study models that are JT-VAE 10 and MolDQN . |
647ad85ebe16ad5c57636b44 | 6 | As with the QED optimization, we compared the molecules generated by optimizing penalized logP using VGAE-MCTS with the molecules in the ZINC dataset and molecules generated using JT-VAE and MolDQN (Figure (A), (B), and (C)). The penalized logP of the molecules generated by VGAE-MCTS (mean: 0.536, median: 0.606) was higher than those in the ZINC dataset (mean: 0.572, median: 0.610) (Mann-Whitney U test: P=6.08×10 -1 ). We found that the molecules generated by VGAE-MCTS had a smaller percentage of low penalized logP values than the molecules in the ZINC dataset and the molecules generated by the previous models. In other words, this suggests that VGAE-MCTS avoids expanding molecules toward the lower penalized logP in the molecular generation using MCTS. VGAE-MCTS was also able to generate molecules with higher penalized logP compared to previous methods, JT-VAE (mean: 0.392, median: 0.309) and MolDQN (mean: 0.472, median: 0.442) (Mann-Whitney U test: P=4.26×10 -14 , P=1.90×10 -3 ). |
647ad85ebe16ad5c57636b44 | 7 | First, the basic performance of VGAE-MCTS in generating molecules was evaluated with the Distribution-Learning Benchmarks in the GuacaMol framework. The molecules generated by VGAE-MCTS had a validity of 100%. This result is due to the fact that the chemical structure is represented as a molecular graph and the MCTS creates molecules by connecting atoms and bonds while protecting the number of valence electrons. The uniqueness and novelty scores of VGAE-MCTS were also higher. These results are thought to be due to the fact that VGAE-MCTS is able to output a wide variety of molecules because the more atoms that make up a molecule, the more molecules are candidates for expansion, and the type of atoms selected is stochastic. |
647ad85ebe16ad5c57636b44 | 8 | Next, molecular generation was performed using VGAE-MCTS to optimize for each of the two types of physicochemical properties values, QED and penalized logP, and in both cases, the accuracy was confirmed to be equal to or better than that of previous studies. The molecules generated by QED optimization are more drug-like than those generated by the models in the previous methods, indicating that VGAE-MCTS is a valuable method for use in drug discovery. |
647ad85ebe16ad5c57636b44 | 9 | Penalized logP is composed of a combination of the three physicochemical properties of logP, SA score, and RingPenalty, and multi-property optimization was relatively successful in VGAE-MCTS. This suggests that VGAE-MCTS can be used to search for molecules considering multiple physicochemical properties. VGAE-MCTS can be expected to be used in practical drug discovery process where multiple conditions are optimized. |
647ad85ebe16ad5c57636b44 | 10 | Finally, we evaluated whether the molecules generated by VGAE-MCTS were able to expand the chemical space from the training data. Because the ZINC dataset is registered for drug-like compounds, many molecules have a large QED, a quantitative measure of drug-likeness (approximately 92% of the molecules in the ZINC have a QED ≥ 0.5). Therefore, we can evaluate whether the generated molecules have expanded their chemical space using the chemical space of drug-like molecules in the ZINC as a reference. or better performance than existing models in the GuacaMol benchmark. We also showed that the performance of the optimization of the physicochemical properties, QED and penalized logP, is comparable or better than previous studies. In addition, to assess the diversity of chemical structures generated, we evaluated the distribution of molecules generated by VGAE-MCTS and several previous models in chemical space. The results indicate that the molecules generated by VGAE-MCTS are distributed in areas that were not well explored by the molecules generated by the previous models. Based on these results, it is expected that our proposed VGAE-MCTS will be able to propose molecules that may have been out of the scope of exploration so far, which will be useful for drug development. |
647ad85ebe16ad5c57636b44 | 11 | GuacaMol's Distribution-Learning Benchmarks . The total number of compounds obtained from ChEMBL was 1,352,672, which were divided into 1,273,104 for training and 79,568 for validation. As the second dataset, compounds were obtained from ZINC , where drug-like compounds are registered, in order to evaluate the capability to optimize the physicochemical properties of molecular generation. The number of compounds obtained from ZINC was 249,456, which were divided into 199,565 for training and 49,891 for validation. |
647ad85ebe16ad5c57636b44 | 12 | Preparation of input data. The molecules of the training dataset are represented in a feature map, which is the data format for input to VGAE. The graph structure representation of a molecule is to convert it into a vector, with atoms represented by nodes and bonds by edges. In the conversion of the molecules to vector representation, node features and edge features were computed using RDKit . Details of node and edge features are shown in Tables and, respectively. The features of the nodes were concatenated to create a feature map of the nodes. |
647ad85ebe16ad5c57636b44 | 13 | The features of the edges were also concatenated to create a feature map of the edges. MCTS generates molecules using the feature maps output from the trained VGAE. In MCTS, the following 1) Selection, 2) Expansion, 3) Simulation, and 4) Update are considered one search and repeated for the number of times specified by the user. When the search is completed for the number of times specified by the user on one feature map, the molecule with the best physicochemical property value at each depth of MCTS is output for numerators below the userspecified depth (minimum_depth). Then, the search is moves on to a next feature map. |
647ad85ebe16ad5c57636b44 | 14 | where s is the score of the node, n is the number of times the node has been visited, c is the search coefficient (c = 1.5 in our case), and N is the number of times the parent node has been visited. Equation ( ) corresponds to Equation of the Upper Confidence Bound 1 (UCB1) , which is well known in reinforcement learning. At this time, the depth is increased by one with the selected node. |
647ad85ebe16ad5c57636b44 | 15 | In addition, to generate realistic molecules, the proposed model introduces two filters, a Steric strain filter and a filter to make it difficult to create a ring structures larger than 7-membered rings. If a node was trapped by at least one of these two filters, our method made it less likely to be selected as a node to be searched by increasing the MCTS reward value by a factor of 10. |
647ad85ebe16ad5c57636b44 | 16 | Optimization of physicochemical property. The Quantitative Estimate of Drug-likeness (QED) and penalized logP were set as the physicochemical properties to be optimized. QED is a quantitative measure of drug-likeness . QED is a quantitative measure of drug-likeness and ranges from 0 to 1, with values closer to 1 indicating that the molecule is more drug-like. When optimizing QED, the 1 -𝑄𝐸𝐷 score was used as the reward function for MCTS. |
647ad85ebe16ad5c57636b44 | 17 | In the evaluation of the optimization of the physicochemical properties, 3,000 molecules were randomly selected from the ZINC data set and 3,000 molecules were selected in the order in which they were generated by each molecule generation model. The distribution of physicochemical properties for the generated molecules was then calculated and evaluated. |
647ad85ebe16ad5c57636b44 | 18 | Statistical analysis. The Mann-Whitney U test was used to test for differences in the distribution of physicochemical property values between methods for molecules generated by optimizing QED and penalized logP. In addition, the significance probability p-values were corrected by Bonferroni's correction . The molecules used for the test were 500 randomly selected from the 3,000 molecules generated for each method. |
60c74efd9abda2258ff8d779 | 0 | COVID-19 following choral rehearsals. However, no direct comparison of aerosol generation from singing and speaking has been reported. Here, we measure aerosols from singing, speaking and breathing in a zero-background environment, allowing unequivocal attribution of aerosol production to specific vocalisations. Speaking and singing show steep increases in mass concentration with increase in volume (spanning a factor of 20-30 across the dynamic range measured, p<1×10 ). At the quietest volume (50 to 60 dB), neither singing (p=0. 19) or speaking (p=0.20) were significantly different to breathing. At the loudest volume (90 to 100 dB), a statistically significant difference (p<1×10 ) is observed between singing and speaking, but with singing only generating a factor of between 1.5 and 3.4 more aerosol mass. Guidelines should create recommendations based on the volume and duration of the vocalisation, the number of participants and the environment in which the activity occurs, rather than the type of vocalisation. Mitigations such as the use of amplification and increased attention to ventilation should be employed where practicable. KEYWORDS: SARS-CoV-2, COVID-19, aerosol generation, airborne transmission, aerodynamic size, singing A novel strain of a human coronavirus was first identified in late 2019, designated severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), and is responsible for the global outbreak termed coronavirus disease 19 (COVID-19). Pandemic status was declared on 11 March 2020 by the World Health Organisation (WHO), with in excess of 21.5 million cases and 767,000 deaths reported worldwide by 17 st August 2020. Early in the pandemic, clusters of COVID-19 were considered to have arisen in several choirs around the world. This rapidly led to many governments restricting or suspending singing. Concerns that woodwind and brass instruments might also be responsible for virus spread led to similar restrictions on the playing of wind instruments. Consequently, large sections of the cultural sector, along with religious institutions and educational establishments, were unable to rehearse and perform, resulting in profound artistic, cultural, spiritual, emotional and social impacts. The livelihoods of many performers have been jeopardised, and the viability of established institutions remains threatened. The economic impact to the United Kingdom (UK) from this sector alone has been substantial, costing the UK economy hundreds of millions in lost tax revenue, usually derived from the £32.2 billion cultural purse. Respiratory particulate matter is expelled during human exhalatory events, including breathing, speaking, coughing and sneezing. The flux generated is proportional to the amplitude of phonation in speech. These actions release a plume of material containing particles of varying size, ranging from macroscopic mucosalivary droplets originating from the oral cavity and pharynx, to microscopic aerosols released by the small airways of the lungs. Traditionally, the division between droplets, which are considered to be of sufficient mass to sediment due to gravity, and aerosols, which remain airborne, is defined arbitrarily at 5 µm diameter. However, particle composition and environmental properties like temperature, humidity and airflow influence the biophysical mechanics of the material released and the extent of transport. Droplets and airway secretions are established vectors of SARS-CoV-2, with expelled infectious material either directly inhaled by an individual in close proximity, or indirectly transmitted through contact with settled-out fomites. The role of airborne transmission by respirable aerosol particles is gaining prominence. Viral RNA has been detected in airborne samples collected both inside and outside the rooms of COVID-19 patients, and SARS-CoV-2 RNA has been reported in size-resolved aerosol distributions in two hospitals in Wuhan, China. Retrospective studies of COVID-19 clusters, including a shopping mall, a restaurant and a high-profile outbreak in an American choir group, found no direct or indirect interaction among the individuals contracting the virus, suggesting airborne transmission. SARS-CoV-2 and other viruses, including severe acute respiratory syndrome coronavirus (SARS-CoV-1) and Middle East Respiratory Syndrome coronavirus (MERS-CoV), are stable in aerosol. Infective airborne potential from human exhalation has been confirmed in other viruses, including respiratory syncytial virus, influenza and MERS-CoV. Several online reports have attempted to examine the quantities of particulate matter expelled by participants performing a range of activities including singing but have struggled to accurately quantify aerosol and droplets because of the large number of background particulates in the environment. This study is the first peer-reviewed study that explores the relative amounts of aerosols and droplets (up to 20 µm diameter) generated by a large cohort of 25 professional performers completing a range of exercises including breathing, speaking, coughing and singing in the clean air environment of an operating theatre with laminar flow ventilation. Measurements of particle number concentration alone would be insufficient to determine the total amount of viral material capable of being transmitted: the total mass of particulate matter produced may be a key factor in assessing the potential risk. Thus, measurements of particle size distributions, as well as concentration, are used to assess the mass concentration. |
60c74efd9abda2258ff8d779 | 1 | singers perform a broad range of genres, including musical theatre (6), choral (5), opera (5), and other genres: gospel (2), rock (2), jazz (2), pop (1), actor with singing interest (1) and soul (1). 6 identified their voice-type as soprano or mezzo-soprano, 7 as alto, 5 as tenor and 7 as bass or baritone. Aerosols and droplet concentrations were measured with an Aerodynamic Particle Sizer (APS, 500 nm -20 µm) in an operating theatre with each participant and researcher required to wear appropriate personal protective equipment. The high air exchange rate, filtration and laminar air flow reduced the pre-existing particle background number concentration to zero cm -3 , enabling the unique and extremely sensitive measurements described. Thus, any particles detected were directly attributable to participant activity, with particle concentrations returning to zero cm -3 between periods of singing, speaking and breathing. |
60c74efd9abda2258ff8d779 | 2 | A standard operating procedure was adopted (see Methods), covering 12 activities over ~1 hour, with each activity involving up to 5 repeat actions, with a 30 s pause between each. These activities included breathing, coughing, singing single notes ("/ɑ/") at different pitches, and speaking and singing the "Happy Birthday" song at different volumes. At the beginning of each action, participants stepped forward to the funnel (Fig. ) such that the dorsum of the nose was aligned to the plane of the base of the cone. Participant position relative to the funnel was monitored to ensure consistency (within 10 cm of the sampling tubes) across all measurements (Extended Data Fig. ). As in previous studies, we report concentrations sampled through the collection funnel, which allows comparison of particle emission rates on a relative basis between activities. In reality, particle concentrations will become rapidly diluted once particles are exhaled, leading to strong spatial variations. |
60c74efd9abda2258ff8d779 | 3 | A sequence of measurements made with one APS for one performer is reported in Fig. ). The bursts of activity, interspersed with periods of no activity, are visible above a zero background in aerosol concentration. participants is reported in Fig. . The statistical analysis is described in Methods and the absolute results summarised in Extended Data Tables and; data normalised to the aerosol concentration from speaking at 70-80 dB are compared in Extended Data Fig. and Table . The distribution of aerosol number concentration generated across all participants is assumed to be log-normal, consistent with the data presented in a previous publication; concentrations must always be positive-valued and a small number of individuals generate a significantly larger aerosol flux than the median. This is particularly apparent for breathing, where measurements from individuals span almost three orders of magnitude. Indeed, 4 participants produced more aerosol in number concentration while breathing than while speaking at 90-100 dB. The reproducibility of concentration from singing a single note (70-80 dB) is not only apparent in single participant data (Fig. ), but also across the cohort with median concentrations in good agreement (0.83 and 0.91 cm -3 at beginning and end, respectively). At the lowest volume (50-60 dB), neither singing (p=0. 19) or speaking (p=0.20) were significantly different in particle production to breathing, with median number concentrations of 0.10, 0.19 and 0.28 cm -3 for speaking, singing and breathing, respectively. In the mixed model, compared to speaking, singing generates a statistically significant (p < 1×10 -5 ) enhanced aerosol number concentration, although this enhancement is small relative to the much larger changes associated with increase in volume (p < 1×10 ). Aerosol number concentration increases by a factor of 10-13 as volume increases from 50-60 dB to 90-100 dB, suggesting that shouting should be associated with little difference in risk to singing at loud volume. compares aerosol number concentrations from speaking and singing at 90-100 dB for male and female participants and for the different genres with the full cohort. Individual participant comparisons are provided in Extended Data Fig. . There are no significant differences in aerosol production either between genders (p = 0.34) or among different genres (p(choral different from "other genres") = 0.46, p(musical theatre different from "other genres") = 0.25, and p(opera different from "other genres") = 0.42). The variability among genres (almost a factor of 2 between the lowest and highest median concentrations) may be attributed to the small cohort sizes for each genre, the sensitivity of number concentration to volume and a minority of participants emitting higher concentrations than others (who could be classed as super-emitters). In addition, there is no correlation between the mean aerosol number concentration generated by an individual participant when singing at 90-100 dB or breathing and the participant's body mass index or peak flow rate (Extended Data Fig. ). Fig. : Comparison of average aerosol number concentrations (linear scale) from speaking and singing at 90-100 dB by the full cohort, males (12), females (13), opera (5), musical theatre (6), choral (5) and other genres (9). |
60c74efd9abda2258ff8d779 | 4 | Comparing the Aerosol Particle Size Distributions and Mass Concentrations. The possibility that singing, speaking and breathing generate aerosol particles of different size cannot be inferred by comparing particle number concentrations alone. Instead, we must compare the aerosol size distributions from these activities. Previously, two overlapping modes in the size distribution of particles from speaking and coughing have been identified. These have been attributed to distinct processes in this expiration process. The mode of lowest size is generated in the lower respiratory tract with a second mode generated in the region of the larynx, expected to be the most important in voicing. Figure reports the variation in mean number concentrations with particle size averaged over the 25 participants and includes the fitted distribution from Johnson et al. reported from a cohort of 15. Our distribution for speaking and singing is in excellent agreement with the shape of the distribution reported by Johnson et al. for particles larger than 800 nm diameter. Although the absolute concentrations are a factor of ~6 larger in our measurements, it should be recognised that the absolute value carries little meaning, reflecting only the instantaneous value recorded by the APS from the sampling funnel, which will depend on the sampling specifications. Measured size distributions for speaking and singing were fitted to bimodal lognormal distributions. The fits all gave the similar mean diameters and variance for both modes, further supporting the conclusion that speaking and singing can be treated similarly (Extended Data Table ). However, both vocalisations generate larger particles than breathing: although the size distribution from breathing is well-represented by a bimodal lognormal distribution, the larger mode is shifted to a smaller diameter and has a narrower variance than for speaking and singing. . |
60c74efd9abda2258ff8d779 | 5 | The consequences of different size distributions are apparent when aerosol mass concentration is reported (Fig. , see Extended Data Table and). This comparison is most important when considering the potential of the different activities to transmit infection. Speaking and singing generate statistically significant differences in mass concentrations of aerosol at similar volumes; however, these are modest (median singing values only a factor of 1.5-3.4 times larger than speaking) relative to the effects of the volume of vocalization (a factor of 20-30 increase). Converting from a number concentration to a mass concentration for breathing results in the mass concentration range shifting to lower values relative to speaking and singing, a consequence of the different size distributions associated with voicing and breathing (median values 24 and 36 times higher for speaking and singing at the highest volume level, respectively, compared with breathing). |
60c74efd9abda2258ff8d779 | 6 | Discussion. This study demonstrates that the assessment of risk associated with the spread of SARS-CoV-2 in large groups due to respirable particles from speaking and singing should consider the number and mass concentrations of particles generated by these activities. The statistically significant, yet relatively modest differences detected between the type of vocalisation at the loudest volume studied, are eclipsed by the effects of volume on aerosol production, which varies by more than an order of magnitude from the quietest to loudest volume studied, whether speaking or singing. By contrast, the number of particles produced by breathing covers a wide range (spanning from quiet to loud volume speaking and singing) but has a size distribution shifted to smaller particle sizes, in principle mitigating some of the potential risk associated with the wider emission range. |
60c74efd9abda2258ff8d779 | 7 | We also find that a minority of participants emitted substantially more aerosols than others, sometimes more than an order of magnitude above the median, consistent with the long-tail of a log-normal distribution when viewed in linear-concentration space. This observation is consistent with a previous study. However, the highest emitters were not consistently the highest across all activities, suggesting the magnitude of emission from an individual may be highly activity specific. It is unclear why some participants emit substantially more than others, and further studies are required to better characterise the variability of aerosol emission across the population, as well as the consistency of emission from an individual over time. |
60c74efd9abda2258ff8d779 | 8 | These conclusions have important policy implications in the context of creating guidelines to reduce transmission of SARS-CoV-2. Breathing produces smaller particles than singing and speaking, suggesting that vocalisation may carry higher risk than breathing if the potential SARS-CoV-2 dose delivered by an individual infected with the virus scales with particle mass. Size distributions are comparable across speaking and singing at the same volume and generate relatively similar, yet statistically significantly different, numbers of particles. Most importantly, number concentrations from speaking and singing rise in parallel with increasing volume. Given that speaking and singing produce numbers of particles of the same order of magnitude, and that increasing volume increases that number by orders of magnitude, guidelines from public health bodies should focus on the volume at which the vocalisation occurs, the number of participants (source strength), the environment (ventilation) in which the activity occurs and the duration of the rehearsal and period over which performers are vocalising. For certain vocal activities and venues, amplification may be a practical solution to reduce the volume of singing by the performers. Based on the differences observed between vocalisation and breathing and given that it is likely that there will be many more audience members than performers, singers may not be responsible for the greatest production of aerosol during a performance, and for indoor events measures to ensure adequate ventilation may be more important than restricting a specific activity. |
60c74efd9abda2258ff8d779 | 9 | The Public Health England Research Ethics and Governance of Public Health Practice Group (PHE REGG) approved this study and all research was performed in accordance with relevant guidelines and regulations of the Ethical Review Board. We recruited 25 healthy volunteers (12 males and 13 females, ranging in age from 22 to 57 years old (mean 38, SD +/-9.8) through contact and collaboration with the entertainment industry. Informed consent was obtained from all participants prior to study participation. |
60c74efd9abda2258ff8d779 | 10 | All participants completed a pre-screening questionnaire including age, gender, professional status, singing training history and COVID 19 symptom status to fulfill inclusion/exclusion criteria. Only participants who self-reported no symptoms of COVID-19 and who had normal temperatures on the day of attendance were included. Each participant's weight, height and peak flow rate were measured before the aerosol measurements. Body mass index was calculated from the height and weight measurement. |
60c74efd9abda2258ff8d779 | 11 | Measurements were performed simultaneously with two APS instruments (TSI 3321) and one Optical Particle Sizer (OPS, 0.3 -10 μm, TSI 3330) sampling from the same custom-printed funnel. A comparison of measurements between the two APS instruments was linear, with a slope that deviated from 1 owing to different sensitivities of the instruments (Extended Data Fig. ). The OPS detected significantly more particles than the APS (up to a factor of 2), a consequence of the lower size detection limit of the OPS (to 300 nm) compared to the APS (to 500 nm) (Extended Data Fig. ). Including these smaller particles in our analysis significantly increases the number concentration but does not significantly change the particulate mass concentrations from the expiratory activities. |
60c74efd9abda2258ff8d779 | 12 | The sampling funnel was 3D printed from PLA (1.75 mm filament) by a RAISE3D Pro2 Printer (3DGBIRE). The funnel was 150 mm wide, 90 mm deep with 3 ports at the neck for sampling aerosol into up to three aerosol instruments (some combination of APSs and OPSs). All tubing was conductive silicone and 130 cm in length (TSI Inc., product number 3001788, inner diameter 0.19 inch, outer diameter 0.375 inch). |
60c74efd9abda2258ff8d779 | 13 | Measured total particle number concentrations were summed over the period of activity and divided by the duration of the activity, reporting a mean concentration (cm -3 ) with a standard deviation, i.e. the average concentration of particles sampled within the funnel volume during the activity. With coughs requiring < 1 s, no averaging across a time-dependent concentration is possible and only the integrated number concentrations per single cough are reported. Further, particle size distributions were recorded by the APS at 1 s intervals with 51 size bins equally spaced in the range 0.5 to 20 m in log(diameter) space. Average size distributions were calculated first by determining the mean size distribution for each participant and then calculating the mean and standard deviation across all participant size distributions for each activity. Mass concentrations were calculated assuming particle density was 1000 kg•m -3 . Our reported number concentrations and particle size distributions for speaking and breathing are consistent with previously published data. The lme package in R-software was used to fit linear random effect models with log-base-e transformed particle concentration or mass as the dependent variable. The independent variables were vocalisation (speaking or singing) and acoustic volume (50-60, 70-80 and 90-100 dB); the random effect was participant identification number. In the Figures, the lower and upper hinges (ends of boxes) correspond to the first and third quartile (the 25th and 75th percentiles). The upper whisker extends from the upper hinge to the largest value but no further than 1.5×IQR (where IQR is the inter-quartile range, the distance between the 1 st and 3 rd quartiles). The lower whisker extends from the lower hinge to the smallest value at most 1.5×IQR. Data beyond the ends of the whiskers are "outlying points" and indicated in red. All components of the box plots were calculated based on the logarithmically-transformed data, owing to lognormal nature of the data, but the plotted and tabular values reported are converted back to linear space for clarity. |
60c74efd9abda2258ff8d779 | 14 | Extended Data Figure : Mean particle number concentration as a function of the distance from the participant's mouth to the apex of the funnel. For this experiment, a participant sang the "Happy Birthday" song for 20 s at 80-90 dB five separate times at each distance. The reported value is the mean and standard deviation for each distance. When the participant vocalised 8-10 cm from the funnel apex, the measured number concentrations did not vary significantly (factor of ~1.5), whereas beyond 12 cm the measured number concentrations decreased by an order of magnitude or greater. For all participants in this study, the distance between the subject mouth and the funnel apex was 8-10 cm. |
60c74efd9abda2258ff8d779 | 15 | Extended Data Figure : Comparison of measurements across 8 participants from the OPS and the APS for which all data are reported in this paper. The OPS measures a larger number concentration because it detects smaller particles, which are generally more abundant than larger particles. Extended Data Table : Measured absolute number concentrations from the series of expiratory activities plotted in Fig. (in cm -3 ). Provided are the statistical parameters visualised by the box plot. Note that these parameters were calculated on the logarithmically transformed data (see Methods). The number of participants for each activity is given by n. Extended Data Table : Lognormal fit parameters for speaking, singing and breathing. For each activity, the size distribution averaged across all participants was fit to a bimodal lognormal fit. |
6710c55312ff75c3a1ad3b8f | 0 | Both operando characterization and predictive-quality multiscale modeling have by now firmly established that significant structural, compositional and morphological evolution of working catalyst surfaces is more the norm than an exception. Less clear is to which degree the dynamics of this evolution couples to the actual reaction chemistry. The very concept of an active site implies that the corresponding geometric and compositional motif prevails over time scales (much) longer than the catalytic cycle. Translated to the level of the involved elementary processes, this means that the processes underlying the operando evolution of the working surface are rare on the time scale of the molecular processes involved in the catalytic reaction. Identification of the active sites is frequently stated as a central objective for mechanistic understanding and rational catalyst design. Seen from the perspective of operando evolution, making this statement thus equates to (knowingly or unconsciously) assuming such a separation of time scales. The motif of a given active site must persist for long enough, say at least a few catalytic turnovers. Otherwise it makes no sense to characterize and discuss it. |
6710c55312ff75c3a1ad3b8f | 1 | To some extent, the widespread use of ab initio thermodynamics to computationally determine the surface structures exposed in given (reaction) environments also belongs to this thinking. In a divide and conquer approach, 3 a classic multiscale modeling ansatz to tackle a possible operando evolution would be to employ this technique to identify the thermodynamically most stable surface termination, and then use this termination as the basis for a subsequently developed microkinetic model for the given reaction (either full fledged to gain mechanistic understanding or reductionist to conduct a high-throughput virtual screening ). The identified termination can indeed differ largely in structure and composition from any bulk-truncated one of the nominal material. Yet, once determined, this termination is typically modeled as static and the elementary processes considered in the microkinetic model only concern the surface reaction intermediates. Notwithstanding, that this divide and conquer ansatz is state-of-the-art is arguably less due to compelling experimental evidence for such a separation of time scales and a concomitant sufficient persistence of active site motifs on the evolving surface. It also arises from limitations in present-day technical modeling capabilities. Including an elementary process into a microkinetic model is a twofold challenge. First, one needs to identify the actual process itself, i.e. realize that the atomistic process could potentially exist and play a role. |
6710c55312ff75c3a1ad3b8f | 2 | Second, within predictive-quality multiscale modeling, its rate constant needs to be computed from first-principles. Presently, for heterogeneous catalysts this is typically performed at the level of density-functional theory (DFT) and (harmonic) transition state theory. While this simplifies the task to the determination of the activation barrier aka transition state (TS) on the potential energy surface (PES), it still comes at a certain computational cost. For these reasons, existing microkinetic models for surface catalysis typically consider only a minimum number of elementary processes, which are hand-selected by the human researcher, and thus primarily focus on more intuitive surface reaction steps. |
6710c55312ff75c3a1ad3b8f | 3 | The computational limitations are less severe for nanocluster catalysts, and recent computational work in this field has in fact highlighted the crucial role of "fluxionality", i.e. a facile dynamic restructuring, for the reactivity of these clusters. It has been exactly demonstrated that there is no ready separation of time scales and that cluster isomerization processes can occur as frequently as the molecular processes of catalysis. At such a dynamic interconversion, different elementary processes of the catalytic cycle can therefore occur on different cluster isomers. The cluster is not a mere static scaffold for the surface chemistry. Instead, there is dynamic coupling between catalyst reactivity and the operando evolution of the cluster. Catalytic function can only be understood by holistically considering both types of processes on equal footing. |
6710c55312ff75c3a1ad3b8f | 4 | Motivated by these findings we return to scrutinize the tacit notion of a separation of time scales in surface catalysis. This is enabled by two recent methodological advances. On the one hand, the emergence of powerful machine-learned interatomic potentials (MLIPs) leads to a substantial reduction in the computational costs and allows for more extensive samplings of the PES without sacrificing the first-principles accuracy. On the other hand, inspired by seminal adaptive kinetic Monte Carlo approaches we have recently developed an automatic process exploration (APE) framework. This framework overcomes the human bias toward intuitive elementary processes and systematically identifies low-barrier processes at minimized PES sampling. |
6710c55312ff75c3a1ad3b8f | 5 | We apply this MLIP-APE approach to the context of oxidation catalysis at Pd(100) as a well-studied prototype for significant operando evolution in form of a restructuring to a ( √ 5 × √ 5)R27 • surface oxide. Specifically exploring the processes involved in the early stages of oxidation of a (110)-type step, we readily find numerous O-mediated restructuring processes with barriers significantly lower than 1 eV. With reported apparent activation barriers for oxidation reactions at Pd catalysts lying close to this value, this suggests that the dynamics of the restructuring process indeed occurs on similar time scales as the catalytic cycle. Fluxionality could therefore be an equally crucial feature at this "hard" inorganic surface. Many of the lowest barrier processes furthermore involve the collective, complex movement of multiple atoms in a non-intuitive manner. While the employed MLIP-APE framework enables efficient identification of these processes, the unearthed complexity challenges the development of appropriate microkinetic models that could assess the impact of this fluxionality on the catalytic function. |
6710c55312ff75c3a1ad3b8f | 6 | The APE framework systematically explores available elementary processes by launching dimer TS searches 49 from a given initial structure. It thereby relies on a fuzzy classification algorithm (DECAF ) to categorize all atoms of the structure into groups with essentially equivalent local atomic environment. In order to establish a maximally diverse set of elementary processes and suppress re-discovery of already identified TSs, the TS searches are then performed centered on one randomly chosen representative of each such equivalence group In simpler systems, APE naturally terminates if all groups in the list are processed and no new local atomic environments are found. In the present system, new equivalence groups in system structures of increasingly higher energy are continuously discovered. We therefore stop the exploration when the energies of these structures continue to be 1 eV higher in energy than the initial structure. As motivated above, we aim for a first unbiased assessment of available low-energy elementary processes for the initial oxidation of the Pd surface. Processes that already start from initial states higher in energy than the barrier used to identify facile processes are therefore beyond the scope of the present study, though we emphasize that a continued oxidation may well have to go through such thermodynamic bottlenecks. |
6710c55312ff75c3a1ad3b8f | 7 | The APE framework executes close to 45,000 successful dimer searches before this stop criterion is reached. This would impose excessive computational costs, if the associated PES evaluations were directly performed at the first-principles DFT level. We therefore combine the APE exploration with an iterative training of an MLIP, specifically a Gaussian Approximation Potential (GAP), as illustrated in Fig. . An initial GAP training set comprised a total of 520 structures, extending over (optimized and rattled) bulk unit-cells, clean and O-covered Pd(100) and Pd(410) surface structures, as well as the full ( √ 5 × √ 5)R27 • surface oxide, 42 cf. the SI for a full list and description. APE is first run on this GAP PES surrogate model until 500 processes have been obtained. We then perform a farthest point sampling on the average kernel of their smooth overlap of atomic position (SOAP) representation separately for the corresponding local minima (LM) and TS structures using the wfl package. The thus identified 50 maximally diverse LM and TS structures are subject to explicit DFT single-point calculations, and subsequently added to the GAP training set. After retraining, the improved GAP is used to explore the next 500 processes, and this iterative learning loop is continued until the APE stop criterion is reached. |
6710c55312ff75c3a1ad3b8f | 8 | Even though a root mean square error of 2.38 meV/atom over the final GAP training set reflects a quite high accuracy, we re-optimized the LMs and TSs for the processes shown in Figures and to ensure a consistent accuracy over all the presented processes. The geometry optimization of the LMs was performed with the conjugate gradient algorithm implemented in VASP, while the optimization to the TS was performed with dimer calculations, starting from the GAP-identified TS structure, along the direction between the initial state and TS structures. Both were performed until residual forces were below 0.01 eV/ Å. Differences in GAP-and DFT-optimized structures and energies are shown in the SI (Figures and). The DFT activation barriers ∆E proc stated below are extracted as the energy difference of the corresponding LM and TS structures, while the corresponding rate constants k proc are determined using the computed local Hessians and harmonic transition state theory. These and all DFT calculations performed for the GAP training employ the Vienna Ab-initio Simulation Package (VASP) version 6.3.2 and the PBE exchangecorrelation functional, with the D3 dispersion correction scheme with full details on the computational settings provided in the SI. |
6710c55312ff75c3a1ad3b8f | 9 | Automated Process Exploration The initial oxidation of Pd surfaces generally proceeds by O decoration of step edges, which then serves as nucleus for the transformation to surface oxides. We exemplarily explore the underlying elementary processes using the Pd(410) vicinal, which exposes four atom row wide Pd(100) terraces that are separated by (110) steps. Cast into a periodic boundary condition supercell in which the surface unit-cell is extended three times along the step edge and twice across the step edges to cover two subsequent terraces, cf. Fig. , this leads to a system size with enough structural flexibility to even allow for O-induced breaking of the step edge and step bunching (see below), but for which comparison with DFT calculations is still manageable. While the more open (110) steps seem more prone to oxidation anyway, we nevertheless note that attempts to explore processes for the alternative (111) steps, as e.g. exposed by (11N ) vicinals, suffered from finite-size effects in corresponding supercell setups. There APE identifies numerous unphysical low-barrier processes involving infinite row shifts along the step edge. Similar processes are also found in the (410) supercell, but due to the different step geometry their end states result in undercoordinated Pd atoms with a corresponding sizable energy penalty. |
6710c55312ff75c3a1ad3b8f | 10 | APE is started from six initial structures, containing from one to six O adsorbate atoms per surface unit-cell that are placed at the preferred adsorption sites first at the upper edge of one of the two steps contained in the surface unit-cell and once these sites are full at the fourfold hollow terrace sites next to these edge sites, cf. Fig. . With this range of stoichiometries, we can thus explore processes that emerge with increasing step decoration. |
6710c55312ff75c3a1ad3b8f | 11 | In total, the APE framework identifies 1,464 inequivalent TSs and a corresponding number of 2,928 unique back and forth elementary processes, excluding an additional 798 TSs that are less than 0.1 eV higher in energy than either the connected initial or final state and that thus correspond to extended shallow PES regions rather than relevant elementary processes. |
6710c55312ff75c3a1ad3b8f | 12 | Systematic process exploration for such a complex system as done here has therefore only become tractable through the massive efficiency gain brought about by the use of a MLIP DFT-surrogate. The surprisingly large number of identified inequivalent processes immediately hints at a level of complexity that stands in stark contrast to prevalent microkinetic models in surface catalysis that typically comprise only 1-2 dozen hand-selected elementary processes. It is not the sheer number that matters, however, but the fraction of low-barrier processes and their corresponding complexity. In other words, are they intuitive enough that a human researcher with sufficient domain knowledge would have considered them, too? With collective motion of multiple atoms typically associated with a certain degree of non-intuitive complexity, we therefore proceed to analyze the number of significantly displaced atoms in each elementary process, using a cutoff of 0.5 Å. Figure summarizes the results, additionally showing the distribution of barriers associated to the processes. As an important marker we thereby employ 1.0 eV. With reported apparent activation barriers of CO or methane oxidation somewhere around or above, we use this value as a rough threshold to identify low-barrier processes that happen on comparable time scales than those of the catalytic processes. |
6710c55312ff75c3a1ad3b8f | 13 | There is thus no separation of time scales into fast and slow processes. (iii) A significant fraction of the identified processes are collective, involving the significant displacement of more than two atoms. Only 43% (1261) of all processes seem simple in the sense that they displace only one or two atoms. (iv) While on average processes that move fewer atoms have lower barriers, the distribution for both few-and many-atom process barriers is wide, and there are multiple collective processes with barriers below the 1 eV threshold. This result also holds when evaluating the prefactors and plotting the distribution of hTST rate constants (Fig. in the SI). There is thus counter-intuitively no attempt frequency penalty for the movement of multiple atoms. It appears that the initial oxidation of the Pd(110) edge may thus indeed involve many fast non-intuitive collective processes. |
6710c55312ff75c3a1ad3b8f | 14 | Complementary insights can be obtained if we instead analyze the processes not by the total number of atoms that are significantly displaced, but by the species type. Figure shows the corresponding 10 such classes with most processes. This reveals that about 40% of all identified processes involve the exclusive motion of Pd atoms. Most of these involve the displacement of Pd 1 or Pd 2 units as one would expect in one-atom surface hopping or two-atom exchange diffusion type surface diffusion processes. However, there are also 148 unique collective processes involving Pd 3+ units. This reflects a high complexity of even just metal surface and edge diffusion well beyond the classical mechanisms -a realization that had also been emphasized in previous works. About 10% of all identified processes concern the mere diffusion of single O atoms (O 1 in Fig. ). Overall, the movement of O is much faster than Pd (see Fig. |
6710c55312ff75c3a1ad3b8f | 15 | The actual relevance of the individual identified processes and the concomitant dominant oxidation mechanism(s) can only be determined through a microkinetic analysis of their interplay within the overall network spanned by all elementary processes at given environmental conditions. For the present system, this would require kinetic Monte Carlo (kMC) simulations that explicitly resolve the atomic arrangement at the surface. Frankly though, present-day kMC simulation packages would not be able to cope with the above described complexity of the APE-identified set of elementary processes. This concerns not only the sheer number, but also that they largely involve atomic positions beyond those of the underlying Pd crystal and corresponding high-symmetry O adsorption sites. In other words, the processes readily reflect that surface oxidation will involve a significant degree of amorphization, whereas existing kMC simulation packages largely draw their efficiency from evaluating lattice models that restrict atomic positions to a specified crystalline lattice. This situation would be further aggravated when aiming to explicitly assess a possible operando fluxional behavior in oxidation catalysis, as this would require to also APE-identify and include the likely at least equally massive number of additional elementary processes involving the reducing reactants and the evolving surface reaction intermediates. |
6710c55312ff75c3a1ad3b8f | 16 | While this calls for substantial methodological advances to boost the microkinetic model- ing capabilities, we can already gain first qualitative insight by analyzing selected sequences of the identified elementary processes. Figure shows a corresponding six-step pathway for the system containing only two O atoms per surface unit-cell and thus characteristic for the very initial oxidation of the step edge. Starting from the smooth (110)-type step of the (410) vicinal, a first lowest-barrier process (0.33 eV) starts to pull out one Pd step atom. This is a typical example of facile O-induced restructuring as the cost of the lower coordination of the Pd atom is mitigated by one of the attached O atoms moving from its original three-fold coordinated upper step edge site to an equally three-fold coordinated anchor site further away from the step. The second still low-barrier process (0.63 eV) continues this pulling out with the peripheral O atom moving to an equivalent site even further out. Of special note is the fact that this resulting state (number 3 in Fig. ) is 0.16 eV more stable than the starting state. In contrast, the analogous rupturing of a Pd atom out of the (100) step without coordinating O atoms is uphill by 0.32 eV and involves a barrier of 0.84 eV. There is no O-free intermediate that corresponds with state 2 (cf. Fig. ). Without O, this structure is not stabilized enough to persist. From system state 3, cf. The next three processes of the pathway equally allow to make contact with previous reports, namely step bunching on vicinal Pd surfaces as observed experimentally under oxidizing conditions by scanning tunneling microscopy and under conditions of CO oxidation by surface X-ray diffraction. Two of these three processes have somewhat higher barriers up to 1.43 eV, but by involving multiple atoms in a collective row-shift motion they efficiently allow to transport Pd atoms from one edge to another, inducing the growth of the terrace with O-decorated step edge at the expense of a retracting neighboring terrace (cf. |
6710c55312ff75c3a1ad3b8f | 17 | Fig. ). This mechanism is again not dissimilar to the ones discussed by Lim et al., but results here from simple visual inspection of the automatically determined processes. We also note that such efficient step bunching would provide a mechanism toward sufficiently wide terraces that have been discussed as crucial for surface oxide formation on highly stepped surfaces. Importantly, there is no strong thermodynamic driving force over the entire pathway. |
6710c55312ff75c3a1ad3b8f | 18 | Through it, the system could thus in principle dynamically interconvert between the involved system states. Given the low barriers of the processes connecting in particular system states 1 through 4, this interconversion between the corresponding smooth and zig-zag step edge structures would be quite fast (see Fig. for the corresponding rates). Such a dynamic O-induced step restructuring would be consistent with a fuzzy resolution of near-step regions in scanning tunneling and transmission electron microscopy at experimental scan rates. Moreover, the barriers are in fact so low that this pathway suggests a potential fluxional behavior of the active surface. The reactivity of the ( ) and (111) type edges is quite different and if these edge types interconvert readily during the catalytic cycle, this operando evolution could be crucial to understand the catalytic performance of the working surface. Very similar conclusions are obtained from another restructuring pathway, now at higher O loading of six O atoms per surface unit-cell. As summarized in Fig. , two processes with barriers below 0.4 eV bring the system efficiently from a step-decorated state into an energetically much lower superbasin. Within this superbasin, most processes are below 0.9 eV, with barriers exiting state 5 at 1.2 eV. These low-energy processes bring the system readily between system states 3 through 6, e.g. at 600 K at rates faster than ∼ 10 3 s -1 (Fig. ). At time scales of the oxidative catalysis, the system could thus dynamically evolve over the corresponding quite different edge structures with their correspondingly different reactivities. States 3 and 4 exhibit e.g. large Pd x O y chains forming a ring that roughens the step edge and pushes it forward, leaving vacancies in the terrace behind the edge. In contrast, state 6 involves fins of Pd 3 O 4 extending from the step edge, without any vacancies within the Pd x O y structures. Time averaged, there is thus a mix of two-and three-fold O coordination in the restructured step edges. This would match with experimental X-ray photoelectron spectroscopy data, while the fast interconversion would again be consistent with the fuzzy images of step edges in scanning microscopies. Notwithstanding, we repeat that only a comprehensive microkinetic analysis could assess the true relevance of this pathway under reaction conditions, and neither can we be sure that the present APE search starting from the step-decorated structures identified all relevant elementary processes. Nevertheless, the framework does provide a more systematic access to the underlying elementary processes and it is baffling that only three non-intuitive and thus hitherto not considered collective processes provide an efficient pathway to interconvert between such radically different step edge structures. |
6710c55312ff75c3a1ad3b8f | 19 | Present-day predictive-quality microkinetic modeling cannot address a possible dynamic impact of the operando evolution of working catalyst surfaces similar to the fluxionality discussed in nanocluster catalysis. With the novel combination of actively learned fast MLIP-surrogate energetics and automated process exploration we can overcome the first methodological reason for this limitation, namely the inability to systematically identify the elementary processes involved in such an evolution. In the application to the initial oxidation of a (110) Pd-step edge this approach readily generated process lists orders of magnitude larger than those commonly employed in state-of-the-art microkinetic modeling. Beyond just the sheer number of identified processes, many of them are also non-intuitive, involving the collective motion of multiple atoms. This level of complexity in detail (revealed here by to the diversity-driven approach inherent to APE) has barely been recognized before. |
6710c55312ff75c3a1ad3b8f | 20 | Simple inspection of selected pathways immediately rationalizes several important experimental observations like step microfaceting, step bunching or step amorphization under oxidative conditions in terms of only a few low-barrier collective processes. The facile nature of these step restructuring processes indicates that some degree of fluxionality could indeed also play a role at this nominally "hard" inorganic catalyst, with dynamic evolution of the catalyst surface able to take place on the same time scale as the catalytic reaction. As a result, different steps of the catalytic cycle may take place at different sites offered by the continuously evolving surface. Scrutiny of these propositions requires future work to also overcome the second methodological hurdle toward a comprehensive microkinetic modeling, namely to be able cope with a correspondingly large number of processes beyond the confines of a trivial crystalline lattice. En route to a fully automated generation of an (off-lattice) microkinetic model from a given list of elementary processes, a pragmatic intermediate step could be systematic coarse-graining approaches that classify the identified processes into a smaller number of qualitatively distinct process classes that can then be captured within manually established approximate models. |
674682dc7be152b1d0fb49f7 | 0 | ABSTRACT: Desorption Electrospray Ionization Mass Spectrometry Imaging (DESI-MSI) is a powerful technique for molecular analysis of surfaces; however, its application of single cell studies has not been previously published. In the current work, a commercial DESI setup (DESI XS) coupled to a mass spectrometer was used to analyze cultured mammalian cells attached onto glass coverslips. A series of experiments have been performed to obtain optimized experimental conditions, and MS images of metabolites with the single-cell level resolution were obtained using 10 µm x 10 µm pixel size. This established method can be readily adopted to extend the power of commercial DESI-MSI techniques in metabolomics studies requiring cellular resolution. |
674682dc7be152b1d0fb49f7 | 1 | Metabolomics is a critical field of study for understanding biochemical processes and disease mechanisms at a molecular level. This research area holds potential for uncovering biomarkers for disease , elucidating metabolic pathways , and advancing medicinal research. Despite significant progress in the field, a profound understanding of cellular heterogeneity and the spatially resolved cell-cell interactions remains a formidable challenge in traditional metabolomics studies. Single-cell analysis is essential for various reasons. In complex tissues, cellular heterogeneity plays a crucial role in functions and responses to environmental stimuli. Tumor microenvironments, for example, consist of a diverse population of cells with distinct metabolic states, contributing to cancer progression and drug resistance. Understanding these variations at the single-cell level can reveal critical information about disease progression and therapeutic targets that are otherwise obscured in bulk analyses. Single-cell metabolomics addresses this challenge by providing insights into the unique metabolic profiles of individual cells, which can differ significantly even within the same tissue or culture. Single-cell mass spectrometry (SCMS) is gaining popularity as a powerful tool for single-cell metabolomics studies due to its label-free approach and ability to sensitively detect a broad range of analytes. Unlike traditional metabolomics methods, such as liquid chromatography (LC)/MS, based on bulk sample analysis, SCMS enables direct studies of biological processes of individual cells in heterogenous populations . |
674682dc7be152b1d0fb49f7 | 2 | Numerous SCMS techniques have been developed to perform single cell metabolomics experiments under vacuum or ambient conditions. Among them, a group of methods are based on mass spectrometry imaging (MSI) techniques. MSI has emerged as a powerful technique for spatial metabolomic analysis, offering the ability to map the spatially-resolved molecular information within biological samples. Unlike traditional mass spectrometry (MS) methods, which provide average measurements of metabolites from homogenized samples, MSI allows for the visualization of spatial distributions of metabolites and their abundances directly within tissue sections or individual cells. This spatial information is invaluable for understanding the microenvironmental context of metabolic processes, revealing metabolic gradients, and identifying cellular subpopulations with distinct metabolic states. |
674682dc7be152b1d0fb49f7 | 3 | Although vacuum-based MSI techniques, including commercialized Matrix-Assisted Laser Desorption/Ionization (MALDI-MSI) and Secondary Ion Mass Spectrometry (SIMS-MSI) , render ultra-high spatial resolution in measurements, challenges (e.g., nontrivial sample preparation and complex mass spectra at low mass) need to be overcome for successful applications. In contrast, ambient-based methods generally provide relatively lower spatial resolutions, but they require little or no sample preparation and provide simpler mass spectra (e.g., fewer interfering ions and fragments) . Particularly, some of these methods, such as nanospray desorption electrospray ionization (nano-DESI) MSI and Single-probe MSI , have successfully achieved high resolutions at the single-cell level. Unfortunately, both nano-DESI and Singleprobe MSI methods are not commercially available, limiting their broad applications. |
674682dc7be152b1d0fb49f7 | 4 | Desorption Electrospray Ionization Mass Spectrometry Imaging (DESI-MSI), an ambient, commercially available MSI technique , presents a promising alternative for single-cell metabolomics. DESI-MS involves charged solvent droplets impacting a sample's surface to desorb molecules and generate analytes' ions for direct MS detection. DESI-MSI combines the benefits of minimal sample preparation with the capability to ionize a wide range of metabolites directly from the sample surface with little or no sample preparation. In addition, DESI does not require the application of matrix molecules, minimizing the interfering ions and simplifying data interpretation. As a commercialized technique, DESI offers a more robust and user-friendly setup, facilitating broader application. |
674682dc7be152b1d0fb49f7 | 5 | Despite these advantages, DESI-MSI also faces several limitations when applied to single-cell analysis. One of the primary challenges is achieving sufficient spatial resolution to differentiate individual cells accurately. The typical spatial resolution of DESI-MSI is limited to around 35 µm, which can be insufficient for single-cell imaging, where cells may be as small as 10-20 µm. For a given analytical technique, sensitivity and resolution (including the spatial resolution) are inversely proportional to each other, i.e., increasing the MSI spatial resolution reduces its sensitivity due to reduced amounts of analytes. An effective approach to increasing the sensitivity of high spatial resolution MSI technique is to increase its efficiencies of molecular ionization and ion transfer. The efficiencies of molecular ionization and ion transfer in DESI-MSI can be affected by multiple factors (e.g., the sample's surface properties, the spray angle and distance, the voltage applied to the spray, the solvent composition, and the flowrate of the solvent) that can lead to variability in signal intensity and complicate quantitative analyses. Because analyte ions need to be transferred from the surface into the mass spectrometer through an ion transfer capillary, ion signal loss can occur due to several factors (e.g., ion neutralization, ion attachment to surfaces within the capillary, and incomplete transfer of ions ) that can reduce the overall ions' intensities. Reducing ion loss during transfer is regarded as an effective strategy. A previous study by Zickuhr et al. (2024) showed that heating the ion transfer capillary from 150°C to 450°C resulted in up to a 1.8-fold increase in overall signal intensity. We aimed to improve the detection sensitivity of the DESI experiments by optimizing the experimental conditions, including heating the ion transfer capillary, to acquire MS images of single cells. |
674682dc7be152b1d0fb49f7 | 6 | In this current work, we performed MSI studies using a DESI XS source (Waters Inc., Milford, MA, USA) integrated with a Waters Synapt G2-Si mass spectrometer and a Waters ACQUITY UPLC M-Class system (Waters Inc., Milford, MA, USA). Without conducting complex instrumentation modification, we achieved single-cell resolution MSI by optimizing experimental parameters such as the temperature of the ion transfer capillary, sprayer-surface distance, voltage, and solvent flowrate. |
674682dc7be152b1d0fb49f7 | 7 | Prior to MSI single cell studies, we optimized experimental conditions using dried cell lysate on PTFE coated glass slides (ref: 041-MICRO-44, Scientific Device Laboratory, Des Plaines, IL, USA). First, we prepared the surface coated with dried cell lysates. To do this, we cultured OVCAR-8 human ovarian cancer cells, prepared cell lysate, and coated multiple spots on a glass slide with 10 µL of cell lysate on each spot and dried in air (Figure ). Second, we performed systematic experiments to optimize the temperature of the ion transfer capillary. Specifically, we wrapped the heating wire (22 gauge, Kanthal, Bethel, CT, USA), covered by a fiberglass mesh sleeve, around the ion transfer capillary (Figure ). This assembly was connected to a thermocouple (to monitor the temperature) and a 144W universal power supply (model: LGY-363000, Dong Guan Shi He Yu Tech, Dongguan, China), which allowed us to apply varying voltages to reach and maintain specific temperatures. We adjusted the voltage applied from the power supply to achieve a series of targeting temperatures: room temperature, 250, 275, 300, 325, 350, 375, 400, and 425°C. Third, for each temperature, the DESI XS stage was manually repositioned to manipulate the distance between the sprayer and spot coated with dried cell lysate on the slide. In general, a short distance can improve ion intensities, whereas an excessively short distance may cause the sprayer to scratch the sample. Last, we optimized the flowrate of solvent (methanol 95%/water 5% + formic acid 1%) delivered by the LC pump system (no column). Optimized experiment parameters include an ion transfer capillary temperature of 375°C, a distance of 1.2 mm between sprayer and surface, and a flowrate of 0.5 mL/min. Other parameters include scan rate of 20 µm/s, mass range of 100-1200 m/z, TOF mode, and positive ion mode. |
674682dc7be152b1d0fb49f7 | 8 | The Total Ion Current (TIC) graph (Figure ) illustrates that high TICs with excellent stabilities were achieved at 375°C. However, the trend of TIC does not represent the ion intensities of individual cellular species. To further evaluate the influence of temperature of the ion transfer capillary on MSI measurements of cellular species, experiments were conducted for m/z 786.2191, an intense ion commonly observed from cells and tissues. Our results indicate that 400°C resulted in significantly improved ion intensities (Figure .). Upon acquiring the optimized experimental conditions, we conducted DESI-MSI studies to achieve single-cell resolution (Figure ). OVCAR-8 cell line was used as the model system. Cells were cultured and attached onto glass coverslips with 50 x 50 µm grid (cat.no:10817, ibidi, Gräfelfing, Germany). To minimize the influence of matrix effect (e.g., induced by salts in the cell culture medium) on ionization efficiency, cells on the coverslips were washed twice with isotonic ammonium formate solution (0.144 M) prior to analysis. A pixel size of 10 µm x 10 µm was used to achieve high-quality MS images with spatially resolved single cells. Figure shows a photo of attached cells that were taken using a brightfield microscope (cat.no: 12575252, Fisher Scientific, Waltham, MA, USA) at 20x magnification. MS images obtained from DESI-MSI with 10 µm x 10 µm pixel size that show the abundances of select ions, which were tentatively annotated (mass error < 8 ppm) based on searching their measured m/z values on LipidMaps (). Darker areas indicate lower abundances, whereas brighter areas indicate higher abundance (Figure ). Excellent correlations of single cells' locations can be observed by comparing the microscopy photo and MS images. Although a finer pixel size of 5 µm x 5 µm has been tested, we were unable to obtain improved quality of MS images, likely due to inadequate signal intensities from smaller sampling areas using our current mass spectrometer. In summary, we have successfully utilized the commercial DESI-MSI technique to obtain MS images of cultured mammalian cells at the single-cell resolution. Experiments were performed using Waters DESI XS course coupled to a Waters Synapt G2Si mass spectrometer to analyze OVCAR-8 cells attached on glass coverslips. A heated ion transfer capillary was implemented into the DESI XS source. Comprehensive experiments were carried out to optimize experimental conditions, including the temperature of ion transfer capillary, distance between sprayer and surface, and solvent flowrate. This established method allowed us to observe the metabolites and their spatial distribution within individual cells with minimal sample preparation. Our method can be readily adopted by other researchers who use the similar instrument with little or no modification. Other analytical methods, such as ion mobility and tandem MS (MS/MS), can be implemented to provide more molecular information. The DESI-MSI technique with the singlecell resolution holds significant potential for advancing our understanding of cellular heterogeneity and the metabolic diversity within individual cells, providing valuable insights into dynamic biological processes, disease mechanisms, and potential therapeutic targets. It is worth noting that our current MSI results using smaller pixel sizes were unsatisfying, very likely due to the inadequate detection sensitivity of the mass spectrometer used in this work. Utilizing later generations of mass spectrometers compatible with the newer version of DESI XS, which is implemented with heated ion transfer capillary, can likely improve its performance for high resolution MSI studies. In addition, implementing new developments in DESI-based methods, such as using He or H2 to replace N2 as the nebulization gas and combining ammonia additive with post-photoionization , for improved ion intensities can potentially promote DESI-MSI studies of single cells. |
64dded8769bfb8925a11f9bf | 0 | Our ability to selectively activate and functionalize C-H bonds in organic molecules is fundamental to countless chemical processes. Despite notable advances in this field, strategies for the selective functionalization of sp C-H bonds of alkenes are underdeveloped. This limitation can be traced back to fundamental selectivity issues that emerge on reaction of alkenes with single-site transition metal complexes. π-Coordination of the alkene to the metal is often kinetically accessible and nonreversible (Figure ). The dominance of this pathway can effectively inhibit available mechanisms for vinylic C-H activation, including oxidative addition. The selectivity contrasts the rich chemistry of aromatic sp 2 C-H or aliphatic sp C-H bonds where substrate binding (π-coordination or σ-complex formation) is typically reversible and a pre-requisite for C-H bond breaking by oxidative addition. For example, Bergman and coworkers have studied the reaction of ethylene with the 16-electron reaction intermediate [IrCp*(PMe)3]. They found that π-complexation of the alkene is thermodynamically favored with respect to oxidative addition of vinylic sp C-H bonds. Moreover, the π-complex was found to not be an intermediate in the lowest energy C-H activation process. Computational studies support the conclusions and suggest that π-coordination and C-H activation of the alkene are separate and competitive pathways. We recently reported a well-defined Fe-Al complex (1) that is capable of selectively breaking the sp and sp C-H bond of pyridine substrates as well as acetonitrile. Herein we present the C-H activation in the vinylic position of styrenes using the same Fe-Al system. These reactions are highly selective resulting in a rare (E)-β-metalation of the alkenes. In contrast to single-site systems, alkene binding appears to initiate the C-H activation and is essential for the reaction to take place. An unusual reaction pathway in which a (2+2) cycloaddition intermediate is directly converted into the hydrido vinyl product is proposed. This new mechanism results in the net oxidative addition of an alkenyl sp 2 C-H bond across the two metal centers and opens new possibilities for selective alkene functionalization by C-H activation using a bimetallic approach. |
64dded8769bfb8925a11f9bf | 1 | Non-reversible (2+2) alkyne binding. Addition of 1 equiv. of a terminal alkyne RC≡CH (R = Ph, SiMe3 or n-Bu, 2-Py) to a solution of 1 in C6D6 at room temperature resulted in an immediate colour change from dark to bright orange, in each case leading to the quantitative formation of the cycloaddition products 2a-d (Figure ). These reactions appear to be non-reversible. 2a-d were all isolated in yields of >95 % and are stable in both solution and in the solid state. 2b was characterised by single crystal X-ray diffraction (Figure ). 2a-d all give rise to very similar and characteristic NMR signals for the coordination of the alkyne. For example, the 31 P{ 1 H} NMR spectrum of 2a exhibits a mutually coupled spin system comprising a doublet at δP = 29.0 ppm (2P) and a triplet resonance at 19.9 ppm (1P) being consistent with the chemical non-equivalence of the axial and equatorial phosphine ligands. In the 1 H NMR spectrum, the bridging hydrides appear as broadened virtual triplet at δH = -14.71 ppm and a doublet of triplets at δH = 10.02 ppm ( 3 JHP = 13.4 and 5.6 Hz) can be found for the ArC≡CH proton of the coordinated alkyne. No evidence for C-H activation of the sp C-H bond was found, despite the acidity of terminal alkynes. |
64dded8769bfb8925a11f9bf | 2 | Reversible (2+2) alkene binding. Styrene substrates were found to bind reversibly to 1 (Figure ). Stepwise addition of excess styrene (10 -40 equiv.) to 1 in C6D6 at room temperature led to the gradual appearance of a new species 3a suggesting an equilibrium established immediately after each addition. A Van't Hoff analysis of the reaction of 1 with excess styrene (21.6 equiv.) was conducted in toluene-d8 over a temperature range of 248-298 K. The formation of 3a was found to be slightly exergonic; ΔH°298 = -12.5 kcal mol -1 and ΔG°298 = -0.4 kcal mol -1 . The formation of 3c appeared to be more energetically favourable. Addition of 4-(trifluoromethyl)styrene (1.5 equiv.) to a solution of 1 in toluene-d8 at -35 °C resulted in an immediate colour change from dark red to bright orange and the NMR spectra recorded at the same temperature revealed that the equilibrium between 1 and 3c has been completely shifted to the product side. 3c could be crystallized from n-pentane at -35 °C and the solid-state structure was analysed by single crystal X-ray diffraction (Figure ). |
64dded8769bfb8925a11f9bf | 3 | In the 1 H NMR spectrum of 3c recorded at -35 °C two broad apparent triplets at δH = -14.92 and -15.93 ppm can be found for the bridging Fe-(μ-H)2-Al hydrides as well as another three broad signals at δH = 2.74, 1.03 and 0.64 ppm for the ArCH=CH2 protons of the coordinated alkene group. In the 31 P{ 1 H} NMR spectrum, the three PMe3 ligands of 3c appear as a well resolved ABX spin system; the AB part centred at δP = 36.5 ppm (JAB = 41.5 Hz) and the X part at 22.2 ppm. Alkene binding, while reversible, only leads to kinetic products in these reactions, over longer time periods activation of the vinylic C-H bonds was observed. |
64dded8769bfb8925a11f9bf | 4 | Vinylic C-H activation. Over the course of 14 days, the room temperature reaction of 1 with styrene (1 equiv.) in C6D6 afforded the vinylic C-H activation product 4a in 78 % NMR yield (Figure ). 4a was isolated in pure form and crystals suitable for X-ray diffraction have been grown confirming the (E)-β-alumination of styrene (Figure ). The new species exhibits a broadened hydride resonance at δH = -15.58 ppm in the 1 H NMR spectrum integrating to 3H and a singlet at δp = 29.0 ppm in the 31 P{ 1 H} NMR spectrum being characteristic for a (PMe3)3Fe-(μ-H)3-Al motif that resulted from the C-H activation reaction. The PhCH=CH protons resonate at δH = 7.86 and δH = 7.34 ppm respectively showing a large coupling constant of 3 JHH = 19.9 Hz diagnostic for an (E)-configuration of the C=C double bond. |
64dded8769bfb8925a11f9bf | 5 | Styrene derivatives containing electron-withdrawing substituents reacted much more quickly than those with electron-donating groups. For example, upon addition of 1 equiv. of 4-(trifluoromethyl)styrene 4c was formed nearly quantitatively within 3 h. In the case of 4a and 4b an excess of the respective styrenes (10 equiv.) was necessary to obtain reasonable reaction rates and full conversion of 1 at room temperature. The formation of the products in these cases was found to follow first order kinetics, with t½ ≈ 24 h (4a) and t½ ≈ 71 h (4b), respectively. The highest reaction rates at room temperature were observed in the reaction of 1 with 2vinylpyridine which affords 4d almost instantly when carried out in the presence of catalytic amounts (1-2 mol %) of MgBr2. MgBr2 appears to act as a Lewis acid catalyst preventing coordination of the pyridine nitrogen to Al and activating the substrate for the C-H activation reaction (vide infra). |
64dded8769bfb8925a11f9bf | 6 | Non-reversible (2+4) addition. In absence of the Lewis acid additive 2-vinylpyridine forms a (2+4) cycloaddition product with 1 (Figure ). Addition of 2-vinylpyridine to a solution of 1 in tolune-d8 at -35 °C resulted in the formation of 5d in ca. 85 % NMR yield alongside with 4d as minor side product. The 1 H NMR spectrum of 5d recorded at -40 °C shows a sharp triplet resonance at δH = 3.81 ppm diagnostic for the α-CH group of the (2+4) bound substrate. Warming the reaction solution to room temperature resulted in the slow decomposition of 5d (see Supporting Information) and did not convert into 2d. These results suggest that the (2+4) addition represents a competitive pathway preventing the C-H activation reaction. The analogue product 5e was obtained from the reaction of 1 with methyl acrylate at room temperature, with the (E)-C-H activation product 4e also being formed in a 15 % NMR yield. In the case of 5e, crystals suitable for X-ray diffraction could be grown confirming the structure of the (2+4) cycloaddition product with the β-CH2 attached to Fe and the oxygen of the former carbonyl group bound to Al (Figure ). |
64dded8769bfb8925a11f9bf | 7 | Alkene binding to 1 was further investigated by ETS-NOCV calculations (see Figure ). Donation of electron density from the former Fe-Al bond into the π* orbital of the substrate (3a: Δρ1 = -420.7 kcal mol -1 ) accounts for >90 % of the total orbital stabilization energy (3a: ΔEorb = -421.0 kcal mol -1 ). NBO analysis identified a σ-bond between iron and the β-carbon of the substrate whereas bonding of Al to the α-carbon is defined as donor-acceptor interaction between a Ccentered lone pair and empty s/p orbitals at Al possessing a partial ionic character. This is underpinned by the NPA charges revealing that the Al-C bond is more polarized (3a: Al +1.80, C -0.88) in comparison to the Fe-C bond (3a: Fe -0.52, C -0.80). |
64dded8769bfb8925a11f9bf | 8 | DFT calculations were also undertaken on the mechanism of the vinylic C-H activation of styrenes. A low energy pathway was identified involving the direct intramolecular C-H activation of the bound styrene in 3a (Figure ). The mechanism is initiated through a concerted but asynchronous addition of the C=C double bond to 1. This first step has a low activation barrier of ΔG ‡ 298K = 15.8 kcal/mol (TS-1a) and is moderately exergonic by -2. TS-3a appears to be a highly asynchronous transition state near the boundary between a concerted and a stepwise mechanism (vide infra). The low imaginary frequency (-92.3 cm -1 ) of TS-3a refers to the reorientation of the bound alkene in which the α-C is moving away from the Al center while the β-C is transferred from Fe to Al. There is no stationary point for the actual cleavage of the C-H bond which occurs along the intrinsic reaction coordinate en route to INT-4a (see Intrinsic Reaction Coordinate in Figure ). The (E)-stereospecific nature of the C-H activation is a direct consequence of the orientation of the substituents in the transition state TS-3a (vide infra). |
64dded8769bfb8925a11f9bf | 9 | To get better insight into the nature of TS-3a an analysis of the key localized molecular orbitals (LMOs) along the intrinsic reaction coordinate (IRC) was carried out. LMOs were calculated following the Pipek-Mezey criterion and a procedure described by Vidossich and Lledóss was used to generate centroids of these LMOs (CLMOs). The CLMOs were used to follow the bond rearrangements around TS-3a (Figure ). For the sake of clarity, depiction of the overall process was separated in a pre-and post-TS stage. From INT-3a to TS-3a (pre-TS), migration of the β-C of the bound alkene goes along with an electron transfer from the Fe-C bond to Al while the electrons from the Al-C are shifted towards the α-C and get delocalized through participation of the arene π-system (vide infra). In the post-TS stage, the most significant rearrangement of the electronic structure occurs during the C-H bond cleavage. Electrons from the C-H σ-orbital are shifted towards Fe forming the Fe-H bond. This process is consistent with a hydride transfer and distinct from the reaction of 1 with pyridines and CH3CN 17 which were found to proceed via a proton transfer (reductive deprotonation). Simultaneously, the delocalized lone pair at the α-C is shifted back towards the β-C to re-establish the π-system of the C=C double bond in the vinyl ligand. The proposed mechanism is supported by deuterium labelling studies. A comparison of the reaction rates of 1 with styrene and styrene-d8 at 323 K from two separate experiments resulted in a kinetic isotope effect (KIE) of 2.64 ± 0.04. The overall KIE is likely affected by an additional equilibrium isotope effect caused by the reversible alkene binding which, however, appeared to be relatively small at room temperature (EIE = 1.06 at 298 K). This contrasts the ortho C-H activation of pyridine for which an unusually large kH/kD value of 14.0 ± 0.2 at 298 K was obtained likely caused by a quantum tunneling and thus diagnostic for a proton transfer reaction. 15 Table . Substituent effect on the activation energies for INT-3 to INT-4 in the reaction of styrenes with 1. |
64dded8769bfb8925a11f9bf | 10 | The C-H activation step was calculated for the entire series of styrene substrates (Table ). The obtained activation barriers correlate well with the Hammett parameters 29 of the respective substituents and reflect the relative reactivities of these substrates observed experimentally. Across the series, the imaginary frequencies of TS-3a-c are getting lower with more positive Hammett parameters indicating flattening of the potential energy surface around the transition state. |
64dded8769bfb8925a11f9bf | 11 | While the IRC for 4-methoxystyrene is almost symmetrical around the transition state it forms a plateau for 4-(trifluoromethyl)styrene flanked by a steep decay marking the onset of the C-H bond cleavage. Computationally, the series was expanded to 4-nitrostyrene due to the large positive Hammett parameter of the NO2 substituent (σ -= +1.27). In this extreme case, complete deconvolution of the rearrangement and C-H activation of the bound substrate into two separate transition states was obtained (TS-3e, ΔG ‡ = 9.4 kcal/mol; TS-3e', ΔG ‡ = 2.1 kcal/mol; see Figure in the Supporting Information). This trend as well as the calculated activation barriers can be traced back to the resonance stabilizing effect of the para-substituent. A comparison of TS-3 down the series shows that the C(α)-C(Ar 1 ) bond distances are getting shorter indicating an increasing double bond character (TS-3b: 1.412 Å, TS-3e: 1.384 Å). At the same time, NBO analysis of TS-3 reveals the negative charge accumulation at the α-C decreases in the same order (TS-3b: -0.59, TS-3e: -0.48). This ability to stabilize the negative charge at the α-C appears to be the key factor for the reactivity. To some extent, these findings resemble recent studies on the transition from a stepwise to a concerted behavior of SNAr reactions. The same effect was observed for the MgBr2-promoted reaction of 2-vinylpyridine with 1 (Figure ). Binding of the Lewis acid to the pyridine nitrogen allows for an extreme resonance structure resulting in a stepwise process and an even lower barrier than for 4-nitrostyrene. The emergent intermediate (INT-3d', ΔG = -3.7 kcal/mol) is just slightly lower in energy than the adjacent barriers ((TS-3d, ΔG = 2.5 kcal/mol; TS-3d', ΔG = 0.4 kcal/mol). |
64dded8769bfb8925a11f9bf | 12 | NBO analysis provided further insight into the nature of this intermediate. In INT-3d', the former π-system of the alkene remains broken resulting in two separated p-orbitals. The β-carbon appears to be sp 3 hybridized carrying a negative partial charge (-1.10) stabilized through coordination to aluminum. The α-C appears to be sp 2 hybridized forming a resonance structure with the adjacent pyridyl group and thus carries a much lower negative partial charge (-0.46). |
64dded8769bfb8925a11f9bf | 13 | Binding of the substrate in INT-3e' is further stabilized through an agostic interaction between one of the β-C-H bonds and the iron center. In TS-3e' this C-H bond is cleaved (vide supra). With the lone pair on the α-C the alkene π-system is re-established forming a C=C double bond which, as a consequence of the anti-periplanar conformation of the C-C bond in INT-3e', adopts the final (E)configuration. |
64dded8769bfb8925a11f9bf | 14 | For monometallic systems the direct oxidative addition of the sp 2 C-H bond in an unactivated alkene is hard to achieve. This is well documented and can be rationalized by the competing πcoordination of the alkene over the formation of a weakly bound σ-C-H complex vital for the bond cleavage at the transition metal centre. By contrast, our findings suggest that alkene binding is essential for the C-H activation to take place in the present bimetallic system. The (2+2) cycloaddition of the styrene substrates is weak and appears to initiate a low energy pathway for C-H activation. We identified an unusual transition state connecting the alkene complex with the final hydrido vinyl product. This key step proceeds through a highly asynchronous transition state near the boundary between a concerted/stepwise mechanism influenced by the resonance stabilization ability of the aryl substituent. Moreover, the geometry of the transition state results in the selective metalation of the (E)-β-C-H bond of the substrate. This concept might stimulate future developments towards catalytic C-H functionalization reactions. |
6719d97383f22e4214f8b821 | 0 | Solvation free energy is one of the key quantities in chemistry and biology as most of the molecular phenomena occur in different solvents. Water being the most versatile solvent, determination or calculation of hydration free energy (HFE) is the most important step in understanding any complex process. The calculation of HFE is typically done using either a quantum mechanical description of the solute in a dielectric continuum or a classical description of the solute where water is treated either with explicit models or implicit models (which are mostly dielectric continuum). The classical forcefields are usually used for macromolecules, and for small molecules interacting with macromolecules, classical forcefields are used for compatibility. Henceforth, all discussions on the calculation of hydration free energy will be with the classical models. Although the physics-based methods are well developed, there are outstanding issues in getting accurate solvation free energy and it has been an active area of new developments and improvements. The most rigorous calculations are alchemical methods like thermodynamic integration and free energy perturbation. However, these calculations are time-consuming and generally speaking, are not suitable for a large set of molecules. |
6719d97383f22e4214f8b821 | 1 | In the PB approach, PB eq. (or Poisson eq. in the absence of any salt concentration) is solved by defining an interior dielectric constant for the solute and an external dielectric constant for the solvent. The Generalized Born approach also defines internal and external dielectric constants; however, here no electrostatic equation is solved, rather an expression, obtained from the generalization of the Born equation for a single ion, is evaluated. In both PB and GB approaches, the molecular surface area is calculated and in GB, the socalled Born radii is calculated, which can be time-consuming. Also, there are some inherent limitations for continuum models as they neglect the molecular nature of water. There have been several attempts to build cluster-continuum models. From an entirely different perspective, several machine learning (ML) models are developed, in the last couple of years, to predict solvation free energy using experimental data in the FreeSolv database. The faster speed of ML models compared to physics-based models is advantageous and can be used for large databases of small molecules used in drug discoveries. However, ML models often suffer from a lack of interpretability and the reasons why they work (or do not work) are often not clear. There have been attempts to define descriptors having clear physical meaning. For instance, Zhang et al. used electron density (obtained from quantum mechanical calculations) based descriptors. In some of the other representative works, Alibakhshi et al. have combined ML models with PCM model to predict solvation free energy in different solvents using the components of the PCM calculations as the features of the ML model, Pattnaik et al. have developed an ML model to predict relative solvation free energy in forty-one solvents. Vyboishchikov has developed a few NN-based models based on the GB model of solvation. The effective Born radii, charges are used as the features in these models. Machine learning has also been used to predict the HFE obtained from MD simulation-based methods. In the current work, our motivation is to use a minimal number of physically interpretable and simple descriptors for predicting HFE. Starting from the GB expression and with an approximate analytical calculation of Born radii, we evaluate the HFE (after adding five more descriptors) and got accuracy almost as good as the paper by Zhang et al. for the FreeSolv dataset. The power of our method is that it is completely physics-based, hence fully interpretable. It has only six descriptors, an electrostatics term (GB term summed with Coulomb electrostatic), polar surface area, number of donor and acceptor atoms, logP , and the number of rotatable bonds. We have used four different models, namely Random Forest (RF), Extreme Gradient Boosting (XGBoost), Gradient Boosting (GradBoost), and Light Gradient Boosting Machine (LightGBM). Our best result is a mean absolute error (MAE) of 0.74 kcal/mol comparable to the work of Zhang et al. This method can be used for large datasets used in drug designing and the reason for specific HFE values can be understood clearly as opposed to most of the ML models. |
6719d97383f22e4214f8b821 | 2 | The experimental hydration free energy database, FreeSolv, prepared by Mobley et al., has been widely used and benchmarked by various physical solvation models as well as machine learning (ML) and deep learning models. The FreeSolv database has 643 small organic molecules with their experimental HFE and their SMILES (simplified molecular-input lineentry system). The database also includes the calculated HFE, enthalpy, and entropy data from explicit molecular dynamics simulations. These calculations utilized the GAFF force field, 42 AM1-BCC partial charges, and the TIP3P water model. The experimental HFE values have the mean and standard deviation as -3.82 and 4.84 kcal/mol, respectively. We have divided the total dataset into nine different groups based on the functional group or presence of a specific atom in the molecule. The eight groups are Alkanol, Alkanone, Alkene, Alkyl Alkanoate, Halo Alkane, Aromatic, Aliphatic cyclic, N-based Aliphatic, and the ninth, misc, is the group for molecules that do not come under any of the previous eight groups. We have assessed the performances of our models both for the whole dataset and these different groups. |
6719d97383f22e4214f8b821 | 3 | One of the primary objectives of this work is to utilize a minimal number of descriptors while ensuring they possess physical interpretability. To achieve this, we have used only six descriptors: polar surface area, hydrogen bond donors, hydrogen bond acceptors, the number of rotatable bonds, logP , and an electrostatic term which we call as the pol term (GB term summed with Coulomb electrostatic). The first five descriptors were calculated using the RDKit 46 package in Python, while the last descriptor is calculated as described below. The simplified polar energy is the sum of two terms, the Coulombic electrostatic energy, and a generalized Born energy term. The generalized Born energy term is calculated by the Generalized Born equation as follows |
6719d97383f22e4214f8b821 | 4 | where ϵ w is the dielectric constant of water (the process being moving a solute from vacuum to water), q i , and q j are the charges of atoms i and j. And f GB is a function, dependent on the distance between the atoms i and j, that interpolates between the distance r ij and the Born radii. Most widely used functional form of f GB is given below |
6719d97383f22e4214f8b821 | 5 | The charge and radius of atoms are taken from the Generalized Amber ForceField (GAFF) forcefield (water radius was taken as 1.4 Angstrom). Although, eq. ( ) is valid only for non-overlapping atoms, this can act as an excellent descriptor in an ML model. In the evaluation of eq. ( ), the numerator of the second term can be negative for overlapping accessible surfaces of the atoms. To circumvent this problem, we have performed the sum over the pairs of atoms with non-overlapping accessible surfaces only. This approximation provides a fast estimation of the polar part of the solvation free energy. Our results show that using this approximation as a descriptor, ML-based methods perform well for the FreeSolv database. We have employed four different machine learning models: Random Forest (RF), 47 Extreme Gradient Boosting (XGBoost), Gradient Boosting (GradBoost), and Light Gradient Boosting Machine (LightGBM). These models are trained on the training set to learn the crucial relationships of the descriptors with the HFE which is the target property. |
6719d97383f22e4214f8b821 | 6 | Although all the above four machine learning models-RF, XGBoost, GradBoost, and Light-GBM -use ensemble techniques, their approaches to prediction differ. Random Forest is a bagging-based technique that builds multiple decision trees independently, averaging their predictions to reduce variance and improve stability. Gradient Boosting applies a boosting strategy to further train the weak learners one after another in a sequence to minimize the loss function while correcting the mistakes of the preceding one. The other two methods -XGBoost, and LightGBM -are more advanced versions of the Gradient Boosting algorithm. |
6719d97383f22e4214f8b821 | 7 | XGBoost uses regularization, parallel processing, and tree-trimming methods to overcome the over-fitting. LightGBM, another variant of Gradient Boosting, was developed to handle large amounts of data, it uses leaf-wise growth and histogram-based learning to speed up and reduce memory usage. Collectively, these models use various techniques to merge decision trees to balance the prediction accuracy and computing time. We have trained and tested our ML models on two classes of data: 1) full data set, and 2) without outliers. We have used the interquartile range (IQR) method to define outliers in the experimental hydration-free energy, with bounds set at Q 1 -1.5 * IQR and Q 3 + 1.5 * IQR. Q1 and Q3 are the first and third quartiles, respectively. This leaves 628 number of molecules in the second dataset set. |
6719d97383f22e4214f8b821 | 8 | We have also utilized GridSearchCV with 5-fold cross-validation to optimize the hyperparameters of our model. This method involves systematically searching through a grid of hyperparameter values to identify the best settings for our model. In conjunction with 5-fold cross-validation, the dataset is divided into 5 folds. The model is trained on 4 folds for each hyperparameter combination and evaluated on the remaining fold. This process is repeated 5 times, each time using a different fold as the test set. We ensure that the selected hyperparameters provide robust and generalizable model performance by averaging the performance across these iterations. In Table in SI, we have listed the optimized parameters of the four models we have used. |
6719d97383f22e4214f8b821 | 9 | To evaluate the performance of our model, we employed several metrics: Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Pearson correlation coefficient (P r), and R 2 score. RMSE and MAE provide insights into the magnitude of prediction errors, while P r assesses the strength of the linear relationship between observed and predicted values, and R 2 indicates the proportion of variance explained by the model. These metrics were used to evaluate our model's accuracy and robustness rigorously, and the results were compared with those from other studies to benchmark our model's performance against existing methods. The feature importance for each descriptor was calculated using the mean decrease of impurity. |
6719d97383f22e4214f8b821 | 10 | The feature importances for the different models highlight the varying roles each descriptor plays in predicting the target variable. We have shown the feature importance for RF and XGBoost in Figure and for Gradient Boosting and LightGBM in Figure in SI. In both the Random Forest and Gradient Boosting models, the polar surface area (psa) emerges as the most important feature, contributing 50%, followed by the pol term, which accounts for approximately 30%. It is to be noted that polar surface area and non-polar surface area are complementary features. In this work, we have taken PSA; however, taking PSA as a feature implicitly includes non-polar surface area also. Hence, the importance of PSA as a feature indicates the importance of polarity of surface areas in general. These polarity of surface area and the pol term dominate the prediction capabilities of these models, suggesting that molecular surface properties play a crucial role in the prediction task. Other descriptors like the number of hydrogen bond donors (n donors), rotatable bonds (nrotb), acceptors (n acceptors), and logP contribute significantly less. This highlights a strong dependence on molecular polarity and surface area in these ensemble tree-based models. |
6719d97383f22e4214f8b821 | 11 | Interestingly, the XGBoost model demonstrates a different feature importance distribution, where the number of acceptors (n acceptors) becomes the most dominant feature, contributing 30% to the predictions. Pol term and psa play smaller but still significant roles, contributing around 18-22%. This indicates that the XGBoost model is more sensitive to hydrogen bond acceptor characteristics compared to the other models. LightGBM also highlights psa, logP , and pol term as the most critical features. These results suggest that while molecular surface and polarity remain crucial across models, each model places a different emphasis on these features based on their algorithmic structure. |
6719d97383f22e4214f8b821 | 12 | The RMSE and MAE for the ninth group i.e. misc (molecules not categorized in any of the previous eight groups) show the highest deviation in the prediction with their values of 0.86 and 0.54 kcal/mol respectively. But the correlation metrics i.e. R 2 and P r show different behaviour than the error metrics. The correlation metrics for this group (R 2 = 0.9 and P r = 0.95) indicate that this group's performance closely agrees with experimental hydration free energy. These two contradicting metrics show that there is a systematic error in the model both in the training and testing phases. The same contradicting trend is also observed in the case of aromatic group. Except for these two groups, our models perform well across different groups with relatively low RMSE (less than 0.53 kcal/mol) and low MAE (less than 0.36 kcal/mol). For the correlation metrics except for alkanone group, all other groups are highly correlated with their corresponding experimental hydration free energy. The R 2 is always more than 0.85 and P is always greater than 0.95 except for alkanone group which signifies the performance of our model across the groups. |
6719d97383f22e4214f8b821 | 13 | In this work, we have developed a physics-based and interpretable machine learning model for predicting hydration free energy of small molecules with only six descriptors. Our results compare well with other works with this dataset. However, the advantage of our method is that the results are fully interpretable, which is often the issue with the ML models. Our models perform well across different chemical groups signifying their applicability to larger databases such as those used in drug discoveries. |
634a5084de2a2174b2a60019 | 0 | Two decades after the seminal publications of List and Macmillan, it is now universally acknowledged that organocatalysts form the third pillar of asymmetric catalysis, along with enzymes and chiral organometallic complexes. Along with the tremendous spurt in the application of organocatalysts to a wide range of different transformations, there has also been a concerted effort to understand the mechanism(s) by which such asymmetric catalytic processes take place, both through experimental, as well as computational studies. This has helped improve the efficiency of the systems. Nevertheless, there are still some interesting aspects of organocatalytic systems that are yet to be completely understood. One such area is the use of additives: it has been observed that the presence of an additive leads to improvement in the yield, enantioselectivity, or both, in a wide variety of organocatalytic systems. Such additives include organic acids, chiral phosphoric acids, alcohols, water, amines, molecular sieves and salts. Of these, the most common are acids and alcohols. It has been hypothesized that their role in such systems is to increase the pKa of the reaction medium through their acidic nature. In some cases, they have been seen as proton shuttles. It has also been hypothesized in some instances that the additive could hydrogen bond with one of the substrates. However, researchers have wondered whether additives could have a more active, defined role in increasing the efficiency of organocatalytic systems. Such a possibility seems particularly pertinent when one considers the beneficial effect one additive class: that of aromatic additives, have played in organocatalytic systems. C-H activation via cross coupling between aryl iodides/bromides and the C-H bonds of arenes, mediated solely by the presence of 1,10-phenanthroline as catalyst in the presence of potassium tert-butoxide as a base, studied by Shi and coworkers, for instance, was seen to benefit from the presence of 1,10-phenathroline. Curran and Studer, commenting on this work, pointed out that the role played by 1,10-phenathroline as additive was unclear. There are several other interesting reports where aromatic alcohols such as phenol (substituted and unsubstituted) and naphthol, have been employed as additives, leading to improved yield and enantioselectivity. Chen and coworkers reported the α-thiocyanation of oxindole with the BzCPD catalyst, using 2-napthol as the additive. Feng and coworkers investigated the asymmetric nitroaldol reaction of αketophosphonates, employing 2,4-nitrophenol as the additive. Zhu, Masson and coworkers reported invertible enantioselectivity with the 6′-deoxy-6′-acylamino-beta-isocupredine catalyst in the asymmetric aza-Morita-Baylis-Hilman reaction, using β-napthol as the additive. Zhu et al. found that the addition of a small amount of β-naphthol in the aza-Morita-Baylis-Hillman reaction of imine and β-naphthyl acrylate led to improved yield and enantiomeric excess (ee). They considered the additive as a hydrogen bonding species. Chen and coworkers studied phenol-additive effects on the aza-MBH reaction of ketimines and acrolein: the addition of (R)or (S)-BINOL to the system as additive was found to dramatically improve the enantioselectivity. Fu and coworkers used 2-chloro-6-methylphenol as an additive in the phosphine-catalyzed double γ-addition of racemic heterocycles to racemic allenoates. Zanardiandcoworkersreportedswitchableregioselectivity in organocatalyticasymmetricadditiontoenals, using an aminecatalystandparanitrophenolasthe additive. Leitner andcoworkersreportedbifunctionalactivationandracemization in thecatalyticasymmetricaza-Morita-Baylis-Hilmanreaction, using PPh3 ascatalystand 3,5bis(CF3)phenolasthe additive. Ooi and coworkers have reported a supramolecular assembly system for the conjugate addition of azalactones to unsaturated acyl benzotriazoles, where three molecules of phenol were envisaged to participate through a hydrogen bonding network, This potential of the aromatic alcohols to stabilize both substrates leads to an important question: is it possible that the additive plays a previously unsuspected role in such systems, by explicitly participating in a cooperative fashion in the reaction, as a second catalyst? This is the question that has motivated the current computational investigation with density functional theory (DFT). |
634a5084de2a2174b2a60019 | 1 | To begin with, we have considered a variety of different organocatalytic systems, many of which have been referred to in the previous paragraph, and determined the stabilization that was gained by the complexation of both the electrophile and the nucleophile to the sandwiched additive, through hydrogen bonding and π•••π stacking interactions. Such stabilization was seen to be quite significant in most of the cases considered, indicating a general phenomenon when the aromatic additive is employed in such systems. This further hinted at a broader role of the additive, implying that such a stable configuration would then lead to increased enantioselectivity of the desired product. That this was the case was then proved for two different organocatalytic systems: the BZCPD cinchona alkaloid catalyst, employing the 2-naphthol additive for the alpha thiocyanation of oxindole, as well as for the asymmetric nitroaldol reaction of α-ketophosphonates using phenol,4-nitrophenol, 2,4-nitrophenol, and 2,4,6-trinitrophenol as the additives. |
634a5084de2a2174b2a60019 | 2 | 2 below, is an example of the same) have been found to be very effective in asymmetric organocatalysis. These systems are known to work though the stabilization of the electrophile and nucleophile through non-covalent interactions, which is the basis of the two standard mechanisms: a) the Wynberg ion-pair hydrogen bonding model and b) the Houk-Grayson hydrogen bonding model, which have been invoked to explain their activity. Our computational studies show that cinchona alkaloid stabilizes the substrates in the fashion of the Wynberg The asymmetric nitroaldol reaction of α-ketophosphonates, using phenol and nitro substituted derivatives of phenol as additives, reported by Feng and coworkers, has also been considered. |
634a5084de2a2174b2a60019 | 3 | The reason for choosing this system for study was because of the interesting, experimentally observed correlation of change in enantioselectivity with the change in additive, in this particular case. This provided us with an opportunity to validate the hypothesis presented in the current work with experiment, and our calculations show that indeed the role of the additive, as proposed here, provides perfect correlation with experiment. The current work, therefore, provides important new insights into the rapidly growing area of additive influenced asymmetric organocatalysis. |
634a5084de2a2174b2a60019 | 4 | One of the most influential characteristics of cinchona-based catalyst systems is their active conformation, because this decides the nature and extent of the stereoselectivity. The conformational space is rather small due to the rigidity and bulk of the substituents and gives rise to a restricted number of conformer populations. Wynberg and co-workers have done a Later, Bürgi and Baiker, through combined NMR experiments and ab initio calculations, revealed that the anti open conformation of the catalyst is the most stable in aprotic solvents. They also observed that in polar solvents, the syn closed and anti closed conformations were preferred over the anti open conformation. As the implicit solvents employed in the current computational study are dichloromethane and dichloroethane, a systematic conformational analysis was required at the outset in order to identify the most stable conformer. In order to carry out such an analysis, it was necessary to identify the principal torsional angles that distinguish the conformers from each other. Cinchona catalysts are rigid molecules containing relatively few rotatable bonds, and torsional angles around rotatable single bonds in the vicinity of the two chiral carbon atoms, namely C8 and C9 (see Supplementary(SI) Fig. ), are crucial structural parameters of the conformational space. Four torsional angles, modelled as α (N1-C8-C9-C4′), β (∠C8-C9-C4′-C3′), γ (∠C8-C9-O10-C9″), and δ (∠C9-O10-C9″-C10″) were seen to be the most relevant in determining the four different conformations that were possible for the rigid structure, and altering them gave rise to the four conformers. All the four conformers were investigated to find the most stable structure. A similar approach has been followed by Wong and co-workers in their computational study of the asymmetric methanolysis of meso-cyclic anhydrides through oxyanion hole stabilization. It is also worth mentioning that BzCPD has a flexible benzyl moiety at C9-O. It was, therefore, necessary to obtain a potential energy surface scan along the dihedral (C8-C9-O10-C11) to find the most stable orientation of the C9-O-benzyl group in BzCPD. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.