Graph neural networks for materials science and chemistry.

Machine learning plays an increasingly important role in many areas of chemistry and materials science, being used to predict materials properties, accelerate simulations, design new structures, and predict synthesis routes of new materials. Graph neural networks (GNNs) are one of the fastest growing classes of machine learning models. They are of particular relevance for chemistry and materials science, as they directly work on a graph or structural representation of molecules and materials and therefore have full access to all relevant information required to characterize materials. In this Review, we provide an overview of the basic principles of GNNs, widely used datasets, and state-of-the-art architectures, followed by a discussion of a wide range of recent applications of GNNs in chemistry and materials science, and concluding with a road-map for the further development and application of GNNs.

Automatic discovery of photoisomerization mechanisms with nanosecond machine learning photodynamics simulations.

Photochemical reactions are widely used by academic and industrial researchers to construct complex molecular architectures via mechanisms that often require harsh reaction conditions. Photodynamics simulations provide time-resolved snapshots of molecular excited-state structures required to understand and predict reactivities and chemoselectivities. Molecular excited-states are often nearly degenerate and require computationally intensive multiconfigurational quantum mechanical methods, especially at conical intersections. Non-adiabatic molecular dynamics require thousands of these computations per trajectory, which limits simulations to ∼1 picosecond for most organic photochemical reactions. Westermayr et al. recently introduced a neural-network-based method to accelerate the predictions of electronic properties and pushed the simulation limit to 1 ns for the model system, methylenimmonium cation (CH2NH2 +). We have adapted this methodology to develop the Python-based, Python Rapid Artificial Intelligence Ab Initio Molecular Dynamics (PyRAI2MD) software for the cis-trans isomerization of trans-hexafluoro-2-butene and the 4π-electrocyclic ring-closing of a norbornyl hexacyclodiene. We performed a 10 ns simulation for trans-hexafluoro-2-butene in just 2 days. The same simulation would take approximately 58 years with traditional multiconfigurational photodynamics simulations. We generated training data by combining Wigner sampling, geometrical interpolations, and short-time quantum chemical trajectories to adaptively sample sparse data regions along reaction coordinates. The final data set of the cis-trans isomerization and the 4π-electrocyclic ring-closing model has 6207 and 6267 data points, respectively. The training errors in energy using feedforward neural networks achieved chemical accuracy (0.023-0.032 eV). The neural network photodynamics simulations of trans-hexafluoro-2-butene agree with the quantum chemical calculations showing the formation of the cis-product and reactive carbene intermediate. The neural network trajectories of the norbornyl cyclohexadiene corroborate the low-yielding syn-product, which was absent in the quantum chemical trajectories, and revealed subsequent thermal reactions in 1 ns.

A simple dynamic model of respiratory pump.

To study the interaction of forces that produce chest wall motion, we propose a model based on the lever system of Hillman and Finucane (J Appl Physiol 63(3):951-961, 1987) and introduce some dynamic properties of the respiratory system. The passive elements (rib cage and abdomen) are considered as elastic compartments linked to the open air via a resistive tube, an image of airways. The respiratory muscles (active) force is applied to both compartments. Parameters of the model are identified in using experimental data of airflow signal measured by pneumotachography and rib cage and abdomen signals measured by respiratory inductive plethysmography on eleven healthy volunteers in five conditions: at rest and with four level of added loads. A breath by breath analysis showed, whatever the individual and the condition are, that there are several breaths on which the airflow simulated by our model is well fitted to the airflow measured by pneumotachography as estimated by a determination coefficient R(2) > or = 0.70. This very simple model may well represent the behaviour of the chest wall and thus may be useful to interpret the relative motion of rib cage and abdomen during quiet breathing.

Variability of end-expiratory lung volume in premature infants.

BACKGROUND: Analysis of breath-to-breath variability of respiratory characteristics provides information on the respiratory control. In infants, the control of end-expiratory lung volume (EELV) is active and complex, and it can be altered by respiratory disease. The pattern of EELV variability may reflect the behavior of this regulatory system. OBJECTIVES: We aimed to characterize EELV variability in premature infants, and to evaluate variability pattern changes associated with respiratory distress and ventilatory support. METHODS: EELV variations were recorded using inductance plethysmography in 18 infants (gestational age 30-33 weeks) during the first 10 days of life. An autocorrelation analysis was conducted to evaluate the 'EELV memory', i.e. the impact of the characteristics of one breath on the following breaths. RESULTS: In infants without respiratory symptoms, EELV variability was high, with large standard deviations of EELV. Autocorrelation was found to be significant until a median lag of 7 (interquartiles: 4-8) breaths. Autocorrelation was markedly prolonged in patients with respiratory distress or ventilatory support, with a higher number of breath lags with significant autocorrelation (p < 0.01) and higher autocorrelation coefficients (p < 0.05). Conventional assisted ventilation does not re-establish a healthy EELV profile and is associated with lower respiratory variability. CONCLUSIONS: In premature infants, EELV variability pattern is modified by respiratory distress with a prolonged 'EELV memory', which suggests a greater instability of the control of EELV.

Calibration of respiratory inductance plethysmograph in preterm infants with different respiratory conditions.

Respiratory inductance plethysmography (RIP) is a method for respiratory measurements particularly attractive in infants because it is noninvasive and it does not interfere with the airway. RIP calibration remains controversial in neonates, and is particularly difficult in infants with thoraco-abdominal asynchrony or with ventilatory assist. The objective of this study was to evaluate a new RIP calibration method in preterm infants either without respiratory disease, with thoraco-abdominal asynchrony, or with ventilatory support. This method is based on (i) a specifically adapted RIP jacket, (ii) the least squares method to estimate the volume/motion ribcage and abdominal coefficients, and (iii) an individualized filtering method that takes into account individual breathing pattern. The reference flow was recorded with a pneumotachograph. The accuracy of flow reconstruction using the new method was compared to the accuracy of three other calibration methods, with arbitrary fixed RIP coefficients or with coefficients determined according to qualitative diagnostic calibration method principle. Fifteen preterm neonates have been studied; gestational age was (mean +/- SD) 31.7 +/- 0.8 weeks; birth weight was 1,470 +/- 250 g. The respiratory flow determined with the new method had a goodness of fit at least equivalent to the other three methods in the entire group. Moreover, in unfavorable conditions--breathing asynchrony or ventilatory assist--the quality of fit was significantly higher than with the three other methods (P < 0.05, repeated measures ANOVA). Accuracy of tidal volume measurements was at least equivalent to the other methods, and the breath-by-breath differences with reference volumes were lower, although not significantly, than with the other methods. The goodness of fit of the reconstructed RIP flow with this new method--even in unfavorable respiratory conditions--provides a prerequisite for the study of flow pattern during the neonatal period.

Respiratory inductance plethysmography is suitable for voluntary hyperventilation test.

The aim of this work was to evaluate the goodness of fit of a signal issued of the respiratory inductance plethysmography (RIP) derivative to the airflow signal during rest, voluntary hyperventilation, and recovery. RIP derivative signal was filtered with an adjusted filter based on each subject airflow signal (pneumotachography). For each subject and for each condition (rest, voluntary hyperventilation, and recovery) comparisons were performed between the airflow signal and the RIP derivative signal filtered with an adjusted filter obtained either on rest signal or on the studied part of the signals (voluntary hyperventilation or recovery). Results show that the goodness of fit was : (1) higher than 90% at almost all comparisons (122 on 132), (2) not improved by applying an adjusted filter obtained on the studied part of the signals. These results suggest that RIP could be used for studying breathing during voluntary hyperventilation and recovery using adjusted filters obtained from comparison to airflow signal at rest.

Enhanced cardiac vagal efferent activity does not explain training-induced bradycardia.

Studies of heart rate variability (HRV) have so far produced contradictory evidence to support the common belief that endurance training enhances cardiac parasympathetic tone. This may be related to the fact that most studies failed to specifically isolate the vagally mediated influence of respiration. This study used a cross-sectional comparison of endurance athletes (n=20; ATHL) exhibiting resting bradycardia and age-matched nonathletes (n=12; CRTL) to indirectly assess training effects on amplitude and timing characteristics of respiratory sinus arrhythmia (RSA). Continuous electrocardiogram (ECG) and ventilatory flows were recorded during spontaneous breathing (SP), as well as during breathing at four cycles less than (M4) or more (P4) than SP, to also examine potential repercussions of training on the sensitivity of the cardiac vagal responses to breathing. A fast Fourier transform procedure was used to quantify the standard spectral high-frequency (HF) and low-frequency (LF) components and a respiratory-centered frequency (RCF) component of HRV. RSA was assessed using a breath-by-breath quantification of the amplitude and timing of the maximum change in instantaneous heart rate. Under baseline SP conditions, heart rate was lower in ATHL (62.6+/-6.5 vs. 75.2+/-9 beats/min; p<0.05) while blood pressure (BP), breath cycle duration, tidal volume, and ventilatory drive were similar in both groups. HRV total spectral power density, LF, HF, or RCF was not different between groups at either the SP, M4, or P4 conditions. Changes in total breath duration similarly affected RSA amplitude in all groups, while HR and BP remained unchanged from SP. RSA phase was not affected by training status or by changes in total breath duration. RSA amplitude was negatively related to breathing frequency in all groups (p<0.05), while the mean slope of the relationship (sensitivity) was not different between groups. In as much as RSA is an adequate marker of cardiac vagal efferent activity, these results add support to a contribution of a decrease in intrinsic heart rate to explain training-induced bradycardia.

Individual differences in respiratory sinus arrhythmia.

To investigate the interindividual differences in respiratory sinus arrhythmia (RSA), recordings of ventilation and electrocardiogram were obtained from 12 healthy subjects for five imposed breathing periods (T(TOT)) surrounding each individual's spontaneous breathing period. In addition to the spectral analysis of the R-R interval signal at each breathing period, RSA characteristics were quantified by using a breath-by-breath analysis where a sinusoid was fitted to the changes in instantaneous heart rate in each breath. The amplitude and phase (or delay = phase x T(TOT)) of this sinusoid were taken as the RSA characteristics for each breath. It was found that for each subject the RSA amplitude-T(TOT) relationship was linear, whereas the delay-T(TOT) relationship was parabolic. However, the parameters of these relationships differed between individuals. Linear correlation between the slopes of RSA amplitude versus T(TOT) regression lines and 1) mean breathing period and 2) mean R-R interval during spontaneous breathing were calculated. Only the correlation coefficient with breathing period was significantly different from zero, indicating that the longer the spontaneous breathing period the lesser the increase in RSA amplitude with increasing breathing period. Similarly, only the correlation coefficient between the curvature of the RSA delay-T(TOT) parabola and mean breathing period was significantly different from zero; the longer the spontaneous breathing period the larger the curvature of RSA delay. These results suggest that the changes in RSA characteristics induced by changing the breathing period may be explained partly by the spontaneous breathing period of each individual. Furthermore, a transfer function analysis performed on these data suggested interindividual differences in the autonomic modulation of the heart rate.

Cardioventilatory changes induced by mentally imaged rowing.

Mentally imaged but unexecuted physical activity has been reported to induce a cardiorespiratory change. In order to test whether the previous experience of the performed exercise was a prerequisite to observe these changes, ventilation and heart rate were recorded during mental imagination of a rowing race in four groups of volunteers: 12 competitive rowers, 10 non-rower athletes, 12 students (22-30 years old) and 12 senior subjects (50-60 years old). Recordings were performed at rest, during the viewing of a rowing race and during mental imagination of this race. Analysis of variance revealed significant condition effect for all cardiorespiratory variables. All subjects increased their breathing rate (mean increase: 16 breaths.min(-1) in rowers, 8 breaths.min(-1) in athletes, 8 breaths.min(-1) in students, and 6 breaths.min(-1) in seniors), 29 decreased their tidal volume (mean decrease: 100 ml in rowers, 102 ml in athletes, 120 ml in students and 26 ml in seniors), with an increase in the resulting ventilation in 38 subjects (mean increase: 14 l.min(-1) in rowers, 3.6 l.min(-1) in athletes, 2.8 l.min(-1) in students, 2.6 l.min(-1) in seniors). Heart rate was increased in 34 subjects (mean increase: 12 beats.min(-1) in rowers, 5 beats.min(-1) in athletes, 6 beats.min(-1) in students and 5 beats.min(-1) in seniors). The number of subjects who exhibited changes was evenly distributed among the four groups. However, mean values of the changes were higher in rowers than in the three other groups, mainly due to three rowers who exhibited extremely large increases in cardioventilatory variables. Analysis of variance showed no significant group effect for heart rate and breathing rate. These results suggest that rowing experience may not be necessary for changes in heart rate and ventilation to be elicited by mentally imagining a rowing race.

A program based on a 'selective' least-squares method for respiratory mechanics monitoring in ventilated patients.

This paper proposes a program for continuous estimation of respiratory mechanics parameters in ventilated patients. This program can be used with any ventilator providing airway pressure and flow signals without additional equipment. Overall breathing resistance, dynamic elastance (E) and positive end expiratory pressure (P(0)) are periodically estimated by multiple linear regression on selected parts of breathing cycles. Experimental validation together with justification of the selection procedure are based on signals obtained while ventilating a lung mechanical analogue with various intensive care ventilators. Clinical validity has been tested on 12 ventilated patients. The quality of estimation has been assessed by mean square difference between measured and reconstituted pressure (MSE), coefficient of determination (R(2)) and the condition number (a confidence index), and by comparison of E and P(0) with corresponding static values. The high R(2) and the low MSE obtained on most clinical cycles indicate that selected parts of cycles obey closely the model underlying parameter estimation. Agreement between static and dynamic parameters demonstrates the clinical validity of our program.

Non-invasive detection of respiratory muscles activity during assisted ventilation.

The instantaneous pressure applied by the respiratory muscles [Pmus(t)] of a patient under ventilatory support may be continuously assessed with the help of a model of the passive respiratory system updated cycle by cycle. Inspiratory activity (IA) is considered present when Pmus goes below a given threshold. In six patients, we compared IA with (i) inspiratory activity (IAref) obtained from esophageal pressure and diaphragmatic EMG and (ii) that (IAvent) detected by the ventilator. In any case, a ventilator support onset coincides with an IA onset but the opposite is not true. IA onset is always later than IAref beginning ((0.21 +/- 0.10 s) and IA end always precedes IAref end (0.46 +/- 0.16 s). These results clearly deteriorate when the model is not updated.