Edited by: Elizabeth B. Torres, Rutgers University, USA
Reviewed by: Sean B. Ostlund, University of California at Irvine School of Medicine, USA; Shinsuke Suzuki, California Institute of Technology, USA
*Correspondence: Henry H. Yin, Duke University, Genome Sciences Research Building II, 103 Research Drive, Box 91050, Durham, NC 27708, USA
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
We recorded activity of dopamine (DA) neurons in the substantia nigra pars compacta in unrestrained mice while monitoring their movements with video tracking. Our approach allows an unbiased examination of the continuous relationship between single unit activity and behavior. Although DA neurons show characteristic burst firing following cue or reward presentation, as previously reported, their activity can be explained by the representation of actual movement kinematics. Unlike neighboring pars reticulata GABAergic output neurons, which can represent vector components of position, DA neurons represent vector components of velocity or acceleration. We found neurons related to movements in four directions—up, down, left, right. For horizontal movements, there is significant lateralization of neurons: the left nigra contains more rightward neurons, whereas the right nigra contains more leftward neurons. The relationship between DA activity and movement kinematics was found on both appetitive trials using sucrose and aversive trials using air puff, showing that these neurons belong to a velocity control circuit that can be used for any number of purposes, whether to seek reward or to avoid harm. In support of this conclusion, mimicry of the phasic activation of DA neurons with selective optogenetic stimulation could also generate movements. Contrary to the popular hypothesis that DA neurons encode reward prediction errors, our results suggest that nigrostriatal DA plays an essential role in controlling the kinematics of voluntary movements. We hypothesize that DA signaling implements gain adjustment for adaptive transition control, and describe a new model of the basal ganglia (BG) in which DA functions to adjust the gain of the transition controller. This model has significant implications for our understanding of movement disorders implicating DA and the BG.
The role of dopamine (DA) in behavior has remained controversial despite decades of research (Cannon and Palmiter,
One limitation of previous work is that detailed movement parameters were rarely measured continuously and quantified. To obtain continuous measures of behavioral parameters, we recently began to combine video tracking and wireless
What is the role of DA in this circuit? There are massive nigrostriatal DA projections which synapse on the neck of the dendritic spines, the sites of glutamatergic corticostrial projections. We hypothesize that DA serves as gain modulation for the movement velocity controller in the BG, e.g., via the corticostriatal projections (Yin,
To test this hypothesis, we simultaneously measured movement kinematics as well as single unit activity from DA neurons in the substantia nigra pars compacta (SNc) in unrestrained mice (Fan et al.,
Eleven male C57BL6/J mice (25–35 g) were used in the electrophysiology experiments. Seven mice (two males, and five females) were used in the optogenetics experiments. All procedures were approved by the Duke University Institutional Animal Care and Use Committee. To make mice perform movements repeatedly, we gave them limited access to water. After the recording session each day, they had free access to water for 1 h. Each mouse received about 0.5 to 1 ml of 10% sucrose during the experimental session. When they had free access to water afterwards, they consumed ~2 ml. The health of all mice was monitored daily.
We used a simple Pavlovian trace conditioning task to study the phasic DA responses in relation to rewards (unconditional stimulus, US) and cues predicting rewards (conditional stimulus, CS). In this task, the mouse stands on an elevated platform (4 × 5 cm, elevated 40 cm) and its movement can be monitored with a camera facing it at 30 frames/s (Figure
Each trial began with the presentation of a tone (100 ms, 4 kHz, 21.6 dB) followed by the delivery of 13 μl 10% sucrose solution dispensed by a Valvelink 8.2 (AutoMate Scientific) and delivered through a spout fixed to the platform. The sucrose solution was delivered ~2 s after the termination of the tone. Each session contained 50–150 trials, with a variable inter-trial-interval of 20–50 s. Each session lasted approximately 1 h.
For sucrose/air puff sessions, either an ABAB design was used, in which reward trials and air puff trials were presented in blocks, or a AB design was used, in which sucrose trials were followed by air puff trials. The same auditory cue was used. Air puffs were 200 ms in duration and delivered from a computer-controlled 1500 series dispenser (EFD, 12 PSI).
Sixteen-channel electrode arrays (Innovative Neurophysiology) were lowered at the following stereotaxic coordinates in relation to bregma: 2.9–3 mm posterior, 1.2 mm lateral, and 4.6 mm below brain surface. Six mice were implanted in left nigra and the other six were implanted in the right nigra. The arrays consisted of 16 tungsten wires, 35 μm in diameter and 7 mm in length, arranged in a four by four configuration, attached to an Omnetics connector. Row spacing was 200 μm and electrode spacing was 150 μm. Electrode arrays were fixed to the skull with dental acrylic. Following the completion of the experiments, all mice were perfused and their brains sliced with a Vibratome and examined under a microscope to verify electrode placement.
The behavioral and electrophysiological data were recorded with a Cerebus data acquisition system (Blackrock) and analyzed with Matlab, Neuroexplorer, and Graphpad Prism. Single unit activity was recorded with miniaturized wireless headstages (Triangle BioSystems International), as described previously (Fan et al.,
Position, velocity, and acceleration are vector quantities with both magnitude and direction. For movement measured with 2D video tracking, this vector has two components (x and y). X and Y head position vectors were differentiated to get X and Y velocity, and the second derivative was taken to obtain acceleration. X and Y velocity and X and Y acceleration were then split into positive and negative components to yield a total of eight kinematic variables: up and down velocity and acceleration, left and right velocity and acceleration.
We then compared the neural activity to kinematic variables. To assess the correlation between neural activity between firing rate and kinematics, we analyzed data from the entire session. A complete record of neural activity and the continuous kinematic variables for each session was analyzed in Matlab with a bin size of 30 ms. The analysis consisted of two steps: first, for each session, cross-correlation was performed between the firing rate of each neuron and each of the eight kinematic variables to determine the shift required for the highest correlation between the two signals. Second, the neural signal and the kinematic signal were shifted accordingly and a Pearson correlation was then performed to determine the correlation between the two signals. Classification of different functional classes of neurons was determined by the strength of the correlation between the kinematic variable and neural activity (
By crossing
Custom-made optic fibers (5 mm length below ferrule, 105-μm core diameter, 1.25-mm-OD ceramic zirconia ferrule; Precision Fiber Products) were lowered into the brain and secured in place with dental acrylic and skull screws (Sparta et al.,
Photo-stimulation was always bilateral. A custom-made commutator was used to split a single laser beam into two beams for bilateral stimulation. During stimulation sessions, mice were connected to a 473-nm wavelength laser by two sheathed fibers (62-μm core diameter, connected by ceramic sleeves, Precision Fiber Products). The total output of the laser was adjusted each day, to obtain ~636 mW/mm2 transmittance.
Following completion of experiments, mice were anesthetized with isoflurane and perfused with ice-cold 4% paraformaldehyde. Brains were post fixed for ~24 h at 4°C, cryoprotected in sucrose solution, and then sliced at 60 μm on a Vibratome. Slices were incubated with primary chicken anti-GFP (1:1000, AbCam) and TH primary rabbit anti-TH (1:1000, Millipore) with 10% goat serum and 0.25% Triton X-100 overnight at 4°C. Secondary antibodies (Alexa Fluor 594 goat anti-rabbit and Alexa Fluor 488 Goat anti-Chicken) were used to visualize TH and GFP, respectively (1:250, Molecular Probes). Slices were imaged with an Axio Zoom.V16 (Zeiss) microscope and processed using Zen software (Zeiss).
The sucrose spout was located next to the platform on which the mouse stands. Following the cue, the mouse started a movement toward the sucrose spout, adjusting its body to prepare for the reward delivery. Once the sucrose solution was delivered, the mouse made another movement to consume the sucrose. Because the mice were not restrained and allowed to move freely within the confines of the small platform, the movement trajectories varied considerably between animals. Although we used a single LED placed on the head to track movements, this does not mean that the mouse only moved its head. There was clear movement of the whole body, as confirmed by pressure pads placed underneath the mouse, and video tracking of tail movements at the time of cue or reward (Figure
We recorded activity of DA neurons in the SNc, the largest DA cell group targeted by classic studies of phasic DA activity (Ljungberg et al.,
We observed phasic activity of DA neurons following the auditory cue and sucrose reward delivery (Figure
The firing rates of DA neurons at the time of cue and reward were correlated with movement velocity and acceleration. Because the movements were time-locked to the cue and reward delivery, it is important to dissociate kinematic variables from these task events. For this reason, rather than selecting only data from the trial, we performed an unbiased correlation between firing rate and the kinematic variables for the entire session, including inter-trial-intervals (Figure
This unbiased method, using cross-correlation of data from the entire session, was used to classify the neurons (Table
Up | 18 | 9 | 31 |
Down | 19 | 13 | 28 |
Left | 13 | 4 | 18 |
Right | 16 | 5 | 20 |
Total | 66 | 31 | 97 |
Cross-correlation analysis allows us to examine the direction specificity of the relationship between neural activity and movements. If a neuron is positively correlated with velocity in a particular direction, it will often show the opposite correlation with velocity in the opposite direction (Figure
Because we used an appetitive behavioral task, it is unclear whether the observed correlation between neural activity and acceleration is specific to reward-related behaviors, whether in reward anticipation or consumption. To address this question, we performed additional experiments with aversive outcomes. We measured the activity of DA neurons on both appetitive and aversive conditioning trials. The same experimental setup was used, with the same trace conditioning procedure, except an aversive air puff was delivered as the unconditional stimulus instead of sucrose solution, from the same location where sucrose was delivered (
Based on data from all recording sessions (appetitive as well as appetitive/aversive), we identified neurons that are correlated with velocity and acceleration in four directions (Table
The DA neurons with positive correlation with velocity or acceleration (Figure
The electrode arrays were always implanted unilaterally, in the left nigra in six mice, and in the right nigra in five mice. Of the movement-correlated neurons, 40 were recorded from the left nigra and 57 from the right nigra. Since many neurons recorded were correlated with either leftward or rightward movements, we examined the distribution of these neurons to see if there is any lateralization of leftward and rightward neurons. As shown in Figure
Although our electrophysiological experiments show striking correlations between movement kinematics and firing rates of DA neurons, such results do not tell us whether DA activity is directly involved in generating movement. To establish a “causal” role for DA neurons in movement, it would be necessary to manipulate the activity of these neurons while measuring movement kinematics. To selectively stimulate DA neurons, we developed a transgenic mouse line in which channelrhodopsin 2, which depolarizes neurons upon stimulation with blue light (Boyden et al.,
We used a blue laser (473 nm) to stimulate two groups of mice, one group with ChR2 expressed in Th-positive neurons (
As shown in Figure
We then varied stimulation frequency (each train lasted ~1 s) and examined the effects on movement kinematics (3
SNr GABA output neurons are known to inhibiti DA neurons. Some have argued that the firing of DA neurons is largely due to disinhibition: bursting is observed when GABA output neurons pause (Tepper and Lee,
Using a Pavlovian trace conditioning task, we first studied the responses of DA neurons following cue and reward presentation. Such responses are similar to what was reported in previous work on monkeys (Schultz,
A key difference between our study and previous work is our use of continuous video tracking to quantify the behavior of the animal, allowing us to examine the relationship between single unit activity and movement kinematics. As shown in Figure
To ascertain the “causal” role of DA neurons in movement, we also mimicked phasic DA activity by photo-stimulation. Using brief stimulation pulses and physiological frequencies (Pan et al.,
Our results therefore suggest that DA is critical for shaping the kinematics of movements, and support the hypothesis that nigrostriatal DA modulates the gain of a closed loop movement velocity controller (Yin,
First, we only examined the activity of nigrostriatal DA neurons in the SNc, which mainly target the dorsal striatum. We did not record from the mesolimbic DA neurons in the ventral tegmental area (VTA), which target the ventral striatum, prefrontal cortex, and other limbic regions. It is possible that VTA DA neurons have different properties, which allow them to encode reward prediction errors. Although we cannot rule out this possibility, classic studies from monkeys concluded that DA neurons from both SNc and VTA have similar phasic responses in relation to behavioral events (Schultz,
In our optogenetic experiments, the Cre driver line which we used for generating the
Pan and colleagues recently reported a subset of optogenetically identified SNr GABAergic neurons with short latency burst responses to salient cues during a trace conditioning task (Pan et al.,
Previous work found burst firing of DA neurons following reward delivery, but this phasic activity occurred earlier in time with training, following any reward predicting cue (Schultz,
Previous studies in support of this hypothesis rarely quantified movements, even though reward-related behavioral variables are not dissociated from movement kinematics. Moreover, most studies used restrained animals; even if they attempted to move by generating the requisite neural signals, it would not have been possible to achieve the actual movements. Inability to move does not mean that the neural signals necessary for movement velocity control are absent. Consequently, in all previous studies on reward prediction errors, there is an important movement confound.
Our results suggest that the previously observed shifts in phasic DA activity over time could reflect changes in movement kinematics. Of course, without using the identical experimental design, a direct comparison between our results and those from previous studies is impossible. Yet none of the standard manipulations (e.g., of reward size, probability, violations of reward prediction) used in previous studies can be free of the movement confound. In all these cases, the experimental manipulations can produce behavioral differences too subtle to be noticed by the casual observer. But we do not know how the animal is moving under the different experimental manipulations.
The reward prediction error hypothesis cannot explain the selectivity of DA neurons for different directions of motion (Figure
The correlation with vector components of velocity and acceleration observed in DA neurons is similar to what we observed in striatal neurons (Kim et al.,
Our finding of direction specificity in DA neurons and striatal neurons suggests considerable specificity in the pattern of nigrostriatal projections. DA neurons that fire during leftward movements, for example, are hypothesized to project to striatal neurons that fire during leftward movements. Unilateral stimulation of the striatum or application of DA agonists after depletion-induced receptor supersensitivity can produce contraversive turning or circling behavior (Ferrier,
Any point in space can be described in Cartesian coordinates, which indicate the direction and distance of this point from some arbitrarily chosen origin. In our model, the origin is the current position of the animal. From the egocentric reference frame, the position change is generated by the action of different classes of striatal neurons, but through a velocity control mechanism.
Movement direction is often assumed to be encoded by a population vector, which combines activity from many motor cortical neurons (Georgopoulos et al.,
The role of DA in scaling performance has long been noted and emphasized by another class of ideas, e.g., DA is thought to reduce sensorimotor threshold (White,
On the other hand, the present results support the model that the BG circuit acts as a transition controller, with DA acting as the gain of this system. Transition, in the sense used here, refers to
A major function of the sensorimotor BG circuit is to control transition in body configurations, which corresponds to movement velocity. This circuit controls the rate of change in proprioceptive transitions by sending descending reference signals to position controllers in the midbrain and diencephalon, which in turn command muscle length and joint angle controllers in the reticulospinal pathway (Yin,
The striatum contains comparators that compare reference and input velocity values, and the resulting velocity error signals are integrated and converted into the nigral output, which represents the reference signals for lower level position controllers.
It is critical to view the BG output as a continuously varying quantity, rather than either on or off. At any time, the output from SNr GABA neurons represents a specific body configuration and orientation. Constant firing rate is associated with a constant reference signal, i.e., no change in posture. When the firing rate does not change, the posture is fixed. To generate a voluntary action, the reference signal can be adjusted by a change in BG output from the SNr GABA neurons. For example, one change associated with steering of the body is a gaze shift, as the foveation target shifts to a different position. The descending command allows steering in any direction by shifting to a new target position. This allows acquisition of any sensory input with the corresponding changes in body configurations.
There are multiple position controllers, found in the tectum, pedunculopontine nucleus, and perhaps the thalamus, that receive descending commands from the BG output. Without a change in the descending BG output, lower level position controllers can independently control position, based on their own reference signals. They can produce outputs to resist any perturbation, even when there is no overt movement. Their outputs vary continuously according to the perturbation signals transmitted via the feedback path. A large output is due to a significant perturbation, as the input transiently deviates from that allowed by the internal reference. This is usually created by a salient stimulus (visual, auditory, or somatosensory), and the resulting behavioral response is traditionally labeled “reflexive.” So-called reflexes are behavioral outputs generated to restore the desired inputs dictated by lower level reference signals, for position, muscle length, tension, etc. Interestingly, artificially induced position error signals from deep layer tectal neurons can also activate DA neurons (Dommett et al.,
One important property of the different types of position controllers is that they will generate movement at a certain velocity, proportional to the position error. For a given position error, the velocity will not vary much. In other words, for behaviors generated at levels below the BG, velocity cannot be regulated independently. To change the velocity, one must change the rate of change in their reference signals, which dictates how quickly the position input should change. This requires a higher level system, a master loop in a cascade control hierarchy. It can be achieved by placing the velocity controller above the position controller, so that the output of the higher loop is the reference signal for the lower loop (
Ascending the hierarchy still further, we must ask what is represented by inputs to the striatum. According to our model, corticostriatal inputs reaching the sensorimotor comparator function in the striatum can either be perceptual representations or reference signals (
Many reference signals at the higher levels are acquired perceptual representations, i.e., memories. These signals represent behavioral goals, yet goals, in the sense used here, do not represent nor determine desired outputs. Instead they represent perceptual variables to be achieved.
To understand the hierarchical control model, it is important to grasp the fundamental difference between this model and traditional models that purport to use closed loop control (Todorov and Jordan,
Thus, according to the traditional view, feedback control means computing the commands (or “control signals”) needed to produce a particular position and then executing these commands. But this analysis is very misleading. Engineers measure the output of a negative feedback controller and compare the measurement with the desired reference value to generate the error signal needed to produce the output. The desired output is in fact some measure of the actual output. Such systems are known as servos, for without their own reference values they can never be autonomous agents. But the prescribed reference value in any living organism does not come from without, say a user adjusting the temperature setting or commanding a position, but from within (Powers et al.,
In the model described here, only perceptual inputs—sensed rate of change in body configurations and other transitions—are controlled. Outputs vary in proportion to the discrepancy between reference and input. Given the anatomical connections of the BG, we hypothesize that the instantaneous position coordinates represented by SNr GABAergic output neurons are reference signals sent to the comparators in various position controllers. These signals are not sensory or motor. Such a classification is meaningless in a closed loop controller, because the output is not driven by sensory inputs or by descending commands; it is a result of a comparison between the two (
In the control hierarchy only the lowest level has access to muscles. The higher order outputs from different controllers are used as descending signals that prescribe the desired amount of perceptual input for the immediately lower level. These are the orders sent to lower levels, not to command a specific output value, but to “request” a specific input value to be reached by the perceptual input functions of the lower level. The descending signals from the SNr GABAergic output neurons are not used to command muscle contractions, but to request specific values from specific position controllers in the tectum and pedunculopontine nucleus, which control orientation and body configuration.
We hypothesize that the sensorimotor striatum is organized as a topographically mapped modules, each dedicated to a specific velocity vector component corresponding to a movement direction (forward/backward, up/down, left/right). There are probably modules for different body parts, as different types of reference signals are generated for different controllers at lower levels (Carelli and West,
DA is hypothesized to serve as the gain in the transition controller. As a neuromodulator, DA does not directly cause firing of target neurons, though it can alter responsiveness of these neurons to inputs, e.g., glutamatergic corticostriatal inputs (Gerfen and Surmeier,
DA can have different effects on striatonigral (D1-expressing) and striatopallidal (D2-expressing) neurons giving rise to the direct and indirect pathways, respectively, but the ultimate effect is to modulate opponent output signals for downstream controllers (Figure
This hypothesis is supported by findings that neurons from both pathways are simultaneously activated during behavior (Cui et al.,
A major source of input is neighboring SNr GABA neurons, which inhibit DA neurons directly. These inhibitory projections come from collaterals of the GABAergic fibers which presumably synapse on other targets such as the tectum and ventral thalamus (Tepper and Lee,
Whereas most DA neurons are correlated with velocity and acceleration, GABA neurons are correlated with instantaneous position coordinates (Barter et al.,
A derivative of the BG output, then, can be fed back to the striatum. This could represent a mechanism for adaptive gain control, in which the gain can vary according to the movement velocity requirement. The gain of controllers for different body parts can also be independently adjusted. Haber and colleagues observed that the nigral output disinhibits a region of the SNc that send DA projections to the origin of the striatonigral projections as well as neighboring regions, forming a striatonigrostriatal loop (Haber et al.,
On the other hand, because the SNr GABA output (position) is a time integral of the striatal output (velocity), the GABA neurons possess cellular properties that implement leaky integration of their inputs (Barter et al.,
Because DA adjusts the gain of the velocity controller, it primarily affects the rate of change in perceptual variables, including those beyond the proprioceptive domain. This could explain common symptoms in various disorders—such as Parkinson's disease and Tourette syndrome—which are associated with abnormal DA levels. To put it simply, too much DA can result in faster movements as well as other types of perceptual transitions, whereas too little DA can result in a slowing down of the same transitions (Yin,
A reduction in DA reduces the magnitude of the velocity error signaled by the striatal projection neurons. Because there is a neural integrator in the SNr, the reduced rate of firing in striatal neurons results in reduced rate of change in body configurations. The total number of spikes from the striatal comparator can be reduced, so that movement amplitude is also reduced. In Parkinson's disease, which mainly involve degeneration of DA neurons in the SNc, both amplitude and speed of movements are reduced, which is why DeLong and colleagues proposed that BG output firing rate could encode both variables (DeLong,
Another implication is suggested by the possible differentiation of the GABA output performed by DA neurons, as a result of the disinhibition mechanism discussed above. Such velocity or acceleration feedback is often used in engineering to prevent oscillations and overshoots, by damping the controller in motion control. The lack of such feedback may explain common symptoms in Parkinson's disease, which is characterized by tremor and neural oscillations (Brown,
One implication of our model is that, in principle, the contribution of the DA signal can be replaced. When DA signaling is insufficient, as in Parkinson's disease, it may be possible to take the derivative from the GABA projection neurons to generate the needed DA signal. The challenge is to find the appropriate place where this signal can be injected. Clearly the striatum is the main candidate, but too little is known about the organization of the functional striatal modules to identify the suitable place in the circuit where more gain can be added. Moreover, to enhance the gain of such a system requires recording neural activity and delivering stimulation in real time based on the recorded data. This could also be challenging, especially in clinical practice.
In conventional studies attempting to relate neural activity with movement kinematics, the observed correlation between kinematic variables and single unit activity is very low (Paninski et al.,
By contrast, we have shown a strikingly linear relationship between neural activity and kinematics. But this is not due to any coding of movement to be used by a decoder. Rather the signals recorded are the signals used in circuits that generate the movements. Strictly speaking, there is neither encoding nor decoding of movement kinematics.
The control of movement velocity is only one type of transition control. The controlled variable is the perceptual input representing the rate of change in kinesthetic variables, e.g., rate of change in joint angle (Mountcastle et al.,
There are multiple parallel cortico-BG networks (Alexander et al.,
The term “reward” can be better defined as the control of some of these sensory representations that are sent to the striatum. The most immediate representation of the rate of food reward comes from controlled perceptual variables in behaviors like chewing and licking. The rate of change in these perceptual transitions can also be controlled in much the same way, though using a different set of effectors. Indeed, for orofacial consummatory behaviors, DA has been implicated in the control of reward rate (Rossi and Yin,
According to our model, what is often called “reward expectancy” can be viewed as an error signal in another higher-order reward rate controller. The reference for this controller is altered by changes in motivational state, i.e., satiety. Its error signal signals one more unit of the reward, just like marginal utility (Alchian and Allen,
Given the known movement deficits following degeneration of DA neurons, it is hardly surprising that phasic DA activity can be related to movement kinematics. What is surprising is how long it took for this relationship to be uncovered. The failure of previous studies to identify the role of DA in shaping movement kinematics could be explained by the lack of continuous behavioral measurements in unrestrained animals (Romo and Schultz,
There is a fundamental difference between our approach and that taken in most previous studies. The traditional measure consists of discrete time stamps that are supposed to represent specific events labeled by the experimenter. This type of measurement creates the impression that behavior comprises a series of pulse-like events, and neural activity is supposed to be found around the time of these events. Although this approach may provide the appearance of rigor, it also makes it all too easy for the experimenter to ignore any behavior that is not described by the task labels, which often reflect his own perception of the experimental events and theoretical biases.
Perhaps there is a deeper reason for the popularity of conventional behavioral analysis in neuroscience (Yin,
It has become customary in neuroscience to attribute behavioral variability to noise in sensory systems (Osborne et al.,
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
This research was supported by a NIH grant (AA021074) to HY and by NSF fellowships to JB and MR. We would like to thank Tatyana Sukharnikova and Stephen Castro for their help with experiments.