Exploring the application of machine learning on the field of optogenetics
By Aizaz Chaudhry MB/ChB
Background: Optogenetics techniques are widely used to investigate functional brain networks, such as feeding networks. Optogenetics offer high specificity; controlling the activity of subpopulations of neurones within regions of the brain and brainstem. However, on its own, the technique provides no information on the global effects on feeding networks. The global effects of optogenetic stimulation of individual neurones can be measured using a technique called optogenetic functional magnetic resonance imaging (ofMRI). In addition to this, more evidence outlining application of machine learning techniques to fMRI data.
Method: We review the literature around optogenetic investigation of feeding networks; ofMRI and machine learning techniques applied to fMRI data to assess the potential application of machine learning techniques such as convoluted neural networks (CNNs) and hierarchical clustering to optogenetics and ofMRI data.
Conclusion: We argue that while techniques that require a large volume of high quality data like CNNs may be more difficult to apply to ofMRI data, hierarchical clustering can act as an aide to indicate regions of interest to follow up with ofMRI study.
Exploring the application of machine learning and optogenetics on feeding networks.
University of Birmingham
Optogenetics is a technique that is integral to the study of specific neurones in the brain to further our understanding about their role in functional brain networks. Their advantage is the specificity of activation, allowing researchers to activate subpopulations of neurones within a region of interest (ROI).
We argue that optogenetic experiments could be enhanced through the application of machine learning. We explore optogenetic fMRI imaging as a potential source of data to train machine learning algorithms for classification and exploration of feeding networks. We also look at the potential for deep learning and unsupervised machine learning as tool to be used alongside optogenetics.
Overview of optogenetics in feeding
Optogenetics involves researchers identifying the specific neurones within the brain region they wish to study. Then, they would insert a virus containing genes of both a fluorescent protein and a channelrhodopsin (ChR) above the region containing these neurones. Infection results in expression, which would then be determined through immunohistochemistry. Then, researchers would surgically insert a light optode above the region of interest (ROI). Stimulation would occur when light was shone through the electrode, resulting in the opening of the ChR, resulting in the depolarisation of the neurone. IN the case of neurones suspected in feeding behaviours, it would be expected to produce an observable alteration to confirm their involvement in said behaviours. (1–6)
This method is chosen because it provides researchers with the ability to activate specific neurones within an area of the brain. This is more important in the field of feeding behaviours, because two different sub-populations of neurones within the same area can mediate opposite effects, and still be involved in regulating feeding. One example of this, as shown by Aponte and colleagues(1), can be observed within the arcuate nucleus. Agouti related peptide (AGRP) neurones within the arcuate nucleus were observed to increase food intake and body weight in mice, whereas pro-opiomelanocortin (POMC) neurones in the same region had the opposite effect.
Using optogenetics, researchers studying the field of feeding behaviours in the brain can gain a better understanding of the precise function not just of each ROI, but the function of specific sub-populations of neurones within each ROI. In Table 1, we present a review of studies using optogenetic techniques to determine the function of neurone sub-populations in relation to feeding behaviour.
One key problem that we can identify immediately from Table 1 is that there are multiple areas of both brain and brainstem that mediate feeding and anorexic behaviours, but how or whether they inter-relate is not determined. This is a significant issue with the technique of optogenetics as a whole: while you do have sub-population specificity, you cannot measure wider network behaviours.
One example of this is the work by Cai and colleagues. Once they had determined that the input from a variety of anorexic signals were processed through PKC d+ neurones, they used a technique called monosynaptic retrograde tracing to determine the inputs to these cells. This technique fluorescently labelled regions activated by anorexic signals with inputs onto these cells (3).
What’s important to note is that a separate experimental technique was required to see which neurones had inputs into the ROI. Additionally, while these regions showed c-Fos expression (indicating increased activity), it was not in response to optogenetic regulation of PKC d+ neurones. Their technique reveals the retrograde pathway from the specific neurones, but gave no indication of any downstream effect of stimulation.
One method in the literature demonstrating a way to map anterograde projections is outlined by Petreanu and colleagues. They describe a method in which cells with expected projections into the contralateral brain are made to express both channelrhodopsin and mCherry fluorescent protein. They severed the axons, then looked for areas of the contralateral cortex expressing fluorescent protein. They then stimulated the axons and measured the electrical activity of neurones in the target area to confirm the existence of a connection.(7)
While Petreanu and colleagues demonstrate a way of determining anterior connection, its still has some deficiencies. First, its applied to areas with a density of fluorescent proteins. Areas with more diffuse connections were not subject to investigation, despite the possibility that those more diffuse projections may have an effect on the network. Second, this study looked at callosal projections involving the somatosensory cortex, which may not have as many connections as the feeding network within the limbic system.
Therefore, other techniques might be needed to get a better understanding of the functional nature of feeding behaviour upon stimulation. We explore the potential for optogenetic functional MRI imaging to fulfill this need later in this review.
Another issue with optogenetics specifically as it applies to feeding is the lack of specificity in the response produced. Feeding is a diverse behaviour, involving a balance between anorexic and hunger input, as well as the relationship between appetite and consummation. So, within the same subpopulation of neurones activated by optogenetics, you still might get a variety of responses.
For example, work undertaken by Jennings and colleagues identified within the same subset of lateral hypothalamus GABAergic neurones, different patterns of activation. Separate neurones were active during either consummation of reward after delivery for behaviour reinforcement or during appetitive behaviour requesting a reward.(5)
This result is most relevant to the underlying assumption made with optogenetics: that the subpopulation of neurones you implement with ChR proteins have homogenous function, therefore stimulation of these neurones produces a homogenous effect. The results from Jennings and colleagues indicate that further subdivision of function is still possible. We touch on how our proposed method may address this issue.
Optogenetic functional magnetic resonance imaging (ofMRI)
As we mentioned when discussing some limitations of optogenetic techniques, optogenetics doesn’t give us a comprehensive view of network involvement during stimulation. This view is advantageous, as it can indicate ROIs that are involved with the network that justify further study, furthering our understanding of that network.
One such technique that may be well suited to this task is optogenetic functional magnetic resonance imaging. This is a technique involving the optogenetic stimulation of channel rhodopsin expressing neurones in a live mouse, but with subsequent fMRI imaging of a cross section of the mouse brain during stimulation. This technique produces blood-oxygen-level-dependant (BOLD) signals.(8)
AN increase in BOLD signalling in the brain indicates an area of high metabolic activity, and is associated with neurone activation(8–10). Specifically, a high BOLD signal indicates an increased supply of blood to the region indicated, as its associated with the vasodilation of venules(11, 12).
BOLD signalling has been extensively used in fMRI imaging of the brain during optogenetic stimulation of a target subpopulation of neurones within an ROI. Voxels of fMRI images that show a statistically significant (assessed through Z score) increase or decrease in BOLD signalling during stimulation can be visualised on an MRI image of a rodents brain, indicating anatomical areas that may be involved in manifestation of behaviour stimulated by optogenetics.(13–15)
An additional application of this technique that gives it an advantage over traditional behavioral analysis is that it can be used to confirm the involvement of a network to its effects.
One example of this was regarding the findings of Drysdale and colleagues. They had used unsupervised hierarchical clustering machine learning techniques to find distinc neuroanatomical and clinical subtypes of depression based on diffuse tensor MRI images of depressed patients. One subtype was more heavily associated with anhedonia type symptoms, and was associated with medial prefrontal cortex (mPFC) activation(16). Ferenczi and colleagues then used ofMRI stimulation of the mPFC in awake mice. They found BOLD signal changes in areas of the brain associated with anhedonia symptoms(17). The work by Ferenczi and colleagues helped strengthen the case made by Drysdale and colleagues regarding the existence of this distinct subtype of depression.
The significance of this application is that it provides an important bridge between specific neuronal control provided by optogenetics and global effects on brain networks. IN this instance, it was a confirmation that the mPFC was involved in the network responsible for anhedonia.
But fMRI has another advantage when applied to optogenetic experiments. Chiefly, it provides two types of data about a global brain network. The first, as mentioned previously, is a heat map of significant BOLD signal change superimposed on an MRI image of a mouse brain. However, one can also derive linear temporal data about variation in BOLD signal change with time from stimulation.
The significance is that both data can be used to train machine learning algorithms in screening involvement of a brain region or even network in relation to a variety of tasks with suspected involvement in feeding behaviour. IN traditional optogenetics, the only data produced are electrical readings from a target region. This data has the advantage of confirming a direct relation from a stimulated region, as mentioned earlier, but cannot be used in further screening for network involvement, an application of machine learning we explore further in depth in the machine learning section of this review.
This method is not without faults. Most concerns surrounding ofMRI have to do with the lack of specificity of the response. This can best be exemplified by asking what is actually being represented by BOLD signal activation.
Some have called into question the reliability of BOLD signals themselves. Work by Christie and colleagues found that using high energies of light can elicit BOLD responses in the absence of ChR insertion, even in deceased mice. They determined that high light intensities resulted in a heating effect in the optic fibre. The temperature of the fibre was associated with BOLD responses (18). Further work by Schmid and colleagues confirmed that light above a 22 mW mm-2 threshold from fibre optic cables were associated with BOLD signals despite no expression of ChR proteins, due to heating effects. Schmid and colleagues also observed that visual pathways in the brain during optogenetic stimulation; an effect that did not disappear after stimulation was removed (19).
On the issue of heat-related BOLD artefacts, the work by Schmid and colleagues also provide a solution. For they find that these artefacts do not remain below the 22 mW mm-2 threshold, and that BOLD signals produced by light stimulation below this intensity is associated with neuronal activation(19). Therefore, if techniques maintain to keep intensity below this threshold, the results may still be valid.
On the issue of visual pathway stimulation, Schmid and colleagues suggest, based on the observation of constant exposure to light desensitises indirect stimulation of visual pathways, that background light should be allowed to enter the MRI machine(19). However, desensitisation of the visual pathway may have an adverse effect on visualisation of a brain network should it involve the visual pathway. Which means, should feeding networks involve then visual pathway, it would require additional measures to separate genuine involvement from visual artefact.
It is also important to note that work done by Takata and colleagues found that astrocytes metabolism may also factor into BOLD signalling. An increase in astrocyte activity could produce BOLD signal change, despite no corresponding change in neuronal metabolism(20).
However, the work undertaken by Takata and colleagues did not investigate whether optogenetic stimulation led to astrocyte stimulation. There remains the possibility that optogenetic stimulation of neurones within a region only results in neuronal activation. Even if it doesn’t, the application of ofMRI in study of feeding networks is done to identify additional ROIs that may be associated with the region under stimulation. BOLD signals produced in other regions of the brain would not be caused by astrocytes unless they were incorporated into the network, as axon projection to those regions would stimulate other neurones responsible for incorporating the signal. Therefore, while astrocyte activity may produce BOLD artifacts, this could be addressed by either only mapping statistically significant BOLD change, as is current practice, or by repeating the experiment amongst multiple models. The latter has an additional benefit of producing a more significant volume of data, which is advantageous for machine learning application, which we discuss in later sections.
Machine learning and its application to ofMRI
Machine learning are algorithms that are design to classify data. Specifically, they process input (called features) data to produce an output (called labels). Labels are usually in the form of classification designations(21).
There are two main types of machine learning: supervised and unsupervised. Unsupervised learning involves training algorithms with data that is not labelled, and includes techniques such as hierarchical clustering. Supervised techniques focus on developing classification algorithms from data that is labelled(21).
A set of machine learning techniques called deep learning algorithms have emerged as a method to classify more high dimensional data. Deep learning employs a series of “hidden layers” in between the features and the labels that process aspects of the data in certain ways to contribute to output classification. Ultimately, each label designated by the algorithm is influence by different weights within the “hidden layer”, that can influence outcome(22). One example of deep learning algorithm that is used to classify medical images (such as MRI imaging) is a convoluted neural network (CNN).
One of the advantages of ofMRI as we mentioned previously is that it produces high dimensional data not only in fMRI BOLD heatmaps, but also changes in BOLD signalling in response to time for optogenetic experiments. Whereas optogenetic experiments look at the electrical behaviour of individual neurones and regions suspected to have a connection, ofMRI provides global data that can be interpreted by machine learning algorithms. Machine learning algorithms trained on ofMRI data can be used for classification; subclassification of networks and improved accuracy of data.
IN this section, we expand on machine learning techniques that could be applied alongside optogenetic study of feeding networks in the brain. Of these, we explore convoluted neural networks and hierarchical clustering algorithms.
Convoluted neural networks (CNNs) are a technique that have risen in popularity. These are form of deep learning algorithms that have been trained and applied on medical imaging and have been used extensively on medical imaging data for classification purposes(23). Work by Zhao and colleagues demonstrated the effectiveness of CNN algorithms at classifying fMRI data based on which one of ten functional networks it belonged to. They trained the algorithm on 3754 sets of fMRI images from the researchers own labelled database (24). They found the accuracy of the CNN algorithm at classification of functional networks to be 94.61%(25).
The work by Zhao and colleagues demonstrates the ability of CNN algorithms to classify functional networks based on fMRI data with a high level of accuracy. The benefit of a highly accurate classification system is that it open the way for CNN algorithms to classify which functional network is active. This way, you could use optogenetic stimulation of neurones suspected to be involved with feeding to determine whether the feeding network is active.
However, a key issue with CNNs is that they require a large database for training. In the paper by Zhao and colleagues, they required 3754 individual fMRI scans from either task based or resting stimuli. These images were obtained from the Human Connectome Project and processed by the Zhao and colleagues(25). To replicate this specifically with feeding would require a large number of participants, which is both resource and time intensive. Work by Hu and colleagues shows that increasing the dataset with which the algorithm is trained increases the accuracy of the algorithm. When using 2000 samples, the accuracy of their CNN algorithm was 71.70%; at 5000 samples, it increased to 79.44%; and at 9950 samples it increased to 83.20%(26).
The application of CNN is unlikely to occur in any context without a large database of fMRI imaging data of feeding functional networks. The danger of applying a CNN network on small number of training samples is overfitting of data. This occurs when the algorithm over-estimates the influence of training set specific features that are poorly translated on other data, thereby reducing the algorithms effectiveness(27).
In addition to this, CNNs may not be the technique of choice if determining whether a network can be subdivided into smaller networks responsible for different stimuli. In feeding networks, as we’ve seen regarding the work by Jennings and colleagues around lateral hypothalamus GABAergic neurones(5), the same neurones may mediate different functions in different contexts. Work by Stachniak and colleagues(6), Nectow and colleagues(2), and Aponte and colleagues(1) has shown that distinct regions can execute opposite functions within feeding networks. CNNs are exceptional at classification of networks, but other machine learning techniques may be better to further understanding the behaviour of brain networks like feeding.
One technique that may be better equipped for understanding the components of an active feeding network in different contexts are hierarchical clustering algorithms. Hierarchical clustering is a form of unsupervised machine learning that seeks to classify images into clusters depending on shared characteristics.
An advantage of hierarchical clustering is the ability to discern distinct functions of the network that involve similar regions but execute different patterns of activity depending on the stimuli. One such application was the subdivision of depression by Drysdale and colleagues. IN their paper, they fed diffuse tensor MRI images of patients with depression into a hierarchical clustering algorithm. The algorithm clustered the patients into 4 sets, which they then found were associated with their own distinct symptom profile(16).
However, hierarchical clustering also suffers from an issue of data. Although its an unsupervised machine learning technique, and therefore doesn’t require labelling of data, it still requires a large amount of data to produce significant results. 711 patient resting state fMRI scans were used to train the hierarchical clustering algorithm used by Drysdale and colleagues(16), similar to the number of patients (n=995 patients for 83.2% accuracy) require to generate large data sets in the training of CNN algorithms by Hu and colleagues(26).
One application of machine learning we have chosen not to include is the application to improve imaging quality and reduce artefacts in imaging data. The reason for this is that this review is concerns with the direct application of machine learning to further understand feeding networks.
The question of how machine learning techniques can be applied to optogenetics remains. In our section exploring ofMRI, we allude to the high dimensional nature of its data. BOLD signalling allows for both spatial and temporal information. This would likely provide a lot of information for CNN algorithms, but would also increase the likelihood of overfitting without a substantial database of fMRI images to train the algorithm on, as mentioned by Angermueller and colleagues(27).
The difficulty with ofMRI is not the number of scans that can be extracted from one model, but rather the number of models itself. Optogenetics requires the surgical insertion of an optical electrode, which would be time consuming and expensive. if the number of models is to be around 500–1000 mouse models.
However, if such a database were generated, one application of the CNN algorithm is the classification of mouse models based on which areas of the feeding network is active. The work done by Zhao and colleagues demonstrated that CNN classification wasn’t just binary, but could classify network activity into 10 classification.
Another problem with CNN application lies in the type of data generated from ofMRI stimulation. Optogenetics involve the stimulation of individual neurones. While this is a strength if investigating roles of specific neuronal sub-populations in a ROI, training algorithms on global data from optogenetics may not reflect the behaviour of feeding networks as a whole. Therefore its possible the algorithm may not be applicable to ofMRI data from stimulation of other areas of the functional network, given they may produce different patterns of BOLD data.
Ultimately, its unlikely that CNN algorithms would be integrated with ofMRI experiments due to the number of individual models required to generate enough data to avoid the effects of overfitting, and the problem with training algorithms with ofMRI data.
However, the prospect of hierarchical learning integration may be more promising. We suggest using hierarchical clustering as a method to generate ROIs for followup ofMRI confirmation. One could apply hierarchical clustering to a set of fMRI images of mice exposed to a variety of stimuli associated with feeding (such as appetitive and consummatory stimuli) to determine if they could be clustered into subcategories based on patterns of network activation. We could then stimulate areas of the network associated with a cluster using optogenetics and measure the behaviour of the mouse as well as the pattern of BOLD signals in the brain to see if the clusters are significant, similar to the method of Ferenczi and colleagues when following up the work of Drysdale and colleagues (16, 17).
While hierarchical clustering has the same require of data size, the application we discuss above requires imaging mice exposed to feeding stimuli, with optogenetics applied to investigate ROIs associated with clusters. Without surgical intervention, animal models may not even be required for this step. Exposing humans to feeding stimuli, then following up with ofMRI investigation on mouse models may prove to be effective. Also by training the algorithm on fMRI of stimulus exposure rather than ofMRI data, the results are more generalizable to feeding networks.
IN conclusion, we argue that hierarchical clustering of fMRI data where subjects are exposed to stimuli suspected of feeding network activation is more useful to optogenetic investigations into feeding networks, as they provide information that more generalizable to the feeding network as a whole and it would be easier to generate a database of fMRI images from subjects.
1. Aponte Y, Atasoy D, Sternson SM. AGRP neurons are sufficient to orchestrate feeding behavior rapidly and without training. Nature Neuroscience. 2011;14(3):351–5.
2. Nectow AR, Schneeberger M, Zhang H, Field BC, Renier N, Azevedo E, et al. Identification of a Brainstem Circuit Controlling Feeding. Cell. 2017;170(3):429–42.e11.
3. Cai H, Haubensak W, Anthony TE, Anderson DJ. Central amygdala PKC-δ(+) neurons mediate the influence of multiple anorexigenic signals. Nature neuroscience. 2014;17(9):1240–8.
4. Carter ME, Soden ME, Zweifel LS, Palmiter RD. Genetic identification of a neural circuit that suppresses appetite. Nature. 2013;503(7474):111–4.
5. Jennings Joshua H, Ung Randall L, Resendez Shanna L, Stamatakis Alice M, Taylor Johnathon G, Huang J, et al. Visualizing Hypothalamic Network Dynamics for Appetitive and Consummatory Behaviors. Cell. 2015;160(3):516–27.
6. Stachniak TJ, Ghosh A, Sternson SM. Chemogenetic synaptic silencing of neural circuits localizes a hypothalamus→midbrain pathway for feeding behavior. Neuron. 2014;82(4):797–808.
7. Petreanu L, Huber D, Sobczyk A, Svoboda K. Channelrhodopsin-2–assisted circuit mapping of long-range callosal projections. Nature Neuroscience. 2007;10(5):663–8.
8. Lee JH, Durand R, Gradinaru V, Zhang F, Goshen I, Kim DS, et al. Global and local fMRI signals driven by neurons defined optogenetically by type and wiring. Nature. 2010;465(7299):788–92.
9. Kahn I, Knoblich U, Desai M, Bernstein J, Graybiel AM, Boyden ES, et al. Optogenetic drive of neocortical pyramidal neurons generates fMRI signals that are correlated with spiking activity. Brain Res. 2013;1511:33–45.
10. Schmid F, Wachsmuth L, Schwalm M, Prouvot PH, Jubal ER, Fois C, et al. Assessing sensory versus optogenetic network activation by combining (o)fMRI with optical Ca2+ recordings. J Cereb Blood Flow Metab. 2016;36(11):1885–900.
11. Chen X, Sobczak F, Chen Y, Jiang Y, Qian C, Lu Z, et al. Mapping optogenetically-driven single-vessel fMRI with concurrent neuronal calcium recordings in the rat hippocampus. Nat Commun. 2019;10(1):5239.
12. Yu X, He Y, Wang M, Merkle H, Dodd SJ, Silva AC, et al. Sensory and optogenetically driven single-vessel fMRI. Nat Methods. 2016;13(4):337–40.
13. Choe KY, Sanchez CF, Harris NG, Otis TS, Mathews PJ. Optogenetic fMRI and electrophysiological identification of region-specific connectivity between the cerebellar cortex and forebrain. Neuroimage. 2018;173:370–83.
14. Weitz AJ, Fang Z, Lee HJ, Fisher RS, Smith WC, Choy M, et al. Optogenetic fMRI reveals distinct, frequency-dependent networks recruited by dorsal and intermediate hippocampus stimulations. Neuroimage. 2015;107:229–41.
15. Palmer HS. Optogenetic fMRI sheds light on the neural basis of the BOLD signal. J Neurophysiol. 2010;104(4):1838–40.
16. Drysdale AT, Grosenick L, Downar J, Dunlop K, Mansouri F, Meng Y, et al. Resting-state connectivity biomarkers define neurophysiological subtypes of depression. Nature Medicine. 2017;23(1):28–38.
17. Ferenczi EA, Zalocusky KA, Liston C, Grosenick L, Warden MR, Amatya D, et al. Prefrontal cortical regulation of brainwide circuit dynamics and reward-related behavior. Science. 2016;351(6268):aac9698-aac.
18. Christie IN, Wells JA, Southern P, Marina N, Kasparov S, Gourine AV, et al. fMRI response to blue light delivery in the naïve brain: implications for combined optogenetic fMRI studies. Neuroimage. 2013;66:634–41.
19. Schmid F, Wachsmuth L, Albers F, Schwalm M, Stroh A, Faber C. True and apparent optogenetic BOLD fMRI signals. Magn Reson Med. 2017;77(1):126–36.
20. Takata N, Sugiura Y, Yoshida K, Koizumi M, Hiroshi N, Honda K, et al. Optogenetic astrocyte activation evokes BOLD fMRI response with oxygen consumption without neuronal activity modulation. Glia. 2018;66(9):2013–23.
21. Camacho DM, Collins KM, Powers RK, Costello JC, Collins JJ. Next-Generation Machine Learning for Biological Networks. Cell. 2018;173(7):1581–92.
22. LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015;521(7553):436–44.
23. Ching T, Himmelstein DS, Beaulieu-Jones BK, Kalinin AA, Do BT, Way GP, et al. Opportunities and obstacles for deep learning in biology and medicine. J R Soc Interface. 2018;15(141).
24. Lv J, Jiang X, Li X, Zhu D, Zhang S, Zhao S, et al. Holistic Atlases of Functional Networks and Interactions Reveal Reciprocal Organizational Architecture of Cortical Function. IEEE transactions on bio-medical engineering. 2014;62.
25. Zhao Y, Dong Q, Zhang S, Zhang W, Chen H, Jiang X, et al. Automatic Recognition of fMRI-Derived Functional Networks Using 3-D Convolutional Neural Networks. IEEE Transactions on Biomedical Engineering. 2018;65(9):1975–84.
26. Hu J, Kuang Y, Liao B, Cao L, Dong S, Li P. A Multichannel 2D Convolutional Neural Network Model for Task-Evoked fMRI Data Classification. Computational Intelligence and Neuroscience. 2019;2019:5065214.
27. Angermueller C, Pärnamaa T, Parts L, Stegle O. Deep learning for computational biology. Mol Syst Biol. 2016;12(7):878.