A fast intracortical brain–machine interface with patterned optogenetic feedback — Paper Summary
Summary #07. A fast intracortical brain–machine interface with patterned optogenetic feedback
📄 [ paper ]
⚡Briefly
The neuroengineers provided a proof-of-concept of a fast closed-loop brain-computer interface (BCI) in mice. They used a control signal from the motor cortex to control a virtual bar and stimulated sensory cortex if the bar “touched” a mouse. The mouse successfully learnt the behavioral task relying solely on artificial inputs and outputs.
⌛️ Prerequisites
The experimental setup included neurostimulation is done with optogenetics. If you are not yet familiar with what it is, check out our summary first: optogenetics
🚀 Motivation
Our body is a closed-loop system: we make actions via motor systems and acquire feedback using sensory signals. In neuroprosthetics one of the current issues is to provide appropriate sensory feedback to the brain to have a sense of touch, temperature and pain of a lost extremity while being able to move a prosthesis. Medical application are mostly concerned with interpreting pathological brain signals and providing appropriate stimulation — a closed-loop design (adaptive DBS, epilepsy closed-loop system, depression closed-loop link). Thus, closed-loop solutions, especially with low latencies, are essential to be developed further.
🔍 Research pipeline
Useful notions:
- vS1 — analogue of a sensory cortex in mice
- vM1 — analogue of a motor cortex
- Mouse whiskers — things like cats moustache, allowing a mouse to sense its environment. Mice rely heavily on touch information (Figure 1)
1) Prepare genetically modified mice
The special strain of mice was used, for which only pyramidal cortical neurons were genetically targeted to express opsins (ion-channels activated by light). As opposed to electrical stimulation, optogenetic stimulation is capable of targeting only a specific subset of neurons, what the researchers took advantage of.
For a deeper explanation please refer to the optogenetics summary.
2) Identify brain regions responsible for each whisker
The researchers did this by touching each whisker separately and measuring brain response via intrinsic imaging (Figure 2).
At the result, specific brain areas responsive for sensing the touch of a whisker were identified (Figure 3). This mapping can later be used to provide sensory input to the brain directly.
3) Implant electrodes into motor area
32 channel multisite electrode (Neuronexus, 8 tetrodes across 4 shanks) was implanted into vM1 to record brain output (Figure 4, left).
4) Link input and output into a closed-loop device
Now, having set up both recording and stimulation, it is possible to set up an autonomous device, reading activity from motor area and projecting to the sensory area (Figure 4, 5).
In a closed-loop output signals are converted into a control variable by an interpreter. It could be machine-learning based solution, simple threshold algorithm or a different means of extracting a meaningful signal. A feedback is generated on the basis of the control signal and fed to an agent. In turn, the agent could adjust its output on the basis of received input.
5) Use this design in the behavioral experiment 🔥
In the case of this study, control signal was an aggregated activity of hundreds of recorded neurons in motor cortex. Mean firing rate interpreted by a computer controlled a location of virtual bar. A bar was chosen, because it can move only by one 1D-axis, touching different columns of whiskers. If a virtual bar “touched” a virtual whisker, sensory feedback was provided to the mouse sensory cortex with optogenetic stimulation (Figure 6).
To test whether mouse could actually navigate in this environment with artificial inputs and outputs, researchers trained a mouse to receive reward only when a virtual bar touched certain whiskers (green range). The only way mouse could achieve it is by somehow increasing or decreasing firing rate of motor neurons. Note, that no actual bar touched the actual mouse whiskers! Nevertheless, mouse has successfully learned this task to receive reward (Figure 7).
It can be seen that without feedback (sense of touch), mouse can not longer perform a task, because it “didn’t know” when virtual bar touched its moustache, even thought the control variable was high enough.
📈 Experimental insights / Key takeaways
- With appropriate training it is possible to train an animal to operate a closed-loop system — if a system provides relevant feedback. The researchers provided a proof-of-concept that optogenetics can be used to deliver targeted stimuli meaningful to the brain.
- This setup operates at very low latencies, enhancing its ability to be interpretable by the brain (Figure 8).
- Motor and sensory cortices of the mouse (and humans) are adjacent, meaning only one compact skull chunk can be surgically removed to access both input and outputs areas of the brain
✏️ My Notes / Future prospects
Limitations & Drawbacks
- Mean firing rate of a neurons is a very simple control signal. With contemporary deep learning solutions more meaningful features can be extracted. Even conventional connectivity, clustering and pattern analysis methods can be applied. In 1D-paradigm, mean firing rate is enough, though.
- Also, several surgical procedures are required to get this setup to work. In some mice, electrophysiological recordings could not be established to have long-term stability. This can be overcomed with optogenetic imaging (Ca++) or using a bio-compatible electrodes, which are yet to be developed.
- The mouse was in fixed position during the experiment. It is not yet clear how to make this solution mobile.
- Lastly, for all of the closed-loop solutions, it is not yet clearly understandable, how exactly to stimulate to convey meaningful information into the brain. In case of mice, large brain areas are responsible for receiving inputs from the whiskers. However, for motor areas there is no strict topographical mapping. Neural code and neural representations must be cracked to further increase complexity of closed-loop BCIs.
Final thoughts
Overall, however, combining optogenetic stimulation, which is precise, fast and accurate, with electrophysiological readout (e.g. Neuropixel, Neuralink), could substantially enhance intracortical BCI abilities, if applied to humans.
For example, readout can be done from language production area and stimulation — to language comprehension zone. I personally believe reinforcement learning algorithms could even overcome the issue of neural code cracking. Not knowing the exact neural code underlying speech production and comprehension processes (or movement/sensing process), the task could still be accomplished by simultaneous collaborative brain and AI training.
😉 Stay tuned and see you in our medium publication
Author: Alexey Timchenko
Collaborator: Alexander Kovalev
Our telegram channel: the last neural cell
If you would like to use this material, please refer to the main “the last neural cell” publication or the author. A lot of work is done for creating concise summaries of interesting papers on a non-commercial basis.